Datasets:

CShorten

/

CORD-19-prototype

like 0

The Trojan Chicken Study, Minnesota

We conducted a study in the summer of 2004 at county fairs in the Midwest to investigate the role poultry exhibits have in spreading avian pathogens to humans. A nearly invisible powder (pathogen surrogate) that fluoresces under UV light was surreptitiously sprinkled each day on 1 show bird at each of 2 fairs. A UV light box was used to daily examine the hands of 94 poultry-exhibit participants (blinded regarding UV box results) for up to 4 days during the poultry shows. Enrollment and end-of-study questionnaires collected data on pathogen risk factors. Eight (8.5%) of 94 participants had evidence of fluorescent powder contamination (95% confidence interval 2.76%–14.26%). This contamination and infrequent handwashing practices suggest that county fairs are a possible venue for animal-to-human pathogen transmission.

We conducted a study in the summer of 2004 at county fairs in the Midwest to investigate the role poultry exhibits have in spreading avian pathogens to humans. A nearly invisible powder (pathogen surrogate) that fluoresces under UV light was surreptitiously sprinkled each day on 1 show bird at each of 2 fairs. A UV light box was used to daily examine the hands of 94 poultry-exhibit participants (blinded regarding UV box results) for up to 4 days during the poultry shows. Enrollment and end-of-study questionnaires collected data on pathogen risk factors. Eight (8.5%) of 94 participants had evidence of fluorescent powder contamination (95% confidence interval 2.76%-14.26%). This contamination and infrequent handwashing practices suggest that county fairs are a possible venue for animal-tohuman pathogen transmission. R ecently, the Centers for Disease Control and Prevention (CDC) declared avian influenza to be the world's number-1 health threat (1); in particular, the wide and rapid spread of the H5N1 strain has heightened concerns. All H5N1 cases to date have been associated with direct contact with poultry, but recently, human-to-human transmission has been purported in Thailand (2) . Previously healthy children and young adults seem to be especially susceptible to this illness (3) . As of February 27, 2006, a total of 173 confirmed human cases of avian influenza A (H5N1) and 93 deaths have been reported to the World Health Organization, for a case-fatality rate of 53.8% (4) . Close contact with live poultry has been implicated in recent outbreaks of avian influenza in humans in Southeast Asia and elsewhere (2, (5) (6) (7) (8) . In the 1997 Hong Kong outbreak, live bird markets were implicated as the source of exposure to the virus (8) . In the United States, live bird markets are a known reservoir for avian influenza (9-11), but thus far they have not been associated with human avian influenza infection. Live bird markets involve a mixing of birds from diverse areas, crowded conditions for humans and livestock, mixing of different species of animal, and often a lack of proper sanitation, thus providing opportunity for outbreaks of disease. Transport of animals to market is a source of stress than can induce increased shedding of infectious agents. Stressed birds are also more susceptible to infections (12) . While live bird markets are uncommon in the Midwest, animal exhibits such as those at county fairs are quite common. Such exhibits are similar to live bird markets in that they involve transport and mixing of animals from different locations, crowded conditions, and a general lack of sanitation. Approximately 125 million people visit agricultural fairs every year in the United States (13) . Fairs usually involve close proximity of food vendors to animal exhibits. Many animal exhibits encourage or allow visitors to touch animals. Small children are frequent visitors to county fairs and animal exhibits, and children also engage in behavior such as nail biting that may make them more likely to ingest infectious agents. Live animal exhibits such as petting zoos and open farms, which are in many ways similar to county fairs, have also been implicated in outbreaks of Escherichia coli O157:H7 and other bacterial diseases (13, 14) . Proper handwashing is recommended to protect persons from infection (15) . However, animal exhibits often lack adequate handwashing facilities, and many persons may be unaware of the risk such exhibits pose. Direct contact with animals, indirect contact with contaminated objects, or inhalation of aerosolized virus could contribute to transmission of pathogens in such settings. Because little is known about the possible spread of pathogens at county fairs, and because most cases of avian influenza have resulted from close contact with poultry, a study was undertaken to model interspecies transmission of pathogens at county fair poultry shows. The specific aims of this study were to determine the proportion of The Trojan Chicken Study, Minnesota human poultry show participants who demonstrate hand contamination by a surrogate marker for an avian pathogen and to determine possible risk factors associated with such contamination. A feasibility study was conducted at a county fair in Iowa (county A) to evaluate study methods. After the feasibility study, human poultry fair participants were enrolled at a larger county fair in Minnesota (county B). Both fairs were held within small cities with populations of ≈100,000. At county fairs, poultry judging often takes place in show areas that are open to the public. Birds are usually placed in cages that are stacked one upon another and set upon tables ( Figure) . Because poultry classes are judged separately and competitors may show their birds in several poultry classes, birds are frequently moved in and out of their cages for grooming and competition. During the competition, birds are moved to competition cages that have previously housed birds from other competition classes. Judges typically handle each bird individually; they take the bird from the exhibitor, examine it, and then hand it back to the exhibitor ( Figure) . Handwashing is not generally performed as the judge moves from bird to bird, nor is handwashing common before or after exhibitors handle their birds. After competition, birds often remain on exhibit for several days, and they may be touched by the general public. This study was reviewed and approved by the University of Iowa's Institutional Review Board and Animal Use and Care Committee. The investigators participated in online human and animal subjects training. Informed consent was sought from participants before they were enrolled. Anyone >7 years of age present in the poultry exhibit area at any time during the period when poultry were on active exhibit was eligible to enroll in the study. Recruitment focused on members of 4-H clubs and openclass exhibitors, their families, and 4-H club staff, but also included other visitors. Enrollment occurred continuously over a 4-day period (Monday through Thursday) while poultry were exhibited at the fairs. A special sign and an information table were used to promote the study. Study participants were recruited for enrollment as they walked through the poultry exhibit area. After providing informed consent, study participants were asked to complete a 1page questionnaire that gathered demographic and poultry exposure data. Participants were also asked to complete a 1-page end-of-study questionnaire after they completed their experience at the poultry exhibit (day 4). This instrument gathered data on handwashing and types of animals handled at the fair. GloGerm (GloGerm Company, Moab, UT, USA), a benign, synthetic, organic colorant A-594-5 that fluoresces under a black light, was used as a surrogate marker for an avian pathogen. This powder (also found in liquid or gel form) is commonly used in handwashing training in hospitals and businesses (16) . Each day, the white powder was surreptitiously applied to the same single chicken at the fair to imitate a single source of pathogen. White broiler chickens were chosen as the exposure birds since the powder was not detectable on their feathers. Each "Trojan chicken" was otherwise treated the same as the other chickens in the poultry shows. While county fair authorities gave permission for the study, neither the judge nor the study participants were aware of neither the surrogate exposure nor which of the chickens were of particular hygienic concern. Instead several participants remarked that they thought the UV light box (see below) in which photographs were taken could somehow detect generic bacterial contamination on the hands. At county A, chicken powdering was conducted early in the mornings of the 3 days of competition, when competitors were not at the poultry exhibit. At county B, the same strategy was followed but the Trojan chicken was also surreptitiously powdered again in the early afternoon for 3 days of the show. During the powdering, approximately one-third cup powder was liberally sprinkled onto the underside of the chicken to imitate fecal shedding of pathogen. The chicken was then returned to its cage. The Trojan chickens each shared their cage with another, very similar, white broiler chicken, since these birds are normally shown in matched pairs. To evaluate potential avian influenza transmission, a 2 × 2 × 2-foot wooden box was constructed from plywood. Three black 1-foot × 18-inch fluorescent lights (15 watts) and 1 white 1-foot × 18-inch fluorescent light (15 watts) were mounted under the lid of this isolation box. Study participants inserted their hands through hand holes in 1 side, and they were blinded as to the result of the fluorescence examination of their hands. From an opening in the box on the opposite site, digital photographs of the ventral and dorsal images of the hands were taken with a digital camera (Figure) . A log was kept to match the sequentially captured photograph numbers with the participants' names (data were later de-identified). Beginning on day 1 of each poultry show, daily photographs were taken of study participants' hands under the black lights (Figure) . Photographs continued to be taken through the afternoon of day 4 (last day of the shows). Statistical analysis was performed with SAS version 8.0 (Cary, NC, USA). Chi-squared analysis and Fisher exact test were used to compare categorical variables with powder contamination. We used t tests to compare continuous variables. Logistic regression modeling was attempted, but the models did not converge. Odds ratios and confidence intervals were calculated by using EpiInfo (CDC, Atlanta, GA, USA) ( Table 1) . Ninety-four persons participated in the study by having their hands photographed. Among these were 30 poultry exhibitors ( Table 2) . Of the study participants, 82 (87.2%) completed the enrollment questionnaire, and 44 (46.8%) completed the end-of-study questionnaire. Of all participants in county B, 29 (30.9%) were male. The mean age of those who completed the enrollment questionnaire was 33 (range 7-79) years. Eighteen participants were poultry exhibitors, who showed 1-10 birds each (mean 3.4). Fifty-five participants (67.1%) of 82 were residents of farms. Eight participants exhibited hand contamination ( Table 1) . Of these, all 8 completed the enrollment questionnaire, and 7 completed the end-of-study questionnaire. Participant gender and hand contamination were not associated. Of participants whose hands were contaminated, 3 were male and 5 were female. None of the persons whose hands were contaminated were exhibitors: 3 were family members of exhibitors, 3 were visitors, and 1 was in the "other" category. In the age group of 7 to 12 years, 1 (7.7%) participant had hand contamination ( Table 1) . None of the participants in the 13-to 21-year age group showed hand contamination. Four participants (10.8%) in the 22-to 50-year age group had contaminated hands, and 2 participants (14.3%) who were >51 years showed hand contamination. Contamination rates did not differ by age group. Our study demonstrated that pathogen transmission is possible through poultry handling at county fairs. A contact transmission proportion of 8.5% (8 persons of the 94 participants had contaminated hands) is high, when one considers the insensitivity of the measure (gross fluorescence) and the number of persons possibly exposed at a county fairs. Both male and female participants were affected, as well as most age and role groups. This study had some unique characteristics. Digital photography of a fluorescent powder on hands was a successful surrogate for contamination. However, this rather gross measure was likely insensitive when one considers how few bacterial or viral particles are needed to cause certain zoonotic diseases. The black light box was also successful in blinding participants to their contamination status, since they were unable to see inside the box, and few seemed to grasp the experimental nature of the study. Some of our study findings were unanticipated. We expected contamination proportions to vary by age, gender, and role because we expected these factors to affect the amount of contact with birds and handwashing behavior. However the rates did not vary by these variables. This finding could be due to the study's limited power to detect such differences. If the differences between those exposed and those unexposed were statistically significant (e.g., also occurring in a similar study with a larger sample size), they might be consistent with studies that suggest that animal handlers (exhibitors) practice better hygiene compared to nonhandlers in the same environment. Alternatively, animal handlers may engage in other behavior that affects their contamination status, such as handling enough animals that the surrogate powder wears away more quickly than it would for someone who does not handle animals. This theoretical model had limitations. Hand contamination with the fluorescent powder was considered a surrogate for pathogen transmission in this study; however, hand contamination of a pathogen does not necessarily lead to transmission. Transmission is dependent upon the amount of inoculated pathogen (dose), the ability of the pathogen to cause disease (virulence), and the ability of the host to defend against infection (host susceptibility) (17) . These variables are complex and difficult to measure in settings such as a county fair. Additionally, such variables often vary by pathogen and host; hence, we measured only surrogate markers for exposure because such exposure is a requirement for disease to occur. GloGerm powder contamination may or may not be reflective of true pathogen transmission. The product is useful in handwashing training because it is generally not visible to the naked eye and persons are usually unaware that they have become contaminated. In our study, GloGerm was additionally useful because study participants were also unaware that a chicken was contaminated. Proper handwashing removes the powder, as it would pathogens. However, the amount of time the powder remains on a person's hands without handwashing varies and may be different from the amount of time that a pathogen would be viable on hands. In addition, dusting the chicken with powder is an attempt to model pathogen shedding, but this practice may not truly reflect the amount of pathogens an infected bird would shed. The undersides of the birds were dusted to model fecal shedding and dispersal of pathogens. However, the amount of powder used may be higher or lower than true pathogen shedding. The study design was further limited in that we did not account for time after exposure when photographs were taken. Since participants could drop by any time of the day, the time after exposure and duration of exposure likely varied between participants. In both the feasibility and pilot studies, the return rate was low, and tracking down participants was difficult. If similar studies are conducted in the future, a reward system might be used to increase compliance. Petting zoos and agricultural fairs are common in the Midwest and attract many thousands of people. While concern about viral and bacterial zoonotic disease transmission in these settings is growing, they are not usually thought of as a public health concern. The observations from this modest study, even with the limitations described above, suggest that live poultry exhibits may pose a disease transmission risk. Of particular concern is the relatively high proportion of powder transmission to poultry show visitors, who have casual and limited exposure to poultry. Larger future studies of similar design might help identify specific risk factors for zoonotic disease transmission and appropriate interventions for such settings. As a minimum contribution, these study data suggest that hygienic educational programs and disease prevention programs are warranted in poultry exhibits.

The Restriction of Zoonotic PERV Transmission by Human APOBEC3G

The human APOBEC3G protein is an innate anti-viral factor that can dominantly inhibit the replication of some endogenous and exogenous retroviruses. The prospects of purposefully harnessing such an anti-viral defense are under investigation. Here, long-term co-culture experiments were used to show that porcine endogenous retrovirus (PERV) transmission from pig to human cells is reduced to nearly undetectable levels by expressing human APOBEC3G in virus-producing pig kidney cells. Inhibition occurred by a deamination-independent mechanism, likely after particle production but before the virus could immortalize by integration into human genomic DNA. PERV inhibition did not require the DNA cytosine deaminase activity of APOBEC3G and, correspondingly, APOBEC3G-attributable hypermutations were not detected. In contrast, over-expression of the sole endogenous APOBEC3 protein of pigs failed to interfere significantly with PERV transmission. Together, these data constitute the first proof-of-principle demonstration that APOBEC3 proteins can be used to fortify the innate anti-viral defenses of cells to prevent the zoonotic transmission of an endogenous retrovirus. These studies suggest that human APOBEC3G-transgenic pigs will provide safer, PERV-less xenotransplantation resources and that analogous cross-species APOBEC3-dependent restriction strategies may be useful for thwarting other endogenous as well as exogenous retrovirus infections.

, bats [2] and chimpanzees [3] . Domesticated animals can also function as zoonotic intermediates (e.g., [4, 5] ). Additional and unprecedented opportunities for zoonoses occur when live cells, tissues or organs are transplanted from one species to another [6] . However, despite risks and technical and immunological challenges, several xenotransplantation procedures have shown preclinical promise for treating diabetes, heart, kidney and other human diseases (e.g., [7] [8] [9] ). Pigs are favourable xenotransplantation sources because of their human-like physiology, large litters, short gestation period and genetic malleability [10] . However, pig to human virus transmission has been a concern since it was shown that porcine endogenous retroviruses (PERVs) could infect human cells in culture [6, [11] [12] [13] . Although PERV transmission has yet to be documented in xenotransplantation patients, significant concerns still exist regarding PERV and other potentially pathogenic viruses [14, 15] . Strategies to reduce the likelihood of PERV transmission have been proposed, such as selective breeding for lower levels of PERV, RNAi transgenesis to knock-down PERV expression or systematic deletion of active PERV copies (e.g., [15, 16] ). The first two are unlikely to be completely effective or risk-free and the third, albeit theoretically feasible, may be overly technical and prohibitively expensive. Therefore, alternative, robust and costeffective methods to reduce PERV transmission and possible xenozoonotic infections are desirable. APOBEC3G is a single-strand DNA cytosine deaminase best understood as a potent inhibitor of HIV-1 replication [17] [18] [19] [20] [21] [22] [23] . It can however also inhibit a variety of other exogenous and endogenous retroviruses/elements (e.g., [17, 18, [24] [25] [26] [27] [28] [29] ). APO-BEC3G engages an assembling retrovirus particle, accesses the RNA genome-containing virus core and, upon reverse transcription, deaminates cDNA cytosines to uracils (C-to-U). Catastrophic levels of uracil either directly inactivate the coding capacity of the virus or trigger the degradation of the viral DNA. The former manifests as genomic strand-specific guanine to adenine (G-to-A) hypermutations (cDNA strand C-to-T transitions). However, in several instances, it is noteworthy that the deaminase activity of APOBEC3G or other APOBEC3 proteins is partly or even completely dispensable (e.g., HIV-1, hepatitis B virus, L1 and Alu [27] [28] [29] [30] [31] [32] ). Interestingly, throughout evolution, the retroviruses of many mammals appear to have become largely immune to APOBEC3G or to the APOBEC3G-like proteins of their hosts. HIV-1 expresses an accessory protein, Vif, which neutralizes APOBEC3G through ubiquitination and degradation [33] [34] [35] [36] . Simian immunodeficiency virus (SIV) uses a similar Vif-dependent mechanism [17, 26] . Foamy viruses employ an unrelated viral protein called Bet, for which the precise neutralization mechanism is currently unclear [37, 38] . Murine leukaemia virus (MLV) and human T-lymphotrophic virus-1 (HTLV-1) may simply avoid APOBEC3 proteins by preventing encapsidation [26, 39, 40] . However, cell-based studies have indicated that an APOBEC3 protein from a mammal to which the virus has not yet adapted may provide an effective strategy for thwarting species-specific viral counter-defenses. For instance, human APOBEC3G can potently inhibit the replication of SIV (except isolates such as SIV cpz which encode a Vif protein closely related to that of HIV-1), feline foamy virus and MLV (e.g., [17] [18] [19] [20] 25, 26, 39, 41, 42] ). Similarly, mouse APOBEC3 can potently block HIV-1 replication (regardless of Vif), although it is completely unable to impede the replication of MLV [26, 41] . One of the most dramatic examples to date used human and mouse APOBEC3 proteins to inhibit the mobilization of the yeast retrotransposons Ty1 and Ty2 [43, 44] . Thus, it is reasonable to hypothesize that cross-species expression of an APOBEC3 protein may be used to create a powerful barrier to impede or perhaps even block retrovirus infection. Here, this rationale is applied to the specific question of whether human APOBEC3G expression can inhibit the transmission of PERV from pig to human cells. The results demonstrate that PERV transmission can be strongly inhibited by APOBEC3G. To determine whether expression of human APOBEC3G would inhibit the transfer of PERV from pig to human cells it was first necessary to establish a long-term co-culture system. A trans-well assay was set up to monitor PERV transmission from pig kidney PK-15 fibroblasts to recipient human embryonic kidney 293T cells ( Figure 1A ). These two cell types were used because transmission from PK-15 to 293 cells had been reported previously [11] . The trans-well system enabled co-cultures to be sustained for several weeks, and it facilitated the recovery of each cell type for downstream analyses. An additional benefit of this co-culture system (not provided by transient assays) is that it enables the simultaneous analysis of multiple, endogenous PERV elements, which are precisely the targets one would want to monitor and ideally inhibit in (xeno)transplantation procedures. At each co-culture passage point, surplus human 293T cells were used to prepare genomic DNA. PERV transmission was monitored by subjecting these samples to quantitative (Q)-PCR. PERV-specific pol gene PCR products could be detected in the human genomic DNA samples after approximately two weeks of continuous co-culture ( Figure 1B) . From the point of first detection onward, the total number of PERV transmissions continued to increase, averaging 190 new events per day per 50,000 cells (10 5 beta-actin copies; SEM = 62; n = 5 experiments). Importantly, pig cells did not breach the 293T cell compartment because Q-PCR analyses of the same human genomic DNA samples failed to detect a concomitant transfer of pig genomic DNA ( Figure 1B ; also see Online Figure S1 for PERV pol gene Q-PCR standard curves, representative PERV-specific datasets and human beta-actin controls). Moreover, PERV copy number did not increase over a two week interval when infected 293T cells were grown in isolation, indicating that PERV was not replicating in the human cells and that the majority of the observed transmission events were derived from the PK-15 cells (i.e., new events). These results combine to indicate that the trans-well assay provided a robust system for monitoring bona fide zoonotic PERV transmissions. The second key step in addressing our experimental question was isolating PK-15 clones that stably expressed human APOBEC3G. Clones expressing human APOBEC3G cDNA or an empty vector control were established in parallel ( Figure 2A ). Immunoblotting identified clones with APOBEC3G levels similar to those in known APOBEC3G-expressing human T cell lines CEM and H9, which are non-permissive for growth of Vif-deficient HIV-1 [19] . Although it was impossible to achieve a physiologic expression level, the comparative immunoblot at least ensured that the levels of APOBEC3G were equal or lower than those present in wellstudied, non-permissive human T cell lines. Co-culture experiments were set up to compare PERV transmission from two independently derived PK-15 clones expressing human APOBEC3G and two vector expressing controls. Remarkably, the human APOBEC3G expressing PK-15 clones showed levels of PERV transmission that were lower than the Q-PCR detection threshold of approximately 10 copies ( Figure 2B ; . This graph summarizes data for two independently derived PK-15 clones, V1 (squares) and V2 (diamonds). All data points were calculated using results from duplicate Q-PCR reactions of genomic DNA from three parallel (but independent) co-cultures. The error bars indicate the SEM. See the Materials and Methods and Online Figure S1 for additional details, representative raw data and controls. doi:10.1371/journal.pone.0000893.g001 Online Figure S1 ). In contrast, the control clones showed high levels of PERV transfer by co-culture day 17 and transmission events continued to accumulate through the duration of the experiment. The kinetics of PERV transmission were similar to those reported in Figure 1B (these results also contributed to the aforementioned transfer rate calculations). These data were further corroborated by additional experiments where PERV transmission was monitored simultaneously by Q-PCR and by reverse-transcriptase ELISA assays (Online Figure S2 ). An ultimate application of the technology described here raises the potential problem that human APOBEC3G may not be subjected to (proper) post-translational regulation in pig cells and it may therefore promote carcinogenesis. Expression of human APOBEC3G in a heterologous system has been shown to trigger elevated levels of genomic C/G-to-T/A transition mutation [44] . Therefore, to help mitigate this risk (in addition to establishing clones that expressed relatively modest APOBEC3G levels; above), we asked whether the predominantly cytoplasmic localization pattern of human APO-BEC3G would be maintained in PK-15 and in a swine testes cell line, ST-IOWA ( Figure 3 ; compare with other APOBEC3G reports [18, 35, 41] ). Unlike some other APOBEC3 proteins such as human APOBEC3B, which is mostly nuclear, both human APOBEC3G and pig APOBEC3F appeared predominantly cytoplasmic in either the human or the pig cell lines (Figure 3 ; [28, 31, 41] ). Both proteins also appeared to concentrate in cytoplasmic punctae, which varied in number and were apparent in some of the cells regardless of species of origin (described previously for APOBEC3G; e.g., [45, 46] ). Overall, these near-identical localization patterns suggested that human APOBEC3G is not aberrantly regulated in pig cells and, interestingly, that these proteins might be subjected to the same cellular regulatory mechanism(s). During the course of these experiments, we reported some of the activities of pig APOBEC3F [41] . It could strongly inhibit the replication of HIV (regardless of Vif) and modestly inhibit the replication of MLV, a gamma-retrovirus phylogenetically related to PERV. Therefore, we wondered whether pig APOBEC3F was expressed in PK-15 and, if so, whether PERV resists this cellular defense. To begin to address this possibility, RT-PCR was used to test PK-15 cells for pig APOBEC3F expression. Pig APOBEC3F mRNA was detected readily (Online Figure S3A ). Full cDNA sequencing revealed that the predicted APOBEC3F protein of PK-15 cells was 98% identical to the variant we reported previously [41] . Eight amino acid differences were found, but both the PK15 and the previously reported APOBEC3F sequences were represented in pig genomic DNA sequences suggesting that these may be breed-specific polymorphisms (R.S.L. and R.S.H., manuscript in preparation). These observations indicated that either PERV resists the endogenous APOBEC3F protein of its host or that the level of suppression by pig APOBEC3F is not sufficient to inhibit PERV transmission. To begin to distinguish between these two hypotheses, PK-15 clones over-expressing pig APOBEC3F were established and used . PK-15 and 293T cell lysates were used as negative controls. CEM and H9 were used as positive controls for APOBEC3G expression. A non-specific (but pan-species) band is shown as a protein loading control (marked by an asterisk). (B) Q-PCR data using genomic DNA prepared from 293T cells co-cultured with two independently derived APOBEC3G expressing PK-15 clones (G1 and G2, circles and triangles, respectively) or two vector control clones (V1 and V2, diamonds and squares, respectively). The experimental parameters are identical to those used in Figure 1B in transmission experiments. The former hypothesis was favored because pig APOBEC3F over-expression did not significantly interfere with PERV transmission (Online Figure S3B ). These results were further supported by PCR experiments showing that PERV could be amplified readily from 293T cells that had been co-cultured with PK-15 over-expressing pig APOBEC3F (unlike the APOBEC3G scenario; below). We further noted that it is highly unlikely that another resident APOBEC3 protein contributes to PERV restriction, because genomic DNA sequencing showed that pigs have only one APOBEC3 gene, APOBEC3F (R.S.L. and R.S.H., manuscript in preparation). These observations combined to indicate that PERV is resistant to the endogenous APOBEC3 protein of its host. In hindsight, this was not particularly surprising given the emerging trend that (successful) retroviruses are selected in part by their ability to evade the APOBEC3 proteins of their host species (see Introduction). The hallmark of APOBEC3G-dependent retrovirus restriction is plus-strand G-to-A hypermutation, which is caused by the deamination of minus-strand cDNA C-to-U during reverse transcription [17, 18, 20, 24, 25] . The deamination of cytosines within singlestrand DNA requires glutamate 259 (E259) of APOBEC3G [47] [48] [49] . Based on homology to structurally defined deaminases, E259 likely functions by helping position the water molecule that ultimately initiates the deamination reaction by attacking the cytosine ring (as a hydroxide; reviewed by [17, 23, 50, 51] ). To determine whether DNA deamination is required the APOBEC3G-dependent inhibition of PERV transmission, we established a new set of PK-15 clones expressing APOBEC3G, APOBEC3G E259Q or a vector control (Figure 4, inset) . Surprisingly, both APOBEC3G and the E259Q derivative diminished PERV transmission to near background levels ( Figure 4) . These data demonstrated that the mechanism of inhibition does not require the DNA deaminase activity of APOBEC3G. These data were further supported by the fact that plus strand G-to-A hypermutations were not apparent in the DNA of the rare transmission events that occurred in the presence of human APOBEC3G (below). To begin to genotype the infectious PERVs and to further probe the mechanism of PERV restriction by human APOBEC3G, the PERV pol gene DNA was amplified from human 293T cells, cloned and sequenced ( Figure 5A ; Online Figure S4 ). Twenty-nine and twentytwo sequences were analysed from APOBEC3G and control experiments, respectively. To minimize possible PCR biases, any sequence that was recovered multiple times was considered one event, unless it arose from independent experiments. These DNA sequence analyses revealed several important points. First, in contrast to vector control and pig APOBEC3F overexpressing co-cultures, PERV pol gene DNA was difficult to amplify from the genomic DNA of 293T cells that had been cocultured with APOBEC3G-expressing PK-15 cells ( Figure 5B and every significant sampling point in our Q-PCR experiments). Taking this together with the observation that APOBEC3G does not effect PK-15 virus production (similar RT levels were observed in cell-free supernatants in the presence or absence of APO-BEC3G; data not shown), we infer that APOBEC3G restricts PERV transmission after virus production but before provirus integration (i.e., between entry and integration). APOBEC3G may restrict PERV at an early reverse transcription stage, possibly by interfering with primer binding, DNA synthesis and/or integration Is Deamination-Independent. PERV-specific Q-PCR data using genomic DNA prepared from 293T cells co-cultured with PK-15 clones expressing APOBEC3G (G; triangles), APOBEC3G-E259Q (GE259Q; circles) or empty vector (V; squares). Two datasets, each with an independent PK-15 clone in three replica co-culture wells, were collected in parallel and averaged for each data point. One standard error of the mean is shown. The experimental parameters are identical to those used in Figure 1B as shown recently for APOBEC3G and HIV-1 substrates (e.g., [30, [52] [53] [54] [55] ). Second, control co-culture PERV transmission events were exceptionally diverse, as 11 unique pol sequences were detected and only 4 were found multiple times (Online Figure S4) . These data suggested that PK-15 cells have at least 11 active PERVs capable of infecting human 293T cells, a number consistent with previous studies that reported the existence of approximately 17-50 PERV copies in total (with only a fraction being replicationcompetent; [11, 12] ). In contrast, the rare PERV sequences derived from the APOBEC3G co-culture experiments showed a much lower genetic complexity. Only three unique sequences were recovered, each differing by a single nucleotide (Online Figure S4) . In parallel experiments with HIV-based viruses, this APOBEC3G expression construct caused approximately 30 G-to-A hypermutations per 1000 bases analyzed (e.g., [41] ). Thus, approximately 12 G-to-A transitions should have been recovered in these PERV DNA analyses (nearly 90 if multiply recovered sequences would have been considered). The absence of hypermutated PERV proviral DNA provided further support for a deaminase-independent mechanism of restriction, which may share features with other instances described previously (e.g., [27] [28] [29] [30] ). We have established a quantitative assay to monitor the zoonotic transmission of PERV to human 293T cells. Expression of human APOBEC3G in the pig PK-15 cell line strongly inhibited PERV zoonoses, while the endogenous APOBEC3F protein of pigs appeared considerably less effective. These data are the first to show that human APOBEC3G can inhibit PERV and the first to demonstrate that APOBEC3 proteins can be used purposefully to reduce if not prevent zoonotic retroviral infections. These results were not anticipated because human APOBEC3G has a relatively weak effect against the PERV-related gamma-retrovirus MLV [26, 39] . Our data indicate that the engineering of pigs to express human APOBEC3G (or an equally potent non-porcine APOBEC3) may result in animals whose cells and tissues are much less likely to disseminate functional PERV. The deamination-independence of the restriction mechanism suggests that a catalytically inert APOBEC3G protein, such as E259Q, may be equally potent and simultaneously reduce the risk of cancer-promoting mutagenesis. APOBEC3G or APOBEC3G-E295Q expressing pigs may therefore constitute safer source animals for pig-to-human xenotransplantation procedures. In contrast to knockdown, knockout (by gene targeting or selective breeding) or most chemical-based anti-viral approaches to neutralize PERV [15] , the APOBEC3 antiviral defense system has several advantages including a potentially broad neutralizing activity (effective against PERV and likely several other endogenous and exogenous viruses) and an applicability to situations where many copies of a virus are already present in a genome. Analogous transgenic applications can be envisaged, such as using cross-species APOBEC3 expression to purposefully impede known viruses (e.g., the AIDS virus HIV-1 or the Hepatitis B virus HBV). Moreover, for humans and other mammals with multiple APOBEC3 proteins, our data encourage the development of methods to induce/up-regulate endogenous APOBEC3 proteins, which have the capacity but may not normally restrict a particular virus (e.g., human APOBEC3B and HIV-1). The porcine kidney PK-15 fibroblast cell line and the swine testes ST-IOWA cell line were obtained from the ATCC and cultured in Dulbecco's modified Eagle's medium (Invitrogen) supplemented with 10% fetal bovine serum (Gemini), and 25 units/ml penicillin and 25 mg/ml streptomycin at 37uC and 5% CO 2 . Human embryonic kidney 293T and HeLa (A. Bielinsky, University of Minnesota) cell lines were grown under the same conditions. The T cell lines H9 and CEM (M. Malim, Kings College London) were cultured in RPMI-1640 supplemented with 10% fetal bovine serum (Gemini), and 25 units/ml penicillin and 25 mg/ml streptomycin at 37 uC and 5% CO 2 . Plasmids encoding human APOBEC3G, human APOBEC3G-E259Q and porcine APO-BEC3F were described previously [41] . The human APOBEC3G and porcine APOBEC3F cDNA sequences used here are identical to GenBank accession numbers, NM_021822 and NM_001097446, respectively. Stable APOBEC3G-or vector control-expressing PK-15 cell lines were constructed by transfection using FuGENE6 according to the manufacturer's protocol (Roche) or by electroporation (BioRad, 250V, 950 mFa). Clones were selected using growth medium containing 1 mg/ml G418 (Roche), and APOBEC3G expressing clones were identified by immunoblotting using a polyclonal antibody toward human APOBEC3G (J. Lingappa, University of Washington). All PK-15 clones were maintained in growth medium supplemented with 250 mg/ml G418 to ensure stable expression. Long-term co-culture assays were performed in 6 well tissue culture plates with inserts (TranswellH, Corning Inc.). This system uses a membrane with 0.4 mM diameter pores, which keeps the two cell types separated physically but simultaneously allows diffusion of nutrients and small molecules including virus particles of approximately 0.1 mM (including PERV). Each experiment was initiated with 75,000 PK-15 cells (insert) and 75,000 293T cells (well) as illustrated ( Figure 1A ). At 72 hr intervals, each cell type was washed with PBS, subjected to mild trypsinization and diluted into 4 parts fresh growth medium. Excess 293T cells were used to prepare genomic DNA (Qiagen DNeasy kit). The rate of PERV transfer was calculated using the pol gene levels from the last two data points (usually spanning a 3 day period). The difference between these levels represents PERV pol gene DNA that has accumulated per 100,000 human beta-actin gene copies (50,000 cells assuming that the 293T cell line has two beta-actin copies) per time period. Individual rates from 5 independent experiments were averaged to determine the overall transmission rate (190+/262 events per day per 50,000 cells). Data from Figures 1B, 2B and S2A contributed to rate calculations. Genomic DNA was isolated from human 293T cells using the DNeasy kit (Qiagen). Duplicate 25 ml PCR reactions consisting of 10 ng of 293T genomic DNA, 100 nM primers and 26 iQ SYBR Green super mix (BioRad) were run on an iCycler iQ Multicolor Real-Time PCR detection System (BioRad). The thermocycler conditions consisted of an initial denaturation of 95uC for 5 min and 50 cycles of denaturation (95uC for 15 sec) and annealing (58uC for 30 sec). After the 50 cycles, a melting curve analysis (55uC to 95uC) was performed to confirm product specificity. The cycle threshold (C T ) was generated using BioRad software and it was used to calculate the amount of target DNA (PERV pol or human beta-actin). A standard curve was generated using the method of Dorak [56] and a dilution series (10 to 10 7 copies) of a linearized plasmid containing the relevant 193 bp PERV pol gene fragment. The equation generated from the standard curve (slope and y intercept) was used to determine the efficiency of the PCR reaction and to quantify the number of PERV pol gene or human beta-actin copies in the Q-PCR reactions. PERV copy numbers were normalized to those of beta-actin using the method of [56] . The primer sets used in this study were: PERV pol (193 bp): 59-AAC CCT TTA CCC TTT ATG TGG AT and 59-AAA GTC AAT TTG TCA GCG TCC TT; Standard reverse transcription (RT)-PCR reactions were performed using RNA prepared from PK-15 cells (TRIzol protocol, Invitrogen), M-MLV reverse transcriptase was used for cDNA synthesis using an oligo dT primer (Ambion) and Taq polymerase was used for PCR (Roche). The primers specific to pig APOBEC3F were 59-TGG TCA CAG AGC TGA AGC AG and 59-TTG TTT TGG AAG CAG CCT TT (175 bp). The semi-nested primer set used to detect plasmid-expressed pig APOBEC3F was 59-CCA AGG AGC TGG TTG ATT TC (exon 6, reaction 1), 59-CTG GAG CAA TAC AGC GAG AG (exon 7, reaction 2) and 59-TAG AAG GCA CAG TCG AGG, with the latter being vector specific (319 bp and 190 bp products, respectively). The mammalian beta-actin primers were 59-CCT TCA ATT CCA TCA TGA AGT G and 59-CCA CAT CTG CTG GAA GGT (236 bp). These primers amplify equally well a 236 bp beta-actin fragment from all mammals tested, including pigs and humans (e.g., Online Figure S3 ). The human APOBEC3G-GFP, pig APOBEC3F-GFP and GFP expression constructs were described previously [28, 41] . The pig and human cell lines were maintained as above. One day prior to transfection, 5,000-20,000 cells were seeded onto LabTek chambered coverglasses (Nunc). After 24 hrs incubation, these cells were transfected with 250 ng of the relevant plasmid construct. After 24 hrs of additional incubation, images of the live cells were acquired using a Zeiss Axiovert 200 microscope at 4006 total magnification. Images were analyzed using Image J software (http://rsb.info.nih.gov/ij). Whole cell protein extracts were prepared from 293T cells by suspending 500,000 cells in PBS, sonicating twice for 5 seconds and clarifying the lysates by centrifugation. Soluble protein levels were quantified using a BioRad Bradford assay. 10 mg of cell lysate was tested for reverse transcriptase activity using a C-type-RT activity assay (Cavidi Tech) following the manufacturers' instructions. Cell-free PK-15 supernatants (PERV-containing) were assayed directly using the Cavidi Tech ELISA assay. Human 293T cell genomic DNA was prepared from terminal cocultures and 50 ng was used for high fidelity, PERV pol genespecific PCR reactions (Phusion polymerase; Finnzymes). 193 bp products were cloned using the Zero Blunt TOPO PCR Cloning kit (Invitrogen) and sequenced (University of Minnesota Advanced Genetic Analysis Facility). Sequence comparisons were performed using Sequencher software (Gene Codes Corp.) and publicly available Clustal W alignment algorithms (http://align.genome. jp/). Figure S1 Quantitative Real-time PCR Analyses. (A) Standard curves depicting Q-PCR data obtained using dilutions of a linearized PERV pol gene plasmid alone (squares) or diluted plasmid plus 10 ng of 293T cell genomic DNA (diamonds). Under both conditions, all template amounts (10 to 10 7 copies) amplified efficiently (the log-linear slope equations are shown). The correlation co-efficiency value (R 2 ), which reports the technical accuracy of the assay, is also indicated. The standard curve data points were the average of 2 independent reactions with deviations smaller than the symbols (i.e., C T errors for each point ranged from 0 to 0.4). (B) Two representative control Q-PCR datasets showing the amplification of PERV pol gene DNA from pig PK-15 cell genomic DNA (circles). Two additional control Q-PCR datasets showing that the PERV-specific primers fail to amplify product from uninfected human 293T cell genomic DNA (squares). The reaction threshold, 10 times the mean standard deviation of the background fluorescence level (BioRad), is indicated. (C) Representative co-culture Q-PCR amplification curves of PERV pol gene DNA. Template genomic DNA isolated from human 293T cells co-cultured with vector expressing PK-15 cells (diamonds) or human APOBEC3G-expressing PK-15 cells (triangles) was used. (D) Representative Q-PCR amplification curves of the 293T cell beta-actin gene, which served as an internal standard for quantifying the real-time PCR data. Raw Q-PCR data will be made available on request. Found at: doi:10.1371/journal.pone.0000893.s001 (9.93 MB TIF) Figure S2 APOBEC3G inhibits PERV transmission. (A) A graph showing the accumulation of PERV pol gene-specific PCR products in 293T cells co-cultured with a control cell line (V3) but not with an APOBEC3G-expressing cell line (G1). The data points were an average of two Q-PCR runs and the difference between each run was smaller than the plotted symbol. The experimental parameters were identical to those used in the experiments shown in Figures 1B and 2B. (B) Relative levels of reverse transcriptase(RT)-activity detected in soluble extracts of day 28 co-cultured 293T cells, which were used to generate the Q-PCR data shown in Figure S2A . Uninfected 293T cell lysates had a relatively high endogenous RT activity. Therefore, to help with the presentation of these data, this level was normalized to one and all of the other data were calculated relative to this value. The level of RT activity in PK-15 extracts was much higher than that of 293T cell extracts (+/2PERV) and it had reached saturation (out of range) when these data were collected. Found at: doi:10.1371/journal.pone.0000893.s002 (4.76 MB TIF) Figure S3 Pig APOBEC3F Is Expressed in PK-15 Cells and its Over-expression Does Not Markedly Inhibit PERV Transmission. (A) An image of an ethidium bromide-stained agarose gel showing the results of an RT-PCR amplification experiment using PK-15 cellular RNA and appropriate controls. The top panel shows that PK-15 and representative PK-15 derived clones all expressed pig APOBEC3F, as indicated by the specific 175 bp pig APOBEC3F PCR product (confirmed by DNA sequencing). 293T cell mRNA and a diluted pig APOBEC3F expression plasmid were used as negative and positive controls, respectively. A larger, non-specific band was apparent only in the 293T cell RT-PCR reactions. The bottom panel shows that a conserved, 236 bp beta-actin gene fragment could be amplified from both PK-15 cells and human 293T cells (but not from diluted plasmid DNA). Note that this primer set differs from the human-specific set used in the Q-PCR experiments. The sizes of the marker (M) DNA bands are shown. (B) An image of an ethidium bromide-stained agarose gel showing expression of plasmid-derived pig APOBEC3F in PK-15 cells after 26 days of continuous co-culture. Non-transfected (NT) cells and diluted APOBEC3F plasmid DNA (pDNA) provided negative and positive controls, respectively. The larger 319 bp (far right lane only) and smaller 190 bp bands are the specific PCR products of the first and second rounds of semi-nested PCR, respectively (confirmed by DNA sequencing). (C) A histogram summarizing the level of PERV transmission that was observed after 23 days of co-culturing human 293T cells with PK-15 cells expressing a vector control or over-expressing pig APOBEC3F. Two datasets, each with an independent PK-15 clone in three replica co-culture wells, were collected in parallel and averaged for each histogram bar. One standard error of the mean is shown. The experimental parameters are identical to those used in Figure 1B . Found at: doi:10.1371/journal.pone.0000893.s003 (8.52 MB TIF) Figure S4 Genetic Variation in Zoonosed PERV pol Gene Sequences. (A) Sequences of the PERV pol gene fragments cloned from 293T cells co-cultured with control vector-expressing PK-15 cells. The number of times that each sequence was recovered is shown (N). Experiments 1 and 2 used genomic DNA prepared from the 293T cells used to generate the data shown in Online Figure S2 (day 28 samples) and Figure 2B (day 23), respectively. The most frequently detected 147 bp PERV pol gene sequence is shown in its entirety (which together with PCR primers makes up the 193 bp product shown in Figure 5 ). Identical nucleotides in other sequences are represented by dashes and non-identical nucleotides by the indicated DNA bases. GenBank accession numbers are shown for pol gene fragments with 100% identity to previously reported sequences. (B) Sequences of the PERV pol gene fragments cloned from 293T cells co-cultured with control APOBEC3G-expressing PK-15 cells. Parameters are identical to those described above. Found at: doi:10.1371/journal.pone.0000893.s004 (0.05 MB DOC)

Experimental infection of H5N1 HPAI in BALB/c mice

BACKGROUND: In 2005 huge epizooty of H5N1 HPAI occurred in Russia. It had been clear that territory of Russia becoming endemic for H5N1 HPAI. In 2006 several outbreaks have occurred. To develop new vaccines and antiviral therapies, animal models had to be investigated. We choose highly pathogenic strain for these studies. RESULTS: A/duck/Tuva/01/06 belongs to Quinghai-like group viruses. Molecular markers – cleavage site, K627 in PB2 characterize this virus as highly pathogenic. This data was confirmed by direct pathogenic tests: IVPI = 3.0, MLD(50 )= 1,4Log10EID(50). Also molecular analysis showed sensivity of the virus to adamantanes and neuraminidase inhibitors. Serological analysis showed wide cross-reactivity of this virus with sera produced to H5N1 HPAI viruses isolated earlier in South-East Asia. Mean time to death of infected animals was 8,19+/-0,18 days. First time acute delayed hemorrhagic syndrome was observed in mice lethal model. Hypercytokinemia was determined by elevated sera levels of IFN-gamma, IL-6, IL-10. CONCLUSION: Assuming all obtained data we can conclude that basic model parameters were characterized and virus A/duck/Tuva/01/06 can be used to evaluate anti-influenza vaccines and therapeutics.

Influenza A (H5N1) virus now becomes a real threat for humans. Since 1997, when first human case of H5N1 HPAI had been reported, more than 317 people were infected and 191 died [1] . Before 2005 attention was attracted to Thailand, Vietnamese and Indonesian viruses. In the beginning of 2005 outbreak on Quinghai lake occurred [2] . Later "Quinghai-like" viruses spreaded to most part of Russia, European countries and Africa and caused numerous outbreaks. Only in Russia more than 1 million of different species and sorts of poultry died and been slaughtered [3] . Confirmed cases in Azerbaijan, Egypt, Iraq, and Turkey was caused by Quinghai-like viruses. Earlier HPAI viruses were investigated in mice [4, 5] and murine models were successively used for reverse genetics made influenza vaccines [6] . It was shown that H5N1 HPAI viruses could have different pathogenicity for mice [7] . Several molecular markers were choused to explain differences. Multibasic cleavage site with 627K in PB2 designate to highly pathogenic phenotype for mice. Also important role of pulmonary cytokines elevation was highlighted [8] . Combination of adaptation for wild waterfowl and high virulence for mammals makes Quinghai-like viruses presumably pandemic. Also, in future, because of ability for rapid spreading for long distances, this group of viruses can appear in North and South America and cause outbreaks. Human disease caused by HPAI viruses can be characterized as acute viral pneumonia aggravated by ARDS, toxic shock and multiple organ failure. System dysfunction mediated by hypercytokinemia and high viral load [9] . To be ready for new influenza pandemy it is necessary to use animal models, in vaccine and antivirals studies, which most closely reflect human disease. Isolates from FRSI SRC VB "VECTOR" repository which were characterized previously were examined for MLD 50 , molecular markers of pathogenicity, sensitivity to amantadines and neuraminidase inhibitors, to be candidates for murine model. Among the investigated isolates A/duck/Tuva/01/06 has best features to be used. Genes of A/duck/Tuva/01/06 were sequenced and analyzed for molecular markers of pathogenicity. Also phylogenetic analysis was performed. Results are presented in figure 1 . A/duck/Tuva/01/06 belongs to group of Qinghailike viruses. HA contains 5 polybasic aminoacids (PQGRRKKKR↓GL) in cleavege site of HA [15] . The receptor binding domen can be characterized as "avian" [16] . High pathogenicity to mammals in general correlates with presence of 627K in PB2 [17] . The analysis of non-structural protein 1 (NS1) which also could be contributed for high virulence of H5N1 viruses revealed deletion of 5 amino acids similar to those in H5N1 viruses of genotype Z which could be contributed to increased expression of TNF-α and IP-10 protein in primary human macrophages [18] . A/duck/Tuva/01/06 con-tained Glu 92 in the NS1 and contained "avian-like" PDZdomain ligand ESEV [19] . It was shown that the most recent H5N1 strains isolated in Southeast Asia were resistant to amantadine and rimantadine; antiviral drugs targeted the M2 ion channels of influenza A viruses [20, 21] . It was also reported about Oseltamivir resistant H5N1 viruses isolation from humans [22, 23] . To determine the potential sensitivity of studied H5N1 viruses to these antivirals, amino acid sequences of the M2 and NA proteins were analyzed. Variants of influenza A viruses resistant to amantadine possessed amino acid substitutions at one of 5 residues (26, 27, 30, 31, and 34) in the M2 protein [24, 25] . Sequence analysis did not reveal any mutations associated with resistance to amantadine. Thus all A/duck/Tuva/01/ 06 is potentially sensitive to this class of antiviral agents. Amino acid residues 119, 274, 292 and 294 in the NA protein (numbering according to the HA of H2 subtype) are crucial for the sensitivity of influenza A viruses to neuraminidase inhibitors [26] ; substitution H 274 →Y in the NA conferred resistance to Oseltamivir was observed in clinical H5N1 isolates [25, 26] . Sequence comparison of the NA protein of A/duck/Tuva/01/06 aligned with the NA of N2 subtype of A/Wuhan/359/95 (H3N2) influenza virus showed phenotype potentially sensitive to neuraminidase inhibitors. A/duck/Tuva/01/06 showed wide cross-reactivity with sera against H5N1 HPAI viruses isolated earlier in South-Eastern Asia. HI results can be found in table 1. These features persuade to use this virus in studies of vaccines made from various H5N1 influenza viruses. First MID 50 and MLD 50 for A/duck/Tuva/01/06 were determined (table 2) . To determine mean time to death (m.t.d) and 2C. In several cases (9 animals totally) the disease was complicated by severe intestine atony, which can independently lead to death or by pressuring on diaphragm can intensify respiratory failure. We also determined virus titers in several organ tissues. As it was expected the highest titers was observed in lungs -5,3 log EID 50 . Brain titers were also high -3,4 log EID 50 . In spleen, liver and kidney tissues virus titers were lower then 1 logEID 50 and considered not significant. We investigated the involvement of several cytokines in immunopathogenesis of experimental H5N1 HPAI infection in mice. Results of ELISA technique revealed alteration of expression both pro-inflammatory and antiinflammatory cytokines after the challenge (figure 3). In general, the most marked changes of cytokine levels were observed before the death of mice. The minimal concentration of IFN-γ was detected on day 5 (14.3 ± 10.8 pg/ml), however, its levels enlarged about 8-fold (256 ± 27 pg/ml) during the course of the infection when compared with uninfected animals. On days 3 and 5 systemic production of TNF-α was below the detection limit of the assay. A peak was reached on day 7 by the cytokine (24 ± 3.2 pg/ml) and its levels remained elevated on day 8. Interestingly, concentrations of IL-1β in mice after the challenge were significantly lower in comparison with the constitutive expression of the mediator in intact animals. An abrupt decrease of IL-1β was detected on day 3 post infection, but was followed by step increase from day 5. After the 2.5-fold enlargement on day 3 the levels of IL-6 decreased dramatically on day 5, and the highest levels of the cytokine were determined at the end of observation period (133 ± 12 pg/ml). The constitutive production of IL-10 was undetectable. The dynamics of IL-10 showed a gradual growth with the maximum level (92.1 ± 6.0 pg/ml) reached before the death of mice. We observed statistically significant increase of IL-12 after the challenge. Concentrations of the cytokine retained constant in infected mice, except the unexpected decline occurred on day 7. The expression of IL-18 could not be detected throughout the entire period of observation. Until 2005 avian influenza was regional problem of sev- [4] and question "why had only some wild waterfowl died?" is still unclear. Most of the outbreaks in Russia associated with wild birds. The same time viruses adapted to wild birds are extremely pathogenic for poultry and mice. This "competitive advantage" makes Quinghailike viruses most probable candidate to be precursor for new pandemic influenza virus. At the same time pathogenesis of different (phylogenetical clades) HPAI reveal common causes. The principal causes of rapid mice death after infecting with HPAI are primary viral pneumonia, ARDS, lesions of central nervous system and multiple organ failure. Our data suggest that A/duck/Tuva/01/06 strain of HPAI caused lethal pneumonia and spread systemically to the brain in BALB/c mice. Lesion of respiratory epithelium and following an activation of monocytes/macrophages results in a release of proinflammatory cytokines (TNF-α, IL-6) which are a hallmark of ARDS in murine model [27] . Despite powerful anti-influenza virus effects of TNF-α in lung tissue, as it was described previously [28] , we consider that elevated production of the cytokines seems to be crucial in the pathogenesis of HPAI infection. Moreover, it was shown that lethal H5N1 viruses are resistant to antiviral effects of interferons and TNF-α [29] . Virus-induced overexpression of TNF-α as well as high IFN-γ lead to activation of endothelium and imbalance in blood coagulation system [30] . This may explain the hemorrhagic syndrome as observed in some of animals. To pay attention that IL-12 is a potent inducer of IFN-γ synthesis by blood mononuclear cells [31] , we concluded the same cytokines hyperproduction reflects macrophage overactivation and subsequent hypercytokinemia. This cascade of events Phylogenetic tree based on full length sequencesof HA Figure 2 Phylogenetic tree based on full length sequencesof HA. Nucleotide sequences were analyzed by using the neighborjoining method with 500 bootstraps. The phylogenetic tree was rooted to the HA gene of A/goose/Guangdong/1/96 (H5N1) virus. including inflammatory mediator production, changes in blood coagulation system and microvascular permeability was denoted as systemic inflammatory response syndrome (SIRS) [32] . On the other hand, we proposed that the prominent production of IL-10 from the early stages of the experimental HPAI infection was the compensatory response to overproduction of proinflammatory cytokines such as TNF-α, IL-6 and IL-12. However, the role of IL-10, which principle function seems to be containment and eventual termination of inflammation [33] , in HPAI pathogenesis is unclear. Also there is an uncertain discrepancy between undetectable expression of IL-18 and high levels of other Th1-cytokines (IFN-γ and IL-12). Summing up, in our study BALB/c mice infected with HPAI, strain A/duck/Tuva/01/06, appeared to be able to produce the innate immune response, which culminated to the development of shock and subsequent multiple organ failure. The main characteristics of our model are comparable to the previously described fatal cases of H5N1 influenza in humans [10, 11] . Proposed model reflects lesions not only same organs but also mediating levels of some (IFN-γ, IL-6, IL-10) cytokines in terminal conditions. The implication of different cytokines in immunopathogenesis of experimental HPAI is beyond question. But to understand exact mechanisms, which determine the disease outcome, further experiments remain to be done. All experiments were performed in BSL 3+ facilities of FSRI SRC VB "Vector" of Rospotrebnadzor licensed for working with highly pathogenic avian influenza viruses. Stock of A/duck/Tuva/01/06 was produced in 9 days-old chicken embryos. Allantoic fluid was aliquoted and stored at -80°C. The infectivity of stock viruses was determined in 10 days-old embryonated chicken eggs; titers were calculated by the method of Reed and Muench [10] and were expressed as log 10 of 50% egg infective dose (EID 50 ) in 1 ml of allantoic fluid. Viral RNA was isolated from virus-containing allantoic fluid with the RNeasy Mini kit (Qiagen, Valencia, CA) as specified by the manufacturer. Uni-12 primer was used for reverse transcription. PCR was performed with a set of primers specific for each gene segment of Influenza A virus [11] . PCR products were purified with the QIAquick PCR purification (Qiagen). Sequencing was done with Beckman Coulter Genom-eLab™ Methods development kit Dye terminator Cycle Sequencing according instructions of manufacturer. Primers for sequence were obtained from E. Hoffman (SJCRH, Memphis, TN). Sequence products were analyzed on automatic sequence analyzer Beckman Coulter CEQ2000. Phylogenetical analysis was done on HA full gene sequence DQ861291 using MEGA 2.1 software. Phylogenetical tree was built by Neighbor-Joining method; matrix of distances was counted with p-distance algorithm. Reliability of clades was checked with bootstrap analysis with 500 replications. Other genes in GenBank DQ861291-DQ861295. Cross-reaction of A/duck/Tuva/01/06 was defined by hemagglutination inhibition test (HI) with 0.5% CRBC [12] with a panel of antisera against H5N1 HPAI. Six-week-old inbred male BALB/c mice (vivarium of FRSI SRC VB "Vector"). Animals were placed to individual cages with food and water available ad libitum. To determine the MLD 50 and MID 50 , mice were anaesthetized by diethyl ether inhalation and infected intranasally with 50 µl 10-fold serial dilutions of allantoic fluidin PBS (pH 7,2). Each group contained 10 animals. Animals were observed daily for 15 days for mortality (MLD 50 ) or sacrificed on day 5 after the challenge with following virus detection in the lungs by inoculation of 10 days-old embryonated chicken eggs (MID 50 ). MLD 50 and MID 50 were calculated by the method of Reed and Muench. Animals from group where 1MLD 50 had been observed were taken to determine virus titers in lung, spleen, kidneys, and liver and brain tissues. Mind time to death (m.t.d) was calculated as previously described [13] . Pathogenicity to chickens was determined by IVPI test [14] . All animal studies were performed according protocols approved by Animal Care & Use committee of FSRI SRC VB "Vector". To determine IFN-γ, TNF-α, IL-6, IL-10, IL1-β, IL-12 we use ELISA R& D Systems kits (Minneapolis, MN, USA). Serum levels of IL-18 were measured using commercial Mouse IL-18 ELISA test kit (MBL, Nagoya, Japan). Detection limits were as follows: TNF-α, less then 5,1 pg/ml; IL1-β, 3,0 pg/ml; IL6, 3,1 pg/ml; IL10, 4,0 pg/ml; IL-18, 25 pg/ml. Sera was taken on 0,3,5,7,8 days and aliquots and stored -80°C upon usage. Day 8 was chosen because m.t.d defined earlier in the work was 8,19 ± 0,18 days. Statistics was performed with Student t-test. Values p < 0,05 considered to be reliable.

Prediction of RNA Pseudoknots Using Heuristic Modeling with Mapping and Sequential Folding

Predicting RNA secondary structure is often the first step to determining the structure of RNA. Prediction approaches have historically avoided searching for pseudoknots because of the extreme combinatorial and time complexity of the problem. Yet neglecting pseudoknots limits the utility of such approaches. Here, an algorithm utilizing structure mapping and thermodynamics is introduced for RNA pseudoknot prediction that finds the minimum free energy and identifies information about the flexibility of the RNA. The heuristic approach takes advantage of the 5′ to 3′ folding direction of many biological RNA molecules and is consistent with the hierarchical folding hypothesis and the contact order model. Mapping methods are used to build and analyze the folded structure for pseudoknots and to add important 3D structural considerations. The program can predict some well known pseudoknot structures correctly. The results of this study suggest that many functional RNA sequences are optimized for proper folding. They also suggest directions we can proceed in the future to achieve even better results.

is a bracket graph developed by Hofacker et al. [1] , where the bracket is matched by an equal and opposite bracket on the 3' side. Figure S2d There are many examples of symmetric I-loops found in RNA structure databases. Some asymmetric I-loops ( n ) can also be found. ))).)))) ((((....(((((....))))))))) g h e f 5' 3' 3' ((((.....) )))..(((((.....))))).)))). (4) For pseudoknots: parts of the structure still satisfy cases 1 through 3. However, in addition, in at least one part of a region between and l such that , there exists some base pairs that satisfy either or k , , ', ' Hence, many more possibilities can be generated once we start allowing pseudoknots. Pseudoknots differ from real knots in the sense that the strand does not pass completely through the loop but only becomes potentially entangled with it. Anyone who has tried to untangle a pair of earphones or tried to untangle a recently, neatly wound up cord can realize that knots naturally occur on long flexible cords. Indeed, great effort seems to be required to avoid tangling a heap of cords. Figure S5a shows a common pseudoknot known as an H-type pseudoknot. This is also known as an ABAB pseudoknot. The structure is notated below with the standard parenthesis notation for the basic secondary structure (here shown in blue) and square brackets for the pseudoknot linkage (here shown in red). Green indicates the regions of free strand that are not forming base pairs (bps). The color distinction of the stems in this example is not important because both stems are the same length and both stems are less than 10 bps. [[[[[[[[[[.) ))))))))))))) . [.) ))))))))))))) . and also known as an ABACBC pseudoknot. (f) The structure shown in bracket notation. (g) A pseudoknot. The difference between (c) and (g) is that the linkage stem is shorter than 9 contiguous bps. (h) The same structure in bracket notation. Figure S5c shows a knot and the corresponding bracket structure is shown below in Fig. S5d . Both stems in Fig. S5c contain 14 base pairs (bp). Since the helical axis makes a rotation of 360 o every 10 bps, this means that the structure in Fig. S5c is tangled in a knot. There is no reason why such a structure cannot form. Indeed, knots are known to form in some rare proteins [2, 3] . However, it is not a pseudoknot and the current approach is not designed to estimate its existence or the likelihood of its formation. One important feature of a pseudoknot is, therefore, that the linkage stem (here shown in red) must be shorter than 10 contiguous bps. Figure S5e shows two stem-hairpins-loops (blue and purple) that are joined by a linkage stem (red). The stems are all short as in Fig. S5a . This is a pseudoknot, sometimes referred to as a kissing loop. It is also known as an ABACBC pseudoknot. It is also observed in a number of places, although less frequently than H-type pseudoknots. The corresponding bracket notation for this structure is shown below (Fig. S5f) . Belvedere knot secondary structure Supplement Figure S6 . A special type of knot (right hand side) that becomes entangled due to the way the structure folds up. The two dimensional nature of the RNA schematics tends to hide curious possibilities as this. The knot is seen in equilibrium between the unknotted structure (left hand side) and the "Belvedere knot" (right hand side). Functional RNA structures that contain knots of the form shown in Figs. S5c or S6 are currently unknown or have not been reported. However, pseudoknots are often observed in functional RNA structures, particularly H-type pseudoknots (Fig. S5a) . Single strand RNA sequences such as messenger RNA with introns can have sequences with lengths that number in the tens of thousands of nucleotides. With such a propensity for a few simple cords to become tangled, and, since the cell can have many thousands of protein and RNA strands present within the cellular environment, this suggests that there is a fair amount of effort made within the cell to prevent or get rid of knots [2] . In this work, we are concerned with the prediction of pseudoknots. These structures have the property that they can be evaluated as structures resulting from reversible folding.

Cell-penetrating peptides as transporters for morpholino oligomers: effects of amino acid composition on intracellular delivery and cytotoxicity

Arginine-rich cell-penetrating peptides (CPPs) are promising transporters for intracellular delivery of antisense morpholino oligomers (PMO). Here, we determined the effect of L-arginine, D-arginine and non-α amino acids on cellular uptake, splice-correction activity, cellular toxicity and serum binding for 24 CPP−PMOs. Insertion of 6-aminohexanoic acid (X) or β-alanine (B) residues into oligoarginine R(8) decreased the cellular uptake but increased the splice-correction activity of the resulting compound, with a greater increase for the sequences containing more X residues. Cellular toxicity was not observed for any of the conjugates up to 10 μM. Up to 60 μM, only the conjugates with ⩾ 5 Xs exhibited time- and concentration-dependent toxicity. Substitution of L-arginine with D-arginine did not increase uptake or splice-correction activity. High concentration of serum significantly decreased the uptake and splice-correction activity of oligoarginine conjugates, but had much less effect on the conjugates containing X or B. In summary, incorporation of X/B into oligoarginine enhanced the antisense activity and serum-binding profile of CPP−PMO. Toxicity of X/B-containing conjugates was affected by the number of Xs, treatment time and concentration. More active, stable and less toxic CPPs can be designed by optimizing the position and number of R, D-R, X and B residues.

Steric-blocking antisense oligonucleotides (AOs) are considered potential therapeutics for genetic diseases such as Duchenne muscular dystrophy (DMD) and b-thalassemia. For their potential to be realized, however, the AOs must be effectively delivered to cell nuclei. Cationic lipoplex-or PEI-based transfection methods used to deliver charged AOs are not suitable for the delivery of uncharged AOs such as phosphorodiamidate morpholino oligomers (PMO, Figure 1 ) (1) and peptide nucleic acids (PNAs) (2) . Conjugation of PMO to short CPPs is a good method to enhance the cytoplasmic and nuclear delivery of PMO because the conjugates are simple to use and because the short peptides and their AO conjugates can be easily manufactured and characterized in a qualitycontrolled manner. Examples of well-studied CPPÀPMO conjugates include those with Tat and oligoarginine peptides (3, 4) Important considerations in the design of effective CPPs include the ability to deliver AO efficiently, stability in living systems and toxicity. We have reported that Tat and oligoarginine peptides are not stable in human serum (5) , and are therefore ill-suited for in vivo applications. Oligoarginine peptides incorporating non-a amino acids have been proven superior to oligoarginine alone. CPPs containing 6-aminohexanoic acid (X) and b-alanine (B) were more stable in human serum than Tat or oligoarginine peptides (5) . A CPPÀPMO conjugate, (RXR) 4 ÀPMO, has been shown to be more efficient in the correction of pre-mRNA mis-splicing (6) and in inhibition of the replication of mouse hepatitis virus in vivo (7) than an oligoarginine peptide. In addition, (RXR) 4 ÀPMO conjugates have been shown to cause effective exon skipping in muscle cells from DMD dogs (8) , in human muscle explants (9) and in mdx mice (10) , as well as inhibiting the replication of various viruses in cell cultures (7, (11) (12) (13) and in mice (7, 13) . The above studies have helped make it clear that unnatural amino acids can confer enhanced stability and activity, and therefore improve the potential of CPPs to deliver therapeutic PMO. In pursuit of CPPs with improved characteristics, we have carried out a structure-activity relationship study to investigate the effects of unnatural amino acid insertions in oligoarginine peptides on cellular delivery, nuclear antisense activity, toxicity and serum-binding characteristics of the resulting CPPÀPMO conjugates. The unnatural amino acids studied here are X, B and D-arginine (r). We chose to study the X amino acid based on the successes of the (RXR) 4 CPP in several studies as shown in the previous paragraph. B and r amino acids were chosen because they have good enzymatic stability (5) . The CPPs are (i) the oligoarginine sequences, R 8 and R 9 , (ii) sequences with RXR, RX and RB repeats, as well as various combinations thereof, and (iii) sequences containing D-arginine, r 8 , (rX) 8 (rXR) 4 , (rXr) 4 and (rB) 8 . The CPPÀPMO conjugates were evaluated for their relative (a) cellular uptake, as determined by flow cytometry, (b) antisense activity, as determined by a splice correction assay (13) and (c) cellular toxicity, as determined by MTT cell viability, propidium iodide membrane integrity and hemolysis assays, as well as by microscopic imaging. CPP nomenclature and sequences are listed in Table 1 . Chemical structures of PMO and (RX) 8 ÀPMO are shown in Figure 1 . The antisense PMO (CCT CTT ACC TCA GTT ACA) is designed to target a b-thalassemic mutant splice site present in the human b-globin intron 2 of a positive-readout antisense activity assay system (13) as described in the Results section. Synthesis of PMO, described previously (15, 16) , and the CPPs, using standard Fmoc chemistry (17) , were performed at AVI BioPharma, achieving purities of 490% as determined by HPLC and mass spectrometry analysis. Conjugation of a CPP to a PMO through an amide linker, described previously (6) , was followed with an additional purification step to remove nonconjugated peptide. Samples were loaded on source 30S resin (Amersham Biosciences, Pittsburgh, PA) in a 2 ml Biorad (Hercules, CA) MT2 column at 2 ml/min with running buffer A (20 mM Na 2 HPO 4 , 25% acetonitrile, pH 7.0) and purified into 45-s fractions with 0-35% buffer gradient (buffer B: 1.5M NaCl, 20 mM Na 2 HPO 4 , 25% acetonitrile, pH 7.0) over 60 min, using a Biorad BioLogic low pressure chromatography system. The desired faction was desalted by a method described previously (6) . HPLC and MS analyses revealed that the final product contained 490% CPP conjugated to fulllength PMO, with the balance composed of CPP conjugated to incomplete PMO sequence, nonconjugated full-length or incomplete PMO. The HeLa pLuc705 (pLuc705) (14) cell line was obtained from Gene Tools, LLC (Philomath, OR). Human liver cell line HepG2 was from American Type Culture Collection (ATCC, Manassas, VA). Cells were cultured in RPMI 1640 medium supplemented with 2 mM L-Glutamine, 100 U/ml penicillin and 10% fetal bovine serum (FBS) (HyClone, Ogden, UT) at 378C in a humidified atmosphere containing 5% CO 2 . All treatments were carried out in OptiMEM medium (Gibco, Inc., Carlsbad, CA.) with or without FBS. Cell uptake assay pLuc705 cells were seeded 20 h prior to treatment in 12-well plates at 100 000 cells/well. Cells were treated with 2 mM fluorescein-tagged CPPÀPMO conjugates for 24 h. After treatment, cells were washed with 500 ml of 7 RXRXRXRXRXRXRXB 7 (RX) 5 RXRXRXRXRXB 5 (RX) 3 RXRXRXB 3 (RXR) 4 RXRRXRRXRRXRXB 5 (rXR) 4 rXRrXRrXRrXRXB 5 (rXr) 4 rXrrXrrXrrXrXB 5 (RB) 8 RBRBRBRBRBRBRBRBB 0 (rB) 8 rBrBrBrBrBrBrBrBB 0 (RB) 7 RBRBRBRBRBRBRBB 0 (RB) 5 RBRBRBRBRBB 0 (RB) 3 RBRBRBB 0 Ã The sequences of peptides are written from N to C terminus. R = arginine, r = D-arginine, X = 6-aminohexanoic acid, B = b-alanine. Each peptide had an acetyl group at the N-terminus and a carboxyl group at the C-terminus. Cell viability assay and microscopy pLuc705 cells were seeded 20 h before the treatment in 96-well plates at 9000 cell/well and then treated with the conjugates. The microscopic phase images of treated cells were visualized by a Nikon Diaphot inverted microscope (Melville, NY), captured by an Olympus digital camera and processed by the Magnafire software (Optronics, Goleta, CA). After imaging the cells, the cell viability was determined by the methylthiazoletetrazolium assay (MTT, Sigma, St. Louis, MO) assay. MTT solution (5 mg/ml) was added to the treatment medium to a final concentration of 0.5 mg/ml and incubated for 4 h at 378C. 85% of the media of each well was then replaced with DMSO containing 0.01M HCl and further incubated for 10 min at 378C and the absorbance measured at 540 nm. Percent cell viability was determined by normalizing the absorbance of each treated sample to the mean of untreated samples. Propidium iodide membrane integrity assay pLuc705 cells were seeded 20 h before treatment in 12-well plates at 100 000 cells/well. Cells were treated by removing the medium, washing with 500 ml PBS and incubating with medium containing CPPÀPMO conjugates. Treatment medium was collected in tubes and cells were washed with PBS once, and then treated with 400 ml of 10% trypsin for 10 min at 378C. Trypsin was neutralized with 500 ml of the serum containing medium. Cells were transferred to the tubes containing previously collected treatment medium, pelleted by centrifugation at 1000g for 5 min, washed with PBS once, and re-suspended in 200 ml of 0.05 mg/ml propidium iodide (PI) in PBS. Cells were further incubated at 378C for 15 min and analyzed by the Beckman Coulter cytometer (30 000 events/sample collected). The hemolytic activities of the conjugates were determined in fresh rat blood according to a method described elsewhere (18) . Cellular uptake of CPPÀPMO conjugates was investigated using the 3 0 -carboxyfluorescein-tagged PMO (PMOF) and flow cytometry. We chose 2 mM as the treatment concentration because none of the conjugates caused any detectable cytotoxicity at this concentration, as demonstrated by the MTT and PI uptake assays. After treating with the conjugates, cells were treated with trypsin (19) to remove membrane-bound conjugates. We found that a heparin sulfate washing step prior to trypsin treatment did not remove additional membrane-bound conjugates but caused some cellular toxicity (data not shown); therefore, only the trypsin treatment step was used in this study. To determine the effect of serum on cellular uptake of the various conjugates, uptake evaluation assays were carried out in the medium containing various concentrations of, or in the absence of, serum. Cellular uptake of CPPÀPMOF conjugates increased with the number of arginines and decreased with the X and/or B residue insertion (Figure 2A and B). The oligoarginine R 9 ÀPMOF had a mean fluorescence (MF) of 662, nearly 3-fold higher than the 234 produced by R 8 ÀPMOF, indicating that a difference of a single arginine can make a substantial difference in the biological properties of a CPP. Insertion of an X or B residue in the R 8 sequence reduced the MF from 234 of R 8 ÀPMO to 42, 70 and 60 of (RX) 8 À, (RXR) 4 À and (RB) 8 ÀPMOF, respectively ( Figure 2A ). The number of RX or RB repeats affected cellular uptake, with conjugates having fewer RX or RB repeats generating lower MF ( Figure 2B) . While the addition of 10% serum to the medium caused a decrease in the uptake of the R 8 À or R 9 ÀPMOF conjugates, it increased the uptake of conjugates containing RX, RB or RXR motifs (Figure 2A and C) . Serum reduced the MF of R 9 À and R 8 ÀPMOF from 662 and 234 to 354 and 158, respectively, and increased the MF of (RX) 8 À, (RXR) 4 À and (RB) 8 ÀPMOF from 41, 70 and 60 to 92, 92 and 111, respectively. These differences were statistically significant (Figure 2A ). However, higher serum concentrations (30 and 60%) decreased the uptake of (RXR) 4 ÀPMOF and oligoarginineÀPMOF ( Figure 2C ). Arginine stereochemistry (D versus L) had little effect on the uptake of CPPÀPMOF conjugates. We compared the MF of R 8 À, (RB) 8 À and (RX) 8 ÀPMOF with their respective D-isomer conjugates, r 8 À, (rB) 8 À and (rX) 8 ÀPMOF and found that there was no significant difference between each pair, as shown in Figure 2D for the (RX) 8 À and (rX) 8 ÀPMOF pair. Nuclear antisense activity. The effectiveness of each CPPÀPMO conjugate was determined in a previously described splicing correction assay (14) , considered a reliable method to assess nuclear antisense activity of a steric-blocking AO. This assay utilizes the ability of steric-blocking AOs to block a splice site created by a mutation in order to restore normal splicing. The luciferase coding sequence was interrupted by the human b-globin thalassemic intron 2 which carried a mutated splice site at nucleotide 705. HeLa cells were stably transfected with the plasmid therefore named as pLuc705 cell. In the pLuc705 system, steric-blocking AOs must be present in the cell nucleus for splicing correction to occur. Advantages of this system include the positive readout and high signal-to-noise ratio. With this system the relative efficiencies of various CPPs to deliver an AO with sequence appropriate for splice-correction to cell nuclei can be easily compared. Oligoarginine, RX, RXR and RB panels. The CPP conjugates with the highest nuclear antisense activities were (RXR) 4 À and (RX) 8 ÀPMO. Figure 3A and B show luciferase activity normalized to protein of cells treated with various conjugates at 1 and 5 mM for 24 h. At both concentrations, (RX) 8 À and (RXR) 4 ÀPMO were more effective than the other conjugates tested, with the difference more prominent in serum-containing medium at 1 mM than at 5 mM. Cells treated with 1 mM of either conjugate had luciferase activity 10-15-fold over the background while the remaining conjugates yielded about a 2-4-fold over the background (Figure 3A ). At 5 mM, all conjugates generated higher luciferase activity Figure 2 . Cellular uptake of the CPPÀPMOF conjugates. pLuc705 cells were treated with carboxyfluorescein-tagged PMOs conjugated to CPPs in OptiMEM with or without 10% serum for 24 h, followed by flow cytometry analysis. Data are presented as mean fluorescence (MF) AE SD of six data points from two independent experiments. (A) Cells treated with 2 mM of R 8 2, R 9 2, (RXR) 4 2, (RX) 8 2 or (RB) 8 2PMOF conjugates. (B) Cells treated with 2 mM of (RX) 8 2, (RX) 7 2, (RX) 5 2, (RB) 8 2, (RB) 7 2 or (RB) 5 2PMOF conjugates in the absence of serum. (C) Cells treated with 2 mM of (RXR) 4 2 or R 9 2PMOF in media containing 0, 10, 30 or 60% serum. (D) Cells treated with (RX) 8 2 or (rX) 8 2PMOF in media containing 10% serum. than at 1 mM, with (RX) 8 ÀPMO and (RXR) 4 ÀPMO again the most effective, followed by (RB) 8 ÀPMO ( Figure 3B ). Figure 3C shows that at 10 mM, the activity of RX or RB conjugates decreased as the number of RX or RB repeats in the CPP decreased. The peptides with 5 or 3 RX or RB repeats, (RX) 5 , (RX) 3 (RB) 5 or (RB) 3 , generated much lower luciferase activity than those with 8 and 7 repeats. Number and position of X residues. Having shown that (RB) 8 ÀPMO had less activity than (RXR) 4 ÀPMO or (RX) 8 ÀPMO, we further investigated the effect of the number and position of X residues on the activity of conjugates. Eleven CPPÀPMO conjugates containing 0, 2, 3, 4, 5 or 8 Xs were compared (Figure 4) . Generally, CPPs containing a higher number of Xs had higher activities. At 2 mM, (RX) 8 ÀPMO (8 X residues) had the highest activity followed by (RXR) 4 ÀPMO (5 X residues) and the conjugates with fewer Xs had lower activities. At 5 mM, three conjugates containing 3 (3d), 4 (4c) and 8 ((RX) 8 ) X residues had the highest activities, suggesting that the position of X residues affects activity. L-Arginine versus D-arginine. Arginine stereochemistry had little effect on the nuclear activity of the R 8 À and (RB) 8 ÀPMO conjugates but affected the (RX) 8 ÀPMO ( Figure 5 ). Replacement of the eight L-arginine residues in R 8 À or (RB) 8 ÀPMO with D-arginine residues did not change the luciferase activity generated over the 1-5 mM ( Figure 5A and B) . However, the replacement did cause a small but statistically significant decrease in the activity for (RX) 8 ÀPMO at 1 mM (P = 0.03) and 2 mM (P = 0.01) ( Figure 5C ). Serum effect on activity. The effect of serum on the antisense activity of the conjugates depended on the CPP sequences, as shown in Figure 3A -D. Addition of 10% serum to the medium decreased the activity of oligoarginineÀPMO conjugates (R 8 ÀPMO and R 9 ÀPMO) but increased activity of conjugates containing RXR, RX and RB repeats. The addition of 10% serum nearly doubled the luciferase activity of (RXR) 4 À, (RX) 8 À and (RB) 8 ÀPMO at 5 mM ( Figure 3B ). We further studied this effect for (RXR) 4 ÀPMO up to 60% serum. While the activity almost doubled as the serum concentration increased from 0 to 10%, it gradually decreased as the serum concentration increased to 60%, at which activity was similar to that in 0% serum which was still significantly above the background. This 'up and down' profile was also observed with the 1 mM (RXR) 4 ÀPMO treatment. Unlike (RXR) 4 ÀPMO, the luciferase activity of R 8 ÀPMO or R 9 ÀPMO (data not shown) only decreased as the serum concentration increased, with an approximately 30% reduction in 10% serum and no activity in 60% serum. R 8 ÀPMO or R 9 ÀPMO did not display any detectable activity at 1 mM, regardless of the serum concentration (data not shown). The cellular toxicity of the various CPPÀPMO conjugates was determined by MTT-survival, propidium iodine (PI) exclusion and hemolysis assays and microscopic imaging. The MTT and PI exclusion assays measure metabolic activity and membrane integrity of cells, respectively. The hemolysis assay determines compatibility with blood. Microscopic images were used to verify the MTT results and observe the general health of the cells. MTT assay. pLuc705 cells were treated at concentrations ranging from 2 to 60 mM for 24 h. As shown in Figure 6 , all conjugates, except (RX) 8 and (RXR) 4 , had no toxicity at up to 60 mM. Up to 10 mM, (RX) 8 and (RXR) 4 conjugates exhibited no toxicity, at higher concentrations they reduced cell viability in a concentration-dependent manner, with (RX) 8 being more toxic than (RXR) 4 ( Figure 6C and D) . Replacement of L-arginine with D-arginine in R 8 À, (RB) 8 À and (RXR) 4 ÀPMO did not change the viability profiles of these conjugates ( Figure 6A-C) . Surprisingly, the L!D replacement in (RX) 8 ÀPMO decreased the toxicity. Cell viability with 60 mM treatment was 40% for (RX) 8 ÀPMO, but 80% for (rX) 8 ÀPMO ( Figure 6D ). The eight conjugates containing fewer than 5 X residues did not inhibit cell proliferation up to 60 mM ( Figure 6E ). Monomers of arginine or X, individually or in combination, at 500 mM each, produced no inhibition of cell proliferation ( Figure 6F ). The toxicities of the CPPÀPMO conjugates, (RXR) 4 ÀPMO, 3dÀPMO and 4cÀPMO, were also evaluated in human liver HepG2 cells. We found that only (RXR) 4 ÀPMO caused dose-dependent inhibition of cell proliferation while other two conjugates had no toxicity up to 60 mM, the highest concentration tested in this study (data not shown). Microscopic images. We sought to verify the MTT results by collecting microscopic images of cells treated with 60 mM of the conjugates. The images correlated well with the cell viability data. Images of (RX) 8 À, (rX) 8 À, (RXR) 3 RBRÀ (4c), (RXR) 4 ÀPMO and vehicle-treated cells are shown in Figure 7 . Cells treated with (RX) 8 ÀPMO and (RXR) 4 ÀPMO appeared rounded and detached from the culture well, and appeared to have fewer live cells. Interestingly, cells treated with (rX) 8 ÀPMO appeared to have normal morphology and cell density. The replacement of one X of (RXR) 4 ÀPMO . Nuclear activity of CPP-PMO conjugates: number and position of X residues. pluc705 cells were treated with the conjugates having 0, 2, 3 (3a, 3b, 3c and 3d), 4 (4a, 4b and 4c), 5 and 8 X residues (see sequences in Table 1 ) in OptiMEM medium with 10% serum for 24 h. Nuclear activity of a conjugate is indicated by relative luciferase activity (RLU) per microgram of protein. Data represent a mean AE SD of 9-12 data points from four independent experiments. Propidium iodine exclusion assay. The effect of the conjugates on integrity of cell membranes was investigated by a propidium iodine (PI) exclusion assay. PI can only permeate unhealthy/damaged membranes, so positive PI fluorescence indicates compromised cell membranes. Only (RXR) 4 À and (RX) 8 ÀPMO conjugates were found to significantly affect membrane integrity at higher concentrations (up to 60 mM tested). Figure 8A shows the histograms of pLuc705 cells treated with (RXR) 4 ÀPMO at 60 mM for 0.5, 5 and 24 h. The PI positive (PI+) region was defined by the cells permeabilized with ethanol (positive control) as indicated by the gate in the histogram. The PI histogram shifts from the PI-negative region to PI-positive region in the longer incubations, indicating the conjugate caused membrane leakage in a time-dependent manner. The 0.5-and 5-h-treatments caused a slight shift towards the PI+ region, while the 24-htreatment produced a distinct peak which corresponded to 57% of cells that were in the PI+ region. Figure 8B shows the histograms of cells treated with (RXR) 4 ÀPMO at concentrations of 2, 10, 20, 40 and 60 mM for 24 h. There was no significant PI uptake at concentrations up to 20 mM. At higher concentrations, the PI+ population appeared and the percentage of PI+ cells increased as the treatment concentration increased, indicating that there were more leaking cells at the higher treatment concentration. Similar concentrationand time-dependent PI uptake profiles were observed for (RX) 8 ÀPMO but not for (RB) 8 ÀPMO and the remaining conjugates (data not shown). Addition of 10% serum to the treatment medium significantly reduced membrane toxicity for the (RXR) 4 À ( Figure 8C ) and (RX) 8 ÀPMO conjugates (data not shown). Hemolysis assay. The (RXR) 4 À and (RX) 8 ÀPMO conjugates were tested in a hemolysis assay and found to be compatible with red blood cells. Fresh rat red blood cells were treated with the conjugates at 60 mM, PBS (background) or 0.005% TX-100 (positive control). The supernatants of conjugate-and PBS-treated samples had small and similar amounts of free hemoglobin released, far lower than that of the TX-100-treated samples ( Figure 8D ). The naturally occurring CPPs such as Tat peptide are not stable in blood and neither are oligolysine/oligoarginine (5) , rendering these CPPs unfavorable as transporters for therapeutic AOs. We reasoned that one approach to improve stability would be to use non-a amino acids or D-amino acids. In this study, we investigated whether incorporation of 6-aminohexanoic acid (X), b-alanine (B) and D-arginine (r) amino acids into the CPP would affect cellular delivery, antisense activity, toxicity and serum binding of the resulting CPPÀPMO conjugates. We found that CPPÀPMOF conjugates containing X/B residues did not enter cells as efficiently as R 8 À and R 9 ÀPMO conjugates. This is consistent with our previous finding for the (RXR) 4 conjugate (6). We have found that cell surface proteoglycans were involved with binding of the TatÀ, R 9 F 2 À and (RXR) 4 ÀPMO conjugates with the (RXR) 4 conjugate having the lowest binding affinity. Insertion of X into an oligoarginine CPP reduces the charge density and may lead to decreased binding affinity for proteoglycans. Despite the lower cellular uptake of X/B-containing CPP-PMO, they generated higher antisense activities in the cell nucleus than oligoarginineÀPMO. We have found that endocytosis was the internalization mechanism (at least primarily) for oligoarginine-and (RXR) 4 ÀPMO conjugates. Indication of different uptake mechanisms was not found among these conjugates (6) . Therefore we hypothesize that X/B-containing conjugates have a greater ability to escape from endosomes/lysosomes than oligoarginine conjugates by a mechanism as yet to be studied. The number of X residues affects both the nuclear antisense activity and the toxicity of conjugates. The CPPÀPMO conjugate with 8 X residues [(RX) 8 ÀPMO] had the highest activity followed by one with 5 Xs [(RXR) 4 ÀPMO] (Figures 1 and 4) . However, these conjugates were toxic to cells at higher concentrations, which may be a concern when considering potential applications for in vivo delivery of PMO. Replacement of all 8 Xs with Bs decreased both toxicity and antisense activity. The combination of 3-4 Xs with several B residues yielded CPPs with no detectable toxicity, and at some concentrations several of them had similar antisense activity as (RX) 8 ÀPMO. We think this type of CPP, having Bs and fewer than 5 Xs, will offer balanced activity and low toxicity as well as the stability, and have considerable potential for delivery of therapeutic AOs. Further investigation into the toxicity and activity versus dosing levels of these CPPs in vivo is warranted. Surprisingly, the replacement of L-arginine with D-arginine enhanced neither uptake nor antisense activity for oligoarginine, or X-and B-containing conjugates. In the case of (RX) 8 ÀPMO, the replacement actually caused a small but statistically significant decrease in activity. Our observation is different from the results reported by others (20, 21) who found that D-CPPs had higher cellular uptake than L-CPPs, although no biological functional cargo was used in their study. The difference between results may be due to the type and size of cargos and the cell lines used for the assays. Whether the use of D-arginine-containing peptides results in superior CPPÀPMO functional activity in vivo remains to be tested. We attempted to understand the nature of (RX) 8 À and (RXR) 4 ÀPMO toxicity. It is apparent that these two conjugates caused little immediate membrane damage with 0.5 or 5 h treatment at concentrations as high as 60 mM (Figure 8 ). However, these two conjugates had dose-dependent toxicity with 24 hr treatment as shown by the leaky cell membranes and fewer cells compared to controls ( Figure 6C&D, Figure 8 ). Interestingly, the replacement of Xs with Bs in (RX) 8 ÀPMO abolished the toxicity, and the replacement of L-arginine with D-arginine reduced the toxicity of (RX) 8 ÀPMO ( Figure 6 ). We have found that (rX) 8 ÀPMO was completely stable and the peptide portion of (RB) 8 ÀPMO was only partially degraded, whereas the peptide portion of (RX) 8 ÀPMO was completely degraded in cells (5) . We wondered whether the difference in toxicity among (RX) 8 , (RB) 8 and (rX) 8 ÀPMO conjugates was caused by differences in intracellular stability, resulting in the metabolized products of (RX) 8 ÀPMO producing toxicity. The identifiable metabolized products of (RX) 8 ÀPMO were XRXBÀPMO and XBÀPMO (5) but neither product had any detectable toxicity as measured by MTT assay (data not shown). It is possible that the CPP portion was degraded into free amino acids and/or smaller peptide fragments which were toxic. However, our investigation revealed that neither free R nor X, alone or in combination, caused cellular toxicity. Another possibility is that because of the high hydrophobicity of X compared to B, X in combination with positively charged arginine residues leads to toxicity not generated by B residue combinations. However, this explanation does not account for the difference in toxicity observed between (RX) 8 ÀPMO and (rX) 8 ÀPMO, which have the same hydrophobicity. Perhaps the toxicity of (RXR) 4 ÀPMO and (RX) 8 ÀPMO was caused by the peptide fragments that we could not identify by mass spectrometry. Unlike the toxicity difference between (RX) 8 À and (rX) 8 ÀPMO, the L!D replacement did not change the toxicity of (RXR) 4 ÀPMO ( Figure 6 ). Substitution of either one R (rXR) or two R (rXr) from the RXR repeat neither reduced nor increased the toxicity profile of (RXR) 4 ÀPMO. At this point, we do not fully understand the mechanisms of (RXR) 4 À and (RX) 8 ÀPMO conjugate toxicity, but look forward to studying this topic further. Serum effect on the activity of a CPPÀAO conjugates is an important issue when considering potential in vivo applications. X/B-containing conjugates were still active in 60% serum while oligoarginine conjugates were not. The greater stability of the X/B-containing conjugates to serum enzymes is likely a factor contributing to their high activity. The loss of activity in high serum concentrations makes oligoarginine CPPs undesirable as potential therapeutic AO carriers. In summary, we have found that the X/B-containing CPPÀPMO conjugates are superior to oligo-arginineÀPMO conjugates for the following reasons: they display higher activity in cell nuclei, are less affected by serum and are more stable in blood (5) . The toxicity of the X/B-containing CPPs can be reduced by keeping the number of X residues below 5 while still maintaining a reasonable delivery efficacy and stability. This study provides a basis for further optimization of CPP sequence using R, r, X and B residues in the interest of further reducing toxicity and increasing antisense activity, which will likely lead to more effective AO transporters for potential therapeutic applications.

Lost in the World of Functional Genomics, Systems Biology, and Translational Research: Is There Life after the Milstein Award?

We've always wanted to save the world from the scourges of virus infection by developing better drugs and vaccines. But fully understanding the intricacies of virus-host interactions, the first step in achieving this goal, requires the ability to view the process on a grand scale. The advent of high-throughput technologies, such as DNA microarrays and mass spectrometry, provided the first opportunities to obtain such a view. Here we describe our efforts to use these tools to focus on the changes in cellular gene expression and protein abundance that occur in response to virus infection. By examining these changes in a comprehensive manner, we have been able to discover exciting new insights into innate immunity, interferon and cytokine signaling, and the strategies used by viruses to overcome these cellular defenses. Functional genomics may yet save the world from killer viruses.

At the time I received the Milstein Award in September 1999, my laboratory was embarking on an exciting new venture. I had become fascinated by the novel opportunities that were becoming available through the sequencing of the human genome and the advent of highthroughput technologies such as DNA microarrays and mass spectrometry [1] . Because these technologies provided a way to obtain a nearly comprehensive view of a biological system, I was convinced they offered the best chance to learn everything there is to know about the complex interplay between viruses and the cells they infect. Over the past eight years, my laboratory has worked tirelessly, and even obsessively, to incorporate genomic, proteomic, and information technologies into our research. Although we certainly don't yet know everything there is to know about virus-host interactions, we've made tremendous progress, and I remain convinced that genomic and proteomic approaches will be a fundamental force for pushing the field of virology forward in the years to come. In this article, we provide a sampling of how we're using these technologies, the discoveries we've made, and how the technologies themselves have continued to evolve. Our initial foray into the world of genomics began with cDNA microarrays that were spotted in-house using a commercially available collection of cDNA clones. Getting this new technology up and running required months of work. But after seemingly endless frustration, Gary Geiss was finally able to use these arrays, representing all of 1,500 human genes, to evaluate the changes in cellular gene expression that occurred in a CD4+ T-cell line infected with human immunodeficiency virus type 1 (HIV-1) [2] . This was a straightforward study in which we analyzed two time points after infection, and we were thrilled to observe the temporal regulation of genes involved in T-cell signaling, subcellular trafficking, and transcriptional regulation. Despite how simplistic this study now appears, it was the first ever to apply genomics to AIDS research and it was the cover article when published in Virology. With the technology now in place, we were eager to apply it to a broad range of experimental systems. The possibilities seemed endless (as indeed they are), and we rapidly began "doing arrays" on cells infected with a variety of different viruses or exposed to double-stranded RNA (dsRNA), interferon, or other treatments ( Figure 1 ). Our next study was a quantum leap in sophistication. Our microarrays now contained over 4,600 cDNAs and our experimental system was more complex. This time, we profiled the cellular gene expression changes that occurred in HeLa cells infected with influenza virus, again analyzing two time points after infection [3] . In addition, we evaluated the cellular response to a heat-inactivated virus, allowing us to identify gene expression changes that were independent of viral replication. Genes involved in protein synthesis, transcriptional regulation, and cytokine signaling were induced by the replicating virus, and many of these changes suggested an active cellular antiviral response. Interestingly, viral replication also resulted in the down-regulation of many cellular genes, particularly at the later time point. The gene expression pattern induced by the inactivated virus included the induction of the cellular metallothionein genes, which may represent a protective response to virus-induced oxidative stress. Our early use of microarrays also included the first systematic attempt to define the full repertoire of dsRNA-regulated genes [4] , and we used arrays, in conjunction with polysome fractionation, to identify cellular and viral mRNAs that increased in abundance and which were selectively recruited to polyribosomes following influenza virus infection [5] . Our dsRNA study identified 175 dsRNA-stimulated genes, which encoded proteins involved in a broad range of cellular functions and metabolic pathways. In addition, this study showed for the first time that many genes are also down-regulated by dsRNA. Our polysome fractionation studies, headed by John Kash, revealed that many of the cellular mRNAs that are recruited to polyribosomes during influenza virus infection are involved in the control of cell proliferation and survival, mitogenesis, or the inflammatory response. In addition, many of these mRNAs contain a putative binding site for GRSF-1, a cellular mRNA-binding protein that also specifically interacts with a discrete and conserved region of the viral 5′ untranslated region [6] . As our use of microarrays extended to an increasing number of experimental systems, we quickly found ourselves having to cope with an avalanche of data. To manage the huge amounts of information generated from multiple array experiments, our bioinformatics group, then headed by Jeff Furlong, developed a custom gene expression database, called Expression Array Manager [7] . In addition to serving as a central data repository, Expression Array Manager functions as our laboratory information management system, provides a streamlined data analysis pipeline, and provides a mechanism for publishing our gene expression data to the scientific community (available at http://expression.viromics.washington.edu). Most of our core genomic and proteomic data analysis functions are now performed using the Rosetta Biosoftware systems Resolver and Elucidator. These systems are designed for managing and analyzing large amounts of gene expression and mass spectrometry data and both contain a suite of higher-order analysis tools. To explore genomic and proteomic data in greater depth, we also use a variety of tools that provide a diversity of analytical choices. As an example, we use Spotfire DecisionSite to format data for further analysis using tools such as Ingenuity Pathway Analysis or MetaCore (GeneGo). These tools allow us to leverage proprietary biological content databases and examine gene expression (or protein abundance) data in the context of complex pathways to identify connections between differentially expressed genes. Sean Proll, Bryan Paeper, and Jon Rue provide the laboratory with expert guidance on the use of these tools and ensure that data moves efficiently through this complex pipeline. We had earlier discovered that hepatitis C virus (HCV) uses its nonstructural 5A (NS5A) protein to repress the interferon-induced protein kinase, PKR, a primary mediator of the antiviral effects of interferon [8] . Our first application of gene expression profiling to HCV research was therefore to evaluate the effects of interferon treatment on cellular gene expression and to determine whether NS5A was able to alter this pattern of expression [9] . Gary Geiss again headed these studies in which we examined the effects of interferon treatment by using several types of human cells, including HeLa cells, liver cell lines, and primary fetal hepatocytes. In response to interferon, 50 of the approximately 4,600 genes examined were consistently induced in each of these cell types and another 60 were induced in a cell typespecific manner. A search for interferon-stimulated response elements (ISREs) in genomic DNA located upstream of interferon-stimulated genes revealed both previously identified and novel putative ISREs. The expression of NS5A partially blocked the interferon-mediated induction of interferon-stimulated genes and inhibited the induction of a reporter gene driven from an ISRE-containing promoter, further suggesting that NS5A may play a role in HCV resistance to interferon. This study also set the stage for our use of genomics to better understand the interferon response and the varied mechanisms used by viruses to evade the effects of interferon. What really accelerated our HCV genomics program was being awarded a National Institute on Drug Abuse (NIDA) P30 Center to do translational and clinically relevant genomics studies (http://nida.viromics.washington.edu/). Maria Smith headed many of our early studies, the first of which looked for novel tumor markers in surgical liver samples from patients with hepatocellular carcinoma (HCC) [10] . This study, which by now used microarrays containing over 13,000 human cDNAs, revealed a set of 50 potential HCC marker genes that were upregulated in the majority of the tumors analyzed. This HCC marker set contained several cancer-related genes as well as a set of genes encoding secreted or plasma proteins that may provide potential serological markers for HCC. We performed a similar set of analyses on surgical material and core biopsy specimens obtained from HCV-infected patients with liver cirrhosis [11] . Importantly, this study also included an analysis of normal liver samples to determine normal physiologic variation in gene expression. To identify markers associated with cirrhosis, genes that were differentially expressed in normal liver, or in HCC, were subtracted from the set of genes differentially expressed in cirrhotic livers. The resulting gene set included genes expressed in activated lymphocytes infiltrating the cirrhotic liver and genes involved in the remodeling of the extracellular matrix, cell-cell interactions associated with cytoskeleton rearrangements, the anti-apoptotic pathway of Bcl-2 signaling, and the interferon response. Together, this analysis identified several potential gene expression markers of HCV-associated liver disease and contributed to our rapidly expanding database of experiments describing HCV pathogenesis. Perhaps one of our most promising applications of genomics to HCV-associated liver disease is our work with serial liver biopsy samples from HCV-infected liver transplant patients. Liver transplant recipients infected with HCV develop recurrent hepatitis soon after transplantation and, in some cases, progress to fibrosis within two years of the transplant. Our goals are to identify molecular processes influencing liver disease progression and to find potential gene expression markers of early fibrosis. To achieve these goals, we are working closely with clinicians in the liver transplant unit at the University of Washington, including Anne Larson, Robert Carithers, and James Perkins, to collect biopsy samples and to integrate the gene expression data obtained from these samples with clinical observations. Our initial analyses were performed on serial liver biopsy specimens obtained from 13 transplant recipients at 0, 3, 6, and 12 months after transplantation [12] . Gene expression data were compared with clinical observations and with gene expression data obtained from 55 nontransplant HCV-infected and uninfected liver samples. Our analyses revealed several specific gene expression patterns, the first of which was unique for the transplant recipients regardless of their infection status. The corresponding genes encoded stress response proteins and blood proteins involved in coagulation that were differentially expressed in response to post-transplantation graft recovery. The second pattern was specific to HCV-infected samples and included the increased expression of genes encoding components of the interferonmediated antiviral response and immune system. This pattern was absent or suppressed in the patients who developed early fibrosis, indicating that disease progression might result from an impaired liver response to infection. Finally, we identified gene expression patterns that were specific for the 12-month biopsy specimens of all four HCV-infected patients who developed early fibrosis after transplantation. The identified gene expression patterns may prove useful for diagnostic and prognostic applications in HCV-infected patients, including predicting early progression to fibrosis. We are continuing to collect and analyze biopsy samples as new patients are recruited into the study, which should enable us to obtain increasing levels of confidence in the predictive power of the markers that we identify. In other recent studies, Kathie Walters and Sharon Lederer have examined the gene expression profiles that differentiate alcohol-and HCV-induced liver cirrhosis [13] and the gene expression patterns associated with HCV-induced pathogenesis in individuals co-infected with HIV-1 [14] . We found that global gene expression patterns vary significantly depending upon the etiology of liver disease and that stages of liver cirrhosis can be differentiated based on gene expression patterns in ethanol-induced, but not HCV-induced, disease. Many of the gene expression changes specifically observed in HCV-infected cirrhotic livers are associated with activation of the innate immune response. In contrast, we found that intrahepatic global gene expression profiles do not differ between HCV-and HCV/HIV-1 co-infected individuals. However, a specific gene expression pattern that may be associated with HCV-induced pathogenesis was identified. This pattern has similarities to the gene expression profiles in transplant patients who progress to fibrosis within one year of transplantation, suggesting it may also be relevant to predicting disease progression. Although clinical samples can provide considerable insight into the changes in cellular gene expression that occur during liver disease progression, they are not necessarily ideal for studies investigating the molecular mechanisms responsible for these changes. These types of studies typically require an animal model that can be experimentally manipulated. In this regard, research in the HCV field has been significantly enhanced by the development of a mouse model of HCV infection. This model, developed by Lorne Tyrrell and Norman Kneteman at the University of Alberta, Canada, is based on the severe combined immunodeficiency disorder (SCID)-beige/albumin (Alb)-urokinase plasminogen activator (uPA) transgenic mouse. Intrasplenic injection of freshly isolated human hepatocytes into these mice leads to the repopulation of the mouse liver with human hepatocytes. The result is a mouse containing a chimeric mouse-human liver and these animals can be persistently infected with a variety of HCV genotypes [15] . The chimeric mouse model provides a number of features that make it a useful system in which to study HCV-host interactions. Because groups of mice are transplanted with hepatocytes from different donors, we have the opportunity to analyze host-specific responses to HCV. In addition, the lack of an adaptive immune response in these animals makes it possible to distinguish between virus-mediated and immune-mediated effects on gene expression. Kathie Walters is using this model to characterize the host transcriptional response to HCV infection. Because liver samples from these animals typically have small percentages of contaminating mouse liver tissue, she first evaluated the level of cross-hybridization of mouse liver mRNA to the human probes present on the microarray [16] . The small set of genes identified (less than 2% of the genes present on the array) is either removed from subsequent data analysis or changes in the expression of these genes are validated using human-specific quantitative realtime PCR. In an initial series of experiments, groups of mice transplanted with hepatocytes from different donors were inoculated with a single source of HCV and gene expression profiling was performed to characterize the host response to infection [17] . Although all HCV-infected animals showed evidence of an interferon response, the level of this response, both in the intensity and number of up-regulated interferon-induced genes, varied between animals depending upon the origin of the donor hepatocytes. These results indicate that host genetic factors contribute to the variation in the host response to HCV, including the activation of innate antiviral signaling pathways. Mice with weak interferon responses also tended to have high levels of intrahepatic HCV RNA, indicating that an ineffective interferon response may allow increased levels of viral replication. These animals also accumulated higher numbers of differentially expressed genes than did mice with strong interferon responses. Interestingly, we have also observed impaired interferon responses and the accumulation of increased numbers of differentially expressed genes in the intrahepatic gene expression profiles of transplant patients who develop fibrosis within one year of transplantation. HCV research and the NIDA P30 Center also provided our entry point into proteomics. The ability to look at the protein content of a cell provides a strong complement to our genomic analyses, allowing us to evaluate how gene expression changes in response to virus infection translate into changes in the abundance of proteins, the molecules that directly carry out biological functions. Moreover, proteomic analysis of body fluids, such as blood, holds the promise of identifying candidate biomarkers for human diagnostic or prognostic applications. Our initial venture into the realm of proteomics was led by Wei Yan in collaboration with Ruedi Aebersold at the Institute of Systems Biology. We began by using tandem mass spectrometry to analyze the global proteome of the Huh7 hepatoma cell line and cultured primary and immortalized human fetal hepatocytes [18] . We also used isotope coded affinity tag (ICAT) technology to measure the changes in protein abundance that occurred in human hepatocytes in response to interferon treatment [19] . These analyses led to the generation of a liver proteome dataset consisting of 2,159 unique proteins. Among the identified proteins were 78 interferon-regulated proteins that play roles in a multitude of cellular functions including antiviral defense, immune response, cell metabolism, and signal transduction. These data also contributed to the development of PeptideAtlas, a public resource for further annotating and validating the human genome by mapping identified peptide sequences to the human genome sequence [20] . For our more recent proteomic work, we have teamed with Richard Smith at Battelle, Pacific Northwest National Laboratory (PNNL). In our first studies with the Smith group, we used high mass accuracy Fourier transform ion cyclotron resonance (FTICR) mass spectrometry, coupled with the accurate mass and time (AMT) tag approach [21] , to perform global proteomic analyses on Huh-7.5 cells containing a full-length HCV replicon [22] . Using this more sensitive technology, we identified over 4,200 proteins. As a first-pass means to detect changes in protein abundance associated with HCV infection, we also performed a semi-quantitative comparison of total peptide identifications. This peptide-spectral count ("peptide-hits") approach revealed HCV-related perturbations in the abundance of a variety of proteins associated with lipid metabolism and oxidative stress. We have also taken advantage of the nanoproteomic platforms available at PNNL to perform proteomic studies on human liver biopsy samples that yield only limited amounts of protein. These studies were performed using a high-sensitivity 11.4-tesla FTICR mass spectrometer coupled with the AMT tag approach and our Huh-7.5 cell protein database. Using this method, we identified over 1,500 proteins from only 2 μg of a protein digest obtained from a liver biopsy sample [22] . The proteins identified included many of relevance to HCV infection and liver disease, including cellular proteins involved in lipid metabolism and the interferon response. The number of proteins identified by this approach is considerably larger than what has been reported in other published studies of the liver proteome, where less-sensitive methods, requiring milligram amounts of protein lysate, have identified only a few hundred proteins. Deborah Diamond leads the proteomics research in our laboratory and has recently authored a review that more fully discusses current developments in the application of proteomics to liver disease research [23] . Influenza research has been a mainstay of the laboratory for over 20 years, and as mentioned earlier, some of our earliest genomic studies were performed on cells infected with influenza virus. Today, with worldwide attention focused on highly pathogenic avian influenza and the looming threat of a new pandemic, influenza research has taken on a renewed vigor. One of the priorities in this research is the development of improved animal models to better understand influenza pathogenesis and to develop new diagnostic, therapeutic, and vaccine strategies, all of which are important components of preparedness for the next influenza pandemic [24] . Over the past two years, we have taken on this challenge by developing a macaque influenza infection model that features genomic and proteomic analyses as a key component of the experimental design. Carole Baskin has taken a leading role in this effort and was the first to report on the use of genomic technologies to characterize experimental influenza virus infection using a pig-tailed macaque model [25] . This study was performed using a mildly pathogenic strain of virus (A/Texas/36/91) with the goal of constructing a blueprint of an uncomplicated influenza infection. Gene expression profiling performed on lung tissue and tracheobronchial lymph nodes revealed that numerous genes associated with the interferon response were differentially expressed, particularly at day 4 after inoculation. We also observed an increase in the expression of genes encoding various mediators of chemotaxis, adhesion, and the transmigration of immune cells, suggesting the trafficking of these cells into the lungs and tracheobronchial lymph nodes, which was confirmed by histopathology. In addition, this study revealed gene expression changes relevant to the processing of antigens on MHC class I complexes, particularly in lung tissue, and many genes encoding proteins relevant to T-cell function were induced at one or more time points. The gene expression profiling performed for this study was done using human cDNA microarrays, which until recently was the only option available for studies using nonhuman primates. However, because of the extensive use of macaques as models for a wide variety of human diseases (and in AIDS research in particular), we have been part of a push to develop genomic resources focused on macaque species [26] . This effort resulted the sequencing of the rhesus macaque genome [27] , and James Wallace headed an effort to use this extensive sequence information to develop an oligonucleotide microarray containing over 17,000 unique macaque sequences [28] . Additional information related to nonhuman primate genomics is also available at macaque.org, a Web site we have developed to disseminate macaque genomic and proteomic data and resources. The macaque oligonucleotide array was used in a follow-up study in which we focused on the early innate immune response of macaques infected with influenza virus [29] . In this study, led by Tracey Baas and Carole Baskin, we also performed genomic analyses on whole blood samples from infected animals to determine whether we could detect gene expression signatures in the blood that would correlate with those detected in the lung. In general, genomic analyses of lung tissue showed an increase in the expression of many genes associated with interferon signaling early after infection, while other immune and cytokine responses were sustained throughout the course of the infection. We also observed a number of genes (many of which were involved in the interferon response) that were differentially expressed in a similar fashion in lung and blood samples, particularly early after infection. This finding suggests that microarray analyses of blood samples may hold promise for diagnostic applications. Finally, this study also included a first of its kind global proteomic analysis of macaque lung tissue. This analysis identified over 3,500 proteins, and consistent with the gene expression data, proteomic data also revealed an increase in the abundance of many proteins involved in the innate immune response. What influenza virologist wouldn't be interested in understanding the molecular mechanisms underlying the extreme pathogenesis of the virus responsible for the 1918 influenza pandemic? Studies related to this virus, which killed and estimated 50 to 100 million people worldwide [30] , have been (and remain) a prominent fixture of the laboratory, and genomic analyses are providing important new insights into what made this virus so lethal. In our first experiments associated with the 1918 virus, we evaluated global gene expression patterns in cells infected with wild-type influenza, engineered viruses lacking all or part of the NS1 gene, or with a virus containing the NS1 gene of the 1918 pandemic virus [31] . Deletion of the NS1 gene increased the magnitude of expression of cellular genes implicated in the interferon, NF-κB, and other antiviral pathways, and a virus with a C-terminal deletion in its NS1 gene induced a cellular gene expression pattern intermediate between the patterns induced by wild-type and NS1 knockout viruses. In contrast, a virus containing the NS1 gene from the 1918 pandemic virus was more efficient at blocking the expression of interferon-regulated genes than its parental virus. Together, these results suggested that the cellular response to influenza virus is significantly influenced by the NS1 gene and that the 1918 NS1 is a particularly effective interferon antagonist. Of course, we have since learned that the story is much more complicated. We next used a mouse infection model to assess the contribution of the 1918 HA and NA genes to viral pathogenesis [32] . These studies, led by John Kash, evaluated mouse-adapted A/WSN/ 33 viruses that were engineered to contain the HA and NA genes from the 1918 virus or from the nonlethal A/New Caledonia/20/99 virus. Microarray analyses performed on lung tissues isolated from all infected animals showed the activation of many genes involved in the inflammatory response, including cytokine, apoptosis, and lymphocyte genes. However, consistent with histopathology analysis, the parental and 1918 HA/NA:WSN recombinant viruses showed increased expression of genes associated with activated T cells and macrophages, as well as genes involved in apoptosis, tissue injury, and oxidative damage that were not observed in mice infected with the New Caledonia HA/NA:WSN virus. These studies documented clear differences in gene expression profiles that were correlated with pulmonary disease pathology and suggested that an intrinsic property of the 1918 HA and NA proteins may be the production of a longer and more severe immune response culminating in a more destructive viral infection. The reconstruction of all eight gene segments of the pandemic virus provided the opportunity to find out what genomics could reveal about the high-virulence phenotype of the 1918 virus. The initial characterization of this virus by Terrence Tumpey and colleagues demonstrated that it highly virulent in mice and most animals died within five days after infection [33] . Genomic analyses, led by John Kash, showed that animals infected with the reconstructed virus showed an increased and accelerated activation of genes associated with pro-inflammatory and celldeath pathways by 24 h after infection [34] . Significantly, these genes remained activated until the death of the animal. This was in contrast to less dramatic and delayed host immune responses (and less severe disease pathology) in mice infected with influenza viruses containing only subsets of genes from the 1918 virus. These findings indicate a cooperative interaction between the 1918 influenza genes and suggest that enhanced inflammatory and celldeath responses may contribute to severe immunopathology. Our most recent work with the reconstructed 1918 virus was done in collaboration with Yoshihiro Kawaoka. In these studies, we evaluated the host response to the reconstructed virus using a cynomolgus macaque infection model [35] . The 1918 virus replicated to high levels and spread rapidly throughout the respiratory tract of infected animals, causing severe damage and masses of infiltrating immune cells throughout the course of infection. Genomic analyses revealed that the 1918 virus triggered an aberrantly high and sustained expression of numerous genes involved in the innate immune response, including proinflammatory cytokines and chemokines. The early and sustained host response in macaques infected with the 1918 virus was similar to what we observed in mice, indicating that critical host responses that influence disease outcome may occur very early after infection. Interestingly, the 1918 virus also appeared to selectively attenuate the expression of specific innate immune response genes, including certain genes associated with the type 1 interferon response. We have not determined the mechanism for this attenuation, but it is possible that the viral NS1 gene may play a role in regulating this response. Although the details remain to be uncovered, the atypical innate immune response induced by the 1918 virus was insufficient for protection and may actually contribute to the lethality of the virus. Our interests in combining genomics and macaque infection models have led to the establishment of the Division of Functional Genomics and Infectious Disease at the Washington National Primate Research Center. Our laboratory is the central component of this Division, which is devoted to developing genomic and proteomic technologies (along with Richard Smith's group at PNNL), immunologic resources (in collaboration with Edward Clark and Murali-Krishna Kaja), and advanced computing and bioinformatics applications to enhance the use of nonhuman primates as models for influenza virus infection and AIDS. Our goal is to develop and apply the tools needed to perform comprehensive and integrated analyses on nonhuman primates and to analyze the response to virus infection at multiple points along the flow of biological information: from the whole animal to DNA to RNA to protein to biological function (Figure 2 ). By integrating these diverse types of data, we have the opportunity to better understand the dynamics of the host response to infection and the molecular mechanisms underlying the progression to virus-mediated disease, immunopathology, or the development of protective immunity. We also have the opportunity to better understand how gene expression changes in response to infection translate into changes in protein abundance and function, and how these changes correlate with clinical outcome. Moreover, we have the opportunity to assess how changes in gene expression and protein abundance affect immune cell function, and how the innate immune response develops and its link to adaptive immunity. Ultimately, we believe this integrated approach will translate into molecular signatures that predict protective immunity or pathology, biomarkers for diagnostic or prognostic assays, and a rational base for improvements to antiviral therapies or vaccine strategies. The studies outlined in this review provide just a glimpse of how we're using genomic and proteomic technologies to unravel the complexities of virus-host interactions. In addition to work with HCV and influenza, we're using genomic approaches to study a variety of other viruses. Tracey Baas heads collaborative research on SARS coraonavirus [36] and herpes simplex virus [37] . When not working on pandemic influenza, John Kash also leads efforts focused on Ebola virus [38] and vaccinia virus [39] , and Matthew Thomas and Michael Agy have explored gene expression changes associated with simian immunodeficiency virus infection [40] . We've also worked with Michael Gale on West Nile virus [41] , and in our work on P58 IPK (a cellular PKR inhibitor), we've even explored gene expression patterns associated with pancreatic β-cell depletion in diabetic mice [42] . Jamie Fornek is continuing the work on P58 IPK , including using P58 IPK knockout mice and genomics to examine the role of this PKR inhibitor during influenza virus infection. Gregory Zornetzer is developing a targeted proteomics approach to identifying protein-protein interactions associated with key viral proteins. The technologies we use continue to improve and data is accumulating at an ever increasing rate. Today's microarrays represent over 18,000 genes (over ten-fold more than when we started) that together with assorted controls are arrayed as an impressive 44,000 oligonucleotides per slide. Barry Robinson is working to incorporate laser capture microdissection into our studies, which will give is our first opportunity to examine the gene expression changes that occur in specific cell types isolated from complex tissues. Bioinformatics capabilities are becoming evermore sophisticated and Jon Rue, Eric Flamoe, and Matthew Harding ensure that we stay atop the most recent advances in data management and analysis. Our Expression Array Manager database now contains information from over 7,000 expression profiles representing more than 50 million measurements of differential gene expression. As we've discussed elsewhere [43] , the future of virology depends on making progress in understanding how things work at a systems level, and we must ensure that genomic and proteomic technologies do more than simply generate ever larger quantities of data without providing clues to underlying function. Nevertheless, we are confident that as technologies and experimental systems continue to progress, and data integration strategies become more mature [44] , functional genomics will make ever greater contributions to virology. As an example, we are continuing to take our use of genomics in new directions and Robert Palermo is working together with Marjorie Robert-Guroff (National Cancer Institute) to pursue genomic approaches to improving AIDS vaccine development. Genomic analyses during vaccine trials may reveal gene expression markers of protective immunity or gene expression changes that are indicative of a predisposition to a particular response to immunization and subsequent challenge. All told, the past eight years have been an exciting journey full of discoveries, and we look forward to new breakthroughs in the future. Indeed, life after the Milstein award hasn't been too bad. Representation of the range of viruses and experimental systems we have evaluated using genomic and proteomic technologies. Highlights of experiments related to the use of many of these experimental systems, particularly for hepatitis C virus and influenza, are summarized in this review. SARS-CoV: severe acute respiratory syndrome-associated coronavirus; SIV: simian immunodeficiency virus; WNV: West Nile virus; HIV-1: human immunodeficiency virus type 1; HSV-1: herpes simplex virus type 1; EAM: Expression Array Manager. The Division of Functional Genomics and Infectious Disease provides the virus infection models and the clinical, genomic, proteomic, and immunologic resources needed to perform comprehensive analyses on nonhuman primates. This allows us to analyze the response to virus infection at multiple points along the flow of biological information. EST: expressed sequence tag; AMT: accurate mass and time.

Electrospray ionisation-cleavable tandem nucleic acid mass tag–peptide nucleic acid conjugates: synthesis and applications to quantitative genomic analysis using electrospray ionisation-MS/MS

The synthesis and characterization of isotopomer tandem nucleic acid mass tag–peptide nucleic acid (TNT–PNA) conjugates is described along with their use as electrospray ionisation-cleavable (ESI-Cleavable) hybridization probes for the detection and quantification of target DNA sequences by electrospray ionisation tandem mass spectrometry (ESI-MS/MS). ESI-cleavable peptide TNT isotopomers were introduced into PNA oligonucleotide sequences in a total synthesis approach. These conjugates were evaluated as hybridization probes for the detection and quantification of immobilized synthetic target DNAs using ESI-MS/MS. In these experiments, the PNA portion of the conjugate acts as a hybridization probe, whereas the peptide TNT is released in a collision-based process during the ionization of the probe conjugate in the electrospray ion source. The cleaved TNT acts as a uniquely resolvable marker to identify and quantify a unique target DNA sequence. The method should be applicable to a wide variety of assays requiring highly multiplexed, quantitative DNA/RNA analysis, including gene expression monitoring, genetic profiling and the detection of pathogens.

Mass spectrometry has much to offer the field of genomic analysis, particularly in terms of multiplexed analysis and accurate quantification. To date, many mass spectrometry-based approaches for genomic analysis have been based on direct detection of nucleic acids particularly using matrix assisted laser desorption ionisation time-offlight (MALDI TOF) MS analysis. MALDI TOF is well suited to this approach due to the high mass range achievable by TOF analysis, however, MALDI TOF instrumentation is relatively expensive and sample preparation can be quite laborious. In addition, direct analysis of nucleic acids by mass spectrometry suffers from problems such as depurination leading to fragmentation (1) or cation adduct formation (2, 3) . These issues notwithstanding, MALDI TOF analysis of nucleic acids has been applied to DNA sequencing (4, 5) , RNA sequencing (6, 7) , analysis of DNA tandem repeats (8) . In particular, PNA probes, with and without noncleavable mass modifiers, have been used for characterization of genomic DNA libraries (9) , detection of DNA methylation (10) and detection of single nucleotide polymorphisms (SNPs) (11, 12) by MALDI TOF mass spectrometry. Electrospray ionisation (ESI) mass spectrometry has also been used for direct detection of nucleic acids (13) (14) (15) . ESI-MS has some advantages over MALDI TOF MS particularly the availability of lower cost instrumentation. Moreover, sample handling can be simpler since most molecular biology assays are carried out in solution and such liquid samples are injected directly into the instrument. Furthermore, very high molecular weight species can be analysed due to the propensity for large molecules, such as PCR products, to form multiply charged ions that have relatively low overall mass-to-charge ratios under electrospray ionisation conditions. However, direct analysis of nucleic acids by ESI-MS still suffers from the same problems as MALDI TOF MS such as cation adduct formation (14) . In addition, the multiply charged ion spectra that are generated for large nucleic acid fragments can be very complicated reducing sensitivity and making multiplexing difficult (16) . As an alternative to direct analysis of nucleic acids, mass spectrometry can also be used to detect nucleic acids indirectly through the use of cleavable mass tags, which avoids many of the limitations of direct analysis of nucleic acids while also offering numerous advantages such as ease of multiplexing, more robust detection of tag species and higher sensitivity and less demanding workup and sample preparation for analysis. Mass spectrometric analysis of mass tags, by ESI or MALDI also offers the possibility of accurate quantification through the use of isotopic tags. This capability of mass spectrometry has not really been exploited in genomic analysis but has been quite widely used in proteomic analysis (17) (18) (19) and is a key advantage of the mass spectrometric approach. Again MALDI TOF analysis of nucleic acids with cleavable mass tags has been demonstrated by various groups (20) (21) (22) (23) (24) but it would be advantageous to be able to take advantage of lower cost ESI-MS/MS instruments and to avoid the laborious sample workup requirements of most MALDI approaches. A matrix-free laser desorption approach, which has reduced workup requirements has been demonstrated (21, 25) but this still requires that the sample be spotted onto a MALDI target or hybridized to an array. To our knowledge, only one mass tagging approach employing ESI-MS analysis has been demonstrated (26) . In this approach, mass tags are photocleavably linked to oligonucleotides and tag detection requires a photo-cleavage step and a tag isolation step outside of the mass spectrometer prior to tag detection, i.e. the workup is not much simpler than that required for MALDI TOF analysis. Here, we describe the synthesis of novel ESI-cleavable Tandem Nucleic acid mass Tag-peptide nucleic acid conjugates and their analysis by ESI-MS/MS. We demonstrate a novel mass tag cleavage method in which source voltages in the electrospray ionisation source are used to cleave an ESI-cleavable linker, by a collision-based process, releasing the mass tag from the oligonucleotide during ionization. This method allows for direct analysis of assay solutions without requiring complex workups to cleave and isolate tags. In principle, this cleavage method would also allow in-line separation, by capillary electrophoresis for example, of labelled nucleic acids with direct spraying of the separated material into the ion source where tag cleavage would take place automatically. In addition, we demonstrate a novel MS/MS-based tandem nucleic acid mass tag (TNT) design and detection process that allows highly specific detection of TNTs in a complex background. The Tandem Nucleic acid mass Tag design also allows easy synthesis of large sets of isotopic tags supporting the development of multiplexed and quantitative assays. We demonstrate the quantitative nature of the TNT approach. Furthermore, we evaluate ESI-cleavable TNT-PNA conjugates as hybridization probes for the detection of target DNA sequences via the use of ESI-MS/MS. The TNTs described here are constructed using FMOC peptide synthesis chemistry. Synthesis of peptide nucleic acid-TNT peptide conjugates is relatively straightforward as PNA synthesis can be achieved using the same FMOC protection groups that are used in peptide synthesis (27) . This means that peptide TNT-PNA conjugates can be synthesized on the same resin in a continuous process. PNA is a useful analogue of DNA and is advantageous for this application due to its high specificity and its neutral backbone, which means that it does not require high concentrations of salt to hybridize making it highly compatible with mass-spectrometry-based detection methods (28) (29) (30) . The outline of the general approach for the detection of DNA sequences via ESI-cleavable TNT-PNA conjugates is presented in Figure 1 . The ESI-cleavable TNT-PNA conjugates consist of a PNA probe portion, which interacts with the immobilized target sequence (DNA or RNA) and a peptide tandem nucleic acid mass tag portion, which is ultimately detected. Note that the TNT shown is merely a representation of the tag and not a real structure ( Figure 2 ). An ESI-cleavable linker connects the PNA probe portion of the conjugate to the tandem nucleic acid mass tag peptide. The complete TNT 'Parent Tag' comprises the red 'Tag Fragment' portion and the blue 'Mass Normalizer' portion shown in Figure 1 . The TNT marker is designed to have a unique combination of parent tag mass and tag fragment mass, released during collision induced dissociation (CID), and it is this pair of masses that serves as the sequence identifier. In a typical scenario (Figure 1 ), a set of PNA hybridization probes labelled with different TNTs is first hybridized to the captured target nucleic acids of interest (step (1)). After stringent washes to remove the non-hybridized probes (step (2)), the probes are denatured from the target (step (3)) and injected into an ESI-MS/MS instrument for detection (step (4)). In the mass spectrometer, the TNTs cleave from the PNAs during electrospray ionization (step (5)). The TNT parent tag ions are then selected from the background, fragmented by CID and finally, daughter, tag fragment ions from the fragmentation are detected (step (6)), confirming that the signals are indeed due to the presence of tagged probes, thereby detecting the presence of the target sequences. The use of this MS/MS-based approach offers high specificity allowing TNT labels to be detected in a background of fragmentation noise. In addition, the MS/ MS detection means that tags can share the same mass as long their tag fragment ions are distinguishable from each other. This means that many TNTs can be detected in a compressed mass range. This feature combined with the relatively low overall mass required for TNTs, means that TNT technology will be able to exploit lower cost compact and portable ESI-MS/MS instrumentation that is currently under development (31) (32) (33) (34) (35) (36) . FMOC-protected peptides were custom synthesized by PepSyn Ltd (Liverpool, UK) using commercially available FMOC-protected amino acids on a Beckman synthesizer. 4-FMOC-piperazin-1-ylacetic acid was obtained from Fluka (Sigma Aldrich, Dorset, UK). The amino acid isotope 13 C 3 , 15 N-FMOC-L-alanine was obtained from Cambridge Isotope Laboratories, Inc (Andover, MA). The sequences of the TNT peptides used for the preparation of ESI-cleavable TNT-PNA probes are shown in Table 1 . Peptide nucleic acid oligonucleotide syntheses were carried out using a 2-mmol cycle on an Expedite 8900 synthesizer (Applied Biosystems, Foster City, CA). For the preparation of peptide-PNA conjugates, the FMOCprotected peptide TNT sequence was synthesized and left on the resin by PepSyn Ltd with the N-terminal FMOC left intact. The resin was extracted from the peptide synthesizer column and was then loaded into the Expedite synthesizer column (2 mmol per column). PNA synthesis was carried out as normal on the preloaded resin. The yield of each purified conjugate was in the range 4.3-32 OD 260 , which corresponds to a 1.5-21% yield based on the 2-mmol synthesis scale. The sequences and yields of the TNT-PNA probes are shown in Table 2 . Biotinylated, fully complementary target 50-mer oligodeoxyribonucleotides for the hybridization experiments were synthesized by Yorkshire Bioscience (York, UK) on a 1-mmol scale. Target sequences are shown in Table 3 . TNT-PNA stock solutions were made up at a concentration of 20 pmol/ml in water. Biotinylated target sequences were made up in stock solutions of 50 pmol/ml in water. Six aliquots of 20 ml of stock solution of the first TNT-PNA probe was mixed with 20, 10, 4, 2, 1 and 0.5 ml of the second TNT-PNA, respectively. To these, water was added to make up the solution to a total of 40 ml. The aliquots were then made up to 80 ml with methanol and formic acid to give a final solution of the TNT-PNA probes in 50:50 water:methanol with 1% formic acid. These samples were then analysed by direct injection ESI-MS/MS. Five aliquots of 50 ml of MyOne Streptavidin C1 Dynabeads (10 mg/ml suspension) were separated from their storage buffer and washed with twice with 50:50 methanol:water to remove potential mass spec contaminants. The beads were then washed with 1 Â Bind & Wash (B&W) buffer. B&W buffer for Dynabead incubation was made up according to the manufacturer's instructions: 20 mM Tris, pH 8.0, 2 mM EDTA, 2 M NaCl. Six aliquots of 20 ml (1 nmol) of stock solution of one biotinylated target was incubated with the Dynabeads. These aliquots were all made up to 40 ml with the addition of 20 ml of 2 Â B&W buffer. The biotinylated targets were then incubated at room temperature with the streptavidin beads for 1 h according to the manufacturer's instructions to immobilize the targets on the beads. The target solution was then removed from the beads and the beads were washed twice with hybridization buffer (20 mM Tris, pH 7.5, 10 mM MgCl 2 , 25 mM NaCl). A sixth aliquot of 100 ml of MyOne Streptavidin C1 Dynabeads was made up in the same way but this aliquot was incubated with 40 ml (2 nmol) of stock solution of the same biotinylated target and 40 ml of 2 Â B&W buffer. Each aliquot of the first five aliquots of bead-captured target was then incubated with 50 ml of a TNT-PNA probe solution comprising 1 nmol of the TNT-PNA probe complementary to the target on the beads. Similarly, the sixth aliquot was then incubated with 100 ml of a second TNT-PNA probe solution comprising 2 nmol of the TNT-PNA probe complementary to the target but with the alternative TNT tag on the probe. Hybridization of all the aliquots was carried out at room temperature for two hours. After hybridization, the probe solution was removed from the beads and the beads were washed three times with ice-cold 70 mM aqueous ammonium citrate solution. After the wash step, the sixth aliquot was resuspended in 100 ml of ammonium citrate. Aliquots of the sixth aliquot were then pipetted into the first six aliquots: 20, 10, 4, 2, and 1 ml, respectively. The mixtures The TNT tags, outlined by the red broken line, are modular comprising different functional components that correspond to individual synthetic components in the automated synthesis of these reagents. Each TNT is composed of two parts. The first part is a tag fragment, drawn in red, that comprises a charge-carrying group, the piperazine moiety outlined by the black broken box, and an alanine mass modifier group, outlined by the green broken box. The tag fragment is linked to the second part of the tag, which is a mass normalization group, drawn in blue, that ensures that each tag in a pair of tags shares the same overall mass and atomic composition. Note that the tags shown above are not strictly isobaric due to an issue with the synthesis discussed in the text. The mass normalization group is essentially uncharged and comprises a second alanine mass modifier group to adjust the overall mass of the tag and a proline group that is part of the electrospray cleavable linker. The ESI-cleavable linker is outlined by another black broken box and comprises an aspartic acid proline linkage. Part (ii) of Figure 2a Tables 1 and 2. were then mixed thoroughly. The hybridized TNT-PNA probes were then denatured from their captured target by incubating the beads at 858C for 20 min in 50:50 water:methanol with 1% formic acid. The elution solutions were then analysed by direct injection ESI-MS/MS. Six aliquots of 125 ml of MyOne Streptavidin C1 Dynabeads (10 mg/ml suspension) were separated from their storage buffer and washed with twice with 50:50 methanol:water to remove potential MS contaminants. The beads were then washed with 1 Â B&W buffer. Six aliquots of 20 ml (1 nmol) of stock solution of one biotinylated target was mixed with 20, 10, 4, 2, 1 and 0.5 ml, respectively of the other biotinylated target. To these, water was added to make up the solution to a total of 40 ml. The aliquots were then made up to a volume of 80 ml by addition of 40 ml of 2 Â B&W buffer. These aliquots were then added to the aliquots of washed beads. The biotinylated targets were then incubated at room temperature with the streptavidin beads for 1 h according to the manufacturer's instructions to immobilize the targets on the beads. The target solution was then removed from the beads and the beads were washed twice with hybridization buffer. Each aliquot of captured target was then incubated with 100 ml of TNT-PNA probe solution comprising 1 nmol of each of two TNT-PNA probes of the same length in hybridization buffer, i.e. 1 nmol of ANTHRAX-12 would be mixed with 1 nmol MOMP-12. Hybridization was carried out at room temperature for two hours. After hybridization, the probe solution was removed from the beads and the beads were washed three times with ice-cold 70 mM aqueous ammonium citrate solution. The hybridized TNT-PNA probes were then denatured from their captured target by incubating the beads at 858C for 20 min in 50:50 water:methanol with 1% formic acid. The elution solutions were then analysed by direct injection ESI-MS/MS. ESI-MS/MS spectra were obtained on a Micromass Q-TOF Micro mass spectrometer (Micromass (Waters), Wythenshaw, UK). The TNT-PNA oligonucleotides were denatured from the Dynabeads into 50:50 Methanol:Water with 1% formic acid. Mass spectra were externally calibrated using the manufacturer's standards and calibration protocols. Tandem nucleic acid mass tag-PNA oligonucleotide probe design and synthesis Two pairs of example TNT-PNA oligonucleotide probes are shown in part (i) of Figure 2a and b. The TNTs are peptides comprising two parts, the tag fragment portion, which carries a charge due to the presence of a tertiary amino-functionality and the mass normalization portion, which remains essentially uncharged. These two portions of the tag are linked by a cleavage enhancement group, a piperazine ring, which also carries the charge of the tag fragment on its tertiary amino-group. The two tag portions both comprise a mass modifier component, which are isotopes of alanine in this tag although a large number of different mass modifiers could be used. It can be seen from Figure 2 that the two tags shown employ the same mass modifier components but the order differs Beta Alanine -Aspartic Acid -Proline -13 C 3 , 15 N-Alanine -Piperazin-1-ylacetic acid -Alanine TNT2 Beta Alanine -Aspartic Acid -Proline -Alanine -Piperazin-1-ylacetic acid -13 C 3 , 15 N-Alanine between the tags. Thus, the overall masses of the tags are the same but the tag fragments have different masses, which are equalized by the mass of the mass normalization portion. These tags are designed so that on analysis by collision-induced dissociation (CID), the tag fragment is released to give rise to a uniquely resolvable ion. Thus, this pair of tags allows a pair of PNA probes to be distinguished by MS/MS analysis. Each tag is linked to a PNA oligonucleotide probe by a second linker, comprising aspartic acid and proline that is easily cleaved by CID (37) . As shown in Figure 2 , the aspartic acid/proline linker is used to cleave the tags from their oligonucleotides during electrospray ionization of the tagged oligonucleotides. The expected structures and mass-to-charge ratios of the cleaved parent tag ions generated by dissociation of the aspartic acid/proline linkage are shown in part (ii) of figure 2a and b (37) . Similarly, the expected structures and mass-to-charge ratios of the tag fragment ions and the structures of the neutral mass normalizer fragments, based on the predictions of the 'mobile proton' model of peptide fragmentation (38, 39) , are shown in part (iii) of Figure 2a and b. The use of FMOC peptide synthesis chemistry to synthesize the TNTs combined with FMOC PNA chemistry to synthesize the oligonucleotide probe sequence means the same resin can be used to synthesize both portions of the tagged probe in a single total synthesis approach. For the probes discussed here, the TNT portion of the probe was synthesized in a commercial peptide synthesizer using standard peptide synthesis resin cartridges. The cartridge was then opened and the resin within was loaded into a cartridge suitable for a commercial oligonucleotide synthesizer that supports PNA synthesis (Expedite DNA/PNA synthesizer), allowing the completion of the synthesis of the probe. The completed probe was then cleaved from the resin and deprotected in one step (TFA/cresol) and then purified. With this approach, only one purification step is required resulting in better yields of finished probe than multi-stage processes. Figure 3 illustrates the ease with which a substantial number of tags can be synthesized from a small set of mass modifier components: nine tags, shown in Figure 3b can be made from three mass modifiers, which are the three commercially available isotopes of alanine shown in Figure 3a . In fact, the number of tags that can be synthesized increases as the square of the number of mass modifier components, e.g. there are at least five isotopes of alanine with different masses that are commercially available which would actually allow the synthesis of 25 tags using the design presented here. One issue that emerged from the experiments presented here was a loss of 13 C isotope from the carboxylic acid of alanine when the alanine isotope was present at the C-terminus of the peptide TNT, i.e. in TNT2 shown in Table 1 . This loss is consistent across every TNT-PNA synthesized with TNT2 and may be an effect of the resin used, which relied on a 4-HydroxyMethyl-Phenoxy Acetic acid (HMPA) linker to the carboxylic acid group of the first amino acid to allow the peptide to be cleaved from the resin at the end of the synthesis. Other resins will be tested in future to avoid the loss of isotope. This has meant that the TNTs synthesized were not completely isobaric as shown in Figure 2 . The mass of the singly charged parent tag ion of TNT1 is 388.2 while that of TNT2 is 387.2. For the MS/MS analysis of these tags, both tags could still be selected simultaneously by the first quadrupole of the Q-TOF instrument, as the mass range that is gated is actually about 3 daltons for the default setting of The TNT approach is similar in principle to other mass tagging techniques and enjoys the same features as other approaches, such as ease of multiplexing and the ability to design tag masses to suit applications, with some additional advantages. TNT tags can be made chemically identical, even sharing the same mass as long as the tag fragments are different, so they can act as more precise reciprocal internal standards, which leads to more accurate quantification and the same behaviour in analytical separations, hybridizations and labelling reactions thus avoiding 'dye effects' that plague fluorescent methods (40, 41) . The use of an MS/MS-based detection method allows TNTs to be selected from background noise thus improving signal to noise ratios. This allows untagged material to be ignored, greatly improving data quality. To confirm that the TNT-PNA oligonucleotide conjugates cleave as they are expected to (see expected fragment structures and mass-to-charge ratios in Figure 2 ), the TNT-PNA oligonucleotides were analysed by ESI-MS/MS on a micromass quadrupole-time-of-flight (Q-TOF) mass spectrometer. The complete TNT-PNA probe molecules were initially ionized with the cone voltage (an accelerating voltage in the ESI source that can be varied on the micromass instrument to control the levels of collision induced dissociation) set to minimize fragmentation of the whole TNT-PNA probe conjugate ions. The whole molecular ions were selected using the first quadrupole of the instrument and were then subjected to collision induced dissociation at different collision energies. Fragmentation of the complete TNT-PNA conjugate ions was carried out as it allows both the cleavage of the TNT parent tag from the PNA and the cleavage of the tag fragment from the parent tag to be seen simultaneously in the TOF analyser of the Q-TOF instrument thus demonstrating both cleavage processes and their relative efficiencies. Typical results are shown in Figure 4 , in which it can be seen that, as the collision energy is increased, the whole TNT-PNA probe ion fragments to release the parent tag ion as the predominant fragmentation product. As the fragmentation energy is increased further, the parent tag ion undergoes subsequent consecutive fragmentation to give the desired daughter fragment ion. As the collision energy increases further, the intensity of the parent tag ion increases. Similarly, the ratio of the daughter ion to parent ion increases more. These results show that the TNT-PNA probe molecules fragment as anticipated with the aspartic acid/proline linkage cleaving more easily than the piperazine linkage. The higher collision energy spectra are quite noisy as these also contain other products from the fragmentation of the TNT-PNA probe molecule, which would not normally be present when the TNT parent tag ions are analysed by themselves. The normal mode of analysis is shown in Figure 5 . Here the cone voltage in the electrospray ion source is increased to 25 V increasing the level of fragmentation during ionization thus releasing the parent tag ion from the TNT-PNA conjugate. The parent tag ion is then selected from background by the first quadrupole in the Q-TOF instrument. The parent tag ion is subsequently subjected to CID in the second quadrupole of the Q-TOF instrument. A collision energy of 25 V was used for CID. MS/MS spectra showing the detected tag fragment ions are shown in Figure 5 . These spectra show ratios of two tags and demonstrate the accuracy of quantification of the TNT technology, which is discussed in the next section. A brief experiment to determine whether there were any obvious size dependent effects on the efficiency of the cleavage of the TNT from the TNT-PNA conjugate during electrospray ionisation was carried out. In this experiment, pairs of TNT-PNA oligonucleotides of different lengths were mixed together in 1:1 concentrations. It might be expected that the larger TNT-PNA probes would fragment somewhat less easily due to their greater size and the consequent ability to dissipate kinetic energy from collisions over many more different modes of vibration. This would mean that the larger probes should be detected with less sensitivity. It turned out that the larger sequences gave slightly more sensitivity with the 16-mer being almost twice as sensitive as the 8-mer in this experiment. This result may reflect the mechanisms of cleavage and detection, which are dependent on protonation of the aspartic acid and the piperazine groups in the tags. The larger probes tend to adopt higher charge states (not shown), i.e. they are more heavily protonated and the availability of more free protons on the larger probe ions may facilitate the cleavage of the tags, masking steric effects. However, only three different size molecules were evaluated, and this will be explored more fully in future. A key feature of the Tandem Nucleic acid mass Tag design is the ease with which large numbers of chemically identical, isotopomeric tags can be made (Figure 3) . Sets of TNT isotopes should have almost identical behaviour in analytical separations (17) and during the ionization process. This means that it should be possible to use these tags to accurately quantify their associated oligonucleotide sequences as the ratios of the intensities of the TNT isotopomer fragments should reflect the ratios of the concentrations of the probes in solution or hybridized on their targets ( Figure 5 ). To demonstrate this feature of the TNT design, various experiments were conducted. In these experiments, pairs of TNT-PNA conjugates are analysed by ESI-MS/MS, where the parent tag ions are cleaved from their probes in the ESI ion source using a cone voltage of 25 V. The cleaved parent tag ions are subsequently selected from background by gating both 387.3 and 388.3 ions with the first quadrupole of the Q-TOF instrument. The gated parent tag ions are then subjected to CID in the second quadrupole of the Q-TOF instrument using a collision energy of 25 V followed by mass separation and detection in the TOF analyser to determine the ratios of the tag pairs. In the first set of experiments, pairs of TNT-PNA oligonucleotide probes of the same length and sequence, but with different tags were mixed in a predefined ratio and diluted to determine how well the ratios are conserved as the concentration of probes is decreased. Results are shown in Figure 6 . It can be seen that the ratios of the isotopic TNT-PNA probes are conserved over the range of concentrations investigated. In a second experiment, pairs of TNT-PNA probes of the same length were mixed in various different ratios. The correlation between the expected and measured quantities of these different TNT-PNA ratios is shown in Figure 7 . It can be seen that there are simple linear correlations between the expected and measured ratios. The blue crosses in Figure 7 indicate the results of experiments where the two TNT-PNA probes with the same sequences but with different isotopomeric tags were mixed, i.e. the whole TNT-PNA probes were isotopes of each other. The measured ratios for these probes closely match the expected ratios. The red squares in Figure 7 indicate the results of experiments where the two different probe sequences of the same length were mixed, i.e. although their TNT labels were different isotopes of each other, the complete TNT-PNA probes were not isotopes of each other. The red line represents a linear regression through these data points. Although the predicted and expected ratios do not match exactly, there is a good correlation between the results indicating that the measurements are quantitative. Since the TNT-PNA probes were actually different in these experiments, it is pleasing to see that there is correlation between the measured and expected quantities and the result suggests that, in future use, the measurements of the quantity of different targets in a sample could be calibrated against an internal control such as a housekeeping gene or, preferably, a known quantity of a spiked target sequence. Quantification of hybridized probes was also evaluated. In the first experiment, external calibration of the quantities of hybridized probes was assessed, i.e. the amount of target in one sample was probed with a TNT-PNA whose abundance was then determined by comparison with a second reference sample comprising a predefined quantity of the duplex of the same target sequence and PNA probe sequence, but with a different TNT isotope conjugated to the probe, after the hybridization. In these studies, aliquots of a synthetic biotinylated 50-mer DNA oligonucleotide target was captured onto avidinated magnetic beads and hybridized with TNT-PNA probes with identical probe sequences but different tags. The hybridized beads were then mixed in different ratios. The target, arbitrarily selected, was a fragment of a sequence from the MOMP gene from Chlamydia pneumoniae. A fixed quantity of one TNT-PNA probe complementary to one of the targets was hybridized to the captured target sequences. The aliquots of beads were then washed extensively to remove probe that had not hybridized. The captured TNT-PNA probe mixture was then eluted into 50:50 water:methanol with 1% formic acid (a solvent suitable for ESI-MS/MS analysis) by thermal denaturation. The eluted TNT-PNA and the spike were then injected directly into the Q-TOF instrument for MS/MS analysis. The ratios of the intensities of the two tags derived from the TNT-PNAs should allow the amount of the target sequences in the pooled samples to be determined. Figure 8 shows the actual correlation between the expected and measured quantities of the target sequences. The results are very similar to the experiments where TNT-PNA probes with the same sequence but different tags are simply mixed together: the measured ratio matches very closely the expected ratio. Negative controls in which the target was absent do not show significant binding of TNT-PNA conjugates to the beads so the probe binding is sequence specific. This gives a clear indication that the probes behave quantitatively in hybridization assays and that TNT-PNA probes can be used for accurate quantification. A further evaluation of the quantification of the TNT-PNA conjugates was carried out to determine whether accurate relative quantification can be derived from TNT-PNA pairs with different PNA sequences, i.e. can a reference sequence in a sample probed with one TNT-PNA be used to quantify a second sequence with a different PNA probe as long as the TNTs used in the probe pair are isotopes of each other. This would enable quantification by internal calibration using spiked sequences, housekeeping genes or similar controls in quantitative expression profiling or diagnostic assays. In these experiments, a pair of biotinylated 50-mer target oligonucleotides was used (MOMP-50 again and a sequence from B. anthracis, ANTHRAX-50; see Table 3 ). These were captured onto streptavidin-coated magnetic beads. The quantity of one target was fixed while the relative quantity of the second was varied. The captured targets were then hybridized at room temperature with a probe solution comprising equal quantities of MOMP-12-TNT1 and Anthrax-12-TNT2 probes (see Table 2 ). Probes of the same length were used together. After the hybridization, the magnetic beads were washed as before and the hybridized TNT-PNAs were eluted from their targets on the beads into 50:50 methanol:water with 1% formic acid and analysed as described earlier. Typical results are shown in Figure 8 . As observed in the simple mixture experiments, the measured TNT ratios show a linear relationship with the expected ratios but the measured quantities do not exactly match the expected ratios when the TNT-PNA probes being compared are not true isotopes but the linear relationship does mean that the measurements are quantitative. These data suggest that with appropriate choice of reference sequences, quantitative internal calibration should be achievable, which would be very useful in situations where suitable reference samples are not available for external calibration. In this article the synthesis, characterization and application of ESI-cleavable TNT peptide-PNA conjugates to the quantitative detection of target DNA sequences by ESI-MS/MS has been described. The conjugates were prepared by first synthesizing the TNT tag peptide sequence in a peptide synthesizer, after which the peptide synthesis resin was transferred to a column compatible with a DNA synthesizer in which PNA can be prepared. The PNA sequence was extended directly from the peptide TNT. The use of PNA has several advantages for this application: (i) the oligonucleotide and the peptide TNT are generated in a single synthesis on the same resin, which means only a single purification step is required after the synthesis is completed; (ii) PNA is approximately 15% lower in mass than a corresponding DNA sequence, which Figure 8 . The relationship between the expected and measured ratio of mixtures of captured target sequences probed with 12-mer TNT-PNA probes. Experiments to quantify targets by external calibration where the PNA probes had identical sequences but different TNTs are shown with blue crosses and the blue regression line while experiments to quantify targets by internal calibration, i.e. using one sequence as a reference to quantify a different sequence are shown as red squares and the red regression line. enhances mass spectrometric sensitivity; (iii) PNA has enhanced binding affinity for its target compared with a corresponding DNA sequence, so a shorter probe can be used to achieve a corresponding level of specificity; (iv) PNA can hybridize under low salt conditions, which is more favourable for ESI analysis of samples as ESI-MS is susceptible to signal suppression by high salt concentrations. We have shown that these TNT-PNA probes hybridize quantitatively and specifically to their targets and that the probes perform reliably over a wide dynamic range providing a new platform for multiplexed and quantitative genomic analysis. The ESI-cleavable TNT mass markers described in this article have several properties that make them very useful for quantitative, multiplex assays, including the following: (i) The TNT portion can be elaborated into very large arrays of tags using only small numbers of starting components. The 20 standard amino acids as well as the large number of isotopic variants of these amino acids that are available provide the possibility of synthesis of numerous marker molecules that are easily resolved by the unique combination of parent and daughter ion massto-charge ratios; (ii) The ESI-cleavability allows direct analysis of solution phase assays without complex workup of the samples unlike MALDI, which requires that samples are spotted onto targets; (iii) The ESI-cleavable linker connecting the DNA and the TNT components cleaves virtually instantaneously during electrospray ionisation that will allow separations such as capillary electrophoresis to be performed in-line with the MS/MS analysis; (iv) In-line separation allows for a further level of multiplexing over and above the large numbers of available tags since the probes can be identified by their elution time as well as by their tag and in future work, we will explore the use of this feature to enable in-line coupling of analytical separations such as capillary electrophoresis; (v) sets of isotopic TNTs can be synthesized that behave identically during separations, hybridizations and labelling reactions enabling accurate measurements of quantities of target nucleic acids without 'dye effects' widening the range of applications for which mass tags can be employed. The use of TNT-PNA oligonucleotide conjugates offer many of the advantages of fluorescent detection such as high specificity, ease and safety of handling and high sensitivity with the additional unmatched advantages that result from being able to generate large numbers of tags with predefined masses and from being able to construct these sets of tags with stable isotopes generating chemically identical entities that will behave the same in labelling reactions and in separation steps. This means that multiplexed analyses with accurate quantification are now enabled in a user-friendly format. Future experiments will be directed towards evaluation of the TNT-PNA probes for post-PCR amplicon detection. In addition, the development of TNT-DNA oligonucleotide conjugates and evaluation of these probes as primers for multiplexed PCR amplification and subsequent detection of PCR amplicons will also be pursued. The ability to employ in-line capillary electrophoresis with immediate cleavage and detection of tags will be of particular interest as many genomics assays, such as restriction fragment length polymorphisms, satellite marker analysis and multiplexed PCR employ size separations and the ability to perform such analyses with the higher levels of multiplexing enabled by this technology will be of great advantage.

Hairpin structure within the 3′UTR of DNA polymerase β mRNA acts as a post-transcriptional regulatory element and interacts with Hax-1

Aberrant expression of DNA polymerase β, a key enzyme involved in base excision repair, leads to genetic instability and carcinogenesis. Pol β expression has been previously shown to be regulated at the level of transcription, but there is also evidence of post-transcriptional regulation, since rat transcripts undergo alternative polyadenylation, and the resulting 3′UTR contain at least one regulatory element. Data presented here indicate that RNA of the short 3′UTR folds to form a strong secondary structure (hairpin). Its regulatory role was established utilizing a luciferase-based reporter system. Further studies led to the identification of a protein factor, which binds to this element—the anti-apoptotic, cytoskeleton-related protein Hax-1. The results of in vitro binding analysis indicate that the formation of the RNA–protein complex is significantly impaired by disruption of the hairpin motif. We demonstrate that Hax-1 binds to Pol β mRNA exclusively in the form of a dimer. Biochemical analysis revealed the presence of Hax-1 in mitochondria, but also in the nuclear matrix, which, along with its transcript-binding properties, suggests that Hax-1 plays a role in post-transcriptional regulation of expression of Pol β.

In eukaryotic cells, DNA polymerase b (Pol b) is essential for base excision repair (BER) and involved in recombination and drug resistance (1, 2) . Pol b expression levels are important for the maintenance of genome integrity. On account of its low fidelity, the overexpression of Pol b, which has been reported in several cancer types, leads to a mutator phenotype associated with genetic instability and decreased sensitivity to anti-cancer chemotherapeutics (3) (4) (5) (6) . Conversely, cells deficient in DNA polymerase b display hypersensitvity to alkylating agent-induced apoptosis and chromosomal breakage (7) . Sugo et al. (8) have demonstrated that Pol b-deficient mice die immediately after birth, due to extensive apoptosis in the developing central and peripheral nervous systems. These data indicate that DNA polymerase b expression level is extremely important and must be tightly regulated. Transcription of the Pol b gene is upregulated by the phosphorylated transcription factor CREB-1 in response to alkylating agent exposure. This upregulation requires the presence of the specific cAMP response element (CRE) in the promoter of the Pol b gene (9, 10) . It has also been shown that the Pol b promoter contains a binding site for the transcription factor SP1 (11) . Although regulation at the transcription level is apparent, expression of the Pol b transcripts might also be controlled at the posttranscriptional level. In rat cells there are two alternatively polyadenylated Pol b transcripts, with 3 0 UTRs of significantly different lengths (12) . The expression of the two transcripts is tissue specific. The short transcript is present in most tissues, with significantly higher expression in testis, while the long transcript is expressed mostly in the brain and lungs (13) . Motifs present in the 3 0 UTRs of the two transcripts are likely to be responsible for regulation of tissue-specific expression. Sequence of the short 3 0 UTR is highly similar in the rat and human genes (up to 90% of homology in the conserved region), but in the further region of the long 3 0 UTR similarity abruptly decreases, which is consistent with the observation that there is only one (short) Pol b transcript in the human cells. Sequence similarity indicates that regulatory motifs present in the short 3 0 UTR of the rat Pol b transcript may act in the same manner and bind similar factors in human cells. Structure prediction analysis (14) of the short 3 0 UTR revealed the presence of a putative hairpin element, $50 nt upstream of the polyadenylation sequence and 40-50 nt downstream of the termination codon (12) . There are several examples of transcripts with similar structures within the 3 0 UTR, constituting cis-acting regulatory elements, involved in mRNA localization. Structural motifs (hairpins) present within the 3 0 UTR of c-fos (15) , c-myc (16) , MT-1 [metallothionein-1, (17) ], slow troponin C (18) and vimentin (19) mRNAs have been shown to be involved in targeting these mRNAs to the perinuclear cytoplasm and, presumably, anchoring them in this location by binding to cytoskeletal elements (20) . Hairpin elements within the 3 0 UTRs were also reported to stabilize transcripts by binding factors that protect against nuclease cleavage [e.g. binding of IRP protein to the IRE sequence in the 3 0 UTR of transferrin receptor mRNA, (21) ] and to re-program translation as in the case of the SECIS element which directs an insertion of selenocysteine into in-frame UGA codons (22) . In this report, we confirm the existence of the hairpin structure within the 3 0 UTR of the Pol b mRNA and demonstrate that this element influences the expression of a reporter gene. We describe the identification of a protein factor binding to this motif-Hax-1, an anti-apoptotic, cytoskeleton-related protein, which is known to bind a hairpin structure within the 3 0 UTR of vimentin mRNA. We demonstrate that binding occurs only for a Hax-1 dimer, though RNA binding is not a prerequisite for the dimerization itself. We confirm the importance of the hairpin structure for binding of Hax-1 by its mutagenic disruption, which impairs the RNA-protein interaction. We also report strong association of Hax-1 with the nuclear matrix, which is a novel finding, consistent with its transcript-binding properties. Taken together, these data suggest that the hairpin element within the Pol b 3 0 UTR represents a novel motif important for posttranscriptional regulation of expression. The template for in vitro transcription encompassing the whole short 3 0 UTR of rat Polb (208 nt) was synthesized by PCR with a forward primer containing the T7 RNA polymerase promoter sequence (5 0 -TAATACGACTCAC TATAGGGCCTGCCCCACCCAGGCCT) and reverse primer (5 0 -AAACCATGGTACTGCGATC). The PCR was performed with the plasmid bearing the Pol b short 3 0 UTR sequence (pGEM-4Z/H), in the following conditions: 948C for 1 min followed by 35 cycles at 948C for 1 s, 608C for 1 s and 728C for 30 s. The transcription reaction was carried out in 50 ml containing 20 pmol of PCR product, 500 mM rNTPs, 3.3 mM guanosine, 40 U of ribonuclease inhibitor RNase Out (Invitrogen) and 400 U of T7 RNA polymerase (Ambion). The reaction was carried out at 378C for 2 h and the transcript was purified from a denaturing 10% polyacrylamide gel, and 5 0 -end-labeled with T4 polynucleotide kinase and [g 32 P]ATP (4500 Ci/mmol; ICN). The labeled RNA was again purified by electrophoresis on a denaturing 10% polyacrylamide gel. Prior to structure probing reactions, the labeled RNA was subjected to a denaturation and renaturation procedure in a buffer containing 2 mM MgCl 2 , 80 mM NaCl, 20 mM Tris-HCl pH 7.2 by heating the sample at 808C for 1 min. and then slowly cooling to reaction temperature. Limited RNA digestion was initiated by mixing 5 ml of the RNA sample (50 000 c.p.m.) with 5 ml of a probe solution containing lead ions, nuclease S1 or ribonucleases T1, T2 or Cl3. The reactions were performed at 378C for 10 min. and stopped by adding an equal volume of stop solution (7.5 M urea and 20 mM EDTA with dyes) and sample freezing. To determine the cleavage sites, the products of the RNA fragmentation reaction along with the products of alkaline hydrolysis and limited T1 nuclease digestion of the same RNA molecule were separated on 10% polyacrylamide gels containing 7.5 M urea, 90 mM Trisborate buffer and 2 mM EDTA,. The alkaline hydrolysis ladder was generated by the incubation of the labeled RNA in formamide containing 0.5 mM MgCl 2 at 1008C for 10 min. The partial T1 ribonuclease digestion of RNAs was performed under semi-denaturing conditions (10 mM sodium citrate pH 5.0; 3.5 M urea) with 0.2 U/ml of the enzyme and incubation at 558C for 15 min. Electrophoresis was performed at 1500 V and was followed by autoradiography at À808C with an intensifying screen. Site-directed mutagenesis SDM disrupting the hairpin structure was performed using the QuickChange Mutagenesis Kit (Stratagene), according to the manufacturer's instructions. pGEM-3Z/H, containing 208 bp of the Pol b 3 0 UTR in the sense orientation, was used as a template for mutagenesis, resulting in the generation of pGEM-3Z/Hmut. Primers used for mutagenesis (changed nucleotides in bold): Mutagenic primer #1 (forward): Mutagenic primer #2 (reverse): 5 0 -GAAGAGGCAATCACCTAAAACACCCAAGGG TTACATAGCAAAGG-3 0 A fragment of 208 bp, containing the whole short Pol b 3 0 UTR sequence, generated as described in Structural analysis of RNA, was inserted downstream of Firefly luciferase in the pCMLuc vector (pCM2 derivative, a gift from Dr D.Weil, Institute Andre Lwoff, Villejuif, France) into the EcoRI site present in the polylinker, generating pCMLucH. The same fragment containing a mutation of the hairpin-forming region was cloned into the EcoRI site of pCMLuc, generating pCMLucHmut. The rat hepatoma FTO-2B cell line was grown on DMEM (Invitrogen) supplemented with 10% FBS (Invitrogen). Transfection of FTO-2B was performed using Lipofectamine 2000 (Invitrogen), according to the manufacturer's guidelines. Two hundred nanogram of the appropriate plasmid was used for transfection, performed in 50 ml of medium in a 96-well plate. The activity of Firefly luciferase in lysates prepared from cells transfected with pCMLuc, pCMLucH and pCMLucHmut was measured using a luciferase reporter assay system (Dual-Glo, Promega). Assays were performed according to the manufacturer's instructions. Light emission was measured with LumiCount microplate luminometer (Packard). Transfection efficiencies were normalized by co-transfection with phRL-CMV vector (Promega) containing the Renilla luciferase. Total RNA from transfected cells was isolated using the NucleoSpin RNA II kit (Macherey-Nagel). The first strand of cDNA was obtained with SuperScript Reverse Transcriptase (Invitrogen) from 1 mg of RNA, according to the manufacturer's instructions. Primers used in the experiment were designed to amplify 209 bp of the firefly luciferase transcript (forward: 5 0 -TCGTTGACCGCCT GAAGTCT-3 0 , reverse: 5 0 -GGCGACGTAATCCACGA TCT-3 0 ) and, as a reference, 232 bp of the Renilla luciferase transcript (forward: 5 0 -TGGAGCCATTCAA GGAGAAG-3 0 , reverse: 5 0 -TTCACGAACTCGGTGT TAGG-3 0 ). Quantitative PCR was performed using ABI Prism 7000 Sequence Detection System (Applied Biosystems). Power SYBR Green PCR Master Mix (Applied Biosystems) was used for detection. PCR was performed in the following conditions: precycling hold at 958C for 10 min, cycles: 958C, 30 s, 568C, 30 s, 728C, 30 s, up to 40. The ÁÁC T method was used for quantity calculations (23) . The slope of the validation curve was 50.1, which ensures, that target and reference efficiencies are approximately equal. Yeast three-hybrid screen cDNA library construction. Total RNA from rat testis (Lewis) was isolated with Trizol reagent (Sigma) and mRNA was purified on Oligo(dT) cellulose columns (Molecular Research Center). Hundred microgram of mRNA was used for cDNA synthesis using the Gibco BRL cDNA Synthesis System, according to the manufacturer's instructions. EcoRI adapters (Gibco BRL) were ligated to the cDNA, followed by insertion into the EcoRI site of the pYESTrp1 vector (Invitrogen). The generated cDNA library contained roughly 1.5 Â 10 6 clones. Generation of R40C-W. The yeast strain R40C (coat) (24) was disrupted in TRP1 gene in order to obtain tryptophane auxotrophy required for transformation with the pYESTrp1 vector. The disruption cassette was constructed in the YDp-W (25) vector by cloning KanMX resistance marker from pFA6akanMX4 (26) into XbaI-HindIII sites (blunted) within TRP1 gene. R40C was transformed with BamHI-BamHI fragment containing the disruption cassette and selected on geneticin (G418, 200 mg/ml) -containing medium. Three-hybrid screen. A hybrid RNA construct (pRH5 0 H) containing the sequence of Pol b 3 0 UTR (from +3 to À208 downstream of the stop codon) was generated by inserting the Pol b sequence fragment into the unique SmaI site downstream of the phage MS2 sequence in the pRH5 0 vector (Invitrogen). The pRH5 0 H (H for hairpin) and cDNA library plasmids were sequentially transformed using the lithium acetate method (27) into the yeast R40C-W strain, which contains both HIS3 and b-galactosidase (b-gal) promoters integrated into its genome. Transformed yeast were plated on YC medium lacking tryptophan, uracil and histidine (YC-WUH) and containing 5 mM 3-aminotriazole (3-AT). Transformants were replicated on 25 mM 3-AT and after 3-6 days large colonies were picked for analysis. Selected transformants were screened out by the expression of beta-galactosidase using the colony lift assay with 5-bromo-4-chloro-3indolyl-beta-D-galactopyranoside (X-gal) as the substrate (CLONTECH Yeast Protocols Handbook). To eliminate RNA-independent false positives, plasmids were isolated (28) from selected blue colonies and used to transform electrocompetent E. coli KC8 cells. This bacterial strain enables complementation of yeast auxotrophy marker genes. Colonies with plasmids bearing the yeast TRP1 gene (pYESTrp1 cDNA clones) were isolated on M9 minimal medium lacking tryptophan. Plasmids were isolated, verified by restriction analysis and re-transformed into R40C-W, previously transformed with pRH5 0 H. Double transformants were checked again by the b-gal colony lift assay. This method of selection eliminated false-positive clones, leaving RNA-dependent positives. Positive clones were sequenced and analyzed by BLAST search. Positive control vectors for the screen: pRH3 0 /IRE and pYESTrp1/IRP, were taken from RNA-Protein Hybrid Hunter Kit (Invitrogen). Western blotting was performed with ECL Plus (Amersham) according to the manufacturer's instructions. Monoclonal anti-HAX-1 antibody (BD Bioscience) was used in 1:250 dilution. The secondary antibody, goat-antimouse (Pierce) was used in 1:2000 dilution. Anti-BclX L mouse antibody (Clone 2H12, Sigma) was used in dilution 1:80 and anti-proteasome 20S subunit alpha 1,2,3,5,6,7 (PW8195, Affiniti Research) mouse antibody was used in dilution 1:1000, both with secondary goat-anti-mouse antibody (Pierce) used in dilution 1:15 000. Anti-matrin 3 antibody (a kind gift from Ronald Berezney, State University of New York at Buffalo, Buffalo) was used in 1:1000 dilution with secondary rabbit-anti-chicken antibody (Pierce) used in dilution 1:15 000. Polyclonal rabbit anti-H2B antibody (UPSTATE) was used in dilution 1:5000 with secondary goat-anti-rabbit antibody (BioRad) used in dilution 1:5000. Hax-1 cDNA was amplified with specific primers (forward: 5 0 -CCAGGATCCGAGCGTCTTTGATCTTTTC CGAGGCT-3 0 , reverse: 5 0 -GCTTGTCGACTCGGGAC CGAAACCAACGTCCTA-3 0 ) and the PCR product was cloned into pET201 (a gift from Csaba Koncz, Max Planck, Institute for Plant Breeding Research, Cologne). pET201 is a non-commercial vector representing a derivative of pET vector series used for expression of recombinant proteins fused to N-terminal bacterial thioredoxin and C-terminal 6xHis tags. pET201 vector encoding for thioredoxin was used concurrently to produce thioredoxin as a control protein. Bacteria were grown at 378C to an optical density of 0.5 (OD 600 ), followed by induction with 1 mM IPTG for 4 h. Recombinant proteins were purified under native conditions according to the Qiagen protocol for His-tag protein purification. Eluted proteins were analyzed by 12% SDS-PAGE and stained by Coomassie. Western blot analysis was carried out using monoclonal anti-Hax-1 antibody (BD Bioscience). Purified proteins were stored in 50% glycerol at À208C. The 208 bp of Pol b 3 0 UTR, encompassing the whole short 3 0 UTR sequence with the hairpin encoding region, was cloned into pGEM-3Z and pGEM-4Z vectors (Promega) in both orientations, generating a set of templates for transcription of sense and antisense mRNAs. In vitro transcription was carried out with polymerases SP6 and T7 (Promega). [a-32 P]UTP (3000 Ci/mmol) was used for labeling. RNA integrity was controlled by polyacrylamide gel electrophoresis. Labeled transcripts were gel-purified (5% polyacrylamide, 8 M urea). In vitro transcribed RNA with heparin (5000 U/ml, Sigma) and 160 U of RNAse inhibitor (RNase OUT, Invitrogen) was incubated on ice for 15 min in cross-link buffer (20 mM Tris, pH 7.9, 2.5 mM MgCl 2 ) with 0.5 mg of recombinant Hax-1 protein, BSA (Sigma) or thioredoxin (recombinant thioredoxin purified from bacteria using pET201 vector in the same conditions as Hax-1 protein). Cross-linking was performed on ice for 30 min in the UV Stratalinker 1800 (Stratagene). After cross-linking, the reaction was incubated in the presence of RNase A (final concentration 0.6 mg/ml) for 20 min. in room temperature. Laemmli loading buffer was added and the samples were heated to 958C for 5 min. Samples were analyzed on SDS-PAGE, with Prestained SDS-PAGE Standards, Kaleidoscope (Bio-Rad). Purified recombinant Hax-1 protein (1 mg for the western blot, indicated amount for silver staining) was suspended in 100 ml of sample buffer (SB: 10 mM HEPES, pH 7.4, 120 mM potassium acetate, 2.5 mM MgCl 2 , 1.6 mM dithiotreitol), incubated for 15 min in room temperature to establish monomer-dimer equilibrium and then incubated for 20 min in 228C with 0.00125% glutaraldehyde (Sigma). The reaction was stopped by the addition of 50 ml of 3-fold concentrated loading buffer (150 mM Tris pH 6.8, 6 mM EDTA, 6% SDS, 30% glycerol, 1 M urea, 0.003% bromophenol blue). Samples were heated (908C) and separated on the SDS-PAGE. Proteins were detected by silver staining (29) and western blot. The lanes in the silver-stained gel were scanned and the areas under the peaks were quantified using Multi-analyst (Bio-Rad). The results were analyzed by nonlinear regression analysis to determine an approximate dimer dissociation constant, as described in (30) . The filter-binding assay was performed by a modified method described by Wong et al. (31) . Briefly, radiolabeled, in vitro transcribed RNA at a concentration of $10 nM was heat denatured, allowed to refold and incubated for 15 min in room temperature in binding buffer (20 mM Tris, pH 7.9, 100 mM KCl, 2.5 mM MgCl 2 ) containing heparin (5000 U/ml, Sigma) and 160 U of RNAse inhibitor (RNase OUT, Invitrogen) with purified Hax-1 protein. The reaction mix was loaded onto a presoaked nitrocellulose membrane (0.45 mm, Osmonic Inc.) on top of a nylon membrane (Hybond-N, Amersham) and filtered under pressure in a slot-blot apparatus. Following filtration, each filter was dried and quantitated on a PhosphorImager (BioRad) using QuantityOne software. Dissociation constants (K d ) for the RNA-protein complexes were obtained by fitting the empirical data to a sigmoidal curve by nonlinear regression analysis, using Maxima 5.4. Fitting was performed in respect to dimer concentration calculated as a function of the total protein concentration, as in (30) . Four hundred milligram of pulverized rat testes were homogenized in 1 ml of 2Â MEB buffer (100 mM Tris-HCl pH 7.5, 20% glycerol, 4 mM DTT, 10 ml/ml protease inhibitor cocktail [Roche]) in a Potter homogenizer, filtered on 50 mm Nitex membrane (Tetko) and centrifuged at 2000g, 48C. The pellet (nuclear fraction) was washed two times with 0.5 ml of 2Â MEB buffer. Nuclei were gently resuspended in 0.5 ml of 2Â MEB (small aliquot was analyzed by DAPI staining), passed through a needle and incubated 30 min on ice with 10 ml of DNase I (Warthington). The supernatant was centrifuged at 10 000g, 48C (mitochondrial fraction). After separation of the mitochondrial fraction, the supernatant was centrifuged at 100 000g in an ultracentrifuge and the supernatant (cytoplasmic fraction) was collected. Procedures were adapted from Reyes et al. (32) . Highsalt method: isolated nuclei were resuspended in 1 ml of CSK buffer (10 mM PIPES pH 6.8, 100 mM NaCl, 300 mM surcose, 3 mM MgCl 2 1 mM EGTA, 1 mM DTT, 1 mM PMSF, 0.5% TritonX-100 with protease inhibitor cocktail [Roche]-one tablet per 50 ml), incubated for 3 min at 48C and spun down at 5000g for 3 min to separate nuclei from soluble proteins. Next the chromatin was solubilized by RNase-free DNase I digestion (0.1 mg, Warthington) in 1 ml of CSK buffer for 15 min at 378C. Next, ammonium sulfate was added to a final concentration of 0.25 M and after a short incubation (5 min at 48C) the samples were centrifugated again. The resulting pellet was extracted three times with 2 M NaCl in CSK buffer in the following steps: resuspension, incubation at 48C for 5 min and centrifugation. In the first step of our study, we undertook the determination of the structural features of the entire short 3 0 UTR of the rat DNA polymerase b transcript (208 nt). For this purpose, we used limited fragmentation of the 5 0 -end-labeled RNA with six well-characterized biochemical structural probes: lead ions and five enzymes ( Figure 1A ). Nuclease S1 and lead ions have no documented nucleotide specificity, while ribonuclease T2 and A recognize all nucleotides but have a higher activity for adenosines and pyrimidines, respectively. RNase T1 exhibits specificity exclusively for guanosines and RNase Cl3 digests C-residues only. All the above structural probes recognize flexible and single-stranded regions in the RNA structure. Prior to the probing experiments, the 5 0 -end labeled transcript of the Pol b short 3 0 UTR was analyzed by non-denaturing gel electrophoresis. The result of this test suggests that only a single conformer is formed by the analyzed RNA molecule (migrated as a single band on a gel, not shown). Probing data demonstrated that the short 3 0 UTR of Pol b transcript forms three separate structural modules, designated M1-M3 ( Figure 1B) . Module M1 represents a small hairpin structure with a C-rich terminal loop. The 3 0 -part of this structure is efficiently hydrolyzed with lead ions, even at a low concentration of the probe, which suggests that M1 hairpin has binding capacity for this metal ion. Module M3 contains the polyadenylation signal, followed by the cleavage and polyadenylation site, so only a part of it is present in the mature short Pol b transcript, which implies minor importance of the whole M3 module as a post-transcriptional regulatory element. Module M2 is formed by nucleotides located between bases G29 and C96 of the analyzed transcript and is composed of three helical regions, two 6 bp and one 14 bp in length. Each of these helical regions is resistant to enzymatic digestion and lead ion hydrolysis. The three helixes are separated by two asymmetric, internal loops (b and c), which are mapped very well by all probes used. The longest helical region includes as many as six non-WC, U-G and G-U base pairs. Four of these exist as two tandems: U-G, G-U and U-G, U-G, which are known to be potential metal-or protein-binding sites. The hairpin structure contains a small 3-nt terminal loop (5 0 -UAU), which is also well recognized by both lead ions and enzymes ( Figure 1A and B) . The M2 hairpin structure is conserved among species (Figure 2) , which suggests its significant role in regulation of transcript fate. In order to demonstrate the essential role of the evolutionarily conserved hairpin structure in the regulation of Pol b expression, we carried out in vitro mutagenesis to disrupt the M2 hairpin element. Mutagenic primers were designed based on an MFOLD (14) prediction of the potential effect of nucleotide change on hairpin structure ( Figure 3 ). Substitution of three C-residues for three G-residues in the stem-forming region (positions 67-69) changed the predicted free energy increment (dG) from À14.1 kcal/mol for the intact structure to À3.6 kcal/mol for the mutant. The mutated sequence (Hmut) was used in luciferase reporter assays and served as a template for in vitro transcription to generate mRNA for subsequent crosslink and filter-binding analysis. The influence of the M2 hairpin structure within the short 3 0 UTR of Pol b mRNA on expression was analyzed utilizing a luciferase reporter system. The rat hepatoma cell line FTO-2B was transfected with reporter constructs bearing the Firefly luciferase gene appended by the Pol b 3 0 UTR sequence containing the M2 hairpin element, (pCMLucH) and by the same sequence containing a structure-disrupting mutation in the hairpin-forming region (pCMLucHmut). A vector bearing the unmodified luciferase gene served as a control (pCMLuc). To asses mRNA levels of the reporter, quantitative PCR was performed for cDNA preparations obtained from cells subjected to the same transfections as for luciferase assays. The results (Figure 4) show a significant decrease of mRNA levels for the construct with the hairpin structure (H), compared with almost unchanged mRNA levels for the mutated hairpin (Hmut). These differences, however, are not reflected at the protein level: both constructs (H and Hmut) caused an increase in luciferase expression, although only for the Hmut is this increase significant. The relative increase of protein levels in respect to mRNA levels is therefore more than 3-fold for the construct with the intact hairpin, while only a slight relative increase ($1.7-fold) was observed for the mutated construct. These changes in expression indicate a complex posttranscriptional regulation in which the hairpin element has a key function. In order to identify proteins interacting with the M2 hairpin element present within the 3 0 UTR of the Pol b transcript, we performed a yeast three-hybrid screen of a rat cDNA library. The IRE-IRP interaction served as a positive control, while negative controls consisted of empty pRH5 0 and pYESTrp1 vectors and single transformants of pRH5 0 H 'bait' plasmid or of the protein-hybrid clone isolated from the library. Out of 63 positive clones obtained after the first selection, only 11 were RNA dependent, and these were sequenced and analyzed. Of these, only one was in the proper reading frame, had the correct in-frame orientation and represented a coding sequence-the terminal 622 bp of the rat Hax-1 mRNA(corresponding to the last 150 aa of the protein), ( Figure 5 ). This clone exhibited strong growth on 25 mM 3-AT medium without histidine and tested positive in the b-gal plate test (Table 1) . UV-cross-linking confirms specific binding of Hax-1 to the mRNA region containing the M2 hairpin structure In order to confirm the results of the 3-hybrid screen, analysis of in vitro binding was carried out using purified (14) . The free energy increment for the two structures varies considerably, and amounts to À14.1 for the M2 hairpin and À3.6 for the mutant. Presence of the hairpin-containing 3 0 UTR confers mRNA instability but also enhances expression at the protein level. Disruption of the hairpin prevents mRNA degradation, but a mutation-containing construct still slightly enhances expression at the protein level. Constructs used for transfections: control vector (pCMLuc), hairpin-containing construct (pCMLucH) and mutated hairpin (pCMLucHmut). mRNA levels were quantitated for each construct by RQ-PCR in cDNA isolated from transfected cells (8 independent transfections), in triplicate, normalized with respect to Renilla luciferase mRNA levels. Firefly luciferase levels were assessed for the same eight transfections (in six repeats) and normalized with respect to Renilla luciferase levels, which served as a control of transfection efficiency. Statistical significance is reported as P-value: Ã P50.01, ÃÃ non-significant. UV-cross-linking was performed using the purified recombinant Hax-1 and in vitro transcripts of the hairpin-containing region (H), a control antisense transcript of the same region (A), and the mutated transcript (Hmut). Bovine serum albumin and purified thioredoxin served as controls for interaction specificity. Hax-1 demonstrated a specific interaction only with the hairpin-containing RNA, and this transcript did not interact with either BSA or thioredoxin ( Figure 6) . The cross-linked band migrated at a molecular weight of $100 kDa, which corresponds to a recombinant Hax-1 dimer (the recombinant protein has a higher molecular mass than endogenous Hax-1 [35 kDa], due to its fusion with the thioredoxin sequence). The protein dimer is likely formed during the UV-cross-linking procedure, as has been previously documented (34) . No band corresponding in size to monomeric recombinant Hax-1 (ca. 50 kDa) was observed, which indicates that the protein interacts with RNA exclusively in the form of a dimer. To confirm that Hax-1 forms a dimer, purified recombinant protein was used for chemical cross-linking with glutaraldehyde. As indicated in Figure 7 , monomers with a molecular mass of around 50 kDa were detected, as were dimers with a molecular mass around 100 kDa. Since dimer formation was detected in an in vitro experiment performed with purified protein, one can conclude that the RNA molecule is not necessary for dimerization. As deduced from the UV-cross-linking experiments, RNA is bound exclusively by dimeric Hax-1, hence the dimerization rate may represent an important factor for RNA-protein complex formation. Glutaraldehyde cross-linking with increasing protein amounts ( Figure 7A ) allowed for the estimation of the monomerdimer equilibrium dissociation constant (K ddim ) at 13.5 mM AE 5.5. To corroborate the UV-cross-link results, filter-binding assays were performed with increasing amounts of Hax-1 protein and a constant amount of the same test RNAs (hairpin-containing region-H, antisense-A and mutant-Hmut). Data were analyzed on a Klotz plot ( Figure 8 ) and the apparent K d for each complex was calculated. The fit of a nonlinear binding curve to the experimental data points was best when calculations were performed in respect to dimer concentration, which indicates that the RNA interacts only with a dimeric form of the protein, thus confirming the UV-cross-link results. The approximate K ddim established from the titrated glutaraldehyde cross-linking experiments (13.5 mM AE 5.5) was in the same range as the K ddim predicted by curve-fitting (fitting performed simultaneously for the H and Hmut data point series in respect to variable K d and constant K ddim values yielded a K ddim of 10 mM AE 4). Transcript H showed the greatest affinity for Hax-1, with a K d of 28 nM AE 7, and significantly lower Table 1 . Three-hybrid screen of a rat testis cDNA library proteins interacting with the M2 hairpin-containing region of the Pol b mRNA 3 0 UTR Plasmids used for transformation of R40C-W Growth on 25 mM 3-AT b-gal plate test The interaction for the single positive clone encoding for the terminal 622 bp of the rat Hax-1 gene (Hax-1-C) was tested for growth on 25 mM 3-AT medium without histidine and using the b-gal plate test. The IRE/IRP interaction served as positive control, while transformation with empty vectors as well as single transformants of tested constructs represented negative controls. Hax-1 localizes in mitochondria, but also in the nuclear matrix While Hax-1 mitochondrial localization has been confirmed in many reports (35) (36) (37) , this is not obviously consistent with its transcript-binding properties. Several other reports have shown localization of the protein to the endoplasmic reticulum (35, 36, 38) , apical membrane of hepatocytes (39) and nuclear envelope (35) . In order to establish Hax-1 cellular localization in rat cells, we performed organellar fractionation and subsequent fractionation of nuclei, followed by SDS-PAGE and western blot. These experiments confirmed the presence of Hax-1 in mitochondria, but also indicated its localization in the nucleus ( Figure 9A ) with only traces of Hax-1 detectable in the cytoplasm. Subsequent nuclear fractionation performed with two different methods, revealed that Hax-1 is present in the fraction containing nuclear matrix proteins ( Figure 9B and C), while it was not detected in the other fractions containing soluble and chromatin-associated proteins. To ensure proper quality of nuclear fractions, western blots with appropriate marker proteins were performed. Fraction 1 in both methods contains soluble, chromosomal proteins, released after DNaseI treatment, represented here by histone H2B. These proteins were washed out in subsequent steps of the preparation (fractions 2-3). The nuclear matrix fraction (fraction 4) was probed with a matrix-specific protein matrin 3 antibody (40) . Close association of Hax-1 with the nuclear matrix sheds new light on its mRNA-binding capacity and may indicate its role in regulation of transcript fate. Stable secondary structures in 3 0 UTRs have been shown to play a role in mRNA sorting and localization (20) . Some of them have been also reported to influence mRNA stability (21) . The existence of a stable structural motif in the 3 0 UTR of the rat Pol b mRNA was predicted previously (12) and it was speculated that it may have a regulatory role. In the present work, secondary structure analysis by lead ion hydrolysis and enzymatic digestion revealed the existence of several motifs in the analyzed sequence, namely, a region of strong lead ion binding (M1), a hairpin-forming and highly evolutionarily conserved region (designated as M2) and the region containing polyadenylation sequence (M3). Only M2 is evolutionary conserved, the rest of the untranslated region has high interspecies sequence variation, which suggests lesser functional importance. The sequence in the conserved region exhibits 86.8% identity to the homologous human sequence (100% in the upper stem region), which may indicate a similar role of this element in human cells. We show here that the hairpin structure within the 3 0 UTR influences the expression of a luciferase reporter gene. Lowering of luciferase mRNA levels for the construct with an intact hairpin structure in contrast with almost unchanged mRNA levels for the mutated structure indicate that the hairpin is in fact an RNA destabilizing element. However, at the protein level, expression for both constructs exceeds the expression of the control. This indicates the presence of at least two regulatory events in which the hairpin structure is involved: (i) mRNA degradation, (ii) enhanced mRNA transport (possibly coupled with mRNA stabilization) and/or enhanced translation. These effects may be the consequence of the competitive binding of trans-acting factors for the binding site within the hairpin. Thus, identification of the hairpin-binding factors is important for assessment of its physiological role. A yeast three-hybrid screen identified Hax-1 as the binding partner for the hairpin structure of the Pol b 3 0 UTR. Hax-1 is an RNA-binding protein, known as an anti-apoptotic factor (36) associated with cytoskeletal proteins and involved in cell migration (38, 41) . Hitherto only one RNA target of Hax-1 has been identified: vimentin mRNA (42) . Data from several reports suggest that Hax-1 binding to the hairpin element within the 3 0 UTR of the vimentin transcript plays a role in its localization to the perinuclear cytoplasm (19, 20, 42) . The importance of proper vimentin transcript localization is illustrated by the fact that its misdirection alters cell morphology and motility (43) . The identification of a second RNA target of Hax-1-the hairpin element present in the Pol b transcript-raises the question as to the identity of other mRNA targets of the protein. One may speculate that there is a pool of such mRNAs, especially because vimentin and Pol b are not functionally or evolutionarily related nor are they involved in the same pathway. Comparison of the hairpin motifs in vimentin and Pol b mRNAs did not reveal any significant similarities, which could suggest substrate requirements for Hax-1 binding. The presence of U-rich single-stranded regions (vimentin: AGUUUU in the terminal loop, Pol b: AGUUAU in the internal loop) represents the only similarity between the two structures, but the helical regions adjacent to this U-rich sequence in the Pol b mRNA are not evolutionarily conserved. The lack of similarities suggests that Hax-1-binding mechanism and affinities might be different for these two structures. It is an open question if Hax-1 is in fact a destabilizing factor or if its actions in respect to the Pol b transcript consist of mRNA stabilization, possibly coupled with its transport and/or localized translation. The fact that Hax-1 binds to the instability element (the hairpin) and does not bind to the stable mutated transcript suggests a role in mRNA destabilization. However, data concerning the role of Hax-1 in the regulation of vimentin transcript, indicating that it facilitates its transport to the perinuclear space, are rather contradictory to its potential functions in mRNA degradation. Another possible explanation is that Hax-1 may stabilize otherwise unstable mRNAs and facilitate their transport or enhance the translation rate, conferring elevated luciferase levels in respect to mRNA levels. If this latter case is true, considering that Hax-1 has been identified as the same trans-acting factor for both vimentin and Pol b mRNAs, the hairpin element within the 3 0 UTR of the Pol b transcript may also represent a motif directing mRNA to the perinuclear space. Perinuclear localization of certain transcripts, and their subsequent translation at this site, could facilitate an efficient nuclear import of newly synthesized proteins (15, 16, 17) . DNA polymerase b as a nuclear protein could also benefit from such a mechanism. To present a satisfactory explanation of the role of Hax-1 transcript binding in the cell, one has to resolve the question of the subcellular location of Hax-1. Hitherto, Hax-1 has been reported to localize predominantly in the mitochondria (35, 36, 37) but it has also been detected in the endoplasmic reticulum (35, 36, 38) , apical membrane (39), lammelipodia (38) and nuclear envelope (35) . In the last case, the presence of Hax-1 in the nuclear envelope was interpreted as a consequence of its association with intracellular membranes (by its putative transmembrane domain), as a continuum of endoplasmic reticulum localization. Our data reveal for the first time, that Hax-1 is associated with the nuclear matrix, which is coherent with its transcript-binding capacity and supports the notion of its role in post-transcriptional regulation. Some new data support Hax-1 association with the nucleus. In a recent report, Kawaguchi (44) shows that Hax-1 is present in the nucleus of systemic sclerosis fibroblasts (but not in normal fibroblasts) and is involved in pre-IL-1a translocation into the nucleus-a process blocked by inhibition of Hax-1. This activity is contradictory to previously reported Hax-1 involvement in the cytoplasmic retention of IL-1a (45) and EBNA-LP (46) . The role of Hax-1 in protein import into the nucleus might suggest that it is shuttling between the nuclear matrix and perinuclear space, transporting different cargo molecules. Participation of the hairpin element in the binding between the Pol b 3 0 UTR and Hax-1 was demonstrated by UV-cross-linking, in which the transcript with an intact hairpin structure bound to the protein, whereas a transcript with a disrupted hairpin did not. Cross-linking also revealed that Hax-1 binds to mRNA only in the form of a dimer. The presence of a band of a molecular weight of 100 kDa indicates that UV exposure cross-linked a complex consisting of RNA bound to a protein dimer. RNA-monomer complexes were not detected. Data from chemical cross-linking lead us to conclude that the RNA molecule is not necessary for dimerization. However, given that only a small percentage of protein dimerizes in vitro in the absence of mRNA, the possibility that RNA binding influences the dimerization rate is tempting, and remains to be assessed. Filter-binding experiments, complementing the crosslinks, showed that the binding, though not completely abolished, is substantially weakened for a mutated sequence with a disrupted hairpin structure. Only the C-terminal part of Hax-1 appears to be involved in mRNA binding, since a truncated peptide bearing only the last 150 aa of the protein suffices for binding the hairpin-containing element, as demonstrated in our experiments utilizing the yeast three-hybrid system ( Figure 5 ). Considering that an interaction with RNA occurs only for dimeric Hax-1, these findings suggest that the domain responsible for the dimerization is also located in the C-terminal part of the protein. From these results, we deduce that BH domains (BH1 and BH2) present in the N-terminal part of the protein do not take part in the dimerization. BH domains and a transmembrane domain (present at the C-terminus of Hax-1) represent the features of Bcl-2 family proteins, but there is no significant sequence homology between these proteins and Hax-1only a weak, partial homology to pro-apoptotic Nip3 (35) . Even though BH domains are known to be important for oligomerization of the proteins from the Bcl-2 family, data seem to exclude the possibility that they are responsible for dimerization of Hax-1. We have demonstrated that the hairpin structure within the 3 0 UTR of the Pol b mRNA represents a posttranscriptional regulatory element. Hax-1 protein, which binds to this element, appears to be an important transacting factor, though the exact mechanism of Hax-1-mediated regulation remains to be elucidated, and the mechanisms implicating its role in control of mRNA stability, transport and/or localized translation must be verified by subsequent experiments. Hax-1 is a multifunctional protein, active in different cellular compartments and involved in various cellular processes. Attention has been focused on its functions in apoptosis and regulation of cell motility, but it seems that it has a more complex mode of action and plays a regulatory role in the context of its specific mRNA targets.

Global public goods and the global health agenda: problems, priorities and potential

The 'global public good' (GPG) concept has gained increasing attention, in health as well as development circles. However, it has suffered in finding currency as a general tool for global resource mobilisation, and is at risk of being attached to almost anything promoting development. This overstretches and devalues the validity and usefulness of the concept. This paper first defines GPGs and describes the policy challenge that they pose. Second, it identifies two key areas, health R&D and communicable disease control, in which the GPG concept is clearly relevant and considers the extent to which it has been applied. We point out that that, while there have been many new initiatives, it is not clear that additional resources from non-traditional sources have been forthcoming. Yet achieving this is, in effect, the entire purpose of applying the GPG concept in global health. Moreover, the proliferation of disease-specific programs associated with GPG reasoning has tended to promote vertical interventions at the expense of more general health sector strengthening. Third, we examine two major global health policy initiatives, the Global Fund against AIDS, Tuberculosis and Malaria (GFATM) and the bundling of long-standing international health goals in the form of Millennium Development Goals (MDG), asking how the GPG perspective has contributed to defining objectives and strategies. We conclude that both initiatives are best interpreted in the context of traditional development assistance and, one-world rhetoric aside, have little to do with the challenge posed by GPGs for health. The paper concludes by considering how the GPG concept can be more effectively used to promote global health.

Although the health of the world's poor has been an apparent humanitarian concern of the world's rich for many years, results based on appeals to such 'humanity' have not been sufficient. Even recent high-profile engagements by the Global Fund Against AIDS, Tuberculosis, and Malaria (GFATM), the United States President's Emergency Plan for AIDS Relief (PEPFAR), the Bill and Melinda Gates Foundation, and WHO disease-targeted programs such as Stop TB and Roll Back Malaria have failed to bring us to the levels of assistance needed to achieve the healthrelated Millennium Development Goals (MDGs). Beginning in the late 1990s, the suggestion emerged to address this situation by encouraging policy makers in rich countries to view health assistance not only as humanitarian but as a selfish investment in protecting the health of their own populations. The key concept underlying this new interpretation is that of 'global public goods' (GPGs) [1] . This paper briefly outlines and clarifies the GPG concept, and identifies two major health GPGs: health research and development (R&D), and communicable disease control. However, just because a problem is global and formidable, or just because the response is multilateral, does not necessarily mean that it has anything to do with the undersupply of GPGs. We show this by considering two major global health innovations, GFATM and the rebranding of traditional health objectives in the form of MDGs. Based on the review, in a concluding section we suggest three ways in which the GPG concept can be more effectively deployed to promote global health development. The GPG concept is an extension of the economic tradition of classifying goods and services according to where they stand along two axes: one measuring rivalry in consumption; the other measuring excludability. Pure private goods are those that we are most used to dealing with in our day-to-day lives, and are defined as those goods (like a loaf of bread) that are diminished by use, and thus rival in consumption, and where individuals may be excluded from consuming them. At the opposite end of the spectrum are pure public goods, which are non-rival (not diminished by use) and non-excludable (if the good is produced, it is freely available to all). Public security is an often-cited example. In between these extremes are 'impure' goods, such as 'club goods', which have low rivalry but high excludability, and 'common pool goods', which have low excludability but high rivalry [2] . Clearly in this case, 'health' itself is a private good, as are the majority of goods and services used to produce health [3] . One of the fundamentals of public economics is that the free market -the interplay of individual supply and demand decisions mediated through the price systemwill result in the provision of less than the collectively optimal level of public goods. Thus, the state has a role to play, either in producing the good directly (the traditional approach) or at least in arranging for its production by a private firm (the increasingly popular 'outsourcing' strategy). Examples of national public goods run from police protection to national security to financial regulation to museums and artistic ensembles. But some goods are quite clearly public at the global level. The classic case is greenhouse gas emission control. A reasonable functional definition would be that a GPG is "a good which it is rational, from the perspective of a group of nations collectively, to produce for universal consumption, and for which it is irrational to exclude an individual nation from consuming, irrespective of whether that nation contributes to its financing" [[3], page 9]. The main issue facing non-national (global or regional) public good provision is how to ensure collective action in the absence of a 'government' to directly finance and/or provide the public good, the response in the case of national public goods [4] . Given the reluctance of voters to support programs some of whose benefits are felt beyond the borders, an aspect deserving special attention is mobilizing non-traditional sources of finance [5] . Cutting across all aspects of the GPG concept is the key fact that collective action is in donor countries' self-interest. The GPG concept thus has a specific meaning within economics. However, it has suffered as it has found currency as an advocacy tool for global resource mobilisation [6] [7] [8] . Since a GPG calls for collective action, then, clearly, one's favourite program must be producing a GPG. This has given rise to "fuzziness" and "trendiness" [ [9] , page 2). The GPG 'tag' is at risk of being attached to anything of particular attraction and importance, to the point that, at the limit, anything promoting development could be considered a GPG. This is to be avoided, as overstretching the concept devalues the validity of the point that there really is a class of GPGs requiring public support or provision [10] . Indeed, Smith et al [11] suggest that the GPG concept may perhaps be most usefully applied to just two aspects of health. The first is research and development (R&D) and the second is communicable disease control (epidemiological surveillance, immunization, and other preventive measures). In the next section, we ask how the GPG concept, particularly the need for collective action, has affected policies and programs in these areas. Health R&D unquestionably has GPG aspects, and there is not enough of it in fields that would benefit poor countries. Historically, the public and the not-for profit sectors have carried out research resulting in new drugs and treatments, but the private for-profit sector now plays the largest role [12, 13] . An important policy question is therefore how to encourage private sector firms to engage in research benefiting poor countries and peoples: the ubiquitous '90-10 problem' (that 90% of global R&D spending in health is targeted at diseases affecting only 10% of the world's population) [14] . A related but distinct question is how to bolster the demand for drugs in low-income countries (and hence firms' willingness to engage in R&D); we touch on this in a section below in which we discuss Advance Purchase Commitments and other financial innovations. A GPG perspective would argue that provision of adequate R&D related to diseases of the poor requires innovative collective action. This no-more-business-as-usual attitude has clearly motivated the explosion in the number of Global Public-Private-Partnerships (GPPPs) undertaking R&D related to diseases of the poor [15] [16] [17] . An order-of-magnitude estimate of GPPPs' annual spending might be US$1 billion [18] . This may be compared with total global health R&D spending on the order of US$100 billion [14] . This sounds small, but when the US$1 billion is compared to the estimated US$2.5 billion spent on health R&D by governments in low-and middleincome countries, the perspective is much more favourable for the important role played by GPPPs. In the area of R&D related to "neglected diseases" of the tropics, GPPPs occupy a decisive position. The GPG perspective supports collective action in the area of infectious disease control when reduction in disease prevalence in Country A has a benefit for Country B as well. Areas in which this is particularly true are diseases for which eradication is feasible (polio) and diseases that are highly transmissible around the world, whether by human carriers (SARS), by trade in products (BSE), or by animal vectors (West Nile Virus, avian influenza). The control of antibiotic resistance is a closely related GPG problem. As in the case of R&D, the GPG perspective has informed a number of major new initiatives to provide (not only develop new means of), communicable disease control. These include the GAVI Alliance (formerly the Global Alliance for Vaccines and Immunization), Stop TB, Roll Back Malaria, and others. The main question such initiatives face is the form that assistance should take. There are basically three types of public health interventions: vertically targeted interventions (focused immunization or disease eradication campaigns, for example), horizontally targeted interventions (universal access to a basic medical care package including vaccination, for example), and sector-wide interventions such as capacity building for improved infrastructure and administration. For many years, donors and health officials in low-and middle-income countries gave emphasis to vertical interventions financed by 'earmarked' funds. Problems with this approach include duplication and lack of coordination among projects, 'recipient fatigue' in health ministries forced to administer multiple grants, distortions in local resource allocation such as poaching skilled personnel, and "crowding out" (of which more below) [19] . All donors and the partner countries have committed themselves, through the 2005 Paris Declaration, to pursue harmonisation of practices, standards, and criteria in foreign assistance. Yet the urge to compete rather than cooperate is strong, and well-entrenched donor practices and protocols disappear only stubbornly. One of the ironies of the current health landscape is that, as many in the public health community moved away from vertical interventions towards broader approaches the GPG perspective has helped to fuel the proliferation of specific infectious disease-targeted programs. Yet the experiences of programs in immunization, malaria control, and tuberculosis control demonstrate that impacts are limited by sector-wide weaknesses such as lack of a cold chain, shortages of skilled personnel, insufficient resources for operating vehicles, etc. The plethora of new vertical initiatives may contain the seeds of its own failure if health systems are not generally strengthened (including the crucial human resources aspect) [20] . The danger is that the GPG agenda will promote focused interventions easy to "sell" to voters at home because they address an identifiable menace, at the expense of broader health system strengthening. One response is to identify the health system as a prime 'access good' -not a GPG itself, but a fundamental requirement for the provision of GPGs [11] . The GPG perspective has contributed to a large number of new programs. However, some caveats are in order. Providing an adequate supply of a GPG requires spending more than would have been spent in the absence of collective action, i.e. "additionality." The additionality debate is complex, and involves at least three questions: -At the individual country level, we may ask whether international assistance for production of GPGs in Country X reduced (or, in development parlance "crowded out") Country X's government spending for production of GPGs. Since most of the countries in question are very poor and local needs take clear precedence over global ones, it is probably safe to answer this question in the negative. -More pressing is the question whether international assistance for production of GPGs in Country X reduced international assistance for production of non-GPGs in Country X. A direct answer to this question is difficult to provide because data on foreign assistance at the level of destination countries are much worse than data by country of origin. Reisen et al [21] , looking at the latter, concluded that the average bilateral donor's allocation of US$1 to GPG production reduced its spending on non-GPG foreign assistance by US$0.25. -Finally, additionality questions may be posed as between donors. The focus of this particular storm con-cerns the activities of the GFATM and PEPFAR. Many have complained that PEPFAR diverts U.S. resources away from GFATM, a multi-lateral agency, to a bilateral and highly politicised program. Questions of donor additionality are not confined to HIV/AIDS. Cohen [17] , in looking at the involvement of GPPPs in health R&D, reports that, in expert interviews, researchers complain that the new availability of philanthropic funds for medical research is "crowding out" funding that would have been received from government agencies such as the U.S. National Institutes for Health. While the additionality of resources is ambiguous, we can answer a closely related issue definitively. Resources being used for the provision of GPGs in the area of health R&D and communicable disease control come from traditional, not innovative, sources. The typical new initiative depends on a philanthropic institution for its start-up, followed by infusions of government support channelled through bilateral aid organizations [15] . Yet, the GPG concept is firmly rooted in the self-interested use of domestic monies and as such sees funding as distinct from current aid and philanthropic flows. While there have been innovative fund raising suggestions ranging from a tax on airline travel to a global lottery, progress has been slow. In the health field, Advance Purchase Commitments and the "front-loading" of international assistance for immunization via the GAVI Alliance's International Finance Facility) represent steps forward [22] . In the first case, donors pre-commit to vaccine purchases if R&D is successful; in the second, future aid commitments are collateralized so that funds can be raised immediately in the international bond markets. These are welcome innovations, but they still tap the same source: international donor agency budgets. Inverting the logic above, in this section we wish to clarify that a number of the major priorities in global health today do not represent GPGs. This does not diminish their importance, but it means that progress in these areas should not be equated with progress regarding the undersupply of GPGs. The two innovations we review are the GFATM and the re-framing of traditional health goals in the form of time-bound Millennium Development Goals. Of the three scourges fought by the GFATM, only tuberculosis can be said with accuracy to represent a global public "bad." Malaria has significant cross-border aspects, but these make malaria control a regional public good (and one requiring collective action at the regional level), not a global one. Apart from R&D aspects, the public good problems associated with HIV/AIDS are regional at most, not global [23] . Despite inflammatory rhetoric heard at the beginning of the pandemic, HIV/AIDS has not proven to be a disease that spreads globally like SARS or pandemic influenza. Obvious cross-border aspects like transmission associated with long-haul truckers and migrant workers in Southern Africa call for a regional, not a global response. Even if its transmission did qualify HIV/AIDS as a GPG, the GFATM response is not dealing with the disease using GPG logic. The main concern of GFATM (and PEPFAR, and the WHO's "Three by Five" program, and the G8 nations' commitment to universal access by 2010) is the provision of subsidized antiretroviral therapy (ART) to AIDS sufferers in low-income countries. ART is rival (therapy made available to one person or nation cannot be made available to another) and excludable (persons can be barred from receiving it). By contrast, AIDS prevention, in the form of media campaigns, condom distribution, voluntary counselling and testing, reduction of sexually transmitted infections, and encouragement of male circumcision, is non-rival (if A remains HIV-negative as a result of a prevention program, his sex partners B and C are protected equally) and non-excludable (no one can prevent C from enjoying the same protection as B). Another argument runs that the destabilizing effect of HIV/AIDS in seriously affected countries gives rise to global impacts, but so do the destabilizing effects of everything else, like unemployment and hunger, that contribute to misery. And again, even it the disease's destabilizing potential did qualify it as a GPG, the international policy response does not prioritize GPG aspects. In a world of finite resources, the provision of ART in lowincome countries must come at the expense of prevention (if resources were not finite, of course, such tragic choices would not need to be made). The lopsided cost-ineffectiveness of ART means that each disability-adjusted life year (DALY) saved by treatment comes at the expense of many, many more DALYs lost in the future because the prevention measures needed to reduce transmission cannot be implemented [[24], p. 139]. If it were the looming scale of the AIDS catastrophe and its global spill over effects that were of greatest concern to donors, they would give priority to prevention, not treatment. The need for collective action against AIDS was not the driving force behind the founding GFATM. It was born rather of frustration, especially on the part of AIDS activists, that good ideas from the field were not receiving deserved support because of donor red tape [25] . The response was to be a funding agency which would not assess proposals itself, relying rather on an independent panel, and would use local accounting firms to monitor implementation. Its comparative advantage would be focusing resources quickly on 'best shot' programs in countries with greatest need, as well as raising the profile of the disease. The hands-off approach to program formulation and implementation, it was argued, would mean that the GFATM would have no agenda of its own; aidrecipient countries would be able (through their representation on the review panel) to set their own priorities. The absence of a programmatic/operational agenda would allow the Fund to concentrate on mobilising and disbursing resources [26] . In sum, the rationale for the GFATM was not so much that a GPG was being undersupplied, but that assistance was not being provided efficiently or in a way consistent with needs. Even if the GFATM is interpreted as a bold initiative -and not all do so, since most proposals still come through governments -we run into the caveats made above. The Fund may have generated additional resources or it may not have -the fact that it has struggled against resource constraints practically since its inception [26] gives some reason to suspect the latter. So do anecdotes making the rounds that the arrival of GFATM in some countries has led bilateral donors (apart from the US, through PEPFAR) to limit their own AIDS efforts. GFATM has not mobilized non-traditional sources of finance. As of mid-2007, out of the US$10.9 billion cumulatively pledged for all three diseases through 2008, only US$707 million comes from private firms, foundations, and individuals; this consists almost in its entirety of a US$650 million grant from the Bill and Melinda Gates Foundation. With its resources being pledged almost entirely from traditional donor countries' aid budgets, the GFATM is replicating the source-structure of existing aid flows. A word on the U.S. PEPFAR is in order. PEPFAR promises US$15 billion (US$9 billion of which are claimed to represent new resources) for the global fight against AIDS but only allocates US$1 billion to GFATM [27] . Over half of the resources are targeted to providing AIDS treatment. Of funds allocated to prevention, a significant proportion is earmarked for faith-based programs encouraging teenaged sexual abstinence and discouraging multiple-partner sex. The ability of PEPFAR to finance activities benefiting key target populations -commercial sex workers and injecting drug users -is tightly constrained by law. It is hard to consider PEPFAR as collective provision of a GPG when the resources it makes available could have been channelled through a genuinely collective institution (GFATM) and when its prevention programs are designed to cater to a domestic political constituency. The main focus of global health development at present is the health Millennium Development Goals (MDGs) (General Assembly, A/55/L.2, September 18, 2000) . The targets associated with the health MDGs are to: (i) reduce child mortality by two-thirds between 1990 and 2015; (ii) reduce the maternal mortality ratio by three quarters between 1990 and 2015; (iii) halt and begin to reverse the spread of AIDS by 2015; and (iv) halt and begin to reverse the incidence of malaria and other major diseases by 2015. Developed countries and the development agencies will, in return for low-and middle-income countries' devoting effort to attaining the MDGs, take primary responsibility to establish a global partnership for development. In the area of health, the associated target is: (v) provide, in cooperation with pharmaceutical companies, access to affordable essential drugs in developing countries. How do MDGs for health relate to the GPG perspective? Some of them, for example, those related to tuberculosis and access to drugs (or at least the R&D aspect of that problem), address GPG problems directly. Others, for example those related to maternal mortality, most child mortality, and HIV/AIDS, respond to humanitarian concerns, not GPG problems. As in the case of the GFATM, when we look carefully at the origins of the MDG approach, we find that GPG logic is absent. The MDGs emerged from profound dissatisfaction with the effectiveness of aid to date and insistence on a "results focus" and improved monitoring and evaluation [28] . This process is the elaboration of a Poverty Reduction Strategy Paper or PRSP [29] ; the PRSP process, in turn is meant to encourage countries to adopt a long-term vision "by bringing out explicit awareness of poverty issues and promoting participation of stakeholders" [ [29] , page 11). PRSPs are meant to be country-driven ('ownership'), results-oriented, and participatory; reflect input of civil society and the private sector [30] . Countries are meant to prioritize the MDGs in accordance with their long-term vision of development needs. The PRSP process is also meant to force explicit linkages between fiscal resource allocation decisions and poverty reduction through the putting-in-place of Medium-term Expenditure Frameworks or MTEFs [31] . "We will recognize country ownership and a partnership of equals," runs the donor governments' position, "if you will deliver results and ensure stakeholder participation." "We will deliver results and ensure stakeholder participation," runs beneficiary governments' position, "if you will acknowledge country ownership and a partnership of equals." This is a laudable win-win outcome -but it responds to a crisis in traditional development assistance, not to the need for collective action to supply GPGs. "Country ownership" and acceding to equal partnership are the last things that would be stressed in an approach built on GPG logic. Far from encouraging donor-country voters to support generous foreign aid programs because they are in their own interest, these discourage them from doing so [25] . To conclude, the two initiatives examined show that massive mobilization of humanitarian assistance in pursuit of common goals should not be confused with collective action to ensure the adequate supply of GPGs. That observation does not lessen the importance of such actions, but it guards us against the fallacy of concluding that, just because there is a multiplication of high-profile innovations, fundamental GPG problems are being effectively addressed. The GPG concept, discovered by the aid community in the late 1990s, can be a powerful tool in promoting global health because it marshals arguments of self-interest. It can be used to identify areas in which global collective action is needed, specify where the costs and benefits will rest and communicate to the public why spending to promote health thousands of kilometers around the world is not a waste of their tax dollars. Yet, we find that the GPG perspective has been a mixed blessing. We looked at two acknowledged GPGs related to health, namely R&D and communicable disease control. While recognition of the need for global collective action has supported a large number of new initiatives, it remains to be determined what the result is in terms of additional funds. Those funds that have been generated have come from traditional philanthropic and public sources. The proliferation of infectious disease initiatives has promoted a vertical, "stovepipe" approach, to the detriment of broad health sector strengthening. We then looked at two of the major global innovations in health, the GFATM (and, closely related, PEPFAR) and the re-packaging of traditional health concerns in the form of MDGs. We concluded that both can be more easily understood as addressing weaknesses in traditional humanitarian aid -red tape, lack of country ownerships, insufficient stakeholder involvement, need for results-based management, etc. -than as addressing problems of GPG provision. All of the new initiatives we have discussed here, and many that we have not mentioned, are funding or doing valuable work. How might the GPG perspective strengthen them and lead to other efforts, as well? First and foremost, within existing programs and when proposing new ones, the aid community should adhere to the strict economic definition and avoid the temptation to use the GPG 'tag' as a general-purpose fund-raiser. If we focus GPG logic on those goods and services where global collective action really is needed, that action is more likely to be achieved. Where humanitarian grounds, not rational self interest, are the main motivation for action -as in providing subsidized treatment for AIDS sufferers in poor countries -we should say so without equivocation. Where general health system strengthening is required to guarantee access to GPGs such as immunization or tuberculosis control, this should be stated explicitly, even if it means that budgets for GPG provision strictly defined may be reduced as a result. Second, the aid community should stress to policy makers that, where the GPG label is appropriate, as in the case of communicable disease control, what is needed is not only new packaging/labeling of existing resources, but resources additional to those already being made available, which means mobilizing innovative sources of financing. The current elevated level of concern over emergent diseases, including pandemic influenza, is an ideal context in which to press for a more pro-active response. So is the rapid development of financial engineering tools related to aid, such as advance purchase commitments, collateralization future aid commitments in the bond market so as to "frontload" aid, etc. Third, the relative ease of financing disease-specific actions, as opposed to broad sector strengthening, should not be allowed to distort health sector policy or dictate the structure of support. Where sector support serves an "access" function, the argument that it is a prerequisite for provision of GPGs (essentially, communicable disease control) can be used to strengthen its claim on resources. The aim of this paper was to provide an introduction to the key concepts, and to consider some innovative developments in global health from the GPG perspective. Hopefully this has illustrated the potential and limitations of the concept, and provided a foundation for further discussion of these.

Outcome of paediatric intensive care survivors

The development of paediatric intensive care has contributed to the improved survival of critically ill children. Physical and psychological sequelae and consequences for quality of life (QoL) in survivors might be significant, as has been determined in adult intensive care unit (ICU) survivors. Awareness of sequelae due to the original illness and its treatment may result in changes in treatment and support during and after the acute phase. To determine the current knowledge on physical and psychological sequelae and the quality of life in survivors of paediatric intensive care, we undertook a computerised comprehensive search of online databases for studies reporting sequelae in survivors of paediatric intensive care. Studies reporting sequelae in paediatric survivors of cardiothoracic surgery and trauma were excluded, as were studies reporting only mortality. All other studies reporting aspects of physical and psychological sequelae were analysed. Twenty-seven studies consisting of 3,444 survivors met the selection criteria. Distinct physical and psychological sequelae in patients have been determined and seemed to interfere with quality of life. Psychological sequelae in parents seem to be common. Small numbers, methodological limitations and quantitative and qualitative heterogeneity hamper the interpretation of data. We conclude that paediatric intensive care survivors and their parents have physical and psychological sequelae affecting quality of life. Further well-designed prospective studies evaluating sequelae of the original illness and its treatment are warranted.

the quality of life (QoL) in survivors and in their families are important outcome measures. Historically, outcome research in paediatrics is either based on an age-specific approach, such as follow-up studies of premature infants [41, 72, 73] , or on a more disease-oriented approach, such as follow-up studies in survivors of cardiothoracic surgery or trauma [15, 55, 64, 70] . These studies have shown substantial physical, psychological and neurocognitive sequelae, interfering with daily life and normal development. In addition, effects on parents and siblings have been shown [26] . Evaluative research of adult intensive care survivors showed the effect of intensive care treatment per se. Irrespective of the underlying illnesses, sequelae on all domains with effects on QoL were found [2, 19, 58, 75] . In multi-disciplinary paediatric intensive care unit (PICU) populations, reports on outcome are scarce [24, 25] . Based on these observations, we believe that follow-up research of paediatric intensive care survivors and their families is needed to evaluate: (1) physical sequelae and their impact during growth and development; (2) psychological sequelae in patients and their families and their impact on the QoL of patients and family members; and (3) the need for treatment and support after discharge. The aim of this article is to provide an overview of the available literature concerning the different domains of QoL (i.e. physical, psychological and social functioning) in children surviving paediatric intensive care, including the effect on parents, and to suggest directions for future follow-up research. To identify studies eligible for this review, we searched Medline , EMBASE (1974 EMBASE ( -2006 , CINAHL (1982 CINAHL ( -2006 , pre-CINAHL and the Cochrane Library (2006) in March 2006. In the search strategy, all terms mapped to the appropriate MeSH/EMTREE subject headings and "exploded" were used; among them were: paediatric intensive care unit (PICU), septic shock, respiratory insufficiency, meningococcal disease, central venous catheterisation, intubation, physical and psychological sequelae, post-traumatic stress disorder (PTSD), QoL, health status and long-term outcome. Functional health is defined as an individual's ability to perform normal daily activities, to fulfil usual roles and to maintain health and well-being. QoL is defined as an individual's perception of their position in life, in the context of the culture and value systems and in relation to their goals, expectations, standards and concerns [1] . Health-related QoL (HRQoL) is defined as QoL in which a dimension of personal judgement over one's health and disease is added [21] . Studies were selected for review if they met two inclusion criteria: (1) study of a representative population of PICU survivors (defined as a population consisting of medical and/or surgical PICU patients <18 years old) and (2) evaluation of physical sequelae, measurement of QoL or functional health >30 days after PICU discharge. Because of the limited number of studies, the measurement tools did not need to be standardised. Studies with a retrospective and prospective design were included. Excluded were: (1) studies in homogeneous PICU populations (e.g. survivors of cardiothoracic surgery and trauma) reporting diagnosis-related outcome in particular but not intensive care treatment as such, and (2) studies evaluating mortality only. Eligible studies and quality of the studies Twenty-seven studies were found in which one or more aspects of long-term sequelae in PICU survivors and/or their families were described. The patient characteristics, populations, measurement tools and outcomes are described in Tables 1 and 2 . The quality criteria are described in Table 3 . None of the studies met all of the quality criteria. In studies describing the same outcome aspect, differences in study population, follow-up time and measurement tools make the comparison and synthesis of results difficult. Physical and neuro-cognitive sequelae (Table 1) In 12 studies that included in total 340 patients, aspects of physical and neuro-cognitive sequelae were evaluated. Neurological evaluation was conducted in five studies including 275 survivors. The majority of the children were neurologically normal. In the remaining children, disabilities such as hearing loss, coordination, cognition and developmental problems turned out to be severe [23, 35, 43, 53, 59] . Pulmonary evaluation was conducted in six studies including 65 patients [6, 14, 22, 30, 48, 74] . Restrictive and obstructive disease and hypoxaemia during exercise was found. Cardiac evaluation was conducted in two studies including 23 survivors [22, 74] . No abnormalities were found, except for left ventricular hypertrophy in one child. Renal evaluation was conducted in one study including 12 survivors [62] . In two children, glomerular filtration was impaired, one had hypertension and one had proteinuria. Psychological sequelae (Table 2) Various questionnaires were used. Cut-off points for the diagnosis of PTSD differed between studies but all of them showed high scores for PTSD in children and parents. Psychological evaluation of children was conducted in five studies including 202 children [40, [50] [51] [52] 61] . Symptoms of PTSD were found in 11 of 74 evaluated children. In one study, a relation was found between invasive procedures and high scores [52] . Psychological evaluation of parents was conducted in six studies including parents of 547 children [4, 8, 20, 40, 50, 61] . Symptoms of PTSD were found in 72 of 295 evaluated parents. In some studies, a relation was found between high scores and illness severity as perceived by parents [4, 50, 61] . In one study, these high scores decreased over time [8] . Functional health and QoL (Tables 1 and 2) Evaluation of functional health was conducted in three studies including 821 children [9, 12, 44] . The majority of the children seemed to have normal functional health; the remainder was found to be seriously impaired. Evaluation of QoL was conducted in four studies including 1,664 children [27, 38, 47, 65] . QoL was evaluated using three different questionnaires. In the majority of children, the QoL was normal or equal to the QoL before PICU admission. In all studies, some of the children had poor QoL. Only 27 studies consisting of 3,444 PICU survivors met our inclusion criteria. The small numbers, heterogeneity of the 4 yes yes yes yes no 6 no no yes yes no 8 no no yes yes yes 9 no yes yes no no 12 yes yes yes yes no 14 yes yes yes yes no 20 yes yes yes yes no 22 no no no yes no 23 no no yes yes yes 27 yes studied populations and the used measurement tools, the frequent use of non-validated measurement tools and the various aspects of outcomes studied make aggregation of the data and, therefore, strong conclusive statements difficult. The reviewed studies report distinct physical sequelae, including neurological abnormalities in PICU survivors. Standardised neurological examination of PICU survivors was validated in 1994 but very few studies have been carried out since [24, 25] . As neurological problems have a great impact on daily life, standardised evaluation and adequate support and rehabilitation seem to be relevant, similar to in NICU survivors [11, 46, 56] . Follow-up studies evaluating lung function in children are hampered by the small incidence of severe respiratory insufficiency in children [49] . In adult respiratory distress syndrome (ARDS), the recovery of lung function is shown during the first year and physical limitations seem to be partly dependent on lung function [34, 58] . In infants and children, post-natal lung growth may contribute to the improvement of lung function after critical illness. In addition to lung function, the long-term effect of small airway disease should be evaluated, for instance, in children with respiratory syncitial virus infection. Data on the structured evaluation of cardiac and renal function in paediatric and adult ICU survivors is not available. In young children, septic shock and the need for vasoactive support of the circulation may interact with the developing myocardium and may have persistent effects on cardiac growth and function [10, 67, 77] . Complications of intensive care procedures per se, (e.g. vascular complications due to intra-vascular catheters and side-effects of ototoxic drugs and sedatives) are not evaluated [5, 18, 32, 33, 45, 54, 57, 63] . One can assume the exact incidence of physical sequelae to be higher than has been reported so far. In the reviewed studies, psychological sequelae have been established in 10-14% of survivors and their parents. The comparison of findings is hampered due to different measurement tools and cut-off points for the diagnosis of PTSD and various follow-up intervals. Risk factors accounting for hampered psychological outcome could be diverse (severity of illness, being removed from one's child, having been witness to the accident, mental health, family functioning, social support, coping strategies and lack of information from the medical team) [17, 26, 29, 31] . Psychological support to improve coping strategies and prevent over-protection might improve psychological out-come in children and parents [3, 28] . Further research is essential to establish the appropriate time and extent of the psychological support needed. Cognitive sequelae have rarely been studied in the reviewed studies. Adequate neuro-cognitive evaluation is both expensive and time-consuming. Studies in neonatal ICU survivors show substantial cognitive dysfunction with great impact on daily life [7] . Consequently, early intervention, education and rehabilitation are expected to improve daily life [11, 46] . A majority of PICU survivors seem to have unchanged functional health and good QoL. In the reviewed studies, functional health is evaluated by telephone interviews [27, 38, 47, 65] . In most of these studies, the physician rather than the child or its parents evaluates functional health. Ideal (HR)QoL questionnaires should measure all aspects of QoL and preferably be filled in by the children themselves. Proxy investigation of functional health and (HR) QoL (in children <6-8 years of age) is second best [36, 37, 39, 66] . Besides, the pre-morbid state is probably an important factor which is difficult to assess [16] . The reviewed studies have a number of methodological limitations. Heterogeneity is the most important one. Consensus on all aspects of follow-up research is essential for well-founded conclusions. For example, structured and standardised evaluation of: (1) organ system function with a validated tool such as the Paediatric Logistic Organ Dysfunction (PELOD) score [13, 42, 60, 71] ; (2) neurocognitive function; (3) complications of PICU treatment; and (4) (HR)QoL are warranted. Multi-centre studies as proposed by the Collaborative Pediatric Critical Care Research Network (CPCCRN) with a uniform approach will provide answers either in general PICU cohorts or in disease-oriented study groups [76] . In conclusion, this review indicates that PICU survivors and their parents may have substantial physical and psychological sequelae interacting with QoL. Because of longer life expectancy, longer follow-up time is warranted, emphasising the consequences for health care in children. We believe that paediatric intensivists and psychologists should be involved as core members of follow-up teams.

An evaluation of Comparative Genome Sequencing (CGS) by comparing two previously-sequenced bacterial genomes

BACKGROUND: With the development of new technology, it has recently become practical to resequence the genome of a bacterium after experimental manipulation. It is critical though to know the accuracy of the technique used, and to establish confidence that all of the mutations were detected. RESULTS: In order to evaluate the accuracy of genome resequencing using the microarray-based Comparative Genome Sequencing service provided by Nimblegen Systems Inc., we resequenced the E. coli strain W3110 Kohara using MG1655 as a reference, both of which have been completely sequenced using traditional sequencing methods. CGS detected 7 of 8 small sequence differences, one large deletion, and 9 of 12 IS element insertions present in W3110, but did not detect a large chromosomal inversion. In addition, we confirmed that CGS also detected 2 SNPs, one deletion and 7 IS element insertions that are not present in the genome sequence, which we attribute to changes that occurred after the creation of the W3110 lambda clone library. The false positive rate for SNPs was one per 244 Kb of genome sequence. CONCLUSION: CGS is an effective way to detect multiple mutations present in one bacterium relative to another, and while highly cost-effective, is prone to certain errors. Mutations occurring in repeated sequences or in sequences with a high degree of secondary structure may go undetected. It is also critical to follow up on regions of interest in which SNPs were not called because they often indicate deletions or IS element insertions.

Genome resequencing is the determination of a genome sequence using an already established genome sequence as a reference. In hybridization-based resequencing, the reference is necessary for the generation of microarray probes and for signal normalization. In other types of resequencing the reference is used as a scaffold for the assembly of short sequence reads. The genome to be rese-quenced must be substantially similar to the reference; otherwise the reference looses its effectiveness. Resequencing is useful for relating phenotype to genotype and for analyzing natural variation. It has been used to study the acquisition of antibiotic resistance [1, 2] , the analysis of variation in pathogenic bacteria and viruses [3] [4] [5] [6] [7] , and to study the experimental evolution of bacteria and yeast [8] [9] [10] [11] . Methods of resequencing utilize micro-arrays [2, 7, 10] , polonies [9] or sequencing-by-synthesis technology [12] . The Comparative Genome Sequencing service provided by Nimblegen Systems Inc. is a hybridization-based method and consists of two steps [2] . In the first step, an experimental and reference genomic DNA sample are labeled and hybridized to microarrays containing ~30-mer 'tiled' oligonucleotides spaced every ~7 bp of the genome on both strands. Probes cover every nucleotide of the genome and are designed to have isothermal hybridization characteristics. Probes showing differences in signal intensity are flagged as Regions Of Interest (ROIs) for further investigation. A second microarray is then designed to interrogate the ROIs at single base-pair resolution, with all four possible nucleotides synthesized for each interrogated position on both strands (8 probes per position). If the sequence difference is a single-base polymorphism, the pattern of probe intensities at that position will be discernibly different than expected. Other types of sequence differences such as insertions/deletions (indels) or differences of more than one bp will not be determined conclusively. In order to attribute phenotype to a genotype it is important to know that all relevant sequence differences were detected. False negatives (i.e. failures to detect mutations that are actually present) can result in erroneous interpretation, especially considering that a single mutation can have a dramatic impact on phenotype. False positives on the other hand (i.e. the reporting of sequence differences that are not actually present) are not a big problem in most cases since even 100 false mutations can be checked and refuted with PCR amplification and Sanger sequencing for less money than a typical resequencing experiment. In order to determine the false negative and false positive rates for CGS, we utilized two related strains of E. coli for which high-quality genome sequences exist. Strain W3110 was resequenced with strain MG1655 as the reference, but this setup could just as easily have been reversed. Our results demonstrate the accuracy of CGS resequencing but also raise caveats about CGS and the instability of bacterial strains. The closely related and fully sequenced E. coli strains W3110 and MG1655 represent an excellent test for the accuracy of resequencing technology. Hayashi et al. [13] compared the genome sequences of these two strains and resolved all discrepancies, generating a pair of highly accurate genome sequences. The genome sequence of strain W3110 differs from MG1655 by seven single nucleotide polymorphisms (SNPs), one 2 bp insertion, 12 IS element insertions, one deletion of 6.6 kb, and an inversion of 783.1 kb [13] . A summary of all sequence differ-ences between strain MG1655 and W3110 is presented in Table 1 . To test the accuracy of CGS, E. coli strains W3110 Kohara and MG1655 were obtained from the E. Coli Genetic Stock Collection (CGSC). DNA was extracted and submitted to Nimblegen Systems Inc. for CGS, using MG1655 as the reference. Nimblegen reported 25 SNPs in strain W3110, 4 of which corresponded to known sequence differences. Regions surrounding the other SNPs were PCR amplified and subjected to Sanger sequencing. In this way, 19 of the remaining putative SNPs were refuted, though 2 were confirmed. These two mutations are not present in the W3110 genome sequence, and must have been introduced into the strain after Kohara et al. [14] made the lambda phage library that was sequenced. In sum, there were 19 false positive SNPs out of 25 reported, or one false positive per 244 kb ( Table 2 ). Figure 1 shows a sample of hybridization signals surrounding three different sequence differences. It can be seen that the ratio of signal intensities near the mutations were distinctly elevated. When these regions were resequenced at single bp resolution, the sequence differences in ycdT and acnA were correctly determined. The intensity of hybridization signals in other areas where there were no sequence differences varied considerably. In addition to SNPs, Nimblegen also provided the location of probes that showed hybridization differences, yet could not be resolved as SNPs ("non-called Regions Of Interest" or ROIs). These probes may indicate the presence of mutations other than SNPs, such as insertions or deletions. Nimblegen reported 1094 ROI probes located in 36 clusters, a cluster being a group of probes that are within 500 bp of each other. One of these clusters corresponded to the known 6.6 kb deletion of intZ thru yffS (b2442-2450) ( Table 1) . Three other clusters corresponded to known sequence differences that were not reported as SNPs -the substitutions in rpoS and crp and the 2 bp insertion in dcuA. Nine clusters corresponded to known IS-element insertions in W3110. Most of the IS-element insertions were evident as clusters consisting of multiple ROI probes, though three were only evident as singleprobe clusters. There were three IS element insertions that were not detected as either SNPs or ROIs. The large inversion in W3110 of all genes between ribosomal RNA genes rrlD and rrlE was not detected as either SNPs or ROIs. One novel cluster of ROI probes was very large, indicating the possible deletion of three genes, ynaJ, uspE, and fnr (b1332-1334). Such a deletion is not present in the genome sequence of W3110. PCR amplification using primers located in adjacent genes confirmed the deletion; a 3.8 kb PCR product was obtained from MG1655 while a 1.4 kb product was obtained from W3110. Deletions of fnr have been noted previously in strains of MG1655 obtained from the CGSC [16] . To determine if the deletion occurred before or after the strain was deposited at CGSC, a culture was obtained from the original lyophil made when stock # 7167 was deposited in 1990 by Akira Ishihama. PCR amplification with the same primers showed two products, one at 1.4 kb and another at 5 kb. Sequencing the ends of the 5 kb band revealed the known IS5 element b1331 to the right of ynaI on one side and a new IS5 element inserted to the left of ogt (b1335) on the other. This result seems to indicate that the strain of W3110 deposited with CGSC has two copies of IS5 on either side of ynaJ, uspE and fnr that undergo recombination with each other at high frequency leading to the deletion of the intervening genes. The remaining 22 clusters of ROI probes were PCR amplified and Sanger sequenced to see if they indicated additional mutations not present in the W3110 genome sequence. Indeed, seven IS-element insertions were discovered, while the other 15 clusters showed no mutations. We note that all but one of the false positive ROIs were single-probe clusters and that all of them contained inverted repeats of between 6 and 13 nt. These repeats may lead to hairpin structures and poor hybridization properties of those probes [15] . In total, 10 mutations (2 SNPs, 7 IS-insertions and 1 deletion) accumulated in strain W3110 in the time period between when Kohara et al. generated the lambda library and when we obtained it from CGSC. The history and handling of the strain before Ishihama deposited it with CGSC in 1990 is not known by the authors of this study, though we speculate that it may have been stored as a stab at room temperature. In this study, we sought to evaluate the accuracy of Nimblegen's microarray-based CGS resequencing technology so that the results obtained from experimental samples can be interpreted judiciously. Our results indicate that the false positive rate for SNPs was one per 244,193 bp of sequence. The false positive rate for clusters of ROI probes was one per 309,312 bp of sequence. These rates have been shown to depend on the thresholds chosen for mutation-calling [17] . In theory, low thresholds should be used for mutation mapping, increasing the number of false positives but reducing the number of false negatives. The methods used by Nimblegen to identify ROI's and to determine sequence differences have been described elsewhere [2] [18] . ROIs are identified by discarding erroneous probes then picking probes for which the ratio on both strands exceeds 3.5 standard deviations of the 80 th percentile within a 1800 bp local window [2] . For sequence determination, Molla et al. [18] . describe a machine-learning algorithm that uses relative positions within "feature space" rather than a training set and describe the effects of changing thresholds on sensitivity. In future work, false positive rates might be reduced by additional development of these algorithms and discard-ing single-probe clusters containing inverted repeats. Error analysis for development of improved algorithms or array designs should take into account whether errors occurred in the first "mapping" step or the second single bp resolution step of CGS. A high false discovery rate is generally not a problem with resequencing projects. The cost of PCR amplifying and Sanger sequencing up to 100 candidate regions is small compared to the cost of resequencing itself. The false negative rate on the other hand is very important. In order to associate genotypic change to phenotypic change it is critical to know with some certainty that all of the important mutations were detected. Nimblegen estimates that the false negative rate for SNPs is less than 5% (Tom Albert, personal communication). We found that only half of the small sequence differences actually present were reported as SNPs. If we consider detection in ROI probes as well, then 7 out of 8 small sequence differences were detected, yielding a false negative rate of 12.5%. Given the small sample size, this rate may be consistent with Nimblegen's claims. CGS is meant for the detection of SNPs, but ROI data also reveals deletions and IS-element insertions. The false negative rate for detection of IS element insertions was double the rate for small sequence differences, possibly because CGS has not been optimized for detection of insertions. Conceptually, insertions and inversions could result in a more pronounced hybridization difference at the insertion site. If the insertion site/inversion breakpoint occurs in the middle of the region covered by a probe, then genomic DNA will only match one half of the probe or the other. Detection of insertions might be improved by manipulation of the hybridization/wash conditions or changes to the algorithm to specifically identify probe series indicative of insertions. The single SNP not detected is a substitution in the gene rrlE, which is duplicated 6 other times in the E. coli genome. In previous work, we noted the failure of CGS to detect a 9 bp duplication, a 1 bp deletion and a 28 bp deletion near a transcriptional terminator [8] . It appears that CGS has trouble in regions of high local secondary a. This number only includes those differences reported in reference 13, and not the new ones discovered in this study. b. The number of non-called ROI's that did not contain mutations was taken as the number of false positives for both IS-element insertions and indels. Sample mutation mapping data Figure 1 Sample mutation mapping data. The ratio of signal of W3110 vs. MG1655 is shown for probes spaced every ~7 bp surrounding an IS5 insertion in dcuC (top), a SNP in ycdT (middle), and a SNP in acnA (bottom). acnA region structure and with sequences repeated multiple times in the genome [15] [17] . In addition, detection of SNPs may be less accurate in areas with low overall intensity. Fortunately, such regions are relatively uncommon, and CGS appears to detect most mutations with a reasonable degree of confidence. Site-directed mutagenesis can overcome uncertainty caused by the possibility of false negatives. In our previous work [8] , we introduced all of the mutations identified into the wild type strain and were able to reconstruct the observed change in growth phenotype in 4 of 5 cases, indicating that all of the important mutations were found in most cases. An additional precaution against false negatives is resequencing multiple replicates from the same experimental setup. Our previous work identified mutations in the gene glpK in 4 of the 5 replicates. Sequencing glpK from the one remaining replicate with other technology showed a 9 bp duplication undetected by CGS. This demonstrates the value of replicates and why genes containing mutations in some replicates should be screened for false negatives in the other replicates. Resequencing technology allows a sequence to be determined at less than one tenth the cost of traditional Sanger sequencing, but it is less accurate with some types of sequences and requires a closely related reference sequence. Other methods of resequencing [9, 12] are likely to differ from CGS in cost and accuracy. The control experiment presented here comparing E. coli strains W3110 and MG1655 can now be performed with these other technologies to allow cross-comparison. We determined the false positive rate for CGS to be one per 244 Kb for SNPs and one per 309 Kb for non-called ROIs. We observed a false negative rate of 12.5% for small sequence differences and 25% for IS-element insertions. The one large deletion present in the genome sequence was easily detected, though the chromosomal inversion was not. We conclude that the accuracy of CGS is sufficient for effective resequencing studies, but with some precautionary notes. All clusters of non-called ROI probes should be PCR amplified and Sanger sequenced to detect IS-element insertions and small indels. Also, multiple replicates and/or site-directed mutagenesis should be used in cases where rare false negatives may affect the scientific interpretation. We made the unexpected discovery of 10 mutations accumulated in strain W3110 after Kohara et al. generated the lambda clone library that was sequenced. This highlights the ability of CGS to reveal unexpected results and the rapid degeneracy of strains under some kinds of storage conditions. Strains with the same name (e.g. W3110, MG1655) often differ depending on their source and storage conditions, so sequenced strains should only be obtained from the laboratories that sequenced them or their designated depositories. E. coli strain W3110 Kohara and MG1655 (seq) were obtained from the E. Coli Genetic Stock Collection (CGSC # 7167 and # 7740, respectively). Cells were grown in liquid LB medium and then DNA was extracted using DNAeasy Tissue Kit (Qiagen). The CGS results provided by Nimblegen are given in Additional File 1. Regions containing putative mutations were PCR amplified using oligonucleotide primers listed in Additional File 2. They were then purified using Qiaquick (Qiagen) and sequenced using Sanger sequencing. Sequence differences were confirmed by manual examination of the trace data.

Antibody-Based HIV-1 Vaccines: Recent Developments and Future Directions: A summary report from a Global HIV Vaccine Enterprise Working Group

The authors discuss humoral immune responses to HIV and approaches to designing vaccines that induce viral neutralizing and other potentially protective antibodies.

T he Global HIV Vaccine Enterprise convened a two-day workshop in May of 2007 to discuss humoral immune responses to HIV and approaches to design vaccines that induce viral neutralizing and other potentially protective antibody responses. The goals of this workshop were to identify key scientific issues, gaps, and opportunities that have emerged since the Enterprise Strategic Plan was first published in 2005 [1] , and to make recommendations that Enterprise stakeholders can use to plan new activities. Most effective viral vaccines work, at least in part, by generating antibodies that inactivate or neutralize the invading virus, and the existing data strongly suggest that an optimally effective HIV-1 vaccine should elicit potent antiviral neutralizing antibodies. However, unlike acute viral pathogens, HIV-1 chronically replicates in the host and evades the antibody response. This immune evasion, along with the large genetic variation among HIV-1 strains worldwide, has posed major obstacles to vaccine development. Current HIV vaccine candidates do not elicit neutralizing antibodies against most circulating virus strains, and thus the induction of a protective antibody response remains a major priority for HIV-1 vaccine development. For an antibody-based HIV-1 vaccine, progress in vaccine design is generally gauged by in vitro assays that measure the ability of vaccine-induced antibodies to neutralize a broad spectrum of viral isolates representing the major genetic subtypes (clades) of HIV-1 [2] . Although it is not known what magnitude and breadth of neutralization will predict protection in vaccine recipients, it is clear that current vaccine immunogens elicit antibodies that neutralize only a minority of circulating isolates. Thus, much progress needs to be made in this area. Also, though virus neutralization is considered a critical benchmark for a vaccine, this may not be the only benchmark for predicting success with antibody-based HIV-1 vaccine immunogens. The main targets for neutralizing antibodies to HIV-1 are the surface gp120 and trans-membrane gp41 envelope glycoproteins (Env) that mediate receptor and coreceptor binding and the subsequent membrane fusion events that allow the virus to gain entry into cells [3] . Antibodies neutralize the virus by binding these viral spikes and blocking virus entry into susceptible cells, such as CD4 + T cells [4, 5] . In order to chronically replicate in the host, the virus exploits several mechanisms to shield itself against antibody recognition, including a dense outer coating of sugar molecules (N-linked glycans) and the strategic positioning of cysteine-cysteine loop structures on the gp120 molecule [6] [7] [8] . These shielding mechanisms, although highly effective, have vulnerabilities imposed by fitness constraints. Information on the precise location and molecular structure of these vulnerable regions could be valuable for the rational design of improved vaccine immunogens. Participants in the workshop identified four areas that, if given proper attention, could provide key information that would bring the field closer to an effective antibody-based HIV-1 vaccine: (1) structure-assisted immunogen design, (2) role of Fc receptors and complement, (3) assay standardization and validation, and (4) immunoregulation of B cell responses. Clinical studies have demonstrated that immunization with the gp120 surface unit of the HIV-1 envelope protein does not lead to the induction of potent or broadly reactive neutralizing antibodies. In order to develop better immunogens, it is likely that we will need a more detailed understanding of the atomic level structure of epitopes on the native envelope glycoprotein. Data on the X-ray crystal structure of liganded and unliganded partial gp120 molecules have provided valuable information about the atomic level interaction of gp120 and neutralizing antibodies [9] [10] [11] [12] . The recent atomic level resolution of monoclonal antibody (MAb) b12 bound to the CD4 receptor binding site of the gp120 molecule provides new insights into how successful neutralizing antibodies access functionally conserved regions of the Env glycoprotein [13] . Crystal structures of complete monomeric gp120 and gp120-gp41 trimer complexes in their native unliganded form need to be elucidated, as these are the natural targets for neutralizing antibodies. This information is needed for multiple genetic subtypes of the virus and for transmitted strains of the virus. Coupled with this effort should be a program to make necessary improvements in electron tomography technology to gain a higher resolution of native Env spikes as they exist on virus particles [14] [15] [16] . An improved understanding of the structural basis of antibody binding to the HIV-1 Env glycoprotein will likely form the foundation for a rational program of novel vaccine design. Ongoing efforts to stabilize gp120 into more immunogenic forms or to scaffold conserved neutralization epitopes into foreign proteins may lead to more promising antibody responses. Induction of an effective neutralizing antibody response will require that a vaccine deliver to the naïve B cell repertoire epitopes that are both immunogenic (i.e., possess favorable properties for B cell inductive pathways) and antigenic (i.e., available for high affinity antibody binding on functional Env spikes). Viral epitopes that are conserved among most viral strains are more likely to generate cross-reactive antibodies. In this regard, researchers have focused on a small number of human MAbs, from clade B HIV-1-infected individuals, that possess broadly cross-reactive neutralizing activity [17, 18] . The cognate viral epitopes for these MAbs have been well characterized and are being evaluated as vaccine immunogens. However, for reasons that are not completely understood, these conserved viral epitopes have either been poorly immunogenic or have elicited antibodies of restricted reactivity. Improvements are being sought by introducing specific structural alterations [19, 20] and by targeting autoreactive B cell pathways [21] . These and other efforts to improve the immunogenicity of conserved neutralization epitopes should remain a high priority. Workshop participants recognized the need to expand efforts to identify and characterize new MAbs, with special attention to MAbs from non-clade B HIV-1 infections. New technologies are now available that might afford an advantage for identifying novel antibody specificities that were previously undetected [22, 23] . In addition to this focus on MAbs, sera from selected HIV-1infected individuals that can broadly neutralize HIV-1 isolates should be studied in detail. New assays allow more precise mapping of the polyclonal antibody response in these sera to better understand the epitopes targeted [5, [24] [25] [26] . Such studies may reveal novel antibody specificities and their associated viral epitopes that could be useful for immunogen design. While there has been considerable interest in conserved epitopes, less attention has been paid to more variable epitopes that might be useful if administered in the form of a polyvalent vaccine. Of particular interest are the epitopes that drive the autologous neutralizing antibody response in infected individuals. These epitopes may be quite variable, but recent evidence suggests that there are constraints on the extent of variation the virus can tolerate in these regions [27, 28] . Detailed molecular and immunologic studies of the autologous neutralization response would enhance our understanding of viral determinants that are vulnerable to antibody attack. Similarly, it is possible that combinations of antibodies will have desirable additive or synergistic effects on virus neutralization [29-32]. An example is seen in how soluble CD4 binding rearranges the structure of gp120 to expose the highly conserved coreceptor binding domain, which allows antibody binding and virus neutralization to occur [33, 34] . Such effects of antibodies might be discovered by applying high throughput screening methods to the plethora of existing MAbs as well as new MAbs that become available in the future. Recent findings have generated renewed interest in so-called "non-neutralizing" antibodies that are unable to directly inhibit free virus entry into target cells, but nonetheless exhibit antiviral activity mediated by the Fc region of the antibody molecule. These antibody effector mechanisms include complement binding and viral lysis, phagocytosis of antibody-coated virions, and antibody-dependent cellular cytotoxicity [35] [36] [37] [38] . Recent studies have suggested examples of Fc-dependent antiviral effects of HIV-1-positive serum in cases where there was little or no detectable activity in conventional neutralization assays [39, 40] . In addition, passive transfer studies in a relevant monkey model suggest that Fc receptor (FcR) binding capacity of a protective antibody makes a substantial contribution to the antibody-mediated protection [41] . Antibody effector functions that mediate complement activation and FcR engagement on macrophages, dendritic cells, natural killer cells, and other cell types need to be evaluated to determine their relevance to HIV-1 vaccines. Assays that measure these antiviral antibodies should be standardized and used to assess biologic relevance in passive protection experiments in animal models using antibodies that exhibit the different effector functions in vitro. In order to adequately monitor neutralization breadth and potency and to compare and prioritize immunogens, assays are needed that are sensitive, quantitative, high throughput, and have correlative value. Substantial improvements have been made in the past several years in assay technology and in available reference reagents. Thus, cumbersome and expensive assays using peripheral blood mononuclear cells (PBMC) and uncloned viruses are being replaced with assays that utilize molecularly cloned Env-pseudotyped viruses and genetically engineered target cells lines [2, [42] [43] [44] [45] . This new technology affords greater sensitivity, reproducibility, high throughput, and cost-effectiveness compared to PBMC assays, and as a result, it has been responsible for an explosion of new data. Steps are being taken by the Collaboration for AIDS Vaccine Discovery to transfer this new technology to multiple laboratories around the world and to implement a validated proficiency testing program to assure inter-laboratory equivalency in assay performance. Recently, several cases were identified where neutralization was considerably more potent or only detected in the older PBMC assay compared to the newer assay technology [43, 46, 47] . This raises important questions about current plans to employ a single assay for routine use, and it points to the need for a better understanding of the mechanisms of neutralization. Thus, it may be necessary to use more than one assay to assure that all neutralizing antibodies are detected. There is a need to standardize and compare neutralizing antibody assays and to decide which assay or combination of assays should be used for standardized assessments of vaccine-elicited neutralizing antibody responses. A major priority is to strengthen the standardization of the PBMC assay, given that it is the only assay that has been at least partially validated in passive antibody experiments in animal challenge models. Important decisions need to be made about the types of antibodies and assays that have greatest relevance to HIV-1 vaccines. Validation experiments in animals models are needed to determine the potential correlative value of new assay technologies that rely on the use of genetically engineered cells lines and Envpseudotyped viruses. Ideally, this would be done by employing several different assays to study the antibody response in a clinical efficacy trial in which the vaccine was at least partially protective. Because no such vaccine is currently available for HIV-1, studies in animal models are the next best choice. In this regard, two animal models are widely used for HIV vaccine development: simian immunodeficiency virus (SIV) and chimeric simian-human immunodeficiency virus (SHIV) infection in monkeys [48] . Quantitative passive transfer experiments in either model with antibodies that exhibit different effector functions could be used to address the biological relevance of in vitro assays. Unfortunately, very few SHIVs are currently available and, among these, most are derived from a single genetic subtype (clade B) and exhibit properties that may not be well suited to assay validation [49] . The creation of new and better SHIVs from non-clade B viruses would facilitate assay standardization as well as vaccine challenge models. This workshop identified several critical gaps in the current understanding of B cell regulatory pathways that impede a more rational development of an effective antibody-based HIV-1 vaccine. For example, broadly neutralizing antibodies in patient serum bind epitopes that are present on monomeric gp120 [25] , yet this is a poor immunogen for neutralizing antibody induction in vaccine recipients. Moreover, as mentioned above, viral epitopes for the known broadly neutralizing MAbs appear to be poorly immunogenic in infected individuals and as vaccine candidates. Insights into the immunoregulation of some of these latter epitopes (e.g., epitopes defined by MAbs 2F5 and 4E10) was provided by recent studies in which the MAbs were discovered to bind one or more self antigens [50, 51] , raising the possibility that these antibody specificities are subjected to negative regulation mechanisms, such as receptor editing or deletion. Thus, Env as an immunogen may bypass key steps in the B cell inductive pathway, or may actively induce negative production or downregulation of production of some broadly neutralizing antibodies [52] [53] [54] . The receptor-ligand interactions and intracellular signaling pathways that govern the production of antibodyproducing plasma cells and the persistence of plasma and memory B cells are poorly understood. Additional information on the mechanisms responsible for B cell migration, selection, and differentiation within and between specialized anatomical sites, particularly within lymphoid follicles, might be used to target suitable Env epitopes to appropriate B cell inductive pathways. An example would be to provide necessary signals to generate long-lived and high affinity memory in the marginal zone B cell compartment. Another example would be to discover ways to modify germinal center formation, positive and negative selection, and B cell differentiation to drive long-lived high affinity antibody responses against key epitopes that tend to be poorly immunogenic. In parallel to these efforts, genetic studies at the population level could provide critical information on the most promising paths to follow. In particular, the recent completion of the International HapMap Project now permits whole genome associated studies to be conducted with a minimum number of single nucleotide polymorphism tags [55, 56] . This powerful new technology could be used to identify genes that are associated with the wide variation in neutralizing antibody responses in HIV-1-infected individuals and in vaccine recipients. A critical question to ask is whether the potent neutralizing antibody response in a small subset of infected individuals is due to unique viral epitopes or to host genetic polymorphisms. Current evidence suggests that both might make a substantial contribution in the context of combined epitope and allelic representations [28, 47, 57] . To date most studies of the humoral responses in HIV infections have investigated immunoglobulins, the final product of B cell responses. Relatively few studies have examined B cell immunopathogenesis. A number of basic questions are still unanswered (e.g., extent and reason for perturbation of B cell subset changes, including memory B cells and plasma cells, in peripheral blood and tissues). Questions also remain about other potential functional contributions of B cells to HIV infections (e.g., role as antigen-presenting cells). In vivo studies should be performed in the nonhuman primate animal model to determine the emergence of pathologic events in the B cell compartment, in particular in lymphatic and gastrointestinal tissues of naïve and vaccinated animals that are challenged with pathogenic SIV or SHIV. These investigations should be done in parallel to detailed analyses of the magnitude and function of HIV-specific immunoglobulin responses determined in plasma and tissue secretions, and of HIV-specific B cells on a single cell basis. The establishment of a research consortium to study fundamental B cell biology as it relates to HIV-1 vaccines is recommended. This program should be structured in a way that asks key scientific questions about B cell regulatory pathways that modulate Env immunogenicity. Studies could address B cell receptor-ligand interactions and intracellular signaling pathways that govern the production of antibodyproducing plasma cells, the persistence of plasma and memory B cells, the mechanism of action of adjuvants, and host genetic associations with immune responses. immune antibody library displayed on yeast yields many novel antibodies compared to selection from the same library displayed on phage. Protein Eng Design

Host Gene Expression Profiling of Dengue Virus Infection in Cell Lines and Patients

BACKGROUND: Despite the seriousness of dengue-related disease, with an estimated 50–100 million cases of dengue fever and 250,000–500,000 cases of dengue hemorrhagic fever/dengue shock syndrome each year, a clear understanding of dengue pathogenesis remains elusive. Because of the lack of a disease model in animals and the complex immune interaction in dengue infection, the study of host response and immunopathogenesis is difficult. The development of genomics technology, microarray and high throughput quantitative PCR have allowed researchers to study gene expression changes on a much broader scale. We therefore used this approach to investigate the host response in dengue virus-infected cell lines and in patients developing dengue fever. METHODOLOGY/PRINCIPAL FINDINGS: Using microarray and high throughput quantitative PCR method to monitor the host response to dengue viral replication in cell line infection models and in dengue patient blood samples, we identified differentially expressed genes along three major pathways; NF-κB initiated immune responses, type I interferon (IFN) and the ubiquitin proteasome pathway. Among the most highly upregulated genes were the chemokines IP-10 and I-TAC, both ligands of the CXCR3 receptor. Increased expression of IP-10 and I-TAC in the peripheral blood of ten patients at the early onset of fever was confirmed by ELISA. A highly upregulated gene in the IFN pathway, viperin, was overexpressed in A549 cells resulting in a significant reduction in viral replication. The upregulation of genes in the ubiquitin-proteasome pathway prompted the testing of proteasome inhibitors MG-132 and ALLN, both of which reduced viral replication. CONCLUSION/SIGNIFICANCE: Unbiased gene expression analysis has identified new host genes associated with dengue infection, which we have validated in functional studies. We showed that some parts of the host response can be used as potential biomarkers for the disease while others can be used to control dengue viral replication, thus representing viable targets for drug therapy.

Although dengue-related disease results in an estimated 50-100 million cases of dengue fever and 250,000 to 500,000 cases of dengue hemorrhagic fever/dengue shock syndrome each year [1, 2] , a clear understanding of dengue pathogenesis remains elusive. Dengue virus is an enveloped, positive-stranded RNA virus of the Flaviviridae family transmitted by the mosquito Aedes aegypti and Aedes albopictus. Four serotypes of dengue virus (DENV1-4) circulate in endemic areas. Although infection with one serotype of dengue virus confers life-long protective immunity to that serotype, it does not protect the host from infection with other serotypes [3] . The initial target cells during dengue infection are believed to be Langerhans cells [4] . Through means not yet fully understood, Langerhans cells spread the virus, via the lymphatic system, to other tissues such as liver, spleen, kidney and blood, whereas monocytes, macrophages and endothelial cells are the major cell types in which the virus replicates [5] . In the majority of symptomatic dengue infections, a fever of 5-7 days duration develops together with bone and joint pain, retro-orbital pain, nausea and fatigue, this is called dengue fever (DF). While the majority of DF patients recover without intervention, 2-5% develop a more severe form of the disease, called dengue hemorrhagic fever/dengue shock syndrome (DHF/DSS), characterized by thrombocytopenia and vascular leakage, causing hypervolemic shock and death if not promptly treated [6] . The cause of DHF/DSS is not clear. Antibody dependent enhancement (ADE) is the most widely supported theory explaining the higher risk of DHF/DSS associated with a heterologous secondary infection [7] . The phenomenon of T cell ''original antigenic sin'' has also been described [8] . Like many viruses, dengue inhibits IFNa and IFNb signaling by suppressing Jak-Stat activation, resulting in reduced host antiviral response [9] . The combination of a reduced host defense, increased uptake of the virus and delayed viral clearance likely synergizes to produce higher viremia resulting in a more severe outcome. Because of the lack of a disease model in animals and the complex immune interaction in dengue infection, the study of host response and immunopathogenesis is difficult. The development of genomics technology has allowed researchers to study gene expression changes on a much larger scale. A summary of published microarray data from infection of different host cell types with bacteria, viruses, yeast, protozoa and helminthes revealed a common host-transcriptional-response consisting of a cluster of IFN-stimulated and immune mediating genes [10] . One particularly successful application of microarray technology was the identification of a novel drug target, c-kit, in endothelial cells infected with Kaposi's sarcoma-associated herpesvirus (KSHV) [11] . With regard to dengue infection, gene expression studies have been carried out in infected human umbilical vein endothelial cells (HUVECs) by differential display reverse transcription (DD-RTPCR) and Affymetrix oligonucleotide microarrays [12] . Genes having a role in the IFN antiviral response and immune defense such as 29-59 oligoadenylate synthetase (OAS), myxovirus protein A (MxA), TNFa, galectin-9, phospholipid scramblase 1 and human inhibitor of apoptosis-1 (IAP1) were shown to be upregulated upon infection [12] . Infection with dengue of a more transformed HUVEC-like cell line, ECV304, was analyzed by a different microarray system containing 7600 cDNA oligonucleotides [13] . In this study, the expression of 15 genes involved in cell cycle, apoptosis, membrane trafficking and cytoskeleton was found to be altered after infection [13] . Using a different gene expression approach, primary macrophages infected with a clinical isolate of dengue were analyzed by a cytokine array containing 375 human cytokine-related genes with approximately 20 genes observed to be either up-or down-regulated [14] . However, the functional importance of these changes, if any, was not studied [14] . We have used Compugen human 19K oligonucleotide arrays to study the host response to dengue infection, initially in a human hepatocytic cell line (HepG2). We identified the upregulation of three major clusters of genes; NF-kB-mediated cytokine/chemokine responses, type I IFN response, and the ubiquitin-proteasome system. Increased expression of selected genes, identified by Compugen array, was confirmed using a taqman low density array (TLDA) and results extended to include an additional, readily infectable, cell line (A549) and PBMC derived from adult dengue fever patients from a Singapore dengue prospective cohort study. Next we sought to confirm the functional effects of these upregulated genes. We confirmed the high levels of two novel chemokines, IP-10 and I-TAC, in dengue infection, in both cell lines and in dengue patients. We also determined that overexpression of viperin, an upregulated gene in the IFN pathway, and proteasome inhibition, with MG-132 or ALLN, reduced virus replication. In summary, we have used microarray analysis to identify new host genes associated with dengue infection, and to show that components of the host response can control dengue viral replication and therefore represent potential targets for drug therapy. Cell lines, A549, BHK-21, C6/36, HeLa, HepG2, HUV-EC-C, K562, SK-Hep1 and THP-1, were obtained from ATCC and maintained as instructed. The type 2 dengue virus strain TSV01 was obtained from a dengue outbreak in Townsville, Australia [15] (GenBank, accession number AY037116) and was propagated in the C6/36 cell line. Heat-inactivated virus was prepared by incubating virus samples in a 55uC water bath for 1 hr. Each cell line was infected with dengue virus at a multiplicity of infection (MOI) of 1 and 10 for 3, 6, 12, 24, 48, and 72 hrs before being evaluated for replication efficiency by plaque assay. Infection of cell lines was not continued beyond 72 hrs as after this point a significant degree of cell death became apparent. BHK-21 cells were cultured overnight in 24 well plates before media was removed and serial dilutions (10-fold) of virus culture supernatants added to individual wells. Plates were incubated for 1 hr before media was aspirated and replaced with 0.5 ml of 0.8% methyl-cellulose medium (with 2% FBS). Plates were then incubated for 5 days before the media was removed and cells fixed in 4% formaldehyde for 20 min, rinsed in water, stained with crystal violet for 20 min then rinsed again. Plaques were counted manually and concentrations of plaque forming units per ml (pfu/ ml) in the cell culture supernatant calculated. For drug intervention studies, HepG2 cells were treated with compound, or DMSO, at the indicated concentration and time. While the culture supernatants were collected for plaque assay, cytotoxicity in HepG2 cells was monitored using fluorescein diacetate (FDA, Acros Organics, Belgium), 10 mg/ml added to cells for 25 min at room temperature, and measured using a fluorescent plate reader (485 nm excitation/535 nm emission). Dengue E-protein was assessed in suspension HepG2 cells by intracellular staining and fluorescence activated cell sorter (Becton Dickinson, USA) and an Alexa-647 conjugated (AlexaFluor conjugation kit, Invitrogen, USA) monoclonal antibody (4G2, ATCC). Cells were scraped and permeabilized using BD FACS Perm/Wash solution (Becton Dickinson, USA) and acquisition and analysis was performed using CellQuest software (Becton Dickinson, USA). Dengue is the most prevalent mosquito-born viral disease affecting humans, yet there is, at present, no drug treatment for the disease nor are there any validated host targets for therapeutic intervention. Using microarray technology to monitor the response of virtually every human gene, we aimed to identify the ways in which humans interact with dengue virus during infection in order to discover new therapeutic targets that could be exploited to control viral replication. From the activated genes, we identified three pathways common to in vitro and in vivo infection; the NF-kB initiated immune pathway, the type I interferon pathway, and the ubiquitin proteasome pathway. We next found that inhibiting the ubiquitin proteasome pathway, or activating the type I interferon pathway, resulted in significant inhibition of viral replication. However, inhibiting the NF-kB initiated immune pathway had no effect on viral replication. We suggest that drugs that target the ubiquitin proteasome pathway may prove effective at killing the dengue virus, and, if used therapeutically, improve clinical outcome in dengue disease. Virus quantification by real time RT-PCR RNA was extracted from infected cells using the RNeasy Mini Kit (Qiagen, Netherland). Dengue serogroup 2 virus TSV01 detection by TaqMan PCR was adapted to the ABI7900 real time instrument using previously published primers and probes [16] . Standard ABI conditions were used, incorporating primers at 900nM. Quantification was achieved by relating viral Ct value to the Ct value on a standard curve of a measured number of copies of a 750 bp section of the virus (forward: 59-AAAGATCAGTGG-CACTCGTTCC-39 and reverse: 59-GCAGGTCTAAGAAC-CATTGCCT-39), cloned into pCR2.1-TOPO (Invitrogen, USA). Human arrays of 19,800 60mer oligonucleotide probes (representing 18861 genes), designed by compugen and manufactured by Sigma-Genosys were used according to the standard protocol described for cDNA microarray [17] . Dye swap was performed for each sample, at every time point and a rigorous quality check was performed before an array was used for downstream analysis [18] . 60 mer oligonucleotide probes were spotted onto poly-L-lysinecoated microscope slides using GeneMachines OmniGrid Microarray Spotter (USA). For fluorescence labeling of target cDNAs, 20 mg of total RNA from universal human reference (Strategene, USA) and experiment samples (RNA extracted from infected cells using Qiagen RNeasy Mini Kit ) were reverse transcribed in the presence of Cy3-dUTP and Cy5-dUTP (Amersham Biosciences, UK) using the Superscript reverse transcription kit (Invitrogen, USA). Labeled cDNA were pooled, concentrated, re-suspended in DIG EasyHyb (Roche, Switzerland) buffer and hybridized overnight (14-16h) in the MAUI Hybridization chamber (BioMicro, USA). The arrays were scanned using a GenePix 4000B Scanner (Axon Instruments, USA) to generate Tiff images. The images were analyzed by GenePix Pro 4.0 software (Axon Instruments, USA) to measure Cy3 and Cy5 fluorescence signals intensity and format data for data base deposition. The array data then underwent lowess normalization [19] available in an R package aroma to remove channel specific biases (R Development Core Team). Differentially expressed genes were selected using a procedure known as Significance Analysis of Microarrays (SAM) [20] , described in brief below. The statistic used in SAM is given as where; the numerator is the group mean difference, s the standard error, and s 0 a regularizing constant. Setting s 0 = 0 will yield a t-statistic. This value, called the fudge constant, is found by removing the trend in d as a function of s in moving windows across the data to reduce false positive results. As the statistic is not t-distributed, significance is computed using a permutation test. Genes with a computed statistic larger than the threshold were considered significant. The false discovery rate (FDR) associated with the given threshold can also be calculated from the permutation data. Quantitative PCR (qPCR) by TaqMan Low Density Array (TLDA) RNA was extracted using RNase Easy kit (Qiagen, Netherland) from infected cells. For patients blood samples (see Patient Samples), 2.5 ml of blood was collected in PAXgene tubes, RNA was extracted using PAXgene Blood RNA Kit (PreAnalytiX, Qiagne, Netherland). RNA was subjected to DNase treatment using an RNase-Free DNase Set (Qiagen, Netherland). 100 ng of total RNA was reverse transcribed using the High-Capacity cDNA Archive Kit (ABI, USA) and processed for TaqMan Micro Fluidic Cards (3M Company, ABI, USA) according to manufacturers instructions, together with data analysis using SDS2.2 software (ABI, USA). Differentially expressed genes were detected as above, using SAM. We analyzed SAM gene lists using the Applied Biosystem online program PANTHER [21] (http://www. pantherdb.org/). Pathway, interaction and Gene Onotology analysis was performed using MetaCore, version 3.2.0 (GeneGo, Inc, USA). Subjects were enrolled from the Early DENgue (EDEN) study, a dengue investigation conducted at a primary healthcare clinic in Singapore, for which local ethical approval had been granted, and informed consent from patients obtained. Participation required a PCR diagnosis of fever with duration less than 3 days (for the details of the PCR diagnosis, see [22] ). The enrolled dengue cases were either of serotype 1 or 3. Ten dengue positive subjects (n = 10, 3 males) were selected to represent the most severe cases of dengue by the criteria of a platelet counts below 30 (x 10 3 /ml; range 8-30, mean 20.5). Their mean age was 43.2 years (range 24-67). The mean fever was 38.3uC (range 37.7-39.1) at the first visit (Time Point 1), with a mean duration of 43 hours (range 14-72) from the onset of the fever. The second visit (Time Point 2) occurred at 80 to 96 hours after the first visit, mean fever was 37.1uC (range 36.2-38.7). Dengue negative patients were enrolled with the same criteria (fever with no respiratory infection symptoms), but were PCR negative for dengue. Ten subjects were taken with (n = 10, 4 males) a mean age of 43.2 years old (range 24-67). Their mean fever was 38.4uC (range 37.6-39.4) with a mean duration 27 hrs (range 8-60) at the first visit (Time Point 1). The convalescence sample was collected at the third visit which was 3 to 4 weeks later (Time Point 3). At each time point, a 2.5 ml blood sample was collected in PAXgene tubes for RNA analysis. A 10 ml blood sample was collected, serum was separated within 5 hrs and stored at 280uC. Serum IP-10 and I-TAC concentrations were measured using ELISA kits from R&D Systems as per manufacturers instructions. A549 cells were transfected with an expression construct encoding viperin using lipofectamine 2000 (Invitrogen, USA). Cells were selected using 500 mg/ml G418 and screened for viperin expression by immunoblotting. Cells were cultured overnight in 6 well plates before IFNb (Glycoferon, Singapore), at a final concentration of 500 U/ml, was added to each well while control wells remained untreated. Twelve hours post IFNb treatment, cells were infected with dengue virus (TSV01; MOI 1) for 48 hrs and the plaque assay was used to determine virus production. While a number of different cell lines have been used as models for dengue infection it is not clear which of these represents the most appropriate model for the analysis of host response by microarray, which requires a high rate of infection. As such, we screened seven human cell lines for their ability to support replication of dengue virus. We used the clinical, dengue serotype 2 isolate TSV01 (Accession number: AY037116) strain for infection. Human cell lines were ranked by maximum plaque forming units (pfu)/ml titer produced with A549.HepG2.SK-Hep1.K562.HUV-EC-C.THP-1.HeLa (data not shown). The same results were obtained with the widely used NGC strain (data not shown). The highest yielding cell lines A549 (a lung carcinoma) and HepG2 (a hepatoma cell line) were used in further studies, with HepG2 as the primary focus because of evidence of liver injury in DHF/DSS and the detection of dengue antigens in hepatocytes in liver [23, 24] . Microarray identification of host responses to dengue virus replication using the HepG2 infection model Viral replication in HepG2 cells infected with dengue virus TSV01 for 3, 6, 12, 24, 48 and 72 hours, compared to heat inactivated virus, was determined by plaque assay of the virus released in the cell culture medium ( Figure 1A ), FACS analysis of infected cells labeled intracellularly with an Alexa 674-conjugated antibody against dengue E protein (4G2) ( Figure 1B ) and real-time PCR analysis of viral RNA ( Figure 1C ). All three methods showed that new viral replication began after 24 hours and reached a plateau at 72 hrs, with FACS analysis showing 28% of cells infected at this point. After 72 hrs a degree of cell death become apparent (data not shown) and the experiment was not continued. Analysis of microarrays, performed in duplicate (dye-swapped) on three biological replicates at each time point, comparing infectious with heat inactivated virus; using a SAM q value (false discovery threshold) of 25%, revealed no significantly differentially expressed genes at 3, 6, 12 or 24 hrs post infection. However, there were 24 transcripts identified at 48 hrs and 124 at 72 hrs (a total of 132 transcripts representing 124 genes; Table S1 ) that were differentially expressed. At both 48 hrs and 72 hrs, clustering of transcripts using PANTHER analysis identified the IFN-mediated immunity pathway as the most significant with a P value of 10 215 (at 72 hrs). Genes that are typically induced after type I IFN stimulation, including OAS1, OAS2, OAS3, OASL, STAT1, STAT2, MX1, IFIH and IFNb, featured prominently in this cluster ( Figure 2 and Table S1 ). Further analysis of the 124 genes suggested the involvement of NF-kB-mediated cytokine/chemokine responses (NFKBIB, NFKB1A, TNFA1P, CCL4, CCL5, IP-10 and I-TAC amongst others) and ubiquitin related genes (HERC5, HERC6, UBE2L6, USP15 and others) ( Figure 2 ). Although not completely overlapping, a core host response to pathogen involving the IFN response and the NF-kB-mediated immune defense response [10] was observed in our array results. Other genes that were significantly changed included those associated with cell signaling, lipid metabolism, cell cycle and vesicular transport (Table S1 ). We did not detect any significantly down regulated genes in our system. All the microarray data were deposited in a public database accessible at http://www.ncbi.nlm. nih.gov/projects/geo (accession number is GSE6048). With the evidence of upregulation of three major pathways (NF-kB, IFN and ubiquitin) by microarray, we decided to pursue these pathways with further confirmation and validation. Firstly, we chose 59 genes from the three major pathways that were upregulated in the HepG2 cell line as shown by microarray. Secondly we selected another 36 genes which had a functional association with these pathways or might otherwise be associated with dengue infection. A Taqman Low Density Array (LDA) was then constructed for these 95 genes in order to confirm their (Table S2) , indicating a high rate of validation of the genes detected by microarray. In order to exclude gene responses that were specific to the HepG2 cell line alone, we decided to test the expression level of these genes in the A549 cell line using the same LDA. In the A549 infection model, plaque assay revealed that peak viral production occurred earlier than in the HepG2 model (1.1610 5 61.4610 4 pfu/ ml at 48 hrs) but was still substantial at 72 hrs post-infection (7.3610 4 63.8610 4 pfu/ml). The quantitative PCR also revealed a higher number of differentially expressed genes, with 63 genes at the 48 hr time point and 82 genes at the 72 hr time point (Table S2 ). Seeking confirmation that these genes were relevant in a physiological setting, we sought to test these genes in dengue patients. Whole blood RNA samples were obtained from ten adult fever patients (age .21 years old) enrolled in the Singapore Early Dengue (EDEN) cohort study in 2005 [22] . Each was PCR diagnosed positive for either dengue serotype 1 or 3 and all showed typical dengue fever symptoms. At the second visit (,5 days after onset of fever, onset of fever considered day 1), their platelet count had dropped below 30 (x 10 3 /ml; range 8-30, mean 20.5), and they were admitted to hospital. When these patient samples were tested for expression of the 95 selected genes by LDA, 67 genes were shown to be differentially expressed comparing blood samples taken at acute fever stage (first visit, 1-3 days after onset of fever) to convalescence (third visit, 3-4 weeks after first visit) (Table S2) . After confirmation of the expression of certain genes in two different dengue-infected cell lines and in dengue patients, we Table S1 . Colour intensity is derived from mean relative expression fold changes of dye swap results in comparison to universal reference (Strategene USA), red for upregulated, green for down regulated, black for no change, and grey for missing data. Time course (3, 6, 12, 24, 48, selected those that were upregulated in at least one time point in HepG2 cells and in at least one time point in A549 cells and in dengue patient samples (comparing acute to convalescent blood samples). Fifty genes fulfilled these criteria (Table S2 ) and, we felt, represented common genes involved in dengue virus response, or virus replication, while excluding responses that were cell type specific. These 50 common genes were mapped by direct interactions using the MetaCore program which illustrated the close clustering and interconnectedness of a network of 29 of genes around NF-kB, TNF-a and STAT1. (Figure 3A) . The NF-kB gene alone was added to the network, despite not being from our common list, to illustrate the connections between those induced by it. For example, the upregulation of IP-10, I-TAC, VEGF, PAI1, B2M, TNFAIP3 and RIG-1 could all be linked to the activation of NF-kB, while NFKB1B and NFKB1A, two feedback control genes for NF-kB activation, were also upregulated reflecting the self containment of this activation ( Figure 3A ). The degree of upregulation varied between the genes in this pathway with I-TAC and IP-10 being the most highly up-regulated (average expression in the two cell types and patients) genes ( Figure 3B) . A number of genes clustered around STAT1 were associated with the IFN pathway, and examination of the common list revealed another four genes (viperin, IFI44, IFIH1, G1P3) related to the IFN pathway that were upregulated but not mapped by the MetaCore program ( Figure 3C ). Finally, a number of genes related to the ubiquitin-proteasome pathway also appeared on the MetaCore map (HERC5, USP18 and Hdm2) with several more (unmapped) genes appearing on the common list ( Figure 3D ) The two most highly upregulated common genes from the NF-kB pathway were IP-10 (or CXCL10/IFN-inducible protein 10) and I-TAC (or CXCL11/IFN-inducible I cell a chemoattractant) ( Figure 3B) . In order to determine if this up-regulation of gene transcription lead to translation and protein release, concentrations of IP-10 and I-TAC protein in cell culture supernatant following dengue infection were determined by ELISA. In both the A549 and HepG2 infection models, dengue infected cells produced moderate, but significant (compared to heat-inactivated virus treated cells), concentrations of IP-10 ( Figure 4A ) and I-TAC ( Figure 4B ) at 72 hrs, but not at earlier time points (data not shown). We next investigated if these, or other, NF-kB induced proteins were influencing dengue virus replication. Adding dexamethasone to the HepG2 infection model to inhibit NF-kB activation [25] prevented IP-10 and I-TAC production, but had no effect on viral replication (data not shown). These results suggest that NF-kB activation, and IP-10 and I-TAC production, do not have a direct effect on viral replication and are, rather, simply part of the immune response to infection. Serum concentrations of IP-10 and I-TAC in the ten dengue fever patients described above, together with ten fever patients who did not have dengue (viral PCR negative) were also determined by ELISA. High concentrations of IP-10 ( Figure 4C ) and I-TAC ( Figure 4D) were present in the serum of dengue fever patients. There was significantly more IP-10 in the serum of the patients during the first (1-2 days after fever onset) and second (4-5 days after fever onset) visits, compared to the convalescent serum (P = 10 215 and P = 10 211 , respectively) as well as to non-dengue fever patient serum (P = 10 29 and P = 10 27 , respectively). I-TAC level was also significantly higher in first visit dengue patient serum comparing to both the convalescent (P = 10 27 ) and non-dengue fever (P = 10 26 ) patients. I-TAC levels in dengue fever patients at the second visit were significantly lower than at first visit (Fig 4D) . We detected a large number of IFN response genes induced in both cell line infections and in dengue patients (see Figure 3C ). In our study, viperin was one of the most highly upregulated genes in the type I IFN response pathway. Viperin has previously been identified as an IFN-induced anti-viral protein in HCMV and HCV infection [26, 27] . In order to investigate the role of viperin in dengue infection, we used an established A549 cell line stably overexpressing viperin (Vip) [26] . Comparing to infection in wild type A549 cells (WT), viperin overexpressing cells were significantly resistant to viral replication, as shown by plaque assay two days after infection ( Figure 5A ), with and without pre-treatment with IFNb (+IFN; 500 U/ml). Although pre-treatment with IFNb had the greater anti-viral effect, viperin over expression alone resulted in a small, but significant, reduction in virus production both with (P = 0.038) and without (P = 0.0004) IFNb pretreatment. These results suggest that viperin is an functional component of the IFN-mediated response to dengue, and demonstrate, for the first time, that viperin could be part of the anti-dengue response. Ubiquitination, another pathway identified in our list of common genes, is a key component of the immune system, the conjugation of single, or multiple, ubiquitin molecules to a protein targets it for defined subcellular localization, or destruction in the proteasome [28] . Components of the ubiquitin-proteasome system have been shown to be required for the maturation and release of a number of viruses (see review [29] ). In order to investigate the role of ubiquitin-proteasome system in dengue infection, we introduced proteasome inhibitors, MG-132 and ALLN, to the HepG2 dengue infection model [30] . At lower concentrations, both MG-132 and ALLN significantly reduced virus replication in the cell lines by over 50% (Figure 5B ). An examination of the effects of these compounds on the integrity and viability of inhibitor-treated cells using fluorescein diacetate revealed no cytotoxicity at these concentrations ( Figure 5C ). Higher concentrations of the inhibitors had an even greater effect (.90% reduction in pfu/ml) but with a degree of cytotoxicity ( Figures 5B, 5C ). Despite the fact that dengue is a major disease affecting the tropical world, little is known of its pathogenesis due, partly, to the lack of a suitable animal model and the complex cell interactions in infected individuals. Using microarray analysis of gene expression and high throughput quantitative real-time PCR, we investigated the host response to dengue infection in cell lines and in DF patients. We further validated our microarray results by functional study of the identified genes. Although dendritic cells, Langerhans cells and monocytes have been proposed as the principle reservoirs of viral infection [4] , it is not clear which other tissues, if any, are targeted. Immunohistochemistry and in-situ hybridisation studies of biopsies from DHF patients have indicated that a range of tissues are infected by the virus, including cells in the liver, spleen, lung, kidney and peripheral blood [5] . We chose the HepG2 cell line as the primary cell model for this study as it readily supported viral replication, and it is derived from the liver, which might be of some clinical relevance. The A549 cell line was included as it was also an excellent supporter of viral replication even though there is, presumably, little or no clinical relevance. The final clinical relevance came from the validation of the genes in 10 DF patients from the Singapore EDEN cohort. The 10 patients were selected based on low platelet count, because of the lack of WHO DHF/DSS manifestation in Singapore adult patients. Despite the limited sample number, and the wide range of collection times, variation between individuals at the same time point was ruled out by statistical analysis. We believe that with more patients, more genes that are differentially expressed would be identified. The overlap of upregulated genes, determined by quantitative PCR, in the two cell systems and in patients removed any responses that were unique to the cell type. In fact, we observed cell type specific gene changes such as IL-8, PAI-1 and RANTES that were upregulated in the cell lines but not the patient samples while the anti-inflammatory cytokine IL-10 was upregulated in the patient samples but not the cell lines, indicative of the effects of multiple cell types in an in vivo system. Similarly, IFNb was upregulated in the cell lines while IFNc was in patient samples. However, it is the large overlap between the in vitro infection and patient samples that warrants most attention. The two most highly upregulated chemokines were IP-10 and I-TAC. IP-10 and I-TAC are both ligands for the CXCR3 chemokine receptor and the production of these chemokines leads to the recruitment of CXCR3 expressing T cells and NK cells [31, 32] . Increased IP-10 and I-TAC expression has been seen in various viral infections, especially viral meningitis [33] . In SARS patients, elevated IP-10 early in infection was shown to be a predictor for a more severe outcome [34] . Neuronal IP-10 was shown to be involved in the recruitment of T cells in West Nile virus encephalitis [35] . Elevated I-TAC mRNA and protein has also been found in the liver of chronic Hepatitis C patients [36] . In dengue infection, various chemokines have previously been found to be induced in dengue patients, including IL-8, MIP-1a, MIP-1b, RANTES and MCP-1, but IP-10 and I-TAC have not been examined (see review [37] ). In a dengue intracerebral mouse infection models, CXCR3 2/2 and IP10 2/2 mice both had higher mortality rate than wild type mice after infection, indicating that IP-10 and CXCR3 receptors are part of the host defense mechanism, most likely in recruitment of T cells to the infection site [38] . IP-10 was also proposed to compete with virus binding on the receptor in vitro [39] . However, the relevance of this study to human patients is not clear. Although IFNc, a/b and NF-kB could all induce the production of IP-10 and I-TAC, prevention of IP-10 and I-TAC via NF-kB inhibition clearly reduced the IP-10 level in HepG2 cells. Furthermore, IFNb and IFNc were not consistently upregulated in the two cell lines or in DF patients in our study (data not shown). Therefore, we believe that NF-kB activation rather than IFN induction played the major role in elevated IP-10 and I-TAC in dengue infection. The inhibition of NF-kB, and consequent IP-10 and I-TAC inhibition, had no demonstrable effect on dengue replication in cell lines, suggesting the more complex role for these chemokines in a multi-cellular in vivo setting. Concentrations of each chemokine were significantly higher during the early stages of dengue fever, while I-TAC levels were reduced at the second visit, IP-10 levels remained high. The persistently high level of IP-10 might also contribute to the immunopathogenesis of dengue although this would require further investigation. It may be that concentrations of IP-10 and I-TAC, in combination with a viral antigen, could be used as an early marker for dengue fever. It is interesting to note that the level of IP-10 and I-TAC were independent of previous history of dengue infection, as half of the ten tested patients were suffering from a secondary infection. The common clinical feature of the ten selected patients was the low platelet count at second visit. Because of the lack of severe dengue cases (by WHO definition) in Singapore, the low platelet count was used as the measure of severity by which the ten patients from the cohort study were classified as having more severe dengue fever. We cannot at this point make a link between the level of IP-10, I-TAC and disease severity because of the small number of cases used in this study, a more detailed clinical study, currently underway, aims to determine if the concentrations of IP-10 and I-TAC during early stages of infection are linked to the progression to more severe forms of disease. The ability of IFN pre-treatment to inhibit subsequent dengue replication has been previously reported [9, 40] , as has the importance of IFN in the anti-viral response [41] . Our results indicated that viperin was one of the most highly upregulated genes in all models following dengue infection. Viperin encodes for an IFN-inducible antiviral protein shown to be associated with the endoplasmic reticulum and redistributed to the Golgi apparatus and cytoplasmic vacuoles following human cytomegalovirus (HCMV) infection [26] . The induction of viperin has been suggested to be anti-viral in both HCMV [26] and hepatitis C (HCV) [27] infections, but this is its first association with dengue infection. The use of viperin overexpressing A549 cells demonstrated that, in addition to being significantly upregulated during infection, viperin is directly involved in the anti-viral response to dengue, as shown by the suppression of dengue replication in the presence of increased expression of viperin. This effect was not as significant as the effect of IFNb alone suggesting that viperin is only a part of the IFN-mediated response to dengue. The molecular mechanism of the action of viperin to counter dengue infection will need to be further studied. The ubiquitin-proteasome system is the cellular machinery involved in the conjugation of single or multiple ubiquitin molecules to direct protein trafficking or degradation (reviewed in [28] ). In HCV infection, E6AP ubiquitin ligase was reported to mediate the ubiquitination and degradation of the virus core protein [42] . HCV polymerase has also been shown to interact with a ubiquitin-like protein leading to its degradation [43] . In dengue, ubiquitin-proteasome genes have not been reported to be involved in the dengue virus life cycle although a recent publication has shown that dengue envelope protein interacts with SUMO-1 conjugating enzyme 9 (Ubc9) and that overexpression of Ubc9 reduces virus production in a cell line [44] . Our TLDA results indicated that there was significant upregulation of a number of ubiquitin-proteasome system related genes during dengue virus replication but it was unclear if this response was anti-viral or if the dengue virus utilised components of the ubiquitin-proteasome system for replication. Use of proteasome inhibitors, MG-132 and ALLN, significantly reduced the release of dengue virus following infection of the HepG2 cell line. Previously, it has been shown that the effect of proteasome inhibitors may be virus specific with these compounds less (or non-) effective against virus such as influenza [45] and equine infectious anemia virus [46] . The effect of proteasome inhibition on viral replication appears to be mediated via a number of different processes including the involvement of ubiquitin or the proteasome in virus assembly [45] , budding [47] and release and maturation [48] . Further studies may elucidate the exact role the ubiquitin-proteasome plays in dengue virus replication which may involve trafficking of the virus to the plasma membrane or the maturation and fusion of the virus upon release [49] . It is instructive to compare our results with those of a published microarray study investigating host responses in patient blood [50] . That study identified genes whose protein products are expressed in the ER to be the most significantly enriched, which directly overlaps with our identification of the ubiquitin pathway; specifically including UBE2 and the PSMB genes. This may represent a process fundamental to viral replication in any system. In addition, Simmons et al. described host responses associated with dengue shock syndrome that clearly overlap with our interferon related gene list. In particular, G1P genes, OAS genes, IFI genes and Mx genes were found in both studies, suggesting that the amount of viral replication may be directly related to clinical outcome. Gene arrays of a number of cell systems during dengue infection revealed many genes, and host response pathways, that were upregulated during dengue infection. Further functional analyses distinguished between host response pathways involved in initiating an innate signaling response (NF-kB mediated genes and IFN pathway) and those involved in virus replication (ubiquitin-proteasome system). Specific components of the response to virus, such as viperin and IP-10 and I-TAC have been implicated in dengue infection for the first time. Further investigation of these components, together with the precise role of the ubiquitin-proteasome system in virus replication, may lead to drug targets for dengue. The use of gene array in multiple cell systems to investigate genes involved in virus replication, used in concert with functional studies, has proved to be a valid approach for discovery of novel markers and genes for understanding the host response to dengue and, ultimately, therapeutics against dengue. Table S1 HepG2 Microarray Gene List. List of HepG2 transcripts (n = 132), identified as differentially expressed by SAM analysis, in response to dengue virus TSV01 compared to heat inactivated virus 48 and 72 hours post infection. Transcripts are placed in groups according to biological processes. qV is the SAM calculated q value for each gene (following SAM significance selection based on a Delta value calculated from the variance between the sample sets, see Material and Methods). F.C. indicates fold change. ''-'' represents no significant change. Genes selected for real-time PCR validation through a TaqMan low density array (TLDA) platform are indicated by a tick. Found at: doi:10.1371/journal.pntd.0000086.s001 (0.28 MB DOC) Table S2 Quantitative PCR by Taqman based low density array in HepG2, A549 and Singapore dengue fever patients. Fold increase in gene expression as determined by quantitative PCR. For HepG2 and A549, the fold change was calculated based on dengue virus infection over heat-inactivated virus infection. For dengue fever patients (Patients), the fold change was calculated using patient blood samples collected at the first visit (,1-2 days after onset of fever) over the convalescence (3-4 weeks after the acute fever). Upregulation is shown in black and down-regulation in red. ''1.0'' represents no significant change (significance determined by q value ,5, see Material and Method). Genes that were significantly up-regulated in at least one time point in HepG2 and in at least one time point in A549 and in Patients are indicated with P-values, calculated by standard student T test and selected based on a cut off at P,0.05. Found at: doi:10.1371/journal.pntd.0000086.s002 (0.19 MB DOC)

Species-specific evolution of immune receptor tyrosine based activation motif-containing CEACAM1-related immune receptors in the dog

BACKGROUND: Although the impact of pathogens on the evolution of the mammalian immune system is still under debate, proteins, which both regulate immune responses and serve as cellular receptors for pathogens should be at the forefront of pathogen-driven host evolution. The CEA (carcinoembryonic antigen) gene family codes for such proteins and indeed shows tremendous species-specific variation between human and rodents. Since little is known about the CEA gene family in other lineages of placental mammals, we expected to gain new insights into the evolution of the rapidly diverging CEA family by analyzing the CEA family of the dog. RESULTS: Here we describe the complete CEA gene family in the dog. We found that the gene coding for the ITIM-bearing immunoregulatory molecule CEACAM1 gave rise to a recent expansion of the canine CEA gene family by gene duplication, similar to that previously found in humans and mice. However, while the murine and human CEACAMs (carcinoembryonic antigen-related cell adhesion molecules) are predominantly secreted and GPI-anchored, respectively, in the dog, most of the CEACAMs represent ITAM-bearing transmembrane proteins. One of these proteins, CEACAM28, exhibits nearly complete sequence identity with the ligand-binding N domain of CEACAM1, but antagonizing signaling motifs in the cytoplasmic tail. Comparison of nonsynonymous and synonymous substitutions indicates that the CEACAM28 N domain is under the strongest purifying selection of all canine CEACAM1-related CEACAMs. In addition, CEACAM28 shows a similar expression pattern in resting immune cells and tissues as CEACAM1. However, upon activation CEACAM28 mRNA and CEACAM1 mRNA are differentially regulated. CONCLUSION: Thus, CEACAM1 and CEACAM28 are the first paired immune receptors identified within the CEA gene family, which are expressed on T cells and are most likely involved in the fine-tuning of T cell responses. The direction of gene conversion accompanied by purifying selection and expression in immune cells suggests the possibility that CEACAM28 evolved in response to selective pressure imposed by species-specific pathogens.

The evolution of immunoglobulin superfamily (IgSF) members is largely influenced by the nature of their ligands. The extracellular part of IgSF members which are predominantly expressed in the nervous system are well conserved, while the extracellular domains of members expressed by immune cells or cells involved in reproduction tend to diversify much more rapidly [1] . The CEA gene family, which belongs to the IgSF, is a tandemly clustered multigene family. Such gene families are subject to rapid evolution due to ongoing gene duplication, deletion and mutational events [2] . At present, the CEA family is subdivided into two main subgroups: the CEA-related cell adhesion molecule (CEACAM) and the pregnancy-specific glycoprotein (PSG) subgroups. The CEACAM subgroup in humans consists of 12 members composed of a single immunoglobulin variable (IgV)-like N-terminal (N) domain followed by zero to six Ig constant (IgC)-like domains of A and B subtypes and one member which consists of two IgC-like domains and two IgV-like domains, one at each end of the molecule. CEACAM1, CEACAM3, CEACAM4, CEACAM18, CEACAM19, CEACAM20 and CEACAM21 are transmembrane molecules while CEA/ CEACAM5, CEACAM6, CEACAM7, and CEACAM8 are linked to the cell membrane via glycosylphosphatidylinositol (GPI) anchors and CEACAM16 is most likely a secreted molecule. For CEACAM1 (also known as CD66a, BGP and C-CAM105) and the recently discovered CEACAM16, CEACAM18, CEACAM19 and CEACAM20, but not for other family members have orthologs been identified in rodents [3] . In addition to these orthologs, the transmembrane-bound CEACAM2 (in mice), CEACAM17 (in mice and rats) and the secreted members CEACAM9, CEACAM10, CEACAM11 CEACAM12, CEACAM13, CEACAM14, CEACAM15 (in mice and rats) exist in rodents [3, 4] . In mice, rats and cattle, at least two CEACAM1 alleles each have been identified, which differ considerably in their N domain sequences [5] . No such allelic variation has been observed in humans. The PSG, which are specifically expressed in the trophoblast, are the most abundant fetal proteins in the maternal bloodstream during late human pregnancy. The PSGs are thought to play a pivotal role in the regulation of the maternal immune response to the fetal semi-allograft [6] . The diversity of the CEACAM/PSG family of proteins is further enhanced by species-specific differential splicing of several family members [7] . As expected from the variable structure of the CEA family members, these molecules have diverse functions. It is well established that various CEA family members play crucial roles in cell-cell adhesion, tumor development, angiogenesis, insulin metabolism, reproduction and, as more recently realized, in immunity [8] . CEACAM1 functions as a natural killer cell inhibitory receptor [9, 10] , regulates T and B cell proliferation [11, 12] , induces dendritic cell maturation [13] and facilitates granulocyte and monocyte survival [14, 15] . In addition, CEACAM3, CEACAM6 and CEACAM8 were found to be pivotal for the regulation of granulocyte activation in humans [16] . While human CEACAM1 contains two immunoreceptor tyrosine-based inhibitory motifs (ITIM) in its cytoplasmic domain, one CEACAM1 ITIM is replaced by an immunoreceptor tyrosine-based switch motif (ITSM) in mice and rats. Human CEACAM3 and possibly CEACAM4, CEACAM19 and CEACAM20 harbor functional immunoreceptor tyrosine-based activation motifs (ITAM) in their cytoplasmic tails [17, 18] . In contrast, except for CEACAM19 and CEACAM20 no such ITAM-containing CEA family members exist in rodents. Remarkably, several pathogens namely, Neisseria sp., Haemophilus influenzae, and Moraxella catarrhalis in humans [19] [20] [21] and mouse hepatitis virus (MHV) [22] use CEACAMs to anchor themselves to or to invade host cells. In addition, it has been suggested that bacterial pathogens can down-regulate immune functions by binding to CEACAM1 on CD4+ T lymphocytes and signaling via the CEACAM1 ITIM motif [23] . It was recently proposed that the human granulocyte-specific CEACAM3 functions as a specifically adapted single-chain ITAM-dependent phagocytic receptor involved in the clearance of CEACAM-binding bacteria by human granulocytes [24] . This indicates that fixation of genes in the human genome coding for ITAM-bearing CEACAMs can be the result of a selective pressure mediated by CEACAM-binding pathogenic microbes. In contrast, one of the two allelic variants of murine CEACAM1 exhibits a significant lower virus receptor activity suggesting that, in mice, mutations in the CEACAM1 gene rather than the fixation of an activating virus receptor were acquired to evade pathogen attack. The availability of a high-quality draft genome sequence of the domestic dog (Canis familiaris) [25] provided the possibility to analyze the complete CEA gene family in a third mammalian order. Here we show that in the dog multiple ITAM-bearing CEACAM1-related CEACAMs evolved and that CEACAM28 and CEACAM1 have undergone concerted evolution by gene conversion. Based on sequence analyses and comparison of expression profiles we suggest that CEACAM28 and CEACAM1 are paired activating and inhibitory immune receptors involved in the regulation of T cell responses. The possibility that sporadic recurrence of selective pressure mediated by species-specific pathogens was the driving force for the evolution of multiple ITAM-bearing CEACAM1-related CEACAMs in the dog is discussed. In an effort to identify all CEA gene family members in the dog genome we employed the BLAST program to search the GenBank and Ensemble databases (CanFam 1.0, July 2004) using cDNA nucleotide sequences from all known human and mouse CEA family members as query sequences. The human and mouse CEA gene families are located within the expanded leukocyte receptor complex on chromosome 19q13 ( Figure 1A ) and chromosome 7, respectively. This approach yielded the sequences of the CEACAM1, CEACAM16, CEACAM18, CEACAM19 and CEACAM20 orthologs. Their predicted full length mRNA sequences (Additional file 1) encode proteins with domain organizations identical with their human and murine counterparts: CEACAM16 (N1-A-B-N2); CEACAM18 (N-A-B-TM-Cyt); CEACAM19 (N-TM-Cyt); CEACAM20 (truncated N-A1-B1-A2-B2-TM-Cyt) [3] . Like their mouse and human counterparts, both CEACAM19 and CEACAM20 contain presumed ITAM consensus sequences in their cytoplasmic domains. However, like in mouse, no CEACAM19 cytoplasmic domain exon 1 present in human CEACAM19 was found [3] . As for human CEACAM18, no cytoplasmic domain exon could be identified due to its small size and lack of corresponding EST sequences. In a second round of search we used the BLAST program and cDNA nucleotide sequences from the identified orthologous canine genes to screen the Gen-Bank and Ensemble databases. This strategy was employed in order to identify dog-specific genes which have evolved after the separation of the dog and primate/ rodent lineage. This approach led to the identification of 7 additional CEACAM genes (CEACAM23-CEACAM29) using CEACAM1 nucleotide sequences as query sequences. CEACAM23, CEACAM24, CEACAM25, CEACAM29, CEACAM28 and CEACAM1 are arranged between XRCC1 and LIPE, while CEACAM26 and CEACAM27 are located between CD79A and TGFβ1 on the opposite strand ( Figure 1) . Surprisingly, no genes corresponding to the PSG genes were found. The murine Psg genes are located between Hif3a and Pglyrp1 spanning nearly 1.8 Mbp. In dog, however, the distance between HIF3A and PGLYRP1 comprises only 0.1 Mbp which rules out the presence of a larger set of genes between these marker genes. Similarly, there is not enough space between XRCC1 and CEACAM23 (0.2 Mbp) to accommodate an expanded PSG gene family as found in humans at the corresponding chromosomal locations location where the distance is approximately 1 Mbp ( Figure 1B) . The canine CEACAM gene loci on dog chromosome 1 exhibit a similar arrangement to that found for the human CEA gene family on chromosome 19 but it is partially different from that of the murine CEA gene family on chromosome 7 ( Figure 1B) . In order to identify the very short cytoplasmic exons of the CEACAM1-like genes which cannot be localized by direct search strategies we first localized the transmembrane exons of all CEACAM genes using transmembrane exon sequences of the human and mouse CEACAM genes as query sequences. Surprisingly, the 6 predicted transmembrane domain exons of the CEACAM1-related canine genes clustered with the sequences of the transmembrane domain exons of human CEACAM3, CEACAM4 and CEACAM21 which are connected to cytoplasmic domain exons encoding a C-terminal ITAM motif (CEACAM3 and CEACAM4) or a homologous truncated cytoplasmic domain with a premature stop (CEACAM21). The sequence of the transmembrane exon of canine CEACAM1 clustered with transmembrane sequences from murine Ceacam1 and Ceacam2 and with a group of exons coding for hydrophobic GPI signal peptides from human CEACAM1, CEACAM5, CEACAM6, CEACAM7, CEACAM8 (Figure 2A and data not shown). CEACAM5-CEACAM8 GPI signal peptide exons are thought to have evolved from the transmembrane exon of an ancestral CEACAM1 by introduction of a stop codon [26] . CEACAM1/Ceacam1 and Ceacam2 encode cytoplasmic domains which contain ITIM motifs. Taken together, this indicates that TM and cytoplasmic domain exons are exchanged together during evolution and predicts the presence of ITAM signaling motifs rather than ITIM in the cytoplasmic tails of the dog CEACAM23-CEACAM25, CEACAM28 and CEACAM29 molecules. Subsequently, we screened about 2,000 bp of nucleotide sequence downstream of the predicted canine transmembrane domain exons for the presence of the short cytoplasmic exons. Indeed, small cytoplasmic domain exons similar to the ones found in human CEACAM3 and CEACAM4 could be identified, which encode ITAM motifs close to the C-terminal end ( Figure 2B ). Although homologous cytoplasmic domain exons 1-3 are present in CEACAM29 the loss of the splice donor site in the cytoplasmic domain exon 3 leads to read through into the adjacent intron and truncation by a stop codon following after five codons. CEACAM26 and CEACAM27 probably represent pseudogenes due to multiple deletions in their extracellular (in particular N exons) and cytoplasmic domain exons (Figure 3) . To determine the degree of relatedness between the primordial canine CEACAM genes with the orthologous genes in mice and man, the nucleotide and encoded amino acid sequences of the leader, the IgV-type N domain, the IgC-type A and B domains, the transmembrane domain and the cytoplasmic domain exons were compared ( Table 1 ). The degree of sequence identity between corresponding extracellular Ig-type domains in dog, humans and mice varies greatly at the amino acid sequence level. The lowest value is observed for CEACAM1 N domains (56% and 42% for human and mouse, respectively), the highest for the CEACAM16 N2 domain (92 and 96% for human and mouse, respectively; Table 1 ). Based on the genomic sequence information we designed gene-specific primers for the amplification of full length CEACAM1-related cDNAs from total RNA, isolated either from liver or spleen. In the published genomic dog sequence, one nucleotide is missing from the A1 exon of canine CEACAM1. Cloning and sequencing of full length CEACAM1 clones from a mix-breed dog revealed a complete 279 bp CEACAM1 A1 exon. In total, we could iden- Expansion of ITAM-bearing CEACAM1-related CEA family members in dog No leader exon could be identified for CEACAM23 because of a sequence gap (indicated by brackets) in the publicly available genomic sequences. The genes are arranged in the order and orientation as found on dog chromosome 1. and 4. The cloned CEACAM25 splice variants also code for proteins with only one N domain followed by a transmembrane domain. Three out of four clones encode cytoplasmic domains which contain the predicted ITAM motif (GenBank accession no. EF137908). In one clone, the absence of the 53 nucleotide cytoplasmic domain exon 1 leads to a frame shift and the usage of an alternative stop codon located in cytoplasmic domain exon3 (GenBank accession no. EF137909). Using supposedly CEACAM28specific primers, two products were amplified which differ in their length by 276 bp. Cloning and sequencing revealed that the CEACAM28 gene codes for a protein with one N domain, two A domains, a transmembrane domain and a cytoplasmic tail which contain an ITAM motif (GenBank accession no. EF137910). The sequence of the cDNAs cloned from the smaller PCR product was different from all genomic sequences published either for the boxer or the poodle breed and, therefore, was named CEACAM30. The differences are more or less randomly distributed within the whole coding region which consists of an N, A, transmembrane and cytoplasmic domain region. Two CEACAM30 splice variants were identified, both containing an ITAM motif in the cytoplasmic tail. They differ, however, in the usage of alternative splice acceptor sites of the cytoplasmic domain exon 1 (Gen-Bank accession nos. EF137911 and EF137912) ( Figure 4 ). To analyze the evolutionary relationship of the newly discovered CEACAM genes, multiple nucleotide sequence alignments of Ig domain exons of human and dog CEACAM genes were performed and phylogenetic trees constructed. As expected, the N domain sequences of the orthologous genes (CEACAM16, CEACAM18-CEACAM20) clustered pairwise, while all new canine N domain exon sequences (CEACAM23-CEACAM30) clustered with that from dog CEACAM1 ( Figure 5A ). Similarly, the primate-specific CEACAM1-related N domain nucleotide sequences (CEACAM3-CEACAM8) clustered with the human CEACAM1 N domain exon sequence except for CEACAM21 N. The close relationship between the new dog CEACAM genes and dog CEACAM1 could also be demonstrated when the A and B domain exon nucleotide sequences were compared ( Figure 5B ). Nine of the 17 IgC-like domains of the new CEACAM1-related canine molecules are of the CEACAM1 A2 type and five each were of the CEACAM1 A1 and B type ( Figure 5B ).CEACAM23 appears to be the only CEACAM1-related gene which contains all extracellular domain exons present in CEACAM1, in particular three IgC-like domain exons (A1, A2 and B). All other members of this subgroup either lack one of these exons or contain exons corrupted by nonsense, frame shift or splice site mutations leading to smaller predicted proteins (Figure 3, 4) . Assuming that the N domain is the most relevant domain for ligand interaction, the ligand specificity of CEACAM28 seems to be most closely related to that of CEACAM1. The N domain amino acid sequence of CEACAM28 differs from that of CEACAM1 in only two positions. These amino acid changes do not affect the center of the CFG face, a topological N domain region of CEACAM members essential for ligand and pathogen interactions ( Figure 6A,B) . Interestingly, CEACAM28 lies next to CEACAM1 in the CEACAM1 gene cluster which hints to an evolutionary recent duplication event ( Figure 1B) . CEACAM25 is the most distant CEACAM1 relative within the group of CEACAM1-related genes, sharing only 66% of its N domain exon encoded amino acid sequence (Figure 5A, 6A; Table 2 ). Most of the amino acid sequence differences cluster to the CFG face of the N domains, indicating a positive selection for ligand diversification. To test this possibility, we calculated the ratio of nonsynonymous (Ka) to synonymous (Ks) nucleotide substitution rates on the branches of an N exon phylogenetic tree. The lowest Ka/ Ks ratio (0.1) is observed for CEACAM28. This indicates that the CEACAM28 N domain is under purifying selection. For all other canine CEACAM1-related CEACAMs, the N domain amino acid changes appear to evolve neu- (51) 56 (40) The trally or to be positively selected for during evolution since the corresponding Ka/Ks values are close to 1 or higher ( Figure 6C ). To identify the CEACAM gene expression pattern and splice variants, their nucleotide sequences were aligned, and regions of the transmembrane domain-encoding exons with sequence differences were selected to design primers that are specific for each of the CEACAM1-related cDNAs except for CEACAM28 and CEACAM30 which were amplified with a common primer pair ( Table 3 ). The primers were used to amplify cDNAs from individual CEACAM by RT-PCR from RNA isolated from liver, bone marrow, spleen, purified peripheral blood mononuclear cells (PBMC), and granulocytes ( Figure 7) . Except CEACAM29, all analyzed CEACAM1-related genes, (CEACAM1, CEACAM23, CEACAM24, CEACAM25, CEACAM28, and CEACAM30) were found to be expressed in spleen. CEACAM1, CEACAM24, CEACAM25, CEACAM28 and CEACAM30 transcripts were detected in bone marrow and CEACAM1, CEACAM24, CEACAM28 and CEACAM30 mRNAs in PBMC. Granulocytes express in addition to different CEACAM1 splice variants (preferentially with the long cytoplasmic tail) all the ITAM-bearing molecules except CEACAM23. In liver, the presence of mainly CEACAM1 and CEACAM23 transcripts could be demonstrated. CEACAM26 and CEACAM27 are not expected to be expressed because the sequences of essential exons appear to be corrupted (see above). To further analyze the expression of CEACAM1-related CEACAMs by immune cells, we isolated T cells from PBMC by MACS sorting using CD4-and CD8-specific mAbs. This procedure resulted in a purity of the T cell populations of >97%, as determined by flow cytometry (data not shown). Activation of T cells by the various stimuli was controlled by analysis of blast formation using flow cytometry ( Figure 8A) . RT-PCR analysis demonstrated that resting T cells express CEACAM1, CEACAM28 and CEACAM30 mRNA simultaneously. In addition, expression of these CEACAMs was also found in T cell-depleted PBMC, which are mainly composed of Bcells ( Figure 8B ). Stimulation of PBMC with anti-CD3 mAb or IL-2 had only a minor effect on CEACAM1 expression in T celldepleted PBMC (data not shown) and T cells. In contrast, the content of CEACAM28 mRNA seem to be reduced in T cells upon activation ( Figure 8B ). In unstimulated T cells the CEACAM1/CEACAM28 cDNA ratio was < 3 whereas the ratio was > 7 and > 11 in T cells stimulated with IL-2 and anti-CD3 mAb, respectively ( Figure 8C ). The content Domain organization of dog CEACAM1-related protein splice variants The amino acid (A; below the diagonal) and nucleotide sequences (N; above the diagonal) of IgV-like domain exons were compared pairwise using the ClustalW software (Weight Matrix BLOSOM). The degree of identity is indicated in %. Figure 8B ). The near identity of the N domains between CEACAM1 and CEACAM28 is most likely due to a recent gene conversion event. Indeed, we could identify a 2,332 bp genomic region in CEACAM1 which shows 99% identity with the corresponding region of CEACAM28. The high similarity of this segment starts at position -984 of the 5'flanking region and ends at position 114 of the intron between the Ndomain and the A1 domain exons ( Figure 9A ). Upstream of this segment, the sequence similarity between homologous regions of CEACAM28 and CEACAM1 (~1,200 bp) drops to 80%. The downstream homologous region of ~1,700 bp demarcated by a SINE element in CEACAM28 exhibits a similarity of 92%. The occurrence of such a regionally restricted gene conversion event is also supported by the high degree of amino acid sequence similarity of the leader (100%) and N domain (98%). The A1 and A2 domains of CEACAM28 encoded outside of the conserved genomic region share only 88% and 86% of their amino acid sequence with that of the homologous CEACAM1 A2 domain ( Figure 9B ). These data suggests that the similarity between CEACAM1 and Relationships of the N-domains of dog CEA gene family members the CEACAM28 ancestor was ~92% before the gene conversion event occurred. Recent molecular studies have established strong evidence for four primary, superordinal mammalian clades: Afrotheria, Xenarthra, Euarchontoglires and Laurasiatheria [27, 28] . The CEA gene family is well known in primates (humans) and rodents (mice and rats). However, both orders belong to the same principal lineage of placental mammals, the Euarchontoglires. Since carnivores belong to the Laurasiatheria clade, we expected to gain new insights into the evolution of the rapidly diverging CEA family by analysis of the CEA family of the dog. First of all, identification of CEACAM1, CEACAM16, CEACAM18, CEACAM19 and CEACAM20 orthologs in the dog genome at corresponding chromosomal locations, strongly indicates that the CEA gene family of the common ancestor of Euarchontoglires and Laurasiatheria consisted of at least five genes. It remains to be seen whether species which represent older mammalian lineages like Afrotheria, marsupials and monotremes contain the same set of primordial genes. Furthermore, the findings of this work supports our previous hypothesis that CEACAM3-CEACAM8, CEACAM21 and CEACAM9-CEACAM15, CEACAM17, none of which is present in the dog genome ( Figure 2A and data not shown), represent primate and rodent-specific genes, respectively [3] . The ancestral members of the CEA gene family are located in three different loci (Figure 1 ). While CEACAM16, CEACAM19 and CEACAM20 are close together (CEACAM16 and CEACAM19 are next to each other), CEACAM1 is separated from this locus by more than 1 Mbp and CEACAM18 is displaced into the SIGLEC gene cluster in humans, mice and dogs. Out of these ancestral genes, probably only CEACAM1 ancestors gave rise to massive gene expansions in a orderspecific manner [3] . The CEACAM1-related genes form compact clusters next to CEACAM1 and near the CD79A gene. In rodents even a third locus (around the Hif3a gene) has been encroached by amplified CEACAM1related genes. With analysis of the CEA family in a third order (carnivores), the inversion event observed for the chromosomal region flanked by Cd79a and Lipe can now be decided to have happened in the rodent rather than in the primate lineage, since the order of genes in this locus is conserved between humans and dogs ( Figure 1 ). CEACAM1-related genes can be subdivided into the PSG and the CEACAM subgroup genes. The birth of the ancestral PSG has been estimated to have occurred some 90 Myr ago [29] , approximately the time of rodent-primate divergence [30] . The rapid, independent expansion of the human and mouse PSG gene families occurred through further gene duplication and exon shuffling events [31] [32] [33] . Surprisingly, we did not find any PSG gene in the dog genome indicating that the PSG function exerted in primates and rodents is dispensable for the dog, possibly because dogs have an endotheliochorial placenta type which is different from the hemochorial placentae of primates and rodents [34] . Semi-allogeneic fetal trophoblast Expression pattern of dog CEACAM1-related genes Figure 7 Expression pattern of dog CEACAM1-related genes. CEACAM1-related transcripts were identified by RT-PCR using gene-specific primers which are located in the N domain and transmembrane exons. For the detection of CEACAM1 transcripts, primers in the N domain and cytoplasmic domain exon 3 were used. The products were separated by agarose gel electrophoresis in the presence of ethidium bromide and visualized by UV illumination. One-kb and 100-bp DNA fragment ladders were used as markers. The possible domain organization of the proteins encoded by the splice variants (number of Ig domains) is indicated in the right margin. Sequence determination of the CEACAM28 PCR products revealed simultaneous detection of CEACAM28 and CEACAM30 cDNAs (till then unknown). C, CEACAM. Expression of the paired receptors CEACAM1 and CEACAM28 in stimulated T cells forward scatter sideward scatter cells are less invasive and concomitantly less prone to encounter the maternal immune system in endotheliochorial placentae in comparison to the hemochorial placentae where the trophoblast cells are bathed in maternal blood [35] . Therefore, PSG could be involved in the regulation of trophoblast invasion and/or modulation of the maternal immune system as has been suggest before [3, 36] . Alternatively, the function of the PSG could have been taken over by an unrelated gene family due to functional convergence at the molecular level, similar to that found for natural killer (NK) cell receptors, where killer cell immunoglobulin-like receptors (KIR) and Ctype lectin-like receptors (Ly49) are used in primates and rodents, respectively. In contrast to the PSG, the CEACAM subgroup of canine CEACAM1-related genes underwent a recent expansion similar to that found in primates and rodents. In general, IgSF members which are expressed by leucocytes are diverging rapidly within their extracellular domains. This results in an average amino acid identity of about 54% between human and mouse orthologs [1] . The conservation of the CEACAM1 N domain, most relevant for pathogen and cell-cell interactions [37] , lies within this range, with an amino acid identity of 56% (dog versus human) and 42% (dog versus mouse). This is consistent with the notion that CEACAM1 plays a key role in various functions of the immune system [8] . In contrast, between mouse and human CEACAM16, CEACAM18 and CEACAM19 orthologs, the N domain amino acid identities are higher ranging from 60% to 91%, which argues for non-immunological functions, especially for CEACAM16, which exhibits an overall amino acid sequence identity of about 90% ( Table 1 ). The dog CEACAM1-related genes are arranged head to tail, like the KIR genes in the human KIR locus [38, 39] . Such an arrangement was found to facilitate unequal crossing over leading to gene expansion and generation of multigene families [40] . Furthermore, the KIR gene family is subject to birth-and-death evolution, domain shuffling and mutational changes [41] . Our analyses indicate that the CEA gene family evolved similarly. One characteristic for this kind of evolution is the presence of multiple pseudogenes, which also exist within the canine CEACAM1like gene locus (CEACAM26, CEACAM27). This type of evolution also facilitates a mechanism for qualitatively changing a receptor's signaling potential. A number of type I transmembrane IgSF members including CEACAMs and KIRs are composed of an amino-terminal extracellular "environmental sensor" and a carboxy-terminal cytoplasmic tail which can contain either inhibitory (ITIM) or activating motifs (ITAM or a positively charged amino acid in the transmembrane region permitting coupling with activating adaptors such as DAP12 [42] ). Unequal crossing-over at such gene loci encompassing the corresponding exons could result, in a single step, in genes encoding receptors with a drastically changed signaling potential by combining the same extracellular domain with a cytoplasmic tail of opposite function. This allows rapid adaptation to environmental changes and the creation of so-called paired immune receptors. An example for such a mechanism is found for the KIR gene KIR3DL/S, where either an inhibitory variant or an activating variant encoding KIR gene is found in different human individuals at the same locus [43] . In addition, a change in the receptor's signaling potential which may be the result of an arms race between pathogens and the host immune system has been reported recently for mice. The murine cytomegalovirus (MCMV) induces in infected cells expression of the virus-encoded GPI-anchored m157 protein, which binds to the activating, DAP12-associating killer receptor Ly49H in MCMV-resistant mice leading to NKcell-mediated cytotoxicity. MCMV-susceptible mouse CEACAM1 and CEACAM28 genomic regions involved in gene conversion Figure 9 CEACAM1 and CEACAM28 genomic regions involved in gene conversion. The 2332 bp genomic region of CEACAM1 which covers nearly 1 kb of the 5'-flanking region (5'-FR), the leader exon (L), intron 1 (IntI), the N domain exon (N) and part of intron 2 (Int II) is highly conserved in CEACAM28 (99%) and, therefore, probably participated in a recent gene conversion event. The homologous sequences upstream and downstream from the conserved region are less conserved (80% and 92%, respectively) (A). This can also be deduced from the degree of conservation of the amino acid sequences encoded by the leader, N and A domains which is much less outside of the gene conversion region. Note that CEACAM28 contains two A2 type domains (B). [44] . It was suggested that the activating receptor gene was derived from the gene encoding the inhibitory receptor to counteract a virus that exploits the inhibitory receptor during infection and that this mechanism may apply to many other activating receptors that have inhibitory counterparts [45, 46] . However, recently it was shown for the human SIGLEC5 and SIGLEC14 genes that the direction of gene conversion that generates paired immune receptors is not always the same. This indicates that also other needs drive the evolution of paired receptors like the fine tuning of immune responses [45] . Host-pathogen interactions involving both proteins with ITIM or ITAM signaling motifs have also been reported for humans. The primate-specific CEACAM1-related protein CEACAM3 which contains an N domain closely related to that of CEACAM1 (88% amino acid sequence identity) and a cytoplasmic tail with an ITAM is exclusively expressed in granulocytes. Due to its close relatedness with the ITIM-bearing gonococcal receptor CEACAM1 it can act as decoy receptor for Neisserial pathogens and is able to induce clearance of gonococcal infections by granulocytes. Mutational analyses demonstrated that the cytoplasmic ITAM is instrumental for this process [24] . To our knowledge, CEACAM1 and CEACAM28 represent the first paired coexpressed immune receptors which contain antagonizing ITIM and ITAM (and not just a transmembrane domain with a positively charged amino acid which serves as a docking site for ITAM-containing adaptor molecules, like DAP12) in their respective cytoplasmic tails. CEACAM28 was probably formed from a CEACAM29 ancestor by a recent gene conversion encompassing the leader and the N exon from CEACAM1. The latter exon encodes the region most instrumental for ligand and pathogen binding in CEACAM molecules. Several findings, presented here, argue for a role of CEACAM28 as a CEACAM1 decoy receptor in the dog. First, the direction of the gene conversion was most likely from the inhibitory receptor to the activating receptor, second, the Ka/Ks ratio of the N domain exons of these family members is indicative for a purifying selection, whereby the specificity of the putative CEACAM1 ligand is probably conserved for CEACAM28, third, the amino acid changes observed for the CEACAM28 N domain most likely have no effect on the structure of that part of the N domain shown to be the pathogen binding site in human and mouse CEACAM1 and fourth, is expressed by immune cells (T, probably B and/or NK cells and granulocytes), a prerequisite for a defense function against a putative pathogen. Based on a highly similar amino acid sequence and expression pattern CEACAM30 can be envisioned to also function, like CEACAM28, as a decoy receptor. However, this hypothesis raises questions about the driving force behind the fixation of the other ITAM-bearing CEACAM1-related genes in the dog. Most likely, selective pressure by pathogens changes with time. In times where the selective pressure exerted by pathogens on a receptor system is low, Darwinian selection will act and change the specificity of a receptor now possibly recognizing an endogenous ligand. Once the receptor has gained a new function it also can change the signaling properties which may explain why older i.e. more diverged ITAM-bearing receptors e.g. like CEACAM24 and CEACAM29 have lost the ITAM or have gained new splice variants which do not contain an ITAM. Taken together, the presence of multiple CEACAM1related dog CEACAMs with ITAMs could reflect a repeated arms race between species-specific pathogens and the dog immune system. Fixation and expression by effector cells of ITAM-containing CEACAMs would result from either bacterial pathogens binding to CEACAM1 or viruses which induce the expression of CEACAM1 ligands by infected cells as found for MCMV. This putatively pathogen-driven evolution of the CEA gene family led to the development of the CEACAM1/CEACAM28 receptor pair which could also be involved in the fine tuning of T cell responses. Different canine tissue samples were collected from healthy mix-bread dogs and flash-frozen in liquid nitrogen. Peripheral blood mononuclear cells (PBMCs) and granulocytes were isolated from blood of healthy mixbread or beagle dogs by density-gradient centrifugation through Ficoll-Paque (GE Healthcare, Freiburg, Germany). Granulocytes located on top of erythrocytes were harvested and remaining erythrocytes were lysed (ammonium chloride buffer). Stimulation of PBMC with PHA (2 µg/ml for 72 hr), CD3 (CA17-2A12; [47] ) (0,5 µg/ml for 96 hr) and human IL-2 (200 U/ml for 7 days) was performed at a concentration of 5 × 10 5 cells/ml in RPMI-1640 supplemented with 10% fetal calf serum (FCS "Gold"; PAA Laboratories, Coelbe, Germany), 2 mM L-glutamine, 100 U/ml penicillin, 100 µg/ml streptomycin, non-essential amino acids and 1 mM sodium pyruvate (GIBCO/Invitrogen, Karlsruhe, Germany). T cells were isolated from either stimulated or unstimulated PBMC by positive MACS sorting following the manufacturer's instructions (Miltenyi Biotec, Ber-gisch-Gladbach, Germany). Stimulated and unstimulated PBMC were incubated with mouse anti-canine CD4 (CA13-1) and rat anti-canine CD8 mAbs (Dog 10-8) [48] and subsequently with bead-coupled anti-mouse Fc mAb which also reacts with rat IgG. For surface staining, cells were suspended in PBS/0.3% (w/v) BSA supplemented with 0.1% (w/v) sodium azide. Cells were incubated with 0.5 µg/10 6 cells of the relevant mAb for 30 min at 4°C. Cells were washed twice and analyzed with a FACScan (BD Biosciences, Mountain View, CA). Dead cells were excluded by propidium iodide staining. The following reagents and mAbs from Serotec were used: FITC-conjugated anti-CD3, FITC-conjugated anti-CD4 and PE-conjugated anti-CD8α. Total RNA extraction was performed using either the TRIzol ® reagent (Invitrogen Life Technologies, Karlsruhe, Germany) or the RNeasy Kit (Qiagen, Hilden, Germany) according to the manufactures' protocols. One µg of total RNA was used for cDNA synthesis by reverse transcription (RT) in a total volume of 20 µl using either the Reverse Transcription System ® (Promega, Mannheim, Germany) or the First-Strand cDNA Synthesis ® Kit (MBI Fermentas, St. Leon-Rot, Germany). The RT product (1 µl) was amplified by polymerase chain reaction (PCR) with Taq polymerase (Qiagen). After an initial denaturation step at 95°C for 45 s, 35 PCR cycles (denaturation: 95°C, 30 s; annealing: 60°C, 1 min; extension: 72°C, 1.5 min) and a final extension step at 72°C for 15 min were performed. The primers were designed using the Primer3 software. Gene-specific primers were selected that exhibited > 2 mismatches at the 3'-end of the oligonucleotide to all other members of the canine CEACAM family, except for CEACAM28 and CEACAM30 (see below). The primers were validated by eye using multiple nucleotide sequence alignments of the aforementioned CEA family members. The primer sequence and the predicted sizes of the amplified products are summarized in Table 3 . Canine GAPDH cDNA was amplified using GAPDH.dogforward 5'-GCCAAAAGGGTCATCATCTC and GAPDH.dog reverse 5'-GCCCATCCACAGTCTTCT primers at a annealing temperature of 56°C and 30 cycles. Eight µl of each PCR were analyzed by electrophoresis on a 1.8 % agarose gel and visualized by ethidium bromide staining. The amount of cDNA was quantified by endpoint determination with the Quantity 1 ® software (Bio-Rad Laboratories, Munich, Germany). Based on published genomic sequences and comparison with human CEACAM1 the 5'-and the 3'-untranslated regions of canine CEACAM1 were predicted and used to design primers for amplification of full length cDNA (CEACAM1cf-5i and CEACAM1cf-3i; Table 3 ). These primers introduced HindIII and XbaI restriction sites at the 5'-and 3'-ends of the PCR products, respectively. PCR was performed with cDNA from dog liver using Pfu DNA polymerase according to the protocol supplied by the manufacturer. After electrophoretic separation, 6 distinct ethidium bromide-stained DNA fragments were excised from the agarose gel and purified using the Perfectprep ® Gel Cleanup Kit (Eppendorf, Hamburg, Germany). The full length cDNAs were digested with HindIII and XbaI and cloned into the pRc/CMV expression vector. Plasmid DNA isolated from 12 clones were analyzed by PCR and sequencing. Full length cDNAs of CEACAM23, CEACAM24, CEACAM25, CEACAM28 and CEACAM30 were cloned using the TOPO TA Cloning ® Kit (Invitrogen Life Technologies). PCR was performed with cDNA from dog spleen using Taq polymerase (Qiagen) and the CfCEACAM1-5i-f forward primer for all CEACAMs combined with gene-specific reverse primers from the 3'untranslated regions (UTR; Table 3 ). Sequence similarity searches were performed using the NCBI BLAST tools [49] and Ensembl BLAST/SSAHA search programs [50] . For the identification of CEACAM genes not yet annotated in the dog genome, sequences from human and mouse CEACAM and PSG cDNAs were run against the WGS databases or whole genome sequences. Multiple sequence alignments were performed and phylogenetic trees were constructed with Clus-talW [51] using default parameters. The three dimensional structures of CEACAM N domains were modeled using the Geno3D-release 2 software based on the published crystal structure of the murine CEACAM1 N-A2 fragment [37, 52] . Leader and transmembrane domains were identified with the aid of programs SignalP and TMHMM, respectively. The NetPhos 2.0 server was used to produce neural network predictions for serine, threonine and tyrosine phosphorylation sites in the cytoplasmic domains of canine CEACAMs [53] . Comparison of the number of nonsynonymous substitutions per nonsynonymous site (Ka) with the number of synonymous substitutions per synonymous site (Ks) was performed with the Ka/Ks Calculation Tool [54] . The substitution rate ratio Ka/Ks measures the molecular selective pressure. If Ka/Ks = 1, the amino acid changes are neutral and will be fixed at the same rate as silent mutations. If Ka/Ks < 1, the amino acid changes are deleterious and purifying selection will reduce the fixation rate. If Ka/ Ks > 1, the amino acid changes are evolutionarily advantageous and positive selection will increase the fixation rate.

Nucleolus: the fascinating nuclear body

Nucleoli are the prominent contrasted structures of the cell nucleus. In the nucleolus, ribosomal RNAs are synthesized, processed and assembled with ribosomal proteins. RNA polymerase I synthesizes the ribosomal RNAs and this activity is cell cycle regulated. The nucleolus reveals the functional organization of the nucleus in which the compartmentation of the different steps of ribosome biogenesis is observed whereas the nucleolar machineries are in permanent exchange with the nucleoplasm and other nuclear bodies. After mitosis, nucleolar assembly is a time and space regulated process controlled by the cell cycle. In addition, by generating a large volume in the nucleus with apparently no RNA polymerase II activity, the nucleolus creates a domain of retention/sequestration of molecules normally active outside the nucleolus. Viruses interact with the nucleolus and recruit nucleolar proteins to facilitate virus replication. The nucleolus is also a sensor of stress due to the redistribution of the ribosomal proteins in the nucleoplasm by nucleolus disruption. The nucleolus plays several crucial functions in the nucleus: in addition to its function as ribosome factory of the cells it is a multifunctional nuclear domain, and nucleolar activity is linked with several pathologies. Perspectives on the evolution of this research area are proposed.

Brief history of the nucleolus An ovoid body visible in the nucleus was probably the Wrst observation of the nucleolus more than two centuries ago by F. Fontana. Since that time, the nucleolus has been the object of intense investigation and interestingly our vision of the nucleolus has evolved with technical progress. During the nineteenth century, using light microscopy, numerous cytologists described the variability of nucleolar morphology with great precision (Montgomery 1898) . In 1934, McClintock proposed that the "nucleolus is organized in the telophase through the activity of ... the nucleolar-organizing body" (McClintock 1934) . Since the nucleolar-organizing body corresponds to a speciWc region of chromosome 6 in Zea mays, this was the Wrst time the nucleolus was related to gene activity. In the 1950's the presence of RNAs in the nucleolus was demonstrated, and in the 1960's in situ hybridization techniques made it possible to identify ribosomal genes (rDNAs) in the nucleolar organizer region (NOR) (Caspersson 1950; Perry 1962; Ritossa and Spiegelman 1965) . During the same period, mass isolation of nucleoli became possible leading to the biochemical characterization of nucleolar components. Based on these results it was proposed that ribosome biogenesis occurs in nucleoli. Given that the nucleolus became a subject of great interest, the "International symposium on the nucleolus-its structure and function" was organized in Montevideo in 1965 and the contributions published in Natl Cancer Inst Monogr no 23 (USA) in 1966. Since 1969, at the initiative of W. Bernhard and H. Busch, "Nucleolar Workshops" on nucleolar organization, the biochemistry of nucleolar proteins, rRNA processing as well as variability in cancer cells were regularly organized. Several books on nucleoli were published; among them, the famous "The nucleolus and ribosome biogenesis" is still a very useful source of information (Hadjiolov 1985) . Between 1980 and 2000, the functional organization of the nucleolus was deciphered in large part due to the improvement of labeling by the electron microscopy (EM). Recently a new Weld of investigation was opened when molecules not involved in ribosome biogenesis were detected in the nucleolus (Carmo-Fonseca et al. 2000; Pederson 1998; Politz et al. 2002; Visintin and Amon 2000) . In accordance with these nucleolar localizations, nucleolar mass spectrometry analyses identiWed »700 nucleolar proteins, some of them not related to ribosome biogenesis (Andersen et al. 2005) . The area of plurifunctional nucleolus was opened. Consequently "The nucleolus", a book that presents the state of the art was published (Olson 2004) as well as several reviews on the multiple functions of the nucleolus (Boisvert et al. 2007; Hernandez-Verdun 2006; Hiscox 2007; Raska et al. 2006 ). General information "The nucleolus: an organelle formed by the act of building a ribosome" (Mélèse and Xue 1995) reveals by its size and organization the eYciency of ribosome biogenesis. For example the nucleolus is a prominent nuclear structure in cycling cells but of limited size in the terminal stages of diVerentiation such as in lymphocytes or chick erythrocytes. If ribosome biogenesis is blocked, reorganization of the nucleolar components is visible in segregated nucleoli. In mammalian cells, the nucleolus is disorganized in prophase and reassembled at the end of mitosis using the nucleolar machineries from the previous cell cycle. On the contrary, in yeast the nucleolus is present and active throughout the cell cycle even though condensation of the rDNAs is necessary for transmission of the nucleolus in anaphase (D'Amours et al. 2004; Sullivan et al. 2004; Torres-Rosell et al. 2004) . The nucleolus is the ribosome factory of the cell. In the nucleolus rDNAs are transcribed, the 47S precursor ribosomal RNAs (pre-rRNAs) are cleaved, processed and assembled with the 80 ribosomal proteins and the 5S RNA to form the 40S and 60S ribosomal subunits (selected reviews Gébrane-Younès et al. 2005; Hernandez-Verdun and Junéra 1995; Scheer et al. 1993; Scheer and Hock 1999; Shaw and Jordan 1995; Thiry and Goessens 1996) . This complex series of maturation and processing events, presently better characterized in yeast than in higher eukaryotes is under the control of about 150 small nucleolar RNAs (snoRNAs) and 2 large RNP complexes: (1) the small subunit (SSU) processome containing the U3 snoR-NAs and 40 proteins or Utps (U three proteins) required for the 40S ribosomal subunit, and (2) the large subunit (LSU) processome required for the 60S ribosomal subunit (de la Cruz et al. 2004; Fatica and Tollervey 2002; Fromont-Racine et al. 2003; Sollner-Webb et al. 1996; Tollervey 1996) . The snoRNAs associated with proteins, function in the maturation of rRNAs creating two types of modiWed nucleotides (2Ј-O-methylation and pseudouridylation) and mediating endonucleolytic cleavages of pre-rRNAs (Gerbi and Borovjagin 2004) . Our objective is to focus this review on the ribosome biogenesis processes occurring in the nucleoli that might help to decipher the global organization of nuclear functions. We describe nucleolar organization and dynamics, propose our view on nucleolar targeting, report the relationship between the nucleolus and the cell cycle, review particular relationships between nucleolus and virus, and nucleolus related to cancer. Three main components in the active nucleolus The nucleolus has been proposed as the paradigm of nuclear functional compartmentalization (Strouboulis and WolVe 1996) . It is the site of ribosome biogenesis and in addition the nucleolar machineries are distributed in diVerent compartments. When observed by EM, three main nucleolar components (compartments) can be discerned in mammalian cells: the Wbrillar centers (FCs), the dense Wbrillar component (DFC) and the granular component (GC) (Fig. 1a) . The FCs are clear areas, partly or entirely surrounded by a highly contrasted region (Goessens et al. 1987) , the DFC. The FCs and the DFC are embedded in the GC, mainly composed of granules of 15-20 nm in diameter. The most contrasted structures in the EM sections stained with uranyl and lead correspond to high concentrations of nucleic acids. The condensed chromatin surrounding part of the nucleolus is visible using standard or preferential staining methods and also as a network within the nucleolus (Fig. 1b) . The global amount of intra-nucleolar chromatin is probably low since by light microscopy, DNA staining by DAPI excludes the nucleolus. It has become apparent that nucleoli of diVerent cell types exhibit a variable number of FCs of diVerent sizes, with an inverse proportion between size and number (Hozak et al. 1989; Pébusque and Seïte 1981) . Generally cells with a high rate of ribosome biogenesis possess numerous small FCs. On the contrary, cells with greatly reduced metabolic and transcription activities, present small nucleoli with one large-sized FC such as in lymphocytes and in inactive mammalian neurons (Hozàk et al. 1994; Lafarga et al. 1989 ). In the more active neurons, one giant FC (GFC) of 1-2 m is observed together with small FCs (Fig. 1c, d) . It was demonstrated that the GFC is enriched in the upstream binding factor, the UBF transcription factor, in a small ubiquitin-like modiWer (SUMO)-1 and Ubc9 but lack ubiquitin-proteasome and 20S proteasome (Casafont et al. 2007 ). However, the possibility that only one FC might play a role in storage and become a GFC during intense nucleolar activity is still an open question. It is also remarkable that the tripartite nucleolar organization is not general since the nucleoli of Drosophila and insects lack FCs (Knibiehler et al. 1982; Knibiehler et al. 1984) . It has been proposed that this diVerence in organization could be linked to the evolution of the rDNAs, in particular to the size of the intergenic sequences (Thiry and Lafontaine 2005) . The localization of the nucleolar machineries is related to their function in the production of the small and large ribosome subunits. These Wndings have led to assigning speciWc functions to speciWc compartments of the nucleolus. Nascent transcripts appear at the junction between the FCs and DFC and accumulate in the DFC (Cmarko et al. 2000; Guillot et al. 2005; Hozàk et al. 1994; Puvion-Dutilleul et al. 1997; Shaw and Jordan 1995) . This was recently conWrmed in the GFC since no transcripts can be detected in these large structures (Casafont et al. 2007 ). Processing of the 47S pre-rRNA starts at the site of transcription in the DFC (Cmarko et al. 2000) and continues during the intranucleolar migration of the RNA towards the GC. The nucleolar proteins that participate in the early stages of rRNA processing, localize in the DFC, such as Wbrillarin and nucleolin along with the U3 snoRNAs (Biggiogera et al. 1989; Ginisty et al. 1998; Ochs et al. 1985b; Puvion-Dutilleul et al. 1991) , whereas proteins B23/NPM (nucleo- Fig. 1 The nucleolus of mammalian cells as seen by electron microscopy. a In the human HeLa cell, the three main nucleolar components are visible in a section of material Wxed in glutaraldehyde and osmium tetroxyde, embedded in Epon and the section contrasted with uranyl acetate and lead citrate. FCs of diVerent sizes are visible and the largest is indicated by an asterisk. The FCs are surrounded by the DFC and are embedded in the GC. b Preferential contrast of DNA using NAMA-Ur staining in a PtK1 cell (courtesy J. Gébrane-Younès). The nucleolus is the gray structure surrounded by highly contrasted chromatin (arrow). Some chromatin Wlaments are also visible inside of the nucleolus (Nu). c, d Nucleolus of rat neurones (courtesy M. J. Pébusque) in the day (c), and during the night (d) which is the active period for the nucleolus of the rat. In the nonactive period (c), the nucleolus is reticulated with small FCs (asterisk). In the active period, one giant FC is visible (d, asterisk) . Bar in a: 0.5 m and bars in b, c and d: 1 m phosmin) and PM-Scl 100 (rrp6 in yeast) that are involved in intermediate or later stages of processing have been localized to the GC (Biggiogera et al. 1989; Gautier et al. 1994) . Recent advances in the isolation of large RNP complexes by tandem aYnity puriWcation and the characterization of their constituents demonstrated that two largely independent processing machineries exist in yeast nucleoli, the SSU processome (Dragon et al. 2002; Grandi et al. 2002) and the LSU processing/assembly factors (Raué 2004) . The SSU/90S processome is localized in the DFC and most of the 60S processing occurs in the GC. There is no particular domain characterized in the GC corresponding to the 43S subunit. This is most probably due to the limited events of 40S processing in the GC since the last step of processing occurs in the cytoplasm. In conclusion it seems that in the nucleoli, the vectorial distribution of the machineries successively involved in ribosome biogenesis correlates with the diVerent processing steps of the biogenesis of the ribosome subunits. When ribosome biogenesis is active, the conWnement of certain machineries in the FCs, DFC or GC makes it possible to reveal these subnucleolar constituents by immunoXuorescence as illustrated for FCs ( Fig. 2A) , DFC (Fig. 2Ba , b), and GC (Fig. 2Bc, d) . The factors associated with the rDNA transcription machinery are distributed in several foci, most frequently inside the nucleolar volume as illus-trated for UBF. These foci correspond to FCs. A distribution within the network inside the nucleolus is typical of the DFC as demonstrated for Wbrillarin. Labeling of the nucleolar volume excluding small areas contained within the volume is typical of the GC as illustrated for B23/NPM. These labeling patterns (FCs, DFC, GC) in the nucleoli provide a good indication of the step of ribosome biogenesis concerned and also reveal the blockage of ribosome biogenesis when this organization is disturbed (see below). Transient association of functionally related components appears necessary to generate a morphologically deWned nucleolus with its three distinct components, thereby maintaining the nucleolus in its usual organization. This suggests that such an organization results from the activity of ribosome biogenesis. Indeed nucleolar reorganization is induced when ribosome biogenesis is impaired either by inhibiting rDNA transcription, or inhibiting rRNA processing and/or transport. Nucleolar segregation is observed upon rDNA transcriptional arrest either in physiological conditions or induced Fig. 2 The subnucleolar constituents revealed by Xuorescence microscopy. (A) The rDNA transcription machinery, illustrated by UBF labeling, is localized in several foci corresponding to FCs in HeLa cells (a). rDNA transcription sites detected by in situ BrUTP incorporation (b), mainly colocalize with UBF as seen by the merge (c). The nucleus is visualized by Dapi staining (d). (B) HeLa cells expressing either Wbrillarin-GFP fusion or DsRed-B23 fusion. Fibrillarin decorates the DFC (a, b) whereas B23 decorates the GC (c, d). In ActDtreated cells nucleoli are segregated, Wbrillarin localizes in caps (e, f) contrary to B23 that localizes in the central body and outside the caps (g, h). Arrowheads point the caps. Bars: 10 m by low doses of actinomycin D (ActD). The segregation of nucleoli is characterized by the separation of the nucleolar components that remain close to each other but no longer intermingle ( Fig. 2Be -h) (for reviews see Hadjiolov 1985; Hernandez-Verdun and Junéra 1995; Scheer and Benavente 1990) . The eVect of ActD on nucleolar organization follows sequential changes: Wrst the Wbrillar components (FCs and DFC) condense and migrate towards the periphery of the nucleolus, after which the nucleolar components segregate to Wnally form a central body associated with caps (Hadjiolova et al. 1995) . In the caps are several proteins related to the RNA polymerase (pol) I transcription machinery such as UBF, close to Wbrillarin-containing caps. In the central body are proteins derived from the GC, some of which are progressively released, such as PM-Scl 100. It was recently demonstrated that certain nucleolar caps of segregated nucleoli could recruit factors involved in mRNA splicing. In this case, localization is induced by inhibition of both RNA pol I (rRNA transcription arrest) and RNA pol II (mRNA transcription arrest) (Shav-Tal et al. 2005 ). This is not observed when only RNA pol I is inhibited indicating that the composition of a segregated nucleolus can be more complex when induced by general transcription inhibition. One question that remains unanswered is how nucleolar components continue to be maintained in segregated nucleoli in spite of the absence of transcription or pre-rRNA processing. Nucleolar proteins may still be capable of forming complexes during inhibition of transcription, but why these complexes remain juxtaposed is presently unknown. Recently, it was reported that re-localization of proteins in speciWc caps of segregated nucleoli (after inhibition of RNA pol I and II transcription) is an energy-dependent repositioning process that requires active metabolism of the cells (Shav-Tal et al. 2005 ) most probably also ATP and GTP. A clue to the question of rRNA degradation in the nucleolus was recently proposed in yeast: a surveillance pathway that eliminates defective 60S pre-ribosomal subunits after addition of poly(A) tails was described (LaCava et al. 2005) . RNA degradation appears to occur preferentially within a subnucleolar structure, the No-body, and is mediated by the exosome (Dez et al. 2006) . Similarly, when the nuclear protein of the exosome rrp6 was deleted, poly(A) rRNAs and poly(A) U14 snoRNAs colocalized in one focus with Nop1 (Wbrillarin in human), most probably the Nobody (Carneiro et al. 2007 ). This body is distinct from the nucleolar body that functions in snoRNA maturation in yeast and could be a compartment where polyadenylation and degradation of nucleolar RNAs take place. This compartmentation would promote eYcient recognition of rRNAs in view of degradation by the exosome (Carneiro et al. 2007 ). During nucleolar segregation induced by ActD in human cells, rRNAs are degraded. However, the formation of one focus containing PM-Scl 100 has not been described; it could be either an early event that was not carefully examined, or rRNA degradation could be diVerent in yeast and mammalian cells. Disconnection between rDNA transcription sites and the late rRNA processing proteins can be induced either by kinase inhibitors or by modiWcations of snoRNA domains (Chan et al. 1996; Colau et al. 2004; David-Pfeuty et al. 2001; Rubbi and Milner 2003; Sirri et al. 2002) . The separation of the DFC and GC can be reversed by removal of a CK2 inhibitor, restoring nucleolar organization. The CK2 kinase is known to phosphorylate several nucleolar proteins (Meggio and Pinna 2003) . We postulate that the connection between DFC and GC is controlled at least in part by phosphorylation of these proteins. This hypothesis was veriWed for B23 by mutation of the major site of CK2 phosphorylation (Louvet et al. 2006 ). In conclusion, the rRNA processing proteins can be disconnected from the rRNA transcription sites indicating that rRNA transcripts are not suYcient to attract the processing proteins in these conditions. The dynamics of nucleolar reformation and the connection between DFC and GC is ATP/GTP dependent, sensitive to temperature, and is CK2-driven. The analysis in living cells of intranuclear dynamics has recently become possible using Xuorescent fusion proteins. Time-lapse videomicroscopy can track the movement of large Xuorescent complexes in the cell, and Xuorescent recovery after photobleaching (FRAP) can measure the intracellular mobility or the residency time of Xuorescent proteins (Lippincott-Schwartz et al. 2001) . Inverse FRAP (iFRAP) quantiWes the loss of Xuorescence of the region of interest (ROI) after complete bleaching outside this region (Dundr et al. 2004 ). This constitutes a direct evaluation of the residency time of the proteins in the ROI. Another approach is the use of photoactivatable GFP (PA-GFP) to follow the traYc of the activated proteins (Patterson and Lippincott-Schwartz 2002) . This process is similar to pulsechase experiments since it makes it possible to follow a pool of labeled proteins starting at time zero. These technologies applied to nuclear dynamics have introduced new dimensions and unexpected concepts concerning nuclear functional compartmentation. The mobility of several GFPtagged nuclear proteins (nucleolar proteins, histones, DNA binding proteins, transcription factors, splicing factors, nuclear receptors) has been estimated by FRAP and the recovery of Xuorescence was slower than predicted for isolated diVusing proteins of similar size. FRAP recovery rates change with inhibition of transcription, decreased temperature and depletion of ATP indicating that recovery is correlated with nuclear activity. It was demonstrated that nucleolar proteins rapidly associate with and dissociate from nucleolar components in continuous exchanges with the nucleoplasm (Phair and Misteli 2000) . The recovery curve of GFP-Wbrillarin (DFC marker) in the nucleolus reached a plateau, 60 s after bleaching and the plateau indicated an immobile fraction of »15%. The diVusion coeYcient of Wbrillarin (estimated between 0.02 and 0.046 m 2 s ¡1 ) was 10 times lower in the nucleolus than in the nucleoplasm (Chen and Huang 2001; Phair and Misteli 2000; Snaar et al. 2000) . This value is proposed to reXect the time of residency of Wbrillarin engaged in nucleolar activity, and could explain the fact that the time of residency of Wbrillarin is shorter in the Cajal body than in the nucleolus (Dundr et al. 2004 ). The nucleolar proteins engaged in rRNA transcription and processing (respectively UBF, B23, Nop52, nucleolin and Rpp29) also move with rapid recovery rates in the nucleolus as does Wbrillarin (Chen and Huang 2001; Louvet et al. 2005) . Conversely the recovery rates of ribosomal proteins are slow (»3 times slower than nucleolar proteins). This could reXect a slower mechanism for ribosome protein assembly compared with transcription and processing (Chen and Huang 2001) , or alternatively, more stable associations of ribosomal proteins with the pre-rRNAs. B23 (also designated NPM) is a multifunctional protein, abundant in the GC of the nucleolus that undergoes diVerent phosphorylation events during the cell cycle. It was recently demonstrated by FRAP that the kinetics of B23 depends on its phosphorylation status (Negi and Olson 2006) . During interphase, the half-time (t 1/2 ) recovery of B23 is 22 s in nucleoli but when the CK2 phosphorylation site is mutated (S125A) the t 1/2 increases to 44 s, and when a mutant mimicking phosphorylation charges of the four sites of mitotic CDK1 phosphorylation, the t 1/2 decreases to 12 s. This could indicate that the S125A-B23 protein has a higher aYnity for the nucleolar components (Negi and Olson 2006) . Alternatively this could correspond to a decreased turner-over in the nucleolar complexes in correlation with the disconnection of the DFC and GC occurring by overexpression of S125A-B23 (Louvet et al. 2006) . Overexpression during interphase of B23 mimicking four sites of mitotic phosphorylation increased the mobility of B23. It is tempting to propose that this results from a defect of aYnity for rRNAs of these B23s as demonstrated for mitotic phosphorylation of B23 (Okuwaki et al. 2002) . Inhibition of pol I transcription by ActD does not prevent traYc of nucleolar proteins. However, if the diVusion coeYcient of nucleolar proteins in the nucleoplasm is similar for active and repressed pol I transcription, the traYc in segregated nucleoli changes diVerently for diVerent nucleolar components. TraYc of UBF in the nucleolus is decreased by ActD, whereas it is similar for nucleolin or increased for ribosomal proteins (Chen and Huang 2001) . In contrast to the well-deWned nucleolar structures visible by EM, all the nucleolar proteins involved in ribosome biogenesis that have been examined, cycle between the nucleolus and the nucleoplasm in interphase cells. To summarize, it is now established that rapid diVusion of nucleolar proteins occurs in the nucleoplasm and recruitment to the nucleolus is permanent. Moreover, the diVerence in kinetics of several proteins shared between the nucleolus and the Cajal body suggests the existence of compartment-speciWc retention (Dundr et al. 2004 ). To be localized or retained within the nucleolus In eukaryotic cells, once imported or diVused into the nucleus, some proteins distribute throughout the nucleoplasm and others are targeted to speciWc nuclear compartments such as nucleoli. Proteomic analyses revealed that at least 700 proteins are localized in nucleoli (Andersen et al. 2002 (Andersen et al. , 2005 Leung et al. 2003) . Whereas the rules and signals that govern the nuclear localization and nuclear export of proteins are now well deWned, those concerning nucleolar localization are still debated. Contrary to the nuclear localization signals (NLSs), nucleolar localization signals or sequences (NoLSs) are not well characterized. Although several NoLSs have been described, no obvious consensus sequence has emerged. Nevertheless all NoLSs reported for nucleolus localizing virus proteins, such as HIV-1 Rev (Kubota et al. 1989 ), HIV-1 Tat (Dang and Lee 1989) and human T-cell leukaemia virus type 1 Rex (HTLV-1 Rex) (Siomi et al. 1988) , and for cellular proteins such as the nucleolar protein p120 , Survivin-deltaEx3 (Song and Wu 2005) and HSP70 (Dang and Lee 1989) are rich in basic residues. The capacity of numerous proteins to adopt nucleolar localization has been correlated with interaction of these proteins with B23. Owing to the ability of numerous nucleolar proteins to interact with B23 and because this major nucleolar protein shuttles constantly between the nucleus and the cytoplasm (Borer et al. 1989) , it was frequently suggested that B23 might be a transporter for nucleolar proteins possessing a NoLS (Fankhauser et al. 1991; Li 1997; Valdez et al. 1994) . Even if this tempting hypothesis has never been demonstrated, recent results obtained using stable U2OS-derived cell lines with reduced B23 expression levels showed that the nucleolar localization of ARF is linked to B23 (Korgaonkar et al. 2005) . Indeed, reduced expression of B23 induced a partial delocalization of ARF from nucleoli to nucleoplasm. The authors therefore concluded that B23 targets ARF to nucleoli in a dose-dependent manner. Nevertheless, this result does not allow discriminating between a role for B23 in the transport of ARF from nucleoplasm to nucleoli and/or in the retention of ARF in nucleoli. A NoLS, i.e. a sequence essential for nucleolar localization, is most probably a sequence involved in nucleolar retention by interacting with a nucleolar molecule such as B23 (Lechertier et al. 2007 ). Indeed, recent analyses of the intranuclear dynamics of proteins in living cells revealed that nuclear proteins could diVuse within the nucleoplasm (Phair and Misteli 2000; Sprague and McNally 2005) . As for the nucleolus, it was demonstrated that nucleolar proteins rapidly associate with and dissociate from nucleolar components in a continuous exchange with the nucleoplasm (Chen and Huang 2001; Dundr et al. 2004; Phair and Misteli 2000; Snaar et al. 2000) . There probably exist compartment-speciWc retention mechanisms for proteins in nuclear bodies, implying that the residency time of a particular molecule in a given nuclear body depends on its speciWc interactions (Misteli 2001) . In support of this possibility, we have recently shown that the fusion of a B23-interacting sequence with Wbrillarin makes it possible to re-localize Wbrillarin from the DFC to the GC of nucleoli where B23 is mainly localized (Lechertier et al. 2007) . Similarly, by fusing the B23-interacting sequence to MafG (part of the nuclear transcription factor NF-E2 composed of both MafG and p45), NF-E2 is redirected from the nucleoplasm to the GC. Therefore, interactions most probably govern the nuclear distribution of proteins and a NoLS is very likely a nucleolar molecule-interacting sequence. However, nucleolar localization of a protein is most probably governed by several factors and the presence of a NoLS in its sequence is not suYcient to predict nucleolar localization of the protein. In particular, a nucleolar protein must Wrst be localized in the nucleus, and consequently all mechanisms that interfere with nuclear import and/or nuclear export of a nucleolar protein will modify its localization at the steady state. A good illustration is provided by the major nucleolar protein, B23. This multifunctional protein is normally mainly located in the GC of nucleoli but exhibits an aberrant cytoplasmic localization in one-third of acute myeloid leukemias due to mutations in its C-terminal coding exon that causes a frameshift and the formation of an additional CRM1-dependent nuclear export signal (NES) (Mariano et al. 2006) . Another example showing the diYculty encountered in predicting nucleolar localization is provided by the box C/D snoRNPs: it seems clear that the nucleolar localization of box C/D snoRNPs is linked to their biogenesis (Verheggen et al. 2001; Watkins et al. 2002) . Indeed, by modifying the conserved stem II of the box C/D motif present in the U14 snoRNA, both the spe-ciWc assembly of the box C/D snoRNP and nucleolar localization are lost (Watkins et al. 2002) . Moreover, genetic depletion of one of the four core proteins, namely 15.5kD, Nop56, Nop58 and Wbrillarin, also inhibits the nucleolar localization of box C/D snoRNPs (Verheggen et al. 2001 ). However, targeting of box C/D snoRNPs to nucleoli is not yet fully understood. Indeed, unexpectedly two nuclear export factors, PHAX and CRM1 appear to be stably associated with the U3 pre-snoRNPs (Boulon et al. 2004; Watkins et al. 2004 ). Boulon and coworkers proposed that U3 precursors bind PHAX, which targets the complex to the Cajal body, and that subsequently CRM1 further targets the U3 complexes to the nucleolus. Even if PHAX and CRM1 play an important role in the transport of box C/D snoRNPs to the nucleus, the possibility that these proteins may also function in the nuclear export of snoRNPs cannot be excluded (Watkins et al. 2004 ). This possibility is reinforced by a recent study showing that in addition to nuclear export factors, the nuclear import factor Snurportin 1 is involved in U8 box C/D snoRNP biogenesis (Watkins et al. 2007) . Nucleolar localization of the components of the box C/D snoRNPs would therefore depend on the biogenesis of the box C/D snoRNP complexes, which would imply nuclear export. rDNA transcription machinery rDNAs are found in multiple, tandem, head-to-tail arrayed copies in the nucleoli of eukaryotic cells (Hadjiolov 1985) . In mitotic human cells, rDNA clusters are localized on the short arm of the Wve pairs of chromosomes 13, 14, 15, 21 and 22 and are termed NORs. Each rDNA unit consists of a transcribed sequence and an external non-transcribed spacer (Hadjiolov 1985; Liau and Perry 1969) in which all the sequences necessary for proper RNA pol I transcription such as proximal promoters, spacer promoters and terminators are located (Hadjiolov 1985) . In the rDNA promoter two important elements have been described, a CORE element and an upstream control element (UCE) (Haltiner et al. 1986; Windle and Sollner-Webb 1986; Xie et al. 1992 ) that function synergistically to recruit a transcriptionally competent RNA pol I complex. This complex contains in addition to RNA pol I, the upstream binding factor (UBF) (Pikaard et al. 1989; Voit et al. 1992) , the selectivity factor protein complex SL1 (Learned et al. 1985 ) also called TIF-1B in mouse cells (Clos et al. 1986 ), consisting of the TATA-binding protein (TBP) and four transcription activating factors [TAF I s110, 63, 48 and 41 (Comai et al. 1994; Gorski et al. 2007; Zomerdijk et al. 1994) ], the transcription initiation factor TIF-IA, the mouse homolog of Rrn3p (Bodem et al. 2000; MooreWeld et al. 2000) and the transcription termination factor TTF-1 (Bartsch et al. 1988 ). The UBF containing HMG boxes (Bachvarov and Moss 1991; Jantzen et al. 1990 ) that confer a high aYnity for DNA structures plays an architectural role on the rDNA promoter (Mais et al. 2005) . It was proposed that UBF activates rDNA transcription because it stabilizes binding of SL1 required to recruit the initiation-competent subfraction of RNA pol I. This recruitment is achieved by interaction of UBF with the RNA pol I-associated factor PAF53 (Schnapp et al. 1994) , and by interaction of SL1 with TIF-1A/ Rrn3p (Miller et al. 2001) . TIF-1A/Rrn3p interacts also with the RPA43 subunit of RNA pol I and thus facilitates linking between RNA pol I and SL1 complexes (Peyroche et al. 2000; Yuan et al. 2002) . Following initiation, TIF-1A/ Rrn3p is released and can associate with another preinitiation complex. Recycling of TIF-1A/Rrn3p requires a posttranslational phosphorylation event that appears to play a role in its initiation activity (Cavanaugh et al. 2002; Zhao et al. 2003) . Moreover, it was proposed that TTF-1 is not only involved in termination of transcription in cooperation with the release factor PTRF (Jansa and Grummt 1999), but also in the remodeling of ribosomal chromatin by recruiting ATP-dependent remodeling factors to the rDNA promoter (Längst et al. 1997 ). The nucleolar remodeling complex (NoRC) (Strohner et al. 2001) , which acts in repression at the rDNA promoter level (Li et al. 2006; Santoro et al. 2002) , and the transcription activator CSB (Cockayne syndrome group B protein), a DNA-dependent ATPase, interact with TTF-1 (Yuan et al. 2007 ). The Wnding that TTF-1 interacts with both CSB and NoRC suggests that competitive recruitment of CSB and NoRC may determine the epigenetic state of the rDNA. It is now established that the presence of a fully active nucleolus depends on cell cycle regulators. rDNA transcription is maximum in the S and G2 phases, silent in mitosis, and slowly recovers in G1. Post-translational mod-iWcations of the RNA pol I machinery are required for the formation of a productive preinitiation complex. The phosphorylation status of several components of the RNA pol I machinery can modify the activity and interactions of these proteins and thus can modulate rDNA transcription during the cell cycle. Concerning silencing of rDNA transcription during mitosis, it is well established that some components of the rDNA transcription machinery such as SL1 (Heix et al. 1998 ) and TTF-1 (Sirri et al. 1999) , are mitotically phosphorylated by CDK1-cyclin B. As shown in vitro, CDK1-cyclin B-mediated phosphorylation of SL1 abrogates its transcriptional activity (Heix et al. 1998 ). Moreover CDK1-cyclin B is necessary not only to establish repression but also to maintain it from prophase to telophase. Indeed, in vivo inhibition of CDK1-cyclin B leads to dephosphorylation of the mitotically phosphorylated forms of components of the RNA pol I machinery and restores rDNA transcription in mitotic cells (Sirri et al. 2000) . On the other hand, rDNA transcription also appears regulated by CDK(s) during interphase: the increase of rDNA transcription during G1 progression depends on phosphorylation of UBF by G1-speciWc CDK-cyclin complexes (Voit et al. 1999) , and CDK inhibitor treatments partially inhibit rDNA transcription in interphase cells (Sirri et al. 2000) . ModiWcations of the phosphorylation status of UBF and/or TAF I 110 aVect the interactions between UBF and SL1 necessary for recruitment of RNA pol I (Zhai and Comai 1999) . In addition to phosphorylation, it has been speculated that acetyltransferases might also regulate the activity of RNA pol I transcription factors. Indeed, two studies have demonstrated that UBF and one of the SL1 subunits are acetylated in vivo (Muth et al. 2001; Pelletier et al. 2000) . Functional studies indicated that acetylated UBF is transcriptionally more active than deacetylated UBF. However, acetylation of UBF does not aVect its DNA binding activity as shown for other transcription factors, and it is unclear how this post-transcriptional modiWcation modulates UBF activity. The TAF I 68 subunit of SL1 is speciWcally acetylated by recruitment of PCAF (p300/CBP associated factor) to the rDNA promoter. In vitro analyses indicate that acetylation of TAF I 68 is likely to increase the activity of SL1 facilitating interaction of the complex with DNA. Sirtuins, the human homologues of the yeast Sir2 (silent information regulator) with NAD-dependent deacetylase and ADP-ribosyltranferase activity, have recently been implicated in the regulation of the RNA pol I machinery. In particular, nuclear sirtuin1 deacetylates TAF I 68 and represses RNA pol I transcription in vitro (Muth et al. 2001) . Conversely, the nucleolar sir-tuin7 is described as activator of rDNA transcription by increasing RNA pol I recruitment to the rDNA, but no substrates of such activity have as yet been identiWed (Ford et al. 2006) . Additional in vivo approaches are necessary to better understand the role of sirtuins in the regulation of rDNA transcription. SUMO modiWcation is reported to inXuence the assembly of transcription factors on promoters and the recruitment of chromatin-modifying enzymes, and is often associated with transcriptional repression (Gill 2004) . Recently, the colocalization of SUMO-1 and UBF in the GFC (Casafont et al. 2007 ) of neuronal cells and the nucle-olar localization of the sentrin/SUMO-speciWc proteases, SENP3 and SENP5 (Gong and Yeh 2006; Nishida et al. 2000) suggest a potential role of sumoylation on the regulation of rDNA transcription. Further studies of the identiWcation of sumoylated nucleolar transcription factors will be necessary to verify this possibility. In higher eukaryotic cells at the beginning of mitosis when rDNA transcription is repressed, the nucleoli disassemble and are no longer observed throughout mitosis. Conversely nucleoli assemble at the exit from mitosis concomitantly with restoration of rDNA transcription and are functionally active throughout interphase. In late prophase when mitotic repression of rDNA transcription occurs, the rDNA transcription machinery remains associated with rDNAs in the NORs as revealed by the analysis of diVerent components at the steady state (Roussel et al. , 1996 Sirri et al. 1999 ). Nevertheless, more recent quantitative kinetic analyses have revealed that some RNA pol I subunits, including RPA39, RPA16 and RPA194, transiently dissociate from the NORs during metaphase and reappear in anaphase (Chen et al. 2005; Leung et al. 2004) . As for the mechanism that governs disassembly of nucleoli in prophase, it may be assumed that it is linked to repression of rDNA transcription, most probably caused by CDK1-cyclin B-directed phosphorylation of components of the rDNA transcription machinery (Heix et al. 1998; Sirri et al. 1999) . At the beginning of prophase, the components of the pre-rRNA processing machinery do not remain in the vicinity of the rDNAs (Gautier et al. 1992 ) but become partially distributed over the surface of all the chromosomes (reviewed in Hernandez-Verdun et al. 1993) . The nucleolar proteins that relocate to the chromosome periphery are components of the DFC and GC of the active nucleolus. In living cells, nucleolar proteins tagged with GFP are concentrated around the chromosomes during mitosis and migrate with the chromosomes (Savino et al. 2001) . However, the mechanisms maintaining interactions of nucleolar processing proteins with chromosomes during mitosis have not been characterized. The colocalization of the diVerent factors involved in pre-rRNA processing suggests that processing complexes are at least to some extent maintained during mitosis. It is as yet unknown whether migration of the nucleolar processing proteins occurring at the onset of mitosis (Fan and Penman 1971) takes place as a consequence of the arrest of pre-rRNA synthesis or whether it is also regulated. Indeed, it is noticeable that (1) during prophase, the components of the rRNA processing machinery appear to be delocalized before total repression of rDNA transcription occurs, and (2) the most recently synthesized pre-rRNAs accumulate as partially processed 45S pre-rRNAs (Dousset et al. 2000) suggesting that total repression of pre-rRNA processing could occur prior to total repression of rDNA transcription. These observations therefore raise the possibility that rDNA transcription and pre-rRNA processing are both repressed during prophase by distinct mechanisms. Nucleoli assemble at the exit from mitosis concomitantly with restoration of rDNA transcription at the level of competent NORs (Roussel et al. 1996) . Until recently it was admitted that transcriptionally active rDNAs, serving as nucleation sites, possessed by themselves the ability to organize the nucleoli (Scheer and Weisenberger 1994) . Results obtained in the laboratory showed that (1) reactivation of rDNA transcription in mitotic cells does not lead to the formation of nucleoli (Sirri et al. 2000) , (2) initiation of nucleolar assembly occurs independently of rDNA transcription (Dousset et al. 2000) , and (3) at the exit from mitosis nucleologenesis is impaired in the presence of a CDK inhibitor even if rDNAs are actively transcribed (Sirri et al. 2002) . Consequently, the formation of functional nucleoli at the exit from mitosis is not governed solely by the resumption of rDNA transcription. Based on previous studies (Sirri et al. 2000 (Sirri et al. , 2002 , we propose that the formation of nucleoli is a process regulated by CDK(s) at two levels: resumption of rDNA transcription but also restoration of rRNA processing. In anaphase, early and late processing proteins (respectively Wbrillarin, and Bop1, B23, Nop52) are homogeneously distributed around the chromosomes. During telophase and early G1, along the translocation pathway between chromosome periphery and transcription sites, processing proteins concentrate in foci designated prenucleolar bodies (PNBs), Wrst described in plant cells (Stevens 1965) . PNB formation is a general phenomenon occurring during the recruitment of the nucleolar processing proteins at exit from mitosis (Angelier et al. 2005; Azum-Gélade et al. 1994; Dundr et al. 2000; Jiménez-Garcia et al. 1994; Ochs et al. 1985a; Savino et al. 2001 ). This appears to be a cell cycle regulated process since when the nucleolar function is established during interphase, recruitment of processing proteins is not associated with PNB formation. Inactivation of CDK1-cyclin B occurring at the end of mitosis induces the Wrst events of nucleologenesis. Strikingly, Wbrillarin concentrates in PNBs and rDNA clusters when decrease in CDK1-cyclin B activity overcomes the mitotic repression of RNA pol I transcription (Clute and Pines 1999) , while Nop52 and other GC proteins are recruited later on transcription sites. This late recruitment is under the control of cyclin-dependent kinases since CDK inhibitors block this process (Sirri et al. 2002) . Thus, it seems that recruitment of the processing machinery at the time of nucleolar assembly is a regulated process most probably dependent on cell cycle progression. This provides a physiological situation to investigate the formation, control and dynamics of nuclear bodies. The dynamics of the processing nucleolar proteins was analyzed at the transition mitosis/interphase using rapid time-lapse video microscopy (Fig. 3) . The Wrst detectable assembly of proteins in foci occurred on the surface of the chromosome during telophase (Savino et al. 2001 ), followed by the progressive delivery of proteins to nucleoli ensured by progressive and sequential release of proteins from PNBs (Dundr et al. 2000) . Based on the observations of diVerent Wxed cells, it was concluded that the early processing proteins are recruited Wrst on transcription sites while the majority of the late processing proteins are still in PNBs (Fomproix et al. 1998; Savino et al. 1999 ). This sequence of events was conWrmed in living HeLa cells. Fibrillarin resides brieXy in PNBs (»15 min) before recruitment to the nucleolus, while Nop52 is maintained longer in PNBs (»80 min) (Savino et al. 2001) . The relative dynamics of early and late rRNA processing proteins at the time of PNB formation was examined using coexpression of GFP-Wbrillarin and DsRed-B23 (Angelier et al. 2005) . Once near the poles, 1-2 min after the onset of telophase, numerous bright Xuorescent foci containing both GFP-Wbrillarin and DsRed-B23 appeared almost simultaneously. For about 10 min, the relative amount of B23 in foci was Wve to six times higher than that of the dispersed proteins whereas the amount of Wbrillarin in the same foci was three to four times higher than that of dispersed proteins. Subsequently, Wbrillarin was released while B23 was still present in the foci. This clearly illustrates the presence of the two types of nucleolar processing proteins in the same PNBs and suggests diVerential sorting of these proteins. Conversely in the same conditions, simi-lar dynamics and Xows of GFP-Nop52 and DsRed-B23 were observed. Thus the processing proteins passed through the same PNBs and were released simultaneously suggesting that these proteins could form complexes in PNBs. Time-lapse analysis of Xuorescence resonance energy transfer (FRET) was chosen to determine whether nucleolar processing proteins interact along the recruitment pathway. The apparatus used to determine FRET performs tdFLIM (time domain Xuorescence lifetime imaging microscopy) by the time and space-correlated single-photon counting method (Emiliani et al. 2003) . This technique directly yields the picosecond time-resolved Xuorescence decay for every pixel by counting and sampling single emitted photons. Positive FRET between GFP-Nop52 and DsRed-B23 in nucleoli indicates that the distance and most probably the interactions between the proteins can be evaluated by this approach (Angelier et al. 2005) . Since it is possible to detect FRET between B23 and Nop52 in nucleoli, FRET was tracked during the recruitment of these proteins into nucleoli from anaphase to early G1. FRET was never detected during anaphase at the periphery of the chromosomes whereas it was registered in 20% of the PNBs at the beginning of telophase, in about 40% at the end of telophase, and in 55% in early G1. Thus, interaction between GFP-Nop52 and DsRed-B23 was established progressively in PNBs, as the number of PNBs exhibiting FRET increased. Such data indicate that Nop52 and B23 did not interact until they were recruited in PNBs. It is noteworthy that a given PNB can alternatively present or not present FRET. Based on the behavior of these two proteins, one possibility is that late rRNA processing proteins already interact in PNBs. Were this to be conWrmed for other rRNA processing complexes, PNBs could be proposed as assembly platforms of processing complexes at this step of the cell cycle. It would be very interesting to establish whether this role can be extended to the early rRNA processing machinery (Angelier et al. 2005) . In conclusion, assembly of the nucleolus requires reactivation of the rDNA transcription machinery, and also recruitment and reactivation of the pre-rRNA processing Fig. 3 At the exit from mitosis, the dynamics of DsRed-B23 is followed in living cell. In telophase (0 min), the B23 signal is visible in small foci. These foci corresponding to PNBs are clearly visible 20 min later. The B23-containing PNBs are distributed in the nucleoplasm and B23 is progressively recruited in the incipient nucleolus (40 min). Nu nucleolus machinery. Indeed cells exiting from mitosis in the presence of a CDK inhibitor exhibit neither relocalization of the late pre-rRNA processing components from PNBs to rDNA transcription sites, resumption of proper rRNA processing, nor formation of functional nucleoli. The link between cell proliferation, cancer and nucleolar activity has been well established during the past several decades (more than 5,000 references). Half of the studies related to the nucleolus and cancer are dedicated to the prognostic value of AgNOR staining, a technique revealing the amount of nucleolar proteins. The aim of this technique is to evaluate the proliferation potential of cancer cells by measuring nucleolar activity. B23, nucleolin, UBF and subunits of RNA pol I were found to be the argyrophilic proteins responsible for the silver-staining properties of nucleolar structures (Roussel et al. 1992; Roussel and Hernandez-Verdun 1994) . In interphase cells, the amount of major AgNOR proteins, B23 and nucleolin, is high in S-G 2 and low in G 1 phases and thus a higher value of AgNOR corresponds to actively cycling cells (Sirri et al. 1997) . Standardization of the AgNOR staining method permits routine application of this technique for clinical purposes. The size of the nucleolus is generally enlarged in cancer cells, and this has been correlated with cell proliferation. A new Weld of research was recently opened by the discovery that several tumor suppressors and proto-oncogenes aVect the production of ribosomes (Ruggero and PandolW 2003) . rRNA synthesis is enhanced by c-Myc (Arabi et al. 2005) and it was proposed that this stimulation is a key pathway driving cell growth and tumorigenesis (Grandori et al. 2005) . On the contrary, the decrease of ribosome production induces apoptosis in a p53-dependent or independent manner (David-Pfeuty et al. 2001; Pestov et al. 2001) and the disruption of the nucleolus mediates p53 stabilization (Rubbi and Milner 2003) . The cross talk between the p53 pathway and the nucleolus is at least in part mediated by localization of Mdm2 in the nucleolus, an E3 ubiquitin ligase involved in p53 degradation. Nucleostemin, a nucleolar protein discovered in stem cells and in cancer cells interacts with p53 McKay 2005, 2002) . It was proposed that nucleostemin might regulate p53 function through shuttling between the nucleolus and the nucleoplasm. The major nucleolar protein B23 is directly implicated in cancer pathogenesis as demonstrated by mutation of the gene in a number of hematological disorders (Grisendi et al. 2005) . Importantly, in acute promyelocytic leukemia, the fusion protein NPM/RAR localizes in the nucleolus indicating a role of this nucleolar protein in this disease (Rego et al. 2006 ). Within the last few years, increasing evidence has revealed that viruses require the nucleus and in particular the nucleolus to target proteins indispensable for their replication. An increasing number of key proteins from both DNA-and RNA-containing viruses are localized in the nucleolus: viruses of the family Herpesviridae, Adenoviridae, Hepadnaviridae, Retroviridae, Rhabdoviridae, Orthomyxoviridae, Potyviridae, Coronaviridae and Flaviviridae, encode such proteins. Viruses have developed diVerent strategies to facilitate targeting of their proteins to the nucleolus: (1) it was reported that the sequences of certain viral proteins harbor NoLS and NES (Harris and Hope 2000; Hiscox 2007; Kann et al. 2007) . Recently it was demonstrated by mutagenesis that the nucleocapsid (N) protein of infectious bronchitis virus (IBV), presents an 8 amino acid-long motif that functions as NoLS, and is necessary and suYcient for nucleolar retention of the N protein and colocalization with nucleolin and Wbrillarin; the NoLS is required for interaction with cell factors. (2) Other viral proteins present sequences rich in arginine-lysine (Ghorbel et al. 2006; Reed et al. 2006) known to be nucleolar retention signals; generally, these sequences overlap the NLS. (3) Some viral proteins that target the nucleolus present motifs with aYnity for double-stranded RNA (dsRNA), for RNA binding or for DNA binding (Melen et al. 2007) . (4) Other studies showed that nucleolar localization of viral proteins, is cell cycle-dependent (Cawood et al. 2007 ); using synchronization studies coupled to live cell confocal microscopy, the authors demonstrated that the concentration of N protein in the nucleolus was higher in the G2/M phase than in other phases, and that in this phase the protein was more mobile in the nucleoplasm. In all the cases examined, the viral proteins depend on cell factors to successfully shuttle between the nucleolus and the cytoplasm. Why must viral proteins target to the nucleolus? The answer to this question is not clear; however, diVerent authors had reported that such viral proteins are involved both in replication of the viral genome, and in transcriptional and post-transcriptional regulation of viral genome expression (Dang and Lee 1989; Pyper et al. 1998) . For example, some plants viruses are known to encode a protein designated movement protein, responsible for long-distance movement of the viral RNA through the phloem (Ryabov et al. 1999) . Movement strictly depends on the interaction of the viral movement protein with the nucleolus and the Cajal bodies, which contain snRNPs and snoR-NPs (Kim et al. 2007a, b) . The open reading frame (ORF) 3 of Groundnut rosette virus is one such protein; it is Wrst localized in Cajal bodies and forms Cajal body-like struc-tures, it is then localized in the nucleolus when the Cajal body-like structures fuse with the nucleolus, and Wnally it exits to the cytoplasm (Kim et al. 2007b) . Another study showed that this shuttling is indispensable to form the RNPs essential for systemic virus infection (Kim et al. 2007a) . In this process, the interaction of the viral ORF3 with Wbrillarin is absolutely required. Interestingly, silencing of the Wbrillarin gene blocks long-distance movement of the virus but does not aVect virus replication and movement via plasmodesmata. Because the mobility of nucleolar components depends on the interactions and functions of the components (Olson and Dundr 2005) , we suggest that targeting of viral proteins to the nucleolus could help viral protein traYc and diVusion of viral infection. The activity of the human immunodeWciency virus (HIV)-1 Rev protein is essential for virus replication. Its subcellular localization is nucleolar, but it has the ability to shuttle continuously between the nucleus and the cytoplasm (Felber et al. 1989; Kalland et al. 1994) . Rev possesses both an NES and an NLS; the NLS is associated with importin-as well as with B23 (Fankhauser et al. 1991; Henderson and Percipalle 1997) . Rev-GFP movement in the nucleolus is very slow, implying that it is attached to aYnity binding sites in this subcellular compartment (Daelemans et al. 2004 ). In addition, the transport of Rev from the nucleolus to the cytoplasm can be aVected negatively by NF90, a cellular protein that colocalizes with Rev in the nucleolus (Urcuqui-Inchima et al. 2006) (Fig. 4) . This indicates that the transport of HIV transcripts by Rev to the cytoplasm is a regulated process. Because Rev is concentrated in the nucleolus, it was suggested that the passage of Rev to the nucleolus is an indispensable step for Rev function, and hence for HIV-1 replication. Indeed, based on HIV-1 RNA traYcking through the nucleolus, this organelle is an essential participant of HIV-1 RNA export (Michienzi et al. 2000) . As discussed above for Rev, it has been shown that the Herpes virus saimiri ORF57 protein is required for nuclear export of viral intronless mRNAs (Boyne and Whitehouse 2006) . In addition, the expression of ORF57 induces nuclear traYcking, which is essential for nuclear export of such RNAs; the human transcription/export protein involved in mRNA export, is redistributed to the nucleolus in the presence of the ORF57 protein. Based on these Wndings, the authors concluded that the nucleolus is required for nuclear export of the viral mRNAs. What are the consequences for the cells of the passage of viral proteins via the nucleolus? It is known that all viruses whether with DNA or RNA genomes interfere with the cell cycle, aVecting host-cell functions and increasing the eYciency of virus replication. The data obtained suggest that targeting of virus proteins to the nucleolus not only facilitates virus replication, but may also be required for pathogenic processes. Recent studies following infection by IBV, revealed a change in the morphology and protein content of the nucleolus . This included an enlarged FC and an increase in protein content; interestingly, the tumor suppressor protein p53, normally localized in the nucleus in virus infected cells, was redistributed mainly in the cytoplasm. The Hepatitis B virus (HBV) core antigen (HBcAg) is responsible for export of the virus with a mature genome (Yuan et al. 1999a, b) . Indeed a change from isoleucine to leucine in position 97 (I97L) of HBcAg causes the cell to release virus particles with immature genomes. HBcAg with a mutation Fig. 4 HIV Rev-GFP and NF90-RFP fusions were expressed in HeLa cells. Both proteins colocalize in nucleoli as seen by the merge. The nucleus is visualized by Dapi staining. Bars: 10 m in position 97 (I97E or I97W) has been detected in the nucleolus colocalizing with nucleolin and B23, and this colocalization was often related with binucleated cells or apoptosis (Ning and Shih 2004) , suggesting that the localization of HBcAg in the nucleolus could perturb cytokinesis. The authors propose that this event may be associated with liver pathogenesis. Some factors expressed by west nile virus (WNV) such as NS2B and NS3 and the WNV capsid (WNVCp) participate in WNV-mediated apoptosis (Oh et al. 2006; Ramanathan et al. 2006) . It is well known that p53 is activated in response to oncogenic or DNA damaging stresses, inducing cell cycle arrest and apoptosis (Harris and Levine 2005) . In normal conditions, HDM2 targets p53 and blocks abnormal accumulation of p53 by HDM2-mediated ubiquitinylation, followed by 26S proteasome-dependent degradation of p53 (Haupt et al. 1997; Kubbutat et al. 1997) . Recently it was demonstrated that WNVCp could bind to and sequester HDM2 in the nucleolus, blocking p53-HDM2 complex formation (Yang et al. 2007 ). This phenomenon causes stabilization of p53 and Bax activation and thereafter apoptosis. In addition, the authors show that WNVCp is able to induce apoptosis-dependent processes, suggesting that the viral protein mediates apoptosis through p53-dependent mechanisms by retention of HDM2 in the nucleolus. The conclusions are based on the perspectives and the tendency that can be anticipated from the present research in the Weld of the nucleolus. We propose that in the future, a better understanding of the complexity and variability of ribosome biogenesis will need to be established. For example, the diVerence between the information available in yeast and mammalian cells is of major importance. The diVerent steps of ribosome biogenesis and protein complexes are well characterized in yeast due to easy access of mutants. Similarly, Miller chromatin spreading for electron microscopy in yeast strains carrying mutations reveals the coupling of RNA pol I transcription with rRNA processing (Schneider et al. 2007 ). Additionally, the compaction into SSU processomes of pre-18S ribosomal RNA before cleavage was observed on Miller spreads (Osheim et al. 2004 ). There is presently no comparable information for the mammalian genes. Yet the tendency is to generalize and suppose that the information is similar in the two models. In the future, diVerences will most probably be revealed as well as the complexity of the regulation in diVerentiated cells. Along this line, it was demonstrated that basonuclin, a cell-type-speciWc rDNA regulator transcribes only one subset of rDNAs of a cell (Zhang et al. 2007b ). In such a diVerentiated cell, it remains to be established how the subset of rDNA repeats is selected. The nucleolus was proposed to be a domain of the sequestration of molecules that normally operate outside this organelle, mainly in the nucleoplasm. Sequestration in the yeast nucleolus of the phosphatase cdc14 and its release into the cytoplasm at anaphase was demonstrated to be a key event in cell cycle progression (for a review see Cockell and Gasser 1999; Guarente 2000; Visintin and Amon 2000) . However, it is important to recall that there is no nucleolus during mitosis in mammalian cells. In mammalian interphase cells, the nucleolus is a domain of retention of molecules related to cell cycle, life span, and apoptosis, and is in particular an actor of the p53-dependent pathway. Recently nucleolar retention of the Hand1 transcription factor was observed in trophoblast stem cells (Martindill et al. 2007 ). Phosphorylation of Hand1 induced nucleolar to nucleoplasm translocation of Hand1 and commitment of stem cells to diVerentiate into giant cells. Hand1 translocation to the nucleoplasm might regulate a crucial step of stem cell diVerentiation into polyploid giant cells but the targets of Hand1 in the nucleoplasm are still undeWned. The nucleolus is generally surrounded by highly condensed chromatin Wrst described in rat hepatocytes and presently known as heterochromatin. By following the movements of chromosome sequences introduced in diVerent sites in chromosomes of living cells, it was demonstrated that loci at nucleoli periphery and nuclear periphery are less mobile than in other sites. Disruption of the nucleoli by a CK2 inhibitor increases the mobility of the perinucleolar loci. It was proposed that the nucleolus and nuclear periphery could maintain the three-dimensional organization of chromatin in the nucleus (Chubb et al. 2002) . Recently the perinucleolar ring of chromatin was brought to the fore when its role in the maintenance of inactive X (Xi) was demonstrated (Zhang et al. 2007a) . During middle and late S phase, Xi contacts the nucleolar periphery when it is replicated during the cell cycle. It was discovered that the perinucleolar chromatin is enriched in Snf2h, the catalytic subunit of a remodeling complex required for replication of heterochromatin. These observations demonstrate the role of the perinucleolar compartment in maintaining the epigenetic state of Xi (Zhang et al. 2007a) . The presence of inactive rDNA repeats in perinucleolar heterochromatin is known in many plant cells and in Drosophila. It was recently demonstrated that disruption in Drosophila of histone H3K9 methylation, a marker of heterochromatin, induced nucleolar disorganization and decondensation, and disorganization of rDNA repeats (Peng and Karpen 2007) . The authors suggest that condensation of a part of the rDNA copies into heterochromatin could be a general strategy against recombination of these highly repeated genes. For long, interest concerning the nucleolus was to establish how eYcient ribosome biogenesis occurs and the link of this function with the cell cycle. More recently the eVect of the disruption of ribosome biogenesis appeared very important when it was proposed that the nucleolus is a sensor of stress (Rubbi and Milner 2003) . Indeed disruption of ribosome biogenesis releases ribosomal proteins from the nucleolus that bind to MDM2 and inhibit p53 degradation (Lindstrom et al. 2007) . A connection between ribosomal stress and p53-dependent cell cycle arrest is now proposed (Gilkes et al. 2006) . Considering the diversity of the recent information gathered on the nucleolus, it is clear that this is a very dynamic and rapidly progressing research area. The most promising aspect is the contribution of new models (pseudo-NORs, Prieto and McStay 2007) , new species (not only yeast), new approaches (Miller spreads using mutants, proteomics) and new questions (for instance the role of siRNAs or antisens RNAs in the activity of the nucleolus).

Antidiabetes and Anti-obesity Activity of Lagerstroemia speciosa

The leaves of Lagerstroemia speciosa (Lythraceae), a Southeast Asian tree more commonly known as banaba, have been traditionally consumed in various forms by Philippinos for treatment of diabetes and kidney related diseases. In the 1990s, the popularity of this herbal medicine began to attract the attention of scientists worldwide. Since then, researchers have conducted numerous in vitro and in vivo studies that consistently confirmed the antidiabetic activity of banaba. Scientists have identified different components of banaba to be responsible for its activity. Using tumor cells as a cell model, corosolic acid was isolated from the methanol extract of banaba and shown to be an active compound. More recently, a different cell model and the focus on the water soluble fraction of the extract led to the discovery of other compounds. The ellagitannin Lagerstroemin was identified as an effective component of the banaba extract responsible for the activity. In a different approach, using 3T3-L1 adipocytes as a cell model and a glucose uptake assay as the functional screening method, Chen et al. showed that the banaba water extract exhibited an insulin-like glucose transport inducing activity. Coupling HPLC fractionation with a glucose uptake assay, gallotannins were identified in the banaba extract as components responsible for the activity, not corosolic acid. Penta-O-galloyl-glucopyranose (PGG) was identified as the most potent gallotannin. A comparison of published data with results obtained for PGG indicates that PGG has a significantly higher glucose transport stimulatory activity than Lagerstroemin. Chen et al. have also shown that PGG exhibits anti-adipogenic properties in addition to stimulating the glucose uptake in adipocytes. The combination of glucose uptake and anti-adipogenesis activity is not found in the current insulin mimetic drugs and may indicate a great therapeutic potential of PGG.

Type 2 diabetes has developed into a worldwide epidemic (1) . Ironically, the dramatic increase in the prevalence of type 2 diabetes can be attributed to the rapid economic development and correlated to changes in lifestyle within the last 50 years. Type 2 diabetes is closely associated with obesity. Up to 90% of the patients in the US with type 2 diabetes are either overweight or obese (2, 3) . It seems likely that the readily available high calorie food and a sedentary life style are major causes for obesity. Obesity contributes to insulin resistance and type 2 diabetes. Reducing obesity and stopping weight gain constitutes a way to slow down the rate of occurrence of type 2 diabetes. Type 2 diabetes is caused by insulin resistance, which is defined as defective insulin signaling and a decreased insulin efficiency to induce glucose transport from the blood into key target cells such as muscle and fat (adipocyte) cells (3) . In general, obesity leads to hyperglycemia, which in turn leads to and exacerbates insulin resistance. Insulin resistance, if not treated, results in hyperinsulinemia and eventually leads to full blown type 2 diabetes (3, 4) . Obesity or excessive adiposity, particularly visceral adiposity, contributes to and worsens insulin resistance (2, 5) . Most antidiabetic drugs are hypoglycemic or anti-hyperglycemic (blood glucose level reducing). However, most of these drugs are, to different extents, weight gain promoting (adipogenic) (6, 7) . Thus, these drugs treat one of the key symptoms of type 2 diabetes, hyperglycemia, but exacerbate the condition of being overweight or obese, one of the leading causes of type 2 diabetes. Therefore, while these drugs are beneficial over the short term, they are not optimal for long term health of type 2 diabetic patient. The most desirable situation would be the development of new types of antidiabetic drugs that are either hypoglycemic or anti-hyperglycemic without the side effect of promoting weight gain (adiposity). Herbal medicines known to be useful in diabetes treatment may be able to lead to compounds with such a combination of ideal therapeutic properties (8) (9) (10) (11) (12) (13) . Lagerstroemia speciosa (Fig. 1) , also called banaba in the Tagalog language of the Philippines, is a tropical plant found in many parts of Southeast Asia including the Philippines, Vietnam, Malaysia and southern China. It is a tree that can grow as tall as 20 m. Despite growing in several countries, only in the Philippines are the dried and shredded banaba leaves known to be used as a treatment for diabetes and kidney disease. It is not clear if banaba plants grown in different countries are equally effective in the treatment of diabetes. Garcia published the first research on banaba's insulinlike, hypoglycemic effect as early as 1940 (14) (15) (16) (17) (18) . Later, the popular use of banaba in the Philippines was noticed and led to its introduction in Japan. However, it was not until 50 years after Garcia's first publication that scientific interest in banaba's potential for the treatment of diabetes resurfaced. Scientists from countries including Japan, the Philippines, Korea and the United States are currently studying banaba. Lagerstroemia speciosa (banaba) has become relatively popular in the form of health-promoting tea products in Eastern Asia and the United States. In 1996, Kakuda et al. (19) studied banaba's antidiabetic activity by preparing water and methanol extracts of the plant. After feeding the extracts to hereditary type 2 diabetic KK-Ay/Ta Jcl mice, they found that food containing either 5% of water extract (BE) or 3% of methanol extract was effective in reducing blood glucose and insulin levels (P50.05) (19) . It is interesting to note that the total cholesterol in treated mice was also significantly reduced, but the plasma triglyceride level remained unchanged (19) . In a second study by Kakuda's research group, food containing 5% banaba water extract was used to feed female obese KK-Ay/Ta Jcl mice. Obese mice treated with banaba extract had a significantly reduced body weight (10%) compared with control mice fed with a regular diet (20) . No change in food intake was observed (20) . Interestingly, it was also discovered that liver triglyceride content was reduced by more than 40% in the banaba extract-treated mice. In addition, the parametrial adipose tissue was 10% lighter (P50.01) (20) . However, as in the previous antidiabetic activity study of BE in animals (19) , both the identity of the effective component(s) and the mechanism for the activity were not studied. Nevertheless, these two studies clearly demonstrated the in vivo antidiabetic and anti-weight gaining efficacy of the extract. In 1993, a group of scientists from Hiroshima University used an Ehrlich ascites tumor cell line coupled with a bioassay guided fractionation to screen compounds isolated by HPLC from banaba extract in order to identify the effective antidiabetic component (21) . Corosolic acid (2a-hydroxyursoloic acid) was identified as the effective compound in the methanol extract of banaba leaves in a glucose uptake assay ( Fig. 2A, 21) . However, this result should be considered with caution since the tumor cell line used in this study is a very unusual and unconventional cell line for diabetes studies or antidiabetic compound screening. Furthermore, the result could not explain the discrepancy that both the banaba water extract and the methanol extract were active in antidiabetic and anti-obesity animal studies since corosolic acid only exists in the methanol extract (19) (20) (21) . Realizing the potential problems associated with the cell line selection and assay method, the researchers acknowledged in a later publication that corosolic acid 'could not represent the whole activity of the banaba extract' (22) . Consequently, the researchers switched their cell model from the original tumor cells to a natural cell target of insulin, adipocytes, in order to allow for the 'isolation and the identification of more active compounds using the improved methodology' (22) . After HPLC purification, ellagitannins were identified in the water extract of banaba as the activators of glucose transport in fat cells with a glucose uptake assay (22) . One of the most potent ellagitannins was named 'Lagerstroemin' (Fig. 2, 22) . In a recent study, the same group reported the activation of the insulin receptor (IR) by Lagerstroemin (Fig. 2, 23) . In this study, Lagerstroemin was able to induce phosphorylation of the b-subunit of IR at 150 mM (23) but the mechanism responsible for IR activation was not found. The researchers speculated that Lagerstroemin could act intracellularly or bind to the insulin receptor (IR) extracellularly. In 2004, a separate group of researchers from several universities found that glucose transporter 4 (GLUT4) translocation from the intracellular microsomal membrane to the plasma membrane was significantly increased in the muscle cells of mice treated orally with corosolic acid (P50.05, 24). This result is both interesting and puzzling. GLUT4 is the major glucose transporter protein in both muscle and adipocytes and GLUT4 is insulin-responsive (4, 25) . However, since corosolic acid does not possess insulin-like glucose transport stimulatory activity, the process that leads to GLUT4 translocation is not known. The GLUT4 membrane translocation mechanism initiated by corosolic acid as described (24) must be independent of the IR mediated signaling pathway since corosolic acid does not use this pathway for its activity. In 2006, the same group of researchers found that corosolic acid significantly reduces blood glucose levels and plasma insulin levels in KK-Ay diabetic mice (26) . Additionally, this group of researchers published another article in 2006 in which they showed that corosolic acid significantly lowered blood glucose levels at the 90 min mark in an oral glucose tolerance test done on type 2 diabetic patients (27) . However, no statistical difference between the treated group and the control group was found for any other time point during the test (27) . Our interest in the isolation and identification of antidiabetic compounds from natural sources initiated our investigation of banaba extracts with 3T3-L1 adipocytes as a cell model and a radioactive glucose uptake assay as a screening method for the identification of potential antidiabetic compounds. In our study, we confirmed that the water and methanol extracts of banaba leaves exhibit glucose transport stimulatory activity (28) . In addition, we showed that the activity induced by the extract had a concentration-activity profile similar to that of insulin, suggesting that the activity of banaba extract (BE) might be triggered via an insulin-like mechanism (28) . Furthermore, we demonstrated that the same BE also inhibits adipocyte differentiation (28) . 3T3-L1 preadipocytes treated with methylisobutylxanthine, dexamethasone, and insulin (MDI) and BE did not differentiate into adipocytes as they normally do under the sole influence of MDI. This result indicated that BE inhibits the adipocyte differentiation activity induced by MDI. Unlike the glucose transport stimulation, the anti-differentiation activity is anti-insulin like, since insulin plays an important role in adipocyte differentiation and is a component of MDI ('I' in MDI stands for Insulin). Thus, BE differs from insulin in that it is anti-adipogenic whereas insulin is adipogenic, which is considered to be a negative side effect of insulin. In order to identify the component(s) in BE responsible for the antidiabetic activity, our first goal was to confirm that corosolic acid was responsible for the glucose transport stimulatory activity. Since tannins comprise up to 40% of the material in the extract, we were interested in separating it from corosolic acid first. To our surprise, after removing tannins by either gelatin or bovine serum albumin tannin precipitations (29), the remaining extract was not able to induce glucose transport. We concluded that the glucose transport activity was caused by the tannin component of the extract, and not corosolic acid (29) . Furthermore, we tested pure corosolic acid and found that it was not able to stimulate glucose transport in our cell model (Fig. 3) . Although we cannot exclude the possibility that corosolic acid may have some antidiabetic activity, we can eliminate corosolic acid from having the insulin-like glucose transport stimulatory activity found in adipocytes. It should be noted that banaba extracts prepared from banaba leaves from different sources may have different chemical compositions, which in turn may lead to different experimental results (our lab used dried banaba leaves from the Philippines). After testing corosolic acid, we also tested the corosolic acid-based banaba extract product glucosol (30) . The glucose uptake assay revealed that glucosol was activating glucose transport. Its activity was dose-dependent (Fig. 4) . However, since corosolic acid itself does not have any insulin receptor mediated glucose transport activity, the effect of glucosol is likely to originate from other chemical compounds in the banaba extract and not from corosolic acid. Tannins comprise a large and diverse class of polyphenolic compounds (31, 32) . Our search for the active components in the tannin fraction of BE was made much easier after we discovered that commercially available tannic acid (TA) shows similar glucose transport stimulatory activity to BE (29) . The main components of TA are gallotannins, a subclass of the tannins usually consisting of a glucose core connected to a variable number of galloyl groups via ester bonds (Fig. 5) . TA is known to exhibit various health-beneficial activities (33) (34) (35) (36) . As a constituent of red wine, it has been shown to effectively reduce blood glucose levels in type 2 diabetes patients (37) and the production of endothelin-1 (38), a key protein factor intimately involved in the development of cardiovascular disease (38) . In our study, TA was found to be much more potent and efficacious than the ellagitannin Lagerstroemin (29) . Therefore, TA was chosen as the focus of our research for isolating active compounds from the tannin fraction Figure 3 . Absence of glucose transport stimulatory activity of corosolic acid in adipocytes. Pure corosolic acid in aqueous solution was added to 3T3-L1 adipocytes grown in wells of 6-well cell culture plates to induce glucose transport by a commonly used procedure (28, 29) . Cells treated with either 1 mM insulin or 30 mM penta-galloyl-glucose (PGG) were used as positive controls. Cells treated with water vehicle served as negative controls. Samples were in triplicate in the experiment, and the experiment was repeated three times. No difference was found by a one-way ANOVA statistical analysis between the negative (vehicle) control and corosolic acid samples at any concentration. of BE. Components of TA were separated by HPLC, and active fractions were identified with a glucose uptake assay in 3T3-L1 adipocytes (29) . The study led to the discovery of penta-O-galloyl-D-glucopyranose (PGG) as the most effective compound in TA (Fig. 5, 39 ). Both anomers of PGG occur in nature (40) . The a-anomer was found to be slightly more active than the b-form in its glucose transport stimulatory activity (39) . Corosolic Acid and Lagerstroemin versus Tannic Acid and PGG: Which are More Bioactive? In the ellagitannin study mentioned earlier, Lagerstroemin exhibits glucose transport stimulation at 40 mM with an EC 50 of 80 mM (22) . In comparison, aand b-PGG exhibit activity at a concentration as low as 10 mM with EC 50 of 17 and 18 mM (39) . In other words, aand b-PGG are about five times more potent than Lagerstroemin in stimulating glucose transport. We would like to emphasize that both aand b-PGG possess the adipogenesis inhibitory activity exhibited by the banaba extract (28) and by TA (29) . This suggests that these two activities are associated with the same molecules. It also indicates the functional differences between insulin (glucose transport inducing and adipogenic) and PGG (glucose transport inducing but anti-adipogenic). Both compounds may have the potential to reduce hyperglycemia without increasing adiposity, a very desirable combination of properties that insulin lacks. The Lagerstroemin study also showed that its glucose transport inducing activity is about 54% of that of insulin (22) . In comparison, aand b-PGG showed 60-70% of insulin's glucose transport inducing activity (39) . . Glucose transport stimulatory activity of glucosol as compared with banana extract. Banaba extract (BE) was prepared in house (28) . Glucosol (30) was purchased. Glucosol was compared with BE in a regular glucose uptake assay (28, 29) . Samples were assayed in triplicates, and the assay was repeated three times. The result of the assays was analyzed with a one-way ANOVA. *P50.05, **P50.01, ***P50.001. All samples were compared with the negative (vehicle) control. Thus, aand b-PGG are at least 30% more effective than Lagerstroemin (Table 1 ). It is interesting that PGG possesses many other health-beneficial bioactivities, such as anticancer (41, 42) , anti-inflammation (43, 44) , antivirus (anti-HIV, 45; anti-SARS, 46) and antioxidant activity (47, 48) . From the known studies, we conclude that tannin molecules are responsible for the insulin-like glucose transport stimulatory activity of the banaba extract. Gallotannins such as PGG appear to be more potent and efficacious than ellagitanins such as Lagerstroemin in IR binding, IR activation and glucose transport induction. Corosolic acid does not possess any insulin-like glucose transport stimulatory activity. If its antidiabetic activity can be confirmed, it is likely to be induced via a noninsulin like, indirect mechanism. Although tannins were identified as the effective component for the insulin like glucose transport inducing activity in banaba extract, the most effective tannin molecule in the extract has not been identified. A well designed bioassay (glucose uptake assay) guided isolation should be able to complete this task. In the past 10 years, other studies regarding banaba extracts or chemicals derived from banaba extracts were reported (49) (50) (51) (52) . These studies indicate that banaba extracts contain interesting biomedical substances that have attracted significant scientific attention. More detailed studies at molecular and cellular levels as well as in animal models are required to elucidate banaba extract's antidiabetic activity and other health beneficial activities such as its anti-adipogenesis activity. Corosolic acid -0 Insulin 1 nM 156 *Activity of b-PGG was arbitrarily assigned as 100.

Transmissibility of the Influenza Virus in the 1918 Pandemic

BACKGROUND: With a heightened increase in concern for an influenza pandemic we sought to better understand the 1918 Influenza pandemic, the most devastating epidemic of the previous century. METHODOLOGY/PRINCIPAL FINDINGS: We use data from several communities in Maryland, USA as well as two ships that experienced well-documented outbreaks of influenza in 1918. Using a likelihood-based method and a nonparametric method, we estimate the serial interval and reproductive number throughout the course of each outbreak. This analysis shows the basic reproductive number to be slightly lower in the Maryland communities (between 1.34 and 3.21) than for the enclosed populations on the ships (R(0) = 4.97, SE = 3.31). Additionally the effective reproductive number declined to sub epidemic levels more quickly on the ships (within around 10 days) than in the communities (within 30–40 days). The mean serial interval for the ships was consistent (3.33, SE = 5.96 and 3.81, SE = 3.69), while the serial intervals in the communities varied substantially (between 2.83, SE = 0.53 and 8.28, SE = 951.95). CONCLUSIONS/SIGNIFICANCE: These results illustrate the importance of considering the population dynamics when making statements about the epidemiological parameters of Influenza. The methods that we employ for estimation of the reproductive numbers and the serial interval can be easily replicated in other populations and with other diseases.

The emergence of the highly pathogenic avian influenza strain H5N1 has raised concerns of an imminent influenza pandemic. Public health workers, government officials and disaster planners have an increasing interest in better understanding the potential impact of an influenza pandemic and possible strategies for containment. Crucial in this planning is an understanding of the basic epidemiology of the disease in various settings. This has led to a growing interest in the analysis and understanding of past epidemics, particularly that of 1918, the most virulent and deadly influenza epidemic of the 20th century. Mortality has been estimated at 50-100 million people worldwide as a result of influenza in the 1918 pandemic [1] . It is reasonable to suppose that by better understanding the transmission dynamics of the highly pathogenic virus in 1918, we can gain greater insight into the dynamics, and thus potential methods of control, for a future pandemic [2] . Important parameters for understanding disease transmission are the reproductive number and the serial interval [3] . The basic reproductive number is defined as the average number of secondary infections created from a primary infection in an entirely susceptible population [4, see also 5] . A more complex, but perhaps meaningful parameter is the effective reproductive number which defines the average number of secondary infections an infected will create at a given point in the epidemic. This parameter takes into account that not all contacts of an infected individual are with susceptible persons, as well as the impact of public health control measures. Control strategies are typically targeted to drive this number below one and maintain it there, as this will lead to eventual extinction of the epidemic. An example of this is herd immunity, or immunity to a disease that is incurred from a sufficiently large proportion of the population being immune to a disease. Modeling techniques are often used to determine the proportion of the population that should be vaccinated in order to have the reproductive number low enough to avoid outbreaks of disease [6] . The serial interval can be defined as the time interval between a primary case presenting with symptoms and its infectee developing symptoms [7, 8] . Thus this quantity is completely observable. This is a mixture of the incubation period and the infectious period, both of which are useful to understand, but difficult to measure. The SARS outbreak of 2003 had a relatively long serial interval, estimated to be between 8 and 10 days on average and following a Weibull distribution [9] making case isolation extremely effective in containing the epidemic. Methods for the estimation of basic epidemiological parameters are still in development phase. [10] provides a thoughtful summary of methods for estimating the reproductive number. One particularly interesting and useful method has been previously described by [7] for estimating the daily reproductive number, R t , or the average number of cases an infected individual on day t would cause. One interesting feature of this method is that for days where no cases are observed, the estimated effective reproductive number is zero. Another observation is that this method essentially estimates a curve for the effective reproductive number that traces the epidemic curve, lagged by the average serial interval length. This nonparametric method presupposes information on the serial interval distribution. This is typical as most methods for estimating the reproductive number rely on knowledge of the serial interval. Few have described analytical methods for estimating the serial interval, making most methodologies dependent on contact tracing data, which is often difficult and expensive to attain. [11] describe a method to estimate the reproductive number that relies on limited contact tracing information but not a full estimate of the serial interval. [12] have recently described a method to estimate the serial interval and then used this estimate with the estimator proposed in [7] of the daily reproductive number and have applied their method to data from outbreaks of avian influenza in poultry farms in Europe. Several researchers have studied the 1918 pandemic and estimated some of these key epidemiological parameters. Estimates have ranged from 2-3 for the basic reproductive number, R 0 , when using an SEIR model with a mean latent period of 1.9 days and infectious period of 4.1 days [13, 14] . Using an exponential model and assuming the serial interval to be four days (somewhat based on the assumptions of [13] ), [15] estimated R 0 to be 2.6-10.6 for confined settings (such as prison and ships) and 2.4-4.3 for community settings. The estimates for the mean latent and infectious periods come from [16] and were used again by [17] and [18] . It appears that the original estimates were derived from epidemic data, although their source is not well documented. In what follows, we introduce new methodology for the estimation of both the daily reproductive number and the serial interval. We apply this method to data from two outbreaks on military ships in the 1918 influenza outbreak, as well as welldocumented outbreaks in five Maryland communities. The results from this method are compared to that of [12] . The results illustrate the differences in infectious disease dynamics between outbreaks in a closed population and a dynamic community. We analyze data from several well-documented influenza outbreaks in 1918. First we consider data from two troop ships that embarked in the late fall of 1918 [19] . The Medic reported two initial cases on November 11. Out of 989 passengers (156 crew members, 829 soldiers, 4 civilians) 313 became sick with influenza over a 40 day period (Attack Rate, AR, = 0.32), though most of the cases occurred within the first fourteen days. The Boonah left Durban and in five days, on November 29, reported the first three definitive cases of influenza. Those who collected the data note that there were likely some initial cases that were not identified. Out of 1095 on board (164 crew members and 931 troops), 470 cases were reported (AR = 0.43) in the 40 days of the epidemic. The United States Public Health Service created special surveys of 18 localities during the pandemic [20] . Reported results from six communities in Maryland are derived from house-to-house surveys requesting the date of onset of influenza for all infected, and the sex and age of each case of pneumonia and influenza. A summary of these populations is provided in Table 1 . We describe a likelihood based methodology for estimating the reproductive number at each day in the epidemic as well as the serial interval. The method builds on that described by [21] . We assume that the population is closed, that all cases are observed, and use daily case counts only (i.e. number of new cases each day). Let N = {N 0 , N 1 , N 2 ,…, N T } represent the daily cases counts of influenza for the T days of the epidemic and X ij represent the number of cases that appear on day j that are infected by individuals that appeared sick on day i. Following is a representation of the disease transmission model in the population. . . We assume that the total number of cases produced by those on day i, X i? , are Poisson distributed with parameter N i R i , where R i is the reproductive number for cases on day i. We further assume that X i = {X i,i+1 , X i,i+2 ,…,X i,i+k } follows a multinomial distribution with parameters X i? , p, k, where p = {p 1, p 2,…, p k } represent the distribution of the serial interval. Using these assumptions we can construct a likelihood function (see details in the Supplemental Information), which, when simplified, yields the following convenient form where m i~Ri ( X k j~1 p j N i{j ) [21] . Maximization of this likelihood with respect to R i and p yields estimates of these parameters. To further simplify this process and create a more parsimonious model, we parameterize p by allowing it to follow a traditional parametric form for a serial interval (for instance a Weibull, Gamma, Log Normal, or Exponential distribution). Then the p j are functions of the parameters of the density (for instance in the case of the Gamma distribution, the p j only depend on the shape and rate parameters of the Gamma). Similarly R i can be modeled parametrically as a function of time. One example of a reasonable model for this is the four parameter logistic curve [22] [23] [24] given by The parameters of this curve describe the initial height of the curve (approximately a+b), the point of inflection (d), the curvature over the inflection (c) and the final height of the curve (a). These parameters have biological meaning in this setting where the initial height corresponds to the values of R i prior to intervention and significant depletion of the susceptible population. The inflection point and its steepness would describe the timing of intervention and the rapidity with which it impacts transmission. The final height would describe the ultimate value of R i, which typically is less than one, indicating that disease transmission is in a sub epidemic state. In our analysis, we also implement the method described by [12] (hereafter referred to as the Garske et al. method) and compare the results of the two methodologies. This method first estimates the generation time distribution using a likelihood based method. Then the effective reproductive number is estimated using the method described by [7] (hereafter referred to as the WT method). We fit the likelihood for both methods using a Nelder-Mead maximization procedure and use 576 starting values in order to ensure that we reach the global maximum. All analyses were done using R 2.4.1. Both methods assume homogenous mixing in the population, no missing data (clearly violated with the data from the Maryland communities), that a primary case experiences symptom onset prior to any cases that it infects and a completely closed system where all cases are infected by a case that has been observed. In the case of the Maryland data, where only a sample of the total number of cases was surveyed, we can observe the efficacy and robustness of these methods with sample data. Certainly results should be interpreted with caution, however, as we will show, the results that are obtained are consistent with previous estimates for influenza. Standard errors were calculated for the MLE method using a parametric bootstrap. One thousand epidemics were simulated using the parameter estimates and estimates were obtained from each of these simulated epidemics. The standard deviation of the 1000 estimates was used as the standard error estimates. We used the method described in [12] to estimate the standard error for their estimates, however our simulations based on their assumption of asymptotic normality yielded a large number of negative estimates for the parameters. It is possible that this is due to the non-independence in the data and lack of theoretical underpinnings for the method that they propose. These results make their standard error estimates infeasible to estimate in this case. Therefore we do not present standard error estimates for the results obtained using their methodology. In order to determine the accuracy and relative merit of the estimates obtained from each methodology, we compute one-stepahead residuals and implement a cross validation approach to analyze the generalizeability of the estimates obtained. The onestep-ahead residuals were calculated by first using the estimates from a particular location along with the data to predict the next days' number of cases,Ñ i as follows: EachÑ i is calculated using N 0 , N 1 , …, N i21 . Then the one-stepahead residuals are calculated as We present these residuals averaged over the T days observed. Generalizeability of the results was studied using an ad hoc cross validation (CV) technique. This is done by using the estimates obtained from one location to calculate the one step ahead residuals for another location. Specifically we use the Boonah ship estimates to calculate residuals with the Medic data and then use the Medic estimates to calculate the residuals for the Boonah data. For the Maryland communities, we report the average of the residuals obtained using the estimates from one community to predict the epidemics in each of the other four communities, creating five CV estimates (one for each community). Table 2 gives the results for the serial interval distribution estimates. Notable in these results is the striking consistency in the estimates of the first moment, with the exception of Cumberland. The second moments vary much more, however. In general they tend to be much larger for the ships when using the Garske et al. method compared to the MLE method. For the communities, we observe that they are consistently around 10 for the Garske et al. method and vary much more for the MLE method. Also of interest in these results are the large error estimates, particularly for Cumberland, but also to a smaller extent for Frederick. This is perhaps indicative of the model not fitting the data as well, for instance the logistic model may not be the best fit in this scenario, or that the lack of census data on cases might be more problematic here. In Table 3 and Figure 1 , we present the results for estimation of the effective reproductive number. Evident in these results, is the large initial reproductive number for the Boonah ship. This is likely due to some of the missing data at the beginning of the epidemic and thus the model attributing the large number of cases that rapidly develop to the few individuals who were initially reported. The logistic model fits this as accurately as possible, but perhaps the important message is the qualitative result, indicating that initial transmission in this susceptible, non-quarantined population was very high and rapidly decreased as many became infected. The result is similar for Medic though the initial value is not high. We also note that the reproductive number dropped to sub epidemic levels rapidly (around 10 days for both ships). In the Maryland communities the initial reproductive number tended to be slightly lower (ranging from 1.34 in Salisbury to 3.21 In Table 4 , we present the results of the residual analysis. We notice here that the Garske et al. method often does better than the MLE method. It is important to point out that the WT method of fitting the effective reproductive model over fits the model and suffers from generalizeability. This method essentially traces the epidemic curve, lagged by the mean of the generation time distribution. Thus, according to the residuals, it appears that the WT method outperforms the MLE. However, considering the importance of external validation and reproducibility, the model suffers somewhat as evidenced by the CV measures. The exceptions to this are in the case of the Boonah where the CV measure is impacted by the large initial MLE estimate of the reproductive number and in Cumberland where it appears that either the parametric model chosen may not represent the best fit to the data or there were sensitivities to the survey data. We have presented results that are informative with regard to the dynamics of the 1918 influenza pandemic in different populations and provide insight into two methodologies for estimating basic epidemiological parameters. Both methods assume that the population is closed, there are no missing cases and no migration to or from the population. The second of these assumptions is clearly violated with the data from Maryland; however the results appear to be reasonably robust to this discrepancy, except in the case of Cumberland. The purpose of this exercise determines to some extent which methodological approach we might favor. If the intent is to simply estimate the parameters for a specific epidemic and better understand what exactly was occurring in that setting, then the method presented by [12] (Garske et al.) appears to provide good fit. The caveat that we see in this method is that by estimating the effective reproductive number with the methodology of [7] (WT) there is an over fit of this parameter and it essentially traces the epidemic curve, lagged by the mean of the serial interval. It is not clear if this is a desirable or informative property. The MLE method has greater promise for generalizeability. While it can be argued that adhering to a parametric definition of the shape of the effective reproductive number leads to a greater chance of lack of fit, it can also lead to a result that can be interpretable for other settings that are similar to that being studied. One can choose any reasonable parametric form for modeling the effective reproductive number. Here we have only shown the four parameter logistic model, and feel that it is suitable in most cases where the epidemic curve has a single peak. It is feasible that this model may not apply well in all situations. Another approach might be to analyze the data using the Garske et al. method and then smooth the plot of the effective reproductive number and from this determine a parametric form that closely approximates the smoothed curve. Multiple models could be implemented, then the residual analysis that we have shown provides a valuable tool for model assessment and comparison. The results of these models can be sensitive to underreporting initially in the epidemic. We see this clearly in Boonah, where it was acknowledged that there was underreporting early on and this led to us getting very high estimates for the initial reproductive number. Similarly, in Cumberland, if we remove the first five days of data (three cases on the first day, six cases on the second and then no cases the following three days) we get much more reasonable estimates (m~6:00,ŝ 2~1 0:32) with smaller residuals (6.00). Therefore, it is important to note that unusual observations in the first few days can impact the estimates and one should pay careful attention to this possibility. Overall both methodologies presented are valuable tools that can be used in tandem for understanding the dynamics of infectious disease epidemics. These methods are easy to implement and interpret. The results that we have presented suggest that the average serial interval for pandemic influenza in 1918 was consistently between three and four, regardless of the setting. The standard deviation for the serial interval distribution varied much more for the MLE method depending on the location. Garske et al. estimates indicate that the value was consistently smaller in the communities than in the ships. It is not clear exactly how to interpret this result. Further, we consistently see a large initial value for the reproductive number. In the ships, this value is higher and rapidly drops off, perhaps due to the close quarters and extremely rapid transmission that could take place in these very vulnerable populations. In the communities, the reproductive number tended to drop off later, typically around day thirty. This could be due to a larger initial susceptible population and more complicated dynamics for the disease to spread, leaving large pockets of susceptible individuals unexposed for a longer period of time than in the ships.

Surfactant therapy for acute respiratory failure in children: a systematic review and meta-analysis

INTRODUCTION: Exogenous surfactant is used to treat acute respiratory failure in children, although the benefits and harms in this setting are not clear. The objective of the present systematic review is to assess the effect of exogenous pulmonary surfactant on all-cause mortality in children mechanically ventilated for acute respiratory failure. METHODS: We searched the MEDLINE, EMBASE, CINAHL and Ovid Healthstar databases, the bibliographies of included trials and review articles, conference proceedings and trial registries. We included prospective, randomized, controlled trials of pulmonary surfactant that enrolled intubated and mechanically ventilated children with acute respiratory failure. We excluded trials that exclusively enrolled neonates or patients with asthma. Two reviewers independently rated trials for inclusion, extracted data and assessed the methodologic quality. We quantitatively pooled the results of trials, where suitable, using a random effects model. RESULTS: Six trials randomizing 314 patients were included. Surfactant use reduced mortality (relative risk = 0.7, 95% confidence interval = 0.4 to 0.97, P = 0.04), was associated with increased ventilator-free days (weighted mean difference = 2.5 days, 95% confidence interval = 0.3 to 4.6 days, P = 0.02) and reduced the duration of ventilation (weighted mean difference = 2.3 days, 95% confidence interval = 0.1 to 4.4 days, P = 0.04). CONCLUSION: Surfactant use decreased mortality, was associated with more ventilator-free days and reduced the duration of ventilation. No serious adverse events were reported.

Acute respiratory failure remains the primary indication for admission to North American paediatric intensive care units (PICUs) and accounts for significant mortality, morbidity and resource utilization [1] . Respiratory infections, in particular pneumonia and severe bronchiolitis, are the most common causes of respiratory failure requiring mechanical ventilation in children [1] . Alterations in endogenous surfactant play a role in the pathogenesis of many causes of acute lung injury (ALI) and acute respiratory distress syndrome (ARDS) [2] . Surfactant dysfunction, destruction and inactivation have also been demonstrated in children with acute respiratory insufficiency due to bronchiolitis [3, 4] . The administration of exogenous surfactant may reduce the need for mechanical ventilation and its associated sequelae by restoring surfactant levels and function. Inspired by the success of surfactants in reducing mortality and the need for mechanical ventilation in neonatal respiratory distress syndrome [5] , investigators have studied exogenous surfactant in other populations with various causes of respiratory failure. Trials of surfactant in adults with ALI and ARDS have not demonstrated a mortality benefit [6] [7] [8] [9] , perhaps due to inherent differences in the aetiology of lung injury in adults, the design features of the trials, the mode and timing of surfactant administration or the type and dose of surfactant used. In children with respiratory failure, the efficacy of exogenous surfactant has been suggested in uncontrolled studies ALI = acute lung injury; ARDS = acute respiratory distress syndrome; FiO 2 = fractional inspired oxygen; PaO 2 = arterial oxygen tension; PICU = paediatric intensive care unit; RSV = respiratory syncytial virus. (page number not for citation purposes) [10, 11] . The relatively low mortality rate, the diversity of the study populations and the shorter duration of mechanical ventilation are factors that make large-scale randomized controlled trials in this population challenging to conduct. Two of the largest trials were stopped early due to slower than expected enrolment [12, 13] . While the use of surfactant in ARDS/ALI has not been previously systematically reviewed, its use in children with bronchiolitis has been [14] . We anticipated that including trials enrolling children with acute respiratory failure from a variety of causes would result in a heterogeneous population and would increase the generalizability of the results. Our confidence in the results of the present review would also be increased if a consistent effect is shown in subgroups and across a spectrum of disease severity. The primary objective of the systematic review is to assess the effect of the administration of pulmonary surfactant compared with no therapy or with placebo on all-cause mortality (at or before hospital discharge) in mechanically ventilated children with acute respiratory failure. We included trials that were prospective, that were randomized, that enrolled children intubated and mechanically ventilated for acute respiratory failure and that compared the intratracheal administration or nebulization of at least one dose of natural or artificial pulmonary surfactant with a placebo or no intervention. We excluded trials exclusively enrolling neonates or patients with asthma. We used the trial authors' definitions of paediatric. The primary outcome measure was all-cause mortality at or before hospital discharge. Secondary outcomes were ventilator-free days to day 28 (a composite of mortality and duration of ventilation, defined as days alive and free from mechanical ventilation) [15] , the duration of mechanical ventilation (from intubation to extubation, death or trial withdrawal), the duration of PICU stay, the use of rescue therapy (such as extracorporeal membrane oxygenation, high-frequency oscillatory ventilation, open label surfactant and nitric oxide), and complications and adverse effects as reported by the trial authors. One of us searched for published and unpublished trials, examining trial registries, conference proceedings and the bibliographies of any identified trials and relevant reviews (the search strategy is available upon request). We polled paediatric intensivists and pharmacists at our institution for additional trials. We selected search terms from the keywords and MESH terms of previous surfactant trials and from the generic and brand names of commercially available surfactants. We imposed no language restrictions. One of us screened the title (and abstract if required) of all citations retrieved. We selected citations for further evaluation if they reported the administration of at least one dose of surfactant to at least one child or if the title or abstract did not give enough information to make an assessment. Two reviewers independently reviewed all citations meeting criteria for further review and applied the inclusion criteria. Disagreements between reviewers were resolved by consensus in consultation with a third reviewer. We considered agreement between reviewers to be acceptable if the kappa value was greater than 0.8. We used the following characteristics to assess the methodologic quality: allocation concealment (sealed envelopes or central randomization were considered adequate), blinding (which of the trial personnel and caregivers were blinded, and the methods used to ensure blinding), completeness of followup (assessed by the number of patients randomized for whom there were no outcomes), similarity of the groups at baseline (with respect to known prognostic factors: age, aetiology, severity of illness as measured by the Pediatric Risk of Mortality score, and immunosuppression), whether a standard or recommended strategy for mechanical ventilation was used, and whether a priori criteria for the use of co-interventions were used. Effective blinding of surfactant is challenging because of the large volumes of milky fluid administered, which can often be seen by caregivers in the patients' ventilator tubing or endotracheal tube, particularly during suctioning. We pretested and refined the developed forms on two trials of surfactant therapy for adults, and clarified definitions based on feedback from the reviewers. Two reviewers then independently used these forms to abstract trial quality, blinded to the authors, the journal, the country of origin and the results. We resolved any disagreements by consensus in consultation with a third reviewer if needed. After pretesting and refining the forms on two trials of surfactant therapy in adults and clarifying definitions based on feedback from the reviewers, two reviewers then independently abstracted the data. Reviewers were only provided with a full-text version of the trials from which the introduction, conclusions and discussion were omitted and from which the author, journal and country of origin were deleted. We thereafter examined these sections of the reports for any missing data. We resolved any disagreements between reviewers by consensus in consultation with a third reviewer if needed. We asked the authors to supply data not included in the published reports. Two reviewers performed data entry in duplicate. We quantitatively pooled the results of individual trials when possible. We expressed the treatment effect as a relative risk for dichotomous outcomes and as a weighted mean difference for continuous outcomes with 95% confidence intervals. We considered effects statistically significant if P < 0.05. A z test was used to statistically test the estimates of treatment effect between groups [16] . We assessed heterogeneity among trials using the I 2 statistic, and considered an I 2 value greater than 50% to indicate substantial heterogeneity [17] . RevMan 4.2 software and a random effects model were used to perform the analyses [18] . We chose the random effects model because it gives a more conservative estimate of the precision of the treatment effects and because the true effect of the intervention probably varies given the different populations enrolled in these trials [19] . A subgroup analysis was planned based on the aetiology of respiratory failure (trials enrolling exclusively patients with respiratory syncytial virus (RSV)/ severe bronchiolitis compared with all other trials) if sufficient data were available, because these trials were likely to enrol a younger, more homogeneous, population with a lower predicted risk of mortality. We also planned sensitivity analysis based on methodological features of the included trials (trials reporting adequate allocation concealment compared with all other trials). We identified 742 unique citations, six of which met our inclusion criteria ( Figure 1 outlines the reasons for exclusion). Most reports excluded enrolled neonates or were retrospective or uncontrolled in design. Chance corrected agreement was excellent (kappa = 0.91, 95% confidence interval = 0.73-1.1). Table 1 presents a complete description of our quality assessment. Only one trial did not report allocation concealment [20] . Although effective blinding of surfactant is challenging, two trials reported blinding of the PICU team [12, 20] . The two Flow diagram of included trials Flow diagram of included trials. RCTs, randomized controlled trials. groups were generally well matched in terms of baseline characteristics in most trials. The most significant imbalance was the numerically higher number of immunosuppressed patients in the placebo group. These patients had higher mortality (56%) than the immunocompetent group (13%). The authors attempted to adjust for this imbalance with logistic regression, which suggested that the treatment effect seemed to be relatively consistent between the two groups [12] . Only one trial reported a priori criteria for rescue therapy [13] . Table 2 describes the included trials. Three trials enrolled exclusively infants with RSV-induced respiratory failure [20, 21] or with severe bronchiolitis [22] . The remaining three trials enrolled a heterogeneous group of patients with ARDS or ALI [12, 23, 24] . While the individual treatment protocols varied, all trials used comparable doses (50-100 mg/kg phospholipids) of natural or modified natural surfactants and each patient typically received one or two doses. A variety of interventions were used in the control groups: no intervention, air placebo or similar sedation and ventilation manoeuvres without a placebo. Although one study [20] used a modified natural surfactant, all the products used contained surfactant proteins B and C. All studies administered surfactant early in the course of respiratory failure; most patients were treated within 12-48 hours of requiring mechanical ventilation. The baseline characteristics of the patients are presented in Table 3 . While there was significant heterogeneity among and within trials with respect to age and cause of respiratory failure, we considered the initial Pediatric Risk of Mortality scores and the initial PaO 2 /FiO 2 ratios to be clinically comparable. Mortality data were available for all six trials, randomizing 311 patients and reporting data for 305 patients. There were no deaths reported in the three RSV/severe bronchiolitis trials; thus our estimate is based on three trials randomizing 232 patients, 64 of whom died. In the pooled analysis, surfactant was associated with significantly lower mortality (relative risk = 0.7, 95% confidence interval = 0.4-0.97, P = 0.04). There was no evidence of heterogeneity (I 2 = 0%) ( Figure 2 ). Ventilator-free days to day 28 The number of ventilator-free days to day 28 was available for six trials randomizing 311 patients and reporting data for 305 patients. In the pooled analysis, surfactant was associated with significantly more ventilator-free days (weighted mean dif- (Figure 3 ). The duration of mechanical ventilation was available for six trials randomizing 311 patients and reporting data for 305 patients. In the pooled analysis, surfactant was associated with a significantly shorter duration of mechanical ventilation (weighted mean difference = 2.3 days, 95% confidence interval = 0.1-4.4 days, P = 0.04) (Figure 4) . The duration of PICU stay was available for five trials randomizing 273 patients and reporting data for 272 patients. In the pooled analysis, surfactant was associated with a shortened duration of PICU stay (weighted mean difference = 2.6 days, 95% confidence interval = 0.02-5.2 days, P = 0.05), but this difference was not statistically significant ( Figure 5 ). Data on the use of rescue therapy were available for six trials randomizing 311 patients and reporting data for 305 patients. In the pooled analysis, the surfactant was associated with a significantly lower use of rescue therapy (relative risk = 0.4, 95% confidence interval = 0.3-0.7, P < 0.0001). There was no evidence of heterogeneity (I 2 = 0%). This summary estimate should be interpreted with caution as only one trial reported a protocol for initiating rescue therapy. The decision to use a rescue therapy, particularly an open-label surfactant, may be influenced by knowledge of the patient's allocation; furthermore, only two trials reported blinded caregivers and the methods used to ensure blinding may not be adequate. Surfactant therapy was well tolerated (see Table 4 ), but only three of the trials reported any definitions or a priori criteria or of collecting adverse events [12, 21, 23] . Transient hypotension and transient hypoxia were the most commonly reported adverse events in the largest trial. These responded to a brief adjustment in ventilation, to a slowing of the rate of surfactant administration or to fluid administration. There was no difference in the incidence of air leaks in the two trials that reported this outcome. No patient was withdrawn from any of the trials because of adverse events. We did not pool the data on adverse events associated with the trial interventions from the six trials because of the inconsistent manner in which the events were documented and reported. The effect of surfactant on ventilator-free days, the duration of mechanical ventilation and the duration of PICU stay was not significantly different when we compared the three trials that enrolled exclusively patients with RSV/severe bronchiolitis with the three other trials (Table 5) . A 100% survival in the bronchiolitis trials subgroup precludes formal subgroup analysis for the primary outcome of mortality. All but one of the included trials reported adequate allocation concealment (defined as sealed envelopes or central telephone randomization). Since there were no deaths in this trial we could not assess the effect of inadequate allocation concealment on mortality. Pooling the five remaining trials did not change the direction of the effect and did not significantly Meta-analysis of trials of surfactant in children with acute respiratory failure: Mortality Meta-analysis of trials of surfactant in children with acute respiratory failure: Mortality. ALI, acute lung injury; ARDS, acute respiratory distress syndrome; 95% CI, 95% confidence interval; RR, relative risk; RSV, respiratory syncytial virus. Meta-analysis of trials of surfactant in children with acute respiratory failure: Ventilator-free days Meta-analysis of trials of surfactant in children with acute respiratory failure: Ventilator-free days. ALI, acute lung injury; ARDS, acute respiratory distress syndrome; 95% CI, 95% confidence interval; RSV, respiratory syncytial virus; SD, standard deviation; WMD, weighted mean difference. change the point estimates for the secondary outcomes of ventilator-free days, duration of ventilation or duration of PICU stay (Table 6 ). In the present systematic review and meta-analysis of the effect of surfactant for critically ill children with acute respiratory failure we found that surfactant therapy significantly reduced our primary outcome of mortality. Surfactant was associated with more ventilator-free days, with decreased duration of ventilation and with less use of rescue therapy as compared with standard therapy. There was no significant difference in the duration of PICU stay. Surfactant therapy was well tolerated; while transient hypoxia and hypotension were reported during surfactant administration, no study reported any serious adverse events. The patients enrolled in these trials are representative of the heterogeneous group of children with early, severe acute respiratory failure that is seen in clinical practice. These patients had similar severity of illness scores and a similar degree of respiratory failure (as measured by Pediatric Risk of Mortality scores and PaO 2 :FiO 2 ratios). The heterogeneity of results for our primary outcome of mortality was low. The presence of significant heterogeneity reduces the strength of inferences we can make regarding the effect of surfactant on the secondary outcomes of ventilator-free days, Meta-analysis of trials of surfactant in children with acute respiratory failure: Duration of mechanical ventilation Meta-analysis of trials of surfactant in children with acute respiratory failure: Duration of mechanical ventilation. ALI, acute lung injury; ARDS, acute respiratory distress syndrome; 95% CI, 95% confidence interval; RSV, respiratory syncytial virus; SD, standard deviation; WMD, weighted mean difference. Meta-analysis of trials of surfactant in children with acute respiratory failure: Duration of PICU stay Meta-analysis of trials of surfactant in children with acute respiratory failure: Duration of PICU stay. ALI, acute lung injury; ARDS, acute respiratory distress syndrome; 95% CI, 95% confidence interval; PICU, paediatric intensive care unit; RSV, respiratory syncytial virus; SD, standard deviation; WMD, weighted mean difference. duration of ventilation and duration of PICU stay. Separately pooling the trials that exclusively enrolled patients with RSV/ severe bronchiolitis and those enrolling patients with ARDS/ ALI from a variety of causes did not significantly reduce the heterogeneity. Changing ventilation strategies and the use of a variety of natural and modified natural surfactants may have increased the heterogeneity of our results. Ventilation strategies, such as the use of lower tidal volumes and earlier use of high-frequency oscillatory ventilation, have evolved significantly in the 10-year span over which the included trials were conducted [25] [26] [27] . The surfactants used in the included trials were all natural or modified natural surfactants; however, these surfactants may have slightly different effects on oxygenation and compliance due to the differences in phospholipid and surfactant protein composition, which may have influenced individual study results. The strengths of the present review include a comprehensive search strategy, broad inclusion criteria (resulting in a representative, heterogeneous population) and abstraction of clinically important outcomes in duplicate, independently blinded to information that may bias evaluation. The strength of the inference we can make from our subgroup analysis is limited because we were unable to extract all subgroup data from these trials. Access to individual patient data would allow better examination of the treatment effect in subgroups of patients and would facilitate further exploration of possible causes of heterogeneity. We found that mortality was very different between the trials that exclusively enrolled patients with RSV/severe bronchiolitis and those that enrolled patients with ARDS/ALI from a variety of causes. We pooled the results because both conditions result in abnormal surfactant function and because of the substantial overlap between the two groups; up to 17% of children in the ARDS/ALI trials had RSV and up to 50% of the children in some bronchiolitis studies also had pneumonia. The reduction in mortality and the increased ventilator-free days have important implications as very few trials in paediatric critical care suggest a favourable impact on mortality [28] . The present review suggests that surfactant could be an important adjunct in the management of paediatric respiratory failure. Uncertainty exists, however, about the reproducibility of treatment effects generated from relatively small unblinded trials; questions remain about adverse affects, which may be undetected or under-reported in this literature. Also, a large proportion of patients and events are reported in one trial [12] . Furthermore, issues of the optimal dose and the timing of administration, and which patients are most likely to derive benefit, should be studied in further adequately powered multicentre trials. The Pediatric Acute Lung Injury and Sepsis Investigators network is planning a large rigorous randomized trial enrolling children with acute hypoxemic respiratory failure to address these issues. Surfactant use decreased mortality, was associated with more ventilator-free days and reduced the duration of ventilation. No serious adverse events were reported. Most trials enrolled small numbers of children, and further well-designed and adequately powered multicentre trials are therefore required. • Surfactant decreased mortality in a heterogeneous population of children with acute respiratory failure. • Surfactant was associated with more ventilator-free days and a reduced duration of ventilation. • No serious adverse events were reported. • Further well-designed and adequately powered multicentre trials are required.

Clinical review: Update of avian influenza A infections in humans

Influenza A viruses have a wide host range for infection, from wild waterfowl to poultry to humans. Recently, the cross-species transmission of avian influenza A, particularly subtype H5N1, has highlighted the importance of the non-human subtypes and their incidence in the human population has increased over the past decade. During cross-species transmission, human disease can range from the asymptomatic to mild conjunctivitis to fulminant pneumonia and death. With these cases, however, the risk for genetic change and development of a novel virus increases, heightening the need for public health and hospital measures. This review discusses the epidemiology, host range, human disease, outcome, treatment, and prevention of cross-transmission of avian influenza A into humans.

Human influenza pandemics over the last 100 years have been caused by H1, H2, and H3 subtypes of influenza A viruses. More recently, avian influenza virus subtypes (that is, H5, H7) have been found to directly infect humans from their avian hosts. The recent emergence, host expansion, and spread of a highly pathogenic avian influenza (HPAI) H5N1 subtype in Asia have heightened concerns globally, both in regards to mortality from HPAI H5N1 infection in humans and the potential of a new pandemic. This paper will review the current human infections with avian influenza and their public health and medical implications. Influenza A, B and C are the most important genera of the Orthomyxoviridae family, casusing both pandemic and seasonal disease in humans. Influenza A viruses are enveloped, single-stranded RNA viruses with a segmented genome (Table 1 ) [1] . They are classified into subtypes on the basis of the antigenic properties of the hemagglutinin (HA) and neuraminidase (NA) glycoproteins expressed on the surface of the virus [1, 2] . Influenza A viruses are characterized by their pathogenicity, with highly pathogenic avian influenza (HPAI) causing severe disease or death in domestic poultry [3] . Molecular changes in the RNA genome occur through two main mechanisms: point mutation (antigenic drift) and RNA segment reassortment (antigenic shift) [4, 5] . Point mutations cause minor changes in the antigenic character of viruses and are the primary reason a vaccination for influenza A is given yearly. Reassortment occurs when a host cell is infected with two or more influenza A viruses, leading to the creation of a novel subtype. The influenza subtypes of the 1957 (H2N2) and 1968 (H3N2) pandemics occurred through reassortment, while the origins of the 1918 (H1N1) pandemic are unclear. The HA glycoprotein mediates attachment and entry of the virus by binding to sialic acid receptors on the cell surface. The binding affinity of the HA to the host sialic acid allows for the host specificity of influenza A [6, 7] . Avian influenza subtypes prefer to bind to sialic acid linked to galactose by α-2,3 linkages, which are found in avian intestinal and respiratory epithelium ( Table 2 ) [8] . Human virus subtypes bind to α-2,6 linkages found in human respiratory epithelium [8, 9] . Swine contain both α-2,3 and α-2,6 linkages in their respiratory epithelium, allowing for easy co-infection with both human and avian subtypes (thus acting as a 'mixing vessel' for new strains) [10] . Humans have been found to contain both α-2,3 and α-2,6 linkages in their lower respiratory tract and conjunctivae, which allows for human infections by avian subtypes [9, 11, 12] . The HA glycoprotein is the main target for immunity by neutralizing antibodies. The NA glycoprotein allows the spread of the virus by cleaving the glycosidic linkages to sialic acid on host cells and the surface of the virus. The virus is then spread in secretions or other bodily fluids. The NA glycoprotein is not the major target site for neutralization of the virus by antibodies. Influenza A viruses infect a wide range of hosts, including many avian species, and various mammalian species, such as swine, ferrets, felids, mink, whales, horses, seals, dogs, civets, and humans [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] [29] [30] [31] . Wild birds (ducks, geese, swans, and shorebirds) are important natural reservoirs of these viruses, and all of the known 16 HA and 9 NA subtypes have been found in these birds [32] [33] [34] [35] . In most cases, these subtypes are found within the gastrointestinal tract of the birds, are shed in their feces, and rarely cause disease [32] . Since 2002, however, HPAI H5N1 viruses originating in Asia have been reported from approximately 960 wild bird species, causing disease in some instances and asymptomatic shedding in others [36] [37] [38] [39] [40] [41] [42] [43] [44] [45] [46] [47] [48] . The virus has now spread across Asia, Europe, the Middle East, and some African countries. Additional species, such as tigers, leopards, cats, stone martens, and humans have also become infected with HPAI H5N1 [49] . This spread of H5N1 into a wide range of animal and avian species may enhance the spread of the virus into the human population as it interacts with animals in a number of ways (increased land use, markets, consumption) [44] . Thus, the potential contact, transmission, and mutability of HPAI H5N1 worldwide will increase as the number of species and their interactions increase, complicating prevention, surveillance and treatment possibilities. The incidence of avian influenza infections in humans has increased over the past decade (Table 3) . Initially, cases of avian influenza (H7N7) in humans occurred in association with poultry outbreaks, manifesting as self-limiting conjunctivitis [30, [50] [51] [52] [53] . Then, in 1997, a large scale HPAI H5N1 outbreak occurred among poultry in Hong Kong, with 18 documented human cases [29, 31, 54, 55] . Two subsequent poultry outbreaks in Hong Kong in 1999 and 2003 with HPAI H5N1 occurred without human cases until 2003 when two members of a family in Hong Kong contracted HPAI H5N1 [56] . In December of 2003, HPAI H5N1 surfaced in poultry in Korea and China, and from 2003 to 2006 the outbreak stretched worldwide in the largest outbreak in poultry history. Human cases of HPAI H5N1 followed the poultry outbreak, with a total of 256 cases and 151 fatalities thus far [57] . Other limited outbreaks have occurred, causing variable human disease (Table 3 ) [52, 58] . However, HPAI H5N1 remains the largest and most significant poultry and human avian influenza outbreak. Epidemiological investigations of human cases of avian influenza show that the virus was acquired by direct contact with infected birds [29] [30] [31] [50] [51] [52] [53] [54] [55] [56] . Influenza A is transmitted through the fecal-oral and respiratory routes among wild birds and poultry [32] . Human interaction with these infected secretions and birds was the major mode of transmission, with contact including consumption of undercooked or raw poultry products, handling of sick or dead birds without protection, or food processing at bird cleaning sites. All birds were domesticated (chicken, duck, goose) and no transmission from birds in the wild (migrating) or contaminated waterways has been documented. In a few cases, limited human to human transmission has been reported among health care workers and family members (Table 4 ) [59] [60] [61] [62] [63] . In each of these cases, no personal protective equipment was used, which is the major factor in transmission between humans [60] . The clinical manifestations of avian influenza in humans has ranged from mild conjunctivitis to severe pneumonia with multi-organ system failure ( [53] . However, with HPAI in Hong Kong in 1997 and in Southeast Asia currently, pneumonia progressing to multiorgan failure, acute respiratory distress syndrome (ARDS), and death are the predominant findings [17, 55, [65] [66] [67] [68] . Rye syndrome, pulmonary hemorrhage, and predominant nausea, vomiting, and diarrhea complicate these cases [68] . Laboratory findings include both thrombocytopenia and lymphopenia [65, 66] . Chest radiographic findings include interstitial infiltrates, lobar consolidation, and air bronchograms. The clinical course of patients with HPAI H5N1 is rapid, with 68% percent of patients developing ARDS and multiorgan failure within 6 days of disease onset [69] . The case fatality rate ranges form 67% to 80%, depending on the case series [17, 55, 65, 66] . Once the patients reached the critical care unit, however, the mortality rate was 90% [69] . The average time of death from disease onset was nine to ten days. Avian influenza A infections in humans differ from seasonal influenza in several ways. The presence of conjunctivitis is Available online http://ccforum.com/content/11/2/209 Number of fatalities (percent) 0 (0) 6 (33) 0 (0) 1 (1) 0 (0) 151 (59) H, hemagglutinin; ILI, influenza like illness; N, neuroaminidase. more common with avian influenza A infections than with seasonal influenza. Gastrointestinal symptoms, as seen with HPAI H5N1, and reports of primary influenza pneumonia and development of ARDS are also more common with avian influenza A infections [65, 67, 69] . Finally, the rapid progression to multi-organ failure and eventually death occurs at a much higher rate with avian influenza A infections [69] . Post-mortem studies have illustrated findings consistent with an overwhelming systemic inflammatory response syndrome, including diffuse alveolar damage, acute tubular necrosis and atrophy, disseminated intravascular coagulation, and multiorgan damage [70, 71] . Interestingly, the virus has been isolated from the lungs, intestine, spleen, and brain, suggesting viremia, but active replication of the virus has been limited to the lungs [71] . This overwhelming inflammatory response, with acute lung injury and ARDS as the predominant features, coincides with the findings of preferential binding of the avian influenza A viruses to α-2,3 linkages in type II pneumocytes of the lower respiratory tract of humans and a vigorous cytokine response, including increased interleukin-6, interleukin-10, and interferon beta release [11, 12, 70, 71] . The clinical diagnosis of avian influenza infection in humans is difficult and relies on the epidemiological link to endemic areas, contact with sick or dead poultry, or contact with a confirmed case of avian influenza (Table 6 ). Since many infectious diseases present with similar symptoms, the only feature significant to the clinician may be contact in an endemic area, through travel or infected poultry, and the clinician should always elicit a detailed patient history. The definitive diagnosis is made from isolation of the virus in culture from clinical specimens. This method not only provides the definitive diagnosis, but the viral isolate is now available for further testing, including pathogenicity, antiviral resistance, and DNA sequencing and analysis. Alternatively, antibody testing can be performed, with a standard four-fold titer increase to the specific subtype of avian influenza virus. Neutralizing antibody titer assays for H5, H7 and H9 are performed by the micorneutralization technique [72] . Western blot analysis with recombinant H5 is the confirmatory test for any positive microneutralization assay [59, 60, 72] . More recently, rapid diagnosis can be performed with reverse transcription-PCR on clinical samples with primers specific for the viral subtype [73] [74] [75] . This test should be performed only on patients meeting the case definition of possible avian influenza A infection. Any suspected case of avian influenza in a human should be investigated by the public health officials in the province or country of origin [39, 76] . Additionally, governmental labs are often equipped with the appropriate biolevel safety 3 laboratories, primer libraries, and associated expertise to confirm the diagnosis quickly and efficiently. Any clinical specimens should be submitted with the assistance of the public health experts. Treatment of avian influenza infections in humans includes antiviral therapy and supportive care. Controlled clinical trials on the efficacy of antivirals (NA inhibitors), supportive therapy, or adjuvant care have never been performed, so current recommendations stem from the experiences of past avian influenza outbreaks and animal models. The adamantanes (rimantadine and amantadine) and NA inhibitors (oseltamivir and zanamivir) are the antivirals used for treatment and prophylaxis of influenza infections in humans. In avian influenza virus infections, adamantanes have no role due to widespread resistance through a M2 protein alteration. In addition, over 90% of isolates of H1 and H3 human subtypes during seasonal influenza have had resistance to the adamantanes [77] . Their role has now been limited to prophylaxis in the community when the circulation strain is know to be susceptible to the adamantanes [78] [79] [80] . NA inhibitors (oseltamivir and zanamivir) have been studied for both treatment and prophylaxis with the human influenza A subtypes H1, H2, and H3 as well as influenza B (Table 7 ) [80] [81] [82] . In animal models with HPAI H5N1, their efficacy has been well documented, with improved survival rates seen after infection [83] [84] [85] . Oseltamivir has been used in avian influenza outbreaks involving H7N7 and HPAI H5N1, and therapy with oseltamivir has been shown to decrease the viral load in nasal secretions in patients infected with HPAI H5N1 [11, 86, 87] . Resistance to oseltamivir has been documented in a HPAI H5N1 subtype in a Vietnamese girl treated with 75 mg daily for 4 days as post-exposure prophylaxis [68] . The NA glycoprotein had a histidine to tyrosine substitution at position 274, conveying a markedly higher IC50 for oseltamivir [68, 88] . In one study, the viral count of HPAI H5N1 in nasal secretions did not decrease with the administration of oseltamivir when the H5N1 isolate carried this resistance mutation [68] . However, resistance produced by this change may be overcome with higher doses of oseltamivir in vitro, and this change has not been documented to confer resistance to zanamivir [88] . The timing of treatment with NA inhibitors is paramount, as early therapy is directly related to improved survival [66, [83] [84] [85] . The greatest level of protection was seen if the NA inhibitors were started within 48 hours of infection, and protection rapidly dropped after 60 hours [78, 79] . These initial studies, however, were performed with seasonal human influenza A and B, where the period of viral shedding is approximately 48 to 72 hours. In HPAI H5N1 cases from Southeast Asia, survival appeared to be improved in patients who received oseltamavir earlier (4.5 days versus 9 days after onset of symptoms) [66] . Both of these time periods are much longer than documented in animal models, so the window of optimal therapy is still unknown, particularly if viral shedding exceeds the average 48 to 72 hour period seen in seasonal influenza A and B infections. Combination therapy with influenza A viruses has not been studied [84] . Ribaviron by inhalation has been evaluated in vitro with some avian influenza A subtypes and has been found to reduce mortality from influenza B in a mouse model [89] . Further animal model studies are indicated to determine if there is a role for ribaviron or combination therapy with avian influenza A viruses. Supportive care with intravenous rehydration, mechanical ventilation, vasopressor therapy, and renal replacement therapy are required if multiorgan failure and ARDS are a feature of disease [69, 90] . Due to the progression of pneumonia to ARDS, non-invasive ventilation is not recommended, and early intubation may be beneficial before overt respiratory failure ensues. Corticosteroids have been used in some patients with HPAI H5N1, but no definitive role for steroids has been determined. Other immunomodulatory therapy has not been reported [91] . Human vaccination for avian influenza viruses has not been widely used, although multiple vaccination trials are underway. Prior avian vaccines in humans have been poorly immunogenic and thus have limited use. An inactivated H5N3 has been tested and was tolerated but with limited immunogenicity [91, 92] . Other H5 vaccines have resulted in the development of neutralizing antibodies, but to a limited degree [93, 94] . Recently, a large randomized trial looked at an H5N1 attenuated vaccine from the Vietnam strain [95] . Only a modest immune response was seen, with microneutralization antibodies being developed at 12 times the dose used in the seasonal influenza vaccine. The side effects were minimal. A number of other industry trials with adjuvant vaccines are currently ongoing. Although promising, human vaccination against avian influenza viruses is still under development. Underscoring this development is the uncertainty of a pandemic strain, which may have vastly different antigenic properties from any developed H5 vaccine. Health care infection control is a crucial component in the management of avian influenza infection or a new pandemic strain. Experience from the severe ARDS outbreak in 2002 has illustrated that appropriate infection control measures are paramount to reduce spread to health care workers and, possibly, the community [96] [97] [98] . Therefore, the World Health Organization (WHO) and Centers for Disease Control and Prevention (CDC) recommend contact and airborne precautions for any initial suspected case of avian influenza in a human [99] . In late October 2006, the CDC released updated interim guidance on the use of masks and respirators in the health care setting (Table 8 ) [99] . In certain high risk procedures, additional protection may be considered given the likelihood of generating aerosol particles that may enhance transmission (Table 9 ) [99] . Respiratory protection should be worn along with an impermeable gown, face shield, and gloves. Initial cases should be placed in a negative pressure isolation room with 6 to 12 air changes per hour. Hand hygiene with antibacterial soap or alcohol based washless gel should be standard, with appropriate basins at each patient room. Seasonal vaccination of all health care workers should be preformed and further emphasized in order to reduce the likelihood of co-infection with two stains of influenza. Visitors and family members should be strictly monitored and their access to the patient limited to reduce the likelihood of spread. Finally, antiviral chemoprophylaxis should be available to any health care workers exposed to an infected individual. Any symptomatic worker should be taken off duty and workplace surveillance should occur. With these aggressive measures, risk to health care workers, patients, and family members will be reduced. Avian influenza viruses have occurred with increased incidence within the human population, reflecting the delicate and tangled interaction between wildlife, domesticated animals, and humans. Disease in humans can be limited to conjunctivitis or an influenza-like illness, but HPAI H5N1 causes mainly severe pneumonia, respiratory failure, and death. Most cases have occurred through direct transmission from infected poultry or waterfowl, with only a few limited cases of human to human transmission. Treatment has been successful with the NA inhibitors if started early, and vaccine development is underway with a more immunogenic attenuated H5N1 virus preparation. Infection control measures are the mainstay for prevention and disease reduction. Avian influenza viruses may constitute part of the next pandemic, so appropriate knowledge, prevention, and treatment will reduce the likelihood of this occurrence. Table 9 High risk aerosol procedures in avian influenza Non-invasive mechanical ventilation Bronchoscopy Humidified oxygen delivery Non-rebreather mask without expiratory filter This article is part of a thematic series on Disaster management edited by J Christopher Farmer. Other articles in this series can be found online at http://ccforum.com/articles/ theme-series.asp?series=CC_Disaster

Clinical review: Mass casualty triage – pandemic influenza and critical care

Worst case scenarios for pandemic influenza planning in the US involve over 700,000 patients requiring mechanical ventilation. UK planning predicts a 231% occupancy of current level 3 (intensive care unit) bed capacity. Critical care planners need to recognise that mortality is likely to be high and the risk to healthcare workers significant. Contingency planning should, therefore, be multi-faceted, involving a robust health command structure, the facility to expand critical care provision in terms of space, equipment and staff and cohorting of affected patients in the early stages. It should also be recognised that despite this expansion of critical care, demand will exceed supply and a process for triage needs to be developed that is valid, reproducible, transparent and consistent with distributive justice. We advocate the development and validation of physiological scores for use as a triage tool, coupled with candid public discussion of the process.

It is widely accepted that conditions exist for the evolution of a new strain of influenza virus with the potential to cause a human pandemic [1] . The biggest challenge in planning for an influenza pandemic is the range of unknown factors; its nature and impact cannot be fully predicted until the pandemic virus actually emerges. Those planning for a pandemic must, therefore, work from a number of assumptions based on knowledge gained from previous pandemics and scientific modelling of a range of potential scenarios. The UK Pandemic Influenza Plan [2] sets out a range of possible scenarios for clinical attack rates and case fatality rates during a pandemic, including the potential for more than one wave. The base scenario assumes a clinical attack rate of 25% and a case fatality rate of 0.37%, giving rise to 53,700 excess deaths in the UK. A reasonable worst case scenario involves a cumulative clinical attack rate of 50% with 2.5% case fatality, causing 709,300 excess deaths. Similarly, the US Department of Health and Human Services predicts that in a "moderate" scenario based on a virus with 1968-like pathogenicity, 865,000 will require hospitalisation and 65,000 (7.5%) will require ventilation. They also outline a "severe" 1918-like scenario with 9.9 million hospitalisations and 743,000 patients requiring ventilation [3] . An influenza pandemic will undoubtedly create a major increase in demand for critical care services. The majority of UK hospital intensive care units (ICUs) are already operating at > 98% bed occupancy. Integral to the success of any emergency planning strategy is 'surge capability', incorporating the ability to scale up the delivery of appropriate specialist care to those that require it [4] . Modelling of the impact of an influenza pandemic on UK critical care services has been carried out using the FluSurge 1.0 programme developed at the US Centers for Disease Control [5] . With simulation of an 8-week epidemic and 25% attack rate the demand for critical care beds from patients with influenza would represent 208% of current combined level 2 (highdependency unit) and level 3 (ICU) bed capacity, and 231% of current level 3 capacity [6] . Even allowing for optimistic estimates of other modulating factors (50% reduction in ICU demand with use of neuraminidase inhibitors and 50% upgrade of level 2 to level 3 beds), level 3 bed occupancy due to the pandemic would remain at 75%. Furthermore, occupancy of level 3 beds by 'flu patients' was unsustainable at approximately 50% in terms of care for other patients even in the most optimistic conditions. SARS outbreak, up to 32% of cases were admitted to ICU, 25% were mechanically ventilated and 28 day mortality for ventilated patients was 45% [13] . In Singaporean SARS patients admitted to ICU, 98% developed ARDS [13] . Properly constructed plans for the delivery of critical care during an influenza pandemic must include the ability to deal with excessive demand, high and possibly extreme mortality, and the risk to the health of critical care staff. The consequences of a pandemic, both in terms of numbers of patients and the effect on the healthcare system, are likely to precipitate a 'major incident' where special arrangements are needed to manage the system while it is under extreme pressure. It is anticipated that there will be an overwhelming demand for critical care services, not only for respiratory support through mechanical ventilation but also for a full range of care to manage multi-organ failure. Assuming that the next pandemic derives from the H5N1 strain, the epidemiological evidence to date suggests extremely high mortality and, although not precisely quantifiable, a significant risk to health care workers. Both of these will undermine the ability to deliver critical care to influenza patients even before consideration is given to the duty of care to other critically ill patients. Coherent incident response requires a robust command and control structure, with the ability to make rapid informed decisions across an organisation and also across a health economy. In the UK, health incident management is based on a 'medallion' structure, with gold, silver and bronze corresponding to strategic, tactical and operational command levels [14] . North American and Asian health institutions tend to use the Hospital Emergency Incident Command System [15] . The common theme in both systems is a clear command and control structure with which healthcare staff should be familiar [4, 14, [16] [17] [18] [19] . Their generic hierarchical structure allows application to a wide range of incidents whilst retaining familiarity gained from training and exercises. The importance of familiarity with the command and control structure was highlighted in a recent Delphi study [20] and European survey [21] . Critical care contingency planning guidance from the UK Department of Health places an expectation on providers to expand their level 3 bed capacity by a factor of 3 but no more. Provision of full multiorgan level 3 support is recognised to be unrealistic, but principally respiratory support is felt to be achievable. Cancellation of elective surgery to minimise alternative sources of demand for critical care, upgrading level 2 to level 3 facilities and recruitment of theatre recovery areas and even operating theatres may allow expansion of ICU-like care capacity. Staff in these areas already have the competencies to manage sedated patients and those receiving respiratory support. Escalating their clinical role should require relatively limited focussed training [22] . Other staff may need to be redeployed and receive training in the management of critical care patients to support fully trained staff, permitting a dilution of the standard critical care nurse to patient ratio [23] . Flexibility around dependency level and staff experience will be required [24] . The expansion of ICU capacity to provide critical care in other areas will require the pre-emptive identification, tracing and maintenance of all usable equipment and potentially the stockpiling of key items to allow for rapid up-scaling of activity in response to demand. It is likely that there will be some variability in the prevalence of influenza across the country during a pandemic wave, with peaks in demand staggered across geographical areas. It may be possible to disperse some of the patient load by interfacility transfer if this occurs to any significant extent. The expansion of ICU facilities during the SARS epidemic in Hong Kong and Singapore was recently described [25] . Infection control is recognised as an overriding priority for the delivery of critical care, including the ability, in the early stages, to cohort cases. This should ideally include the use of separate entrances and exits, isolation rooms with negative pressure ventilation and dedicated separate healthcare staff. The Toronto experience identified 21 secondary cases of nosocomial transmission of SARS in ICU from an initial index case before infection control measures were introduced. Even following the introduction of extensive protective equipment, nine healthcare workers developed SARS as a result of being present in the room during the intubation of a single patient. In terms of personal protection, planning and practice in the donning of protective equipment (PPE) and prior fit testing is essential [26] . The practicalities of being able to manage patients when fully attired must be understood and consideration given to the fact that any procedure or task will take longer. This will impact on care efficiency and the staff to patient ratio. While beds can be scaled up and extra areas recruited to provide critical care, without trained staff the planning will be ineffective. Staff illness rates and the risk to staff must be factored into the planning process. In the UK, staff illness has been estimated at 30% with work absences of up to 8 days [2] . Normal working patterns may need to be revised and facilities provided for staff to stay on site rather than go home to their families. Staff absence tends to be greater the longer special circumstances apply and the greater the impact on the lives of the staff [27] . The preventive effectiveness of neuraminidase inhibitors may make focussed chemoprophylaxis a strategy for reducing staff illness in critical care areas [28] . The evolution of a new pandemic strain of influenza will inevitably result in a major increase in demand for critical care services. It is likely that these services will rapidly reach their capacity and even their contingency arrangements for extended facilities will be overwhelmed. Excessive demand where resources are finite creates an ethical dilemma and many emergency plans apply a utilitarian approach of 'best care for the greatest number' [29] . There is a legitimate debate about how limited capacity can best be utilised, but a number of themes are recurrent. There needs to be a legal and ethical framework for the process decided in advance, the rationale for triage should be fair and transparent and it should meet the principles of distributive justice [30] [31] [32] . Triage can conflict with human rights legislation and even humanitarian laws but 'accountability for reasonableness' can temper the disagreements about priority setting [33] . The decision making process needs to be valid and reproducible. Although there are a number of triage systems available for mass casualty incidents, there has been little validation of any of them in the field [34] , and what there has been relates to 'big bang' single incidents and the apparent unreliability of triage [35, 36] . While it does not need to be explicit ahead of time, the decision thresholds should be based on both the cumulative evidence about the disease process and prognosis, and the number of patients and severity of illness making the demands on the service [31] . In effect, triage may result in a gradual degradation of care with the increasing scale of the incident and become a 'societally mandated Do Not Resuscitate order'. On these grounds the process needs to be carefully considered at an appropriately senior level and applied consistently [32] . Allowing for the utilitarian approach, it is recognised that in mass casualty incidents, the standard of care for all patients, including those not immediately related to the incident, may need to be adjusted and reduced. While this may infringe individual rights, the higher ethical principle of 'wellness of society as a whole' allows for the direction of resources to those where it is felt most effective. It may also allow for an expansion in the scope of practice of non-physicians [37] . It may be unrealistic and impractical to expect that senior medical intensive care staff will make all decisions regarding instituting critical care and there will be a need to empower more referring general clinicians to do so. This is at odds with the need for decision making by the most senior person [32] and will require a change in practice for many clinicians; it is not current practice in the UK. The use of track and triage protocols will be essential to direct this decision making and ensure its consistency. Ardagh [38] has developed a set of pragmatic questions for the clinician facing acute problems of resource allocation; the only point lacking in his assessment process is a tool for the 'ranking' of patients in terms of likelihood of benefit from the limited resources. We believe that the basic criteria for a system for triage to critical care in a pandemic are fourfold; it should identify patients sick enough to require higher level care at some stage in their illness, it should be able to recognise those patients who are too acutely or chronically unwell to benefit from critical care, it should be consistently applicable by healthcare professionals and support workers from a variety of backgrounds within the constraints of the pandemic and should ideally also be scalable to reflect any mismatch between need and capacity. In order to fairly allocate resources across both flu and non-flu patients it should also be disease non-specific and allow prognostic comparisons across disease categories. A number of scoring systems have been advocated for use in a pandemic. The UK Department of Health currently recommends a six-point pneumonia severity score [2] . Although US guidelines emphasise the importance of triage in primary influenza, specific tools are only recommended for assessment of post-influenza bacterial pneumonia [39] . The majority of available potential scores were developed as mortality indicators and perform less well for predicting critical care usage. Amongst ICU admissions with community-acquired pneumonia in Massachusetts in 1996 to 1997, 10/32 scored CURB-65 1 or 2 (that is, low risk) and 5/32 were classified as PSI (Pneumonia Severity Index) class III (intermediate risk) [40] . Even amongst patients with pneumonia included in the PROWESS study, only 90.5% were PSI class IV or V, and only 70.3% had a CURB-65 score of 3 or above [41] . There is no guarantee that pandemic influenza will be primarily pneumonic in its presentation; case reports have documented H5N1 influenza presenting with diarrhoea [42, 43] and coma [43] and a World Health Organisation summary has described absence of respiratory symptoms in a number of cases [44] . The utility of disease-specific pneumonia scores may also be limited by mortality from comorbidities such as cardiovascular disease. A number of intensive care scoring systems have demonstrated their power in using physiological derangement to predict mortality or higher resource requirements, whatever the presenting diagnosis [45] [46] [47] [48] [49] . Physiological scores have also been demonstrated to be good predictors of requirement for higher level care on hospital wards [50] , in medical assessment units [51, 52] and in the Emergency Department [53] . We have demonstrated that a purely clinical score incorporating acute physiological derangement and chronic health and performance status can reliably predict requirement for critical care [54] . It is inevitable that if an influenza pandemic reaches the scale of some predictions, some patients who, in normal circumstances, would benefit from critical care will not be offered it. Critical care triage will need to evolve from a process of identifying cases who need high level care to one that determines those patients most likely to benefit from the limited resources available and distinguishes them from those where care is likely to be futile. This is recognised by the Emergency Medicine community and the US administration in terms of disaster triage [37, 55] . The American Thoracic Society adopted the utilitarian principle a decade ago, stating that "the duty of health providers to benefit an individual patient has limits when doing so unfairly compromises the availability of resources needed by others" [56] . The problem now facing policymakers and clinicians is defining a process for resource allocation that meets the requirements of distributive justice and accountability for reasonableness [33] . As the Working Group on Emergency Mass Critical Care of the Society for Critical Care Medicine recognised, "an ideal triage system is based on data collected at hospital admission, requires little or no laboratory testing, and has been proven to predict hospital survival" [57] . The Ontario Ministry of Health Long-term Care working group have courageously taken the first steps in defining a triage protocol for critical care [58] and their use of serial Sequential Organ Failure Assessment (SOFA) scores to place a ceiling on care provided to non-responding patients is to be supported. However, it is unlikely to be feasible for all patients to have a trial of inotropes and/or ventilation and some way of screening out the sicker patients at ward/floor level will be required. We are not aware of the use of objective prognostic scores to allocate or refuse critical care resources at present and indeed most research demonstrates the ad hoc nature of admission decision-making [59] . However, if, as is likely, review by experienced critical care physicians is impractical, decision support will be required for the non-critical care specialist. Emergency physicians, for example, had a positive predictive value (PPV) of only 73% in identifying those with a low chance of survival, as opposed to critical care fellows (PPV 83%) and the Mortality Probability Model (MPM 0 ; PPV 86%) [60] . SOFA scoring has previously been demonstrated on a multinational basis to predict high risk of mortality (a SOFA score of over 15 was 98.9% specific for mortality) [61] . Other critical care scoring systems show comparable performance in mortality prediction; discrimination as measured by area under Receiver Operator Characteristic (ROC) curve was 0.825 to 0.901 for Acute Physiology and Chronic Health Evaluation III (APACHE III) [62] [63] [64] [65] , 0.79 to 0.846 for Simplified Acute Physiology Score II (SAPS II) [62, 64, 66] , and 0.928 for the Multiple Organ Dysfunction Score [67] . However, calibration of these scores to give absolute risks of mortality has not always been reliable [65] and has required customisation for international use [68, 69] . Concentrated work is clearly required to amend and validate existing scoring systems so that they are suitable for use as triage tools. We suggest that this should be done on two levels. While disease specific scoring systems are valuable and should continue to be refined, there is a need to develop an appropriately generalisable scoring system for as unselected a group of patients as possible. To have the discriminating power, it will need to take place on a multicentre or, preferably, on a multi-national basis. It is a general principle of major incident planning that procedures should not be changed at precisely the moment when the system or institution is under its greatest stress, so planning for pandemic flu needs to make use as much as possible of systems and procedures already in place. Development of a triage system and tool needs to be accompanied by planning for hospital command and control (to dictate scalability as related to available resources) and by training for staff whose roles may change. Researchers, clinicians and policymakers in the field need to analyse systems and scores already in existence and improve and validate them as triage tools (though this may not be the purpose for which they were originally developed). At the same time ethical principles require transparency and consistency in the decision-making process, and involvement of public in its development. In reality, perhaps the question we need to address is the action required when critical care services are overwhelmed. The scalability of triage tools may aid in decision making by objectively altering the threshold for admission to critical care. However, the time may come when we need realistically to evaluate the effectiveness of critical care in influenza. If survival with the benefit of critical care is marginal (for example, <10%) and there is a significant cross-infection risk, perhaps critical care should then close and concentrate its efforts on outreach to other areas, including wards. Direction and support from professional bodies and health departments will be required to support the medical staff with such difficult decisions possibly against a ground swell of media-driven public opinion. DW is a member of the UK Department of Health Critical Care Contingency Planning Working Group. The other authors declare that they have no competing interests.

Avian influenza outbreak in Turkey through health personnel's views: a qualitative study

BACKGROUND: Avian influenza threatens public health worldwide because it is usually associated with severe illness and, consequently, a higher risk of death. During the first months of 2006, Turkey experienced its first human avian influenza epidemic. A total of 21 human cases were identified, 12 of which were confirmed by the National Institute for Medical Research. Nine of the cases, including the four fatal ones, were from the Dogubeyazit-Van region. This study aims to evaluate the efforts at the avian influenza outbreak control in the Van-Dogubeyazit region in 2006 through the experiences of health personnel. METHODS: We conducted in-depth interviews with seventeen key informants who took active roles during the avian influenza outbreak in East Turkey during the first months of 2006. We gathered information about the initial responses, the progress and management of the outbreak control, and the reactions of the health professionals and the public. The findings of the study are reported according to the topics that appeared through thematic analysis of the interview transcripts. RESULTS: Following the first suspected avian influenza cases, a Van Crisis Coordination Committee was formed as the coordinating and decision-making body and played an important role in the appropriate timing of decisions. The health and agriculture services could not be well coordinated owing to the lack of integrated planning in preparation for outbreak and of integrated surveillance programs. Traditional poultry practice together with the low socio-economic status of the people and the lack of health care access in the region seemed to be a major risk for animal to animal and animal to human transmission. The strengths and weaknesses of the present health system – primary health care services, national surveillance and notification systems, human resource and management – affected the inter organizational coordination during the outbreak. Open communication between the government and the public played an important part in overcoming difficulties. CONCLUSION: Although there were problems during the avian influenza outbreak in Turkey, the rapid responses of the central and regional health authorities and the performance of the health workers were the key points in controlling the epidemic. The lessons from this outbreak should provide an opportunity for integrating the preparation plans of the health and agricultural organizations, and for revising the surveillance system and enhancing the role of the primary health care services in controlling epidemic disease. Developing successful strategies based on knowledge and experience may play a valuable role in delaying an avian influenza pandemic.

The highly pathogenic Avian Influenza A (H5N1) virus has caused more than ten outbreaks worldwide and many related human fatalities have occurred since 1997. A total of 258 laboratory-confirmed cases have been reported in South East Asia, North Africa and Europe by the World Health Organization (WHO) and 154 people with confirmed avian influenza died between 2003 and 2006 [1,2] . Most patients have similar features; they are children, they have a history of close contact with poultry or wild birds, and they live in some of the poorest areas of the world. The human cases in Turkey and Azerbaijan were the first confirmed reports of human avian influenza infection outside Asia and Africa. This was alarming not only for Turkey but also for the European Union and other countries. After considering the epidemiological and laboratory evidence, WHO has maintained its pandemic alert at Phase 3 (of 6), indicating that the new influenza strain causes human infections with no or very limited human to human transmission [3] . Effective surveillance, early warning systems and containment measures based upon a general capacity for health care have been recognized by WHO as the essential action strategies [4, 5] . In May 2002, "The Global Agenda for Influenza", which was issued by WHO, stressed the necessity of expanding animal influenza surveillance, and noted that studies at the domestic-wild bird and humandomestic bird interfaces are part of this activity [6] . The H5N1 type of avian influenza virus outbreak was lived in Turkey on January 2006; a total of 21 human cases were identified [3] , 12 of which were confirmed by the National Institute for Medical Research [3, 7] . Nine of the cases, including the 4 fatal ones, were from the Dogubeyazit-Van region. The confirmed cases were people aged 3-16 years who had close contact with ill poultry in the rural area in Dogubeyazit or in migrant wards in Van [3, 7, 8] . The timeline of events was summarized in Table 1 . In this period, more than one hundred suspected human cases (21% of total suspected patient in YYU Hospital) underwent prophylactic therapy in the outpatient clinic and twenty five percent of patients who had a history of contact and/or clinical findings were hospitalized in YYU Hospital [7] . Tens of thousands of chickens were culled [8] . Avian influenza infections in Turkey provide an example of concurrent animal and human avian influenza epidemics. These were the first occurrences of human cases in the country, and the impact of the epidemic was quite strong. It was the main agenda for the local and national institutions of health, agriculture and other sectors, the community and the media during the first months of 2006. This study aims to understand the course of the avian influenza outbreak control in Turkey through the views of health care providers. It will also highlight what local health personnel in various positions experienced during that period. This qualitative study was carried out in Dogubeyazit and Van in May 2006. We interviewed seventeen key informants who took active roles during the outbreak of avian influenza in East Turkey during the first months of 2006. The twelve of interviewees were medical doctors (directors, specialists and general practitioners), three were allied health personnel (one director and two health officers), and two were midwife and nurse (primary health care provider and director). Most of the informants were senior staff and had primary responsibilities for the management of the outbreak control. Five informants were from the Van Provincial Health Directorate (PHD). These informants were the Director and the Deputy Director of the PHD, who were the main coordinators of the avian influenza outbreak intervention, two department chiefs who were involved in implementing the intervention and one health officer who is responsible for transportation of the samples and record keeping. Four informants were health personnel in primary health care centers; three general practitioners and one midwife, who have active roles in surveillance in Van and Dogubeyazit. One of the general practitioners made the initial diagnosis of the avian case at a primary health care center in Van. Two general practitioners were staff of the primary health care center in Dogubeyazit. One of the doctors were the chief of the primary health care center and worked as a coordinator of avian influenza outbreak intervention and the other one was assigned to the avian influenza surveillance. Three informants were staff of the State Hospitals in Van and Dogubeyazit. The first was an infectious disease specialist working in Van State Hospital. The second was the Director of Dogubeyazit State Hospital during the outbreak. The third was a pediatrician working in Dogubeyazit State Hospital who pre-diagnosed the avian cases in Dogubeyazit. They worked as if gate-keepers between the primary health care center and the YYU Hospital during the outbreak. Five informants were Van YYU hospital staff. They were the Director of the University Hospital, the chief of nursing staff, two specialists and one health officer. The infectious disease specialist is on duty in the infectious disease control committee of YYU Hospital, who was a counselor in the Van Crisis Coordination Committee (CCC) and communicated with WHO representatives for the outbreak control. One internal medicine specialist who was also a native of Dogubeyazit helped the Provincial Health Directorates and WHO representatives to communicate with families from Dogubeyazit. The health officer was assigned to the emergency services during the outbreak and he was also the representative of the Health Workers Union in Van. We used a semi structured guide consisting of open ended questions during in-depth interviews. All interviews were audio-taped and field notes were taken afterwards from each interview for data-gathering. The appointments with the informants were arranged by the second author of this study, who worked in Van Yuzuncu Yil University, and the interviews were conducted by the first author, who worked outside the region. The authors clarified the purposes of the study and the interviews were conducted with the informants' consent for audio-taping and for the use of their words in the report. In the institutions where the authors were affiliated, interview studies are not subject to permissions from ethical committees. Thus, the authors did not apply for approval of these institutions. Following the interview guide, we asked the participants to inform us about the events chronologically: the first responses of the organizations, the management of outbreak control, the reactions of the health personnel and the resident population, the inter-organizational coordination, and interviewees' roles during the avian influenza outbreak in Turkey. We also asked them to evaluate the strength and weakness of the outbreak control, and the lessons from the avian influenza outbreak. Records were transcribed verbatim. The authors read and identified codes for major themes according to their interest for the study. These codes were then marked to lines of text and the relevant codes were collected one under the other for preliminary analysis. The second analysis [7] involved recovering the relevant codes from the text. The results of the study are reported according to the topics revealed by the analysis. In the avian influenza outbreak, the studies by health organizations were defined by two features as suggested by the interviewees: it was perceived as a new and unknown disease; and it was encountered as a regional, national and international crisis by the health authorities. The preparedness of animal and human health organizations for outbreak-human resource, management, surveillance approach and notification systems-affected inter organizational coordination during the outbreak. The comments related to follow up show us that there is an inconsistency between the pandemic influenza action plan and avian influenza surveillance approach. Although a circular of Turkish MoH was announced on October 2005 including a flow chart to evaluate the suspected cases, the exclusion of primary health care centers and hospitals in the sentinel surveillance and the lack of regional preparedness might have caused health care workers to miss some cases in the referral and reporting procedures. The cost of the culling animals was partially paid to the owners to compensate their economic loss. But in practice the official records couldn't be made exactly, therefore most of poultry owners could not receive the compensations and they had been subject to high rates of financial loss. The situation was expressed by the general practitioner from the primary health care center in Dogubeyazit:"Now about 90-100 thousand chickens were slaughtered. Five Turkish liras were to be paid for each but only 30% of this price was paid because 70% of them were taken without any official record so people couldn't claim their rights." After the declaration of avian influenza outbreak by the Turkish MoH, people's reactions to the disaster and patient applications increased to health care centers. The fear and panic were experienced among health care providers. This situation was explained by a health officer from YYU hospital: "When cases arrived in Van According to the health care providers' experiences, to deal with an outbreak, the health system must be strengthened, health services should be coordinated, surveillance system and the notification for communicable diseases must be operated efficiently, human and animal health care should be integrated. Health personnel suggested that the preparedness of the health organizations is very important for an outbreak in terms of equipment, human resources and information gathering The preparation studies should not only be on the management level, but also throughout the health organization and the experience of public health workers must be taken into account: "My opinion is that continuing education of health care providers should be provided, so the practitioner who notices a suspected case can take the details of the patient's history. If it is done like this, the system will work better, a recording system is formed". According to the experiences of health personnel, sectors and institutions related to the outbreak control must work in collaboration and be coordinated: "(...) avian influenza control absolutely a multi-sector activity. No one should try to be a hero. Every health organization continues with its task; if the organizations of agriculture, municipality, gendarme, police, mufti, education are not included, your success will be overshadowed, and your aim is not achieved." The subjects that most stressed by the health personnel were the prevention of transmission from animal to animal and from animal to human. The health personnel agreed that besides culling animals, studies on preserving animal health must be developed and sustained. In place of traditional poultry farming, alternatives suggested included changes to modern poultry farming at nearby settlements: "At this time our advice to the Van PAD was that since all of them were culling, new projects must be developed and information should be gathered on poultry farming. Of course, this should be done with government support." "The origin oriented control should be. Firstly, a surveillance system at our borders must work well, especially in terms of animal surveillance, but also mainly before disaster happens, not only when a pandemic or epidemic occurs (...)." As the health personnel, communication and information gathering is very important for the outbreak control. The public's psychology and needs must be taken into account to enable them to cooperate with the health teams. The general practitioner who has a close relationship with the local people emphasized the importance of people's psychology and environmental factors. "I mean, why wasn't 'a child playing with chicken heads like puppets' put on the agenda? When she was discharged from the hospital, she came to visit me and I bought her a doll and advised her not to play with chickens, which are microbial things. She slept with the doll for many nights. Why didn't anyone underline this occasion? Why was the people's psychological status not taken into account? Why wasn't attention paid to the problems related to the housing infrastructure or the lack of clean water? I still can't accept this; it still affects me." The Turkish Avian Influenza outbreak is an example for other countries in respect of the experiences and lessons achieved by the health organization managers and health care providers who actively worked in the process. The health care providers who participated in our study think that the avian outbreak reached national and regional crisis level because they faced an unknown emerging infectious disease and the threat to health that it was not predicted. According to the health care providers, the most critical issue was the coordination between the agriculture and health organizations during crisis control. The main reason for this problem was the lack of organizational preparation compatible with central or regional plans. To be prepared for an outbreak, it is vital to define the central, regional and local organizational framework, to maintain close cooperation and collaboration between the health and agriculture sectors and to share information on surveillance, evaluating the risk to humans and planning interventions at the time when a case occurs [9] [10] [11] [12] . WHO indicated that Turkey's preparedness plan provided a framework for action, even if not yet fully developed [3] . Indeed the Turkish MoH published the country's first national pandemic influenza action plan in October 2005 and a circular was sent to health organizations about possible avian influenza case descriptions and precautions [3, 8] . On the other hand, the results of this study indicates that, the circulation of the avian influenza action plan was insufficient, and regional and national activities related the outbreak control only started after following the verification of H5N1 virus in poultry. As preventive human and animal health care, organizational preparations were incomplete. WHO states that there was no active surveillance in neighbouring provinces after a domestic animal outbreak was confirmed in Igdir in December 2005 [3] . According to the health care providers who participated in our study, a reference laboratory could not meet the demand for the animal samples and after the human cases arose, the Turkish MoA decided to cull all the poultry without searching for possible or confirmed cases. Because there were not enough agriculture personnel for the culling at Van, the Turkish MoA purchased support from private veterinary clinics. This shows us the size of crisis and the lack of preparedness such as reviewing organizational facilities and operational plans in pre-outbreak phase. As a result, the work of outer-organizational teams, which proceeded in their own region, was not integrated with that of other regional health teams that continued surveillance. The evaluation of animal health services at that period is limited in this study because we did not interview agriculture personnel. Future research should focus on the experiences of the animal health service personnel. Evidently, chaos is unavoidable in primary health care services that do not include central and organizational level intervention plans for emerging avian influenza-like infectious diseases, as in the Turkey example. The health care providers think that the strengths and weaknesses of the health organizational structure before the outbreak affected the success of the intervention and problems were encountered in coordination between health institutions. While a failure is expected even in the health services of developed countries following a possible avian influenza pandemic [13, 14] , the chaos can be bigger in developing countries due to the weaknesses of health organization. Therefore, strengthening the healthsystem in developing countries shouldbe considered as a factor in delaying the worldwide spread of avian influenza. National and international network and partnerships have to maintain current public health and animal health infrastructures and resources for construction, modernization, enhancement and recruitment for unprecedented emerging and reemerging disease [15] . The most critical issue related to the health system is the surveillance system. The notification of communicable diseases in Turkey was changed in recent years. In the new system, influenza group diseases are subject to sentinel surveillance. According to the sentinel surveillance, primary health care centers have no obligation to notify influenza group diseases including the possible avian influenza cases. Only the training and research hospitals in some selected provinces have the duty for notification of confirmed influenza cases. During the Turkish avian influenza outbreak, Van and Dogubeyazit (Agri) were not included among the selected provinces. The health care providers who participated in our study emphasized that the widespread of the outbreak requires follow-up and assessment of all influenza cases who applied to health care centers. As the interviewees stated, trace-back investigations were not performed because of the current surveillance system mentioned above. The need for an active surveillance system is clear for an epidemic disease such as avian influenza, which is seasonal, regional and closely related to epidemic diseases related to agricultural practice. In the WHO Report dealing with lessons from the avian influenza outbreak in Turkey, the need for active surveillance of animal and human avian influenza outbreaks was underlined [3] . A well integrated effort such as lining up a network between animal and public health laboratory system is needed to define an epidemic earlier, ensure more effective control measures [7, 15] . This will possibly make us gain valuable time by delaying a pandemic. Practitioners, midwives and nurses, who are responsible for population-based health care services in the primary health care system, became the most important human resource for surveillance during the avian influenza outbreak in Turkey. This experience clearly shows that primary health care centers should be included in the surveillance system. For a efficient surveillance system, upgrading the sentinel physician network by enlisting and retraining more participants and [16] the coordination among public health workers, clinicians and managers is most necessary. It is accepted that public health workers will play an integral role in an influenza pandemic [17] . The health managers and providers whom we interviewed believe that their experiences must be reflected in national avian influenza preparedness plans. Another problem encountered during the avian influenza outbreak in Turkey was the fear of contamination risk among the health care providers who have no protective equipment against H5N1 when the outbreak started. It is as important to train the health personnel to be prepared for an outbreak as it is to stock protective equipment for them at the sites of preparation studies [14, 15, 18] . Interviewees believe that the delay in supplying protective equipment and the lack of orderly planning in the distribution of it caused concerns among health care providers about contamination risk, which affected their work output. Health personnel must be accepted as the primary group in prevention, and mortality and morbidity among them must be minimized to allow for efficient intervention by the maximum number of personnel during an avian influenza pandemic. Even if epidemiological evidence about the spread of the H5N1 virus from patients to health personnel is limited [19] , studies indicate that nurses who work in hospitals and public health workers touching the suspected avian influenza cases may experience fear and anxiety for their own and their families' health and can face ethical dilemmas when deciding between continuing their work and maintaining their families' health [20, 21] . Several problems were experienced during intervention in the outbreak by health care providers who participated in the study. They may be classified as problems related to the socio-economic conditions of people living in the region, and those related to the health care system, including the management and surveillance systems: The health care providers emphasized that people who lived in the rural part of Dogubeyazit and the suburban areas of Van that gather migrants lacked basic needs such as health, education and infrastructure. According to the interviewees, the outbreak in Turkey was influenced by factors closely related to traditional poultry farming and poverty-line economics, as was the case in Asian countries [22] . According to the health personnel who participated in our study, the infection spread rapidly because domestic birds belonging to neighbourhood backyards were able to walk around freely. It is possible that wild migrant birds may be in contact with poultry in the area for food or water. Family-based small-scale poultry farming is a major source of income and nutrition for the region. All our interviewees thought that the most important factor in the transmission of the H5N1 virus from animal to human was the sharing of shelter according to the low socioeconomic status of the family. Especially during fall and winter, family members share their one sleeping and dining room with poultry. From the year 2004, the contact histories of children and young adults who are described as a risk group for avian influenza cases throughout the world resemble the ones in Turkey [23] . Health care providers who worked in Van and Dogubeyazit stated that people had eaten ill chickens before death and young adults had taken part in the cleaning, cooking and slaughtering of the chickens. Children had played with sick animals and corpses. The low socio-economic status of the people was accompanied by insufficient primary health care services in the region. The first avian influenza cases were diagnosed after lower respiratory tract infection symptoms appeared. All the fatalities were among those who had recently applied to a health center and been diagnosed as avian influenza. Informants have differing explanations for the late diagnosis of the patients. According to some, the first admission of the families was late. Others claim that the family of the first cases was sent back home during the first admission, and avian flu was diagnosed only after their second admission. The community was affected psychologically and economically from the disaster. Avian influenza created the fear of a mysterious disease on the people. Most of the people were subject to a financial loss because of the slaughtering of their poultries and were impoverished relative to previous status. In this study we found that the communication between community and health personnel is important for establishing public support for outbreak control and to overcome obstacles such as lack of confidence in governmental organizations. After the avian cases were verified in Turkey, the declaration by the Turkish MoH was evaluated affirmative by the health care providers. It is very important that managers be clear about the known and unknown issues by the first stage of the avian influenza crisis, inform the people about health risks and precautions as early as possible, and be sensitive the worries of people [24, 25] . According to the health personnel interviewed, the reason why poultry were kept as a supply of food and living in Van and Dogubeyazit at the time the culling started was the lack of public education about preventable health problems in the region, as much as economic security. The experiences of health personnel indicate that the people participated more after the crisis centers were established by the health and agriculture boards, phone counselling services were made available, good communication was established between health personnel and the people during surveillance, and information studies were applied by media channels. Turkey experience in avian influenza outbreak shows that preparation planning and surveillance systems should be rational and sustainable. The lack of organizational emergency disease plans delineating the tasks and responsibilities of health care providers in the event of a possible avian influenza outbreak, and the lack of training about preventive care, caused health personnel to be caught unprepared by the outbreak. The problems in supplying and distributing protective equipment, reflecting another dimension of the lack of organizational preparedness, influenced the efficiency of the health care providers' work because they were anxious about their own and their families' health. During the preparation and updating of national epidemic and pandemic plans, and during outbreak management, public health workers should participate effectively in decision-making and risk communication. A well designed communication strategy people's participation in control and may relieve from the psychological effects of the outbreak. Animal and human health care services could not be well coordinated owing to the lack of integrated preparation and planning. Because the sentinel surveillance of influenza infections in Turkey was based on the confirmation of the cases diagnosed at the training and research hospitals in selected provinces, detection of possible human avian influenza cases at the primary health care level was hindered and trace-back investigation was precluded. To evaluate the spread of the avian influenza outbreak accurately, evaluation and follow-up of influenza cases needs to be done in all health care centers. Such evaluation also necessitates integrated human and animal surveillance and control measures. The causes of transmission of the H5N1 virus from wild to domestic animals in Turkey are closely related to socioeconomic conditions and traditional poultry farming. Poverty in the region played a predisposing role in the outbreak and the outbreak also increased poverty. Limited access to education, shelter, water supply and waste removal formed the basis for virus transmission from animals to animals and animals to humans. To reduce the risks rising from these predisposing conditions, infrastructure needs to be built specially at the village level. Although the limited access to health care services and the high contact rate with ill chickens were among the characteristics of the vulnerable group, the rapid organization of the health authorities during intervention studies prevented an increase in the mortality rate. Despite the above-mentioned problems in preparedness about coordination during the avian influenza outbreak control, the rapid response and performance of the health workers played an important role in controlling epidemic. the manuscript. TE arranged appointments with the informants during the course of data gathering and contributed to analyzing the data. Both authors read drafts of the manuscript, made comments and approved the final version.

Influenza activity in Europe during eight seasons (1999–2007): an evaluation of the indicators used to measure activity and an assessment of the timing, length and course of peak activity (spread) across Europe

BACKGROUND: The European Influenza Surveillance Scheme (EISS) has collected clinical and virological data on influenza since 1996 in an increasing number of countries. The EISS dataset was used to characterise important epidemiological features of influenza activity in Europe during eight winters (1999–2007). The following questions were addressed: 1) are the sentinel clinical reports a good measure of influenza activity? 2) how long is a typical influenza season in Europe? 3) is there a west-east and/or south-north course of peak activity ('spread') of influenza in Europe? METHODS: Influenza activity was measured by collecting data from sentinel general practitioners (GPs) and reports by national reference laboratories. The sentinel reports were first evaluated by comparing them to the laboratory reports and were then used to assess the timing and spread of influenza activity across Europe during eight seasons. RESULTS: We found a good match between the clinical sentinel data and laboratory reports of influenza collected by sentinel physicians (overall match of 72% for +/- 1 week difference). We also found a moderate to good match between the clinical sentinel data and laboratory reports of influenza from non-sentinel sources (overall match of 60% for +/- 1 week). There were no statistically significant differences between countries using ILI (influenza-like illness) or ARI (acute respiratory disease) as case definition. When looking at the peak-weeks of clinical activity, the average length of an influenza season in Europe was 15.6 weeks (median 15 weeks; range 12–19 weeks). Plotting the peak weeks of clinical influenza activity reported by sentinel GPs against the longitude or latitude of each country indicated that there was a west-east spread of peak activity (spread) of influenza across Europe in four winters (2001–2002, 2002–2003, 2003–2004 and 2004–2005) and a south-north spread in three winters (2001–2002, 2004–2005 and 2006–2007). CONCLUSION: We found that: 1) the clinical data reported by sentinel physicians is a valid indicator of influenza activity; 2) the length of influenza activity across the whole of Europe was surprisingly long, ranging from 12–19 weeks; 3) in 4 out of the 8 seasons, there was a west-east spread of influenza, in 3 seasons a south-north spread; not associated with type of dominant virus in those seasons.

Influenza has an important impact on societies each season. Surveillance data not only provide valuable information on the burden of disease in the population [1, 2] , but also enables an assessment of whether the vaccine is a good match with the circulating virus [3, 4] . Surveillance may help to plan and allocate health care resources and is important for pandemic preparedness [5] . In addition, the surveillance infrastructure can be used to monitor new emerging respiratory diseases, like SARS or avian influenza in humans [6] . Countries in Europe have shared detailed clinical and virological data via the European Influenza Surveillance Scheme (EISS) since 1996 [7] . This collaborative project is partially funded by the European Commission through the European Centre of Disease Prevention and Control (ECDC) and currently includes 30 countries. The scheme covers a total population of about 450 million inhabitants and an area of roughly 12 million square kilometres. EISS collects two types of data on influenza activity each season: 1) clinical and virological data collected by sentinel GPs and 2) virological data from non-sentinel sources. In the present study the EISS dataset was used to characterise important epidemiological features of influenza activity in Europe during eight winters (1999) (2000) (2001) (2002) (2003) (2004) (2005) (2006) (2007) . In recent seasons there have been indications that influenza activity first appeared in the west of Europe and then moved east across Europe. As an example, during the 2003-2004 season, activity began in Ireland, the United Kingdom, Portugal and Spain in early August and reached Poland in February 2004 [8] . These observations led us to assess three related questions: 1. Are the sentinel clinical reports a good measure of influenza activity? 2. How long is a typical influenza season in Europe? 3. Is there a west-east spread and/or south-north spread of influenza in Europe? Data from a median of 17 (14) (15) (16) (17) (18) (19) (20) (21) (22) (23) (24) (25) (26) (27) (28) . Countries were included in the analysis if they were at least 5 years active member of EISS and if weekly data were available for the full season. The assessment of influenza activity presented in this paper is largely based on data reported by sentinel GPs. The GPs report clinical cases of influenza-like illness (ILI) and/or acute respiratory infection (ARI) to a central registry and take respiratory specimens that are sent to a national reference laboratory for testing. This ensures that the clinical data reported by the sentinel physicians are validated by virological data on influenza. The national reference laboratories also report laboratory test results on non-sentinel respiratory specimens e.g. specimens from hospitals or non-sentinel physicians. These data were collected to have an additional indicator of influenza activity and to validate the sentinel virological data. The national reference laboratories participate in the 'Community Network of Reference Laboratories for Human Influenza in Europe' (CNRL), which is coordinated by EISS [9] . CNRL works closely with the WHO through its network of National Influenza Centres and collaborates with the Centre for Reference and Research on Influenza at Mill Hill, London, UK. In the current study only the time points of highest clinical and virological activity were used, because only peak levels represent undisputable markers of activity in a given country. The reason not to take whole incidence curves into account is that incidence rates in Europe vary considerably, because: 1) case definitions are not yet harmonised across Europe; most countries report cases of ILI, but some report cases of ARI, 2) the denominator calculations vary by country and, 3) consultation rates for ILI and ARI vary among countries. They not only depend on cultural factors, but also on the delivery of health care. For example, in some European countries a doctor's certificate is required for a single day of absence from work (leading to a higher consultation rate), whilst in others a certificate is only required after absence of 5 days or more, leading to a lower consultation rate. The weeks of peak activity were selected by plotting the clinical and virological data available for each country. If the clinical and virological activity was very low during a season (e.g. below or around the baseline level; defined as level of influenza activity in the period when no influenza virus was detected), it was difficult to identify the peak week and no peak was selected. Most countries reported cases of ILI to EISS (13 out of the 17 countries); four countries used the less restrictive case definition of ARI (Czech Republic, France, Germany and Romania). Since 2004 the Czech Republic and Romania also report cases of ILI [10] . The case definitions of ILI and ARI have been described by EISS and discussed by Aguilera et al. [11, 12] . Briefly; the general criteria for ILI are: sudden onset of fever > 38°C, with respiratory (i.e. cough, sore throat) and systemic symptoms (headache, muscular pain); the criteria for ARI are: sudden onset of respiratory symptoms, accompanied by fever and headache in the absence of other diagnosis. For the validity analysis of sentinel reports, we defined a good match as a situation where the sentinel and virological peaks occurred in the same week, or when there was a difference of only one week. For example the peak of the incidence of ILI consultations in the Netherlands during the 2002-2003 season was week 10 and the peak of positive laboratory reports of the dominant influenza virus was week 9. This difference of 1 week was considered to be a good match. Therefore, a time difference between peaks of 1 week or less was taken as a measure for a good match (irrespective of which peak presented first); a difference of 2 weeks was taken as a reasonable match and a longer period as a poor match. The analysis was based on the percentage of countries fulfilling the criteria of a good or moderate match during 8 influenza seasons. In a second validity analysis, using similar criteria, the sentinel clinical incidences were compared with non-sentinel laboratory reports. Because the case definitions of ILI and ARI differ considerably, the validity analyses were performed separately for countries using ILI (n = 13) and countries using ARI criteria (n = 4). Differences in matching percentages between ILI and ARI were statistically evaluated using the Chi-square test or Fisher exact test if the expected value in one of the cells was less than five. The length of an influenza season was roughly calculated by subtracting the earliest and latest week of peak clinical activity across Europe for each season. Per season the aggregated data of participating countries were used. Knowing that using peak weeks as indicator for activity would lead to underestimation of the length of the epidemic, because periods of high activity at the beginning and the end were not taken into account, 4 weeks were added: 2 weeks before the earliest peak and 2 weeks after the last peak. These periods still represent a rather conservative estimate of the slopes of increased activity around the incidence peaks. Eight countries participated throughout the 8 seasons (their longitudes ranging from -4 to 15.3); 6 countries during 7 seasons (longitudes -8 to 19.3) and 14 countries were included during less than 7 seasons (longitudes -3 to 25). The sequence of peak activity of influenza in the various European countries was taken as a measure for the spread of influenza across Europe. We are well aware that the use of the word 'spread' is based on the liberal assumption that the sequence of peak activity across Europe parallels the actual spread of influenza. Therefore, as we have no clear insight into the dynamics of influenza between countries, in the present study 'spread' should be appreciated with some caution and in a very general context. In order to assess a possible west-east spread or a south-north spread of influenza activity in Europe, the peak week data of influenza activity in EISS countries were plotted against the longitude and latitude of the central point in each country. For the purpose of finding the appropriate geographic centre of a country, rounded longitude and latitude figures were used, based on The Gazetteer of Conventional Names, third Edition, August 1988 [13] . For Northern Ireland, Scotland, Wales and England such central points were not available. Therefore we took the capital cities of these regions: Belfast, Edinburgh, Cardiff and London as best substitutes. Considering it was difficult to identify a peak during seasons of low influenza activity, the sentinel virological data, if available, were used to select the peak. If no sentinel virological data were available, no peak was selected. Regression analysis and analysis of significance was performed using SPSS 11.5. The variance was expressed as squared correlation coefficients (R 2 ), interpreted as follows: < 0.1 very weak correlation; 0.1-0.25 weak; 0.25-0.50 moderate; 0.5-0.75 strong; 0.75-0.9 very strong; > 0.9 exceptionally strong correlation. Table 1 presents the clinical data in 17 countries during eight seasons. When using the norm of +/-1 week, the mean total overlap of sentinel clinical and virological data was 72% (median 71%; range 25%-100%) for countries using ILI as case definition, and 71% (median 73%; range 57%-83%) for countries using ARI as case definition. When using the norm of +/-2 weeks, the mean overlap for ILI-countries rose to 84% (median 86%; range 50%-100%) and rose to 85% (median 86%; range 71%-100%) for countries using ARI as case definition. Table 1 also compares the sentinel clinical data with the non-sentinel virological data. When using a match of +/-1 week, the mean overlap of sentinel clinical and non-sentinel virological data for countries using ILI was 63% (median 67%; range 0%-100%), and 46% (median 50%; range 43%-50%) for countries using ARI. When accepting a match of +/-2 weeks, the overlap rose to 84% for ILI (median 86%; range 50%-100%), and to 68% for ARI (median 60%; range 50%-100%). The differences between ILI and ARI were statistically not significant (1week match, p = 0.26; 2-week match, p = 0.19). Table 2 No correlation was found between length of the epidemic and type of virus or virus combinations. Table 2 assesses the west-east and south-north spread of influenza activity in Europe after respectively plotting the longitude and latitude of each country to the peak clinical level of activity for each season. As an example the plot of the 2003-2004 season is shown in Figure 1 . The assessment was based on calculation of the squared correlation coefficient. In four consecutive seasons (2001) (2002) (2003) (2004) (2005) , there was a moderate to strong correlation between longitude and peak influenza activity, indicating a west-east spread of influenza. It is noteworthy that during these four seasons A(H3) was the dominant virus. However, this was also the case during the 1999-2000 and 2006-2007 seasons, in which there was no indication for a west-east spread. We repeated the analysis for the latitudes of each country and found a moderate correlation in three season, favouring a south-north spread. During two of these seasons (2001-2002 and 2004-2005) there was also a clear west-east spread of influenza. Taking all eight seasons into account the overall indication for a west-east spread was twice as strong than for a south-north spread (mean R 2 west-east: 0.277; mean R 2 south-north: 0.137). In the present study routinely collected surveillance data were used to assess the timing, length and spread of influenza activity in Europe during eight winter seasons. It first tried to establish the validity of the sentinel reports and then used these data to assess the length of an average season in Europe and whether or not there is a general westeast spread of influenza activity. The analysis in this paper is largely based on clinical sentinel reports and it was therefore important to validate this S: Sentinel, information provided by sentinel general practitioners NS: Non-sentinel, data obtained from non-sentinel physicians, hospitals and institutions ILI: influenza-like illness used as case definition by 13 countries ARI: acute respiratory disease used as case definition by 4 countries 1 week overlap: occurrence of peak incidences differ 1 week or less (good match) 2 week overlap: occurrence of peak incidences differ 2 weeks or less (moderate match) Percentages represent the mean overall match aggregated from data of most participating countries obtained during 8 influenza seasons data source, first with the virological sentinel data and then with non-sentinel laboratory data. We found that the data were valid, irrespective of whether ILI or ARI was used as case definition. There was an overall match of 72% (+/-1 week) between the clinical and virological sentinel data and a 60% match (+/-1 week) between sentinel clinical data and the non-sentinel virological data. Allowing a larger overlap (+/-2 weeks) provided a match of 84% and 80%, respectively. The results indicate that the sentinel system is very adequate in estimating influenza activity in a continent, irrespective of case definition. The strength of the system is that it combines communitybased clinical and virological data and also can provide age-specific data. The sentinel approach has already been implemented in other continents, our results imply that it should be considered in other parts of the world [14] . It is not surprising that there was a close relationship between the clinical and virological sentinel data. This relationship should be close, because the diagnosis of ILI or ARI and the subsequent collection of respiratory specimens is done by the same person (sentinel GP). Sampling of specimens is also usually at its highest during the period of increased influenza activity, which will lead to increased numbers of positive specimens. Some of the non-matches of sentinel data occurred during the Christmas/New Year period when the clinical and virological surveillance systems are affected by holidays. The second comparison we performed was between the sentinel clinical and the non-sentinel virological data. This was an important validity check because two independent surveillance systems were compared. Many countries (e.g. the U.S.) do not collect sentinel virological data, but base the virological assessment of influenza activity solely on data from non-sentinel sources [15] . Our study found a good match between sentinel clinical and nonsentinel virological sources. Again, the clinical rates can sometimes be unreliable during seasons when activity is very low, which may lead to mismatches. On the other hand, a large outbreak in a major hospital or region may be another important factor that can affect the non-sentinel data. Such localized outbreaks may lead to increased non-sentinel reports that are not picked up by the sentinel clinical reports, resulting in a mismatch. Taken together, we do favour the sentinel collection of virological data because the approach is more systematic, less prone to pre-selection and differs less among countries than clinical sampling. The mean length of a typical influenza season in Europe based on the peak activity levels of ILI/ARI was 15.6 weeks. It is a conservative estimate, because it does not include the period of increased influenza activity. If this period is also taken into account the average European influenza season lasts about 4.5 months. This is a important finding as it highlights the fact that influenza activity occurs for a long period of time in Europe each season. It also highlights the need to present country-specific data, in order to get insight into the diversity of activity [16] . Because the spread of influenza depends of many factors one might not expect a particular pattern. Still, four out of eight seasons showed a clear west-east spread of influenza. In three seasons there was a south-north spread. The overall indication for a west-east spread was stronger than for a south-north spread, which means that a west-east spread of influenza is a more common, but far from consistent phenomenon in Europe. It is noteworthy that the westeast spread occurred in 4 consecutive seasons during which the more virulent A(H3) was the dominant virus. It should be noted that the data for 1999-2000 and 2000-2001 were not complete: a number of important countries to the east of Europe (Latvia, Lithuania, the Slovak Republic and Poland) were not included in the analysis because data were not available. This might have affected our westeast analysis. In an analysis of the spread of influenza in the US over 30 years, Viboud et al. observed a consistent early onset of the epidemic in California, which is the most populous state in the U.S. [17] . Interestingly, the onset in California was earlier than in 3 populous Eastern states, suggesting that in addition to population factors also geographical and climate factors, such as:, mountain ranges, plains, lakes, and predominant wind direction may drive early epidemic activity. Some of these factors may be involved in a west-east spread in Europe: the western of Europe is the most populated part and factors that contribute to spread, such as commuting and airline travel consequently are more intense in the western than in the eastern part of Europe [18] . We have used a rather crude method to assess the geographic spread of influenza activity in Europe. A more refined method would be to collect doctor specific data as routinely is done in France [19] . As these data are not available yet within EISS we had to use the methodology of analysing peak levels of activity in each country. In 2005 EISS initiated a European Mapping Project, in which data from Germany and the Netherlands were brought together to map the spread of influenza on a weekly basis, using data from about 500 physicians in Germany and 84 in the Netherlands [20] . This project has now been extended to 7 countries and hopefully the upcoming data will allow us to better assess the spread of influenza [21] . These results have important consequences for public health: it allows better planning of health care resources at a local level, and at a European level better recommendations can be made about the timing of vaccination. Our analysis has demonstrated that the sentinel clinical data, the main indicator used to measure influenza activity in Europe, is a valid indicator for influenza activity. We also found that the length of influenza activity in Europe was surprisingly long. Finally, during 4 out of 8 seasons there was a clear indication of a west-east spread and in 3 seasons a moderate indication of a south-north spread of influenza activity across Europe.

Diagnosis and treatment of severe sepsis

The burden of infection in industrialized countries has prompted considerable effort to improve the outcomes of patients with sepsis. This has been formalized through the Surviving Sepsis Campaign 'bundles', derived from the recommendations of 11 professional societies, which have promoted global improvement in those practices whose primary goal it is to reduce sepsis-related death. However, difficulties remain in implementing all of the procedures recommended by the experts, despite the apparent pragmatism of those procedures. We summarize the main proposals made by the Surviving Sepsis Campaign and focus on the difficulties associated with making a proper diagnosis and supplying adequate treatment promptly to septic patients.

Severe sepsis and septic shock are currently among the most common causes of morbidity and mortality in intensive care, and their incidences have increased during the past decade as the population has aged [1, 2] . The emergency department (ED), where patients are treated for community-acquired infection, many of whom require intensive care unit (ICU) management [3] , has been identified as a setting in which these syndromes and their outcomes may readily be observed. Despite dramatic improvements in diagnostic and treatment procedures, mortality rates among patients with sepsis remained unchanged from the 1960s through to the late 1990s. Diagnostic algorithms have therefore been developed to identify at-risk populations, and professional societies have worked to implement treatment procedures that focus efforts on early intervention. The Surviving Sepsis Campaign proposed management procedures that differentiate between 'resuscitation bundles' for the first 6 hours and 'management bundles' to be applied until the end of the 24th hour [4] . These procedure recommendations have been disseminated worldwide and are focused on global improvement in practices whose primary goal it is to reduce sepsis-related death. As a consequence of the recommendations, a trend toward decreasing mortality has been observed during the past few years. Difficulties remain, however, in applying all of the procedures recommended by the experts. This article summarizes the main proposals raised by the Surviving Sepsis Campaign and focuses on the difficulties associated with applying these guidelines in an appropriate time frame. Definitions of sepsis, severe sepsis, and septic shock were proposed 15 years ago. They were based on expert advice and used criteria that identify progression of the infection along with appropriate responses [5] . However, these criteria are clearly inadequate in terms of allowing detection of severe infections in routine daily practice. A study of a large multicenter cohort of ICU patients with infection [6] concluded that simply categorizing an infectious process as 'sepsis' or 'severe sepsis' did not predict prognosis. A high score indicating a septic condition did not necessarily predict a patient's outcome, even though that outcome might be affected by sepsis-related organ dysfunction. With regard to patients presenting at the ED because of community-acquired pneumonia (CAP), a recent report from Dremsizov and coworkers [7] illustrated the limited value of the well established criteria for 'systemic inflammatory response syndrome' (SIRS) in predicting outcome. That work emphasizes the inability of the SIRS designation to identify which infected patients were at risk for developing severe sepsis or shock. These findings prompted experts to propose new scoring systems aimed at identifying patients who are at [9] outcomes in patients presenting at an ED with infection, and calculation of this score requires data that are immediately available in the ED. Despite its ability to predict all-cause death in the study population, the accuracy of the MEDS score has not been tested at the individual patient level; its use at the bedside has not been evaluated, and therefore this tool should not be used in decisions regarding triage and ICU referral [10] . Most of the other newly developed scoring systems appear to have only marginal utility in daily routine practice because they require microbiologic identification and 24-hour clinical evaluation; hence, they lack the immediacy that is required for decision making in emergency medicine [6] . To date, the Pneumonia Severity Index is the only scoring system that is considered to help physicians to assess severity of illness in the ED [11] . Using this score at the bedside allows better triage of lowrisk patients [12] [13] [14] , but it does not alter outcomes in more severe pneumonia [15] , in which it is only slightly more effective than the inadequate SIRS classification [7] . Evaluation of biologic factors also may help in determining the severity of illness. Cady and coworkers [16] proposed use of the arterial blood lactate level to identify patients with severe illness and to assess the severity of sepsis. The Surviving Sepsis Campaign Management Guidelines Committee [4] , and the American College of Chest Physicians and the Society of Critical Care Medicine Consensus Conference Committee [17] have also proposed guidelines that help to identify those patients who are at greater risk for sepsis. Recent reports from Shapiro [18] and Nguyen [19] and their colleagues have emphasized the importance of lactate clearance in identifying those patients who will respond to treatment and have a favorable outcome. Lactate clearance was shown to be a better prognostic factor than a single lactate determination performed on ED admission [18, 19] . However, a single venous lactate measurement above 4 mmol/l predicted short-term and in-hospital risk for death in patients presenting at the ED with suspected infection [20] , even in those with normal arterial blood pressure [21] . A single lactate dosage is thus a valuable tool that may facilitate early detection of at-risk patients. Plasma procalcitonin may also be valuable in this setting. Procalcitonin is a more specific test than C-reactive protein [22] and interleukin-6, and can help the physician to detect sepsis [23] . Higher levels of procalcitonin are sufficiently specific to identify those septic patients who will develop severe sepsis, but it is not sensitive enough for routine use in ED triage [24] . It is clear that the site of infection should be managed promptly in patients with severe infection, including emergency surgery when applicable. However, efforts should also focus on early and carefully controlled antimicrobial therapy. Minimizing the delay between admission and beginning antimicrobial treatment is key to achieving a successful outcome. The potential influence of delayed antibiotic therapy was first evaluated in patients with CAP. In a series of 18,209 Medicare patients older than 65 years admitted because of CAP [25], the antibiotic regimen used saved lives when the first dose was administered before hour 4 after admission. Of note, fewer than 50% of patients received antibiotics during the first 4 hours in this study and as many as 17% received antimicrobial treatment after hour 6. Those patients in whom administration of antimicrobial agents was delayed were elderly people with an atypical CAP presentation, or they exhibited clinical features inconsistent with a diagnosis of sepsis, such as the absence of fever and altered mental status [26] (specifically, patients in whom the diagnosis of infection was not obvious). Such a lack of aggressive and early antimicrobial therapy has been identified in various settings in which patients were being treated for such conditions as meningitis, cancer, CAP, and nosocomial pneumonia [27] [28] [29] [30] [31] [32] [33] . A recent retrospective analysis quantified the impact of delayed antimicrobial treatment in patients with severe sepsis. Kumar and coworkers [34] demonstrated that every additional hour without antibiotics increased the risk for death in hypotensive septic patients by 7.6% during the first 6 hours. Early antibiotic therapy has been incorporated into the Surviving Sepsis Campaign recommendations [35] , and we expect compliance with this component of the guidelines to increase from its current low level [36] . The focus of infection is sometimes difficult to ascertain, but treatment must effectively target the responsible pathogen, from among a wide range of potentially etiologic agents [37] . Initial selection of an antimicrobial agent with good activity against the causative organism is crucial for survival. A prospective evaluation of sepsis [38] emphasized that, other than comorbidity, the factor most strongly associated with death was ineffectiveness of antimicrobial treatment against the micro-organism identified in blood cultures. Several large reports corroborated the relation between ineffective antibiotic treatment and poor prognosis. Consequently, broadspectrum antibiotics have been recommended, and the agent selected should provide coverage against the microorganisms that are usually involved in the suspected focus of infection [35] . Supportive clinical evidence for use of broadspectrum antibiotics will probably remain sparse [36] , but effective antimicrobial management requires good microbiologic sense. Adherence to such guidelines regarding use of antibiotics may positively influence prognosis [39] , but efforts to improve detection of pathogens should continue because enhanced specificity allows one to focus treatment on the responsible micro-organism and so limit the spectrum of coverage. The usual microbiologic techniques of detection may lack effectiveness. The use of urine antigens to Streptococcus pneumoniae and Legionella pneumophila type 1 can help in patients with pneumonia. Apart from their good sensitivity, the presence of these antigens can be detected long after an infection and, in the case of pneumoccocal related infection, may reflect carriage of the micro-organism in the upper respiratory tract [40] . Sensitive genomics tools are now available to detect both bacteria and viruses, and multiplex platforms allow screening of a wide range of micro-organisms [41] . The position of these techniques in the diagnostic armamentarium is yet to be defined, but efforts to improve antimicrobial therapy must continue so that our practices and therefore outcomes may be improved in the future. Among the symptomatic treatments, need for hemodynamic management is the most apparent, but modalities continue to be discussed and the scientific literature abounds with studies in this area. Efficient restoration of circulating blood volume is the primary goal of resuscitation in septic patients [42] . Albumin was the first product to be broadly used for intravenous fluid loading, but a meta-analysis comparing albumin with other fluid loading agents [43] identified an increased risk for death among patients who received albumin for supportive treatment during shock. However, subgroup analysis (septic patients with hypoalbuminemia) [44] revealed a trend toward greater efficiency of fluid loading with albumin. The cost-benefit balance is another factor that has restricted use of albumin, but in their recent report Guidet and colleagues [45] indicated that albumin infusion was potentially cost-effective in patients with sepsis. Thus, use of albumin should be considered with caution; it currently lacks the support needed for it to be recommended for use in patients with septic shock. Transfusion of packed red cells may also be considered in septic patients because transfused hemoglobin may contribute to improved oxygen transport and delivery. Few controlled studies have tested this option, however, and it has been reported that liberal transfusion is potentially ineffective [46, 47] . Since the publication of the findings of Rivers and coworkers [48] , use of packed red cells has been regarded as a valuable approach to improving tissue oxygenation, but the specific indications for transfusion of packed red cells in this setting remain unclear. Although controversy persists in this area, preferential use of crystalloids rather than colloids is supported by the available literature. For the same amount of volume expansion, there is no difference between these two treatments in terms of ejection stroke volume or oxygen delivery [49] . Systematic reviews and meta-analyses that included patients with sepsis and other types of patients concluded that crystalloids and colloids were generally similar in effect; an exception was one study that identified an advantage for crystalloids [50] . This finding received support from a randomized trial [51] that found that patients with septic shock receiving colloids had greater renal impairment. A recent study [52] was conducted to compare colloid with crystalloid volume resuscitation, with the aim being to identify the safest choice for use in patients with sepsis. This study, which employed a prospective randomized multicenter design, compared the influence on outcome of Ringer's lactate versus hydroxyethyl starch and of intensive versus conventional insulin therapy in patients with severe sepsis and septic shock. Experts have already criticized this study on the grounds that its design confounds applicability of the findings to routine daily care [53] . To summarize, although infusing fluids is a cornerstone of supportive care during sepsis, the optimal modalities and volume are difficult to determine and choices should be driven by objectives in the individual patient [48] . A solid rationale explains the utilization of vasopressors in daily practice [54] , but the few comparative studies and the combination of different molecules account for their practical selection. Combining norepinephrine (noradrenaline) and dobutamine improved hemodynamic parameters of hepatosplanchnic circulation [55] but required invasive monitoring procedures, without clinical benefit. Dopamine and epinephrine are vasoconstrictors that also increase cardiac output, but their metabolic effects may be harmful [56, 57] . In addition, use of vasopressors has been associated with poorer outcomes in septic patients, but their influence on mortality was unclear [58] . To assist physicians in their use of vasoactive drugs, professional associations have proposed guidelines that allow an opportunity to administer epinephrine or a combination of norepinephrine and dobutamine to more severely ill patients [4] . A recently reported study [59] indicated that these two strategies were equivalent in terms of both efficacy and safety. Interest in vasopressin is reflected in a growing number of publications, but the available evidence does not allow its integration into a global therapeutic scheme. However, recent data [60] may justify reconsideration of vasopressin in severe sepsis management guidelines in the near future. The VAsopressin in Septic Shock Trial (VASST) study [61] is currently comparing vasopressin with norepinephrine as initial vasopressor in septic shock patients. Because the study is not yet completed, no analysis or definite conclusions can yet be drawn from this trial. Whichever drug is selected, introduction of vasopressors should be considered after optimal fluid loading; these agents may allow therapies to be applied earlier and more aggressively in order to improve physiological parameters and ultimately outcomes [48, 62] . In the initial management of patients with sepsis, improving physiological parameters such as blood pressure and tissue oxygen delivery is a clear goal, as has been emphasized by experts since the late 1990s [63] . Previous studies underscored that applying an early goal-directed therapy (EGDT) approach could improve survival. The landmark study conducted by Rivers and coworkers [48] emphasized this concept in the field of sepsis. Its publication in 2001 prompted a debate in basic medical practice centered on the question, is it possible to improve outcomes in septic patients by increasing tissue oxygenation parameters during the first 6 hours of management? The protocol proposed by Rivers and coworkers involves attainment of physiological levels of hemodynamic parameters (arterial blood pressure and central venous oxygen saturation [ScvO 2 ], by using fluid loading, vasopressors, packed red cells, and early initiation of mechanical ventilation) as rapidly as possible. The overwhelmingly positive results of this EGDT study prompted a number of ED and ICU teams to change their daily care in accordance with the study protocol. Some papers [62, [64] [65] [66] reported partial or absolute adherence to the procedures evaluated by Rivers and coworkers. Others proposed adapting the procedure to their medical system with either less aggressive therapy or by forming 'sepsis teams' specifically tasked with managing patients with severe infection [67] [68] [69] [70] [71] [72] . The overall result of these reports was a trend toward improved outcomes in septic patients [73] . However, these findings have been tempered by a number of barriers. Not all EDs have access to the same equipment, and ability to monitor hemodynamic parameters invasively varies widely [74] . Another unresolved issue is that not all ED physicians have the necessary resuscitation skills to administer optimal treatment, as observed in ICUs [75] . Additionally, a number of recent reports have identified the fact that EDs are increasingly overburdened. This can compromise the quality of care delivered to patients, especially those who require highly technical care that many ED physicians do not have time to practice because of everincreasing numbers of patients [76] [77] [78] . Finally, studies are now emerging that indicate how few of the recommendations have been implemented. Early administration of antimicrobial therapy was poorly adhered to, even in recent reports. In these, although the Surviving Sepsis Campaign proposals were implemented, the mean delay to first infusion of antibiotics remained in excess of 3 hours [62] , and as many as 68% of patients did not receive their first dose within this period [79] . Only a few EGDT validation studies have been conducted in EDs applying aggressive treatment outside the ICU. However, even in those EDs, mortality sometimes remained at 31% before and after the institution of procedures to improve coordination between ED and ICU [80] . In addition, effort should be maintained after the initiation of an EGDT strategy because performance dramatically decreased after initial implementation [81, 82] . In addition to the pragmatism of this therapeutic approach, the optimal tools with which to evaluate attainment of physiological goals have also been subject to debate. Although ScvO 2 is a valuable parameter when it is abnormal, it may be in the normal range even in severely septic patients [73] . The hemodynamic presentation, of which there are many, depends on comorbidities and stage of sepsis [83] . In addition to ScvO 2 , central venous pressure may also provide useful information. A low central venous pressure indicates hypovolemia, and a high central nervous pressure with a low ScvO 2 indicates myocardial suppression or mismatch of supply and demand. In any clinical situation, the findings must be interpreted alongside other clinical data. Other indicators may help, and systolic volume and pulse pressure variation of 10% or above also provide valuable information regarding blood volume [84] . Relatively liberal use of packed red cells to improve ScvO 2 may be offset by its potential harm [46, 47] but in the setting of severe sepsis and septic shock the theoretical risks appear balanced by the benefits in terms of tissue oxygenation [85] . Although use of central venous pressure and ScvO 2 to evaluate attainment of physiologic goals can be debated [73] , it is clear that defining reasonable goals to treat sepsis is important whatever the local organization and the available means to achieve those objectives are [86] . For the past two decades therapeutic trials attempting to elicit a change in the host response to infection have failed to improve patients' conditions despite positive preclinical data [87] [88] [89] [90] [91] . However, the results of two recent studies have led to a more promising approach to this problem with recombinant human activated protein C (rhAPC) and low-dose steroids. The hemodynamic effects of steroids have been widely discussed since their use was found to allow early withdrawal of vasopressor treatment in a prospective doubleblinded, multicenter study [92] . The positive effects of steroids on adrenergic receptor cycling and sodium and water balance have been proposed as explanations for this efficacy. Their anti-inflammatory role as well as their anticoagulant effect, caused by limiting membrane expression of tissue factor, may contribute to the clinical benefit. A major difficulty lies in defining adrenal deficiency in septic shock patients, and a number of definitions have thus far been used. A recent retrospective multicenter cohort study conducted by the Corticus study group [93] emphasized the importance of cortisol variation after corticotropin stimulation. That study additionally raised the possibility of a deleterious effect of etomidate on hormonal response and outcome, a concern that was previously reported by others [94] . This specific point is still subject to debate [95] . Efforts are currently being made to define the best strategy for use of steroids during sepsis. The efficacy of rhAPC has been tested in a large multicenter study, the results of which have been widely debated. This compound was initially designed to compensate for a deficit in the natural anticoagulant protein C during sepsis, and thus it limited organ failures and improved the survival of septic shock patients [96] . Since then a number of studies have demonstrated that it has additional beneficial effects on complex interactions with inflammation, innate immunity, and apoptosis [97, 98] . rhAPC also protected animals and healthy volunteers from hypotension after lipopolysaccharide challenge. A similar finding was also reported in the PROWESS (Recombinant Human Activated Protein C Worldwide Evaluation in Severe Sepsis) study, with more rapid improvement in hypotension and vasopressor withdrawal [99] . These clinical effects could be related to endocrine modulation (adrenomedullin was implicated in this regard) and vasoactive capacity. Mechanisms, efficacy, and safety of rhAPC are discussed in other reviews included in this supplement. Despite the strong evidence base, use of adjunctive therapies has remained sparse in the setting of sepsis. Questionnaire surveys have attested to the under-use of such adjunctive therapies. Once again, the need for medical adherence to new therapies must be promoted by implementation of local guidelines that are inspired by the recommendations of the Surviving Sepsis Campaign (Figure 1 ). New standards of care, such as low tidal volume mechanical ventilation and tight blood glucose control, have recently emerged and are now a cornerstone of treatment for critically ill patients. Low tidal volume (≤ 6 ml/kg) as compared with 'standard' mechanical ventilation (12 ml/kg) has improved survival in patients with acute respiratory distress syndrome in independent studies [100, 101] . Two landmark studies by van den Berghe and colleagues [102, 103] suggested that aggressive insulin therapy improved 30-day survival in critically ill surgical patients, and reduced morbidity indicators such as weaning from mechanical ventilation and hospital days in medical ICU patients. Whereas occurrence and management of hypoglycemia appeared irrelevant in the main papers and additional data, hypoglycemia has been identified as potentially causing harm by others [104] . Even if these standards are still discussed and do not specifically impact on sepsis, they may also contribute to quality-of-care improvement and finally to patients' successful outcome [83, 84] . Guidelines that were proposed through the Surviving Sepsis Campaign to improve outcome in septic patients are difficult to apply routinely in most EDs. Attempts to apply these procedures fully have varied widely; diagnosis may be problematic because of atypical or unspecific presentations, biomarkers are of little help at the start of treatment and are unspecific, supportive treatment often depends on local supply of resources, and specific devices are often absent in EDs for initial therapy and monitoring. Even adherence to early administration of antibiotic therapy is poor, with delays being common. Our goal is now to improve the level of care by applying evidence-based procedures. J-FD was an investigator in the PROWESS study and is a consultant for Eli Lilly and Company. Y-EC declares that he has no competing interests.

Simian virus 40 vectors for pulmonary gene therapy

BACKGROUND: Sepsis remains the leading cause of death in critically ill patients. One of the primary organs affected by sepsis is the lung, presenting as the Acute Respiratory Distress Syndrome (ARDS). Organ damage in sepsis involves an alteration in gene expression, making gene transfer a potential therapeutic modality. This work examines the feasibility of applying simian virus 40 (SV40) vectors for pulmonary gene therapy. METHODS: Sepsis-induced ARDS was established by cecal ligation double puncture (2CLP). SV40 vectors carrying the luciferase reporter gene (SV/luc) were administered intratracheally immediately after sepsis induction. Sham operated (SO) as well as 2CLP rats given intratracheal PBS or adenovirus expressing luciferase served as controls. Luc transduction was evaluated by in vivo light detection, immunoassay and luciferase mRNA detection by RT-PCR in tissue harvested from septic rats. Vector abundance and distribution into alveolar cells was evaluated using immunostaining for the SV40 VP1 capsid protein as well as by double staining for VP1 and for the surfactant protein C (proSP-C). Immunostaining for T-lymphocytes was used to evaluate the cellular immune response induced by the vector. RESULTS: Luc expression measured by in vivo light detection correlated with immunoassay from lung tissue harvested from the same rats. Moreover, our results showed vector presence in type II alveolar cells. The vector did not induce significant cellular immune response. CONCLUSION: In the present study we have demonstrated efficient uptake and expression of an SV40 vector in the lungs of animals with sepsis-induced ARDS. These vectors appear to be capable of in vivo transduction of alveolar type II cells and may thus become a future therapeutic tool.

Sepsis is the leading cause of death in critically ill patients [1] . Despite advances in treating the sepsis syndrome, the incidence and mortality of sepsis remains high (35-45%) [2] . Lung is the organ most often involved, with lung injury taking the form of Acute Respiratory Distress Syndrome (ARDS) [3] [4] [5] , carrying 40% mortality [6] . The pathological hallmark of ARDS is airspace flooding with proteinaceous fluid, basement membrane disruption, hyaline membrane deposition, surfactant depletion, interstitial swelling, and interstitial neutrophilic infiltration [7] . Histological sections of the lungs from patients dying of ARDS and from animal models of the disease reveal: interstitial edema, followed by extensive necrosis of alveolar epithelial cells [8] . Alveolar epithelium in the adult lung consists of two cell types: Type I, differentiated cells that facilitate gas exchange and Type II, metabolically active cells involved in surfactant secretion and epithelial repair, serving as progenitors for injured type I cells [9] . Injury to type II cells impairs gas exchange by reducing the surfactant supply and limits type I cells regeneration, their preservation being essential for recovery from ARDS [10] . Nowadays therapy for ARDS remains mainly supportive, designed to prevent secondary injury [11] . However, recent studies demonstrated that organ damage in sepsis involves an alteration in gene expression. Decreased transcription of surfactant proteins [12] as well as profound pulmonary epithelial dysregulation together with altered levels of Hsp70 were found in animal models of ARDS [13] . We also showed in our previous studies that severe sepsis induced by cecal ligation and puncture (2CLP) precipitates ARDS [13, 14] . Furthermore, we demonstrated that enhanced Hsp70 expression in pneumocytes using an adenoviral vector (AdHSP) not only decreased histological abnormalities in the lung but also improved shortterm outcome [15] . Specifically, we showed that AdHSP limited sepsis-induced acute inflammation by suppressing NF-κB activation [50] . One approach to correct deficient protein expression is to use viral mediated gene transfer [16] . An optimal vector for gene therapy should be safe, efficient, nonimmunogenic, and available in high titers [17] . There is currently no single vector with all these advantages. SV40 based vectors are efficient gene delivery vehicles for a wide spectrum of ex vivo and in vivo targets, including hematopoetic, liver and kidney cells [18, 19] . The vector was shown to be potentially effective in a number of clinical models, including Crigler Najjar syndrome [20] , HIV, [21] SV40 evades host immune response most likely by caveolar endocytosis, followed by vesicular transport to the endoplasmic reticulum [27], thus avoiding the more common endosomal-lysosomal pathway used by most viruses [28] . SV40 vectors were found to be nonimmunogenic, allowing repeated administrations and long survival of transduced cells [29, 30] . Replication incompetent SV40 vectors in which the viral T-antigen is replaced by the gene of interest are produced in vivo in cells that supply T-anti-gen in trans-, such as COS or COT cells [31] , and high vector titers may be readily prepared [32] . Cloning capacity of these vectors is limited; nevertheless, many potentially therapeutic genes may be accommodated, as shown by the wide spectrum of applications already investigated. The use of SV40 vectors for gene transfer to the lung has not been explored. In this work we found that SV40 vectors may be used for lung cell transduction. After in vivo vector delivery, we established and measured expression of the reporter gene in rats with sepsis induced ARDS. Our results lead the way for studies regarding gene delivery to the lung, in order to modulate pulmonary disease processes. SV/luc is a T-antigen replacement vector carrying the firefly luciferase (luc) reporter gene [33] . Preparation of the SV/ luc vector was performed as previously described [34] . Recombinant E1, E3-deleted adenoviral vectors (Ad/luc) were propagated in HEK293T cells, by commonly used methods [35] . Animal procedures were approved by the Institutional Animal Care Ethical Committee. Under Ketamine/Xylasine/Isoflurane anesthesia, severe sepsis was induced in Sprague-Dawley rats using cecal ligation double puncture (2CLP) [14] . 1.25 × 10 8 IU/ml (infectious units) of SV/luc in 300 μl PBS were administered via a tracheal catheter to 2CLP and sham operated (SO) rats immediately after the procedure and to unoperated (UO) rats. In the positive control group, 10 9 IU/Ad/luc in 300 μl PBS were administered in the same way to 2CLP animals [15] . Negative control animals received PBS only. The reporter gene used encodes the luciferase protein, an enzyme that converts luciferin in presence of oxygen and ATP to a bioluminescent substance. In vivo light detection was performed using the Roper Chemiluminescence Imaging System, (CCCD-cooled coupled charged camera) model LN/CCD-1300EB equipped with ST-133 controller and a 50 mm Nikon lens (Roper Scientific, Princeton Instrument, Trenton, NJ). The system enables detection of an internal light signal emerging from mammalian tissue. The measurement is the sum of the integrated light signal subtracted the background light emission of an area of equal size [36] . A pseudo color image represents light intensity spectrum (from blue -least intense, to redmost intense). Forty eight hours after vector administration, rats were reanesthetized. 125 mg/kg Beetle Luciferin (Promega, Madison, WI) was administered intraperitoneally, and luciferase activity was measured by photographing the animals first in light (to obtain the animal's image) and then in the dark (to measure light emission). A composite photograph was obtained by superimposing the two images. The average amount of fluorescence measured in light units per area, as detected by the CCCD camera represents signal intensity. After the light emission measurement, at 48 hrs, the animals were sacrificed and internal organs harvested. One lung was removed, preserved in formalin, embedded in paraffin, cut at 5 μm thickness and stained with hematoxylin and eosin. H&E sections were evaluated for lung injury, degree of injury and its distribution within the lungs [14] . Part of the spleen, liver, heart and one kidney were also harvested and prepared in the same way. The remaining part of the spleen, heart, liver, the second kidney and lung were frozen in fluid nitrogen and preserved at -80°C for RNA. Immunostaining was performed using the procedure described by Lavon et al [37] . For luciferase we used a primary rabbit anti-mouse polyclonal antibody (Cortex, San Leandro, CA) at 1:50 dilution, followed by a secondary goat anti-rabbit IgG antibody (biotin conjugated, Zymed, San Francisco, CA). VP1 capsid protein detection was performed using a rabbit anti-SV40 polyclonal primary antiserum [38] followed by FITC-conjugated goat antirabbit IgG secondary antiserum (Zymed). DAPI (blue) was used as nuclear counterstain. Double immunostaining for surfactant protein C (ProSP-C) and VP1 was performed in order to detect internalization of VP1 viral capsid protein into the alveolar cells. Serial lung sections were immunostained for VP1 using the same rabbit polyclonal antibody as above followed by goat anti-rabbit Cyte 5, and for ProSP-C using rabbit anti-ProSP-C polyclonal primary antiserum (Alomone Inc, Israel) followed by FITC-conjugated goat anti-mouse IgG. Immunohistochemical detection of Lymphocyte (CD3+ T cells) infiltration was performed using the immunoperoxidase avidin biotin methodology. A primary CD3+ T cell antibody (monoclonal mouse anti-rabbit, affinity purified, Biosource, Carlsbad, California) at 1:100 dilution was used, followed by a secondary goat anti-rabbit IgG antibody (biotin conjugated, Zymed). Total RNAs from the frozen tissues were extracted using Trireagent (Sigma, Saint Louis, MO) according to the manufacture's instructions. A quantity of 2.5 μg of extracted RNA (for each sample) was subjected to reverse transcription using 400 u/μl RNAse, 0.5 μg/μl oligo dT, 200 u/μl M-MLV RT and 2.5 mM dNTPs. 2 μl of the reacting cDNA products were used as template for PCR. PCR was performed using PCR primers specific for Luciferase: 5'-TGGTCTGCCTAAAGGTGTCG-3' (forward) and 5'-ATGTAGTCTCAGTGAGCCC-3' (reverse), and carried out for 35 cycles. pGL3 Luciferase reporter vector (Promega) served as a positive control. GAPDH was used as a house keeping control gene, using PCR primers specific for GAPDH: 5'-ACCACAGTCCATGCCATCAC-3' and 5'-TCCACCACCCTGTTGCTGTA-3'. In the present experiment, ARDS was induced secondary to intra-abdominal sepsis (2CLP). Starting 12 hours after 2CLP, and more prominently at 24 and 48 hours, 2CLP rats displayed the typical signs of sepsis: intense pallor/ cyanosis of the mucous membranes, dirty fur with erected hairs, distended abdomen, diarrhea, tachypnea (40-60 breath per minute). Animal behavior was grossly abnormal: periods of agitation followed by periods of sleepiness along with limited movements and inability to feed. Histological examination of the lung sections from 2CLP animals showed changes consistent with ARDS. Macroscopically: lungs were less aerated, covered with white fibrin patches and pleural fluid was found in different quantities. H&E stained sections depicted: alveoli filled with proteinaceous fluid, septal thickening, and interstitial neutrophilic infiltration ( Figure 1 ). Mortality was present only in the septic (2CLP) animals. The 48 hrs mortality rate following 2CLP was similar to previously published numbers in sepsis induced ARDS [13] [14] [15] . SV/luc or Ad/luc vectors were directly administered into the trachea of 2CLP and SO rats, immediately following the procedure. Using this route we achieved maximal vector concentration and distribution to the lung, limiting the systemic spread. Histopathologically there was no difference between the lungs of the septic rats given PBS or SV/luc ( Figure 1 ). Mortality rate was similar in 2CLP animals given either PBS or SV/luc, suggesting that the vector itself did not add to the sepsis/ARDS induced mortality. At 24 and 48 hours, all the animals were reanesthetized and photographed with the CCCD camera to detect light emission. At 48 hours all animals were sacrificed and the lungs were harvested. At 24 hours, Luciferase activity was not detected in any animal groups. We explain this finding by the time needed for the reporter gene to express. However, at 48 hours, luc activity (detected as luminescence by the CCCD camera) was seen over the lung areas in 2CLP animals given SV/luc (Figure 2a) . Low levels of luc activity were detected in the tracheostomy region, confirming a degree of vector affinity for airway epithelium (Figure 2b ). No luc activity was detected in rats given PBS. No significant in vivo luc activity was detected in the heart, liver, spleen or kidneys of the animals given SV/luc. Fluorescence over the lungs of SO rats was significantly lower than that seen in 2CLP animals (Figure 2a ). Intratracheal administration results in little vector uptake in the normal lungs of SO rats, while changes in the intracellular structure occurring during ARDS may be responsible for enhanced viral expression. Similar findings were observed with adenoviruses, and explained by an ARDS-induced exposure/expression of the CAR and integrin receptors that mediate adenoviral entry into the alveolar cells [14] . This mechanism may also apply to the SV40 vector, although different receptors or increased endocytosis may be involved. [39] [40] [41] . Based on our previous findings [14, 15] , a comparison of luc activity between SV/luc and Ad/luc vectors was performed. As expected, luc expression was lower in the 2CLP-SV/luc than in the 2CLP-Ad/luc group (Figure 2c, 2d) , because of the natural tropism of adenoviral vectors for the pulmonary epithelium [14] . Another factor may have been the higher titer of the adenoviruses used, a total of 10 9 IU/ml adenoviral vector as compared to 1,25 × 10 8 IU/ml of the SV40 vector. Luc immunostaining was nonspecific, involving both alveolar type I type II cells. It was moderate in septic animals and low in SO animals ( Figure 3 ). This may be due to moderate infectivity of the SV40 vector in the lungs or weak activity of the SV40 promoter in alveolar cells. This finding is also consistent with the literature, describing less transgene expression of some non -mammalian, non -vertebrate encoded proteins (e.g. luciferase, GFP, lac Z) commonly used as markers for transduction in SV40 vectors, than is usually seen with other vector systems (e.g., adenovirus) [18] . Immunostaining for luciferase in the liver, kidney, spleen and the heart was negative in both 2CLP and SO animals groups given SV/luc, suggesting that direct intratracheal administration of the vector results in negligible systemic spread despite alterations in membrane permeability and capillary leak seen in ARDS. In order to test if the low level of Luc immunostaining in lung tissue is due to a low vector penetration into the alveolar epithelium or to low expression of the reporter gene, we tested for the presence of the VP1 capsid protein by fluorescent immunostaining. A significantly higher number of alveolar cells contained viral capsid proteins compared with the number of luc positive cells (Figure 4a , 4b, and Lung pathology Figure 1 Lung pathology. H&E stained lung tissue, shown at ×20 and ×100 magnifications. Untreated control rats show normal lung histology; SO rats given intratracheal SV/luc also show normal histological appearance; 2CLP rats given intratracheal PBS or SV/ luc show distinct ARDS pathology: alveoli filled with proteinaceous fluid, septal thickening, interstitial lymphocytic and neutrophilic infiltration. compare with Figure 3 ). This suggests that vector penetration into lung epithelial cells is efficient, but transgene expression in many of the cells is below detection level. It is possible that the SV40 promoter used in the present vector is weak and therefore gene expression is lower than gene delivery. A stronger, lung specific promoter may significantly improve expression of a therapeutic gene in alveolar cells. Alveolar type II cells participate in the regenerative process of the lung. Therefore their transduction is critical for successful ARDS gene therapy. To examine this we performed co-immunostaining of lung tissue for surfactant protein C (Pro SP-C) a marker for alveolar type II cells and VP1 viral capsid protein (Figure 5 ). The figure depicts co-localization of Pro SP-C and VP1. These findings support our conclusion that the vector is internalized into alveolar type II cells. The ability to transduce these cells is essential because of their impor- tance in alveolar epithelial repair processes and lung recovery in ARDS [10] . RT-PCR was performed on frozen tissues of the lung, kidney, liver, spleen and heart. High levels of mRNA were detected in the lungs of 2CLP animals, similar to those found in the plasmid pGL3 used as positive control for luciferase. Minimal levels were present in the liver, kidney spleen and none at all in the heart, confirming minimal systemic spread of the vector after direct intratracheal instillation. Interestingly, in T0 and SO controls, minimal Detection of the VP1 capsid protein in lung tissue Figure 4 Detection of the VP1 capsid protein in lung tissue. 4a: Positive staining for VP1 protein appears as red intracytoplasmatic coloration: moderate for SO rats (above) and high for 2CLP rats (below). DAPI (blue) was used as nuclear counter-stain. 4b: Negative control for VP1 immunostaining. Above: lung tissue from control septic rats given intratracheal PBS, and below: normal lung tissue from a control T0 rats. The minimal staining seen represents the background. DAPI (blue) was used as nuclear counter-stain. Shown at ×40 magnification. Figure 3 Luciferase immunostaining. Was performed on the lung tissue harvested from 2CLP and SO rats given intratracheal SV/luc. Positive immunostaining appears as brown intracytoplasmatic coloration (Black arrows). Shown at ×100 magnification. expression was noticed in the lungs, in spite of the same route of vector administration, suggesting that the pathological processes taking place in the sepsis-induced ARDS lungs (as in 2CLP animals) enhances vector penetration and expression by an unknown mechanism. Regarding the kidney, liver, spleen and heart of T0 and SO animals, the same minimal expression was observed as in the septic animals ( Figure 6 ). Direct intratracheal administration of SV/luc induced no immunological response as measured by lymphocytic and neutrophilic infiltration of the H&E stained lung sections compared to PBS treated septic lung (Figure 1 ). To further confirm this we performed immunohistochemical detection for lymphocytic (CD3+ T cells) infiltration ( Figure 7a, 7b) . As expected, inflammatory cell infiltration was higher in the ARDS compared to the SO lungs. However lymphocytic infiltration in the SV/luc and PBS treated septic rats was similar, indicating that the SV40 vector does not elicite an excessive cellular immune response. These findings support our previous results which demonstrated that there was no cellular immune response against the vector or the transgene following liver transduction by the SV/luc vector, measured by lymphocyte proliferation assay at 84-110 days following vector administration [30] . Moreover, literature data reported that neutralizing antibody activity against an SV40 vector was not detected in the serum of treated mice, even after 8 consecutive intraperitoneal or subcutaneous inoculations [29] . In our previous study we detected marginal humoral immune response against the vector only at very high vector concentration [30] . Two factors appear responsible of this behavior: the uncommon cell entry route (SV40 vectors evade immune surveillance most likely by entering cells via caveolar endocytosis, followed by vesicular transport directly into the endoplasmic reticulum, bypassing the endosomal-lysosomal pathway) [27] , and deletion of T antigen of the viral genome (renders the vector nonimmunogenic) [32]. The results presented in this article are only preliminary data. Currently, the most common vectors for gene transfer to the lung are replication-deficient adenoviruses. The major advantage of adenovectors is their excellent efficiency in gene transfer. However, gene expression is transient and the immunogenicity prevents repeated administrations. Moreover, the preparation of a gutless adenovirus with a more extended expression is technically demanding [42] . Figure 5 VP1 detection in type II cells. The cells were double-stained for VP1 and Pro-SP C as described in Materials and Methods. An SO lung is shown at the top row and ARDS lung in the bottom row. VP1 was detected by Cyte5 (red fluorescence, left panels) and Pro-SP C by FITC (green fluorescence, middle panels). The upper and lower right images show co-localization of both stainings inside alveolar type II cells (yellow). Shown at ×100 magnification After a detailed search of the literature data we can confirm that this is the first experiment in which an SV40 based vector was used for transduction of lung cells. The goal of our research was to provide a basis for the use of these vectors in face of their advantages over adenoviral vectors, as well as their limitations. In spite of the two major barriers met by the viruses after intratracheal administration: the mucociliary clearance system and the glycocalix [42] , in the present study we have demonstrated efficient uptake of an SV40 vector in the lungs of animals with sepsis-induced ARDS. Although nonspecific, these vectors appear to be capable of in vivo transduction of alveolar type II cells, key cells involved in regenerative process of the lung [10] . Acute respiratory disorders like ARDS are the result of a variety of endogenous and exogenous influences, arising rather as a misbalance between protective and destructive mechanisms. Transient gene therapy may thus help restore the homeostatic balance by short term expression of protective genes or suppression of damaging genes [43] . In the near future we plan taking this research a step further by replacing the reporter gene with a therapeutical one (encoding one of the surfactant proteins, Hsp70 or interleukin 6) and trying to alter the course of the disease. SV40 vectors are generally considered non-integrating, transiently expressing vectors [44] . Although others reported a prolonged expression of rSV40 vectors (1 month -1 year) [30], we did not estimate the duration of transgene expression, as our study was limited to 48 hours. Short term or transient gene expression may be sufficient and even advantageous for the treatment of an acute condition like ARDS, where a therapeutic protein or enzyme may be only needed during the acute state. However, as these vectors elicit marginal immune response, repeated administrations, if required, would be possible. Our results show moderate expression of the luciferase reporter gene. Use of a CMV promoter in the construction of the SV40 vector may improve transgene expression. However, low level of proteins expression may not be entirely disadvantageous. Supra-physiological production of specific proteins may lead to unexpected side effects. In contrast, low expression may be augmented by repeated administrations of the vector [40] . With the SV40 vector RT-PCR for luciferase mRNA detection Figure 6 RT-PCR for luciferase mRNA detection: High levels of mRNA were detected in the lungs of 2CLP animals (plasmid pGL3 used as positive control for luciferase). Minimal levels were present in the liver, kidney spleen and none in the heart. Minimal expression in the lungs, kidney, liver, spleen, and heart of T0 and SO animals. GAPDH was used as a house keeping control gene, using PCR primers specific for GAPDH. repeated administrations are possible due to its low immunogenicity. Safety considerations require that the vectors do not contain replication associated T antigen sequences. For this reason our vector is T antigen deleted. Vector propagation is achieved in COT18 cells, cell lines with minimal sequence identity with the vector. Reacquisition by rSV40 of T antigen DNA and reemergence of wtSV40 virions remains extremely unlikely [30]. Another major advantage with SV40 vectors is that they may be constructed in vitro by packaging DNA of choice in recombinant capsids [45] . These vectors, which combine efficient gene delivery of viral vectors with safety and flexibility of non-viral vectors, were shown to have similar tropism as standard SV40 vectors such as the SV/luc described here [30] . Furthermore, these vectors accommodate plasmids as large as 17 kb, significantly larger than SV40 DNA, and may be therefore used to deliver a variety of genes with complex regulatory signals [46] . Furthermore, these vectors do not require any SV40 sequences, providing additional safety margin [47] [48] [49] . The present study suggests that in vitro constructed SV40 vectors may be tailored to deliver genes of choice with lung-specific promoters for treatment of acute respiratory diseases like ARDS. Immunostaining for CD3+ T cell Figure 7 Immunostaining for CD3+ T cell. 7a: Spleen tissue from 2CLP rats serving as positive control for CD3+ T cell immunostainin 7b: left: minimal lymphocytic infiltration of the lung tissue in SO rat. Middle: moderate infiltration in a 2CLP lung rat after SV/luc administration. Right: moderate infiltration in a 2CLP lung rat not treated with SV/luc. Note that the degree of lymphocyte infiltration is similar in the middle and right panels, suggesting that it is part of the disease process rather than being induced by the vector.

Role of receptor polymorphism and glycosylation in syncytium induction and host range variation of ecotropic mouse gammaretroviruses

BACKGROUND: We previously identified unusual variants of Moloney and Friend ecotropic mouse gammaretroviruses that have altered host range and are cytopathic in cells of the wild mouse species Mus dunni. Cytopathicity was attributed to different amino acid substitutions at the same critical env residue involved in receptor interaction: S82F in the Moloney variant Spl574, and S84A in the Friend mouse leukemia virus F-S MLV. Because M. dunni cells carry a variant CAT-1 cell surface virus receptor (dCAT-1), we examined the role of this receptor variant in cytopathicity and host range. RESULTS: We expressed dCAT-1 or mCAT-1 of NIH 3T3 origin in cells that are not normally infectible with ecotropic MLVs and evaluated the transfectants for susceptibility to virus infection and to virus-induced syncytium formation. The dCAT-1 transfectants, but not the mCAT-1 transfectants, were susceptible to virus-induced cytopathicity, and this cytopathic response was accompanied by the accumulation of unintegrated viral DNA. The dCAT-1 transfectants, however, did not also reproduce the relative resistance of M. dunni cells to Moloney MLV, and the mCAT-1 transfectants did not show the relative resistance of NIH 3T3 cells to Spl574. Western analysis, use of glycosylation inhibitors and mutagenesis to remove receptor glycosylation sites identified a possible role for cell-specific glycosylation in the modulation of virus entry. CONCLUSION: Virus entry and virus-induced syncytium formation using the CAT-1 receptor are mediated by a small number of critical amino acid residues in receptor and virus Env. Virus entry is modulated by glycosylation of cellular proteins, and this effect is cell and virus-specific.

The CAT-1 receptor mediates the entry of ecotropic gammaretroviruses into rodent cells. Virus properties that rely on receptor recognition such as host range or pathogenicity could potentially be affected by polymorphisms that alter the receptor or the receptor binding domain (RBD) of the virus. In previous studies we identified two unusual ecotropic mouse leukemia virus (MLV) variants [1, 2] . Both of these viruses have altered host range, both are cytopathic, and both have amino acid substitutions at the same site in their RBDs. Spl574 is a Moloney MLV (MoMLV) variant with the substitution S82F, and F-S MLV is a Friend MLV (FrMLV) variant with the substitution S84A. Both viruses cause the formation of large multinucleated syncytia on cells derived from the wild mouse species M. dunni two days after infection, and syncytium formation is accompanied by the accumulation of large amounts of unintegrated viral DNA [2] . These two viruses also differ from each other and from their respective parental MLVs in host range. Spl574 replicates efficiently only in M. dunni cells and very inefficiently in other mouse cells such as NIH 3T3 and SC-1 cells. F-S MLV shows no unusual pattern of infectivity in mouse cells, but is capable of infecting hamster cells that are normally resistant to ecotropic MLVs. The fact that these two viruses are only cytopathic in M. dunni cells suggests involvement of the receptor-virus interaction for two reasons. First, the amino acid residue that is modified in both viruses has been identified as one of the critical amino acids forming the receptor binding site [3, 4] . Second, M. dunni cells differ from other mouse cells in their resistance to MoMLV [5] , and these cells are known to carry a modified CAT-1 receptor (dCAT-1). The dCAT-1 gene of M. dunni cells differs from the prototypical CAT-1 gene of the laboratory mouse (mCAT-1) in that the third extracellular loop that contains the virus binding region has a substitution (I214V) as well as an inserted glycine after Y235, a residue critical for receptor function [6] (Fig. 1A) . In this study, we examined the role of the dCAT-1 receptor in syncytium formation and susceptibility to infection by different ecotropic MLVs. We generated an expression vector containing dCAT-1 and transfected either this clone or the mCAT-1 gene into cells of non-rodent species that are not normally infectible by ecotropic virus. The transfected cells were then evaluated for susceptibility to infection by ecotropic MLVs and for virus induced syncytia. While virus induced syncytia were only seen in the dCAT-1 transfectants, a different panel of virus isolates was capable of efficiently infecting and/or inducing syncytia in these transfectants suggesting that virus-cell fusion and cell-cell fusion are distinct receptor mediated phenomena. The possible contribution of differential glycosylation to these phenotypic differences was evaluated using Western analysis, treatment by glycosylation inhibitors and mutagenesis to remove glycosylation sites. HA-tagged mCAT-1 and dCAT-1 clones were transfected into three cell lines that are not naturally susceptible to infection by ecotropic mouse gammaretroviruses: MA139 (ferret), Tb-1-Lu (bat lung), and MDCK (canine kidney) cells. As a control, mCAT-1 was transfected into M. dunni cells. Pools of stably transfected cells were used for analysis along with single cell derived clones of transfected MA139 cells. mCAT-1 and dCAT-1 expression in transfected cells was confirmed by Western analysis (Fig. 1B) . Consistent with previous observations [7] , CAT-1 was detected as a heterogeneously glycosylated protein in each cell line. The size range distribution for the mCAT-1 and dCAT-1 proteins was similar for each cell line, but the size range and band patterns were variable between cell lines suggesting cell specific differences in glycosylation. Thus, for example, the molecular weight range of CAT-1 was lower in MDCK cells (not shown) and Tb-1-Lu cells than in MA139 cells and M. dunni cells (Fig. 1B) . These stable transfectants of MA139, Tb-1-Lu and MDCK cells were infected with a panel of ecotropic gammaretroviruses including two, Spl574 and F-S MLV that induce multinucleated syncytia in M. dunni cells but not in other mouse cell lines. The infected cells were examined for cytopathicity over a period of 2-5 days. Transfectants of all 3 cell lines expressing mCAT-1 showed no signs of cytopathicity following virus infection as shown for the MA139 and MDCK transfectants in Fig. 2 . In contrast, dCAT-1 expressing MA139 and MDCK cells (Fig. 2 ) as well as Tb-1-Lu cells (not shown) formed syncytia within two days of infection with Friend virus isolate F-S MLV. Several separate pools of MA139 transfected cells were generated and tested. Virus-induced syncytia were observed in two independently derived pools of dCAT-1 transfected MA139 cells as well as three independently isolated clonal lines (FerrD2, N65FerrC2 and N65FerrB6), but not in 3 independently derived pools of mCAT-1 transfected MA139 cells. Among the ecotropic isolates tested, F-S MLV was most efficient in inducing syncytia in all dCAT-1 transfected cells (Fig. 2 ), but syncytium formation was also observed following infection with the Friend MLV isolates FBLV and FrMLV57. Infection with MoMLV or Spl574 occasionally resulted in syncytium formation in these transfectants, but these syncytia were smaller and fewer in number, and, appeared 1-2 days after the appearance of syncytia in parallel cultures infected with the most cytopathic isolate, F-S MLV. Thus, cells expressing the dCAT-1 receptor can, like M. dunni, produce syncytia in response to virus infection, but these transfectants differ from M. dunni in their relative insensitivity to Spl574 and their sensitivity to syncytium formation by virus isolates that are not typically cytopathic in M. dunni cells. Virus induced syncytium formation in M. dunni cells was previously shown to be accompanied by the appearance of high levels of unintegrated viral DNA [2] , a phenomenon also observed for other pathogenic retroviruses [8] . To determine if the transfected cells show this same response to cytopathic virus, we extracted Hirt DNA from CAT-1 transfected MA139 cells 3 days after infection with F-S MLV (Fig. 3A ). At the time of DNA extraction, FerrM cells, expressing mCAT-1, showed no cytopathic response and the observed level of unintegrated viral DNA was low (4.0% of M. dunni, Fig. 3A ,B). In contrast, virus-induced syncytia were observed in all 3 dCAT-1 transfectants. 2 of these 3 transfectants had large multinucleated syncytia that involved >50% of the cells in infected cultures (Fig. 3C) ; levels of unintegrated linear DNA in these cells were high (53% and 68% of M. dunni) (Fig. 3A,B) . The third dCAT-1 transfectant showed fewer and smaller syncytia ( dCAT-1-g : : : : V : : : : : : : : E : : : : : V : : : : : : G : : : : Thus, viral DNA accumulation is observed in dCAT-1 but not mCAT-1 transfectants, increased viral DNA is associated with virus-induced cytopathicity in the transfected cells, and the amount of viral DNA varies with the severity of the cytopathic response. To define the relationship between syncytium formation and productive virus infection, transfected cell lines carrying either mCAT-1 or dCAT-1 were tested for susceptibility to a panel of ecotropic MLVs using the XC plaque overlay test (Table 1 ). In this assay, clusters of infected cells expressing ecotropic Env glycoprotein are identified by plaques of syncytia formed by overlaid rat XC cells [9] . For the cytopathic viruses Spl574 and F-S MLV, the number of syncytia induced directly by these viruses in susceptible cells is approximately equivalent to the titer determined by this XC overlay assay; for example, parallel cultures of infected M. dunni cells produced an XC titer of 10 5.1 (Table 1 ) compared to Spl574 syncytium titer of 10 4.6 . Transfected M. dunni cells expressing mCAT-1 in addition to the endogenous dCAT-1 gene were significantly more susceptible to MoMLV infection than untransfected M. dunni cells (Table 1) , consistent with a previous study indicating that the dCAT-1 sequence variation is responsible for M. dunni resistance to MoMLV [6] . No difference was noted in the XC plaque titer of Spl574 in M. dunni cells expressing mCAT-1 in addition to the endogenous dCAT-1, and no viruses other than Spl574 and F-S MLV were cytopathic in the mCAT-1 transfected M. dunni cells. The differences between M. dunni and NIH 3T3 cells in susceptibility to ecotropic viruses were not reproduced in MA139 cells expressing dCAT-1 (FerrD2) or mCAT-1 (FerrM). In fact, there were no significant differences in the XC titers of different MLVs in FerrD2 and FerrM (Table 1) . FBLV, F-S MLV, and, surprisingly, MoMLV efficiently infected both FerrM and FerrD2 with slightly higher XC titers for all viruses in FerrD2. Also, even though Spl574 efficiently replicates in M. dunni, Spl574 produced comparably low XC titers in both FerrD2 and FerrM. Thus, FerrD2 does not resemble M. dunni cells in its susceptibility to infection by MoMLV and Spl574; this difference suggests the involvement of additional factors independent of the CAT-1 receptor sequence. The cytopathicity of different virus isolates did not always correlate with the efficiency of virus replication in FerrD2 as determined by XC virus titer. While on the one hand, Spl574 produced low XC titers on FerrD2 (Table 1) and was also poorly cytopathic, high XC titer viruses did not all produce syncytia in these cells. Thus, the most cytopathic virus in FerrD2 cells, F-S MLV, produced an XC titer comparable to that of the rarely cytopathic MoMLV. Efficient virus replication is thus not sufficient to generate a cytopathic response. To further investigate the observed differences in XC titers for cells expressing different CAT-1 genes, we assessed infectivity using viral pseudotypes in a single round infectivity assay ( The Spl574 pseudotype is restricted in NIH 3T3 cells as is the Spl574 virus (Tables 1, 2) suggesting that this restriction is entry related. In contrast, the Spl574 pseudotype was not restricted in FerrD2 or FerrM cells although Spl574 virus produces low XC titers in both of these transfectants. This shows that the failure of Spl574 to replicate efficiently in the transfected cells is not entry related and suggests the involvement of factor(s) restricting post-entry stages of Spl574 virus replication in ferret cells. M. dunni and FerrD2 cells express the same dCAT-1 receptor, but these cells differ in their relative infectivity by MoMLV and Spl574, and they produce syncytia in response to different virus isolates. One possible explana- tion for these differences is that CAT-1 may undergo different post-translational modification in the two cell lines. It has been shown that resistance of M. dunni cells to MoMLV infection is reduced by treatment with the inhibitor of glycosylation, tunicamycin (Tu) [10] . The involvement of glycosylation is also suggested by the observation that the CAT-1 glycosylation patterns differ in transfected MA139 and M. dunni cells ( Fig. 1B ; lanes e,f). To determine if glycosylation contributes to the observed differences, we generated a dCAT-1 clone from which the N-glycosylation sites had been removed. The CAT-1 protein has two glycosylation sites, and both carry N-glycans [7] . Both sites are in the third extracellular loop which also contains the residues implicated in virus binding and entry [11, 12] . Both glycosylation sites were removed by PCR mediated site-specific mutagenesis from the dCAT-1 variant (Fig. 1) , and the resulting clone, dCAT-1-g, was transfected into MLV, nor did it result in syncytium formation by viruses not cytopathic in M. dunni cells. The transfectants, however, showed increased susceptibility to MoMLV compared to untransfected M. dunni cells (Table 3 , Exp.1), as also shown above for M. dunni(mCAT-1) ( Table 1) ; the transfectants showed no increase in their susceptibility to other ecotropic viruses. This is consistent with the conclusion that glycosylation of dCAT-1 is associated with MoMLV resistance. MA139 cells expressing dCAT-1-g resembled FerrD2 in their susceptibility to virus infection ( Table 3 , Exp. 2 and Table 1 ) and sensitivity to F-S MLV-induced syncytia (Fig. 4B ). The cells with the unglycosylated receptor were, like FerrD2, efficiently infected by MoMLV. Syncytia were produced in these transfectants with the same viruses that are cytopathic in FerrD2, and no cytopathic response was observed with viruses that are also noncytopathic in FerrD2. Thus, the complete absence of N-glycans on dCAT-1 did not alter the ability of the dCAT-1 receptor to mediate virus induced syncytium formation in MA139 cells, nor did it alter the panel of viruses that were cytopathic and/or infectious in the transfectants. The glycosylation inhibitor tunicamycin (Tu) was previously shown to reduce resistance to MoMLV in M. dunni cells [10] . We tested the ability of multiple glycosylation inhibitors to alter infectivity of ecotropic MLVs in mouse cells expressing the two functional CAT-1 variants: mCAT-1 (NIH 3T3 cells) and dCAT-1 (M. dunni cells). The 6 inhibitors included Tu which blocks generation of the carbohydrate-dolichol precursor needed for N-linked glycosylation, the sugar analog 2-deoxy-D-glucose (2DG), and 4 inhibitors which inhibit different enzymes involved in oligosaccharide trimming: castanospermine (CST), deoxymannojirimycin (DMM), deoxynojirimycin (DNM) and swainsonine (Sw). Western analysis of M. dunni cells transfected with HA-tagged mCAT-1 (Fig. 5A) showed that none of the inhibitors had a significant effect on expression levels, although all inhibitors reduced the size range of the mCAT-1 glycoprotein. Effect of glycosylation inhibitors on expression of HA-tagged mCAT-1 in M. dunni cells Because the resistance of NIH 3T3 cells to Spl574 infection is comparable to the resistance of M. dunni cells to MoMLV, we treated NIH 3T3 cells with 5 different glycosylation inhibitors before Spl574 infection (Table 4 ). All 5 inhibitors significantly reduced resistance to Spl574 replication, but inhibition of N-glycosylation did not affect the XC titer of other ecotropic viruses in NIH 3T3 cells, as shown for MoMLV. Resistance of SC-1 cells to Spl574 [1] is similarly relieved by glycosylation inhibitors (data not shown). M. dunni cells were also treated with the same set of glycosylation inhibitors prior to virus infection ( Table 4 ). All inhibitors reduced the resistance of M. dunni cells to infection with MoMLV, but no comparable increase in titer was noted with Spl574. To confirm that this effect is on entry, DMM-treated M. dunni cells were infected with LacZ pseudotypes of MoMLV; pseudotype titer was 10 3.6 on DMMtreated cells compared to no detectable LacZ expressing cells in untreated M. dunni. To determine if altered infectivity results from inhibitormediated changes in cell surface receptor levels, we measured biotinylated CAT-1 in M. dunni cells transfected with HA-tagged mCAT-1. As shown in Figure 5B , surface mCAT-1 in DMM-treated cells shows the expected reduction in size because of the predominance of smaller highmannose N-glycans, but quantitation of this expression by densitometric scanning shows that the level in DMM treated cells is not significantly different from the untreated control. These results, taken together, indicate that N-glycans can impede ecotropic MLV entry in cells expressing mCAT-1 as well as cells expressing dCAT-1, and that these N-glycans obstruct different ecotropic isolates in NIH 3T3 and M. dunni cells. Also, the fact that the effect on entry is seen with inhibitors other than Tu suggests that inhibition may be due to N-glycan type or size. Three factors contribute to the observed variations in host range and/or cytopathicity of mouse ecotropic gammaretroviruses: specific sequence differences in the viral env, differences in the CAT-1 receptor, and glycosylation of cellular proteins. The role of specific env sequence variations in virus-induced syncytium formation was previously suggested by our identification of two MLV isolates that are uniquely cytopathic in M. dunni cells. Both isolates have amino acid substitutions at the same RBD residue that is critical for receptor binding: S82F in Spl574 and S84A in F-S MLV. That mutations in the viral receptor binding site contribute to cytopathicity is also supported by the observation that a third MLV variant, TR1.3, is cytopathic in SC-1 cells and brain endothelial cells because of a single substitution, W102G [13] , at a site that together with S82 and D84 forms the receptor binding site [3, 4] . The involvement of CAT-1 in the cytopathic response in M. dunni cells was suggested by the specific sequence differences that distinguish the dCAT-1 receptor variant from mCAT-1. These 2 receptors differ by 4 amino acids of which two are within the third extracellular CAT-1 loop that contains the virus binding site: I214V, and a glycine insertion within the YGE virus binding site [6] . As shown in the present paper, all cells expressing the dCAT-1 variant and none expressing mCAT-1 are susceptible to virusinduced syncytium formation. This indicates that one or both of these two amino acid changes, I214V and Δ236G, are responsible for the cytopathic response mediated by this receptor variant. Previous studies with cytopathic retroviruses such as HIV have identified the accumulation of unintegrated DNA as a hallmark of cytopathicity [8] . Analysis of MA139 cells expressing the naturally occurring mouse receptor types, mCAT-1 and dCAT-1, shows that receptor type also correlates with this aspect of cytopathicity, and that in different dCAT-1 transfected lines the amount of unintegrated DNA corresponds to the extent of syncytium formation. This cell-virus system may thus be useful in further studies on the mechanisms thought to be involved in this cell killing such as endoplasmic reticulum stress induced apoptosis [14] . It is known that the glycans on various cell surface receptors can modulate virus entry (for example, [15] ). The CAT-1 receptor is glycosylated at two sites, and previous studies have shown that glycosylation inhibitors reduce resistance to ecotropic MLV infection in rat and hamster cells expressing the rCAT-1 and haCAT-1 receptor variants [16] [17] [18] [19] , as well as resistance to MoMLV in M. dunni cells with dCAT-1 [10] . It has also been shown that in mink cells expressing mCAT-1, glycosylation affects SU binding and the down-modulation of receptor by virus infection [20]. Our results show that glycosylation modulates virus entry mediated by the laboratory mouse CAT-1 receptor, mCAT-1, in NIH 3T3 cells. This resistance is specific to Spl574 and is not seen in heterologous cells expressing mCAT-1. The control of this differential sensitivity of mCAT-1 to a specific ecotropic isolate by cell specific glycosylation has not been previously described. The present study also considered whether altered glycosylation could explain why two cells expressing the same dCAT-1 receptor, M. dunni and FerrD2, produce syncytia in response to different viruses. As shown by the inhibitor results, however, while N-glycans contribute to the restriction of MoMLV entry into M. dunni cells, comparisons of ferret transfectants expressing the dCAT-1 or dCAT-1-g receptor variants produced no evidence that N-glycans modulate virus infectivity or virus-induced cytopathicity in the MA139 cells. N-glycans can have high mannose, complex or hybrid structures. The various glycosylation inhibitors target different steps in protein glycosylation and can be used to manipulate the carbohydrate composition of glycoproteins. The inhibitor CST blocks glucose trimming, and DMM and SW inhibit successive steps in mannose trimming. The fact that all of these inhibitors along with the sugar analog 2DG and glycosylation inhibitor Tu relieved the resistance of M. dunni cells to MoMLV and of NIH 3T3 to Spl574 suggests that these viruses are most effectively blocked by the large complex oligosaccharides produced in the terminal stages of glycosylation. These results, taken together, suggest roles for N-glycans in virus entry that are virus-specific and cell-specific, and also indicate that this regulation may be sensitive to small sequence changes in both virus and receptor. These results indicate that N-glycans broadly regulate ecotropic gammaretrovirus interactions with the CAT-1 receptor in cells of their natural host [21] , although it is possible that glycosylated proteins other than CAT-1 may contribute to this resistance. Our demonstration that not all infectious viruses are cytopathic in M. dunni and FerrD2 cells supports the idea that virus-cell fusion and cell-cell fusion are distinct receptormediated phenomena. A similar lack of correlation between infectivity and syncytium formation has been reported, for example, in a mouse cell line that is unusual in its resistance to HTLV Env-mediated syncytium formation although it is highly susceptible to virus infection [22] . It has also been shown that, for a transformed NIH 3T3 cell line subject to MoMLV-induced syncytium formation, chloroquine treatment blocks MoMLV entry but does not also block syncytium formation [23] . Our results further distinguish cell fusion and virus entry as separate receptor functions. Finally, these studies also identify differences between M. dunni and FerrD2 cells that are clearly not receptor mediated. Use of LacZ pseudotypes shows that Spl574 Envs efficiently mediate entry into FerrD2 cells, but XC titers in Spl574 virus infected FerrD2 cells are clearly reduced as is virus-induced syncytium formation. This indicates a postentry block to virus replication leading to reduced surface Env, and the nature of this block is under investigation. The CAT1 receptor mediates ecotropic gammaretrovirus entry and the cytopathic response to virus infection. Use of virus env variants, receptor mutations, and inhibitors of glycosylation demonstrate that both of these virus-receptor interactions are modulated by a small number of critical amino acid residues in virus and receptor, and that Nlinked glycans can modulate entry for specific virus-cell combinations. Virus stocks were made by collecting culture fluids from infected or transfected cells. These stocks were titered by the XC overlay test [9] following infection of NIH 3T3, SC-1 [24] , or M. dunni [5] and cells transfected with CAT-1 receptor. Cells were plated at 1-2 × 10 5 cells/60 mm dish and infected with 0.2 ml of appropriate dilutions of virus stocks in the presence of polybrene (4 ug/ml; Aldrich, Milwaukee, WI). Cells were irradiated 4 days after virus infection with ultraviolet light from germicidal bulbs (30 sec at 60 ergs/mm 2 ) to kill the cells but not the virus, and were then overlaid with 10 6 XC cells/plate. XC cells produce plaques containing syncytia in response to focal areas of virus infected cells. Plates were fixed and stained 3 days later and examined for plaques of syncytia. To screen for the formation of multinucleated syncytium in virus infected cells, 2 × 10 4 cells in six-well tissue culture plates or 10 5 cells in 60 mm plates were infected with virus-containing medium in the presence of 4 ug/ml polybrene. After 2-4 days, the cells were examined by light microscopy using objective lenses of 4×-20× and photo-

The presence of the TAR RNA structure alters the programmed -1 ribosomal frameshift efficiency of the human immunodeficiency virus type 1 (HIV-1) by modifying the rate of translation initiation

HIV-1 uses a programmed -1 ribosomal frameshift to synthesize the precursor of its enzymes, Gag-Pol. The frameshift efficiency that is critical for the virus replication, is controlled by an interaction between the ribosome and a specific structure on the viral mRNA, the frameshift stimulatory signal. The rate of cap-dependent translation initiation is known to be altered by the TAR RNA structure, present at the 5′ and 3′ end of all HIV-1 mRNAs. Depending upon its concentration, TAR activates or inhibits the double-stranded RNA-dependent protein kinase (PKR). We investigated here whether changes in translation initiation caused by TAR affect HIV-1 frameshift efficiency. CD4+ T cells and 293T cells were transfected with a dual-luciferase construct where the firefly luciferase expression depends upon the HIV-1 frameshift. Translation initiation was altered by adding TAR in cis or trans of the reporter mRNA. We show that HIV-1 frameshift efficiency correlates negatively with changes in the rate of translation initiation caused by TAR and mediated by PKR. A model is presented where changes in the rate of initiation affect the probability of frameshifting by altering the distance between elongating ribosomes on the mRNA, which influences the frequency of encounter between these ribosomes and the frameshift stimulatory signal.

The precursor of HIV-1 structural proteins, Gag, and the precursor of the viral enzymes, Pol, are translated from the full-length viral messenger RNA (mRNA). Gag is produced by conventional translation whereas Pol requires a programmed -1 ribosomal frameshift during the elongation step of translation, which generates the fusion protein Gag-Pol (1, reviewed in 2, 3) . Previous studies showed that a 2-to 20-fold increase in the Gag-Pol to Gag ratio prevents viral infectivity (4-7) and our group showed that a decrease in the frameshift efficiency as low as 30% severely impairs the replication of the virus in cultured cells (8) . The Gag-Pol to Gag ratio is therefore critical for viral infectivity and the programmed -1 frameshift that determines this ratio represents an interesting target for the development of novel antiretroviral agents against HIV-1. The HIV-1 frameshift event requires two cis-acting elements in the viral mRNA: a slippery sequence, UUUUUUA, where the frameshift occurs (1, reviewed in 2,3), followed by an irregular stem-loop (9) (10) (11) , the frameshift stimulatory signal, that makes the ribosomes pause over the slippery sequence and controls the frameshift efficiency. Only a fraction of the ribosomes that encounter the stimulatory signal make a frameshift. After the pause, the ribosomes unfold the signal, which can reform after their passage. HIV-1 can use a cap-dependent mechanism to initiate translation of its mRNAs, like most eukaryotic mRNAs (for a review on translation initiation, see [12] [13] [14] [15] . There are two major control steps in eukaryotic cap-dependent translation initiation (see details in Figure 1A ). One is the binding of the initiator tRNA, Met-tRNA i Met , to the 40S ribosomal subunit, which requires the participation of the initiation factor 2 (eIF2) associated to GTP. The other one is the binding of the 40S subunit bearing the ternary complex to the 5 0 cap structure of the mRNA, which is controlled by the eIF4F complex. Double-stranded RNA (dsRNA), such as the TAR RNA structure, can modify the rate of translation initiation. TAR, the transactivation response element, is a 59-nt stem-bulge-loop structure present at the 5 0 and 3 0 end of all HIV-1 mRNAs in the nucleus and the cytoplasm (reviewed in 16) . It is also present under a free form of 58-66 nt in the cytoplasm of cells infected with the virus (17, 18) . In the nucleus, TAR mediates transcription activation by binding to the viral Tat protein and the cellular cyclinT protein (19, 20) . In the cytoplasm, a low concentration of TAR activates PKR, the dsRNA-dependent protein kinase, whereas a higher concentration of TAR inhibits this kinase by blocking its dimerization, which is essential for its activity (reviewed in 21) . When PKR is activated, it phosphorylates the a subunit of eIF2, interfering with translation initiation, whereas, when it is inhibited, the amount of eIF2 phosphorylated decreases and the rate of translation initiation increases. In this study, we investigated whether the presence of TAR affects HIV-1 frameshift efficiency in relationship with the changes it causes in the rate of cap-dependent translation initiation. To this end, we used a dualluciferase construct (8) which expresses the Renilla luciferase (Rluc) and the firefly luciferase (Fluc) separated by HIV-1 frameshift region as a fusion protein. Rluc is expressed following conventional rules of translation whereas Fluc expression requires a -1 frameshift in the HIV-1 frameshift region. This type of construct is adapted from Grentzmann et al. (22) , who pioneered the use of a dual-luciferase reporter for studying recoding signals. CD4+ T cells (Jurkat) or 293T cells were transfected with the dual-luciferase plasmid and TAR was added either in cis or in trans of the reporter mRNA. Several conditions were assayed to characterize the effect of TAR on frameshift efficiency and the involvement of PKR in this effect, such as the introduction of a small or a large amount of TAR in the cells, the use of mutants of TAR that cannot perturb PKR activity and the silencing of PKR expression with short interfering RNA (siRNA). Our results show that HIV-1 frameshift efficiency increases at a low concentration of TAR, when capdependent translation initiation is slowed down, whereas it decreases at a high concentration of TAR, when translation initiation is stimulated. These effects were shown to be dependent on PKR. A model is presented which relates the effects of TAR on frameshift efficiency to changes in the spacing between the elongating ribosomes on the mRNA caused by changes in the rate of translation initiation. Such changes affect the frequency of encounter between the ribosomes and the frameshift stimulatory signal. To measure HIV-1 frameshift efficiency, we used the dualluciferase reporters pDual-HIV(-1) and (0) (8) . These plasmids are derived from pcDNA3.1Hygro+ (Invitrogen) and contain the HIV-1 frameshift region inserted between the coding sequences of the Renilla (15) . The figure is adapted from Gebauer and Hentze (14) . Only the factors we refer to in the text are named. The 40S ribosomal subunit associates with the ternary complex [initiation factor 2 (eIF2) plus GTP plus the initiator tRNA, Met-tRNA i Met ] and with other factors, and binds to the 5 0 cap structure of the mRNA. This binding requires the eIF4F complex formed by three initiation factors: eIF4E, the cap-binding protein, eIF4G, a scaffold protein and eIF4A, a RNA helicase that unfolds secondary structures. After each round of initiation, eIF2 is released from the ribosome in association with GDP. Phosphorylation of the a subunit of eIF2 (eIF2-a) prevents the recycling of eIF2-GDP in eIF2-GTP, blocking translation initiation. Thapsigargin induces endoplasmic reticulum stress, which stimulates the PERK kinase that phosphorylates eIF2-a, reducing the level of functional eIF2 (55) (56) (57) . Rapamycin shuts down the mammalian target of rapamycin (mTOR) pathway, which blocks the phosphorylation of the translation repressor 4E-BP, and hypophosphorylated 4E-BP sequesters the initiation factor eIF4E (58, 59) . Hippuristanol is a selective inhibitor of eIF4A (60) , which interferes with the binding of the 40S subunit to the mRNA. luciferase (Rluc) and the firefly luciferase (Fluc). Expression of these genes is under control of a CMV promoter, which is followed by a T7 promoter. Plasmid pDual-HIV(0) differs from pDual-HIV(-1) by the addition of an adenine after the slippery sequence in the frameshift region. Derivatives of pDual-HIV(-1) and (0) were constructed where the TAR sequence was inserted after the CMV and T7 promoters. A TAR-containing fragment flanked with HindIII sites obtained from pcDNA3-RSV-TAR-Rluc plasmid (23), a kind gift from L. DesGroseillers (Universite´de Montre´al), was cloned in the HindIII site of pDual-HIV to produce pDual-HIV-TAR(-1) and (0), where the TAR sequence is located at a distance of about 40 nt from the 5 0 end of the reporter mRNA. To produce pDual-HIV-50TAR(-1) and (0), where the TAR sequence is at a larger distance from the 5 0 end of the reporter mRNA, a cassette of a 50-nt noncoding sequence was inserted in the AflII site of pDual-HIV, followed by the insertion of TAR immediately after these 50 nt, in the HindIII site. The oligonucleotides for the cassette were cass50nt-fwd and cass50nt-rev (see the sequence of all the oligonucleotides used in this study in Table 1 of the Supplementary Data). Plasmid pTAR, which expresses the free TAR sequence in trans from the reporter mRNA, was made by inserting the TARcontaining fragment flanked with HindIII sites into the HindIII restriction site of pcDNA3.1Hygro+. Derivatives of pTAR, pTARuucg à and pTARibulge à , which express mutants of TAR, were constructed by cloning oligonucleotide cassettes (cass_TAR-uucg à fwd and cass_TAR-uucg à rev or cass_TAR-bulge à fwd and cass_TAR-bulge à rev) between the two NheI restriction sites present in the TAR sequence of pTAR. In the first mutant, the upper loop, CUGGGA, is replaced with UUCG and, in the second mutant, the bulge UCU preceding the upper loop is deleted. Plasmid pCGNiC [a generous gift from N. Hernandez, Cold Spring Harbor Laboratory (24) ] expresses a mutant of the TAR-binding protein Tat (Tat à ), named TatC30,31A. Jurkat cells (CD4+ T cells) were maintained in RPMI 1640 medium (Wisent) supplemented with 10% (v/v) FBS (Wisent) and HEK 293T cells (human embryonic kidney cells transformed with adenovirus and simian virus 40 large-T) were maintained in DMEM (Gibco) supplemented with 10% (v/v) FBS. Transfections were performed with polyethylenimine (PEI) (Polysciences, Inc.) in six-well plates containing Jurkat cells (1.2  10 6 ), 293T cells (4.0  10 5 ) or 293T stable transfectants (6.0  10 5 cells) expressing a dual-luciferase HIV reporter (see subsequently). PEI was added drop-wise to serum-free medium and incubated 10 min at room temperature. In parallel, serum-free medium was added to DNA. The diluted PEI was added to the DNA solution (PEI to DNA ratio of 2:1) and incubated at least 15 min at room temperature. An empty plasmid, pcDNA3.1Hygro+, was added, when required, to maintain an equivalent DNA input. Translation inhibitors were added as follows: rapamycin (Fisher), 16 h post-transfection (final concentration: 25 nM), hippuristanol (a generous gift from J. Pelletier, McGill University), 24 h before harvest (final concentration: 400 nM) and thapsigargin (Sigma), 4 h before harvest (final concentration: 300 nM). Transfected cells were harvested 48 h post-transfection. Non-adherent cells were centrifuged at 3000 g for 5 min, washed with PBS and lysed in 100 ml of Cell Passive Lysis Buffer (Promega). Adherent cells were washed with PBS and lysed in 400 ml of Cell Passive Lysis Buffer. Cell lysates were centrifuged 2 min at 13 000 g at 48C to remove cell debris, before luciferase assays. Plasmids pcDNA5-Dual-HIV(-1) and (0) were made by inserting the HindIII-ApaI fragment from pDual-HIV(-1) or (0), respectively, into pcDNA5-FRT (Invitrogen), which contains a resistance gene to hygromycin B. An in-frame construct without the HIV-1 frameshift region was generated by cloning an oligonucleotide cassette (inframe-fwd and inframe-rev) into the KpnI and BamHI restriction sites of linearized pDual-HIV. In pDual-in-frame, the luciferase coding sequences are in the same reading frame and separated by a short linker. The HindIII-ApaI fragment from pDual-in-frame was cloned into pcDNA5-FRT. Cell lines stably expressing the (-1) or (0) dual-luciferase HIV reporter, or the in-frame construct, were generated following the manufacturer's instructions, using 293T Flp-in TM cells (Invitrogen). Individual clones that stably incorporated the plasmids were selected on the basis of their resistance to hygromycin B (Wisent) (250 mg/ml) and maintained in hygromycin B. 293T transfectants (6.0  10 5 cells) stably expressing the (-1) and (0) dual-luciferase HIV reporter were transfected with 150 ng of the PKR ShortCut Õ siRNA Mix or the eGFP ShortCut Õ siRNA Mix (New England BioLabs), using PEI. The TAR-expressing plasmids were transfected 24 h after the transfection with a siRNA mix. Cells were harvested 48 h after this second transfection and luciferase assays were performed. 293T transfectants, transfected with a siRNA mix, as described above, were harvested 48 h after the transfection, washed in PBS and lysed in 100 ml of Ripa-Doc (final concentration: 140 mM NaCl, 8 mM Na 2 HPO 4 , 2 mM NaH 2 PO 4 , 1% Nonidet P-40, 0.5% sodium deoxycholate and 0.05% sodium dodecyl sulphate), containing a cocktail of protease and phosphatase inhibitors. Equal amounts of proteins (15 mg) were separated on a 10% SDS-PAGE gel, transferred on a nitrocellulose membrane and immunoblotted with a mouse anti-PKR hybridoma supernatant (clone F9) (a generous gift from A. Koromilas, McGill University) and a horseradish peroxidase-conjugated goat anti-mouse secondary antibody (Amersham) diluted 1/1500. After detection of the antigen-antibody complexes, the membrane was washed with 25 ml of stripping buffer (final concentration: 0.08 M b-mercaptoethanol, 2% sodium dodecyl sulphate and 0.06 M Tris-HCl, pH 6.9) for 30 min at 508C, and immunoblotted with a mouse antia-tubulin monoclonal antibody (clone B-5-1-2 Sigma) diluted 1/5000 and a horseradish peroxidase-conjugated goat anti-mouse secondary antibody diluted 1/1500. Antigen-antibody complexes were detected with an enhanced chemiluminescence (ECL) system. The Fluc versus the Rluc activities of the (-1) and (0) constructs were measured as relative light units with a Berthold Lumat LB 9507 luminometer, as previously described (8) . A Dual-Luciferase Reporter Assay System kit (Promega) was used for Jurkat cells and home-made reagents (25) were used for 293T cells. The Rluc activity is used to normalize the Fluc activity (Fluc/Rluc). The frameshift efficiency is equal to: ½FlucðÀ1Þ=RlucðÀ1Þ= ½Flucð0Þ=Rlucð0Þ þ FlucðÀ1Þ=RlucðÀ1Þ: Our aim was to investigate whether the presence of TAR affects HIV-1 frameshift efficiency in relationship with its effect on cap-dependent translation initiation. To this end, we used a dual-luciferase construct, pDual-HIV(-1), which contains the Rluc and the Fluc reporter genes separated by the HIV-1 frameshift region ( Figure 1B ). In this construct, the Fluc is produced only by ribosomes that make a -1 frameshift when translating the HIV-1 frameshift region. To assess the frameshift efficiency, we used a control construct, pDual-HIV(0), in which an adenine is added after the slippery sequence in the frameshift region, so that the Fluc coding sequence is in-frame with the Rluc coding sequence. The Rluc is synthesized by conventional translation in both (-1) and (0) constructs. Before investigating the effect of TAR, we verified that changes in cap-dependent translation initiation affect HIV-1 frameshift efficiency. Jurkat cells, a CD4+ T-cell line, were transfected with pDual-HIV(-1) or (0) plasmids and treated with thapsigargin, rapamycin or hippuristanol, three inhibitors perturbing a different step of capdependent translation initiation ( Figure 1A ). The frameshift efficiency, which is 5.1 AE 0.4% in the absence of inhibitors, was increased about twofold in the presence of either one of these three inhibitors ( Figure 1C) . The presence of a high amount of TAR decreases HIV-1 frameshift efficiency We next assessed the effect of TAR on the frameshift efficiency. TAR (Figure 2A ) was inserted at about 40 nt from the 5 0 end of the mRNA in pDual-HIV, generating pDual-HIV-TAR(-1) and (0) ( Figure 2B ). We avoided placing TAR at the very end of the mRNA, since such a position could interfere with the binding of the 40S subunit to the messenger (23,26,27 and references therein). We first examined the effect of a high amount of TAR that inhibits PKR and stimulates translation initiation (21) . The frameshift efficiency was assessed in Jurkat and 293T cells. When 2 mg of pDual-HIV-TAR were delivered into the cells, the frameshift efficiency was decreased to 70% of its value in absence of TAR in either Jurkat or 293T cells ( Figure 2C and D) . Under the conditions of these assays, the frameshift efficiency in absence of TAR was 6.1 AE 0.2% in Jurkat cells and 11.3 AE 0.9% in 293T cells. These values, and the value of 5.1 AE 0.4% observed in the experiment described in the preceding section with Jurkat cells that were transfected under slightly different conditions (see details in 'Materials and Methods' section), are comparable to the values obtained with different heterologous systems containing the HIV-1 frameshift region, which were shown to range between 2 and 10% in mammalian cultured cells (8, 22, 28, 29) . It can be recalled here that several groups observed that the absolute value of the frameshift efficiencies changes, depending upon various parameters such as the conditions used for the assay and the type of cultured cells (30) . We then investigated whether the decrease in frameshift efficiency observed with pDual-HIV-TAR was influenced by the position of TAR in cis or in trans from the reporter mRNA. Two other constructs were used, pDual-HIV-50TAR, where the distance between TAR and the 5 0 end of the reporter mRNA was increased by 50 nt compared to pDual-HIV-TAR, and pTAR, that provides TAR in trans from the reporter mRNA expressed from pDual-HIV ( Figure 2B ). The frameshift efficiency was decreased to 75 and 60%, respectively, in Jurkat cells and 293T cells transfected with pDual-HIV-50TAR compared to the value in absence of TAR. When Jurkat and 293T cells were co-transfected with 2 mg of pDual-HIV and 2 mg of pTAR, the frameshift efficiency was reduced to 70% of its value in absence of TAR, a decrease similar to that observed when TAR was present in cis of the reporter mRNA ( Figure 2C and D) . These results indicate that it is the presence of TAR in the cells and not its presence in the reporter mRNA that decreases HIV-1 frameshift efficiency. The effect of TAR on the frameshift efficiency was confirmed when using an infection system to deliver the reporters into the cells (see Figure 1 in the Supplementary Data). To verify that PKR was involved in the changes in HIV-1 frameshift efficiency observed with a high amount of TAR, we created two constructs, pTARibulge à and pTARuucg à , expressing mutants of TAR that cannot bind PKR (31) ( Figure 3A ). When Jurkat cells were co-transfected with pDual-HIV and plasmids generating these TAR mutants, the frameshift efficiencies were similar to that obtained in absence of TAR and significantly higher than the value obtained in the presence of wild-type TAR ( Figure 3B ). This result supports that PKR is involved in the changes of frameshift efficiency observed in the presence of TAR. To further confirm that inhibiting PKR decreases HIV-1 frameshift efficiency, a plasmid expressing Tat, a HIV-1 viral protein, was co-transfected with the dualluciferase plasmids. In addition to its well-characterized transactivation effect on transcription of the viral mRNAs by binding to TAR, Tat influences translation by inhibiting PKR, either directly by binding this kinase or indirectly by blocking the binding of TAR to PKR (32, 33) . We used a Tat mutant (Tat à ) that can bind TAR and inhibit PKR but cannot transactivate transcription, and, thereby, that does not affect mRNA levels (24) . Jurkat cells were co-transfected with the plasmid coding for this Tat mutant and with pDual-HIV, pDual-HIV-TAR or pDual-HIV-50TAR. In the presence of Tat à , the frameshift efficiency was decreased to approximately 60% of its value in absence of Tat à ( Figure 3C ). The decrease with Tat à was the same, whether TAR was present or not, which suggests that Tat à and TAR both act via the same mechanism, the inhibition of PKR. Next, we investigated the effect of a small amount of TAR, which activates PKR and thus interferes with translation initiation (21) . We used stable 293T transfectants expressing a dual-luciferase HIV reporter. Stable transfectants expressing a (-1) or (0) dual-luciferase HIV reporter were transfected with pTAR, pTARibulge à or pTARuucg à in amounts ranging from 0 to 2.3 mg. Figure 4A shows the effect of wild-type TAR. In the presence of a small quantity of TAR, the frameshift efficiency increases to about 140% of its value in absence of TAR but with a larger quantity of TAR, the frameshift efficiency decreases to about 80%, a decrease comparable to that observed with a transient transfection of pDual-HIV ( Figure 2 ). As a control, we used stable 293T transfectants expressing Rluc and Fluc in-frame, separated by a linker instead of the HIV-1 frameshift region. The ratio of Fluc activity to Rluc activity in lysates from these transfectants was unchanged in the presence of pTAR (data not shown), confirming that changes in the Fluc to Rluc ratio observed with stable transfectants expressing the dual-luciferase HIV reporter are due to variations in the frameshift efficiency. When the stable 293T transfectants expressing the dual-luciferase HIV reporter were transfected with plasmids producing TAR mutants that cannot bind PKR, the frameshift efficiency was unaltered ( Figure 4B) . The effect of a low amount of TAR was also assessed by transient co-transfection of Jurkat cells with pDual-HIV and different quantities of pTAR, ranging from 0 to 2 mg, the ratio of pTAR to pDual-HIV being equal or inferior to 1:1. The frameshift efficiency also increases under the conditions corresponding to low amounts of TAR, the highest increase being $140% of the frameshift efficiency without TAR (data not shown). We investigated the involvement of PKR in the changes in frameshift efficiency observed with a low amount of TAR. To this end, PKR expression was silenced by transfecting a PKR siRNA mix into stable 293T transfectants expressing a dual-luciferase HIV reporter. After 24 h, cells were transfected with pTAR in different amounts and harvested 48 h later. As a negative control, an eGFP siRNA mix targeting GFP was used. In the presence of the eGFP siRNA, the frameshift efficiency increases when TAR is present. However, when PKR expression is silenced, this effect disappears, supporting that it is related to PKR activation ( Figure 5A ). Effective silencing of PKR is achieved under the conditions of the assay as shown in Figure 5B . It can be noted that the response of the cells to the increase in the amount of TAR appears to differ from that in Figure 4 . This is due to a difference in the experimental protocol resulting in a lower ratio of the quantity of transfected pTAR to the number of cells (see 'Materials and Methods' section). Using a dual-luciferase reporter system in Jurkat and 293T cells, we showed that the presence of TAR alters HIV-1 frameshift efficiency. The addition of a high amount of TAR, in cis or in trans of the reporter mRNA, decreases the frameshift efficiency. This effect is related to an inhibition of PKR. Conversely, a low amount of TAR increases the frameshift efficiency, by activating PKR. Activation or inhibition of PKR is well-known to affect translation initiation via changes in eIF2 phosphorylation (reviewed in 21). However, it is also known that transformed cells, such as those we used in this study, tolerate a certain degree of endoplasmic reticulum stress leading to a certain level of phosphorylation of eIF2 via PERK, a kinase functionally homologous to PKR (34) . Our experimental conditions do not drastically affect the expression of our reporters, implying that the changes in the translation initiation rate caused by activation or inhibition of PKR are small and that the changes in eIF2 phosphorylation should be modest. Using western blotting, we could not detect significant variations in the phosphorylation level of eIF2 in 293T or Jurkat cells transfected with different quantities of TAR (data not shown). We nevertheless suggest that the effect of PKR on HIV-1 frameshift efficiency results from changes in eIF2 phosphorylation that are too small to be detected in presence of the endogenous signal for phosphorylated eIF2 in these cells. However, we cannot exclude that PKR could also influence HIV-1 frameshift efficiency via another yet undiscovered mechanism. Contradictory effects were seen in previous observations on the influence of the translation initiation rate on the frameshift efficiency. The frameshift efficiency of a plant virus, the beet western yellow virus (BWYV), was higher in a reticulocyte lysate than in a wheat germ extract, which has a lower rate of translation initiation (35) . Also, the frameshift efficiency of the human T-cell leukemia virus type II (HTLV-2), when measured in a reticulocyte lysate, was higher with capped than with uncapped mRNAs, which have a lower rate of translation initiation (36) . These observations disagree with our results that show a negative relationship between the rate of translation initiation and the frameshift efficiency. However, Paul et al. (37) , when comparing the frameshift efficiency of the barley yellow dwarf virus (BYDV) with capped and uncapped mRNAs in a yeast extract, found that increasing the translation initiation rate decreased the frameshift efficiency. Furthermore, Lopinski et al. (38) , who investigated in vivo the effect of a reduced translation initiation rate on the frameshift efficiency of the L-A virus of S. cerevisiae, found that this efficiency was increased under these conditions. The results of Paul et al. (37) and Lopinski et al.(38) are in perfect agreement with our findings, and, in line with them, we present the following model that explains our results ( Figure 6 ). When a ribosome translates the HIV-1 frameshift region, it encounters the frameshift stimulatory signal and makes a pause, its decoding center covering the slippery sequence (39) (40) (41) . During the pause, the ribosome can shift or not the reading frame, and, after the pause, the ribosome unfolds the frameshift stimulatory signal and translation continues. If the upstream ribosome reaches the frameshift region before the signal has refolded, the probability that the frameshift occurs is extremely weak. The spacing between ribosomes translating the HIV-1 frameshift region, which is determined by the rate of translation initiation [basal rate estimated to about one initiation event every 6.5 s (42)], could thus affect the frameshift efficiency. Therefore, if we assume an average elongation speed of five amino acids per second per ribosome, corresponding to a displacement of 15 nt per second on the mRNA (42) , the minimal distance between the decoding centers of two ribosomes translating a mRNA would be of about 100 nt. A ribosome covers about 32 nt on the mRNA and heel-printing studies showed that the first base of the P-site codon is at a distance of 12 nt from the 5 0 edge of the ribosome and of 20 nt from the 3 0 edge (43) . From these calculations, there would be about 70 exposed nt between two elongating ribosomes. Thus, the HIV-1 frameshift region, including the 43-nt frameshift stimulatory signal, would be exposed after the passage of the first ribosome. The signal would then re-form, which takes only a few microseconds (44), before the upstream ribosome reaches the region of the mRNA containing the sequence of this signal. However, the pause made by the first ribosome when encountering the signal decreases the distance with the following ribosome, which has continued to progress during the pause of the first ribosome. This second ribosome could reach the region corresponding to the stimulatory signal before this signal could refold, being still partially covered by the first ribosome. A pause of about three seconds for the first ribosome is sufficient to prevent the refolding of the stimulatory signal. The second ribosome would thus avoid frameshifting and the spacing between this ribosome and the third ribosome would not be altered. As a consequence, the third ribosome would encounter the stimulatory signal and pause, and frameshifting would be possible. This analysis shows that the signal affects every other ribosome under basal conditions. According to this model, an increase in the rate of translation initiation would decrease the frameshift efficiency, since ribosomes would be closer to each other and a smaller proportion of ribosomes would encounter the folded frameshift stimulatory signal. Conversely, a decrease in translation initiation would increase the frameshift efficiency since ribosomes would be further apart and it is very likely that each ribosome would encounter the folded signal. Interestingly, Lopinski et al. (38) , when studying the effect of a reduced translation initiation rate with the L-A virus in yeast cells, observed that the frameshift efficiency doubled, independently of the severity of the initiation defect. Their interpretation was that every other ribosome encounters the signal under basal conditions and that, with a reduced initiation rate, every ribosome encounters this signal. Our analysis fully supports this interpretation. Figure 6 . Changes in the rate of translation initiation influence the frameshift efficiency by modifying the spacing between elongating ribosomes. This model shows elongating ribosomes that reach the frameshift region and explains how the rate of translation initiation, which determines the spacing between these ribosomes, affects the frameshift efficiency (see the text). Note that a ribosome must encounter a folded frameshift stimulatory signal to make a frameshift, but this encounter does not ensure that frameshifting will occur. Although HIV-1 does not induce a rapid and dramatic global shutdown of host cell translation following infection, in contrast to other viruses such as poliovirus, cap-dependent translation initiation is decreased due to cellular stress following infection by this virus (25, 45) and this decrease can be related to PKR activation (46) . Our results suggest that a change in cap-dependent translation initiation could affect HIV-1 frameshift efficiency in infected cells. As mentioned in the 'Introduction' section, the virus replication appears to be exquisitely sensitive to changes in frameshift efficiency. Given the detrimental effect of such changes, the virus likely uses various strategies to counteract this effect. One strategy is inhibition of PKR (reviewed in 25, 47) to stimulate translation initiation. HIV-1 uses two major ways to inhibit PKR: its Tat protein inhibits PKR and its TAR RNA structure blocks PKR dimerization when present in large quantities. TAR is located at the 5 0 and 3 0 end of all HIV-1 mRNAs and is also present under a free cytoplasmic form of 58-66 nt (17, 18) . All these forms of TAR can participate in the inhibition of PKR. However, inhibition of cap-dependent translation initiation can occur independent of PKR activation. Indeed, the HIV-1 Vpr protein is capable of inducing G2 arrest in cultured CD4+ T cells (48, 49 and references therein), and, during such arrest, cap-dependent translation initiation is severely impaired (50) . Another possible strategy to circumvent the problem caused by this situation is the use of a cap-independent mechanism by HIV-1 to initiate the translation of its full-length mRNA (25, 45) . The virus would thus continue to express Gag and Gag-Pol and would maintain a frameshift efficiency that is optimal for its replication. An internal ribosomal entry site (IRES) was identified in the 5 0 UTR region of HIV-1 full-length mRNA (51) and another IRES was found in the beginning of the gag coding sequence (52) . IRES have also been found in HIV type 2 (53) and in simian immunodeficiency virus (54) , two viruses related to HIV-1. However, the use of an IRES by HIV-1 in the context of replication-competent viruses remains to be proven (45) . The two strategies that are described above are not mutually exclusive. HIV-1 could first counteract changes in cap-dependent translation initiation by inhibiting PKR, until a larger stress in the cellular environment severely perturbs cap-dependent initiation. The virus would then switch to an IRES-driven mode to translate its full-length mRNA. This scheme is deduced from studies in cultured cells and it will now be important to investigate the frameshift efficiency in the context of a viral infection. A detailed understanding of the mechanisms used by HIV-1 to control its frameshift efficiency will provide valuable information for the design of drugs targeting the frameshift event.

Establishing a nationwide emergency department-based syndromic surveillance system for better public health responses in Taiwan

BACKGROUND: With international concern over emerging infectious diseases (EID) and bioterrorist attacks, public health is being required to have early outbreak detection systems. A disease surveillance team was organized to establish a hospital emergency department-based syndromic surveillance system (ED-SSS) capable of automatically transmitting patient data electronically from the hospitals responsible for emergency care throughout the country to the Centers for Disease Control in Taiwan (Taiwan-CDC) starting March, 2004. This report describes the challenges and steps involved in developing ED-SSS and the timely information it provides to improve in public health decision-making. METHODS: Between June 2003 and March 2004, after comparing various surveillance systems used around the world and consulting with ED physicians, pediatricians and internal medicine physicians involved in infectious disease control, the Syndromic Surveillance Research Team in Taiwan worked with the Real-time Outbreak and Disease Surveillance (RODS) Laboratory at the University of Pittsburgh to create Taiwan's ED-SSS. The system was evaluated by analyzing daily electronic ED data received in real-time from the 189 hospitals participating in this system between April 1, 2004 and March 31, 2005. RESULTS: Taiwan's ED-SSS identified winter and summer spikes in two syndrome groups: influenza-like illnesses and respiratory syndrome illnesses, while total numbers of ED visits were significantly higher on weekends, national holidays and the days of Chinese lunar new year than weekdays (p < 0.001). It also identified increases in the upper, lower, and total gastrointestinal (GI) syndrome groups starting in November 2004 and two clear spikes in enterovirus-like infections coinciding with the two school semesters. Using ED-SSS for surveillance of influenza-like illnesses and enteroviruses-related infections has improved Taiwan's pandemic flu preparedness and disease control capabilities. CONCLUSION: Taiwan's ED-SSS represents the first nationwide real-time syndromic surveillance system ever established in Asia. The experiences reported herein can encourage other countries to develop their own surveillance systems. The system can be adapted to other cultural and language environments for better global surveillance of infectious diseases and international collaboration.

groups starting in November 2004 and two clear spikes in enterovirus-like infections coinciding with the two school semesters. Using ED-SSS for surveillance of influenza-like illnesses and enteroviruses-related infections has improved Taiwan's pandemic flu preparedness and disease control capabilities. Taiwan's ED-SSS represents the first nationwide real-time syndromic surveillance system ever established in Asia. The experiences reported herein can encourage other countries to develop their own surveillance systems. The system can be adapted to other cultural and language environments for better global surveillance of infectious diseases and international collaboration. With the recent global concern over emerging infectious diseases (EID) and the challenges of the 2003 SARS epidemics, government health officials in SARS-affected countries have begun to consider various measures of improving their infectious disease surveillance systems [1] [2] [3] [4] . Infectious disease epidemiologists and several leading public health administrators at the Centers for Disease Control in Taiwan (Taiwan-CDC) becoming aware of the importance of early detection of EID or bioterrorism, started developing an automatic alert system. Therefore, the Automatic Syndromic Surveillance Planning Task Force Committee was created and recruited infection physicians, epidemiologists, biostatisticians, and information technology (IT) experts in July 2003 to oversee the initiation and development of Taiwan's first medical informatics-based emergency department syndromic surveillance system (ED-SSS). To prepare for this project, we reviewed the syndromic surveillance systems of other countries and officials of health informatics at Taiwan-CDC started collaborating with the Real-time Outbreak and Disease Surveillance (RODS) Laboratory at the University of Pittsburgh to develop a real-time syndromic surveillance system for Taiwan in August 2003 [1, [4] [5] [6] [7] [8] . RODS, used during the 2002 Olympic Winter Games, is the first commonly used syndromic surveillance system in the United States and has been found to efficiently process and analyze data in a timely manner [9] [10] [11] . Together, the task force and the RODS group aimed to establish a nationwide syndromic surveillance system within six months to meet the challenges of potential avian flu outbreaks for up-coming winter seasons and other future EIDs. To gain more operational level experiences, we also visited the Department of Health in New York City, where syndromic surveillance system was established and has been in daily operation since 2001 [12] . There, the task force members observed routine workflow processes and became familiar with other practical concerns of operating an ED-SSS on a daily basis. Based on these experiences and high population density in Taiwan, we decided to create a nationwide surveillance system. To this nationwide ED-SSS, we added geographical information system (GIS) technology, meant to facilitate epidemiological investigation and feedback between data providers and decision-makers [13] . Using the electronic data from the health information systems already in place in about eighty percent of the hospitals in Taiwan required by the National Health Insurance Payment Program and the technical support of the RODS Laboratory at the University of Pittsburgh, Taiwan's ED-SSS has been in operation since March, 2004 [14] . It is the first time in Taiwan that information technology and timely data directly from hospitals has been used with systematic approaches to facilitate public health surveillance. This report shares our experience of establishing an ED-SSS in a non-English-speaking country. It covers the process of taking into account the various needs at different levels of hospitals, discusses the stages of developing the system, and highlights the characteristics of ED-SSS data collected during the first year. The experiences reported here may benefit other countries seeking to establish or improve their own surveillance systems for infectious disease. Initially, two Taipei City Municipal Hospitals that kept electronic files of their emergency department patients' medical information were selected as pilot sites. From these two pilot hospitals, we had learned work flow involved in the process of data transfer, format of Chinese chief complaints, practical concern of ED (such as heavy workload etc.) during and after the 2003 outbreak of SARS, and available electronic ED information from nationwide emergency care hospitals. To obtain more representative data from various geographical areas, we gradually enlisted the cooperation of 189 hospitals nationwide, all offering emergency healthcare. Because many outbreaks of EIDs require emergency health care, these emergency care-designated hospitals were required transmit their ED triage and patient data to the Taiwan-CDC electronically on a daily basis ( Figure 1 ). Nurses at the triage stations at all participating hospitals generated the bulk of the information needed for syndromic surveillance. That information included time and date of admission, date of birth, gender, home zip-code, body temperature, triage categories and chief complaints for patients admitted to their EDs. Because the National Health Insurance requires hospitals to keep clinical data written in ICD-9 codes based on the criteria of the international classification of diseases, 9th revision, clinical modification (ICD-9-CM) for billing purposes, we were also able to collect clinical data from initial assessment of each case by an ED physician. These fundamental data were provided in either the health level-seven protocol (HL-7) format or extensible Markup Language (XML) format, if HL-7 was not used by a hospital during the time of the study. Thus, ED-SSS is capable of accepting these data of above-mentioned variables (Table 1) , including the patients' clinical and demographical information, and hospital identification numbers, in either format. All the sentinel hospitals recruited into our system had independent MySQL servers on which their data were saved, plus a remote connecting program for automatic transfer of data. Data files generated by the 189 hospitals, including data from their triage classification systems, hospital information systems (HIS) and clinical information systems (CIS), were firstly de-identified and then transferred hourly to a Microsoft SQL Server 2000 at the Taiwan-CDC. Hyper text transfer protocol over secured sockets layer (HTTPS) or secured file transfer protocol (SFTP) was used in this process. All communication histories were recorded in a log file in the SQL server at the Taiwan-CDC and monitored daily by health informatics personnel. A program was written into the system so that each transfer attempt to the Taiwan-CDC would automatically generate an e-mail to the hospital notifying them whether the transfer had been successful or not. Three different data storage tables were designed to process the data in the Taiwan-CDC's syndromic surveillance database. All information received is initially fed into the first table, with a serial number generated in an additional column for each case. The system picks up the data from the first table every five minutes and moves it into a second temporary table for a logic check and data cleansing. At this point, the system checks for unambiguously erroneous data, e.g., a birth date later than the admission date or other variables such as body temperatures that fall outside of reasonable ranges. The data cleansing work is accomplished through a system algorithm written with SQL commands. The cleaned-up data are transferred to the third table for further epidemiological analysis, aberration detection, and then sent them to related local public health agencies. Although only a few variables were collected from each hospital on consecutive days, one major difficulty we had was the data presented by discontinuous data, i.e. data that sometimes be there and sometimes not. Sometimes data were repeated. To handle this problem, specific criteria of data cleansing were used for different variables, including the logic checks described above and double checks for possible presence of duplicate patient records. If data in the chief complaint field was written as "test" or the field was left empty or if the ICD-9 field was written as "test" or left empty, they were deleted before data analysis. Hospital ID, date of birth, admission time, gender, and home zip-code were used as key indicators of whether a listing is a duplicate listing and be deleted as repeated data. The system was capable of performing frequent and rapid checks of any subject of hospital identification code and time format of all time fields. It was capable of moving erroneous data to an "error table" for storage. Incorrectly formatted ICD-9-CM data were also moved to the error table. All deletion and removal operations were recorded in the log file for monitoring. In certain situa-tions in case possible systematic errors were found (i.e., aberrant number of ED visits on certain days or occasionally inconsistent formats of ICD-9 codes), the data examiner would contact the medical informatics officers of those specific hospitals to discuss improving data entry. After the data cleansing, we categorized ED visits into 11 different syndromic groups important in Taiwan. There were: (1) fever, (2) respiratory, (3) skin, (4) neurological, (5) upper gastrointestinal (GI), (6) lower GI, (7) haemorrhagic, (8) influenza-like illness (ILI), (9) asthma, (10) enterovirus-related infection (EVI) syndrome, and (11) syndrome for severe illness or death. Since only about 25% of all chief complaints were written fully in English and the grouping of syndromes by chief-complaints due to Chinese language barriers would have effects on the outbreak detection ability, we first analyzed our data according to the ICD-9 coded syndrome groups [15] . Definitions of clinical syndromes were based on two different sources: (1) those associated with bioterrorism-related agents as announced by the Centers for Disease Control and Prevention (CDC) in the U.S.; and (2) those identified as important by the ED-SSS Advisory Committee in Taiwan, whose members include infectious disease physicians, emergency doctors, pediatricians, and epidemiologists [16] . For example, because the epidemics of enterovirus 71 caused severe fatal cases in Taiwan in the years of 1998, 2000 and 2001, the EVI syndrome group was considered as an important syndrome group locally (ICD-9 codes listed in Table 2 ) [17] . All patient information was de-identified and only aggregated data was used for data analysis. The protocol for this study was approved by the Research Ethical Committee (Institutional Review Board) of National Taiwan University. 20 Coxsackie carditis, unspecified 074. 21 Coxsackie pericarditis 074. 22 Coxsackie endocarditis 074. 23 Coxsackie myocarditis, Aseptic myocarditis of newborn 074. 3 Hand, foot, and mouth disease Vesicular stomatitis and exanthem 074. 8 Other specified diseases due to Coxsackie virus, Acute lymphonodular pharyngitis Although the ED-SSS data started transferring on March 10, 2004, we confined our analysis to data collected between April 1, 2004 (when the data became more stabilized) and March 31, 2005 . Data were organized using statistical programs to perform a descriptive analysis of the daily and weekly plots of different syndrome cases and obtain a baseline pattern for each syndrome in Taiwan. We initially generated the SQL commands for data querying and data grouping into the 11 different syndromic groups. To increase the sensitivity of this ED-SSS in monitoring regional patterns of these 11 syndrome groups, we categorized ED-SSS data by four different geographical areas (northern, central, southern and eastern Taiwan), based on major regional variations in the types of infec-tious diseases. In analyzing the seasonal patterns of ED visits, the correlation between the ILI syndrome and respiratory or asthma syndrome was assessed by the value of Pearson's coefficient (R). the Taiwan-CDC on a daily basis. The greatest challenge as began to develop this system was communication with different hospitals. Because different hospitals were using different information systems and inputting different data with various formats, it took long time to agree which variables and their data format should be collected. At the very beginning to build the ED-SSS, we had at first intended to collect information on a large set of epidemiologically useful variables, including occupation, travel history, family clustering, other exposure-related factors, and address for each patient to help detect possible zoonosis. However, such data were not collected during routine medical examination and care. Finally, we decided to capture only parameters usually collected by the hospitals during examination, intake and care. During the planning phase, it was necessary to gain a full understanding of what not only just public health personnel expected but also what the medical staff at participating hospitals expected from the ED-SSS. Public health officials tended to prefer a timely and sensitive surveillance system able to detect all possible outbreaks of emerging or known infectious diseases. Mostly concerned over the limited public health resources, they wanted more evidence to prove the cost-effectiveness of ED-SSS and fewer false positive signals from pilot studies before integrating the system into routine public health surveillance workflow. On the other hand, the hospitals and their medical staff had three major expectations. First, the hospitals expected an easily operated feedback mechanism and quick feedback of useful information for better decision-making. Those who had experienced nosocomial infection of SARS during the 2003 SARS outbreak were particularly interested having analyzed information, based on their own hospital or regional/national hospital data, quickly fed back them. They believed that this would provide incentive for them to share hospital data and routinely maintain the high quality of their data for public health usage. Second, the hospitals anticipated two-way communication with public health agencies, as they frequently been requested or even forced to send data on short notice when they were too busy or too involved in emergency care. What made matters worse, despite their compliance; they had difficulty in obtaining useful feedback information from the public health agencies so that they could improve their care of patients at the time of an outbreak crisis. Third, the hospital decision-makers wanted immediate firsthand feedback, particularly with regard to control of nosocomial infection and hospital management in order that their health-care workers could be protected during regional outbreaks. Considering the expectations of both public health agencies and hospitals, we learned that the syndromic surveillance system should provide efficient means of feedback and effective two-way communication. From April 1, 2004 to March 31, 2005 , data transmitted from the 189 hospitals on 2,692,325 ED visits were collected and stored in the Taiwan-CDC database. Initially, we appointed two computer engineers to cleanse the data by checking the log files and inform the hospitals by telephone on weekdays to correct errors or provide missing data. Then, these cleaned data on daily counts of ED visits collected from ED-SSS were analyzed. The time series plot of rough data on daily numbers of ED visits in our nationwide ED-SSS during the study period is shown in Figure 3 . The computer system shut down twice during this period. The first time occurred from August 8 th to August 9 th , when no daily procedure was installed to monitor the quality of uploaded data. To help both hospitals and public health agencies perform routine data quality checks, we installed a computer program having check-up procedures of data quality after each data transfer from the hospital to the Taiwan-CDC for automatic quality control of data. This program records all the logs of each data sending from the participated hospitals. If Taiwan-CDC doesn't receive data from hospitals, program will send e-mail automatically to inform the personnel of health informatics in that hospital about failure sending. System maintenance personnel need to check the log daily and make a phone call to the hospital to verify successful data transfer and quality of data if there are no data transfers in two consecutive days. Only 5.04% of hospitals failed to send ICD-9-CM data, but almost half (47.4%) failed to send chief complaints. For example, of the 239,617 sets of cleaned data received in July 2005, about 7.1% of ICD-9-CM information was lacking or filled out as 'null' and 54.82% of cases did not include chief complaints. Additionally, certain hospitals transmitted the patients' chief complaints in Chinese, further complicating the analysis. Because of these difficulties, this report focuses on the data based on ICD-9-CM diagnostic codes only. The time-series plots of the 11 syndrome groups (Figure 4 , 5) that may correlate to infectious diseases and important health problems in Taiwan (e.g., asthma has become an important pediatric problem in recent years) were analyzed. Understanding the characteristics and patterns of numbers of ED visits over time from our established ED-SSS in Taiwan is very crucial before we set up appropriate threshold levels of different syndrome groups for outbreak detection, There was a significant difference in daily counts between weekdays and weekends, which occurred on a weekly basis. ED visits were 1.288-fold-higher on weekends than on weekdays (p < 0.001), while national During this first year study, the ED-SSS found patients throughout Taiwan were more likely to seek emergency medical services at medical centers than at district or local hospitals. Most patients visiting ED were 60 years old or older (21.46%) or below the age of 10 years old (18.55%). These two groups were also at greatest risk for various infectious diseases, especially influenza (Table 3) . Young adults between 20 to 39 years old ranked the third most frequent visitors to ED (17.43%), though traffic accidents, not infectious diseases, were the main reasons for their visits. These age distributions suggest that the newly established ED-SSS was capable of providing information for the age groups most at risk for severe cases of infectious diseases in Taiwan. Male ED patients slightly outnumbered female ED patients (male : female = 1.12:1), which approximates the general distribution of gender in Taiwan (male/female ratio = 1.10). For elderly ED patients (age > 65 years old), the male/female ratio in our sample was the same as general population (male: female = 1.10:1). To understand the epidemiological characteristics of the eleven important syndrome groups and asthma syndrome in ED-SSS, data of their daily and weekly counts were plotted and shown in Figure 4 and 5, respectively. As can been seen in time series plots in Figure 6 , the seasonal patterns for ED visits due to respiratory syndrome and ILI syndrome were quite similar with high correlation (R = 0.98), while that for asthma syndrome, which had distinct peaks, was nonetheless not highly correlated with ILI syndrome (R = 0.78) nor respiratory syndrome (R = 0.77). Importantly, the ED-SSS was able to detect peaks of these respiratory-related syndromes even occurred in or around the summer season (July to September), though their most noticeable peaks were found during the Chinese new year holidays ( Figure 5B, 5C, 5D ). Like respiratory infection syndrome, visits due to the fever syndrome also showed another peak during the summer (July to September) ( Figure 4A ). ED visits for respiratory syndrome peaked earlier (mid-September) than those visits for ILI (mid-October) and asthma syndrome (end of September, e.g. the transition period between summer and autumn). Visits for pediatric asthma syndrome for children 12 years old and below peaked in mid-autumn, around mid-October (data not shown). Later during winter season (between November and February), there was an increase Respiratory Syndrom Asthma Visits In gastrointestinal (GI) syndromes, visits due to upper GI or lower GI or total GI started increasing November 2004 and peaked during the Chinese New Year holidays ( Figure 4F, 4G, 4H) . Interestingly, cases of hemorrhagic syndrome also increased slightly in the winter season ( Figure 4I ). For those syndrome groups with severe symptoms, including skin rash, neurological symptoms and death/ coma that might be related to bioterrorism attacks, there was no significantly cyclic or seasonal patterns ( Figure 4J , 4K and 4L). One spike of syndrome cases with clinical severity (severe syndrome) appeared in mid-May, but that occurred as a result of one hospital sending duplicate data that escaped from our check algorithm ( Figure 4L ). Taiwan has very high population density (Taiwan, 632.23 persons/km 2 ; Taipei, 9662.53 persons/km 2 ), which increases the spread of many human-to-human infectious diseases [19] . This makes the surveillance of infectious diseases very important for this island. Taiwan's ED-SSS is the first syndromic surveillance system to be implemented in Asia. It also represents the first time that Taiwan's public health agencies have attempted active nationwide surveillance. Its automatic data collection mechanism is capable of capturing comprehensive population-based health information and providing important details on current disease epidemics at the community level. The information it provides can also be used as community baseline data for further infectious disease modeling and can also improve the detection of emerging infectious diseases. In addition to the information that the ED-SSS can provide for disease control, it can open avenues for further investigation. For example, in addition to the neurological syndrome, asthma syndrome, and syndrome for severe symptoms, there were clear and consistent weekend and holiday increases in the visits of other nine syndrome groups. Because cost of ED visits in Taiwan is not as expensive as it is in other countries, especially the United States, it is very likely that many patients seek ED medical care when local clinics are closed on weekends and holidays. It is also possible that the gathering of people on the holidays would increase the transmission of certain pathogens, particularly on cold days in closed spaces where respiratory viruses including influenza virus are easily transmitted. Therefore, future research might want to investigate the effect of holidays on the aberration detection of outbreaks and prediction of number of cases for certain infectious diseases using Taiwan's ED-SSS. We also found differences in seasonal trends in visits due to symptoms/signs related to respiratory, influenza-like illness and asthma syndromes. Our ED-SSS found a sum-mer peak in visits for cases with influenza-like illnesses in 2004. This has seldom been found by the previous passive surveillance systems used in Taiwan. These summer cases of influenza-like illness occurred before annual vaccinations usually done in October or November. Therefore, a further longitudinal analysis of influenza-like syndrome patterns is needed to formulate the best vaccination policy on human influenza. Another epidemiological finding from our ED-SSS was an increasing trend in visits due to gastrointestinal syndrome starting late autumn 2004. Such trends have not been detected by other infectious disease surveillance systems in Taiwan. There are two possible explanations for this finding. One reason for the increases might be related to the increased activity of certain pathogens, including rotavirus or norovirus, during winter season, as was found during the winter of 2006 in both Japan and Taiwan [20] [21] [22] [23] [24] [25] [26] [27] . Another reason might be the social habits of Taiwanese who like to dip raw meats and seafood into boiling water fondues and eat from the chafing pot during the winter season. This would increase the change that inexperienced or careless diners would consume undercook seafood or use chopsticks contaminated by raw seafood. The findings of our ED-SSS, the first time in Taiwan to use daily rather than weekly data, suggest further directions for research into GI syndrome and many other diseases of significant interest to public health. For example, during the 1998, 2000 and 2001 enterovirus 71 epidemics, children aged 3 years and younger who were at higher risk of severe or fatal cases of the disease were identified for more effective prevention only after the occurrence of several cases of sudden deaths from weekly sentinel physician surveillance and later retrospective epidemiological data analysis on those cases when sample size became larger. Therefore, prospective monitoring of daily ongoing data of EVI syndrome in this high risk age group and early ED-SSS detection of enterovirus activities by local public health personnel might help minimize social panic among parents. Furthermore, the results on seasonal pattern of enterovirus-like infection in our ED-SSS was consistent with the previous epidemic patterns in Taiwan, again demonstrating the usefulness of ED-SSS to avoid future large-scale or severe epidemics caused by enteroviruses [18] . In summary, these initial findings suggest that it is necessary to develop algorithms capable of detecting aberrations for different syndrome groups from patients in different geographical areas of Taiwan, taking into account variations in the levels of medical care and the effect of weekends and holidays on ER visit. The ED-SSS did not, however, reveal obvious trends in all syndrome groups. For example, it was hard to find seasonal patterns or secular trends in cases of coma/death, skin rash, or neurological symptoms -the three syndrome groups that might be useful in the detection of severe outbreaks caused by bioterrorism, e.g., anthrax, during the study period without bioterrorism attacks [28, 29] . Certainly, continuous monitoring for these syndrome groups at both local and national levels will be very helpful in detecting possible bioterrorism or EIDs in future years. Using those trends in coma/death and other syndrome groups of clinical severity or unexpected symptoms/signs, our ED-SSS data have provided directions for further research in the areas of pathogen detection, epidemiological clues, and improvement in public health policies. Therefore, future investigations have to control the weekend and holiday effects of ED visits for better aberration detection even during long holidays. In daily public health practice to monitor the data of ED-SSS, careful verification and systematic management is needed once the aberration signals are detected. The server needs an automatic error feedback system function instead of the original use of engineers to double check for data errors would increase the efficiency and completeness of surveillance. Future efforts require closer collaboration between computer-science professionals and medical informatics personnel at the Taiwan-CDC to establish a system with the standard operating procedures (SOP) for database maintenance and to provide more continuous on-job training for both hospital users and local and central public health agencies [30] . The major difficulty in developing our ED-SSS was diverse formats for different types of data, including categories of chief complaints, the ways to fill out ICD-9-CM codes, and even the different number of digits used in home zip codes in different participated hospitals. For example, most Taiwan hospital ED physicians/nurses EDs only write down one chief complaint, which is very different from the ED reports made by most U.S. hospitals which list all possible complaints (with text format) in English. Several participating hospitals only had a paper system for recording triage chief complaint data. A standard format for select syndromes and variables needs to be established and continuously reevaluated to improve data quality and stability of data transmission. There are needs to have more research into the chief complaints with Chinese styles, the suitability of chief complaints vs. ICD-9 codes, how to combine symptoms/signs and link data to improve sensitivity. With regard to the current epidemics of avian influenza H5N1 in China and many other southeast Asian countries, an ED-SSS like the one we developed in Taiwan may play an important role in detecting an outbreak possibly caused by human-to-human transmission even when cluster size is small [31] [32] [33] [34] . Through early detection, ED-SSS may help minimize the adaptation of avian influenza virus to human populations. Because of the large volume of business traffic, international travelers, and workers from Southeast Asia coming to Taiwan, it has previously been difficult to do real-time surveillance for imported infectious diseases, including dengue, malaria, acquired immunodeficiency syndrome (AIDS) and SARS. However, using the ED-SSS to monitor health status at the community level may help public health decision-makers handle unexpected health threats. Because countries are so interconnected today, it is imperative that we share our health information and experiences with other countries if international health is to be guarded. Our ED-SSS has equipped Taiwan the ability to closely monitor avian influenza and other potential EIDs in Asia and worldwide. We hope that by sharing our experiences developing ED-SSS, other countries can be encouraged to develop and improve their own surveillance systems for infectious disease. The author(s) declare that they have no competing interests. TSJW was in charge of epidemiological data analysis, improvement of the ED -Syndromic Surveillance System, and manuscript writing. FYFS initiated the thoughts on Syndromic Surveillance for detecting emerging infectious diseases in 2003 and contributed to syndrome groupings, selection of variables for ED-SSS, and system improvement based on clinical data analysis. MYY contributed to syndrome groupings, initiating the standard format for collecting Chinese Chief-Complaints in our ED-SSS, and system improvement with regard to clinical aspects. JSJW initiated the thoughts on Syndromic Surveillance for Emerging Infectious Disease in 2003. SWL worked on computer programming on ED-SSS and health informatics for surveillance systems of infectious disease at Taiwan-CDC. KCMC was a leader and coordinator on health informatics for surveillance systems of infectious diseases at Taiwan-CDC and gave the most administrative support on systematic improvement in health informatics. CH provided statistical consultation and chose the best statistical modeling method for outbreak detection in the initiation stage of establishing the ED-SSS. JHC was the Deputy Director at Taiwan-CDC in charge of improving surveillance of infectious diseases, coordinated the 189 hospitals designated for emergency health care to participate the ED-SSS, and provided strong administrative support on ED-SSS. YTC was a research assistant of ED-SSS in charge of the data analysis and project administrative assistance. HC was the Director of Department of Health in Taipei City and participated in the task force meetings from planning to implementation of ED-SSS in the perspective of the local government. CHC was the Section Chief of Department of Health in Taipei City in charge of surveillance, prevention and control of infectious diseases in Taipei City and gave suggestions health informatics and practical concerns from the viewpoints of local government. FCRT helped set up the Real-time Outbreak and Disease Surveillance (RODS) system at Taiwan-CDC. MMW introduced our public health officials and scholars in Taiwan to the practical applications of RODS in the USA and informed us of recent progress of RODS. IJS, the former director of Taiwan-CDC, had the vision to invite scholars to discuss the improvement of infectious surveillance system in Taiwan right after the 2003 outbreak of SARS. CCK, involved in the improvement of infectious disease surveillance in Taiwan for more than twelve years, initiated research on syndromic surveillance, coordinated each trouble-shooting step as the ED-SSS was developed and implemented, and was involved the revision of the manuscript. All authors read and approved the final manuscript. Publish with Bio Med Central and every scientist can read your work free of charge

Amino Acid Similarity Accounts for T Cell Cross-Reactivity and for “Holes” in the T Cell Repertoire

BACKGROUND: Cytotoxic T cell (CTL) cross-reactivity is believed to play a pivotal role in generating immune responses but the extent and mechanisms of CTL cross-reactivity remain largely unknown. Several studies suggest that CTL clones can recognize highly diverse peptides, some sharing no obvious sequence identity. The emerging realization in the field is that T cell receptors (TcR) recognize multiple distinct ligands. PRINCIPAL FINDINGS: First, we analyzed peptide scans of the HIV epitope SLFNTVATL (SFL9) and found that TCR specificity is position dependent and that biochemically similar amino acid substitutions do not drastically affect recognition. Inspired by this, we developed a general model of TCR peptide recognition using amino acid similarity matrices and found that such a model was able to predict the cross-reactivity of a diverse set of CTL epitopes. With this model, we were able to demonstrate that seemingly distinct T cell epitopes, i.e., ones with low sequence identity, are in fact more biochemically similar than expected. Additionally, an analysis of HIV immunogenicity data with our model showed that CTLs have the tendency to respond mostly to peptides that do not resemble self-antigens. CONCLUSIONS: T cell cross-reactivity can thus, to an extent greater than earlier appreciated, be explained by amino acid similarity. The results presented in this paper will help resolving some of the long-lasting discussions in the field of T cell cross-reactivity.

Each T cell expresses thousands of T cell receptors (TCR) of a single specificity that allows inspection of peptide fragments bound by major histocompatibility complex molecules (MHC) on the surface of other cells. Peptides originate as the product of intracellular protein turnover, and both foreign and self-peptides are able to form peptide:MHC complexes (pMHC). Presentation of peptides for which the inspecting CTLs have not been tolerized, triggers a cytotoxic response. Although much has been learned about peptide processing and MHC presentation [1, 2] it is still largely unknown why roughly half of all natural foreign pMHC are ignored [3, 4] . The processing and MHC binding of naturally processed foreign peptides is a primary requirement for the initiation of a cellular immune response. However, the availability of a suitable TCR further determines if a peptide is immunogenic. The structural mechanism of T cell recognition is a highly debated subject in the immunological literature and a consensus view of the promiscuous peptide recognition has not yet been reached (see e.g., [5] ). The core problem is that T cells seem to combine high specificity with the ability to recognize a surprisingly large number of dissimilar antigens. Two terms are often used to describe this nature of T cell recognition. Poly-specificity is used to emphasize TCR's ability to recognize multiple distinct/unrelated pMHC ligands with high specificity (with little or no tolerance to substitutions of the ligands) [6, 7] . Cross-reactivity is a term that was originally used to indicate unexpected reactivity to targets that differed from those used to initially define the T cell clone [8] . Several studies suggest that T cells can recognize seemingly dissimilar epitopes (for a summary see [6] ), while other studies have established that substitutions affect peptide recognition in a predictable and additive manner [9] suggesting that the majority of cross-reactive pMHC complexes share structural similarities. One outstanding question in T cell biology is therefore whether T cell cross-reactivity is mostly a stochastic phenomenon induced by unpredictable structural constraints or, whether we can predict which peptides should be cross-reactive. Previous studies of crossreactivity have focused on limited data covering a single or a few T cell clones. Here, we investigate a simple model of T cell crossreactivity and perform a large-scale analysis spanning both a broad set of experimental settings, heterogeneous pathogens, MHC molecules and T cell clones. We use this benchmark to investigate whether cross-reactivity is either generally predictable or mostly random. Finally, we test whether the degree of host mimicry is negatively correlated with immunogenicity. By analyzing a large set of known HLA-A2 restricted HIV epitopes, we investigate if potential HIV epitopes with high similarity to self are able to trigger detectable immune responses. Our results suggest that amino acid similarity, rather than identity, is a predictive measure of cross-reactivity. We analyzed public data on CTL sensitivity and created a visualization of how CTLs react to single amino acid substitutions. Lee et al. [10] analyzed the specificity of CTL responses against the immunodominant HLA-A2 restricted HIV Gag epitope SLFNTVATL (SFL9). IFNc production was measured in response against all 171 single mutant variants of SFL9. Abrogated TCR responses were mostly due to loss of TCR binding as the majority of SFL9 variants retained binding to MHC. The cross reactivity data for the three data sets: G10, T4 and PBMC were converted into a position-specific-scoring-matrix (PSSM) as described in Materials and Methods. The recognition motif of the T4 clone (the PSSM matrix) is visualized in Fig. 1A as a Logo plot [11] . The plot shows a stack of the possible amino acid mutations on each position in SLFNTVATL (x-axis). The height of the stack is reciprocal to the number of tolerated mutations (i.e., it indicates the degree of T cell recognition specificity at this position, see Materials and Methods). Few tolerated mutations translate into tall stacks while many tolerated mutations show up as short stacks (bars). For example, in position five, only one variant is tolerated (T5S) and is shown as a tall bar. The remaining variants on position 5 were unable to bind the TCR even though the binding to MHC was mostly preserved. On the contrary, in position one, 18 out of 19 variants preserved the TCR recognition. The logo plot for CTL clone T4 given in Fig. 1A suggests that the central peptide positions are most important for peptide-TCR binding, which is in agreement with earlier data [12, 13] . The average Shannon information [14] plot for T4, G10 clones and PBMC is shown in Fig. 1C . This figure also indicates that positions 2-6 and 8-9 are most important for peptide recognition whereas position one is consistently of little importance. Position 2 and 9 are the main positions determining peptide binding to the HLA-A*0201 molecule, see Fig. 1B . Thus, positions 3-6 and 8 were consistently involved in the primary TCR recognition motif. Moreover, the sequence motif for T4 clone suggests that tolerated substitutions tend to be conservative with respect to the original epitope sequence, SLFNTVATL. Examples are F3Y (both non-polar and aromatic), T5S (both polar), and V6I (both aliphatic). Similar observations on tolerated substitutions were made for the other two CTL clones (data not shown). Taken together, these data suggest that amino acid similarity could be a major component of T cell recognition. Using the information in the TCR amino acid position specific scoring matrices, we estimate the number of ligands recognized by a given T cell clone by assuming that recognizable peptides contain only those amino acids giving a detectable ELISPOT response in the Lee et al. [10] study. Non-recognized peptides are the ones containing at least one prohibited amino acid for which no response was detected. The number of recognizable peptides was computed by the following procedure. The degeneracy of a TCR on a single position was measured as the diversity of amino acids present at that position defined in terms of the Simpson index (see Materials and Methods). This diversity measure yields a value between 1 and 20. Here, 20 means that all amino acids are used with equal frequency at a position, and 1 means that only a single amino acid is found. The higher the diversity the more degenerate the TCR is at this position. In the binding motif of T4 clone (Fig. 1A ) the first position diversity is very high, 13.26, as [14] which is a measure of how conserved a position is. Rigid positions have few but tall letters, while very degenerate positions have many but very short letters. For example, position 1 was mutated 19 times of which 18 variants preserved TCR binding, only the S1R variant compromised TCR binding while the MHC binding was preserved (see [38] ). The frequency of amino acids occurring in this TCR motif can also be used to estimate the number of distinct ligands this T cell clone can recognize (see text for details). (B) Sequence motif of HLA-A2 binding peptides (277 HLA-A2 restricted peptides were extracted from the SYFPEITHI database [15] ). (C) The average Shannon information at each position, for the CTL clones: G10 and T4, and PBMC. doi:10.1371/journal.pone.0001831.g001 expected, because this position is highly degenerate. In the conserved position five, the Simpson diversity drops to 1.29. The product of the tolerated amino acid diversity at each position can provide an estimate of the number of ligands a T cell clone can recognize. For T4, we estimate a total of 5.6?10 5 ligands in this way and this value is in good agreement with previous estimates [8] . For the G10 clone, we estimate 3.2?10 6 ligands, suggesting that this clone is more degenerate. Similarly, one can estimate the number of ligands that can bind to a MHC molecule. For example, the HLA-A*0201 molecule (see Fig. 1B for the binding motif) can bind 4.8?10 9 distinct peptides [15] . Thus, measured in this way the CTL binding event is three orders of magnitudes more specific than that of the MHC. The above calculation suggests that a single T cell receptor can recognize as many as 10 6 ligands. How related are these ligands, and is the cross-reactivity of a T cell clone predictable? A few studies suggest that cross-reactivity is not completely random [9, 16] , while others argue that T cells can recognize unrelated ligands (see e.g. [17] ). Here, we investigate whether TCR peptide cross-recognition can be predicted by a quantitative model of peptide similarity using amino acid similarity matrices (SM) as explained in detail in Materials and Methods. The peptide similarity score is unity for two identical peptides, and 0 for peptides of maximum dissimilarity, as defined by the SM. Note, that this simple model does not differentiate between positions. Below, the predictions made from this model are tested on several independent data sets, and compared against the performance of random predictors. [10] ). The Pearson correlation coefficients between their relative SFU and our peptide similarity score were: 0.40, 0.39, and 0.35 for G10, T4, and PBMC data, respectively (p,0.0001, Monte Carlo randomization exact estimate). Since PBMC consist of two clones, where one clone is dominant [10] , the prediction performance on this data set is similar to the performance on the single clonal data. These significant correlation coefficients suggest that peptide crossreactivity can, to some degree, be estimated from peptide similarities. Thus the proposed model of peptide similarity is capable of producing significant predictions of the loss of recognition due to single amino acid substitutions. Iversen et al. [18] measured IFNc secretion by T cells specific for SLYNT-VATL (SYL9), when they are stimulated with naturally occurring (i.e, patient derived) variants of SYL9. Data consisted of 21 variants of SYL9. Each variant peptide had between 1 and 3 mutations with respect to SYL9. Fig. 2B presents a scatter plot of the data from Iversen et al. [18] for the T4 clone, where the peptide similarity is plotted on the y-axis against the relative IFNc secretion (x-axis). Using the BLOSUM35 matrix to calculate the peptide similarity score (see Materials and Methods) the Pearson correlation was 0.65. Similar results were obtained using BLOSUM matrices 35-90 (data not shown). For the remaining CTL clones (G10, C-3, C-4, C-22 and C-32) tested by Iversen et al. [18] correlations were 0.49, 0.47, 0.55, 0.60 and 0.57 respectively (all values are significantly different from zero with p,0.02 Monte Carlo randomization exact estimate). This model of peptide similarity (or cross-reactivity) was thus able to explain around 20240% of the IFNc secretion. Still, a number of SYL9 variants, for which we predict rather high peptide similarity to SYL9, hardly induce an IFNc response, e.g., A7S, A7V, T5A mutants given in the upper left corner of Fig. 2 . Part of this discrepancy is due to the fact that our model is not position specific, and thus underestimates the effect of mutations in the central positions, which are crucial for T cell recognition (see Fig. 1A ). When more data becomes available, the peptide similarity model can be extended with a weighting accounting for the relative importance of the peptide positions. We were able to achieve similar performances while testing the model on other peptide scanning data, e.g. La Rosa et al. [19] (HLA-A2 restricted CMV epitope, data not shown). Thus, our . The x-axis shows the relative IFNc secretion measured for 171 single mutants of SLFNTVATL (SFL9). Immunogenicity was grouped in four bins with average ELISPOT responses of 0, 0.15, 0.50 and 0.85 of maximal ELISPOT for SFL9. In both figures the y-axis shows the predicted CTL recognition in terms of BLOSUM35 similarity scores (see Eq. 2). Unfavorable (non-conservative) substitutions (low x) are associated with a low similarity score (low y) whereas conservative substitutions (high x) in general are associated with higher similarity scores (high y). (B) Observed and predicted recognition of patient derived SLYNTVATL (SYL9) variants with 0-3 mutations. The axis shows the relative IFNc and peptide similarity scores. Note, that the IFNc response falls to a half when peptide similarity is around 0.85. doi:10.1371/journal.pone.0001831.g002 model was able to predict cross-reactivity of T cell clones measured in at least two different peptide-scanning library studies. Striking examples of T cell cross-reactivity have been reported for CTL responses to viruses [17, 20] . It was shown that CTLs that were elicited during a primary viral infection might also respond when the same mice are re-infected with unrelated viruses. By mapping the different viral epitopes to which a particular T cell clone can respond, it was demonstrated that these cross-reactive epitopes can share very little sequence identity [17, 20] leading to the conclusion that CTLs are extremely non-specific [8, 17, 20] . Reviewing the literature, we compiled a set of 19 cross-reactive epitopes in Table 1 . These epitopes are restricted to the K b , K d , D b , HLA-A1, HLA-A2, and HLA-B62 MHC alleles. Some of the epitopes share only a few amino acids; one is even different on all positions, while others share the majority of the amino acids. We assumed that the first epitope (x) in a cross-reactive pair (x,y) is the original epitope for which the cross-reactive CTL clone was first raised, and that it was observed to respond to y later (see Table 1 ). To test whether these cross-reactive epitopes that differ markedly in their sequence could nevertheless have structurally similar amino acids on the non-identical positions, we did the following. First we computed the similarity of the cross-reactive epitopes S O . Then we constructed an ensemble of random peptides that have the same identical positions as the cross-reactive epitope pair but otherwise consist of random amino acids (see Materials and Methods for details). We then computed the baseline (or the expected) peptide-similarity as the average random similarity denoted S E . In 16 out of 19 pairs the observed similarity S O exceeded the expected baseline similarity S E (see Table 1 and Fig. 3A, p,0 .02, Fisher's exact test). Fig. 3A shows the observed (S O ) versus the baseline expected similarity (S E ) and the solid line presents the case where S O = S E . This plot demonstrates that crossreactive epitopes are significantly more similar than unrelated peptides with the same level of sequence identity. Thus, in crossreacting T cell ligands non-identical positions are significantly more conservative than random. Fig. 3B shows this more explicitly. The 19 epitope pairs were split in two groups according to the level of sequence identity; less than 50% and larger than or equal to 50% identity. For both groups we compute the percent excess observed similarity of the cross-reactive constituents defined as 100?(S O 2S E )/S E From Fig. 3 , we clearly see that for ''seemingly'' unrelated sequences (identity,50%) the excess observed similarity (y-axis) is on average 25.8% +/2 10.8%, i.e., when sequence identity is low, the observed similarity is much higher than the expected similarity. Conversely, for epitopes sharing more than half the amino acids, excess similarity drops markedly (2.0% +/2 4.3%) probably because cross-reactivity is maintained by the more numerous identical positions. The difference in excess observed similarity between the groups is highly significant (p,0.001, rank test), which suggests that amino acid identity is a poor measure for estimating physicochemical similarity, and thus T cell crossreactivity. In summary, the above results demonstrate that biochemical similarity plays a large role in defining CTL crossreactivity when sequence identity is low. In such cases, crossreactivity is observed for non-identical, but conservative, substitutions preserving structural and/or physiochemical properties satisfying the idiosyncratic binding constrains of the responding TCR. The columns are as follows: 1) MHC restriction, 2) source pathogen and protein for initial infection, 3 Another open question in T cell response is why roughly half of all foreign cell surface-presented antigens fail to raise a T cell response [3, 4, 21] . Tolerance to self-antigens could explain this lack of immunogenicity, in which case the degree of similarity to self-antigens should predict which foreign antigens are likely to be non-immunogenic. We examined this effect of self-tolerance on immunogenecity using our cross-reactivity model. First, a large set of self-antigens was defined, and secondly, a list of non-self (e.g., HIV) antigens was built, labeled as either immunogenic or nonimmunogenic according to experimental evidence (data obtained from the Los Alamos HIV database, see Materials and Methods). The expectation was that T cell clones, with high affinity for HIV peptides similar to self peptide(s), have been tolerized during thymic education via negative selection [22, 23] . Such TCRs should therefore not be present in the functional T cell repertoire thus causing tolerance to molecular mimics of self-peptides. We define a score of cross-reactivity to self as the maximum peptide similarity between the non-self antigen and the set of all selfantigens (see Materials and Methods) and test whether nonimmunogenic peptides have a higher cross-reactivity score to self when compared to immunogenic ones. We downloaded the human proteome from the NCBI website and identified a set of 230,460 potential HLA-A2 self-antigens (see Materials and Methods). Next, we downloaded the HIV proteome from the Los Alamos HIV database and predicted a set of potential HLA-A2 epitopes. 33 of the 91 predicted HIV candidate epitopes were annotated as A2 supertype restricted epitopes in the Los Alamos database of CTL HIV epitopes, while the remaining 54 of the HIV peptides were never identified as epitopes. Another four peptides were found to be immunogenic for other HLA alleles than HLA-A2. Since it would be wrong to tag these epitopes as ''non-immunogenic'', they were excluded from the data set. The 33 confirmed HLA-A2 epitopes were labeled: confirmed HIV epitopes and the remaining 54 possible non-immunogenic peptides were given the label: putative, non-immunogenic HIV peptides. It is possible that future studies reveal that a number of the putative nonimmunogenic HIV peptides do in fact elicit CTL responses in HLA-A2 + patients. Nevertheless, this set of HIV peptides should be enriched in HIV peptides that fail to generate CTL responses. Maximal similarity scores were computed between all 87 HIV peptides (33 immunogens and 54 putative non-immunogens) and the set of 230,460 predicted HLA-A2 self-epitopes. Fig. 4 shows a scatter plot of the 33 HIV immunogens (black diamonds) and 54 putative non-immunogens (open circles). The x-axis shows the predicted antigen presentation score (NetCTL) while the y-axis shows the estimated maximum similarity to the self-antigens S SELF (x,y) (see Materials and Methods). Immunogenic peptides tend to be less self-like, although the difference between immunogens and non-immunogens is not significant (p = 0.2, Mann-Whitney). Drawing a horizontal line at y = 0.85 separates the most self-similar HIV antigens from the rest (the results presented in Fig. 2B suggest that IFNc response would fall to a half when the similarity drops to 85%.) For the antigens that have a self-similarity score above 0.85, most (14/16) are classified as nonimmunogenic HIV antigens i.e. predicted epitopes not confirmed by experimental evidence (p-value,0.05, Fisher's exact test). Note, that the NetCTL score does not correlate with the maximal selfsimilarity score (p-value = 0.42, exact estimate) and the above difference between the immunogenic and non-immunogenic antigens is therefore not explained by the difference in the NetCTL scores. Repeating our analysis for HLA-A3 and HLA-B7, we found similar tendency of more-self-likeness among nonimmunogenic HIV-1 peptides (p,0.3 and p,0.45 respectively). Summarizing, these results suggest that similarity to self-antigens plays a role in discriminating immunodominant from cryptic peptides. Many studies have suggested that T cells can recognize totally unrelated peptides and a new term, poly-specificity was coined to express the high specificity of T cell receptors to unrelated peptides [6] . The ''unrelatedness'' of the peptides was defined as low sequence identity, however, sequence identity might not be able to account for the total amount of structural similarity that drives TCR recognition. Here, we demonstrate that this is indeed the case: T cell receptors recognize biochemically and structurally related peptides and cross-reactivity is, up to a degree, predictable. The loss of recognition simply depends on the number and similarity of non-identical amino acid between cross-reactive constituents. We find that the majority of the seemingly ''unrelated'' cross-reactive peptides have a significantly higher biochemical similarity to each other than what would be expected from truly ''unrelated'' peptides. This is especially true for peptides with very limited identity. To our knowledge this is the first study that analyzes a large set of cross-reactive peptide-MHC combinations and demonstrates that the cross-reactivity can, up to a certain extent, be predicted. Because negative selection of immature thymocytes remove high affinity TCR specific for self-antigens [7] , we expected that this should leave a ''hole'' in the T cell repertoire around negative selecting self-antigens. Hence, if an infected cell presents a nonself antigen that is highly similar to a negative selecting self-antigen, then this foreign antigen might not be matched by any available TCR which could provide an explanation for why around half of foreign pMHC do not generate a T cell response [3, 4, 21] . We tested this hypothesis for HLA-A2 restricted HIV-1 response and showed that the absence of T cell response to part of the non-confirmed (i.e. putative non-immunogenic) HLA-A2 restricted HIV-1 peptides can be explained by their similarity to self antigens. These results are in agreement with a recent study by Rolland et al. [24] , who showed a trend of more-selflikeness (measured in terms of number of shared amino acids) among HIV peptides with no detectable CTL responses in a large study group. We predict that the correlation found by Rolland et al. would be stronger if amino acid similarity is taken into account. If peptide similarity can be used to describe T cell reactivity, what would then be the best model to describe similar peptides? We have chosen for the simplest model (given by Eq. (2) in Materials and Methods) because systemic data on cross-reactive peptides is very limited. An obvious extension of this model would be to add position dependence, i.e., account for the fact that central positions play a larger role. The number of potential antigens exceeds the number of T cells in the immune system and the ability to recognize multiple ligands is required to mount at least a few responses to all potential pathogens [8] . Here, we demonstrate that the number of expected T cell ligands is not necessarily reduced by restricting T cell recognition to cover only similar peptides: it is still possible for a T cell to recognize 10 5 210 6 peptides. These estimates of the complexity of CTL recognition are well within the bounds of earlier estimates [8] . In summary, the results presented here quantify, to our knowledge for the first time, that the basis of T cell recognition is amino acid similarity, defined in terms of biochemical properties of amino acid side chains. Lee et al. [10] analyzed the specificity of CTL responses against the immunodominant HLA-A2 restricted HIV Gag epitope SLFNTVATL (SFL9). IFNc production was measured in response against all 171 single mutant variants of SFL9 for two T cell clones (G10 and T4), and for purified Peripheral Blood Mononuclear Cells (PMBCs) using an ELISPOT assay. Purified PBMCs consisted of just two clones where one was dominant. CTL responses were reported as the percentage of maximal IFNc (I) obtained for the reference ELISPOT of SFL9 and discretized in Peptide similarity score The un-normalized peptide similarity score A(x,y) between reference epitope: x = {x 1 , L, x N } and peptide: y = {y 1 , L, y N } of the same length N is defined as the sum of substitution scores along the sequences expressed by the relation: W(xi, yi) is the amino acid substitution matrix, e.g., BLOSUM35, providing a measure of how conservative substitutions are. The peptide-similarity score for the reference peptide x spans the interval: where the length of the interval |I| = A max 2A min depends on references peptide x and matrix W. Two different intervals (I x , I y?x ) are not comparable per se. Thus, we define the normalized peptide similarity using the relation This equation constitutes the model of peptide similarity used throughout subsequent analysis. S(x,y) measures how much peptide y resembles x in terms of the number and magnitude of conservative substitutions. A x max is the auto-peptide-similarity score of x. Thus, A x max = A(x,x). If peptide y is a mimic of x then S(x,y) should be close to 1. The other extreme value (A x min ) is found by comparing x with a peptide x, where on each position the amino acid in x corresponds to the substitution in x with the smallest value i.e. the least likely substitution. In this way 0#S(x,y)#1 for all peptides y. The peptide-similarity score is asymmetric i.e. S(a,b)?S(b,a) despite W being symmetric. The reason is that the extreme values (A min ,A max ) cannot be guarantied to be identical for any pair of peptides (a,b). Reference peptides x which are dominated by amino acids like tryptophan that hardly ever substitute, will have few highly similar peptides (y) which satisfy the condition: S(x,y) < 1. In contrast, reference peptides which are enriched in amino acids that are more likely to substitute (given the matrix W) have a greater number of highly similar peptides. This property is captured by the asymmetry of S. The observed similarity S O of pairs of experimentally verified cross-reactive epitopes (x,y) is to be compared to unrelated peptides (z i ) which retain the sequence identity of (x,y) but have an otherwise random amino acid on non-identical positions. The procedure to compute the ''unrelated'' or ''baseline'' expected similarity is best illustrated by an example: The HLA-A2 epitopes x = GLCTLVAML and y = GILGFVFTL from EBV and influenza-A share 3 identical positions: G1, V6 and L9. We first compute the observed similarity S O = S(x,y) between epitopes x and y using Eq. (2). Then we generate a set of N = 10.000 random peptides, z 1 , z 2 , L, z N , with the same identical positions, i.e. we have z i = G......V...L where a dot can be any amino acids avoiding identity with x at that position. The expected similarity between the primary (original) epitope (x) and the unrelated but semi-identical artificial peptides z is then defined as the average similarity to the ensemble of unrelated peptides as: S E~1 N P N i~1 S x,z i ð Þ. The HIV-1 HXB2 sequence for Env, Pol, Vpu, Rev, Tat, Vif, Vpr, p17, p24 and p2p7p1p6 and the HIV-1 clade B consensus sequence for Nef (due to a stop codon in HXB2-Nef) were downloaded from the Los Alamos database (www.hiv.lanl.gov). There was also one stop codon in HXB2 sequence for TAT, however, no TAT peptides were predicted to be HLA-A2 epitopes and thus the stop codon did not interfere with our results. Out of 3,063 HIV nonamers, 91 were predicted to be HLA-A2 epitopes using NetCTL version 1.2 [25, 26] and default selection threshold (0.75). NetCTL predicts the level of antigen presentation by combining three separate predictions of: proteasomal cleavage, TAP affinity and MHC biding. Four out of the 91 predicted HIV epitopes were found to be immunogenic for other supertypes than HLA-A2 and were filtered out. These were: QLQARILAV, RILAVERYL (class II, DPW4.2), TLYCVHQRI (HLA-A11) and SINNETPGI (HLA-A25). Of the remaining 91-4 = 87 peptides, 33 were confirmed HLA-A2 epitopes by cross-referencing the records of the LANL CTL epitope summary table (downloaded December 2006). Thus, the epitope prediction resulted in the identification of 87 possible HLA-A2 restricted HIV epitopes where 33 (38%) were confirmed and 54 (62%) were not. The human proteome was downloaded from the NCBI website (www.ncbi.nlm.nih.gov/Genomes/date: 29 march 2006) and contained 34.460 protein sequences. The removal of proteins containing the words: predicted, hypothetic or isoform in the protein description label lead to a final core human proteome of 14,034 human protein sequences. We predicted A2 self-antigens using NetCTL version 1.2 [25, 26] for all these protein sequences (default epitope selection threshold). Repeats were removed, along with a small set of self-peptides, which contained the unknown amino acid (X). The final set consisted of 230,460 predicted human HLA-A2 restricted self-antigens each of length 9. HIV self-similarity (HLA-A2) The maximal similarity between predicted HIV antigens (x) and the set of human 230,460 self-antigens (y) was defined as the selfsimilarity score S self (x) = max(S(x,y)) for HIV peptide x. Self-similarity scores were obtained for all confirmed HIV epitopes and putative HIV antigens. Because no identical matches were found between HIV epitopes and self-antigens, self-similarity scores were always S(x,y),1. Confirmed HIV epitopes and putative non-immunogenic HIV peptides were ranked on maximum self-similarity, and the combined ranking was split in two parts: a) The peptides with a selfsimilarity score greater than 0.85 and b) and peptide with a selfsimilarity score below 0.85. We used Fisher's exact test to compute the significance of the difference in the frequency of putative HIV epitopes in the top versus the bottom.

Distinguishing Molecular Features and Clinical Characteristics of a Putative New Rhinovirus Species, Human Rhinovirus C (HRV C)

BACKGROUND: Human rhinoviruses (HRVs) are the most frequently detected pathogens in acute respiratory tract infections (ARTIs) and yet little is known about the prevalence, recurrence, structure and clinical impact of individual members. During 2007, the complete coding sequences of six previously unknown and highly divergent HRV strains were reported. To catalogue the molecular and clinical features distinguishing the divergent HRV strains, we undertook, for the first time, in silico analyses of all available polyprotein sequences and performed retrospective reviews of the medical records of cases in which variants of the prototype strain, HRV-QPM, had been detected. METHODOLOGY/PRINCIPLE FINDINGS: Genomic analyses revealed that the six divergent strains, residing within a clade we previously called HRV A2, had the shortest polyprotein of all picornaviruses investigated. Structure-based amino acid alignments identified conserved motifs shared among members of the genus Rhinovirus as well as substantive deletions and insertions unique to the divergent strains. Deletions mostly affected regions encoding proteins traditionally involved in antigenicity and serving as HRV and HEV receptor footprints. Because the HRV A2 strains cannot yet be cultured, we created homology models of predicted HRV-QPM structural proteins. In silico comparisons confirmed that HRV-QPM was most closely related to the major group HRVs. HRV-QPM was most frequently detected in infants with expiratory wheezing or persistent cough who had been admitted to hospital and required supplemental oxygen. It was the only virus detected in 65% of positive individuals. These observations contributed to an objective clinical impact ranging from mild to severe. CONCLUSIONS: The divergent strains did not meet classification requirements for any existing species of the genus Rhinovirus or Enterovirus. HRV A2 strains should be partitioned into at least one new species, putatively called Human rhinovirus C, populated by members detected with high frequency, from individuals with respiratory symptoms requiring hospital admission.

Human rhinoviruses (HRVs) are the most frequently detected pathogens in acute respiratory tract infections (ARTIs). HRVs have been associated with lower respiratory tract (LRT) illness and more serious clinical outcomes within pediatric and other vulnerable populations [1] . Despite this, HRV strains continue to be commonly defined, en masse, by their most prolific and currently most well-defined role in causing the 'common cold' [2, 3] . Classified within the family Picornaviridae, the genus Rhinovirus consists of 100 serotyped strains divided into two species, Human rhinovirus A (HRV A; n = 75) and Human rhinovirus B (HRV B; n = 25) [4] . Species classification was initially based on the susceptibility of strains to capsid binding antivirals [5, 6] and subsequently confirmed by phylogenetic studies [4, 7] . HRVs are also subdivided by receptor usage; major group HRVs use an intercellular adhesion molecule (ICAM-1; n = 88 strains), which interacts within a depression of the viral surface known as the canyon [8] . Minor group HRVs (n = 12) receive a molecule from the low density lipoprotein receptor family (LDLR) at protrusions along the 5-fold axis of the capsid surface [9] [10] [11] [12] . Recently, heparan sulphate was identified as a pH dependent, low efficiency receptor for HRV-54 [13] . A relationship between receptor usage and species identity is only evident among minor group HRVs; all are HRV As. Major group HRVs reside in both species. Studies have identified four neutralizing immunogenic sites on the surface of a major group HRV (HRV-14; NImIA, NImIB, NImII and NImIII) [14] and three sites on a minor group HRV (HRV-2; site A, B and C) [15] , with some overlap [16] . These sites consist of discontinuous amino acid sequence within VP1, VP2 and VP3. The human enteroviruses (HEVs) are the closest genetic relatives of the HRVs but they belong to the genus Enterovirus. HEV strains are classified into four species (A-D) and some have been implicated in ARTI [17] . Similar to the major group HRVs, some HEVs have been shown to use ICAM-1 as their primary cellular receptor [18] . Other host molecules employed by HEVs include the decay accelerating factor (DAF) [19, 20] , poliovirus receptor (PVR) [21] and coxsackie-adenovirus receptor (CAR) [22] . Under a proposal currently before the International Committee on Taxonomy of Viruses (ICTV), HRV and HEV strains will be combined into the genus Enterovirus, at the expense of Rhinovirus, an issue of contention for decades [23] . Picornavirus genomes typically consist of a 7.2 to 7.5 kb positive-sense, single-stranded RNA molecule located within a non-enveloped icosahedral capsid. The genomes contain a single open reading frame flanked by 59and 39untranslated regions (UTR). The polyprotein is first cleaved into three precursor polyproteins, P1, P2 and P3 ( Figure 1 ). P1 is further cleaved into the structural proteins VP4, VP2, VP3 and VP1. Union of a single copy of each structural protein constitutes a viral protomer; the HRV capsid consists of 60 protomers arranged in a T = 1, pseudo T = 3 (i.e. T = p3) conformation. P2 and P3 are cleaved into seven non-structural proteins including two proteases (2A pro and 3C pro ) and an RNA-dependant RNA polymerase (3D pol ) [24] . Recent studies have generated 27 additional HRV A and 12 additional HRV B genomes [25, 26] also resulting in the proposal that individual HRV genomes are under purifying selective pressure and that significant intra-strain variation occurs only at the antigenic sites [26] . This is in stark contrast to the HEVs for which recombination in the non-structural regions is a significant contributor to viral diversification [27, 28] . Untypeable rhinoviruses have been described using PCR-based tools, since 1994 [29] but it was only in 2007 that the complete coding sequences of a number of novel and apparently divergent HRV strains were reported, permitting a for identification of the first new strains in two decades [30] . The initial strain described, HRV-QPM, was identified from children with presenting symptoms often suggesting bronchiolitis [31] . Subsequently, 'HRV-Xs' were found in an adult asthma study in the United States [32] and 'HRV-Cs' were described from a study of pediatric infection by human bocavirus in Hong Kong [33] . Some notable features associated with these six putative viruses include the reliance upon molecular methods for their detection, their apparent endeminicity, their failure to produce cytopathic effects in culture, their frequent detection in subjects with expiratory wheezing and their limited sequence identity with existing HRV strains but high shared sequence identity. Studies employing subgenomic sequences proposed that there were many of these putative viruses occupying a distinct position within the genus Rhinovirus, which we collectively called HRV A2 [34] . It has been proposed that these strains may constitute a new picornavirus species [33] . Data have not yet been presented which appropriately address classification using current ICTV criteria and the scope of the clinical impact of these divergent strains has not yet been studied in detail. Because it appears that the members of this clade are frequently associated with LRT illness we sought to better characterise them in the hope of finding features that might prove useful in predicting the outcome of HRV A2 infection. Intensive genomic analysis and computer-based modelling resulted in comprehensive characterisation of all HRV A2 polyprotein sequences and, by superimposing the sites of known receptor contacts, we were able to propose a more robust taxonomic placement for these newly identified viruses (NIVs). Detailed medical chart reviews on HRV-QPM-positive (prototype HRV A2) individuals were augmented by a severity scoring system which quantified the outcome of infection by this apparently uncharacteristically pathogenic clade of HRV strains. We undertook a comparative analysis of all HRV A2 complete coding regions (n = 6; HRV-QPM, -X1, H2, -C024, -C025 and C026) against all HRV A (n = 34), HRV B (n = 13), most HEV sequences (n = 71, including species A-D) and two non-human enteroviruses (simian, SEV and porcine, PEV). HRV A2 sequences had a higher G+C content (average, 42.4%) than all other HRVs (average 38.7%) or HEV Ds and a shorter coding region (average, 2144 aa) than any other picornavirus investigated. The final residue of the HRV A2 genomes was an isoleucine, distinguishing them from all other picornavirus genomes we investigated, which ended at the preceding phenylalanine; the biological implications of this feature are unknown. We identified 10 protease sites (data not shown) dividing the HRV A2 polyproteins into four structural and seven nonstructural proteins ( Figure 1A ). However, unlike previous studies that relied on amino acid alignments to determine cleavage sites [33] , we obtained additional evidence that these sites were recognised by 2A pro and 3C pro by using a neural network protease cleavage-site prediction tool. A single cleavage by 2A pro was predicted between an A or L and a G residue at the VP1/2A junction. The remaining cleavages, predominantly occurring between Q and G residues, were mediated by 3C pro . Three motifs crucial for RNA polymerase binding (YGDD, TFLKR and SIRWT) were identified in all HRV A2 sequences [35, 36] . Using SimPlotß [37] , the HRV A2 sequences were found to be most similar to those from HRV A strains (54% identity) whereas identity with HEV species ranged from 46%-49% ( Figure 1B ). Dips in identity occurred across all regions but only the structural region was investigated further because its components contain antigenic sites and receptor contact residues, or 'footprints'; the latter being significant domains for defining the major and minor group HRVs [38] and other picornaviruses. Although identifiable using amino acid alignments [33] , the scope and impact of the many HRV A2 sequence deletions could not be fully qualified or quantified in this form. However, the inability to isolate the HRV A2 strains in vitro will hinder crystallography studies seeking to identify structural features, so we created structure-based alignments of the VP1-4 sequences from the prototype HRV A2 strain, HRV-QPM and used them to predict the presence of a-helices and b-sheets. These were identified in the eight, anti-parallel b-sheet 'jellyroll' conformation common among picornavirus proteins VP1, VP2 (including the EF 'puff') and VP3 [39] (Figure 2 ). Deletions in the VP1 protein imparted the most dramatic changes, reducing the protrusion of the BC, DE and HI loops compared with the other HRVs and introducing an additional C-terminal sheet-loop-sheet structure (arrow, Figure 2B ). Similar deletions were noted in all HRV A2 strains but only HRV-C026 was predicted to encode a similar, Cterminal structure in VP1. We next assembled the HRV-QPM structural proteins into a protomer by associating them with the closest sequence match (HRV-16) and then replicated this structure in silico into a viral pentamer. RMSD values were derived from the VP1, VP2 and VP3 structures of all HRVs with crystallography-derived data available (HRV-1A, -2, -3, -14 and -16) and compared to the predicted HRV-QPM structure. HRV-QPM and HRV-16 shared the highest conformational similarity (0.145, 0.142 and 0.147 Å for the VP1, VP2 and VP3 structures respectively), reflecting the use of HRV-16 in predicting the structure HRV-QPM. The poorest agreements were between HRV-1A and HRV-3 in VP1 (0.793 Å ) and VP3 (0.740 Å ) and between HRV-2 and HRV-14 in VP2 (0.697 Å ). At the ICAM-1 footprint, HRV-QPM values ranged from 0.064-0.106 Å and 0.709-0.779 Å in comparison to the same regions on HRV-16 and -14, respectively. Values could not be determined for HRV-2 because of the impact of deletions. We also examined the conformational agreement between an HEV footprint e.g. the CAR site employed by CV-B3, and the predicted HRV-QPM structure. A space filling model of the pentamer was next used to display SimPlot data in order to visualise regions of sequence conservation and diversification with a focus on the major (ICAM-1; Figure 3A ) and minor (VLDL-R; Figure 3C ) group domains. Overall, greater sequence identity was apparent between HRV-2 than HRV-14 reflecting the genomic similarity to HRV A strains, however within the domains defining HRV groups, HRV-QPM was most similar in its antigenic sites to the major group representative, HRV-14. Most sequence diversity was found in the VP2 protein while the most conserved protein in these comparisons was VP3. To study these regions in greater detail, we mapped known HRV antigenic sites and receptor footprints onto ribbon depictions of pentamers derived from empirically determined structural data (major group, HRV-14, Figure 3B ; minor group, HRV-2, Figure 3D ) and compared their locations and structures to the predicted HRV-QPM pentamer. HRV-QPM and HRV-14 appeared to share a number of similar structures in the region of the ICAM-1 footprint [6, 33, 40] ( Figure 3B ). The noteworthy structural disparities were two small helices present in the HRV-QPM sequences; one in the VP1 and the other in the VP3 ICAM-1 footprint ( Figure 3B ; VP1 and VP3, ICAM-1 arrows). Similar sequences were infrequent among the other HRV strains we examined. The HRV-QPM NImIA and IB sites appeared to differ structurally due to the shortening of the BC and DE loops in VP1 while the NImII and NImIII sites appeared structurally similar. Comparison to HRV-2 ( Figure 3D ) also revealed the impact of VP1 deletions. The resulting loops were shorter, which ablated any analogue of antigenic site A and may hamper binding of the VLDL-R molecule, the only LDLR family member with a welldefined footprint. The area in VP1 forming site B was altered due to the presence of more hydrophilic amino acids at the equivalent positions in HRV-QPM ( Figure 3D ; VP1 site B, arrow). Deletions were apparent but did not confer obvious structural changes to the VP2 portion of site B. Furthermore, in antigenic site C (VP2), a single deletion and a hydrophobic to hydrophilic amino acid change contributed to a protrusion in HRV-QPM compared to the same region in HRV-2. Given the comparative differences we observed when locating and comparing the HRV domains, we also investigated the HRV- The deletions in the VP1 BC loop had a less obvious impact on HEV receptor footprints which were mostly located on the bsheets rather than the loop (Figure 4 ). The EF loop in HRV-QPM VP1 provided the greatest observable identity with any HEV footprint while the C-terminal sheet-loop-sheet structure may interfere with binding to the HEV receptors we investigated. All the HEV VP2 receptor footprints occured in the EF 'puff' but in this region of HRV-QPM a number of deletions and residue substitutions were apparent ( Figure 4 ) that affected the appearance of predicted structural overlap. HEV footprints in VP3 differed from HRV-QPM at the N-terminus and the CD loop due to amino acid substitutions and loop-shortening deletions but they overlapped in the GH loop. HRV-QPM protomers were replicated and mapped onto an icosahedral lattice (T = 1 configuration; Figure 5A ), rendered in 3D and depth cued in order to predict capsid structures. Since HRV-QPM protein and protomer structures were inferred by homology to their closest available relatives, these capsids (HRV-16 and HRV-14) were produced from the available empirical data, using the same approach. Visual comparisons revealed that the deletions in the HRV-QPM VP1 most obviously reduced the size of protrusions around the 5-fold axis of the capsid. Previously, phylogenies had been estimated using subgenomic sequences from HRV A2-like strains [31] [32] [33] . To test whether this approach accurately represented inter-strain relationships, we [10] . Attachment of the VLDL-R involves adjacent VP1 molecules. Magnified VP1 area represents one half of a VLDL-R footprint [42] . Amino acid substitutions (arrowed) contributed to the differences between minor group sites B and C. doi:10.1371/journal.pone.0001847.g003 compared the entire polyprotein sequences from 120 previously described and, for the first time, all six newly-identified picornavirus strains ( Figure 6 ). We confirmed that the newlyidentified HRVs occupy a distinct phylogenetic position with the genus Rhinovirus [34] and are most closely related to the HRV A strains (also supported by SimPlot mapping; Figure 3 ). We also found support for our earlier data [31] that this clade is divided into two distinct subgroups. In the 2C and 3CD regions, these subgroups share less than 70% amino acid identity, which also occurs among members of the existing HRV species. Previously, preliminary and potentially subjective notes made by physicians upon first contact with ill individuals had been used for an approximate determination of the clinical impact of HRV-QPM [31] . To better address clinical impact, a comprehensive review of available patient charts was undertaken and two objective severity scoring tools were applied (Table 1) . Most HRV-QPM-positive individuals (76.5%) were admitted to hospital ( Table 2 ). Five patients were admitted for 96 hours or longer; two with severe illness and only HRV-QPM detected (variants 005 and 012) and one with moderate illness positive for HRV-QPM (variant 017) and HCoV-NL63. The remaining two had underlying medical conditions and multiple infections (HRV-QPM variant 009 and HCoV-229E with cystic fibrosis deterioration and HRV-QPM variant 008, HBoV and HCoV-NL63 with a viral URTI during admission for cardiac surgery). Oxygen therapy was required by almost half of all HRV-QPM-positive individuals studied. In addition to those described previously [31] , a new co-detection was identified with the newly identified polyomavirus, WUPyV [41] in a patient exhibiting a persistent cough, already positive for HRV-QPM (variant 013) and HBoV. In silico data obtained from this study permitted an enhanced analysis of HRV A2 strains using the ICTV criteria which assign a member to the genus Rhinovirus or Enterovirus [42] (Table 3) . The requirement common to both genera was for .70% amino acid identity in the P1 and 2C+3CD regions; this was not met by the HRV A2 genomes ( Figure 1B ). The remaining Rhinovirus species criterion was classification based on capsid-binding antiviral susceptibility and we found here, building upon earlier data [31] , that all HRV A2 strains contain a Thr 191 . This residue reportedly contributes to conveying resistance to the capsid-binding antiviral Pleconaril [4, 43] , placing the divergent HRVs into antiviral Group A [5] . Comparison to the Enterovirus species criteria revealed several features that may also be applicable for classifying HRV A2 strains. Firstly, although we could not identify a specific receptor using these in silico methods, the HRV-QPM models most closely accommodated interaction with an ICAM-1-like molecule, similar to some HEV C viruses [44] . Secondly, HRV A2 genomes exhibited #2.5% variation in G+C composition compared to HEV C and HEV D and ,5% compared to HEV A and HEV B. Thirdly, HRV-and HEV-like protease cleavage sites and locations suggested similar proteolytic processing ( Figure 1A ). Although we have been unable to find any indication of recombination in HRV-QPM [31] , the differences in amino acid identity within the novel HRVs, particularly in the 2C and 3CD region of HRV-X2, -C024 and -C025 compared to HRV-QPM, -X1 and -C026 (data not shown), may indicate a relevant event requiring further study. In sum, these data indicate that the HRV A2 strains cannot be assigned to an existing HRV or HEV species and should be assigned to a new species, tentatively called Human rhinovirus C. In 2007, the sequences of a number of divergent HRV strains were reported as a result of worldwide molecular investigations into cases of suspected ARTI [26, 31, 33] . Because these strains formed a distinct clade within the existing HRV As, we previously assigned to all Australian and similar New York strains, the title of HRV A2 [31] . Kistler et al. identified similarly divergent strains [32] and Lau et al. used sequence-based criteria to propose that some HRV A2-like strains could be classified as a new species [33] . Our ongoing efforts to characterise additional HRVs led us to catalogue the distinguishing molecular and clinical features of these HRV A2-like strains and to better address their taxonomic placement. We undertook in silico analyses, for the first time using all complete HRV A2 coding sequences and performed a retrospective review of the medical records of cases from which variants of the prototype strain, HRV-QPM, had been previously detected. To date the studies reported herein are the first of their kind to have been conducted on a single HRV strain. During our genomic analysis we found HRV A2 strains shared only 50%-53% average amino acid sequence identity with other HRV strains ( Figure 1) ; less with any strain from an HEV species. Average HRV A2 G+C content (42.4%) fell between the rhinovirus (38.7%) and enterovirus (45.6%) genera, suggesting an intermediate evolutionary path for HRV A2 strains. Nonetheless, a large number of HRV motifs, including those recognized by the RNA-dependent RNA polymerase, were retained by HRV A2 strains. This supported the proposal that some HRV genomic regions are under purifying selection [26] , even among the newly identified and divergent strains. HRV A2 strains were most similar to the HRV As overall, but amino acid sequences were as different from any traditional HRV or HEV species as the species in those genera were from each other. Previous subgenomic HRV A2 nucleotide phylogenies [31] [32] [33] were confirmed and expanded herein reinforcing that these NIVs reside within a distinct clade of the genus Rhinovirus, branching from the existing HRV A species ( Figure 6 ). The inability to propagate an HRV A2 virus could be due to many factors including the age, storage, site of origin and amount of the inoculum, its handling during processing, how long after symptoms appear before sampling occurs and the culture conditions used. It is also possible that an unknown primary receptor is employed or that binding of a secondary receptor is required for successful viral attachment and entry; or a mix of both. We sought to approach receptor-related issues using homology models derived crystallography data which already exist for a small number of HRV strains. Models of the component structural proteins from the prototypical HRV A2 strain, HRV-QPM, predicted that the characteristic 'jellyroll' conformations were retained (Figure 2 ) despite frequent and distinguishing deletions and inter-and intra-genus diversity at the sites of receptor contact and antigenicity (Figure 3 ). This also supported findings that HRV variability is mostly localized to receptor and antigenic sites [24, 33, 45] and begins to extend the data to include the more divergent strains. The VP1 BC and HI loops receive the VLDL-R molecule among some minor group HRV strains [12, 46, 47] . The comparative reduction in size of protrusions along the 5-fold axis, resulting from deletions in the BC, DE and HI loops of the HRV-QPM VP1, together with structure-altering deletions and substitutions affecting some antigenic sites previously identified from a minor group strain (HRV-2), particularly site A, lead us to propose that HRV A2 viruses are unlikely to behave as members of the minor group. Mapping the known ICAM-1 contact sites onto the predicted HRV-QPM structure revealed that most of the footprint was retained in structure by HRV-QPM, although it differed in sequence compared to other major group strains. SimPlot data revealed that sequence variation was not uncommon among strains within the major groups. Two small a-helices predicted to form in the HRV-QPM VP1 and VP3 footprints may disrupt receptor binding but since ICAM-1 is predicted to attach at differing angles in different HRV strains [8] , it may be suitably flexible to overcome such obstacles. Although NImIA and NImIB, first identified in a major group virus (HRV-14), were affected by the deletions in VP1, additional support for HRV-QPM belonging to the major HRV group came from structural similarities in the NImII (VP2) and NImIII (VP3). Amino acid substitutions occurred at these sites (ERG and GRT respectively) rather than insertions or deletions. These variations may result in distinct immunogenicity. Localisation of four known HEV receptor footprints to the predicted HRV-QPM structure identified at least one major sequence and/or structural divergence in each ( Figure 4 ). However, like the ICAM-1 interactions with HRV-14 and -16, HEV receptors could still bind at the same locations but utilize different amino acids on the capsid of HRV A2 strains [20] . Recent findings by Bartlett et al have shown that once past the receptor, HRV replication can occur within non-human tissues [48] exemplifying the importance of receptors in moderating the potential for more widespread HRV replication and illness. The high detection frequency of HRV A2 strains from patients with acute LRT illness may be another indication that these divergent viruses employ a tissue-specific receptor or it may be that all HRV strains cause similar illness, they simply have not been subjected to sufficient individual study. The LDLR family of molecules are located on many tissues in many species [10] and yet all these locations do not, to our knowledge, host HRV replication. Identifying an HRV A2 receptor in vitro cell culture system will be important for empirically addressing receptor-and replicationrelated issues and the outcomes may have broad implications for what is currently known of the attachment, entry and replication of HRV strains. Based on our predictive in silico studies of HRV-QPM, an ICAM-like molecule may be an early candidate for the role of receptor. Future in silico studies could examine all available HRV polyprotein sequences to predict the extent of flexibility in receptor interactions 'normally' occurring in the major and minor groups. This would provide a context for the structural differences we observed in HRV-QPM. The sequence of most HRV genomes is unknown and our understanding of HRV structure is mostly due to crystallography data derived from less than 5% of HRV strains. There are limited structural data to support the wholesale extrapolation of receptor binding behaviour and capsid structure to the other HRV strains. Without data to the contrary, in silico analyses appear to provide a suitable surrogate to address this issue. By employing RMSD calculations to examine the conformational similarity between predicted HRV-QPM structures and those from HRV structures determined empirically, we could be confident that our predictions approximated experimental data. All comparative values (0.064-0.766 Å ) were well below the maximum resolution used to originally determine HRV crystal structures (2-3 Å [14,15,49] ). Comparison to an HEV footprint also returned values below this threshold. While the inability to cultivate divergent HRV strains is a defining feature of the HRV A2s it is also a hindrance to classifying them. In our experience (data not shown), the earliest PCR-based methods [50] are already efficient at detecting these divergent strains. By extrapolating from published subgenomic sequence data, this clade is populated by a great number of divergent strains which are distributed both geographically and temporally, suggesting an endemic aspect to these previously uncharacterised viruses. No evidence exists to suggest that the HRV A2 strains are emerging viruses, rather it appears that they are newly identified strains that have been contributing to respiratory illness, in the absence of detection using culture-based diagnostic methods, for many years. There has now arisen a need for reliable protocols to characterise these putative viruses but in the absence of an HRV A2 isolate, neutralization, acid sensitivity and antiviral susceptibility studies, important for taxonomic placement, cannot be obtained. We have catalogued a series of features using methods which, while requiring further validation, contribute to our understanding of the biology and pathogenesis of HRV A2 strains. Additional molecular approaches will also be useful to classify other recently described divergent HRV strains [51] . Our data demonstrated that the newly identified HRVs, residing within the HRV A2 clade, do not satisfy the criteria for assignment to any existing HRV or HEV species and most likely constitute one or more novel species in the current genus Rhinovirus, tentatively called Human rhinovirus C (HRV C). Until recently, the genus Rhinovirus has been frequently neglected from clinical laboratory diagnosis and to date its members are not sought independently. Improved diagnostic methods, the broader application of existing PCR technology, new sequence data and many new insights have greatly enhanced our understanding of human rhinoviruses and the selective pressures impacting on their evolution [25, 26] . Similarly, the identification of a putative new HRV species, HRV C, the characterisation of HRV-QPM's molecular features and epidemiology and the identification of similar viruses around the world [31, 34, 52] have potential to reinvigorate HRV research. We found that the variants of one strain of the HRV Cs, HRV-QPM, sought in a well characterised population of specimens collected over all months of a single year, were more often associated with LRT illness than is commonly reported. The illness frequently presented as exacerbation of expiratory wheezing or persistent cough, with a requirement for supplemental oxygen, steroids and bronchodilator treatments. Despite exhaustive PCR-based investigation for other respiratory viral causes in our previous retrospective studies [31, 41] , HRV-QPM detection, in the absence of any other pathogen, occurred among infants with mild, moderate and severe (in both instances) LRT illness and most cases (88.2%) were admitted to hospital. Associated illness was usually (92.3% of the time) scored as mild to moderate and 58.8% of the HRV-QPM positive cases were children aged 12 months or less. The same age group represented 43.2% of the total study population originally screened for HRV-QPM (n = 1,244). While it is to be expected that a paediatric hospital-based population would overestimate the severity of HRV C clinical impact compared to a communitybased study, our study may in fact better represent the impact of first infection with HRV strains, the major pathogens detected during the first year of life [53] . It will be important to further characterize these NIVs and determine whether their differences confer distinctive biological behaviours, such as unique growth properties, antiviral resistance and discrete clinical outcomes among infected individuals. Strainfocussed studies could also identify how many distinct HRVs circulate in a single respiratory season, how often a given strain recurs in a population and just how many HRV strains there are. Nucleotide and amino acid compositions were determined using BioEdit Sequence Alignment Editorß v7.0.5.3. Picornavirus coding sequences (obtained from GenBank or picornaviridae.com) were determined by multiple alignments and subsequent truncations of the 59and 39UTRs. Discrepant nucleotides were substituted with the most frequently employed nucleotides, as determined by consensus alignment. Alignments available upon request. Predicted picornavirus cleavage sites were identified by amino acid sequence submission to the NetPicoRNA World Wide Web server [54] . Amino acid similarity was determined using SimPlotß v3.5 [37] employing the Hamming method on a 50% consensus of each group, a sliding window of 100 bp and a 1 bp step. Similarity data were mapped to the HRV-QPM capsid residues on the original sequence alignment and visualized using Chimera [55] . Picornavirus polyprotein sequences were compiled, translated, and aligned with the program BioEdit [56] . A neighbour-joining tree was generated using the Kimura two-parameter estimation in MEGA 3.1 [57] . Nodal confidence values (%), noted at the relevant nodes, indicate the results of bootstrap resampling (n = 1000). Because of the sequence diversity among the viruses analyzed, we undertook an additional analysis on a subgroup of strains whereby we gap-stripped the alignment, removing all gaps/ highly divergent regions and re-estimated the phylogeny (data not shown). This confirmed the phylogenetic position determined in the first instance. Secondary structures in the HRV A2 structural proteins were predicted using amino acid sequences submitted to the Jpred web server (http://www.compbio.dundee.ac.uk/,www-jpred/ submit.html). The location of a-helices and b-sheets were revealed by comparison to other picornavirus sequences using Cn3D 4.1 (http://130.14.29.110/Structure/CN3D/cn3d.shtml) (data not shown). To determine the predicted protomer structure of HRV-QPM, DeepView/Swiss-Pdb Viewer v3.7, [58] were used to locate and individually thread HRV-QPM amino acid sequences into HRV reference structures. Of the five HRV strains with empirically determined structural data available, HRV-14 [14] and -16 [49] shared the highest sequence identities with HRV-QPM in the regions chosen. Proteins VP1 to VP3 were aligned and threaded through HRV-16 (1ayn1, 1ayn2, 1ayn3 respectively) and VP4 through HRV-14 (1na1). Models were submitted to the SWISS-MODEL server for final threading analysis. The resulting HRV-QPM protein data bank files (PDB) were then matched and structurally aligned to a template HRV protomer (HRV-16, 1AYN) . Other picornavirus reference PDB files used in this study: HRV-2 (1FBN), HRV-3 (1RHI), HRV-14 (4RHV), CV-A21 (1Z7S), E-11 (1H8T), PV-1 (1HXS) and CV-B3 (1COV). Structural matching, alignments, ribbon figures, capsid predictions and root mean square deviation (RMSD) data were produced using Chimera [55] .

Epithelial Cell Apoptosis and Neutrophil Recruitment in Acute Lung Injury—A Unifying Hypothesis? What We Have Learned from Small Interfering RNAs

In spite of protective ventilatory strategies, Acute Lung Injury (ALI) remains associated with high morbidity and mortality. One reason for the lack of therapeutic options might be that ALI is a co-morbid event associated with a diverse family of diseases and, thus, may be the result of distinct pathological processes. Among them, activated neutrophil- (PMN-) induced tissue injury and epithelial cell apoptosis mediated lung damage represent two potentially important candidate pathomechanisms that have been put forward. Several approaches have been undertaken to test these hypotheses, with substantial success in the treatment of experimental forms of ALI. With this in mind, we will summarize these two current hypotheses of ALI briefly, emphasizing the role of apoptosis in regulating PMN and/or lung epithelial cell responses. In addition, the contribution that Fas-mediated inflammation may play as a potential biological link between lung cell apoptosis and PMN recruitment will be considered, as well as the in vivo application of small interfering RNA (siRNA) as a novel approach to the inhibition of ALI and its therapeutic implications.

The aim of this review is, therefore, to elucidate mechanisms of ALI pathogenesis, focusing on two main theories. The first, that neutrophils (PMN) can play a central role in driving ALI and the second, that lung epithelial cell apoptosis represents an important pathological aspect inducing lung injury. In this regard, we discuss the data in the field as well as some of our own findings using a clinically relevant double-hit mouse model of indirect ALI induced by hemorrhagic shock and subsequent polymicrobial sepsis. In addition, a novel strategy, i.e. the use of siRNA in mouse lungs in vivo, has served us in broadening our understanding of the pathology of ALI. The pathophysiologic mechanisms of ALI are insufficiently understood. In the common finale of this heterogeneous entity, the alveolo-capillary barrier is compromised allowing consecutive edema formation in the interstitium as well as alveoli, thus compromising gas exchange leading to organ dysfunction and respiratory failure. Histological evaluation of lungs from ALI patients indicated substantial accumulation of activated PMNs, diffuse alveolar damage, loss of epithelial integrity, and increased pulmonary edema (16, 17) . There are several proposed mechanisms which are believed to account for this phenomenon. Among leukocytes, PMNs are considered the first line of defense against microorganisms. Rapidly recruited to the inflammatory site/organ, they exert a variety of primarily beneficial functions (phagocytosis, production of reactive oxygen species [ROS] , and nitric oxide species [NOS] , degranulation of lytic enzymes, etc.), that, when well orchestrated, enable clearance of the invading pathogen. However, it is also hypothesized that activated PMN may possess harmful potential when these same functions are directed at otherwise normal host tissue, culminating in injury and organ damage (Figure 1 ). There is sub- Figure 1 . Proposed mechanisms of acute lung injury through hemorrhagic shock (HEM) "priming" for inflammation (Infl./red color)/apoptosis (Ao/ gray color)/injury, and "triggered" by a subsequent infectious insult: The resting lung (A) is primed by divergent inflammatory mediators released during an initial event (for example, shock, inflammation, etc.) that acts on a number of cells in the blood (MΦ: monocytes, PMN: neutrophils) and lung (EP: epithelial cells, EC: endothelial cells, AMΦ: alveolar macrophages) (B). These cells, in turn, stimulate, either separately or concomitantly, the pro-inflammatory response and/or the Ao of a small number of EP, both through Fas-FasL activation. The release of chemokines like MCP-1 then primes the AMΦ. When, at a later time, a subsequent inflammatory/infectious ('trigger') event takes place (C) the local EC, AMΦ, and/or EP become activated, release chemokines and activating agents that recruit the primed and now activated leukocytes (PMN, MΦ) to the lung (D). These activated leukocytes then transmigrate into the interstitium and alveoli where they perform their effector roles (in the absence of infection, the effector response may be solely injurious). In addition, they may propel the inflammatory/apoptotic response into a vicious cycle by further activating Fas through FasL on their cell surface (E). stantial evidence that the lung is particularly susceptible for PMN accumulation. In contrast to most other organs, where PMN sequestration occurs at post-capillary venules, pulmonary PMN retention takes place within the pulmonary capillaries, representing a complex interconnecting network of short capillary segments where the course from arteriole to venule crosses numerous alveolar walls and often includes more than 50 capillary segments. The blood in this complex network contains 50 times more PMN compared with most other vascular beds (reviewed in 18) . PMN accumulation has been observed early in lung tissue (19, 20) as well as in bronchoalveolar lavage fluids (BALF) of ARDS patients (16) . When developed during neutropenia, ALI rapidly progressed once PMN counts were restored (21, 22) . Furthermore, the degree of neutrophilia in BALF has been correlated with poor prognosis in septic ARDS (23) . In an experimental setting of indirect ALI stemming from hemorrhagic shock (HEM) followed by polymicrobial sepsis, we found that, in response to HEM, circulating PMNs exhibited an ex vivo increase in respiratory burst capacity and a decrease in apoptosis. This is consistent with the concept that shock/injury can produce in vivo "priming," the significance of which could be seen when HEM then was followed by sepsis, as recruitment of these PMNs into the lung occurred, along with the development of ALI (24) . Also, when these HEM-primed PMNs were injected intravenously into PMN-depleted animals, which subsequently underwent cecal ligation and puncture (CLP) to induce sepsis, ALI again resulted (24) . Upon this background, using different models of ALI, we (25) and others (26) (27) (28) (29) (30) (31) (32) have found that depletion of PMNs actually may serve to decrease injury associated with ALI. Thus, depletion of PMNs prior to HEM and sepsis markedly reduced the extent of lung inflammation and ameliorated lung protein influx and the severity of ALI (25) , which is in line with findings during transfusion-induced ALI (29) . Inflammation is closely linked to the pathogenesis of ALI. Levels of Interleukin (IL)-8 and IL-1 have been found increased in the lungs of ALI patients (33, 34) and the persistent increase of inflammatory cytokines in the lung correlated with poor outcome in ALI (35) (36) (37) . Inflammatory mediators exert several effects on PMNs. For example, complement protein C5a activates PMNs, thus contributing to their pulmonary sequestration and mediating lung tissue injury (38, 39) . Several inflammatory mediators, including complement, cytokines, chemokines, and lipid mediators also are able to induce the expression of surface molecules on the endothelium, which further promote PMN recruitment (40, 41) . Tumor necrosis factor (TNF)-α and IL-1β are two important cytokines that are found regularly in the BALF of ARDS patients (36, 42, 43) at higher concentrations than in their plasma, thus supporting their local origin (35, 44) . However, anti-TNF and anti-IL-1 therapies have failed to protect septic patients from ALI (44, 45) . Proinflammatory cytokines also are regularly expressed by activated PMNs traveling to the lung (32) . They appear to be a relevant source of IL-1β, favoring the subsequent release of other mediators, such as TNF-α, MIP-2, and IL-8 (46) (47) (48) . However, an early anti-inflammatory response driven by IL-1-, TNF-and IL-6-receptors, with early peaking IL-10 levels, was also shown in ARDS patients (41) . In addition, PMN stimulation with lipopolysaccharide (LPS) or TNF-α showed an activation of nuclear factor kappa B (NF-κB), p38, and AKT (49, 50) . The degree of nuclear translocation of NF-κB in peripheral PMNs from patients with septic ALI was linked to ventilator time and survival (50) . PMNs on Their Way to Lung-Sequestration, Adhesion, Migration, Activation, and Tissue Injury. First, in response to inflammatory mediators, PMNs migrate to the pulmonary capillaries in preparation for extravasation ( Figure 1 ). Initial changes in the cyto-skeleton prevent PMNs from deforming, making it less likely for them to pass through the pulmonary capillaries (51) . In this regard, activated PMNs from ARDS patients appeared even more rigid than those from septic patients (52) . Second, emigration of PMNs and passage through the endothelium can be regulated through adhesion molecules (51) . While their initial sequestration appears to be independent from adhesion molecules, the durability of this process could rely on them. Thus, while L-selectin-deficient mice showed a normal pool of marginated PMNs and cleavage of L-selectin from the PMN surface did not alter this margination (53, 54) , L-selectin deficient mice exhibited only a very short and transient leukopenia in response to complement activation (53, 55) . PMNs can adhere to endothelial cells using CD11/CD18 interacting with intercellular adhesion molecule-1 (ICAM-1) on the endothelial side. Thus, PMN emigration in response to Escherichia coli (E.coli), E.coli lipopolysaccharide, Pseudomonas aeruginosa, immunoglobulin (Ig) G immune complexes, and IL-1 was mediated through CD11/CD18, but this did not appear to be the case in response to Streptococcus pneumoniae, Group B Streptococcus, Staphylococcus aureus, hyperoxia, or C5a (56) . Even in CD11/CD18-dependent migration, blocking of CD18 reduced PMN emigration only by 60% to 80%, suggesting that other redundant mechanisms mediated these effects (56) . Cytokines and chemokines such as IL-8 also have been shown to be involved in increasing β2-integrin avidity (57) . It is important to understand that activation of PMNs is associated closely with the endothelial interaction mediated by the adhesion molecules (57, 58) . Rolling PMNs might well integrate the sum total of inputs received while scanning the endothelium. If an activation threshold is reached, β2-integrins switch to the high-affinity conformation, redistribute on the cell surface, and trigger arrest and adhesion (59, 60) . However, integrins also are involved in PMN migra-tion, phagocytosis, respiratory burst, and even cytokine production (57, 61, 62) . As eluded to above, chemokines are critically involved in the activation and recruitment of PMNs to the lung, potentially contributing to their harmful effects on an organ level. In this regard, associations between chemokines, lung neutrophil influx, and alveolo-capillary dysfunction have been reported (63, 64) . In our experiments, we observed that during hemorrhage-induced septic ALI (24) , as well as after experimental blunt chest trauma (65) , chemokines are markedly upregulated locally in the lung of the animals, as well as systemically. Blockade of the CXCR (chemokine [CXC motif] receptor) 2 using antileukinate, a hexapeptide inhibitor, following HEM, reduced lung PMN influx in response to subsequent sepsis and additionally decreased lung inflammation and lung protein leak, while pulmonary IL-10 levels were markedly increased (66) . We also administered anti-KC or anti-MIP-2 antibody into mice immediately following HEM. The adoptive transfer of PMN isolated from the anti-MIP, but not from the anti-KC-treated donors, exhibited a reduction of lung inflammation and lung PMN influx in the recipients in response to sepsis when compared with control Ig-treated donors (67) . These data demonstrated that hemorrhage-induced priming of PMNs not only mediated experimental ALI, but also that this process was differentially effected by MIP-2 and KC, although both signal through CXCR2 in mice. However, as humans utilize IL-8 in place of the two mouse α-chemokines and because receptors for IL-8, i.e. CXCR1 and CXCR2, also are utilized differently, these murine chemokine data may not translate directly to the patient setting of ALI/ARDS (68) . Once activated PMNs have reached the lungs, they possess the potential to induce tissue injury. The release of ROS/NOS has long been thought to be a central effector mechanism of PMNmediated lung injury (69) . In this regard, while the inhibition of the nicotinamide adenine dinucleotide phosphate (NADPH) oxidase attenuated sepsisassociated ALI (70) , surprisingly, NADPH oxidase-deficient animals were not protected from complement-induced ALI (71) . On the other hand, inhibition/ deficiency of the nitric oxide synthetase showed protection in complement-/ endotoxin-induced ALI (71, 72) . However, ROS and NOS also play additional roles as signaling molecules modulating kinases and phosphatases, receptors and transcription factors (reviewed in 73) complicating the interpretation as to whether their net pulmonary effect is one of harm or help. Proteolytic enzymes also are a part of the PMN pathogen-clearing arsenal. While neutrophil elastase, cathepsin G, and metalloproteinases all have important roles in clearing invading pathogens and function in aspects of tissue remodeling/wound healing, they also are able to mediate neutrophil-induced tissue damage and can degrade extracellular matrix efficiently. Thus, while inhibition of neutrophil elastase improved respiratory function during septic ALI (74), its deficiency also rendered mice more susceptible to gram negative, but not gram positive sepsis (75) . In addition, deficiency of elastase and/or cathepsin G was accompanied by increased susceptibility to fungal infections, but decreased susceptibility to endotoxin-mediated inflammatory ALI (76) . Activated gelatinase A correlated with protein accumulation in the epithelial lining fluid in ARDS patients and might, therefore, be a potential marker for ARDS (77) . The inhibition of matrix metalloproteinases and elastase prevented lung dysfunction and reduced septic ALI in pigs (78) . The detailed mechanisms by which PMN mediate ALI via the involvement of proteolytic enzymes are reviewed in detail elsewhere (79, 80) . Regulation of PMN Apoptosis. Neutrophils, once mature, exhibit a constitutive form of programmed cell death with a life span between 6 to 12 h in circulation. Under normal circumstances, activated PMNs are eliminated fairly quickly once the invading pathogen has been cleared. Several inflammatory agents, such as LPS, TNF, IL-8, IL-6, IL-1, Granulocyte [macrophage] colony-stimulating factor (G [M] -CSF), etc., can inhibit PMN apoptosis (81) (82) (83) (84) . The delayed apoptotic response provides PMNs with a longer life span, which, in turn, allows them to accumulate at local tissue sites of inflammation/infection. In this respect, PMNs obtained from patients following major trauma (85, 86) , burn injury (87), sepsis (88) , and ARDS (89) all showed evidence of decreased apoptosis. The anti-apoptotic effect of ARDS plasma on PMNs appears to be mediated through the GM-CSF receptor (90) . In sepsis, the delay in apoptosis was associated with a decrease in caspase-3 and -9 activity and a prolonged maintenance of the mitochondrial membrane potential (88) . The delay of PMN apoptosis also involved the active regulation of CXC receptors by PMNs themselves (91) . However, controversy persists as to whether this sustained activation actually contributes to organ injury. Activated PMNs have been shown to exert damaging effects in the lungs (68, 92) . However, following G-CSF treatment in pneumonia patients, no differences in outcome or time of recovery were noted, while it appeared that complications such as ARDS were decreased (93) . These data were complemented by studies that indicated that GM-CSF in the air spaces was associated with improved survival in patients with ARDS (94) . In our model of hemorrhage-induced septic ALI, it was evident that prolonging the lifespan of activated PMNs by using transgenic mice that overexpress the anti-apoptotic protein Bcl-2 in a myeloid-restricted fashion (95) did not exacerbate acute lung injury, although a prolonged presence of activated PMNs in the lung was evident (96) . In contrast, we observed an initial survival benefit when the life span of PMNs was further prolonged, likely due to an enhanced capacity to clear bacteria as evidenced by the lower bacteria counts in the lung (96) . In addition, in an inflammatory/ non-infectious environment, such as SIRS, mimicked by intraperitoneal injection of E. coli-LPS, prolonging the lifespan of activated PMNs was detrimental for the animals' survival and was associated with an exacerbation of lung injury (96) . Additionally, during non-infectious/ inflammatory ALI, failure to clear PMNs from the lungs contributed to increased inflammation and mortality (97) . Together, these data imply that the tissue environment (infectious versus inflammatory) that the neutrophil encounters plays an important role in determining whether PMNs mediate organ damage or not. However, the mechanisms which determine whether PMN turn bad or good are not understood well and require further investigation. The above information makes a strong case for PMNs playing a significant role in ALI. However, it is interesting to note that ARDS has been described as developing in patients with neutropenia (98) (99) (100) . In addition, following G-CSF therapy in pneumonia and sepsis patients, the incidence of ARDS was not increased (93, 101) and, during experimental ALI or pneumonia, PMNs have been shown to migrate to the lungs without exerting deleterious effects (102, 103) . Thus, other mechanisms also are likely to be involved in the pathogenesis of ALI. In the development of ALI/ARDS, the loss of epithelial cells in the lungs with consecutive impairment of the integrity of the alveolo-capillary barrier is commonly noted (19, 104) . Type I epithelial cells make up approximately 90% of the pulmonary surface, and, while type II epithelial cells account for only 10%, they are critical because they produce surfactant and may differentiate into type I cells (17) . Impairment of the alveolar epithelial barrier is, in many ways, important in the development of ALI. On the one hand, the epithelial barrier is, under physiologic conditions, less permeable when compared with the endothelial barrier, thus, destruction of its integrity prompts a progressive influx of proteinrich fluid into the alveoli (17, 102) . On the other hand, the loss of epithelial integrity represents an impairment of the physiologic trans-epithelial fluid transport and further inhibits the re-absorption of the alveolar edema (105, 106) . Thus, while an isolated injury to the endothelium still leaves intact the capacity of the epithelium to counteract pulmonary edema formation, the fluid balance becomes rapidly disturbed upon injury of the lung epithelium (107, 108) . Therefore, it is not surprising that the extent of epithelial damage and impaired edema re-absorption was linked clinically to outcome in ALI (109) . Additionally, the accompanying reduction of surfactant production led, via an increase in the surface tension, to an augmented respiratory effort affecting gas exchange (110, 111) . Programmed cell death/apoptosis of lung epithelial cells represents a potentially important mechanism contributing to the loss of this cell type in the development of ALI (Figure 1 ). Bachhofen et al. reported that patients who died from ARDS exhibited excessive apoptotic alterations in the chromatin of their alveolar type II cells (19) . Subsequent studies then confirmed these DNA fragmentations (112) as well as an increased expression of the pro-apoptotic protein Bax (Bcl-2 associated X protein) (113) . More recent studies also have suggested a role for lung epithelial cell apoptosis in pediatric ARDS (114) . Importantly, epithelial cell and PMN apoptosis is differentially regulated during ARDS. While, as mentioned earlier, PMNs show a decreased rate of apoptosis, it is increased markedly in epithelial cells (107) . Local mediators in the lungs of these patients may regulate these opposing effects (115, 116) . Interestingly, the induction of lung epithelial cell apoptosis appears to lead to the development of ALI, which is further aggravated, when apoptosis is additionally induced in macrophages (117) . Furthermore, inhibition of apoptosis, using a caspase inhibitor, protected mice from the lethal consequences of endotoxin-associated ALI (118) . In addition, Bcl-x(L) treatment reduced albumin leakage and lung tissue damage in LPS-mediated ALI (119). In the induction of lung apoptosis during ALI, Fas-(an apoptotic death receptor pathway) mediated cell death also appears to play a major role (Figure 2 ). In this regard, during endotoxin-induced ALI, an increase of Fas expression on epithelial cells, as well as an immigration of Fas-ligand (FasL) expressing cells in(to) the lung, was observed (120). Human lung epithelial cells also expressed Fas and were, particularly in the distal airways, sensitive to Fas-mediated apoptosis (121) . Caspase-cleaved cytokeratin-18, a marker for epithelial cell apoptosis, also was increased in BALF during ARDS (122) . During ALI and ARDS, an increased concentration of Fas and FasL in patients' BALF and lung tissue was detected (122, 123) and BALF from ARDS patients was shown to induce apoptosis in healthy lung epithelial cells (116) . Furthermore, ARDS non-survivors exhibited markedly higher FasL concentrations when compared with survivors (116) . Particularly during septic ALI, an infection-severity dependent activation of the Fas-FasL system in the lung was observed (124) . FasL concentrations have been reported to be much higher locally than systemically, suggesting its origin is pulmonary (122, 123) . In this respect, there appear to be several potential sources of FasL: infiltrating monocytes have been implicated (but not alveolar macrophages) (125), PMNs have been implicated (126) , and FasL may be cleaved from cell membranes by the activation of matrix metalloproteinases 3 and 7 (127, 128) . However, anti-apoptotic proteins such as soluble Fas (129) also appear to be increased during ARDS (122, 123) and have been described to correlate with clinical features such as the PaO 2 /FiO 2 ratio (122) . This may point toward a simultaneous activation of protective anti-apoptotic mechanisms during ALI. Under experimental conditions, the instillation of Fas-activating antibody into murine lungs induced apoptosis of lung epithelial cells, PMN recruitment and impairment of the alveolo-capillary barrier (130) . This development of Fas-dependent lung injury was linked to Fas-expression on non-myeloid but not myeloid cells (131) . Furthermore, blocking of the Fas-FasL system reduced the development of endotoxin-mediated ALI (120) . Our experiments further demonstrated the relevance of Fas-mediated apoptosis in the lung during HEMinduced septic ALI (132) . Fas and FasL mutant animals exhibited less pulmonary epithelial cell apoptosis in response to the insult when compared with background animals. In addition, the extent of ALI as assessed histologically and by protein influx was diminished significantly. This was associated with a survival benefit for Fas mutant mice when compared with background animals (132) . It also has been reported that Fas-FasL deficiency is associated with a reduction in the degree of injury seen during direct ALI in E. coli, S. aureus, or S. pneumoniae sepsis models (133) . Similar results were described for sepsis induced by Legionella pneumonia (134) . In contrast, data from pulmonary sepsis fol-lowing lung infection with P. aeruginosa suggest a reduction in lung apoptosis in the absence of Fas signaling as well, but a decreased survival rate was observed, associated with an increased dissemination of the bacteria (135) . It should be noted that Fas-mediated apoptosis in the lung appears to be modulated by several factors such as surfactant protein A, Angiotensin II, transforming growth factor (TGF)-beta, decoy receptor 3, etc. (reviewed in 108) , and that other apoptotic pathways, for example, mediated by TNF-α (136) or mitochondria (137) , appear to play a role under certain circumstances, too. Recent studies suggest that activation of Fas serves not only to induce apoptosis, but also to induce the secretion of cytokines and chemokines by a variety of cell types (138) (Figure 2 ). In murine lungs, activation of Fas initiated a pro-found inflammatory response with an early generation of chemokines and subsequent recruitment of neutrophils (130, 132, 139) . This response could be reduced by antagonizing Fas ligand (139) . Triggering of Fas also led to increased concentrations of TNF-α, MIP-1 α, MIP-2, monocyte chemotactic protein (MCP)-1, and IL-6, and compromised the alveolo-capillary barrier (133) . LPSmediated pulmonary inflammation also appears to be regulated through Fas (140) . Mice, genetically altered to express Fas either on myeloid or on nonmyeloid cells in the lung, presented no marked increase of inflammatory mediators following Fas activation (131) , in contrast to non altered mice lungs (130, 132, 139) . With respect to cell types involved, it has been indicated that activation of Fas on monocyte cell lines induced production of MIP-2 (141) . The activation of an inflammatory program by Fas also was shown for peritoneal macrophages (142) , monocytes and macrophages (143) , endothelial cells (144) , and others (145) (146) (147) (148) (149) (150) (151) . Our own experiments with the Fas and FasL mutant mice showed a marked decrease in inflammation in the lungs of these animals in response to hemorrhageinduced septic ALI, when compared to background animals (132) . This phenomenon also was evident following Fas silencing in the mouse lungs using siRNA (152) and was associated, in both cases, with a reduction of PMN immigration into the lungs (132, 152) . Further experiments revealed that lung epithelial cells were capable of secreting MIP-2, KC, and MCP-1 in vitro in response to Fas activation, through mechanisms involving ERK and, potentially, FLIP (132), complementing the findings of other researchers who demonstrated that a NF-κB dependent mechanism appears to underlie this response (151) . Instillation of a Fasactivating antibody into transgenic murine lungs in which lung macrophage numbers are markedly reduced displayed a similar inflammatory response as seen in background animals, further supporting a role for Fas-mediated, ep- ithelial cell-induced pulmonary inflammation in vivo (132) . The fact, that inhibition of Fas activation modulated both PMN recruitment into the lung as well as pulmonary epithelial apoptosis, intriguingly points at its value as a potential therapeutic target in ALI. The use of silencing RNA (siRNA) represents not only a potentially powerful experimental approach to allow us to better understand the evolving pathology of ALI, but may possibly represent a novel therapeutic approach to the treatment of this condition. The history of siRNA, its discovery, development, the mechanisms involved, as well as its successful initial uses in mammals in vivo are described elsewhere (153) (154) (155) (156) . With respect to its application in vivo, the lung appears to be a good candidate, as it can be accessed straightforwardly intranasally (i.n.) or intratracheally (i.t.). Nevertheless, nucleic acid transfer efficiency can be diminished substantially by phospholipids and proteins, which are major components of the airway surface liquid (157) . Interestingly, unlike systemically, delivery of naked siRNA into the lungs has proven very efficient (152, 158, 159) , thus potentially complex and costly approaches using vector systems or chemical siRNA modifications may not be necessary. In this regard, the delivery of siRNA inhibitors of SARS coronavirus in 5% glucose was superior, even in reducing the severity of the disease, than when given in modified calf pulmonary surface active material (160) . As early as 2004, the feasibility of a surfactantbased or naked siRNA approach in the mouse lung targeting glyceraldehyde-3phosphate dehydrogenase (GAPDH) or heme oxygenase-1 during ischemiareperfusion, respectively, has been demonstrated (161, 162) . Phase I/II trials for Respiratory Syncytial Virus have been conducted and are now on their way to evaluate the safety, tolerability, and antiviral activity of siRNA treatment in human lungs (155, 158, 163) . With respect to the application of siRNA in the in vivo setting of the lung, initially we attempted to extend our observation that blockade of MIP-2 or KC with conventional antibodies effected the development of ALI (67) by using an in vivo approach silencing these same chemokines locally in the lung (164) . Subsequently, we initiated studies with antiapoptotic siRNA against Fas and caspase-8 assessing their capacity to protect the lung from the detrimental effects of hemorrhage-induced septic ALI (152) . However, to establish the initial feasibility of in vivo siRNA administration, mice overexpressing green fluorescent protein (GFP) were chosen to receive a single intratracheal instillation of a GFPsilencing RNA. What we found was that green fluorescence in the lungs of these animals at 18 h post-instillation was reduced by at least 65% when compared with vehicle-treated GFP mice, while no decrease in fluorescence was seen in other organs outside the lungs (152, 164) . The use of siRNA at these concentrations neither induced substantial activation of type I interferons (165, 166) via activation of TLR-3 (167), TLR-7 (168), TLR-8, TLR-9, and protein kinase-R(PKR) pathways (169) (170) (171) nor via more classic proinflammatory processes (169, 171) . This is in line with reports demonstrating that intravenous delivery of naked siRNA did not induce an interferon response (170) . However, the induction of STAT (signal transducers and activator of transcription)-1 in this context appears to be dose dependent (165) , and thus it cannot be ruled out that different dosages, delivery, or timing of administration might have provided dissimilar results. To gain some insight into the cellular localization of siRNA in the lung, we followed the intrapulmonary deposition of Cy-5 fluorochrome labeled siRNA uptake by confocal immunofluoresence microscopy. Interestingly, labeled siRNA was found to co-localize only with lung epithelial cells counterstained with anti-cytokeratin-18 at 24 h post instillation, but not alveolar macrophages (152) . However, negative results for alveolar macrophages may not preclude the uptake of siRNA by this cell type as one can imagine different kinetics in the various cell types with respect to siRNA uptake and processing. Relative to this, the feasibility of gene silencing in macrophages using siRNA has been described in vitro (172) (173) (174) , however, silencing of typically macrophagederived molecules such as TNF-α and IL-6 during indirect murine lung injury remained unsuccessful in our hands (unpublished observations). Whether this might be attributable to the underlying cell type or perhaps to the degree of upregulation of a certain gene during pathological conditions remains to be elucidated. Irrespective these and the experiments below represent the first studies to demonstrate the feasibility of in vivo lung siRNA as treatment for developing ALI. To finally assess the therapeutic capacity of such a siRNA approach, experiments were designed to modulate PMN immigration based on the neutrophil hypothesis using in vivo siRNA constructs against the murine chemokines KC and MIP-2 during the development of indirect septic ALI. To this end, suitable siRNA constructs were instilled into the lungs of animals 2 h following HEM and prior to CLP. ALI was assessed 24 h after the induction of sepsis. Silencing of MIP-2 markedly reduced tissue and plasma IL-6 concentrations, tissue MIP-2, as well as lung PMN influx, interstitial edema, alveolar congestion, and disruption of lung tissue architecture (164) . In contrast, KC-siRNA treatment, while reducing plasma KC, tissue KC, and tissue IL-6, produced neither a significant reduction in plasma IL-6 nor lung PMN influx nor lung damage (164) . Alternatively, based on the epithelial cell hypothesis, siRNA sequences against Fas and caspase-8 were intratracheally instilled during septic ALI. Interestingly, while these sequences markedly diminished lung Fas and caspase-8 expression in a gene specific fashion, when lung apoptosis was assessed, pulmonary tissue caspase-3 activity was reduced only in response to Fas but not caspase-8 silencing. As silencing here, as commonly observed, did not result in a total depletion of the target protein, but rather in a diminished expression, it is possible that sufficient active caspase-8 was still present to mediate the apoptotic effects either through direct activation of caspase-3 or indirectly through cleavage of the proapoptotic protein Bid (152, 175) . Our data also indicated that silencing of Fas on lung epithelial cells was associated with a reduction in lung inflammation, PMN influx, and a diminished extent of lung injury (118) , thus emphasizing the relevance of epithelial cell integrity during indirect septic ALI. Keeping in mind the diversity of clinical scenarios, we use the inclusive acronym ALI, one could think that it might be exactly this imprecise global view that prohibits the development of novel therapeutic approaches for ARDS. The diversity of underlying pathologic mechanisms prohibits the formulation of a unified pathophysiology of this clinical entity. Whether different forms of ALI (for example, direct versus indirect) are more of a response to a certain patho-mechanism than another remains to be determined. However, while there may be numerous diverse stimuli that can initiate the pathogenesis of this clinical entity, the final steps in ALI/ARDS, such as compromise of the alveolocapillary barrier function, appear to be somewhat common. Here we have reviewed the data which support the role of the activated PMN in lung injury as well as the contribution of epithelial cell death as independent entities contributing to ALI. Further, we have shown how recent studies looking at epithelial cells and the silencing of Fas-induced apoptosis and/or inflammation in these cells may serve as a lynchpin linking these two pathological processes. Upon early Fas activation, alveolar macrophages and lung epithelial cells might produce chemokines in the lung, thus attracting activated and potentially harmful neutrophils, monocytes, or even T-lymphocytes to the site of injury and potentiating the degree of injury. Finally, we trust we have illustrated how the application of siRNA-induced gene knockdown in the lung not only has produced novel insights into pathogenesis of ALI, but may represent a potential exciting therapeutic approach for its treatment.

The cucumovirus 2b gene drives selection of inter-viral recombinants affecting the crossover site, the acceptor RNA and the rate of selection

RNA–RNA recombination is an important pathway in virus evolution and has been described for many viruses. However, the factors driving recombination or promoting the selection of recombinants are still unclear. Here, we show that the small movement protein (2b) was able to promote selection of RNA 1/2–RNA 3 recombinants within a chimeric virus having RNAs 1 and 2 from cucumber mosaic virus, and RNA 3 from the related tomato aspermy virus, along with heterologous 2b genes. The source of the 2b also determined the selection of the acceptor RNA and the crossover site, as well as affecting the rate of selection of the recombinant RNAs. The nature of the RNA 3 also influenced the selection of the recombinant RNAs. A 163-nt tandem repeat in RNA 3 significantly affected the rate of selection of the recombinant RNA, while a single nucleotide within the repeat affected the crossover site. The recombination occurred in a non-random manner, involved no intermediates and probably was generated via a copy-choice mechanism during (+) strand RNA synthesis.

RNA-RNA recombination is one of the most important pathways in virus evolution. It was first discovered in the early 1960s in poliovirus (1, 2) and has been documented in various genera of animal viruses, plant viruses and bacterial viruses group. Studies conducted on viruses during the last 40 years indicated that RNA-RNA recombination could occur between RNAs of the same or different strains of one species (3, 4) , of different species (5) or between viral and host cellular RNAs (6, 7) . RNA hairpins or mutations in a replicase domain have been implicated to play an important role in promoting RNA-RNA recombination (8) (9) (10) (11) . While none of those studies demonstrated a role for nonreplicase proteins in viral RNA-RNA recombination or selection of recombinants, recombination without replication has been described (12, 13) . RNA recombination has been studied extensively in the plant virus genus Cucumovirus. Those studies have involved two of the three recognized viral species in this genus, Cucumber mosaic virus (CMV) and Tomato aspermy virus (TAV). These viruses both contain a single-stranded and positive-sense RNA genome divided into three species designated RNAs 1, 2 and 3 (14, 15) . RNA 1 encodes the 1a protein involved in virus replication (16, 17) . RNA 2 encodes two proteins: the 2a, also involved in virus replication (16, 17) and the 2b protein, involved in virus movement and the suppression of RNA silencing (18) (19) (20) (21) . RNA 3 also encodes two proteins, the 3a movement protein and the capsid protein, both of which are also involved in virus movement (22) (23) (24) . Protein 2b is translated from a subgenomic RNA of RNA 2, RNA 4A (25, 26) , while the capsid protein is translated from a subgenomic RNA of RNA 3, RNA 4 (27) . The 3 0 terminal 305-310-nt non-translated region (NTR) is nearly identical in each of the three genomic RNAs of each virus and represents promoters for (À) viral RNA synthesis during viral replication (23, 28, 29) . The location of most of the recombinants described involving CMV and TAV is within these 3 0 NTR sequences. RNA-RNA recombination was first described in cucumoviruses isolated from tobacco plants after multiple passage of a pseudorecombinant (reassortant) virus containing RNAs 1 and 2 from CMV, and RNA 3 from TAV (30) . The progeny RNA 3 consisted mostly of TAV RNA 3 but with the 3 0 terminal sequence from CMV RNA 2. In addition, in an earlier passage, there was an intermediate RNA 3 present containing a duplicated 21-nt sequence at the crossover site (30) . Subsequently, several other cucumoviral recombinant RNAs were also discovered in plants infected with a so-called quadripartite hybrid virus having RNAs 2 and 3 from CMV, and RNAs 1 and 2 from TAV (31) , with a mixture of wild-type (wt) CMV and wt TAV (32, 33) , with mixed CMV isolates (34) (35) (36) , or within a single wt CMV (37) . These characterized recombinant RNAs had non-templatederived nucleotides at the crossover site (31) and had CMV sequences at one end of the crossover site and TAV sequence at the other end (32) or had the same 3 0 -terminal NTR sequence for all three genomic RNAs (37) . Some of the recombinants had crossover sites localized within a short region of high sequence similarity (32) . Others were predicted to occur in regions of high secondary structure (33, 34, 38) . TAV RNA 3 itself has a direct repeat of 163 nt (differing by one nucleotide) in the 3 0 NTR, which is not essential for infection (39) and in some cases, was deleted during long-term incubation (40) . While some of the above studies examined recombinants formed in the directly inoculated leaves and thus avoided detecting only those replicable recombinants that might be selected (32, 33) , other studies examined those recombinants detected after slow selection (31, 38) and found that some recombinants could co-exist in the same plants (38) . However, what factors may have led to the selection of particular recombinants has not previously been examined. Here we show that infection and passage of a chimeric virus having RNAs 1 and 2 from CMV, with RNA 3 and the 2b gene from TAV consistently resulted in the slow selection of recombinant viruses. We demonstrate that the 2b gene plays a key role in promoting this selection. Plasmid cDNA clones, with transcription in planta controlled by the cauliflower mosaic virus (CaMV) 35S promoter and terminator, were used for plant inoculation. A pGEX2T-based plasmid (Amersham Pharmacia Biotech), pGEX-T2b, in which the gene encoding the TAV 2b protein was fused with sequences encoding GST, was used to express GST/TAV2b for RNA-binding assay. Various plasmids used in this study are shown in Figure 1 . Plasmid cDNA clones C1 (previously named pQCD1), C2 (pQCD2), C2 T2B (pQCD2qt), C2 T2BC2a (pQCD2qt1), C2 W2B (pQCD2qw), C3 (pQCD3), T1 (pCass1T1), T2 (pCass1T2), T3 (pCass1T3) and T3 Á163(A) (pCass1T3 Á163(A) ) have all been previously described (26, 39, (41) (42) (43) . The other plasmid cDNA clones were constructed as follows. Plasmid cDNA clones C1 Á23 , C2 T2BÁ23 and T3 Á163(A)Á23 were constructed through three steps: first, amplifying two fragments with appropriate primer pairs (see Supplementary Table 1 Figure 1 . Schematic representation of constructs used in this article. The constructs regulated from the 35S promoter and 35S terminator are indicated. The open reading frames encoding various cucumoviral proteins on the constructs are shown. The 23 nt of sequence identity between all genomic RNAs is indicated by a short thick bar. The tandem repeat of 163 nt is indicated by an arrow. The crosshatch lines represent the amino acid sequences different from the native sequence for either 2a or 2b. T3 Á163 (A)Á23 ; and finally, by ligating the two digested fragments into the appropriate sites created by digestion with the same enzymes. Plasmid cDNA clones C1 3081-3390 , C2 W2B2757-3065 , T3 Á163(A)Á141 , T3 1-2060 , T3 1-2223 , T3 1902-2386 and T3 2065-2386 were constructed via two steps: first amplifying appropriate fragments, and then ligating these fragments into the pCass1 vector (43) . Plasmid cDNA clones C2 W2BC2a and T3 Á163(G) were constructed with a QuickChange Site-Directed Mutagenesis Kit (Stratagene) according to the manufacturer's instruction. All the amplified fragments as well as their joined regions in the plasmids were sequenced to ensure whether they were correct. Two plant species, Nicotiana clevelandii and Nicotiana glutinosa, were chosen for inoculation of the plasmid cDNA clones. Unless specified otherwise, all inocula used were a mixture of the three genomic cDNA clones (corresponding to RNAs 1, 2 and 3), each at an equal concentration of 10 mg/10 ml per plant. For each test, at least five, four-leaf plants of each of the above two species were inoculated mechanically after a 24-h dark treatment. Each test was repeated at least once. Total plant RNAs were extracted from infected plants and analysed by northern blot hybridization as described previously (26) . The nylon membranes were hybridized with a 32 P-labelled transcript probe, complementary to the 3 0 terminal sequences of either TAV RNA 3 (nucleotide 2287-2386) or RNA 2 of the Q-strain of CMV (Q-CMV; nucleotide 2871-3035). These two probes would hybridize to all the genomic and subgenomic RNAs of the corresponding viruses (39) . An additional probe complementary to nucleotide 1-1884 of TAV RNA 3 was specifically used to hybridize to the 5 0 -terminal region of TAV RNA 3. Viral particles and virion RNAs were purified according to the method of Peden and Symons (44) . Individual viral RNAs were purified from the virion RNAs first via an agarose gel and then via a polyacrylamide gel (45) . The purified, single, viral RNAs were polyadenylated with Escherichia coli poly(A) polymerase and then RT-PCR was done using an oligo dT primer as described previously (43) . The amplified products were cloned into the pBluescript SK + vector and sequenced from both orientations. The purified, single, recombinant RNAs, total RNAs extracted from plants and total virion RNAs were also reverse transcribed using primers T3 0 and C3 0 , respectively. T3 0 is complementary to the 3 0 -terminal sequence of TAV, while C3 0 is complementary to the 3 0 -terminal sequence of Q-CMV (see Supplementary Table 1 for details). The synthesized first-strand cDNAs from the templates were amplified with different pairs of primers, C3 0 /T3-1187, C3 0 /C2-980, C3 0 /C1-2032, T3 0 /T3-1187, T3 0 /C2-980 and T3 0 /C1-2032 (see Supplementary Table 1 for details), using the thermocycler program: 1 min at 948C, 1 min at 528C and 2 min at 728C for 30 cycles. The amplified products were also cloned into the pBluescript SK + vector and sequenced from both orientations. At least five clones were sequenced from each PCR product. DNA sequencing was carried out using 4 pmol of a specific primer and 0.1 mg of a cDNA clone. The sequence reaction was incubated at 378C for 10 min and then analysed via a manual sequencing gel system. RNA sequencing was carried out under the same conditions as the DNA sequencing except that 10 mg of the total plant RNAs, 2 mg of the virion RNAs or 1 mg of a single RNA were used as a template and that the sequence reaction was incubated at 428C for 30 min. The plasmid pGEX-T2b was constructed via two steps: amplifying the PCR fragments first and then cloning the PCR fragments into appropriate vectors (as above). PGEX-T2b was transformed into E. coli BL21 (DE3). Protein expression was induced with 0.1 mM IPTG (isopropyl-b-D-thiogalactopyranoside). Purification of the protein was carried out using glutathione beads. The cleavage of GST/TAV2b protein was carried out using thrombin (Novagen) as recommended by the manufacturer. The purity of the purified protein was analysed by SDS-PAGE. The amount of purified protein was measured using the Bio-Rad protein assay. Preparation of full-length TAV RNA 3 transcripts DNA templates representing a full-length RNA 3 of TAV and including T7 or T3 promoter were generated by PCR using the infectious cDNA clone of TAV RNA 3 (T3) as a template and primer pairs T3 0 and T7-TR3(+) (a primer used for transcription of the sense RNA), or T5 0 (nucleotides 1-27 of TAV RNA 3) and T3-TR3(À) (a primer used for transcription of the anti-sense RNA; see Supplementary Table 1 for the primer sequences). Labelled transcripts from the DNA templates were obtained using [ 32 P]-a-UTP and T7 or T3 DNA-dependent RNA polymerase. The free nucleotides in the transcription reaction were removed using P-30 micro Bio-Spin columns (Bio-Rad), while the DNA templates were removed by incubation with DNase I. The labelled TAV RNA 3 transcripts were purified through a 4% denaturing polyacrylamide gel and quantified by UV spectrophotometry (Beckman). Three different amounts (150, 300 and 450 ng) of the TAV 2b protein purified from E. coli were each incubated with 4 ng of each of the [ 32 P]-labelled TAV RNA 3 transcripts in a binding buffer [50 mM Tris-HCl (pH 8.2), 10 mM MgCl 2 , 1 mM EDTA, 10% glycerol, 200 ng of yeast tRNA (Sigma), and 2 U of RNase inhibitor (Promega)] at 258C for 30 min (46) . After incubation, potential protein-RNA complexes were analysed via electrophoresis on a 4% non-denaturing polyacrylamide gel and detected by autoradiography. Pseudorecombinant viruses formed between RNAs 1 and 2 of CMV and RNA 3 of TAV have been shown gradually to yield stable, recombinant viruses after multiple passage (30, 31, 38) . However, the effects of the various viral genes on the selection of these recombinant viruses have not been examined. Previously, using interspecies hybrid viruses, we examined the roles of different 2b genes on virulence and virus accumulation and showed that while the nature of the 2b gene influenced virulence, the RNA 3 (and therefore the two genes encoded by RNA 3; the 3a movement protein and the capsid protein) did not (21, 47) . Therefore, the hybrid viruses generated previously, C1C2 T2B C3 and C1C2 T2B T3, where C indicates Q-CMV, T indicates TAV, 1, 2 and 3 indicate RNAs 1, 2 and 3, and the subscript T2B indicates that the 2b gene of Q-CMV was replaced precisely with that of TAV ( Figure 1 ), were subjected to long-term, multiple passages in N. glutinosa to determine whether the 2b gene had any effect on the selection of recombinant viruses. The progeny viral RNAs were examined by northern blot hybridization, with probes specific to the 3 0 terminal NTR sequences of either the three Q-CMV RNAs or the three TAV RNAs (Figure 1 ). This analysis indicated that in the multiple passage plants TAV RNA 3 derived from C1C2 T2B T3 had recombined with CMV RNA, as the progeny RNA 3 was detected by both the Q-CMV-specific and TAV-specific probes (Figure 2A and B, compare lanes 9 and 20). By contrast, RNA 3 derived from the same virus maintained for several weeks in the plasmidinoculated plants was not a recombinant RNA, as it was only detected by the TAV-specific probe, but not by the Q-CMV-specific probe (Figure 2A and B, compare lanes 5 and 16). The specificity of the CMV-specific and TAV-specific probes was demonstrated in that both probes only detected their own specific viral RNAs ( Figure 2A and B, compare lanes 2 and 3, and lanes 13 and 14) . The other RNAs of C1C2 T2B T3 were still detected by their specific probes in both the inoculated plants and the multiple passage plants (Figure 2A and B) . To confirm that RNA 3 was a recombinant, the RNA 3 derived from C1C2 T2B T3 was gel-purified, reverse transcribed with primers C3 0 and T3 0 , respectively, and then amplified with different pairs of primers as described in the Materials and Methods section. Only the primer pair C3 0 /T3-1187 was able to generate a specific RT-PCR product (data not shown). T3-1187 corresponded to sequences ending with nucleotide position 1187 of TAV RNA 3, while C3 0 was complementary to the 3 0 -terminal sequence of Q-CMV. This indicates that the RNA 3 was indeed a recombinant, having a TAV RNA 3 sequence in the capsid protein region and a CMV RNA sequence at the 3 0 terminus. To determine the nature of the recombinant RNAs derived from C1C2 T2B T3, the recombinant RNAs derived from this virus were purified from infected plants, subjected to RT-PCR and the RT-PCR products were then cloned into the pBluescript SK + vector for sequencing. The sequencing results from five independent clones showed that there was one type of RNA 3 recombinant derived from C1C2 T2B T3, containing 2370 nt, of which the 5 0 terminal 2060 nt was derived from the 5 0 terminal 2060 nt of TAV RNA 3 ( Figure 3A right, open bar), while the 3 0 terminal 310 nt was identical to the 3 0 terminal 310 nt of Q-CMV RNA 1 ( Figure 3A right, stippled bar). The crossover site on the TAV RNA 3 sequence was 7 nt 5 0 of a homologous sequence of 23 nt conserved in all three genomic RNAs, and 3 nt 5 0 of this conserved 23 nt sequence in the Q-CMV RNA 1 sequence. The TAV RNA 3 sequence in the recombinant RNA 3 retained the first repeat [in order of 5 0 to 3 0 on the (+) sense strand] of two 163-nt tandem repeats (39) , but lacked the second one, which originally was present in the TAV RNA 3 inoculum. The difference between the two 163-nt repeats is an A at position 1966 in the first repeat, with a G at the corresponding position (2129) in the second 163-nt repeat ( Figure 3A left) (39) . The repeated sequences showed no difference in viral pathogenicity when only one or the other was present (39) . The 3 0 terminal 310 nt of the recombinant RNA 3 was not only the same as the 310 terminal nucleotides of Q-CMV RNA 1, but was also the same length as Q-CMV RNA 5. Q-CMV RNA 5 is a mixture of subgenomic RNAs derived from the 3 0 terminal NTR of RNAs 1, 2 and 3 [(48); our unpublished data]. The level of RNA 5 has a strong effect on the severity of symptoms; the presence of more RNA 5 in the plants results in less severe symptoms (our unpublished data). Repeated multiple passages of virus generated from the plasmids comprising C1C2 T2B T3 in 10 N. glutinosa plants or 10 N. clevelandii plants resulted in the appearance of the same recombinant RNA 3s with the same border sequences among the sequenced clones. These recombinant RNAs were not detected by northern blot hybridization in the inoculated parental plants, even if the plasmid-inoculum was increased 5-fold (50 mg/10 ml) or decreased 20-fold (0.5 mg/10 ml), although the same recombinant RNAs were detected again in the multiple passage plants, when the initial inoculum concentration was changed (data not shown). The observation that recombinant RNAs were detected when the 2b gene of Q-CMV was substituted by that of TAV (Figure 2A , lane 9), and not when the 2b gene of Q-CMV was present in the interspecific hybrid virus C1C2T3 (Figure 2A , lane 11) suggested that the 2b gene has some role in the generation and/or selection of the recombinant viruses. However, as the replacement of the 2b sequence also changed the sequences encoding the C-terminal 41 amino acids of the 2a protein, it cannot be ruled out that the C-terminal 41 amino acids of the 2a protein might play a role in the recombination and/or selection. Therefore, to determine whether the C-terminal 41 amino acids of the 2a protein encoded by the overlapping TAV 2a gene sequence had such a role, we introduced a stop codon upstream of the TAV 2b gene in C2 T2B to prevent the expression of the C-terminal 41 amino acids of the 2a protein (see C2 T2BC2a in Figure 1 ). Bioassay of C1C2 T2BC2a T3 onto N. glutinosa showed that the virus induced symptoms similar to those induced by C1C2 T2B T3 and northern blot hybridization and RT-PCR analyses showed that recombinant RNA 3 still accumulated in the multiple passage plants, but not in the inoculated plants (data not shown). Moreover, the recombinant RNA 3 generated was the same as that derived from C1C2 T2B T3 (data not shown). Therefore, the C-terminal 41 amino acids of the 2a protein encoded by the TAV 2b sequence had no role in the recombination and/or selection, indicating that the TAV 2b gene was required for the recombination and/or selection. That the 2b protein rather than the 2b RNA sequence itself affected the generation and/or selection of recombinant RNA 3s could not be established conclusively, since expression of the 2b gene was shown to be necessary for the cell-to-cell movement, but not the replication of the pseudorecombinant virus C1C2T3 (21) . However, if the 2b protein was actually involved in the recombination/ selection process, then at a minimum the 2b protein should be capable of binding viral RNA. To determine 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 C1C2T3 whether there was some interaction between TAV RNA 3 and the 2b protein, we performed an RNA-proteinbinding assay. The TAV 2b protein ( Figure 4B , lane 3) purified from E. coli ( Figure 4A ) was able to bind to the T7-transcribed (+) sense TAV RNA 3 ( Figure 4C , lanes 2-4) as well as to the T3-transcribed (À) sense TAV RNA 3 ( Figure 4D , lane 3). The binding ability increased with an increasing amount of the 2b protein ( Figure 4C , compare lanes 1-4). By contrast, comparable amounts of GST alone, which was expressed and purified under the same conditions as the TAV 2b protein ( Figure 4B , lane 2), did not bind to TAV RNA 3 (data not shown). This demonstrated that RNA-binding activity was not due to a co-purified, contaminating protein from E. coli. The significance of the more efficient binding of the 2b protein to (+) versus (À) sense RNA is unknown. The different binding abilities probably are due to different sequences and/or structures between RNA 3(+) and RNA 3(À). Since the 2b protein was designated a movement protein (19) and was shown to be required for the movement of pseudorecombinant viruses containing C1C2T3 (21), the 2b protein probably also binds to TAV RNA 3 in vivo. Another approach to determine the role of the 2b protein in the generation and/or selection of the recombinant RNAs, was to use a different 2b gene in the heterologous interspecific hybrid virus. To this end, we replaced the whole 2b gene of (the subgroup II strain) Q-CMV with that from a subgroup IA strain of CMV, WAII-CMV, generating the hybrid virus C1C2 W2B T3 (see C2 W2B in Figure 1 ). The WAII-CMV 2b gene and the Q-CMV 2b gene are only 53.4% identical at nucleotide level and the WAII-2b protein is 10 amino acids longer than the Q-2b protein. Co-inoculation of C2 W2B , C1 and T3 onto N. glutinosa plants followed by multiple passages at 1-2 weekly intervals showed that C1C2 W2B T3 induced symptoms as severe as those induced by C1C2 T2B T3, which were much more severe than those induced by either of the parental viruses (21) . Northern blot hybridization showed that RNA 3 derived from C1C2 W2B T3 also had recombined with a CMV sequence in the passage plants, but not in the inoculated plants ( Figure 2C and D, compare lanes 27 and 34, and lanes 29 and 36). This was supported by RT-PCR analysis (data not shown). These data strongly indicate that the 2b protein was required for recombination and/or selection or the recombinant RNA 3s, and that this process was not TAV 2b protein-specific. To determine the nature of the recombinant RNAs derived from C1C2 W2B T3, the recombinant RNA 3 was subjected to RT-PCR, cloning and sequence analysis. Surprisingly, the RNA 3 recombinants derived from C1C2 W2B T3 differed from those derived from C1C2 T2B T3 or C1C2 T2BC2a T3 in the following respects: First, the recombinant RNA 3 derived from C1C2 W2B T3 was 22 nt longer than that derived from C1C2 T2B T3 or C1C2 T2BC2a T3 (2392 versus 2370). Second, the 3 0 terminal NTR of the recombinant RNA 3 derived from C1C2 W2B T3 contained a Q-CMV RNA 2 NTR sequence, rather than the Q-CMV RNA 1 NTR sequence (compare Figure 3A right and B right) . Third, the crossover sites on RNA 2 and on RNA 3 were different in the recombinant RNA 3 derived from C1C2 W2B T3 versus either C1C2 T2B T3 or C1C2 T2BC2a T3 (compare Figure 3A and B). And fourth, the recombinant RNA 3 derived from C1C2 W2B T3 contained an imperfect repeat of 19 nt near the crossover site preceded by an A between the TAV RNA 3 sequence and the 3 0 CMV RNA sequence ( Figure 3B) . Moreover, the recombinant RNA 3 derived from C1C2 W2B T3 also was missing the second 163-nt repeat. These similarities on one hand and differences on the other suggested that the recombinant RNA 3s derived from C1C2 T2B T3 and C1C2 T2BC2a T3 versus C1C2 W2B T3 might be generated using different templates, but occurred via a similar mechanism. Taken together, the above results indicate that the 2b protein determined either the RNA species participating in the recombination and the crossover site, or selected for particular recombinant RNAs generated at random that have the above characteristics. The 2b protein and a 163-nt tandem repeat on RNA 3 both affect the rate of recombinant selection while a single nucleotide within the repeat determines the crossover site on RNA 3 Both the TAV 2b protein and WAII-CMV 2b protein were able to promote the recombination and/or selection of RNA 3 in the passage plants, but not in the inoculated plants. To determine when the recombinant RNA 3s could first be detected, a time-course experiment was done involving three viruses: C1C2 T2B T3, C1C2 T2BC2a T3 and C1C2 W2B T3. The appropriate plasmid combinations were inoculated onto 10 N. glutinosa plants and sap was then passaged onto healthy N. glutinosa plants for 10 passages, at 6-day intervals. The inoculated leaves of the passage plants were detached 3 days after inoculation, total RNAs were extracted and the RNAs were analysed by northern blot hybridization. These analyses showed that the recombinant RNAs were only detected from the fourth passage and onwards in plants infected with C1C2 T2B T3 or C1C2 T2BC2a T3, or from the sixth passage and onwards in plants infected with C1C2 W2B T3 (Supplementary Figure 1A, data not shown) . These results indicated that selection of the recombinant RNA 3s is a slow process and therefore it seems more likely that the 2b proteins were selecting for specific recombinants, rather than being involved in the generation of the recombinant RNA 3s. This is also consistent with the different subcellular localization of the replication proteins and the 2b protein (49) (50) (51) and the lack of detectable interaction of the replicase proteins with the 2b protein (52) . In addition, virus containing the TAV 2b protein showed selection of the recombinant RNA 3s two passages earlier than did virus containing the WAII-CMV 2b protein. Therefore, it seems likely that the source of the 2b protein also affected the rate of selection. To examine the role of the two 163-nt repeats in the recombination and/or selection, we deleted each repeat independently in C1C2 T2B T3. The resultant deletion mutants were designated C1C2 T2B T3 Á163(A) and C1C2 T2B T3 Á163(G) (Figure 1) , where the subscripts Á163(A) and Á163(G) represent the first 163-nt repeat deletion and the second 163-nt repeat deletion, respectively. Bioassay of C1C2 T2B T3 Á163(A) and C1C2 T2B T3 Á163(G) onto N. glutinosa plants showed that both deletion mutants induced similar symptoms to those induced by C1C2 T2B T3 (data not shown). Northern blot hybridization analysis showed that neither deletion affected the presence of recombinant RNA 3s in the multiple passage plants (Figure 2A, lanes 8 and 19 , data not shown). However, both deletion mutants yielded recombinant RNA 3s in both the inoculated plants and the passage plants (Figure 2A and B, compare lanes 4 and 15, and lanes 8 and 19, data not shown). Repeated inoculation and northern blot hybridization analysis gave the same results as earlier. This suggests that the two 163-nt repeats affected the rate of the recombination and/or selection of recombinants. To confirm the role of the two 163-nt repeats in affecting the rate of recombination and/or selection, we generated two other deletion mutants, C1C2 W2B T3 Á163(A) and C1C2 W2B T3 Á163(G) . Bioassay of C1C2 W2B T3 Á163(A) and C1C2 W2B T3 Á163(G) showed that these two deletion mutants also induced symptoms similar to those induced by C1C2 W2B T3 (data not shown). Northern blot hybridization analysis showed that the deletions also resulted in the generation of the recombinant RNA 3s in both the inoculated and passage plants ( Figure 2C and D, compare lanes 26 and 33, and lanes 28 and 35, data not shown). These results showed that the two 163-nt repeats indeed affected the rate of the recombination and/or selection. To determine the nature of the recombinant RNA 3s derived from the 163-nt repeat-deletion mutants, we again purified the recombinant RNA 3s of C1C2 T2B T3 Á163(A) , C1C2 W2B T3 Á163(A) , C1C2 T2B T3 Á163(G) and C1C2 W2B T3 Á163(G) , used RT-PCR to amplify the recombinant sequences, cloned the RT-PCR products and sequenced them. The sequences of the recombinant RNA 3 derived from C1C2 T2B T3 Á163(A) contained the 5 0 -proximal 2082 nt from TAV RNA 3 and the 3 0 proximal 310-nt 3 0 NTR of CMV RNA 1 ( Figure 5A ). The CMV RNA 1 NTR sequence was identical to that present in the recombinant RNA 3 derived from C1C2 T2B T3. However, the TAV RNA 3 sequence was 22 nt longer than that of the recombinant RNA 3 derived from C1C2 T2B T3 (compare Figures 3A and 5A) , suggesting that the first 163-nt repeat might have a role in determining the crossover site on one of the participating RNAs, despite not affecting the crossover site on the other participating RNA. By contrast, the sequence of the recombinant RNA 3 derived from C1C2 T2B T3 Á163(G) was exactly the same as that derived from C1C2 T2B T3 (compare Figures 3A and 5B) , indicating that the second 163-nt repeat had no role in determining the crossover sites on either RNA. The recombinant RNA 3 derived from C1C2 W2B T3 Á163(A) consisted of the 5 0 proximal 2082 nt from TAV RNA 3 and the 3 0 proximal 309-nt 3 0 NTR of CMV RNA 2 ( Figure 5C ). Although the 2082 nt sequence from TAV RNA 3 was exactly the same as that in the recombinant RNA 3 derived from C1C2 T2B T3 Á163(A) , it was 18 nt longer than the one seen in the recombinant RNA 3 derived from C1C2 W2B T3 (compare Figures 3B and 5C ). This result supports the conclusion that the first 163-nt repeat might indeed affect the left border of the crossover site. The CMV RNA 2 sequence present in the recombinant RNA 3 derived from C1C2 W2B T3 Á163(A) was 19 nt shorter than the CMV RNA 2 present in the recombinant RNA 3 derived from C1C2 W2B T3 (due to the absence of the additional A and the first 19-nt repeat shown in Figure 3B , but with an additional U residue, nucleotide 2757), suggesting that the first 163-nt repeat might also affect the right border of the crossover site. In the recombinant RNA 3 derived from C1C2 W2B T3 Á163(G) , the sequence was exactly the same as that of the recombinant RNA 3 derived from C1C2 W2B T3 (compare Figures 3B and 5D) , confirming that the second 163-nt repeat had no role in determining the position of the crossover sites on either RNA. Since only one nucleotide differed in the two 163-nt repeats (A versus G), the above results indicate that the position of the crossover sites is actually affected by this single nucleotide difference within each repeat. To determine whether the recombinant RNA 3s derived from C1C2 W2B T3 or C1C2 W2B T3 Á163(G) , both of which contained a 19-nt repeat sequence, might represent an intermediate, buffered extracts from plants infected by viruses containing these recombinant RNA 3s were passaged onto five healthy plants for more than 20 passages done at 1-week intervals. RT-PCR and sequence analyses of the recombinant RNA 3s from the passage plants showed no sequence difference from the original recombinant RNA 3s described above, indicating that the 19-nt repeat was still present in the same position as before (data not shown). These results indicate that recombinant RNA 3 derived from C1C2 W2B T3 or C1C2 W2B T3 Á163(G) containing the 19-nt repeat was not an intermediate, but the stable recombination end-product. To examine how soon after infection this 19-nt repeat was generated and whether there was an intermediate generated before this 19-nt repeat appeared, a timecourse experiment was done, in which two viruses, C1C2 T2B T3 Á163(G) and C1C2 W2B T3 Á163(G) , were inoculated onto N. clevelandii. Nicotiana clevelandii was chosen as the host because the symptoms are more readily seen on this host than on N. glutinosa (unpublished data) and C1C2 T2B T3 Á163(G) and C1C2 W2B T3 Á163(G) were chosen because both viruses showed more rapid appearance of recombinant RNA 3 (see above). Moreover, the recombinant progeny virus of C1C2 W2B T3 Á163(G) contained the 19-nt repeat, which could be used as a marker to determine whether there was an intermediate before this final product. After inoculation of each virus to 200 plants, leaves were detached from groups of five inoculated plants at 1-day intervals for up to 40 days. Northern blot hybridization of the total RNAs extracted from the detached leaves showed that recombinant RNA 3s derived from any of the above viruses did not accumulate to a detectable level before day 35; the earliest time when a recombinant RNA 3 could be detected by northern blot hybridization was at day 35 in the C1C2 T2B T3 Á163(G)infected plants (Supplementary Figure 1B) . The recombinant RNA 3 derived from C1C2 W2B T3 Á163(G) was not detected during the whole time-course of the experiment (data not shown). However, by using RT-PCR with the primer pair C3 0 /T3-1187, the recombinant RNA 3 derived from C1C2 T2B T3 Á163(G) was detected as early as day 22 (Figure 6 upper panel) , while the RNA 3 recombinant derived from C1C2 W2B T3 Á163(G) could be detected at day 35 (data not shown). These results also confirmed that the TAV 2b protein promoted the recombination/selection faster than the WAII-CMV 2b protein. To characterize the RT-PCR products further, Southern blot hybridization was done using CMV-specific and TAV-specific probes. All the RT-PCR products on the gel were the same size ( Figure 6 middle and lower panels) and the size was also the same as for the RT-PCR products obtained from the passage plants (data not shown). In addition, all the RT-PCR products were able to hybridize with both the Q-CMV-specific probe and the TAV-specific probe ( Figure 6 middle and lower panels). Sequencing of these RT-PCR products showed that the RT-PCR products from the C1C2 T2B T3 Á163(G) -inoculated plants contained 874 nt derived from TAV RNA 3 (up to nucleotide 2060 of TAV RNA 3) and 310 nt from CMV RNA 1. This sequence was exactly the same as that derived from the same virus in the passage plants and also the same sequence as that derived from C1C2 T2B T3, either in the inoculated plants or in the passage plants ( Figures 3A and 5B ; data not shown). All of the RT-PCR products derived from C1C2 W2B T3 Á163(G) also were exactly the same in sequence, having 878 nt from TAV RNA 3 and 326 nt from CMV RNA 2 including the 19-nt repeat ( Figure 5D ; data not shown). The sequence of the RT-PCR products also were the same as those from the passage plants, or those derived from C1C2 W2B T3 in both the inoculated plants and passage plants ( Figure 3B ; data not shown). These results indicate that the recombinant RNA 3s accumulating in the infected plants did not contain a detectable intermediate, and that the crossover sites on both donor and acceptor RNAs occurred at the same position. The conserved 23-nt sequence and the 3' NTR of RNA 3 are critical for infection The conserved 23-nt sequence at the beginning of the 3 0 NTR has been implicated in the recombination and nucleotides 1-20 of the 23-nt sequence also have been suggested to be an internal subgenomic promoter site (33, 38) . This sequence also was always near the crossover sites in the recombinant RNA 3s observed here, and the 19-nt imperfect repeat also comes from this sequence. Therefore, we examined the requirement for this sequence in infection. Three deletion mutants were constructed: C1 Á23 , C2 T2BÁ23 and T3 Á163(A)Á23 , where the conserved 23-nt sequence was deleted from each of C1, C2 T2B and T3 Á163(A) , respectively. Co-inoculation of each deletion mutant in the combinations C1 Á23 with C2 T2BÁ23 and T3 Á163(A)Á23 , C1 Á23 with C2 T2B and T3 Á163(A) , C1 with C2 T2BÁ23 and T3 Á163(A) , or C1 with C2 T2B and T3 Á163(A)Á23 , onto N. glutinosa and N. clevelandii plants, showed that none of these combinations was able to infect the plants (Table 1) . However, when either C1 Á23 was combined with C2 and C3, or T3 Á163(A)Á23 was combined with T1 and T2, the mixtures were infectious (Table 1) , indicating that the presence of this 23-nt sequence is essential for infection of the pseudorecombinant viruses, but not the parental viruses. Since the pseudorecombinant viruses containing the 23-nt sequence were viable without detectable recombination for some time (Figure 2) , these results indicate that the 23-nt sequence is not essential for replication of the pseudorecombinant virus, but may be essential for maintaining infection of the pseudorecombinant viruses through the plants, just as the 2b protein was required for this task, but not for the maintenance of infection of the parental viruses (19, 21) . The above results showed that the CMV sequence in the recombinant RNA 3 was similar or identical to the CMV RNA 5 sequence. Correspondingly, the TAV RNA 3 sequence in the same recombinant RNA 3s lacked the second 163-nt repeat and the rest of the 3 0 NTR (which together corresponds to the TAV RNA 5 sequence). To determine whether CMV RNA 5 itself or the 3 0 NTR of CMV RNAs 1 and/or 2 could be used as a template for the synthesis of the recombinant RNA 3, we prepared the appropriate templates that would be expressed transiently in inoculated plants and assessed these for the presence of recombinant virus. In the templates, we deleted 326 nt from the 3 0 NTR of T3 to make T3 , and deleted the first 163-nt repeat plus 141 nt from the 3 0 NTR of T3 to make T3 Á163(A)Á141 (Figure 1 ). These two deletions had the same length of the TAV RNA 3 sequence as in the recombinant RNAs. In addition, we also made four RNA 5-like cDNA clones, C1 3081-3390 , C2 W2B2757-3065 , T3 1902-2386 and T3 2065-2386 . The former two were based on the CMV RNA 1 and C2 W2B sequences in the recombinant viruses, respectively, while the latter two were based on the sequence of TAV RNAs 3B and 5, respectively (39) . All of these clones were placed into the pCass1 vector, under transcriptional control of the CaMV 35S RNA promoter and terminator sequences. The infectivity results of the various plasmid mixtures tested on N. glutinosa and N. clevelandii plants are shown in Table 1 (Table 1 ). These results indicate that the sequences containing the 3 0 NTR are essential for the viral infection and could not be separated functionally from the rest of the TAV RNA 3 sequence. In addition, the results also indicate that the recombination observed in other experiments did not occur during the synthesis of (À) sense RNA from (+) sense templates, since there was no repair of the 3 0 NTR from synthesis initiated on either genomic RNAs 1 or 2, or subgenomic RNAs 3B or 5 (Table 1) and template switching to TAV RNA 3 lacking the 3 0 NTR. Such repair recombination is common among tombusviruses (8, 53) and occurs with the related brome mosaic virus (54) (55) (56) . Therefore, an alternative explanation is that the recombination occurred during (+) sense RNA synthesis on the (À) sense RNA templates, with recombination occurring as a result of template switching of the viral replicase from the donor (TAV RNA 3) strand to the acceptor (CMV RNA) strand. Based on putative secondary structures for the 3 0 NTRs of CMV and TAV, models have been proposed to account for the generation of some of the other recombination products previously observed to occur between CMV and TAV (33, 38) . These models are based on template switching of the viral replicase on a stem-loop structure present in the (À) RNA 3 (L1/L2 in Figure 7 ) that may be the promoter for the synthesis of subgenomic RNAs 3B and 5 (33, 38) . Similar structures can be used to explain the various recombinant products obtained here and the effect of the 163-nt repeats and the single G to A substitution on the recombination event. This model assumes that intramolecular recombination is a more frequent event than intermolecular 1-2060 T3 1902-2386 À À C1C2 T2B T3 1-2060 T3 2065-2386 À À '+' indicates infection, while 'À' indicates no infection. The nature of the plasmid constructs is described in Figure 1 . recombination. Therefore, generation of TAV RNA 3 progeny with deletion of one of the 163-nt repeats is more likely to occur than recombination between TAV RNA 3 and another viral genomic RNA. That is, if one repeat is missing, then only intermolecular recombination can occur, which may explain why the recombinants were seen to accumulate more rapidly in the plants directly inoculated with virus containing only one repeat ( Figure 2 ). On the other hand, loss of a repeat was not a prerequisite for intermolecular recombination, since intramolecular recombination would involve loss of the first repeat [163(A)], while the recombinants formed from the full-length T3 all retained the first repeat. Intramolecular recombination would be expected to occur during (+) TAV RNA 3 synthesis, if the polymerase jumped from the beginning of the first repeat in stem-loop L2 ( Figure 7A ) to the beginning of the second repeat in stemloop L1 ( Figure 7B ). These sequences also contain the putative RNA 5 subgenomic promoter sequence (33, 38) that would facilitate this recombination and those described previously that involved this 'hot spot' for intermolecular recombination between TAV and CMV (33, 38) . The presence of another stem-loop structure (K1/K2 in Figure 7 ) immediately adjacent to the putative subgenomic promoter may slow the polymerase down sufficiently to stutter and in some cases switch templates. In the case of the first TAV-CMV RNA 3 recombinant described (30) , this occurred just before another stem-loop structure in the first repeat (J2 in Figure 7A ; see arrow), but in that instance, the polymerase jumped to the putative RNA 5 subgenomic promoter of (À) CMV RNA 2, leading to an initial recombinant that had a duplication of the two, putative subgenomic RNA promoters, which then led to a second recombination event, with the intervening TAV RNA 3 sequences deleted, giving the final, stable recombinant RNA 3 described (30) . The full-length TAV RNA 3, as well as the T3 Á163(G) variant gave rise to similar recombination products, with the exact switchover position on the donor RNA of the recombinant determined by the 2b gene present (T2B versus W2B), while the switchover position on the receptor RNA is the beginning of the RNA 5 sequence in either CMV RNAs 1 or 2, again determined by the specific 2b gene present. In the case of C1C2 T2B T3 and C1C2 T2B T3 Á163(G) , the recombination occurred immediately after nucleotide 2060. For C1C2 T2B T3, switchover would have had to occur at the base of stem-loop L1 ( Figure 7B ), while for C1C2 T2B T3 Á163(G) , switchover would have had to occur within stem-loop G ( Figure 7B and C), since the sequences comprising stem-loop L1 (plus stem-loops K1 to H1) would not exist in C1C2 T2B T3 Á163(G) . Since stem-loops L1 and G are quite different in most of their sequence and structure, as well as stability (ÁG = -1.3 kcal/mol for stem-loop G and À6.9 kcal/mol for stem-loop L1), it seems likely that the common structures preceding these stem-loops (J2-I2-H2 versus J1-I1-H1; ÁG = À3.8, À4.8 and À2.9 kcal/mol, for each stem-loop structure) may have slowed the polymerase sufficiently to allow it to pause at nucleotide 2060, in either stem-loop L1 or G ( Figure 7B ). In the case of C1C2W 2B T3 Á163(G) , switchover occurred following nucleotide 2064 in stem-loop G, but did not immediately proceed to the (À) CMV RNA 2 as the receptor RNA. Rather, the polymerase jumped back to the subgenomic promoter stem-loop (L2) and copied this through the A residue in the gap between stem-loops L2 and K2 ( Figure 7A) , before pausing and jumping to the same subgenomic promoter sequence on (À) CMV RNA 2. The small difference in ÁG between the 6 bp stem-loop K2 in (À) TAV RNA 3 (ÁG = À2.7 kcal/mol) and a weaker 3 bp stem-loop (K 0 ) structure adjacent to the subgenomic promoter in (À) CMV RNA 2 (ÁG = À2.1 kcal/mol) may be sufficient to prevent the polymerase from pausing at the base of stem-loop K 0 in (À) CMV RNA 2. This explains the duplication of the 19 nt seen in Figure 5D , following the A residue (from copying the first nucleotide, U-1902 of the RNA 5/3B promoter). In the case of C1C2 W2B T3, the same recombinant progeny RNA 3 is seen, but it must have been generated from a different starting point, since the second repeat is present in the template but missing in the recombinant progeny RNA 3. Here, template switching occurred either immediately after nucleotide 2084 (the A residue in the gap between stem-loops L1 and K1), again leading to a partially duplicated subgenomic promoter with the 19 nt repeat, or it occurred at nucleotide 2064, in the lower region of stem-loop L1 ( Figure 7B ), and then proceeded as described above for C1C2 W2B T3 Á163(G) . In contrast to C1C2 T2B T3 Á163(G) , C1C2 T2B T3 Á163(A) and C1C2W 2B T3 Á163(A) underwent recombination 22 nt further along the TAV RNA 3 molecule after nucleotide 2245, adjacent to stem-loop E (in Figure 7C) , regardless of the nature of the 2b protein. Therefore, while the nature of the receptor RNA and the apparently slow selection of the recombinants (Figures 2 and 6) were determined by the 2b gene, the generation of recombinants at position 2245 was determined by the absence of the first repeat. Since the first and second repeats differ by only one nucleotide (underlined blue letters in Figure 7A and B, at positions 1966 and 2129), we propose that there must be a difference in structure between the virus containing only one repeat, when this is the second repeat versus the first repeat, and that this difference in structure affects pausing by the polymerase, leading to switchover to another template. This difference could be the presence (or absence) of a weak pseudoknot structure formed between the three A residues in the loop of stem-loop G and the three U residues in the non-base-paired region between stem-loops K2 and J2 (see blue letters in Figures 7) . This pseudoknot would not be stable in T3 Á163(A) , since it would only contain two A:U base pairs. This weak pseudoknot structure may be sufficient to affect pausing by the polymerase in the T3 Á163(G) variant, if it can re-form after the polymerase has gone through it the first time and then transcribed through the various, following stem-loop structures (stem-loops J, I and H in Figure 7 ), leading to template switching in stem-loop G. However, as this pseudoknot is absent in T3 Á163(A) , the polymerase would continue through stem-loop G until it pauses at the base of stem-loop E ( Figure 7C ; ÁG = À6.8 kcal/mol), before switching to the receptor RNA. It would seem likely that the polymerase is as equally able to use RNA 1 as RNA 2 for the receptor RNA, since the polymerase is essentially the same in all cases, and the presence or absence of the 2a protein C-terminal sequences (where the recombinant polymerases showed some differences in sequence) did not affect the nature of the recombinant RNAs. However, since only one type of recombinant progeny was observed in each case, this demonstrates that there must have been selection for a particular progeny RNA, which depended on the source of the 2b gene and its encoded protein. How could the nature of the 2b protein affect the selection of recombinant RNAs? It is conceivable that the selection of recombinant RNAs is a manifestation of the RNA silencing suppression-related functions of the 2b proteins, since some of these also vary between strains of CMV (57, 58) . However, based on the spatial separation of the 2b protein from the replication machinery within the cell (49) (50) (51) , it is difficult to envision how different 2b proteins provide the selection observed through their RNA silencing suppressor functions. In the case of the tombusvirus cucumber necrosis virus, inhibition of expression of the 20-kDa protein now known to be the RNA silencing suppressor of this virus, led to the rapid appearance of defective-interfering RNAs (59) . Thus, in that, the absence of the silencing suppressor led to enhanced generation or selection of recombinant RNAs, while in the case of CMV, the source of the 2b protein determined the nature of the recombinant RNAs selected and pseudorecombinant viruses could not be selected in its absence (21) . Moreover, in the case of influenza virus, where the NS1 gene encodes an RNA silencing suppressor (60) , mutation in a different gene (NS2) led to the rapid appearance of defective interfering particles (61) . At the time it was thought that the influenza NS2 protein might be involved in replication. In fact, the NS2 protein is involved in selecting ribonucleoprotein complexes for transport from the nucleus to the cytoplasm and is now called the nuclear export protein (62) . Based on previous observations that the 2b protein is also a movement protein (19, 63) , that the presence of the 2b protein was necessary for the cell-to-cell movement of pseudorecombinants viruses (21) , and that the 2b protein is a viral RNA-binding protein (Figure 4) , it seems likely that the 2b binds to and selects specific recombinant viral RNAs for cell-to-cell movement. The different forms of recombinant RNAs generated may have different affinities for the various 2b proteins, leading to selection of one recombinant form over another by a particular 2b protein, which leads eventually to selection of the recombinant RNA over the parental RNA. Since selection to reach detectable levels was only observed after several weeks, the recombinant viral RNAs do not appear to be under strong selection pressure, although there is clearly a gradual selection for the recombinant RNAs over the parental RNA 3. This was seen in the slow appearance of the recombinant RNAs in both sets of time course experiments ( Figure 6 and Supplementary Figure 1 ). The rate of selection could be tested further by direct competition experiments. The Q-CMV 2b protein did not select for any recombinants, while the TAV and WAII-CMV 2b proteins selected for different recombinants, possibly reflecting differences in their RNA-binding properties. It would also be interesting to determine whether various 2b mutants would show differences in their selection of different recombinants.

Epigrass: a tool to study disease spread in complex networks

BACKGROUND: The construction of complex spatial simulation models such as those used in network epidemiology, is a daunting task due to the large amount of data involved in their parameterization. Such data, which frequently resides on large geo-referenced databases, has to be processed and assigned to the various components of the model. All this just to construct the model, then it still has to be simulated and analyzed under different epidemiological scenarios. This workflow can only be achieved efficiently by computational tools that can automate most, if not all, these time-consuming tasks. In this paper, we present a simulation software, Epigrass, aimed to help designing and simulating network-epidemic models with any kind of node behavior. RESULTS: A Network epidemiological model representing the spread of a directly transmitted disease through a bus-transportation network connecting mid-size cities in Brazil. Results show that the topological context of the starting point of the epidemic is of great importance from both control and preventive perspectives. CONCLUSION: Epigrass is shown to facilitate greatly the construction, simulation and analysis of complex network models. The output of model results in standard GIS file formats facilitate the post-processing and analysis of results by means of sophisticated GIS software.

Epidemic models describe the spread of infectious diseases in populations. More and more, these models are being used for predicting, understanding and developing control strategies. To be used in specific contexts, modeling approaches have shifted from "strategic models" (where a caricature of real processes is modeled in order to emphasize first principles) to "tactical models" (detailed representations of real situations). Tactical models are useful for cost-benefit and scenario analyses. Good examples are the foot-and-mouth epidemic models for UK, triggered by the need of a response to the 2001 epidemic [1, 2] and the simulation of pandemic flu in differ-ent scenarios helping authorities to choose among alternative intervention strategies [3, 4] . In realistic epidemic models, a key issue to consider is the representation of the contact process through which a disease is spread, and network models have arisen as good candidates [5] . This has led to the development of "network epidemic models". Network is a flexible concept that can be used to describe, for example, a collection of individuals linked by sexual partnerships [6] , a collection of families linked by sharing workplaces/schools [7] , a collection of cities linked by air routes [8] . Any of these scales may be relevant to the study and control of disease spread [9] . Networks are made of nodes and their connections. One may classify network epidemic models according to node behavior. One example would be a classification based on the states assumed by the nodes: networks with discretestate nodes have nodes characterized by a discrete variable representing its epidemiological status (for example, susceptible, infected, recovered). The state of a node changes in response to the state of neighbor nodes, as defined by the network topology and a set of transmission rules. Networks with continuous-state nodes, on the other hand, have node' state described by a quantitative variable (number of susceptibles, density of infected individuals, for example), modelled as a function of the history of the node and its neighbors. The importance of the concept of neighborhood on any kind of network epidemic model stems from its large overlap with the concept of transmission. In network epidemic models, transmission either defines or is defined/constrained by the neighborhood structure. In the latter case, a neighborhood structure is given a priori which will influence transmissibility between nodes. The construction of complex simulation models such as those used in network epidemic models, is a daunting task due to the large amount of data involved in their parameterization. Such data frequently resides on large geo-referenced databases. This data has to be processed and assigned to the various components of the model. All this just to construct the model, then it still has to be simulated, analyzed under different epidemiological scenarios. This workflow can only be achieved efficiently by computational tools that can automate most if not all of these time-consuming tasks. In this paper, we present a simulation software, Epigrass, aimed to help designing and simulating network-epidemic models with any kind of node behavior. Without such a tool, implementing network epidemic models is not a simple task, requiring a reasonably good knowledge of programming. We expect that this software will stimulate the use and development of networks models for epidemiological purposes. The paper is organized as following: first we describe the software and how it is organized with a brief overview of its functionality. Then we demonstrate its use with an example. The example simulates the spread of a directly transmitted infectious disease in Brazil through its transportation network. The velocity of spread of new diseases in a network of susceptible populations depends on their spatial distribution, size, susceptibility and patterns of contact. In a spatial scale, climate and environment may also impact the dynamics of geographical spread as it introduces temporal and spatial heterogeneity. Understanding and predicting the direction and velocity of an invasion wave is key for emergency preparedness. Epigrass is a platform for network epidemiological simulation and analysis. It enables researchers to perform comprehensive spatio-temporal simulations incorporating epidemiological data and models for disease transmission and control in order to create complex scenario analyses. Epigrass is designed towards facilitating the construction and simulation of large scale metapopulational models. Each component population of such a metapopulational model is assumed to be connected through a contact network which determines migration flows between populations. This connectivity model can be easily adapted to represent any type of adjacency structure. Epigrass is entirely written in the Python language, which contributes greatly to the flexibility of the whole system due to the dynamical nature of the language. The geo-referenced networks over which epidemiological processes take place can be very straightforwardly represented in a object-oriented framework. Consequently, the nodes and edges of the geographical networks are objects with their own attributes and methods (figure 1). Once the archetypal node and edge objects are defined with appropriate attributes and methods, then a code representation of the real system can be constructed, where nodes (representing people or localities) and contact routes are instances of node and edge objects, respectively. The whole network is also an object with its own set of attributes and methods. In fact, Epigrass also allows for multiple edge sets in order to represent multiple contact networks in a single model. Figure 1 Architecture of an Epigrass simulation model. A simulation object contains the whole model and all other objects representing the graph, sites and edges. Site object contaim model objects, which can be one of the built-in epidemiological models or a custom model written by the user. These features leads to a compact and hierarchical computational model consisting of a network object containing a variable number of node and edge objects. It also does not pose limitations to encapsulation, potentially allowing for networks within networks, if desirable. This representation can also be easily distributed over a computational grid or cluster, if the dependency structure of the whole model does not prevent it (this feature is currently being implemented and will be available on a future release of Epigrass). For the end-user, this hierarchical, object-oriented representation is not an obstacle since it reflects the natural structure of the real system. Even after the model is converted into a code object, all of its component objects remain accessible to one another, facilitating the exchange of information between all levels of the model, a feature the user can easily include in his/her custom models. Nodes and edges are dynamical objects in the sense that they can be modified at runtime altering their behavior in response to user defined events. In Epigrass it is very easy to simulate any dynamical system embedded in a network. However, it was designed with epidemiological models in mind. This goal led to the inclusion of a collection of built-in epidemic models which can be readily used for the intra-node dynamics (SIR model family). Epigrass users are not limited to basing their simulations on the built-in models. User-defined models can be developed in just a few lines of Python code. All simulations in Epigrass are done in discrete-time. However, custom models may implement finer dynamics within each time step, by implementing ODE models at the nodes, for instance. The Epigrass system is driven by a graphical user interface(GUI), which handles several input files required for model definition and manages the simulation and output generation (figure 2). At the core of the system lies the simulator. It parses the model specification files, contained in a text file (.epg file), and builds the network from site and edge description files (comma separated values text files, CSV). The simulator then builds a code representation of the entire model, simulates it, and stores the results in the database or in a couple of CSV files. This output will contain the full time series of the variables in the model. Additionally, a map layer (in shapefile and KML format) is also generated with summary statitics for the model (figure 3). The results of an Epigrass simulation can be visualized in different ways. A map with an animation of the resulting timeseries is available directly through the GUI (figure 4). Other types of static visualizations can be generated through GIS software from the shapefiles generated. The KML file can also be viewed in Google Earth™ or Google Maps™ (figure 5). Epigrass also includes a report generator module which is controlled through a parameter in the ".epg" file. Epigrass is capable of generating PDF reports with summary statistics from the simulation. This module requires a LATEX installation to work. Reports are most useful for general verification of expected model behavior and network structure. However, the LATEX source files generated Workflow for a typical Epigrass simulation Figure 3 Workflow for a typical Epigrass simulation. This diagram shows all inputs and outputs typical of an Epigrass simulation session. Epigrass graphical user interface Figure 2 Epigrass graphical user interface. by the module may serve as templates that the user can edit to generate a more complete document. Building a model in Epigrass is very simple, especially if the user chooses to use one of the built-in models. Epigrass includes 20 different epidemic models ready to be used (See manual for built-in models description). To run a network epidemic model in Epigrass, the user is required to provide three separate text files (Optionally, also a shapefile with the map layer): 1. Node-specification file: This file can be edited on a spreadsheet and saved as a csv file. Each row is a node and the columns are variables describing the node. 2. Edge-specification file: This is also a spreadsheet-like file with an edge per row. Columns contain flow variables. 3. Model-specification file: Also referred to as the ".epg" file. This file specifies the epidemiological model to be run at the nodes, its parameters, flow model for the edges, and general parameters of the simulation. The ".epg" file is normally modified from templates included with Epigrass. Nodes and edges files on the other hand, have to be built from scratch for every new network. Details of how to construct these files, as well as examples, can be found in the documentation accompanying the software, which is available at at the project's website [10] In the example application, the spread of a respiratory disease through a network of cities connected by bus transportation routes is analyzed. The epidemiological scenario is one of the invasion of a new influenza-like virus. One may want to simulate the spread of this disease through the country by the transportation network to evaluate alternative intervention strategies (e.g. different vaccination strategies). In this problem, a network can be defined as a set of nodes and links where nodes represent cities and links represents transportation routes. Some examples of this kind of model are available in the literature [8, 11] . One possible objective of this model is to understand how the spread of such a disease may be affected by the pointof-entry of the disease in the network. To that end, we may look at variables such as the speed of the epidemic, number of cases after a fixed amount of time, the distribution of cases in time and the path taken by the spread. The example network was built from 76 of largest cities of Brazil (>= 100 k habs). The bus routes between those cities formed the connections between the nodes of the networks. The number of edges in the network, derived from Epigrass output visualized on Google-Earth Figure 5 Epigrass output visualized on Google-Earth. Figure 4 Epigrass animation output. Sites are color coded (from red to blue) according to infection times. Bright red is the seed site (on the NE). the bus routes, is 850. These bus routes are registered with the National Agency of Terrestrial Transportation (ANTT) which provided the data used to parameterize the edges of the network. The epidemiological model used consisted of a metapopulation system with a discrete-time SEIR model (Eq. 1). For each city, S t is the number of susceptibles in the city at time t, E t is the number of infected but not yet infectious individuals, I t is the number of infectious individuals resident in the locality, N is the population residing in the locality (assumed constant throughout the simulation), and n t is the number of individuals visiting the locality, Θ t is the number of visitors who are infectious. The parameters used were taken from Lipsitch et al. (2003) [12] to represent a disease like SARS with an estimated basic reproduction number (R 0 ) of 2.2 to 3.6 ( Table 1) . To simulate the spread of infection between cities, we used the concept of a "forest fire" model [13] . An infected individual, traveling to another city, acts as a spark that may trigger an epidemic in the new locality. This approach is based on the assumption that individuals commute between localities and contribute temporarily to the number of infected in the new locality, but not to its demography. Implications of this approach are discussed in Grenfell et al (2001) [13] . The number of individuals arriving in a city (n t ) is based on annual total number of passengers arriving trough all bus routes leading to that city as provided by the ANTT (Brazilian National Agency for Terrestrial Transportation). The annual number of passengers is used to derive an average daily number of passengers simply by dividing it by 365. Stochasticity is introduced in the model at two points: the number of new cases is draw from a Poisson distribution with intensity and the number of infected individuals visiting i is modelled as binomial process: where n is the total number of passengers arriving from a given neighboring city; I k, t and N k are the current number of infectious individuals and the total population size of city k, respectively. δ is the delay associated with the duration of each bus trip. The delay δ was calculated as the number of days (rounded down) that a bus, traveling at an average speed of 60 km/h, would take to complete a given trip. The lengths in kilometers of all bus routes were also obtained from the ANTT. Vaccination campaigns in specific (or all) cities can be easily attained in Epigrass, with individual coverages for each campaign on each city. We use this feature to explore Vaccination scenarios in this model (figures 6 and 7). The files with this model's definition(the sites, edges and ".epg" files) are available as part of the Additional files 1, 2 and 3 for this article. To determine the importance of the point of entry in the outcome of the epidemic, the model was run 500 times, randomizing the point of entry of the virus. The seeding site was chosen with a probability proportional to the log 10 of their population size. These replicates were run using Epigrass' built-in support for repeated runs with the option of randomizing seeding site. For every simulation, statistics about each site such as the time it got infected and time series of incidence were saved. The time required for the epidemic to infect 50% of the cities was chosen as a global index to network susceptibility to invasion. To compare the relative exposure of cities to disease invasion, we also calculated the inverse of time , for all k neighbors elapsed from the beginning of the epidemic until the city registered its first indigenous case as a local measure of exposure. Except for population size, all other epidemiological parameters were the same for all cities, that is, disease transmissibility and recovery rate. Some positional features of each node were also derived: Centrality, which is is a measure derived from the average distance of a given site to every other site in the network; Betweeness, which is the number of times a node figures in the the shortest path between any other pair of nodes; and Degree, which is the number of edges connected to a node. In order to analyze the path of the epidemic spread, we also recorded which cities provided the infectious cases which were responsible for the infection of each other city. If more than one source of infection exists, Epigrass selects the city which contributed with the largest number Cost in vaccines applied vs. benefit in cases avoided, for a simulated epidemic starting at the highest degree city (São Paulo) Figure 6 Cost in vaccines applied vs. benefit in cases avoided, for a simulated epidemic starting at the highest degree city (São Paulo). Cost in vaccines applied vs. benefit in cases avoided, for a simulated epidemic starting at a relatively low degree city(Salvador) Figure 7 Cost in vaccines applied vs. benefit in cases avoided, for a simulated epidemic starting at a relatively low degree city(Salvador). of infectious individuals at that time-step, as the most likely infector. At the end of the simulation Epigrass generates a file with the dispersion tree in graphML format, which can be read by a variety of graph plotting programs to generate the graphic seen on figure 8. The computational cost of running a single time step in an epigrass model, is mainly determined by the cost of calculating the epidemiological models on each site(node). Therefore, time required to run models based on larger networks should scale linearly with the size of the network (order of the graph), for simulations of the same duration. The model presented here, took 2.6 seconds for a 100 days run, on a 2.1 GHz cpu. A somewhat larger model with 343 sites and 8735 edges took 28 seconds for a 100 days simulation. Very large networks may be limited by the ammount of RAM available. The authors are working on adapting Epigrass to distribute processing among multiple cpus(in SMP systems), or multiple computers in a cluster system. The memory demands can also be addressed by keeping the simulation objects on an objectoriented database during the simulation. Steps in this direction are also being taken by the development team. The model presented here served maily the purpose of illustrating the capabilities of Epigrass for simulating and analyzing reasonably complex epidemic scenarios. It should not be taken as a careful and complete analysis of a real epidemic. Despite that, some features of the simulated epidemic are worth discussing. For example: The spread speed of the epidemic, measured as the time taken to infect 50% of the cities, was found to be influenced by the centrality and degree of the entry node (figures 9 and 10). The dispersion tree corresponding to the epidemic, is greatly influenced by the degree of the point of entry of Spread of the epidemic starting at the city of Salvador, a city with relatively small degree (that is, small number of neigh-bors) Figure 8 Spread of the epidemic starting at the city of Salvador, a city with relatively small degree (that is, small number of neighbors). The number next to the boxes indicated the day when each city developed its first indigenous case. Effect of degree(a) and betweeness(b) of entry node to the speed of the epidemic Figure 9 Effect of degree(a) and betweeness(b) of entry node to the speed of the epidemic. Effect of betweeness of entry node on the speed of the epi-demic Figure 10 Effect of betweeness of entry node on the speed of the epidemic. the disease in the network. Figure 8 shows the tree for the dispersion from the city of Salvador. Vaccination strategies must take into consideration network topology. Figures 6 and 7 show cost benefit plots for three vaccination strategies investigated: Uniform vaccination, top-3 degree sites only and top-10 degree sites only. Vaccination of higher order sites offer cost/benefit advantages only in scenarios where the disease enter the network through one of these sites. Epigrass facilitates greatly the simulation and analysis of complex network models. The output of model results in standard GIS file formats facilitates the post-processing and analysis of results by means of sophisticated GIS software. The non-trivial task of specifying the network over which the model will be run, is left to the user. But epigrass allows this structure to be provided as a simple list of sites and edges on text files, which can easily be contructed by the user using a spreadsheet, with no need for special software tools. Besides invasion, network epidemiological models can also be used to understand patterns of geographical spread of endemic diseases [14] [15] [16] [17] . Many infectious diseases can only be maintained in a endemic state in cities with population size above a threshold, or under appropriate environmental conditions(climate, availability of a reservoir, vectors, etc). The variables and the magnitudes associated with endemicity threshold depends on the natural history of the disease [18] . Theses magnitudes may vary from place to place as it depends on the contact structure of the individuals. Predicting which cities are sources for the endemicity and understanding the path of recurrent traveling waves may help us to design optimal surveillance and control strategies.

Autoimmune Cholangitis in the SJL/J Mouse is Antigen Non-specific

Primary biliary cirrhosis (PBC) is an autoimmune disease characterized by intrahepatic bile duct destruction and the production of anti-mitochondrial antibodies (AMA). The absence of an animal model has been a striking impedance in defining the molecular basis of disease. Previous work has suggested that SJL/J mice immunize with the pyruvate dehydrogenase complex (PDC-E2), the major mitochondrial autoantigen of PBC, leads to the development of lymphoid cell infiltration in portal tracts and a model system coined autoimmune cholangitis. We hypothesized that this pathology would be augmented if immunization occurred in the presence of IFN-γ injections. Accordingly, SJL/J mice were immunized with PDC-E2 and, for purpose of control, α-casein. Subgroups of mice were also treated with exogenous IFN-γ. As expected, mice immunized with PDC-E2, with or without IFN-γ, developed high titer AMAs. In contrast, mice immunized with α-casein, develop antinuclear antibodies. More importantly, the livers from mice immunized with PDC-E2 and/or those immunized with α-casein all displayed lymphoid cell infiltration to the portal tracts, irrespective of bile duct size. Indeed, there was no significant difference between the experimental and the control groups by histologic analysis. Thus, autoimmune cholangitis in these mice is antigen non-specific.

Primary biliary cirrhosis (PBC) is a chronic inflammatory, organ specific autoimmune disease of the liver characterized by high titer anti-mitochondrial antibodies (AMA) and intrahepatic bile duct destruction (Nakanuma and Ohta, 1979; Scheuer, 1994; Kaplan, 1996) . The etiology and pathogenesis of PBC are unknown. High titer AMA react with a series of highly conserved intramitochondrial proteins including the E2 subunit of the pyruvate dehydrogenase complex (PDC-E2), the major autoantigen of PBC. Whether liver injury is caused by AMA (Malmborg et al., 1998; Feehally and Allen, 1999) , PDC-E2 specific T cells (Van de Water et al., 1997; Shimoda et al., 1998; Inada et al., 2000) , NK cells (Tsuneyama et al., 1998; Iwata et al., 2000) or pathological alterations of the bile ducts themselves (Yasoshima et al., 1995; Kimura et al., 1996; Leon et al., 1997) are all being investigated. However, it is important to bear in mind that PBC is an autoimmune disease that most likely develops as a multi-hit disease, not just one insult. It has been theorized that PBC, in a genetically susceptible host, may be initiated by a molecular mimic (Coppel and Gershwin, 1995; Shimoda et al., 2000) or a xenobiotic (Uibo and Salupere, 1999; Long et al., 2001) . However, to this end, no definitive causative agent has, as yet, been identified. The first reported animal model for PBC was observed in an inbred rabbit strain with spontaneous nonsuppurative cholangitis (Tison et al., 1982) . However, the specificity and repertoire of AMA in these rabbits failed to mimic human disease. More recently, PBC epithelial cells were injected intravenously into neonatally thymectomized A/J mice on the premise that altered epithelial cells may initiate disease (Kobashi et al., 1994; Masanaga et al., 1998) . Also, injection of C57BL/6 mice with purified PDH in LPS or injection of recombinant polypeptides of PDC-E2 (Ide et al., 1996; Kimura et al., 1996) have been attempted. However, although AMAs were induced in all of these proposed models, pathological changes in the liver similar to human PBC were not induced. In contrast, a GVHD model for PBC has also been developed which induces histological pathology but serological responses are not specific for mitochondrial antigens (Tanaka et al., 1999; Zeniya, 2000) . As a whole, all animal models developed to date fail to accurately mimic human PBC by the induction of both AMA and intrahepatic liver damage. With these observations in mind, Jones et al. (2000) reported the appearance of autoimmune cholangitis in SJL/J mice following the immunization of bovine PDC-E2, emulsified in CFA. The SJL/J mouse was chosen because it has been successfully used as models of other autoimmune disease, including experimental allergic encephalomyelitis, systemic lupus erythematosus and mercury induced immunopathology (Vidal et al., 1994; Abedi-Valugerdi and Moller, 2000; Winer et al., 2001) . Further, T cells isolated from affected mice displayed a mixed Th1/Th2 response when stimulated with autoantigen in vitro (Jones et al., 1999) . IFN-g is a Th1 cytokine involved in the generation of antigen-specific T cells and has already been shown to be involved in induction of experimental myasthenia gravis (Wang et al., 2000) . Therefore, we hypothesized that injection of exogenous IFN-g would enhance the development of a PBC-like disease in SJL/J mice immunized with bovine PDC-E2. However, we report herein that while SJL/J mice do demonstrate evidence of significant cholangitis, that such cholangitis is antigen non-specific regardless of IFN-g injection and thus not reflective of PBC. A mitochondrial fraction isolated from bovine heart tissue was prepared as described previously (Fregeau et al., 1990) ; native PDC-E2 was thence purified (Stanley and Perham, 1980) . As a negative control, a-casein, was purchased from Sigma Inc. (St Louis, MO). A lipoated, huPDC-E2 peptide (p163) spanning amino acids 163-176 (GDLLAEIETDKATI), previously shown to be a major T cell epitope (6 -8), was synthesized by Alpha Diagnostic International (San Antonio, TX). Peptide purity was confirmed by reversed phase HPLC. Female SJL/J mice were obtained from Jackson Laboratories (Bar Harbor, ME) and all procedures conducted according to the principles outlined in the Guidelines of the Committee on Care and Use of Laboratory Animals. In the first set of experiments, mice were immunized at 12 weeks of age with a single 200 ml intraperitoneal injection of 500 mg of either bovine PDC-E2 or a-casein emulsified in complete Freund's adjuvant as described (Jones et al., 2000) ; the complete Freund's adjuvant (CFA) used contained 10 mg/ml Mycobacterium tuberculosis strain H37Ra (Difco) in incomplete Freund's adjuvant purchased from Sigma Inc. (St Louis, MO). In subgroups of mice, 3000 units of murine recombinant IFN-g was administered intraperitoneally twice a week for four weeks (Kyuwa et al., 1998) . The mice were divided into the following groups: (A) PDC-E2; (B) PDC-E2 and IFNg; (C) a-casein; and (D) a-casein and IFN-g in groups of 8 -12 animals (Table I) . Mice were serially observed and by 30 weeks of age, mice were sacrificed. This time period was chosen based on earlier observations by Jones et al. (2000) . Sera were assayed for autoantibodies. In addition, the liver, salivary gland, thymus, lung, heart, kidney, spleen, stomach, small and large intestine, were all removed for evaluation. Tissue specimens were fixed with 10% neutral buffered formalin and embedded in paraffin. Deparaffinized thin sections from each paraffin block were stained with hematoxylin and eosin for histology. Because some pilot data suggested the possibility of amyloid deposition, a Direct Fast Scarlet stain was performed and tissues viewed using polarized microscopy. In the second set of experiments, two groups of five mice were immunized at 8 weeks of age with a single 200 ul intraperitoneal injection of the following: (E) 50 ug of p163 in incomplete Freund's adjuvant (IFA); or (F) IFA alone. Mice were serially observed and at 36 weeks of age, they were sacrificed. Sera were assayed for autoantibodies. The livers were removed for evaluation as described above. Antimitochondrial antibodies were assayed using a combination of ELISA, immunoblotting and immunohistochemical staining of HEp-2 cells. For ELISA, microtiter plates were coated with 10 mg/ml PDC-E2 suspended in carbonate buffer (pH 9.6) and incubated overnight at 48C. After washing three times with 0.05% Tween-20 in PBS (PBS/Tween), plates were blocked with 3% milk in PBS for 1 h at room temperature (RT). Diluted mouse sera (1:200) was distributed into individual wells and incubated for 1 h at RT. After washing, the plates were incubated with 0.1 ml of HRP-conjugated goat anti-mouse immunoglobulin (Zymed, South San Francisco, CA) for 1 h. Reactivity was visualized by addition of 40 mM 2,2 0azinobis (3-ethylbenzthiazoline-6-sulfonic acid) dissolved in 0.05 M citric acid buffer (pH 5) containing 0.5 M H 2 O 2 . The reaction was stopped after 15 min by the addition of 5% SDS. Reactivity was measured using a Wallac spectrophotometer at a wavelength of 405 nm. Known positive and negative sera were included with each assay. For immunoblotting, samples were suspended in 250 ml of sample buffer (125 M Tris -HCl (pH6.8) containing 4% SDS, 20% glycerol and 5% 2-ME), boiled for 5 min, and resolved by SDS-PAGE using 1.5 mm-thick slab gels with a 4.75% stacking gel and a 10% separating gel. Separated proteins were transferred electrophoretically to nitrocellulose filters purchased from Micron Separations (Westboro, MA). After transfer, nitrocellulose filters were cut into strips, blocked with 3% milk in PBS for 1 h at RT and probed by incubation for 1 h with mouse sera diluted at 1:200. After washing with PBS/Tween, strips were incubated for an additional hour with HRP-labeled goat anti-mouse or anti-human polyvalent immunoglobulin (1: 2000). Strips were washed and visualized with enhanced chemiluminescent substrate purchased from Pierce (Rockford, IL). Known AMA positive PBC sera and control negative sera were used throughout these studies as controls. Immunohistochemical staining of Hep-2 cells were performed using Hep-2 cell slides purchased from Antibodies Inc. (Davis, CA). Slides were incubated at RT for 1 h with SJL/J mice sera, human PBC sera or control sera diluted at 1:40. After incubation, slides were washed in PBS for 10 min. The reaction was followed by a 30 min incubation with FITC-conjugated goat anti-mouse or anti-human IgG diluted in PBS. Once again, known positive and negative sera were used throughout. One defining serologic characteristic of PBC is the development of high titer AMA specific for the E2 subunit of PDC-E2 . Mice were bled and screened for the development of AMA. All sera collected at 7 weeks from SJL/J mice immunized with PDC-E2 (group A) showed positive reactivity against PDC-E2 by immunoblot (Fig. 1, lanes 1 and 2) . In contrast, sera isolated at 7 weeks from mice immunized with a-casein (group C) failed to react with PDC-E2 (Fig. 1, lane 3) . All responses from mice immunized with PDC-E2 (group A) were similar to responses of mice immunized with PDC-E2 plus IFN-g (group B), including reactivity with the conserved inner lipoyl domain of PDC-E2 (data not shown). At subsequent time points, through sacrifice (30 weeks), sera reactivity was very much similar to that determined at 7 weeks post-immunization ( Fig. 1, lanes 4 and 5). In contrast, sera isolated at 30 weeks from mice immunized with a-casein (group C) failed to react with PDC-E2 (Fig. 1, lane 6) . Results performed by immunoblot were similar to those obtained by ELISA. Further, sera from mice primed with PDC-E2 (Groups A and B) displayed a typical speckled (punctate) antimitochondrial staining pattern (Fig. 2a) . However, a single mouse primed with PDC-E2 and who received IFN-g (group B) demonstrated an anti-nuclear antibody pattern at a 1:40 dilution. Sera isolated from mice immunized with a-casein (groups C and D) also displayed weak antinuclear antibody staining patterns (Fig. 2b) . At 7 and 30 weeks post-immunization, chronic portal inflammation and granulomas were observed in all mice, irrespective of the antigen used (Table I, Fig. 3a and b) . Mild bile duct injuries were prevalent as demonstrated by the irregular arrangement of the biliary epithelium (Fig. 3c) . Moreover, significant lymphoid cell infiltration into the biliary epithelium and isolated apoptotic biliary FIGURE 1 Immunization of SJL/J mice with PDC-E2 induces AMA. Sera isolated from SJL/J mice immunized with PDC-E2 as well as PBC patient sera recognized PDC-E2 (74 kD) and OGDC (48 kD). Lanes 1 and 2 represent sera isolated at 7 weeks from mice immunized with PDC-E2. Lane 3 represents sera isolated at 7 weeks from a mouse immunized with a-casein. Lanes 4 and 5 represent sera isolated at 30 weeks from mice immunized with PDC-E2. Lane 6 represents sera isolated at 30 weeks from mice immunized with a-casein. Lane 7 is serum from a PBC patient. epithelial cells were observed in all groups. Of note, this pathology was observed in both small and large bile ducts (data not shown). Loss of bile ducts and destruction of the portal tracts leading to cirrhosis failed to develop in any mice. Interestingly, focally globular amyloid depositions developed along the walls of hepatic vessels in mice immunized with bovine PDC-E2 regardless of IFN-g treatment (Fig. 3d) . These amyloid depositions were characterized as AA type as the amyloid was sensitive to KMnO 4 treatment. In contrast, amyloid deposition was not detected in mice injected with the a-casein and there was no significant difference in histology between a-casein groups treated with or without IFN-g administration (data not shown). Although PDC-E2 is a ubiquitous antigen found in every cell in the body, pathology is localized to small bile ducts and salivary glands. Thus, it was important to confirm the specificity of the immune response induced by injection of bovine PDC-E2 by not only analyzing liver tissue, but also selected control tissues. It was interesting that there was lymphoid cell infiltration associated with mild duct injuries in the salivary glands of mice in all groups (Fig. 3e) . Moreover, after 30 weeks, lymph node swelling was evident in all mice. However, in contrast to the liver and salivary glands, there were no indications of significant inflammatory responses in other organs. Thus, immune responses to bovine PDC-E2 and a-casein with or without IFN-g were localized to the ducts of the liver and salivary glands and appeared to be due to CFA or impurities in our antigen preparations. The initial study by Jones et al. (23) indicated that immunization of SJL/J mice with an immunogenic huPDC-E2 peptide (p163) was as effective as whole, bovine PDC in inducing hepatic histologic changes characteristic of PBC. To eliminate differences in antigen and CFA preparation as a reason for the difference between our results and those of Jones et al., studies were conducted using synthesized p163 and IFA. At 36 weeks post-immunization, we observed significant chronic portal inflammation with bile duct invasion and damage in 2 of 5 mice injected with lipoated p163 þ IFA (Table II) . No granulomas were detected in any of these livers. Low titers of anti-PDC-E2 antibodies were detected by immunoblotting against purified human PDC-E2 only in the sera of the two mice with histologic changes (data not shown). Pre-incubation of sera with p163 inhibited immunoblotting of huPDC-E2. However, portal inflammation with bile duct invasion and damage, though less severe, was also observed in 3 of 5 mice injected with IFA alone. Granulomas were observed in two of these mice. Anti-PDC-E2 antibodies were not detected in the sera of these mice. Serum alkaline phosphatase levels were elevated only in the mouse (E4) with the most severe portal tract damage (data not shown). Portal inflammation with bile duct damage occurred as frequently in SJL/J mice injected with IFA alone as compared to mice injected with p163 and IFA. PBC is an autoimmune disease associated with an inflammatory response localized to the small bile duct cells of the liver Ludwig, 2000) . While attempts to establish an animal model for PBC in a number of animal species, including several strains of mice, have had with little success (Krams et al., 1989; Zeniya, 2000) , Jones et al. have proposed an animal model for PBC termed "experimental autoimmune cholangitis" in which SJL/J mice are immunized with purified bovine PDC-E2 (Jones et al., 2000) . Both the use of SJL/J mice and bovine PDC-E2 may have been important factors in establishing this proposed model of PBC. Unclear susceptibility factors in SJL/J mice have made these mice useful for the establishment of models of other autoimmune diseases (Vidal et al., 1994; Abedi-Valugerdi and Moller, 2000; Winer et al., 2001) . Additionally, purified bovine PDC-E2 may be more efficient at inducing an immune response for several reasons. In our hands, native PDC-E2 isolated from bovine heart tissue induces a stronger response than E. coli-produced recombinant protein when injected into mice (unpublished data). First, it has been suggested that lipoic acid is essential for optimal antigenicity (Fregeau et al., 1990; Bassendine et al., 1998; Koike et al., 1998) . Native protein is lipoated while synthetic peptides and recombinant proteins may not be lipoated. Second, it is important to note that bovine PDC-E2 differs from murine PDC-E2 by 28% at the amino acid level as determined by a protein blast. However, the antigenic region of bovine PDC-E2, including lipoic acid, has high homology with the murine protein and thus select regions may be recognized as a foreign antigen or altered self by the murine immune system. It thus may induce a stronger immune response against the autoantigen than a recombinant self protein or peptide (Coppel and Gershwin, 1995; Mackay, 2000) . In the study herein, we conducted a thorough serologic and histologic review to evaluate the premise that the SJL/J mouse is a suitable animal model for PBC. We also injected purified bovine PDC-E2 as the antigen to initiate the immune response in SJL/J mice. Our control group was not injected with CFA alone, as in the study by Jones et al., but with a-Casein and CFA. Mild bile duct pathology was induced by immunization with bovine PDC-E2 emulsified in CFA. However, similar pathology was also noted in the mice immunized with the control protein (a-Casein) emulsified in the same adjuvant. Antigen specific AMA were successfully induced in experimental SJL/J mice but not control mice. However, the characteristic bile duct injury typical of PBC failed to develop even after 30 weeks, although mild bile duct pathology and portal inflammation persisted. Furthermore, inflammation and diffuse bile duct pathology were observed, irrespective of bile duct size. In contrast, only small bile ducts are damaged in human PBC and affected bile ducts are distributed sparsely, not diffusely (Sasaki et al., 2000) . To magnify the inflammatory environment in SJL/J mice, we injected exogenous IFN-g but such injections failed to alter the pathology of PBC. The addition of IFN-g did not change the specificity or intensity of histologic or serologic effects. These results are not surprising since the pathology observed herein is antigen non-specific. Hence, this system is different from the disease associated features of IFN-g when NOD mice are injected with exogenous cytokines (Xiang and Zaccone, 1999) . Alternatively, treatment of IFN-g may cause classswitching of antigen specific AMA to IgG1. It is possible that IgG1 is not pathogenic. In fact, recently our lab has demonstrated that IgA may significantly contribute to the pathology of PBC (Reynoso-Paz et al., 2000) . In this respect, we should note that mice do not have IgA receptors on their bile duct cells and, in fact, have a significantly different mucosal immune system than that of humans . To eliminate non-specific effects caused by impurities in our antigen or adjuvant preparations, SJL/J mice were also injected with a lipoated, immunogenic huPDC-E2 peptide (aa 163 -176) emulsified in IFA or with IFA alone. Portal inflammation with bile duct invasion and damage occurred with similar frequency in both groups. There is no indication that the development of portal inflammation was specific to a loss of tolerance to PDC-E2, though the severity of inflammation was greater in mice expressing anti-PDC-E2 autoantibodies. Thus, both the development of autoantibodies and portal tract inflammation may simply reflect the severity of the inflammatory response following ip injection. SJL/J mice appear to be predisposed to the development of portal inflammation with minimal provocation. It is of interest that systemic amyloidosis of AA type was induced in most of the SJL/J mice immunized with PDC-E2 but not in controls with a-casein. Generally, AA type amyloidosis occurs secondary to systemic chronic inflammatory changes such as rheumatoid arthritis and tuberculosis. Heretofore, AA amyloidosis in mice has been previously induced in a number of strains of mice by repeated injections of a-casein, chemicals and endotoxin (Skinner et al., 1977; Baumal, 1979; Rokita et al., 1989; Kisilevsky and Young, 1994; Sipe, 1994) , and in a transgenic mouse model carrying human IL-6 gene with mouse metallothionein-I promoter (Solomon et al., 1999) . Spontaneous amyloid deposits have also been noted in aged SJL/J mice. In fact, the injection of amyloid enhancing factor (AEF) together with a-casein in CFA induced AA amyloidosis in SJL/J mice, while a-casein alone did not induce amyloidosis (Rokita et al., 1989) . Our findings are in agreement with these findings but not those of Jones et al. (1999) . The SJL/J mouse is an interesting model for a number of diseases and it remains possible that with further study, significant information may be forthcoming on the inflammatory environment within the liver and salivary glands. However, it appears that such responses are antigen non-specific.

La Crosse virus infectivity, pathogenesis, and immunogenicity in mice and monkeys

BACKGROUND: La Crosse virus (LACV), family Bunyaviridae, was first identified as a human pathogen in 1960 after its isolation from a 4 year-old girl with fatal encephalitis in La Crosse, Wisconsin. LACV is a major cause of pediatric encephalitis in North America and infects up to 300,000 persons each year of which 70–130 result in severe disease of the central nervous system (CNS). As an initial step in the establishment of useful animal models to support vaccine development, we examined LACV infectivity, pathogenesis, and immunogenicity in both weanling mice and rhesus monkeys. RESULTS: Following intraperitoneal inoculation of mice, LACV replicated in various organs before reaching the CNS where it replicates to high titer causing death from neurological disease. The peripheral site where LACV replicates to highest titer is the nasal turbinates, and, presumably, LACV can enter the CNS via the olfactory neurons from nasal olfactory epithelium. The mouse infectious dose(50 )and lethal dose(50 )was similar for LACV administered either intranasally or intraperitoneally. LACV was highly infectious for rhesus monkeys and infected 100% of the animals at 10 PFU. However, the infection was asymptomatic, and the monkeys developed a strong neutralizing antibody response. CONCLUSION: In mice, LACV likely gains access to the CNS via the blood stream or via olfactory neurons. The ability to efficiently infect mice intranasally raises the possibility that LACV might use this route to infect its natural hosts. Rhesus monkeys are susceptible to LACV infection and develop strong neutralizing antibody responses after inoculation with as little as 10 PFU. Mice and rhesus monkeys are useful animal models for LACV vaccine immunologic testing although the rhesus monkey model is not optimal.

proteins in overlapping reading frames: the nucleoprotein (N) and a non-structural protein (NS S ) which suppresses type 1 interferon (IFN) in the mammal host. The M segment encodes a single polyprotein (M polyprotein) that is post-translationally processed into two glycoproteins (G N and G C ), and a non-structural protein (NS M ) [3] . G N and G C are the major proteins to which neutralizing antibodies are directed [4, 5] . The L segment encodes a single open reading frame for the RNA-dependent RNA polymerase (L) [6] . The virus is transmitted by hardwood forest dwelling mosquitoes, Aedes triseriatus, which breed in tree holes and outdoor containers. Ae. triseriatus feed on Eastern chipmunks (Tamias striatus grinseus) and Eastern gray squirrels (Sciurus carolinensis) which serve as amplifying hosts for LACV [7] [8] [9] . Interestingly, the virus can be maintained in the mosquito population in the absence of vertebrate hosts by transovarial (vertical) transmission, thus allowing the virus to over-winter in mosquito eggs [9] . More recently, LACV has been isolated from naturally infected, non-native Aedes albopictus mosquito [10] . The infection of Ae. albopictus may represent a shift in virus ecology and increases the possibility for generation of new reassortants [11] . LACV was first identified as a human pathogen in 1960 after its isolation from a 4 year-old girl from Minnesota who suffered meningoencephalitis and later died in La Crosse, Wisconsin [12] . In humans, the majority of infections are mild and attributed to the "flu" or "summer cold" with an estimated 300,000 infections annually, of which there are only 70-130 serious cases reported [1, 2, 13, 14] . Isolation of virus is rare and has been made from post-mortem brain tissue collected in 1960, 1978, and 1993 [15-18] . Two isolates from non-fatal LACV cases were collected in 1995, one from a brain biopsy of a child and one from cerebrospinal fluid [16, 19] . Histopathologic changes in two human cases, one from 1960 and one from 1978, were characteristic of viral encephalitis. Inflammatory lesions consisted of infiltration of mononuclear leukocytes either diffusely or as microglial nodules. The largest inflammatory foci were observed in the cerebral cortex, including the frontal, parietal, and temporal lobes, and foci could also be found in the basal ganglion and pons. In these two cases, there was a lack of inflammatory lesions in the posterior occipital cortex, cerebral white matter, cerebellum, medulla, and spinal cord [17] . Brain biopsy from a non-fatal LACV infection contained areas of perivascular mononuclear cuffing and focal aggregates of mononuclear and microglia cells [16] . RT-PCR analysis of neural tissues from the 1978 case could only detect viral RNA in the cerebral cor-tex and not in the medulla, cerebellum, spinal cord, basal ganglion, or temporal lobe [20] . In children and adults, severe LACV encephalitis clinically mimics herpes simplex virus encephalitis with fever, focal signs, and possible progression to coma [16, 21, 22] . Confirmatory diagnosis has been made by RT-PCR of cerebrospinal fluid to exclude herpes simplex virus and by fluorescent staining for LACV antigen in brain biopsy sections [16] . Children who recover from severe La Crosse encephalitis may have significantly lower IQ scores than expected and a high prevalence (60% of those tested) of attention-deficit-hyperactivity disorder [13] . Seizure disorders are also common in survivors [23] . Projected lifelong economic costs associated with neurologic sequelae range from $48,775 -3,090,398 per case [24] . Currently, a vaccine or specific antiviral treatment is not available, but could serve to reduce the clinical and economic impact of this common infection. Although evidence of LACV infection has been reported for several species, only limited research has been done to understand LACV pathogenesis in its natural host or experimental rodents [8, [25] [26] [27] . LACV administered subcutaneously to suckling mice first replicates in muscle, and viremia develops with virus invading the brain across vascular endothelial cells [28] [29] [30] [31] . Virus replication in muscle was confirmed by immunohistochemical (IHC) staining, and the predominant cell type infected in the CNS is the neuron [28, 32] . The virulence of LACV for mice decreases with increasing age, similar to humans in which it causes CNS disease predominantly in pediatric populations [13, 28] . As an initial step in the establishment of animal models useful for vaccine development for humans, we sought to better characterize the tissue tropism of LACV in mice by identifying the tissues that support LACV replication after peripheral inoculation. We have previously described Swiss Webster mice as suitable for characterization of LACV infection at birth and at 3weeks of age [6] . Here we inoculated 3-week old Swiss Webster mice with either 1 or 100 LD 50 of virus intraperitoneally. Twenty tissues were individually collected for six consecutive days and processed for virus titration, immunohistochemical staining, and histopathology studies. Since experimental infection of non-human primates with LACV has not been reported, we also sought to determine if rhesus monkeys were susceptible to LACV infection. Rhesus monkeys were chosen since they are susceptible to a variety of neurotropic arboviruses, including flaviviruses [33] . In this study, rhesus monkeys were infected intramuscularly or subcutaneously with a mosquito or human isolate of LACV. These two isolates were used since preliminary genomic sequence analysis indicate that there are host specific differences between LACV isolated from humans and mosquitos [6] . LACV was found to be highly infectious for rhesus monkeys, but infection did not result in viremia, disease, or significant changes in blood chemistries or cell counts. However, high titers of neutralizing antibodies developed in all monkeys indicating that rhesus monkeys, although not optimal, will be useful for studying the infectivity and immunogenicity of LACV vaccine candidates. Previous evaluation of LACV in suckling mice revealed that it first replicated in striated muscle cells from which it seeded the blood and next invaded the CNS, where it replicated in neurons [28, 32] . In developing a rodent model for our live attenuated LACV vaccine development program, we sought to characterize LACV infection in outbred weanling mice, rather than suckling mice, since older mice can be used to study both the level of attenuation and the immunogenicity of our LACV vaccine candidates. To identify key steps in pathogenesis of LACV in weanling mice, we evaluated LACV tissue tropism after peripheral inoculation of wild-type virus and sought to identify tissues in which virus replicated efficiently. Weanling Swiss Webster mice (21-23 days-old) were inoculated intraperitoneally with 1 or 100 LD 50 (2.5 or 4.5 log 10 PFU) of LACV/human/1960, and the tissues indicated in Figures 1 and 2 were collected from 3 mice per day on days 1-6 post inoculation. Following inoculation of either dose, virus could first be detected in tissue near the inoculation site such as inguinal lymph nodes, spleen, and ovaries/uterus. Virus was detected in serum on days 1-3, but rarely on subsequent days even in moribund mice. By day three, virus distribution was widespread and could be found in the majority of tissues sampled, albeit at very low levels with titers rarely exceeding those found in serum. The highest virus titers detected were present in nasal turbinates, brain, and spinal cord. Respiratory tissue, including lung and nasal turbinates, contained virus following inoculation with 100 or 1 LD 50 beginning on days 1 and 2, respectively. CNS infection appeared to follow respira-Tissue distribution of La Crosse virus following intraperitoneal (IP) inoculation of Swiss Webster mice with 10 2.5 PFU Figure 1 Tissue distribution of La Crosse virus following intraperitoneal (IP) inoculation of Swiss Webster mice with 10 2.5 PFU. a Percent of mice positive by plaque assay represented by shading: 100% black, 66% dark gray, 33% light gray, 0% no data entry. Mean virus titer calculated only for virus positive tissues. Areas left blank indicate virus titer below detection limit of 0.7 log10 PFU/tissue. tory tissue infection, with virus being present in the brain 5 days after infection with 1 LD 50 and on day 2 after infection with 100 LD 50 . At either virus dose, infection appears early in the lymph nodes and major organs, with subsequent infection of the upper respiratory tract (nasal turbinates) followed by infection of the brain and eventually the spinal cord. By day six, mice began to succumb to infection in the high dose group showing signs of paralysis whereas mice in the low dose group failed to show clinical signs at this time, but would have succumbed later in infection. These results indicate that LACV replicates to low to moderate levels in peripheral tissues in weanling mice, with the nasal turbinates rather than striated muscle being the major site of replication. Since LACV replicated to high titers in the nasal turbinates, we sought to determine if intranasal inoculation of mice with LACV could lead to infection. Three-weekold Swiss Webster weanling mice (n = 6/dose) were inoculated intranasally (IN) (10 µl volume) or intraperitoneally (IP) (100 µl volume) with serial dilutions of LACV/ human/1960, and the LD 50 and 50% infectious dose (ID 50 ) were determined. Clinical disease served as a surrogate for lethality and mice were promptly euthanized prior to succumbing to LACV disease. In both groups, clinical disease was first noted on day 6 ( Figure 3 ). Twenty days post-inoculation, the LD 50 was determined. All surviving mice were tested for the development of a neutralizing antibody response. respectively) were slightly lower than the LD 50 titers, indicating LACV can cause a subclincal infection in weanling mice, but only at low doses. To further characterize the LACV infection in weanling mice, an additional group was inoculated intraperitoneally with 100 LD 50 of LACV/human/1960 and selected tissues (serum, muscle, nasal turbinate, brain, and spinal cord) were collected for virus quantitation (n = 5, daily for six days) to confirm titers found in Figure 2 and for histopathological and immunohistochemical (IHC) examination (n = 3, daily for six days), and the data is summarized in Table 1 and Figures 4 and 5. Virus titers for nasal turbinates, brain, spinal cord, muscle, and serum were in agreement with findings in Figure 2 (data not shown). Histopathologic lesions in the brain (including areas of the olfactory bulb, cerebral cortex, thalamus, hippocampus, and medulla oblongata) and spinal cord included perivascular cuffing ( Figure 4A ), neuronal degeneration ( Figure 4B -C), necrosis (either single cell or small foci), and apoptosis (4F). There was an infiltration of CD3+ lymphocytes and macrophages ( Figure 4D -E). The most extensive lesions occurred in the medulla oblongata and were associated with perivascular cuffing. Histopathological changes were minimal outside the CNS. In the spleen, lymphoid atrophy was only observed on days 3-4 post-infection. Plasmacytosis in the spleen was observed on days 4-6 and areas of necrosis were observed on days 4-5 with myeloid hyperplasia on day 5. Histopathological changes were not observed in respiratory nasal epithelium or muscles of the upper rear limb. To determine the location of cells expressing viral antigen, tissues were immunostained for La Crosse virus antigens. Viral antigens were not observed in the nasal turbinates, muscle, spleen, or pancreas. However, viral antigen was detected in the olfactory bulb of the brain, thalamus, cerebrum, medulla oblongata, and spinal cord. On day 5, viral antigen could be detected in all sampled brain regions with a greater number of cells and regions positive for viral antigen than for overt histologic lesions. Mild perivascular cuffing was seen in a few areas and single cell necrosis was seen in areas stained positive for viral antigen. At this time and at later days, the olfactory bulb was more extensively involved ( Figure 5A ), although viral antigen was not detected in sections of the underlying nasal epithelium. Viral antigen could be detected in the brain tissues of all mice on day 5, but only small foci of antigen were seen in the spinal cord of one mouse suggesting brain infection proceeds spinal cord infection. By day 6, viral infection in the CNS was widespread and extensive throughout all regions of the brain and spinal cord ( Figure 5B ). Neurons were the main cell expressing viral antigens, but supporting glia also appeared positive ( Figures 5D-F) . Single cell necrosis, focal necrosis (more than one cell), and apoptotic bodies were prominent throughout the lesions ( Figure 4C ) and apoptotic bodies stained positive by TUNEL staining ( Figure 4F ). Degenerative neuronal changes were also commonly seen, includ- To develop a non-human primate model of LACV infection for pathogenesis studies and for testing of future vaccine candidates, rhesus monkeys were inoculated with 10 5 PFU of biologically cloned (terminally diluted) human or mosquito LACV (LACV/human/1978-clone, LACV/mosquito/1978-clone). Illness was not observed following virus administration, and virus was not detected at any time in serum samples ( Table 2 ). Virus was present at a low titer in lymph nodes on days 6, 8, and 12, however, virus replication in these tissues could not be identified by IHC staining. Despite the low level (or absence) of viremias and highly restricted replication in the tissues sampled, all monkeys developed neutralizing antibody responses that were first detected on days 6-8 indicating that the each monkey was infected. Twenty-eight days after inoculation, neutralizing antibody titers (PRNT 60 ) for each group were in the range of 1:560 -1:2186 (Table 2) . Low-level cross-reactive antibodies were present in two monkeys (CL6E and DBOH) at the start of the experi-ment. A boost in antibody titer in these monkeys was not observed compared to other monkeys, suggesting that this experimental LACV infection was a primary infection. Complete blood counts (CBC) with differential and blood chemistries were analyzed at each blood collection. Since LACV infection in monkeys was asymptomatic and also since differences between the four virus groups indicated in Table 2 were not observed, the CBC and blood chemistry data were averaged for the 16 animals to detect changes in blood values during the course of infection (Table 3) . Days in which specific parameter values for a significant number of monkeys were greater than 1 standard deviation from normal appear boxed in Table 3 with mean values for each test shown. After day 6, the majority of monkeys experienced a slight anemia, which may in part be associated with the overnight fast in preparation for anesthesia prior to each blood collection. This analysis suggests that infection of major organs such as liver was minimal or absent. To estimate the minimum dose required to infect a monkey, rhesus monkeys were inoculated with LACV/mosquito/1978-cl at 10 1 or 10 3 PFU subcutaneously. Blood was collected on days 0, 21, 28, and 42, and serum neutralizing antibody titers were determined. Mean neutralizing antibody titers were 1:355 and 1:82 for the 10 1 or 10 3 PFU dose groups, respectively, and all monkeys seroconverted by day 28 (PRNT 60 ≥ 40) (Table 4 ). Thus, LACV is highly infectious for rhesus monkeys even at a dose of 10 1 PFU and results in the induction of a high level of neutralizing antibody. However, LACV infection did not result in identifiable clinical abnormalities in this group of 24 monkeys. As an initial step in development of a live attenuated LACV vaccine, we sought to better characterize LACV infection in weanling mice because at this age mice can be used to evaluate both the level of attenuation and immunogenicity of candidate vaccine viruses. Previous LACV pathogenesis studies in suckling mice inoculated subcutaneously with a mosquito isolate of LACV demonstrated that viral antigen was detected in the cytoplasm of striated muscles, the interscapular brown fat, and the endothelial and smooth muscle cells of small arteries and veins [28] . When virus was first detected in the brain, it was confined to cerebral vascular endothelial cells but later spread to neurons. The authors suggest that the late infection of the dorsal route ganglion indicates that the virus does not penetrate the CNS via nerves but rather by vascular endothelial cells [28] . This previous model therefore suggests that virus first replicates in muscle cells leading to the development of viremia with subsequent hematogenous spread to the CNS, and we sought in the present study to examine if this pattern of infection also occurred in weanling mice. In weanling mice inoculated intraperitoneally with 1 or 100 LD 50 of LACV, virus was first detected on days 1 -3 in tissues near the inoculation site. At either dose, virus was no longer detectable by days 4 and 5 in these tissues, suggesting that it was rapidly cleared by the innate immune system. Interestingly, the virus was not detected in muscle tissue until day 3 post inoculation and clearly did not preferentially infect this tissue. Rather, outside the CNS, the virus replicated to highest titers in the nasal turbinates and appears to spread from this site into the brain. Immunohistochemical staining of the nasal turbinate tissue was not sensitive enough to identify the LACV infected cells, but is thought LACV may gain access to the CNS via cells in the nasal turbinates. This suggestion is offered with the caveat that respiratory epithelial cells could also have been infected, but the magnitude of the infection could not be detected with the IHC staining. To travel from the nasal olfactory epithelium to the olfactory bulb, the virus a Days with significant number of monkeys 1 SD from the pre-inoculation mean (day -7 and 0 combined) are boxed (chi square, p < 0.05, n = 16), mean values displayed. b Abbreviations: MCV-mean corpuscular volume, MCH-mean corpuscular hemoglobin, MCHC-mean corpuscular hemoglobin concentrations, ASTaspartate aminotransferase, ALT-alanine aminotransferase, ALP-alkaline phosphatase, BUN-blood urea nitrogen. c Pre-inoculation mean determined from samples collected on days -7 and 0 post inoculation. would follow olfactory neurons into the brain, as infection is first detected in the rostral section of the brain. Although virus replication in nasal turbinate tissue was detected, we were unable to identify the cells that were infected. Viruses such as vesicular stomatitis virus, rabies virus, mouse hepatitis virus, Borna disease virus, pseudorabies virus, and herpes simplex virus have all been demonstrated to enter the mouse CNS via olfactory neurons [34] [35] [36] [37] [38] [39] [40] . It is important to note that 2 of 62 mice tested had detectable virus within the brain without detectable virus in the nasal cavity (individual data not shown) suggesting that more than one route might be used to gain access to the CNS. We were surprised by the ability of the virus to infect intranasally, and found that the LD 50 and ID 50 were almost identical by either route by of inoculation. The kinetics of the development of clinical disease that occurred following intranasal administration of virus was similar for virus given IP or IN. The finding that LACV is able to infect very efficiently via the nasal route has possible implications for the ecology of the virus. It is possible that infectious virus is present in water collections containing Aedes mosquito larvae, e.g., tree holes, since the virus has been shown to be transmitted from infected mosquitoes to larvae via eggs. A mammalian host that drinks the water would intake both fluid, which might contain free virus from lysed larvae, and infected larvae, either of which could initiate an infection in the mammalian host. Thus, exposure to such water could lead to an alternate route of infection for the natural hosts, i.e., the oral/nasal route in addition to the vectorbourne route. This hypothesis needs to be confirmed experimentally but remains an interesting possibility. In the CNS of the weanling mouse, LACV infects predominantly neurons (some microglial cells are also infected) with spread in a rostral to caudal direction eventually reaching the lumbar spinal region. In both mice and humans, virus has been detected in the cerebral cortex, however infection appears more widespread in the mouse CNS with virus detection in the medulla oblongata, cerebellum, thalamus, olfactory bulb, and all regions of the spinal cord. The virus used in this study, LACV/human/ 1960, was isolated and passaged twice in C6/36 mosquito cells and was not previously adapted to growth in mouse neural tissue. One surprising difference between human and mouse infection was the detection of virus replication by IHC and the development of lesions in the mouse spinal cord [17] . Taken together, these data suggest that in weanling mice the virus first replicates in the periphery near the inoculation site. If the infection is not quickly controlled, the virus disseminates, most likely via blood, to the nasal turbinates. The detection of virus and lesions first in the rostral brain suggest the virus may utilize olfactory neurons to gain access to the CNS. The differences in findings between our study and previous work may be the result of differences in virus strain (mosquito vs. neurovirulent human isolates), mouse strain (outbred white vs. Swiss Webster), mouse age (suckling vs. weanling), inoculation route (subcutaneous vs. intraperitoneal) and dose (1000 TCID 50 vs. 1 or 100 LD 50 ). Although sequence data is not available for the strain used in the previous mouse pathogenesis work, it is known that the virus was a mosquito isolate and not directly linked to human disease [28] . In humans the majority of infections are asymptomatic, but children hospitalized with severe disease present with fever (86%), headache (83%), vomiting (70%), disorien- [13] . Like the majority of human infections, rhesus monkeys in the current study experienced a subclinical infection without the development of systemic disease or neurologic symptoms. A much greater number of monkeys would probably need to be tested to detect neurologic symptoms after peripheral inoculation. Nevertheless, all monkeys were infected with LACV and developed neutralizing antibody responses, even after inoculation with as little as 10 PFU. Future work will include the intracerebral inoculation of rhesus monkeys to determine if LACV is neurovirulent in this species, but this will wait until vaccine candidates have been identified. It is still unclear why so many human LACV infections remain asymptomatic. In our mouse model, infection with 1 LD 50 of virus resulted in delayed CNS infection compared to mice receiving 100 LD 50 . Mice were able to control virus infection at doses at or below the LD 50, and developed strong neutralizing antibody responses. The LACV ID 50 for humans is unknown, but if human exposure is limited to a small dose, the virus may be effectively controlled by the immune system and CNS infection may be averted. If, like our mouse model, an individual is exposed to a greater dose of virus, virus growth may outpace control mechanisms leading to CNS infection. To further support the role of immune control affecting human disease outcome, it has been shown that individuals residing in endemic areas with major histocompatibility complex molecule B49 on CD8+ cytotoxic T lymphocytes (HLA-B49) had a greater relative risk of developing encephalitis after infection (relative risk 17.65, χ 2 = 7.3, P < 0.1). Infected individuals with HLA-DR5 had a lower relative risk of developing seizures (relative risk 0.22, χ 2 = 5.10, P < 0.025) [41] . In weanling Swiss Webster mice, the LD 50 and ID 50 are similar, indicating that most infections lead to a lethal outcome at this age. LACV first replicates in tissues near the inoculation site, enters the blood stream, infects the nasal turbinates, and gains access to the CNS, presumably via olfactory neurons. This model will be useful to identify attenuated vaccine candidates that are deficient in the ability to disseminate from the site of inoculation, to replicate to high titers in the nasal turbinates, or to establish infection in mouse CNS. The CNS infection of mice appears more widespread than described for human infection, both in the brain and spinal cord. LACV is highly infectious both by the IP and IN routes suggesting that infection of natural mammalian hosts such as the chipmunk or squirrel might occur by the oral/intranasal route in addition to the bite of an infected mosquito. In rhesus monkeys, LACV is highly infectious with as little as 10 PFU resulting in the development of neutralizing antibodies, but clinical disease is not observed at any dose tested, suggesting that rhesus monkeys will be useful for studying the infectivity and immunogenicity of live attenuated virus vaccine candidates, but will be of limited usefulness for the evaluation of their level of attenuation. C6/36 cells (Aedes albopictus mosquito larvae) were maintained in Earle's MEM supplemented with 10% fetal bovine serum (HyClone), 2 mM L-glutamine (Invitrogen), and 1 mM non-essential amino acids. Vero cells (African green monkey kidney) were maintained in Opti-PRO™SFM medium (Invitrogen) supplemented with 4 mM L-glutamine (Invitrogen). LACV/human/1960 was isolated from post-mortem brain tissue collected from a Minnesota patient hospitalized in La Crosse, Wisconsin and passaged two times in C6/36 cells. LACV/mosquito/1978 was isolated from mosquitoes collected in North Carolina and passaged once in mouse brain and three times in Vero cells. LACV/human/ 1978 was isolated from post-mortem brain tissue collected in Wisconsin and passaged once in mouse brain, twice in BHK-21 cells, and once in Vero cells. Biological clones were generated by terminal dilution in Vero cell cultures. Passage histories and complete genomic sequences for all stocks used in this paper have been previously published (EF485030-EF48538) [6] . Weanling Swiss Webster mice (Taconic Farms, Germantown, NY) were inoculated with LACV/human/1960 diluted in L15 media (Invitrogen) intraperitoneally (100 µl volume) or intranasally (10 µl volume) while under isofluorane anesthesia. Mice were observed daily for 20 days for clinical disease including tremors, seizures, and limb paralysis. Moribund mice were promptly euthanized. Serum was collected 20 days after inoculation for determination of the presence and titer of neutralizing antibodies. The replication of LACV virus was evaluated in 3-week-old weanling female Swiss Webster mice (Taconic Farms, Germantown, NY). All animal experiments were carried out in accordance with the regulations and guidelines of the National Institutes of Health. The Swiss Webster mice, were inoculated IP with 1 or 100 LD 50 in a volume of 100 µl [6] . The tissues indicated in Table 1 were collected individually (3 mice per day for 6 days at each dose of virus), weighed, homogenized in L15 with SPG buffer (final concentration 218 mM sucrose, 6 mM L-glutamic acid, 3.8 mM dibasic potassium phosphate, pH 7.2). Homogenates were centrifuged for 10 minutes at 3000 rpm to remove cell debris and aliquots were frozen at -80°C until virus quantitation was performed. All mice were carefully observed twice daily for clinical disease including tremors and limb paralysis. Mice exhibiting clinical disease were promptly euthanized. Monolayer cultures of Vero cells grown on 24-well plates were infected in duplicate with ten-fold serial dilutions of tissue homogenate, and the cells were overlayed with OptiMEM (Invitrogen) supplemented with 1% methylcellulose, 5% FBS, 2.5 µg/ml amphotericin B, and 20 µg/ml ciprofloxicin. Five days after infection the overlay was removed and cells were washed twice with PBS. The cells were fixed and stained for 10 minutes with crystal violet solution, virus plaques were enumerated, and tissue titers were expressed as mean log 10 PFU/g tissue. Weanling Swiss Webster mice were inoculated intraperitoneally with 100 LD 50 of LACV/human/1960, and tissues were collected for six consecutive days for virus titration (n = 5) and pathology (n = 3) per day. Virus titrations were performed to confirm previous virus kinetics. Tissues collected for pathology were fixed in 10% neutral buffered formalin (NBF) for a minimum of 72 hours, embedded in paraffin and sections were prepared at 4-5 (µm). When bone was present in the tissue, as with muscle, nasal cavity, and spinal cord, tissues were decalcified using Immunocal (Decal Chemical Corp. Tallman, NY). Sections were stained with hematoxylin and eosin (H&E). A serial section was saved for immunohistochemical staining (see below). For immunohistochemical analysis, fixed serial sections of mouse tissues were mounted onto slides and deparaffinized with xylene, rehydrated in a series of ethanol solutions (100%, 95%, 70%, 50%), and washed with deionized water. TUNEL Staining to detect apoptosis was preformed using the "DeadEnd™ Colorimetric TUNEL System" (Promega USA, Madison, WI). Sections were pretreated with Pro K enzyme (provided in the kit, diluted 1:500 with PBS). Strepavidin (also provided in the kit) was diluted at 1:500 in PBS. To detect CD3+cells, slides were steam hydrated and pretreated with Diva Solution (Biocare Medical Concord, CA) for 20 minutes. The primary antibody, rabbit antihuman CD3, (A-0452, Dako Corporation, Carpentaria, CA) was used at a dilution of 1:300. The detection system was the Standard ABC kit (Vector Laboratories, Burlingame, CA), with 3,3'-diaminobenzidine (DAB) as the chromogen and a modified Harris hematoxylin counterstain. To detect macrophage antigen 2 (MAC-2) expressing cells, slides were stream hydrated with citrate buffer for 20 minutes. The primary antibody, rat anti-Mac-2 (TIB-166, ATCC, Mannasas, VA) was used at a 1:200 dilution followed by a biotinylated goat-anti-rat IgG secondary antibody and developed with Streptavidin HRP (Biocare, Concord, CA). Rhesus monkeys were inoculated with 10 5 PFU of biologically cloned human or mosquito LACV (LACV/human/ 1978-clone, LACV/mosquito/1978-clone). Each virus was inoculated intramuscularly (n = 4) or subcutaneously (n = 4). Blood samples were collected on days -7, 0, 2, 4, 6, 8, 10, 12, 14, 21, 28 post inoculation for blood chemistries, virus titration, and neutralizing antibody titration. Axillary or inguinal lymph nodes were surgically excised on days 4, 6, 8, and 12 post inoculation for viral titer or fixed in buffered formalin for histopathology and immunohistochemical analysis. Monkeys were observed daily for clinical disease. To determine LACV infectivity at low doses, rhesus monkeys (n = 4) were inoculated with 10 1 or 10 3 PFU LACV/mosquito/1978-clone subcutaneously. Blood samples were collected on days 0, 21, 28, and 42 post inoculation for neutralizing antibody titration. Neutralizing antibody in mouse and monkey serum was quantitated by a plaque reduction neutralization assay. Test sera were heat inactivated (56°C for 30 minutes) and serial 2-fold dilutions beginning at 1:5 or 1:10 were prepared in OptiMEM (Invitrogen) supplemented with 2% FBS, 50 µg/ml gentamicin, and 0.5% human albumin (Talecris Biotherapeutics, Inc., Research Triangle Park, NC). The homologous LACV was diluted to a final titer of 500 PFU/ml in the same diluent and 10% guinea pig complement (Cambrex Bioscience Walkersville, Inc., Walkersville, MD) was added to equal volumes of the serum dilutions and mixed well. Serum/virus mixture was incubated at 37°C for 30 minutes, added to confluent monolayers of Vero cells, and incubated for 1 hour to allow virus attachment. Cells were overlayed with 1% methylcellulose and incubated for 5 days at 37°C. After incubation, the overlay was removed, and the monolayers were washed twice with PBS and stained with crystal violet to allow for the enumeration of virus plaques. A 60% plaque-reduction neutralization titer was calculated.

The intrinsically disordered C‐terminal domain of the measles virus nucleoprotein interacts with the C‐terminal domain of the phosphoprotein via two distinct sites and remains predominantly unfolded

Measles virus is a negative‐sense, single‐stranded RNA virus within theMononegavirales order,which includes several human pathogens, including rabies, Ebola, Nipah, and Hendra viruses. Themeasles virus nucleoprotein consists of a structured N‐terminal domain, and of an intrinsically disordered C‐terminal domain, N(TAIL) (aa 401–525), which undergoes induced folding in the presence of the C‐terminal domain (XD, aa 459–507) of the viral phosphoprotein. With in N(TAIL), an α‐helical molecular recognition element (α‐MoRE, aa 488–499) involved in binding to P and in induced folding was identified and then observed in the crystal structure of XD. Using small‐angle X‐ray scattering, we have derived a low‐resolution structural model of the complex between XD and N(TAIL), which shows that most of N(TAIL) remains disordered in the complex despite P‐induced folding within the α‐MoRE. The model consists of an extended shape accommodating the multiple conformations adopted by the disordered N‐terminal region of N(TAIL), and of a bulky globular region, corresponding to XD and to the C terminus of N(TAIL) (aa 486–525). Using surface plasmon resonance, circular dichroism, fluorescence spectroscopy, and heteronuclear magnetic resonance, we show that N(TAIL) has an additional site (aa 517–525) involved in binding to XD but not in the unstructured‐to‐structured transition. This work provides evidence that intrinsically disordered domains can establish complex interactions with their partners, and can contact them through multiple sites that do not all necessarily gain regular secondary structure.

Nucleoproteins of Paramyxoviridae are divided into two regions: a structured N-terminal moiety, N CORE (aa 1-400 in MV), which contains all the regions necessary for self-assembly and RNA-binding Curran et al. 1993; Bankamp et al. 1996; Liston et al. 1997; Myers et al. 1997b Myers et al. , 1999 Karlin et al. 2002a; Kingston et al. 2004b) , and a C-terminal domain, N TAIL . N TAIL is intrinsically unstructured (i.e., it lacks any stable secondary and tertiary structure in physiological conditions) ) and is exposed at the surface of the viral nucleocapsid (Heggeness et al. 1980 (Heggeness et al. , 1981 . The presence of a flexible region protruding from the viral nucleocapsid allows the establishment of interactions with numerous different viral partners and with several cellular proteins (Moyer et al. 1990 ; De and Banerjee 1999; tenOever et al. 2002; Zhang et al. 2002; Laine et al. 2003) . In Morbilliviruses and Respiroviruses, N TAIL is responsible for binding to P (Curran et al. 1993; Harty and Palese 1995; Bankamp et al. 1996; Liston et al. 1997; Longhi et al. 2003; Kingston et al. 2004b) , to the polymerase complex P-L, and to the matrix protein (Coronel et al. 2001) . Beyond viral partners, N TAIL also interacts with several cellular proteins, including the interferon regulatory factor 3 (tenOever et al. 2002) and the heat-shock protein Hsp72 (Zhang et al. 2002) . Moreover, N TAIL within viral nucleocapsids released from infected cells also binds to a yet unidentified protein receptor expressed at the surface of human thymic epithelial cells (Laine et al. 2003 (Laine et al. , 2005 . The P protein of Paramyxovirinae plays multiple roles in both transcription and replication: It is an essential subunit of the viral polymerase complex and acts as a bridge between the nucleocapsid template (N NUC ) and the polymerase complex L-P. P is a modular protein, consisting of an intrinsically unstructured N-terminal moiety (PNT) (Karlin et al. 2002b , and of a well conserved C-terminal moiety (PCT) which contains all the regions required for transcription (Curran 1996) . Paramyxovirinae PCT have a modular organization, consisting of alternating disordered and structured regions . In particular, they possess a coiled-coil domain (referred to as PMD, for P multimerization domain) responsible for both oligomerization and binding to L (Smallwood et al. 1994; Liston et al. 1995) , and a C-terminal globular region (referred to as XD), that is involved in binding to both monomeric and assembled forms of N (Curran et al. 1994 (Curran et al. , 1995a Harty and Palese 1995; Tuckis et al. 2002; Johansson et al. 2003; Kingston et al. 2004b) . Figure 1 shows a schematic representation of the MV N NUC -P complex, which highlights the role of N TAIL in the recruitment of P via the binding to XD. We have previously reported the crystal structure of the C-terminal globular domain of MV P (XD, aa 459-507): It consists of a monomeric protein composed of three a-helices, forming an anti-parallel three-helix bundle . We have also shown that N TAIL undergoes an induced folding in the presence of XD and that this unstructured-to-structured transition implies a gain of a-helicity . Using a combination of computational and biochemical approaches, we have identified within N TAIL an a-helical molecular recognition element (a-MoRE, aa 488-499) involved in binding to P and in the induced folding of N TAIL (Bourhis et al. 2004) . The a-MoRE has been modeled in the crystal structure of XD, in the long hydrophobic cleft delimited by helices a2 and a3 ). In this model, the contact of N TAIL with a hydrophobic patch at the surface of XD would be the driving force in the induced folding of the a-MoRE of N TAIL by burying apolar residues at the protein-protein interface. Very recently, Kingston et al. (2004a) have reported the crystal structure of a chimeric protein composed of MV XD and a peptide corresponding to residues 486-505 of MV N. The structure of the complex is a four-helix bundle in which the ahelix of N is bound in the reverse orientation with respect to the model proposed by Johansson et al. (2003) . Using NMR and crystallographic studies, Kingston et al. (2004a) deduced that the induced folding of N TAIL is restricted to only 18 residues (out of 125), although the analysis was restricted to the N region encompassing residues 477-505. (B) . The encapsidated RNA is shown as a dotted line. PMD is represented with a dumbbell shape according to Tarbouriech et al. (2000) . The tetrameric P (Rahaman et al. 2004 ) is shown bound to N NUC through three of its four C-terminal XD ''arms,'' as in the model of Curran and Kolakofsky (1999) . The L protein is shown as a rectangle contacting P through PMD by analogy with SeV (Smallwood et al. 1994) . In this paper we examine this localized induced folding event in the context of the entire N TAIL domain. A low-resolution structural model of the complex between XD and N TAIL shows the presence of a bulky globular region and of an extended and elongated shape. The model shows that the N-terminal region of N TAIL (residues 401-488) remains predominantly unfolded, and the 489-525 region is packed against XD, thus suggesting that beyond the a-MoRE the C terminus may play a role in the interaction with XD. Indeed, we present several lines of experimental evidence confirming that beyond the region encompassing the a-MoRE, an additional N TAIL region encompassing residues 517-525 also contributes to binding to XD, although without undergoing any gain of regular secondary structure. Small angle X-ray scattering (SAXS) is a valuable technique for the study of flexible, low compactness macromolecules in solution, which has already been successfully used to characterize N TAIL ). Therefore, we used SAXS to study the N TAIL -XD complex. To this endeavor, we cloned, expressed, and purified from the soluble fraction of Escherichia coli an N-terminally histidine-tagged form of N TAIL (i.e., N TAILHN ) (data not shown). The identity of the recombinant products was confirmed by immunoprecipitation (IP) studies using anti-N Cl25 and anti-hexahistindine tag monoclonal antibodies (mAbs) (data not shown). Beforehand, we checked whether XD possesses the same structure in solution as in the crystal. SAXS experiments performed on XD showed that it has a globular shape with a radius of gyration (R g ), extrapolated at a nil concentration, of 12.1 6 0.8 Å and a maximum diameter D max of 41 6 1 Å (data not shown). Comparison with the crystal structure using the program CRYSOL (Svergun et al. 1995) indicates that the scattering profile calculated for the crystal structure (PDB code 1OKS) is identical to the experimental one, fitting the data with a 2 of 1.4 and giving a theoretical R g of 12.3 Å (data not shown). These results suggest that the overall conformation of XD in solution is similar to that observed in the crystal. The scattering profile of the N TAILHN -XD complex was obtained as described in Materials and Methods. Analysis of the curve in the low q region with the Guinier approximation gave an R g of 32.7 6 0.7 Å . The molecular mass (MM) calculated from the forward scattering intensity I(0) is 23.5 6 2 kDa, in agreement with the value expected for a 1:1 stoichiometric complex (21.3 kDa), thus suggesting that the complex did form in solution. The high value of R g indicated that the overall structure of the N TAILHN -XD complex is not compact. The distances distribution function inferred from the scattering curve of the N TAILHN -XD complex exhibits a maximum at 20 Å , with a shoulder at about 30 Å and a long tail up to 146 Å , typical of an elongated object (see Fig. 2A ). The bump most probably corresponds to the intramolecular distances within the globular portion of the complex (see below), while the tail indicates that N TAIL possesses regions with an extended conformation. The overall envelope of the complex was restored ab initio from its scattering profile using the program GASBOR ). Several independent runs yielded different shapes with recurrent features: they were all elongated, with a globular cluster of always the same size at one extremity and an elongated protuberance with varying bends and cross-sections. The quality of the fit to the experimental data was similar in all cases, with a 2 of 1.3 to 1.6. The globular part most probably corresponds to XD packed against the folded region of N TAIL , while the outgrowth corresponds to unfolded regions of the latter. The crystal structure of the XD-N TAIL486-505 complex (PDB code 1T6O), which encompasses the a-MoRE (residues 488-499), was inserted in the shape using the program SUPCOMB (Kozin and Svergun 2001) (Fig. 2B ). Interestingly, only one protuberance was observed, corresponding to the N-terminal area of N TAIL , and no other protuberance was restored from the shape determination in the opposite region of the globular cluster, thus suggesting that the C-terminal region of N TAIL (aa 506-525), although not visible per se, is packed inside the bulky portion of the complex (Fig. 2B ). In order to get further insights into the conformation that the regions of N TAIL encompassing residues 401-485 and 506-525 adopt in the complex with XD, we used the program package CREDO, which is an extension of GASBOR and calculates the position of missing loops in crystal structures. The results obtained with several independent runs using the crystal structure of XD-N TAIL486-505 as the template, revealed at one side a peptide chain in an extended conformation and at the other side a chain packed against the bulky portion of the XD-N TAIL486-505 complex. The extended conformation, which protrudes from the globular cluster of the complex and points to the solvent, corresponds to the 92-residue-long N-terminal region of the N TAILHN construct. In contrast, the more compact peptide chain, corresponding to the 20-residue-long C-terminal region of N TAIL (aa 506-525), always packs against the XD-N TAIL486-505 complex but at varying distances and positions. An example of one of the calculated conformations is shown in Figure 2C . The experimental data were fitted by these models with similar 2 of about 0.8-1.0 (see Fig. 2C , inset). The various possible conformations provided by the program CREDO together with the extended structural properties of the complex (as indicated by the distance distribution function, the R g and the restored shape, with respect to the number of amino acids of the complex) are typical of disordered polypeptide chains lacking a stable regular structure. Accordingly, the structure shown in Figure 2C represents only one possible conformation found by CREDO among many others. Indeed, CREDO restores only one conformation per run, and therefore cannot account for multiple conformations due to disorder when fitting the data. Therefore, although a unique conformation of the N TAIL region encompassing residues 506-525 cannot be derived, all the solutions provided by CREDO consistently revealed a packing of the Cterminal region of N TAIL against XD. The recurrence of these packed conformations suggests that the Cterminal region of N TAIL interacts with XD. Moreover, we checked the influence of the Box3 conformation on the scattering curve by calculating the scattering profile of pseudomodels of the complex. In these pseudomodels, the N-terminal region of N TAIL is in the same conformation as in the credo model shown in Figure 2C . The only difference concerns the orientation of Box3, which points out of the complex, in different solventexposed conformations. Noteworthy, the theoretical scattering curves calculated from these pseudomodels using CRYSOL (Svergun et al. 1995) poorly fit with the experimental curve, with an average 2 of $ 7.0 (data not shown). This confirms that the C terminus of N TAIL does contribute to the observed scattering profile, thus attesting the reliability of the model provided by CREDO. In order to further explore the possible contribution of N TAIL regions other than the a-MoRE to XD binding, we have looked at the regions of homology conserved amongst members of the Morbillivirus genus (Diallo et al. 1994) . These regions of homology are herein referred to as Box1, Box2, and Box3 and span N residues 401-420, 489-506, and 517-525, respectively (see Fig. 3A ), with the a-MoRE being located within Box2. We have then designed three deletion constructs bearing different combinations of these homology boxes (Fig. 3A) . The gene fragments encoding the different N TAIL deletion proteins were cloned into the pDest14 vector (Invitrogen) to yield N-terminally histidine-tagged and C-terminally flag-tagged recombinant products, the expression of which is under the control of the T7 promoter. In all cases, most recombinant protein was recovered from the soluble fraction of bacterial lysates (Fig. 3B , lanes SN). The N TAIL deletion proteins were purified to showing the location of the a-MoRE, is given below. The inset shows the experimental scattering curve (black circles) and fit (red line) obtained by CREDO with the low-resolution model. homogeneity (>95%) in two steps: Immobilized Metal Affinity Chromatography (IMAC), and gel filtration (Fig. 3B ). The identity of the recombinant products was confirmed by IP studies using anti-N Cl25, anti-flag, and antihexahistindine tag mAbs (data not shown). As shown in Figure 3B , the three N TAIL deletion proteins migrate in SDS-PAGE with an apparent MM of either 20 kDa (N TAILD3 and N TAILD1 ) or 18 kDa (N TAILD2,3 ) (expected MMs are 14.5, 13.4, and 11.5 kDa, respectively). This abnormal migratory behavior has already been documented for N TAIL , where mass spectrometry analysis and N-terminal sequencing gave the expected results . The anomalous electrophoretic mobility is therefore due to a rather high content of acidic residues, as frequently observed in intrinsically disordered proteins (Tompa 2002) . Likewise, the behavior of the truncated N TAIL proteins can probably be accounted for by this sequence bias composition. The number of different conformations of the N TAIL deletion proteins is limited, as indicated by the sharpness of the peaks observed in gel filtration (data not shown). As expected for intrinsically disordered protein subdomains, the Stokes radius (R S ) values, as inferred by gel filtration (27 6 3 Å , 22 6 3 Å , and 27 6 3 Å for N TAILD3 , N TAILD2,3 , and N TAILD1 , respectively), are consistent with extended conformations (see Materials and Methods). Thus, these deletion proteins share similar hydrodynamic properties with full-length N TAIL , all possessing an elongated shape. In order to directly measure the contribution of the different N TAIL boxes to binding, we have studied binding reactions between XD and N TAIL deletion proteins. Changes in surface plasmon resonance were monitored in real time as the N TAIL proteins passed over sensor chips to which XD was covalently coupled. This analytical approach is ideally suited to study reversible, lowaffinity protein-protein interactions that typify interactions involving intrinsically disordered proteins (Wright and Dyson 1999; Dunker and Obradovic 2001; Uversky 2002) . Moreover, reaction rate and equilibrium constants were calculated for reactions with different protein substrates, allowing differences in XD binding affinities to be readily quantified. Binding affinities between XD and N TAIL constructs were established using 180-225 RU of immobilized XD and N TAIL concentrations ranging from 0.1 to 10 mM (see Materials and Methods). Dosage-dependent binding was observed in this range. Reactions conformed to a 1:1 ligand-substrate (Langmuir) binding model, exhibiting an excellent fit (i.e., a 2 value <1 and residuals within the range of 62) following global analysis of sensorgrams. Binding reactions between XD and N TAILHNFC exhibit an equilibrium dissociation constant of 80 nM (see Table 1 ). The XD binding affinity for N TAILD1 (50 nM) is similar to that of N TAILHNFC indicating that Box1 does not participate in binding. In contrast, removal of either Box3 alone or Box2 plus Box3 results in a strong decrease (three orders of magnitude) in the equilibrium dissociation constant, where N TAILD3 and N TAILD2,3 display similar binding affinities (see Table 1 ). The strong decrease in the affinity resulting from removal of Box3 clearly indicates that Box2 is not the sole region involved in binding to XD, and attests that Box3 also plays a role in the interaction with XD, as already suggested by SAXS studies. The far-UV circular dichroism (CD) spectra of N TAIL deletion proteins at neutral pH are typical of unstructured proteins, as seen by their large negative ellipticity at 198 nm and very low ellipticity at 185 nm ( Fig. 4A -C). The solvent 2,2,2-trifluoroethanol (TFE) mimics the hydrophobic environment experienced by proteins in protein-protein interactions, and is therefore widely used as a probe to unveil disordered regions having a propensity to undergo an induced folding (Hua et al. 1998 ). Thus, we have recorded CD spectra of N TAIL deletion proteins in the presence of increasing concentrations of TFE ( Fig. 4A -C). All proteins show an increasing gain of a-helicity upon addition of TFE, as indicated by the characteristic maximum at 190 nm and minima at 208 nm and 222 nm ( Fig. 4A -C). In the case of N TAILD3 and N TAILD1 , most unstructured-to-structured transitions take place in the presence of 20% TFE, a concentration at which the a-helical content is estimated to be about 13% (using the ellipticity at 222 nm) (Fig. 4D ). On the other hand, in the case of N TAILD2,3 , TFE concentrations as high as 30% are required for most pronounced unstructured-to-structured transitions to take place (see Fig. 4B ). Moreover, in the presence of 30% TFE, the a-helical content of N TAILD2,3 is not only lower (11%) than that (16%) of N TAILD1 and N TAILD3 , but also lower than the a-helical content observed at 20% TFE for the two other N TAIL deletion proteins (Fig. 4D) . Therefore, N TAILD2,3 displays the lowest a-helical potential, while N TAILD1 and N TAILD3 exhibit a folding propensity similar to that observed for N TAIL Bourhis et al. 2004 ). These results, beyond confirming that the region spanning residues 489-516 affects the folding potential of N TAIL (Bourhis et al. 2004) , suggest that neither Box1 nor Box3 contribute to the a-helical propensity of N TAIL , in agreement with the secondary structure prediction provided by PSI-PRED (McGuffin et al. 2000) and PHD (Rost 1996) , which both predict an a-helix (residues 489-504, see In order to investigate whether the N TAIL deletion proteins retained the ability to undergo induced folding in the presence of XD, we have used far-UV CD spectroscopy. The N TAILHNFC protein, bearing an N-terminal hexahistidine tag and a C-terminal flag was purified from the soluble fraction of E. coli (data not shown) and used as the reference in these experiments to allow direct comparison with the N TAIL truncated proteins. Noticeably, the CD spectrum of N TAILHNFC is fully superimposable on that of N TAILHN (data not shown), thus ruling out the possibility that the flag sequence might affect the folding properties of the protein. The far-UV CD spectrum of XD ( Fig. 5A -D, gray line) is typical of a structured protein with a predominant ahelical content, as indicated by the positive ellipticity between 185 nm and 200 nm, and by the two minima at 208 nm and 222 nm. After mixing N TAILHNFC with different molar excesses of XD, the observed CD spectra differed from the corresponding theoretical average curves calculated from the individual spectra. Since the theoretical average curves correspond to the spectra that would be expected if no structural variations occur, deviations from these curves indicate structural transitions. The results obtained in the presence of a threefold molar excess of XD indicate a random coil to a-helix transition, as judged by the much more pronounced minima at 208 nm and 222 nm, and by the higher ellipticity at 190 nm of the experimentally observed spectrum compared to the corresponding theoretical average curve (see Fig. 5A ). Although even more dramatic structural transitions of full-length N TAIL have been previously observed with a twofold molar excess of XD , the results obtained with a threefold molar excess have been selected for presentation in order to allow direct comparison with the N TAIL deletion proteins (see below). Figure 5 , panels B-D, show the results obtained for the three N TAIL deletion proteins in the presence of a threefold molar excess of XD, a condition leading to the most dramatic structural transitions. Removal of Box3 significantly reduces, but does not abrogate, the folding ability of N TAIL . Indeed, the experimental CD spectrum does not significantly deviate from the average curve in the 200-260 nm region, but it considerably deviates from the average curve in the 185-195 region (58% mean increase of ellipticity) (Fig. 5B ), thus supporting partial folding ability of N TAILD3 in the presence of XD. Further removal of Box2 results in a truncated N TAIL form, which has completely lost its ability to fold in the presence of XD, as indicated by the good superimposition between the experimental and the average spectra (Fig. 5C) . Conversely, removal of Box1 does not affect the folding ability of N TAIL , as the deviations from the average spectrum are similar to those observed with the full-length form (Fig. 5, cf. D and A) . The mean increase in ellipticity in the 185-195 nm region (77%) and the decrease in the ellipticity value at 220 nm (46%) observed with N TAILD1 , are comparable to the corresponding values observed with N TAILHNFC (70% and 46%, respectively). As a control, we recorded CD spectra of N TAILHNFC in the presence of lysozyme (data not shown). The absence of significant structural variations even with molar excesses as high as 5, confirms the specificity of the deviations observed upon addition of XD to the N TAIL proteins. A two-and threefold molar excess of XD is required to induce the most pronounced structural transitions of fulllength and truncated forms of N TAIL , respectively. In contrast, an equimolar amount of PCT is sufficient to produce the same effect Bourhis et al. 2004) . This difference can be accounted for by a lower affinity of N TAIL towards XD, compared to PCT, rather than to the formation of a 1:2 or 1:3 stoichiometric complex. Indeed, formation of a 1:1 stoichiometric complex between XD and a peptide corresponding to residues 477-505 of N, has been documented using isothermal titration calorimetry (Kingston et al. 2004b ). The higher affinity of N TAIL for PCT compared to XD could be ascribed either to cooperativity phenomena among the different XDs within the PCT tetramer (Rahaman et al. 2004) or to the possible contribution of other PCT regions to binding. This latter possibility can be ruled out based on recent data pointing out XD as the sole PCT region contributing to binding (Kingston et al. 2004b; S. Longhi and M.J. Oglesbee, unpubl.) . In conclusion, these results indicate that (1) Box1 is fully dispensable for binding to XD and induced folding, (2) Box2 is strictly required for induced folding to take place, and (3) Box3 contributes to binding to XD, as pointed out by the reduced ability of N TAILD3 to undergo induced folding. Moreover, the fact that N TAILD3 , N TAILD1 , and N TAIL Bourhis et al. 2004) share similar folding propensities suggests that the contribution of Box3 to the interaction can be accounted for more in terms of binding rather than of induced folding. In order to further characterize the contribution of Box3 to binding, we have used fluorescence spectroscopy. Accordingly, we have designed an N-terminally hexahistidine-tagged N TAIL variant form bearing a tyrosine to tryptophan substitution at position 518 (see also Fig. 3A ). Introduction of a tryptophan residue in Box3 allowed binding events to be followed by fluorescence spectroscopy, while maximizing the conservative nature of the substitution (note that neither N TAIL nor XD contain any tryptophan residue). Most recombinant product was purified from the soluble fraction of the bacterial lysate (Fig. 6A) . The identity of the recombinant product was confirmed by IP studies using anti-N Cl25, and anti-hexahistindine tag mAbs (data not shown). As already observed in the case of full-length N TAIL and N TAIL deletion proteins, N TAILW518 migrates in SDS-PAGE with an apparent MM higher than expected (Fig. 6A) . The mutated protein, displays the same gel filtration elution profile as the wt form (data not shown), leading to an estimated R S of 27 6 3 Å . As shown in Figure 6B , the far-UV CD spectrum of N TAILW518 is almost perfectly superimposable on that of N TAILHN , thus indicating that the introduction of the tryptophan residue does not affect the overall secondary structure content of the protein. In order to investigate whether the variant form retained the ability to undergo induced folding in the presence of XD, we have added various molar excesses (ranging from 1 to 4) of XD to N TAILW518 , and recorded the corresponding far-UV CD spectra (data not shown). These latter studies indicate that the tyrosine to tryptophan substitution does not affect the ability of the protein to undergo induced folding in the presence of the partner, thus supporting the biochemical relevance of this variant form. Fluorescence spectroscopy studies showed that N TAILW518 has a maximum of emission at 356 nm, indicating that Trp 518 is fully exposed to the solvent (data not shown). Addition of gradually increasing XD concentrations triggers an increase in the fluorescence intensity in a dose-dependent manner, which indicates a modification in the pattern of interactions with neighboring groups. At the same time, addition of XD causes a progressive shift in the emission maximum from 356 nm to 352 nm (data not shown), thus indicating that Trp 518 becomes only slightly less exposed to the solvent. No significant variations are observed in the fluorescence spectrum obtained after addition to N TAILW518 of a 2 mM solution of an irrelevant protein (SARS virus, unclassified protein 5) of similar size and devoid of tryptophan residues (data not shown). After plotting the relative fluorescence intensity increase as a function of the XD concentration (Fig. 6C) , an apparent constant equilibrium dissociation (K Dapp ) value of 133 6 32 mM is derived. This value is in good agreement with the value obtained by surface plasmon resonance studies with wt N TAIL . Although the environment of Trp 518 remains mostly polar upon binding to XD, the observed increase in the fluorescence intensity indicates that the chemical environment of Trp 518 is affected, thus further suggesting that the C terminus of N TAIL interacts with XD. In order to further explore the nature of the interaction established between Box3 and XD, we have used NMR spectroscopy. To this endeavor, we have recorded a HSQC spectrum of 15 N uniformly labeled N TAILHN either alone or in the presence of a twofold molar excess of XD, as well as of 15 N uniformly labeled N TAILD3 in the same conditions. This analysis allowed a quantitative estimation of the number of residues involved in the interaction with XD by following chemical shift changes in the backbone amide and proton resonances upon addition of unlabeled XD. Upon addition of XD to N TAILHN , 16 correlation peaks are displaced. Among them, 11 undergo an upfield shift (see Fig. 7 , stars), which indicates a random coil to a-helix transition, and two correspond to the side chain of either a Gln or an Asn. Additionally, at least seven additional peaks undergo a less dramatic displacement (see Fig. 7 , diamonds). Because of its small amplitude, this shift most likely does not reflect a conformational change. Nevertheless, it is indicative of a change in the chemical environment of these residues resulting either by direct interaction with the partner or by local magnetic perturbations of residues spatially close to the newly formed a-helix. Based on this experiment, however, the determination of N TAIL residues involved in these two types of shifts is not possible. The precise identification of these residues would require the full assignment of the N TAIL HSQC spectrum, which is complicated by the strong signal overlapping inherent in the predominantly unfolded nature of N TAIL . Comparison of present results to those of the NMR spectroscopy studies of Kingston et al. (2004a) suggest that the 11 residues undergoing the random coil to a-helix transition are most likely located within the 486-503 region. In order to assess whether the additional N TAIL residues undergoing a less pronounced displacement upon binding to XD are located within Box3, we recorded a HSQC spectrum of 15 N uniformly labeled N TAILD3 either alone or with a twofold molar excess of XD. As shown in Figure 7 , the same peaks undergoing the large displacement in the N TAILHN -XD complex are also observed in the N TAILD3 -XD complex. In particular, among these peaks, the occurrence in both complexes of the 11 peaks that undergo the random coil to a-helix transition (see Fig. 7 , stars), provides further support that the helical folding occurs within Box2. On the other hand, the seven peaks that undergo a less dramatic displacement upon complex formation with XD (see Fig. 7 , diamonds) are not present in the spectrum recorded on the N TAILD3 -XD complex. This latter observation indicates that these seven peaks correspond to residues that are located within Box3. In conclusion, these experiments reveal that complex formation between N TAIL and XD implies two types of interaction: one, mediated by residues belonging to Box2, involves a significant gain of a-helicity, while the other, attributable to Box3 residues, is not accompanied by a significant structural transition. In this paper we show that N TAIL remains predominantly unfolded after binding to XD, with two distinct sites being involved in the interaction. Although the XDinduced gain of regular secondary is restricted to Box2 (aa 489-504), the region encompassing residues 489-504 is not the only N TAIL region involved in binding to XD. In particular, we present several lines of evidence indicating that the extreme C terminus of N TAIL (Box3, aa 517-525) also contributes to binding, without however gaining any regular secondary structure. The C terminus of N TAIL is an additional XD binding site SAXS studies suggest that the C terminus of N TAIL interacts with XD. In particular, the calculated overall envelope of the complex shows the presence of a globular cluster of invariant size at one extremity and of an elongated protuberance with varying shapes. The elongated protuberance corresponds to the 92-residue-long N-terminal region of N TAIL , while the more compact region accommodates the structure of the XD-N TAIL486-505 complex as well as the 20 C-terminal residues of N TAIL . Data from overall shape calculations thus clearly indicate that the C terminus is not protruding towards the solvent, and remains close to XD. On the other hand, attempts to more precisely model the conformation adopted by the C-terminal region of N TAIL within the complex led to an ensemble of solutions. In all these solutions, the C-terminal region of N TAIL always packs against the XD-N TAIL486-505 complex, rather than being extended and exposed to the solvent. Identification of the possible gain of regular secondary structure within the C-terminal region of N TAIL is beyond the resolution limits of SAXS. However, results provided by CD, fluorescence spectroscopy, and heteronuclear NMR all converge to suggest that the C-terminal region of N TAIL does not gain any regular secondary structure (see below). Besides supporting a role for the C-terminal region of N TAIL in the interaction with XD, these data also indicate that most of N TAIL remains disordered within the complex. The prevalent disorder of N TAIL within the complex is in agreement with other data reported in the literature where an intrinsically disordered protein largely preserves its overall extended conformation even after interaction with a binding partner (see Tompa 2002 , and references cited therein; Permyakov et al. 2003) . Finally, it is noteworthy that in the SAXS model the N terminus of XD is exposed to the solvent (see Fig. 2B ), a position that would accommodate the remaining part of P. The involvement of additional N TAIL regions other than the a-MoRE in binding to XD is confirmed by surface plasmon resonance studies, where the contribution of Box3 to XD binding can be quantitatively estimated. Removal of either Box3 alone or Box2 plus Box3 leads to a strong decrease (three orders of magnitude) in the affinity as compared to full-length N TAIL , thus indicating that Box3 contribution to binding is similar to that of Box2. Spectroscopy studies showed that Box3 contributes to binding to XD but does not undergo any gain of regular secondary structure. Although Box3 does not affect the folding potential of N TAIL , as indicated by CD studies in the presence of TFE, the removal of Box3 significantly reduces the ability of N TAIL to undergo induced folding in the presence of XD. This supports a role for Box3 in binding to XD. Further removal of Box2 results in a truncated form that has a significantly decreased folding potential and has lost the ability to undergo induced folding in the presence of XD. These results are consistent with the unique a-helical forming potential of Box2 and its role as primary binding site for XD. Contribution of Box3 to binding without dramatic structural change is supported by fluorescence spectroscopy data, which show an increase in the fluorescence intensity of N TAILW518 upon addition of XD. Increases in the fluorescence intensity upon binding to a partner/ ligand have been already documented, for both folded (Bette et al. 2002) and intrinsically unstructured proteins (Raggett et al. 1998) , and indicate that the chemical environment of the Trp, although remaining mostly polar, is changed as a result of the addition of the partner. However, the fluorescence data do not enable us to discriminate between a direct Trp 518-XD interaction and secondary effects resulting from local perturbations triggered by a direct Box2-XD interaction. That complex formation with XD involves both Box2 and Box3, and that only Box2 undergoes a gain of a-helicity, is confirmed by NMR studies. Heteronuclear NMR experiments show that upon complex formation with XD, 11 N TAIL residues, belonging to Box2, undergo a random coil to a-helix transition, while at least seven additional residues undergo a shift in their chemical environment not accompanied by the gain of regular secondary structure elements. Our data show that these latter peak displacements correspond to residues belonging to Box3, as indicated by the absence of such peaks in both the spectra of N TAILD3 alone and in complex with XD. Thus, the NMR studies, beyond confirming the role of Box3 in the interaction with XD, also highlight that binding of N TAIL to XD implies two types of interaction, where gain of regular secondary structure is restricted to only one of the binding determinants, Box2. The accommodation of the 486-505 region of N within XD (Kingston et al. 2004a ) triggers some minor rearrangements at the surface of the latter, compared to the crystal structure of the uncomplexed form . In particular, the largest movements are observed for the side chains of residues Arg15 and Glu17 of XD (corresponding to residues 472 and 474 of P). The Arg15 NH2 atom moves 9.7 Å away in the chimeric structure compared to the structure of XD alone, while the Glu17 C atom undergoes a shift of 2.5 Å . These two residues are both located within the loop connecting a1 and a2 helices, and occur at least 9 Å away from Box2, thus ruling out the possibility that the observed spatial rearrangements they undergo could be ascribed to Box2 embedding. The significant displacement they undergo in the complex therefore indicates that binding of Box2 induces local conformation changes in XD that could favor interaction with Box3. The surface displaced residues of XD could, in fact, be part of a Box3 binding site. We tentatively propose that binding to XD might take place through a sequential mechanism involving successive binding of disordered domains, as it has been recently reported for p27 upon binding to the CdK2-cyclin A complex (Lacy et al. 2004 ). The K D value between N TAIL and XD, as measured by both fluorescence spectroscopy and surface plasmon resonance, is in the 100 nM range. Surprisingly, this value is considerably lower than that reported by Kingston et al. (2004b) (13 mM) and derived from isothermal titration calorimetry studies. A weak binding affinity, associated with a fast association rate, would ideally fulfill the requirements of a polymerase complex, which has to cartwheel on the nucleocapsid template during both transcription and replication. However, a K D in the mM range would not seem to be physiologi-cally relevant considering the low intracellular concentrations of P in the early phases of infection. Moreover, such a weak affinity is not consistent with the ability to readily purify nucleocapsid-P complexes using rather stringent techniques such as CsCl isopycnic density centrifugation (Robbins and Bussell 1979; Stallcup et al. 1979; Robbins et al. 1980; Oglesbee et al. 1989) . Our data support a higher affinity between P and N, resulting in a stable P-N TAIL complex that would be predicted to hinder processive movement of P along the nucleocapsid template. In agreement with this prediction, Box3 has been shown to have an inhibitory role upon transcription and replication, as indicated by previous minireplicon experiments, where deletion of Box3 enhanced basal reporter gene expression (Zhang et al. 2002) . We can speculate that the transient nature of the N TAIL -XD interaction might be ensured by the possible intervention of cellular and/or viral cofactors. Indeed, the requirement for cellular or viral cofactors in both transcription and replication has been already documented in the case of measles (Vincent et al. 2002) , respiratory syncytial (Fearns and Collins 1999) , and Ebola viruses (Hartlieb et al. 2003) . These cofactors may serve as processivity or transcription elongation factors and could act by modulating the strength of the interaction between the polymerase complex and the nucleocapsid template. Such a mechanism may explain the stimulatory effect of hsp72 on MV transcription and genome replication, an effect mediated by Box3-hsp72 interaction (Zhang 2002) . In this capacity, hsp72 may neutralize the contribution of Box3 to a more stable complex between P and N TAIL , thereby promoting successive cycles of P binding and release that are essential to polymerase processivity. Using different physico-chemical approaches we have shown that the interaction of N TAIL with XD involves an additional, previously unreported site located at the extreme C terminus of N TAIL . While the primary site (i.e., Box2) gains a-helical structure upon binding to XD, this additional site does not gain any regular secondary structure elements. Nevertheless, it plays a crucial role in stabilizing the complex. We have previously shown that N TAIL belongs to the premolten globule subfamily, e.g., it possesses a certain extent of residual secondary and/or tertiary structure Bourhis et al. 2004 ). In agreement with these findings, NMR studies showed that the N TAIL region encompassing residues 486-503 is not a statistical random coil (Kingston et al. 2004a ). These results, supporting a restricted conformational freedom of this N TAIL region, are in agreement with the speculation that residual structure within N TAIL may play a role for efficient binding to P Bourhis et al. 2004) . That residual structure within intrinsically disordered proteins may favor the folding process triggered by binding to a partner has already been reported (Bienkiewicz et al. 2002; Fuxreiter et al. 2004; Csizmok et al. 2005) . N TAIL provides an interesting model system for the study of the interaction of an intrinsically disordered protein and its partners. Its multiple-site mode of interaction well illustrates the complexity of the contacts established by intrinsically disordered proteins and emphasizes the need for thorough investigations of the molecular mechanisms underlying recognition of the partner. Indeed, binding of disordered regions to their targets involves different types of interactions, such as binding coupled to folding, binding with no gain of regular secondary structure, or coexistence of both interaction mechanisms (for a review, see Dyson and Wright 2005) . Binding coupled to folding has been reported for the phosphorylated kinase-inducible domain (pKID) of CREB, which undergoes a coil to a-helix folding transition upon binding to the KIX domain of the transcription coactivator CBP (CREB Binding Protein) (Radhakrishnan et al. 1997) . The amphipathic helix aB of pKID interacts with a hydrophobic groove defined by helices a1 and a3 of the partner. The other pKID helix, aA, contacts a different face of the a3 helix. This mode of interaction is reminiscent of that occurring between N TAIL and XD. Similarities concern the involvement of two distinct sites within the disordered domain (where Box2 and Box3 resemble helices aB and aA, respectively) and embedding of an a-helix within a hydrophobic cleft delimited by a-helices from the structured partner. The difference concerns the mode of interaction of the second site, where Box3 of N TAIL does not gain any regular secondary structure element, contrary to aA of pKID. Nevertheless, complexes with dual interaction have already been described in the literature. Among them, we mention the case of the binding of p27 to the cyclinA-Cdk2 complex (Lacy et al. 2004) , and that of the activation domain of CITED2 to the TAZ1 domain of CBP (De Guzman et al. 2004 ). On the other hand, binding without any concomitant gain of regular secondary structure, has been also reported. This is the case of the unfolded proteins 4E-BP1 (4E binding protein 1), an inhibitor of translation. Its binding to eIF4E does not involve any transition to stable regular secondary structure (Fletcher et al. 1998) . Interestingly, when eIF4E binds to another partner, namely eIF4G, it induces a coil-to-helix transition in the latter (Marcotrigiano et al. 1999) , highlighting the diversity of the structural transitions that can be triggered by a structured partner. Finally, there are a few examples of proteins that undergo different structural transitions as a function of the partner they bind. Notably, HIFa (hypoxia-inducible factor a ) possesses an intrinsically disordered domain, which can interact with two different partners. This unstructured domain adopts an a-helical structure when bound to the TAZ1 domain of CBP. However, binding of the same region of HIFa to an asparagine hydroxylase results in a highly extended conformation, which is required for the enzymatic activity of the latter. These findings provide an elegant example of the plasticity that intrinsically disordered proteins display for different partners involved in different functions. We can speculate on a similar behavior in the case of N TAIL . In particular, Box2 and Box3 of N TAIL interact both with at least two distinct partners, P and Hsp72, and competition between XD and Hsp72 for binding to N TAIL has been recently shown (Zhang et al. 2005) . It is conceivable that binding of N TAIL to P or to Hsp72 may involve a different structural transition, as in the case of HIFa. Last, but not least, our findings, beyond contributing to elucidate the dynamics of the interactions established by intrinsically disordered proteins, provide an interesting target site for mutational studies aimed at exploring further the mode of interaction between N TAIL and XD. The E. coli strains DH5a (Stratagene) was used for selection and amplification of DNA constructs. The E. coli strains Rosetta [DE3] pLysS (Novagen) and C41 [DE3] (Avidis) were used for expression of recombinant proteins. E. coli was grown either in Luria-Bertani (LB) medium, or in minimal M9 medium supplemented with 15 NH 4 Cl. Pfu polymerase was from Promega. Primers were purchased from Invitrogen. The anti-hexahistidine tag mAb was purchased from Qiagen. The anti-flag mAb was purchased from Sigma. The anti-N Cl 25 (Giraudon et al. 1988; Buckland et al. 1989 ) mAb was kindly provided by D. Gerlier. The XD gene construct, encoding residues 459-507 of the MV P protein (strain Edmonston B) with an hexahistidine tag fused to its C terminus, has already been described . All N TAIL constructs were obtained by PCR using the MV N gene, strain Edmonston B, as template. The N TAILHN gene construct, encoding residues 401-525 of the MV N protein with a hexahistidine tag fused to its N terminus, was obtained using the plasmid pET21a/N (encoding the MV N protein; Karlin et al. 2002a ) as template. Forward primer (5 0 -gatagaac catgCATCATCATCATCATCATactactgaggacaagatcagtaga-3 0 ) was designed to introduce a hexahistidine tag encoding sequence (upper case) at the N terminus of N TAIL , while reverse primer (5 0 -ggggaccactttgtacaagaaagctgggtcttagtctagaa gatttctgtcattgta-3 0 ) was designed to introduce an AttB2 site (bold). The PCR amplification product was further used as template in a second PCR step, using forward primer (5 0 -ggggacaagtttgtacaaaaaagcaggcttcgaaggagatagaaccatgCAT CATCATCAT-3 0 ), designed to introduce an AttB1 site (bold), and reverse primer as above. After purification (PCR Purification Kit; Qiagen), the PCR product was cloned into the pDest14 vector (Invitrogen) using the Gateway recombination system (Invitrogen). The final construct is referred to as pDest14/N TAILHN . The N TAILHNFC gene construct, encoding residues 401-525 of the MV N protein with an N-terminal hexahistidine tag and a C-terminal Flag sequence (dykddddk) (Brizzard et al. 1994) , was obtained using the plasmid pDest14/N TAILHN as template. Forward primer (5 0 -ggggacaagtttgtacaaaaaagcaggcttcgaaggagata gaaccatgCATCATCATCAT-3 0 ) was designed to introduce an AttB1 site (bold), and reverse primer (5 0 -gtcttaTTTGTCGTCAT CGTCTTTATAATCgtctagaagatttctgtcattgta-3 0 ) was designed to introduce a Flag encoding sequence (upper case). The PCR amplification product was further used as template in a second PCR step, using the same forward primer as above, and reverse primer (5 0 -ggggaccactttgtacaagaaagctgggtcttaTTTGTCGTCAT CGTCTTT-3 0 ), which was designed to introduce an AttB2 site (bold). After purification, the PCR product was cloned into the pDest14 vector to yield pDest14/N TAILHNFC . The N TAILD3 gene construct, encoding residues 401-516 of the MV N protein with an N-terminal hexahistidine tag plus a C-terminal Flag sequence, was obtained using the plasmid pet21a/N as template. Forward primer (5 0 -gatagaaccatgCA TCATCATCATCATCATactactgaggacaagatcagtaga-3 0 ) was designed to introduce a hexahistidine tag encoding sequence (upper case) at the N terminus of N TAIL , and reverse primer (5 0 -gtcttaTTTGTCGTCATCGTCTTTATAATCtataggggtgtcc gtgtctgagcc-3 0 ) was designed to introduce a Flag encoding sequence. The PCR amplification product was further used as template in a second PCR step, using forward primer (5 0 -ggggacaagtttgtacaaaaaagcaggcttcgaaggagatagaaccatgCATCA TCATCAT-3 0 ) and reverse primer (5 0 -ggggaccactttgtacaa gaaagctgggtcttaTTTGTCGTCATCGTCTTT-3 0 ), which were designed to introduce an AttB1 and an AttB2 site (bold), respectively. After purification, the PCR product was cloned into the pDest14 vector to yield pDest14/N TAILD3 . The N TAILD2,3 gene construct, encoding residues 401-488 of the MV N protein with an N-terminal hexahistidine tag plus a C-terminal Flag sequence, previously referred to as N TAIL2 (Bourhis et al. 2004) , was obtained using the plasmid pet21a/ N as template. Forward primer (5 0 -gatagaaccatgCATCATCA TCATCATCATactactgaggacaagatcagtaga-3 0 ) was designed to introduce a hexahistidine tag encoding sequence (upper case) at the N terminus of N TAIL , and reverse primer (5 0 -gtctta TTTGTCGTCATCGTCTTTATAATCactgtcctgcggatcttggctg ga-3 0 ) was designed to introduce a Flag encoding sequence (upper case). The PCR amplification product was further used as template in a second PCR step, using the same pair of primers as in the second PCR step, which yielded the N TAILD3 amplification product. After purification, the PCR product was cloned into the pDest14 vector to yield pDest14/ N TAILD2,3 . The N TAILD1 gene construct, encoding residues 421-525 of the MV N protein with an N-terminal hexahistidine tag plus a C-terminal Flag sequence, was obtained using the plasmid pet21a/N as template. Forward primer (5 0 -gatagaaccatgCA TCATCATCATCATCATcacggtgatcaaagtgagaatgag-3 0 ) was designed to introduce a hexahistidine tag encoding sequence (upper case) at the N terminus of N TAIL , and reverse primer (5 0 -gtcttaTTTGTCGTCATCGTCTTTATAATCgtctagaagattt ctgtcattgta-3 0 ) was designed to introduce a Flag-encoding sequence (upper case). The PCR amplification product was further used as template in a second PCR step, using the same pair of primers as in the second PCR step, which yielded both N TAILD3 and N TAILD2,3 amplification products. After purification, the PCR product was cloned into the pDest14 vector to yield pDest14/N TAILD1 . The N TAILW518 gene construct, encoding residues 401-525 of the MV N protein with a Tyr ! Trp substitution at position 518 and with a hexahistidine tag fused to its N terminus, was obtained by PCR, using the plasmid pET21a/N as template. Two separate PCR steps were carried out in parallel, yielding amplification products A and B. Product A was obtained using forward primer (5 0 -gatagaaccatgCATCAT CATCATCATCATactactgaggacaagatcagtaga-3 0 ), designed to introduce a hexahistidine tag-encoding sequence (upper case) at the N terminus of N TAIL , while reverse primer (5 0 -aaga tttctgtcattCCAcactatTggggtgtc-3 0 ) was designed to introduce a Trp at position 518 (bold and upper case) and a silent mutation at nucleotide position 1545 (upper case) thus resulting in the introduction of a BstXI site (underlined). Product B was obtained using forward primer (5 0 -gacaccccAatagtgTGG aatgacagaaatctt-3 0 ) designed to introduce a Trp at position 518 (bold and upper case) and a silent mutation at nucleotide position 1545 (upper case) thus resulting in the introduction of a BstXI site (underlined), and reverse primer (5 0 -ccgggcatgc atccggatatagttcctcctt-3 0 ) designed to anneal with nucleotide positions 1755-1775 of pet21a/N (where the A of the ATG codon of the N ORF was set as nucleotide position 1). A further PCR step was carried out using amplification products A plus B as template, to yield the N TAIL W518 amplification product. Forward primer (5 0 -ggggacaagtttgtacaaaaaagcaggct tcgaaggagatagaaccatgCATCATCATCAT-3 0 ) was designed to introduce an AttB1 site (bold), while reverse primer (5 0 -ggggac cacttttgtacaagaaagctgggtcttagtctagaagatttctgtcattCCA-3 0 ) was designed to introduce an AttB2 site (bold) and a Trp at position 518 (upper case). After purification, the PCR product was cloned into the pDest14 vector using the Gateway recombination system. The final construct is referred to as pDest14/ N TAILW518 . Candidate clones bearing the desired mutation were selected on the basis of the ability of their recombinant plasmids to be restricted by BstXI, a unique restriction site introduced by PCR together with the Tyr to Trp substitution. The sequence of the coding region of all expression plasmids was verified by sequencing (MilleGen). E. coli strain Rosetta [DE3] (Novagen) was used for the expression of N TAIL constructs. Since the MV N gene contains several rare codons that are used with a very low frequency in E. coli, coexpression of N TAIL constructs with the plasmid pLysS (Novagen) was carried out. This plasmid, which supplies six rare tRNAs, carries also the lysozyme gene, thus allowing a tight regulation of the expression of the recombinant gene, as well as a facilitated lysis. Cultures were grown overnight to saturation in LB medium containing 100 mg/mL ampicilin and 17 mg/mL chloramphenicol. An aliquot of the overnight culture was diluted 1/25 in LB medium and grown at 37 C. At OD 600 of 0.7, isopropyl b-D-thiogalactopyranoside (IPTG) was added to a final concentration of 0.2 mM, and the cells were grown at 37 C for 3 h. The induced cells were harvested, washed, and collected by centrifugation. The resulting pellets were frozen at À20 C. Isotopically substituted ( 15 N) N TAILHN and N TAILD3 were prepared by growing transformed bacteria in minimal M9 medium supplemented with 15 NH 4 Cl (0.8 g/L). A 50-mL preculture grown overnight to saturation in LB medium containing 100 mg/mL ampicillin and 17 mg/mL chloramphenicol, was harvested, washed in minimal M9 medium, and inoculated into 1 L of minimal M9 medium supplemented with ampicillin and chloramphenicol. The culture was grown at 37 C. At OD 600 of 0.5, IPTG was added to a final concentration of 0.2 mM and the cells were grown first at 37 C for 3 h, and then over night at 28 C. The induced cells were harvested, washed, and collected by centrifugation. The resulting pellets were frozen at À20 C. Expression of tagged XD was carried out as described in Johansson et al. (2003) . Cellular pellets from bacteria transformed with the different N TAIL expression plasmids were resuspended in 5 volumes (v/ w) buffer A (50 mM sodium phosphate at pH 8, 300 mM NaCl, 10 mM Imidazole, 1 mM phenyl-methyl-sulphonyl-fluoride [PMSF]) supplemented with lysozyme 0.1 mg/mL, DNase I 10 mg/mL, protease inhibitor cocktail (Sigma) (50 mL/g cells). After a 20-min incubation with gentle agitation, the cells were disrupted by sonication (using a 750 W sonicator and four cycles of 30 sec each at 60% power output). The lysate was clarified by centrifugation at 30,000g for 30 min. Starting from a 1-L culture, the clarified supernatant was incubated for 1 h with gentle shaking with 4 mL Chelating Sepharose Fast Flow Resin preloaded with Ni 2+ ions (Amersham Pharmacia Biotech), previously equilibrated in buffer A. The resin was washed with buffer A, and the N TAIL proteins were eluted in buffer A containing 250 mM imidazole. Eluates were analyzed by SDS-PAGE for the presence of the desired product. The fractions containing the recombinant product were combined, and concentrated using Centricon Plus-20 (molecular cutoff, 5000 Da) (Millipore). The proteins were then loaded onto a Superdex 75 HR 10/30 column (Amersham Pharmacia Biotech) and eluted in either 10 mM sodium phosphate at pH 7 or 10 mM Tris/HCl at pH 8. The proteins were stored at À20 C. Purification of histidine-tagged XD was carried out as described in Johansson et al. (2003) . All purification steps, except for gel filtrations, were carried out at 4 C. Apparent molecular mass of proteins eluted from gel filtration columns was deduced from a calibration carried out with LMW and HMW calibration kits (Amersham Pharmacia Biotech). The theoretical Stokes radii (R S ) of a native (R S N) and fully unfolded (R S U) protein with a MM (in Daltons) were calculated according to (Uversky 1993) : log(R S N) = 0.369 Á log(MM) À 0.254 and log(R S U) = 0.533 Á log(MM) À 0.682. Protein concentrations were calculated either using the theoretical absorption coefficients e (mg/mL Á cm) at 280 nm as obtained using the program ProtParam at the EXPASY server (http://www.expasy.ch/tools), or the Biorad protein assay reagent (Bio-Rad). IP experiments were carried out using the anti-N Cl 25, the anti-flag, and the anti-hexahistidine tag mAbs, and bacterial lysates expressing N TAIL proteins as described in Longhi et al. (2003) . Small angle X-ray scattering All protein samples were prepared by dilution of the purified N TAILHN and XD solutions in buffer 10 mM Tris/HCl at pH 8, 40 mM NaCl, with 1 mM DTT as radiation scavenger. The complex N TAILHN -XD was prepared by mixing XD and N TAILHN with a molar ratio of 2:1 in the same buffer and at a final protein concentration of 10 mg/mL. The samples were filtered prior to each measurement (Millex syringe filters 0.22 mm, Millipore) to eliminate possibly existing large aggregates. SAXS experiments were carried out on beamline ID02 (Narayanan et al. 2001 ) at the European Synchrotron Radiation Facility (ESRF), Grenoble, France. The wavelength was 1.0 Å and the sample-to-detector distance was 3.0 m and 1.0 m, leading to scattering vectors q ranging from 0.02 to 0.20 Å À1 and 0.05 to 0.40 Å À1 , respectively. The scattering vector is defined as q = 4 p/l sin, where 2 is the scattering angle. The detector was an X-ray image intensified optically coupled to an ESRF developed FReLoN CCD camera. Forty successive frames of 1.0 sec with a 5-sec pause between each frame were recorded for each sample. The protein solution was circulated through an evacuated quartz capillary between each frame. Thus, no protein solution was irradiated longer than 1.0 sec. Each frame was then carefully inspected to check for possible bubble formation or radiation-induced aggregation. No such effect was observed, and individual frames could then be averaged. Absolute calibration was made with a Lupolen sample. A series of measurements at different protein concentrations ranging from 1.8 to 10 mg/mL were performed for every protein (XD, N TAILHN , and the mixture N TAILHN -XD) to check for interparticle interaction. Background scattering was measured before or after each protein sample using the buffer solution and then subtracted from the protein scattering patterns after proper normalization and correction from detector response. All the experiments were carried out at 20 C. The data acquired at both sample-to-detector distances of 3 m and 1 m were merged and extrapolated to zero concentration for the calculations using the entire scattering spectrum. The scattering pattern of the N TAILHN -XD complex was obtained following this process: to avoid any possible bias on the absolute intensities due to the concentration, the experimental merged scattering curves obtained as described above were normalized by their theoretical I(0). The scattering pattern of XD was then subtracted twice from the scattering pattern of the mixture N TAILHN -XD (molar ratio 1:2). Given the rather low K D between N TAIL and XD ($100 nM; see Results section), we assumed that the concentration of uncomplexed N TAILHN in solution is negligible at the concentration used in the SAXS experiments (2 to 10 mg/mL). The scattering pattern of the complex thus obtained was then used for further calculations. The value of the R g was derived from the Guinier approximation (Guinier and Fournet 1955) : I(q) = I(0) exp(Àq 2 R g 2 /3), where I(q) is the scattered intensity and I(0) is the forward scattered intensity. The R g and I(0) are inferred, respectively, from the slope and the intercept of the linear fit of Ln[I(q)] versus q 2 at low q values (q R g < 1.0). The distance distribution function P(r) is the histogram of all the interatomic distances within a molecule. This function also provides the maximum dimension D max of the molecule, which is defined as the point where P(r) becomes zero. The P(r) function was calculated by the Fourier inversion of the scattering intensity I(q) using GNOM (Svergun 1992) and GIFT (Bergmann et al. 2000) on the entire scattering spectra. The low-resolution shape of the N TAILHN -XD complex was determined ab initio from the scattering curve using the program GASBOR ). This program restores lowresolution shapes of protein and calculates a volume filled with densely packed spheres (dummy residues) fitting the experimental scattering curve by a simulated annealing minimization procedure and considering the protein as an assembly of dummy residues centered on the C a positions (spheres of 3.8 Å diameter) with a nearest-neighbour distribution constraint. Several independent fits were run with no symmetry restriction with an input of 188 dummy residues, corresponding to the 56 residues of X D and 132 residues of N TAILHN . The program package CREDO (CHADD and GLOOPY) was used to restore the low resolution model of N TAILHN in complex with XD, with the crystal structure of the chimeric protein between XD and the region of N TAIL encompassing residues 486-505 (XD-N TAIL486-505 ) (Kingston et al. 2004a ) as template. CREDO is an extension of the original program GASBOR. It calculates the structure of missing domains or loops of crystal structures from the experimental scattering curve of the entire particle and represents them by an ensemble of dummy residues forming a chain-compatible model. Binding between purified XD and purified N TAIL proteins was analyzed by using BIAcore 3000 (Amersham Pharmacia Biotech), a system for real-time biomolecular interaction analysis that is based upon surface plasmon resonance technology. Purified XD (1.4 mg/mL in acetate buffer at pH 5.5) was covalently bound to carboxy-methyl groups of CM5 sensor chips using amine-coupling chemistry (Biosensor AB, Amersham Pharmacia Biotech). The levels of immobilized XD were comprised between 180 and 225 RU (1000 RU equal a change in mass of 1 ng/mm 2 on the sensor surface) (Zhang et al. 2002) . Kinetic and equilibrium constants were calculated from global analysis of reactions with multiple analyte concentrations (0.1, 0.2, 0.5, 1, 2, 5, and 10 mM). Reactions were performed at 25 C. Remaining flow channels on the sensor chip included a control for a nonspecific interaction with the sensor chip (i.e., activated/ blocked flow channel) and a control for nonspecific interactions with irrelevant protein targets (i.e., lactoferrin conjugated flow channel). Protein analytes were passed over the sensor chip in HBS-P buffer (0.01 M HEPES at pH 7.4, 0.15 M NaCl, 0.005% surfactant P-20) containing 2.5 mM magnesium acetate and 2.5 mM ATP. Sensorgrams plotted changes in surface plasmon resonance (measured in RU) as a function of time. Multiple sensorgrams representing various analyte concentrations were analyzed by using BIAevaluation 3.1 software. Background interaction of N TAIL proteins (i.e., analytes) with the sensor surfaces were measured on flow channels that were activated and subsequently blocked under buffer conditions used to immobilize XD. This background was subtracted from all binding curves (i.e., sensorgrams) prior to global analyses. Immobilized lactoferrin was used as a specificity control, and the resultant sensorgrams ruled out high affinity interactions between N TAIL constructs or peptides and this irrelevant protein ligand (data not shown). Global fitting of experimental data to well-characterized binding reactions was used to define reaction rate and equilibrium constants. Signal changes on the activated/ blocked control channel were subtracted from the peptide-XD and N TAIL protein-XD interactions using in-line reference and the subtracted sensorgrams were analyzed. Curves generated with serial analyte concentrations were applied globally to the 1:1 Langmuir binding model with or without correction for baseline drifting depending on baseline status. 2 and residual values were used to evaluate the quality of fit between experimental data and individual binding models. Plots of residuals indicate the difference between the experimental and reference data for each point in the fit. The 2 value represents the sum of squared differences between the experimental data and reference data at each point. A good fit between experimental and reference data has small residuals in the À2 to +2 range that randomly distribute about the X-axis, and 2 values are less than 10. The CD spectra were recorded on a Jasco 810 dichrograph using 1-mm thick quartz cells in 10 mM sodium phosphate (pH 7) at 20 C. Structural variations of N TAIL proteins were measured as a function of changes in the initial CD spectrum upon addition of either increasing concentrations of TFE (Fluka), or different amounts of XD or lysozyme (Sigma). CD spectra were measured between 185 and 260 nm, at 0.2 nm/min and were averaged from three independent acquisitions. Mean ellipticity values per residue ([Q]) were calculated as [Q] = 3300 m DA/(l c n), where l (path length) = 0.1 cm, n = number of residues, m = molecular mass in daltons, and c = protein concentration expressed in mg/mL. Number of residues (n) are 140 for N TAILHNFC , 131 for N TAILD3 , 103 for N TAILD2,3 , 120 for N TAILD1 , 132 for N TAIL W518, 56 for XD, and 129 for lysozyme, while m values are 15,626 Da for N TAILHNFC , 14,523 Da for N TAILD3 , 11,539 Da for N TAILD2,3 , 13,440 Da for N TAILD1 , 6690 Da for XD, and 14,300 Da for lysozyme. Protein concentrations of 0.1 mg/mL were used when recording spectra of both individual and protein mixtures. In the case of protein mixtures, mean ellipticity values per residue ([Q]) were calculated as [Q] = 3300 DA/{[(c 1 n 1 / m 1 ) + (c 2 n 2 /m 2 )] l}, where l (path length) = 0.1 cm, n 1 or n 2 = number of residues, m 1 or m 2 = molecular mass in daltons, and c 1 or c 2 = protein concentration expressed in mg/mL for each of the two proteins in the mixture. The theoretical average ellipticity values per residue ([Q] Ave ), assuming that neither unstructured-to-structured transitions nor secondary structure rearrangements occur, were calculated as follows: [Q] Ave = [([Q] 1 n 1 ) + ([Q] 2 n 2 R)]/(n 1 + n 2 R), where [Q] 1 and [Q] 2 correspond to the measured mean ellipticity values per residue, n 1 and n 2 to the number of residues for each of the two proteins, and R to the excess molar ratio of protein 2. The a-helical content was derived from the ellipticity at 220 nm as described in Morris et al. (1999) . Fluorescence intensity variations of the single tryptophan in N TAILW518 was measured by using a Cary Eclipse (Varian) equipped with a front-face fluorescence accessory at 20 C, by using 2.5 nm excitation and 10 nm emission bandwidths. The excitation wavelength was 290 nm and the emission spectra were recorded between 300 nm and 540 nm. Titrations were performed in a 1-mL quartz fluorescence cuvette containing 1 mM N TAILW518 in 10 mM sodium phosphate buffer at pH 7, and by gradually increasing the XD concentration from 0.05 mM to 3 mM. Experimental fluorescence intensities were corrected by subtracting the spectrum obtained with XD alone (note that XD is devoid of tryptophan residues). Data were analyzed by plotting the relative fluorescence intensities at the maximum of emission at increasing XD concentrations. The dissociation equilibrium constant (K Dapp ) value was determined from data fitted to a single exponential equation, by using the PRISM 3.02 nonlinear regression tool (GraphPad). 2D-HSQC spectra (Mori et al. 1995) were recorded on a 500-MHz DRX Bruker spectrometer either on 0.5 mM uniformly 15 N-labeled N TAILHN or on 0.125 mM uniformly 15 N-labeled N TAILD3 in 10 mM sodium phosphate buffer at pH 7.0 containing 10% D 2 O (v/v) alone or after addition of a twofold molar excess of XD. The temperature was set to 283 K and the spectra were recorded with 2048 complex points in the directly acquired dimension and 256 points in the indirectly detected dimension, for 6 h each. Solvent suppression was achieved by the WATERGATE 3-9-19 pulse (Piotto et al. 1992) . The data were processed using the UXNMR software; they were multiplied by a sine-squared bell and zero-filled to 1 K in first dimension prior to Fourier transformation. For the spectra recorded after addition of XD, the number of scans was multiplied by 4. The accession number of MV N is P04851. Secondary structure predictions were carried out using both PSI-PRED (McGuffin et al. 2000) and PHD (Rost 1996) . L. Kingston for kindly providing us with the PDB file of the chimeric construct between XD and the N TAIL peptide (PDB code 1T6O). This study has been carried out with financial support from the Commission of the European Communities, specific RTD program ''Quality of Life and Management of Living Resources,'' QLK2-CT2001-01225, ''Towards the design of new potent antiviral drugs: Structure-function analysis of Paramyxoviridae RNA polymerase.'' It does not necessarily reflect its views, and in no way anticipates the Commission's future policy in this area. This work was also in part supported by the CNRS and by funds from the National Institute of Neurological Disorders and Stroke (R01 NS31693).

A survey of knowledge, attitudes and practices towards avian influenza in an adult population of Italy

BACKGROUND: Several public health strategic interventions are required for effective prevention and control of avian influenza (AI) and it is necessary to create a communication plan to keep families adequately informed on how to avoid or reduce exposure. This investigation determined the knowledge, attitudes, and behaviors relating to AI among an adult population in Italy. METHODS: From December 2005 to February 2006 a random sample of 1020 adults received a questionnaire about socio-demographic characteristics, knowledge of transmission and prevention about AI, attitudes towards AI, behaviors regarding use of preventive measures and food-handling practices, and sources of information about AI. RESULTS: A response rate of 67% was achieved. Those in higher socioeconomic classes were more likely to identify the modes of transmission and the animals' vehicles for AI. Those older, who knew the modes of transmission and the animals' vehicles for AI, and who still need information, were more likely to know that washing hands soap before and after touching raw poultry meat and using gloves is recommended to avoid spreading of AI through food. The risk of being infected was significantly higher in those from lower socioeconomic classes, if they did not know the definition of AI, if they knew that AI could be transmitted by eating and touching raw eggs and poultry foods, and if they did not need information. Compliance with the hygienic practices during handling of raw poultry meat was more likely in those who perceived to be at higher risk, who knew the hygienic practices, who knew the modes of transmission and the animals' vehicles for AI, and who received information from health professionals and scientific journals. CONCLUSION: Respondents demonstrate no detailed understanding of AI, a greater perceived risk, and a lower compliance with precautions behaviors and health educational strategies are strongly needed.

The first known direct avian to human transmission of influenza A (subtype H5N1) viruses was reported during an outbreak in Hong Kong in 1997 and exposure to infected poultry was identified as the probable route of transmission [1] [2] [3] . Since then, outbreaks of the H5N1 highly pathogenic avian influenza strain have been identified in birds, wild and domestic poultry, in several countries, particularly in Vietnam, Indonesia, Thailand, China, Cambodia and, more recently, in Turkey and Iraq. In Italy, no human cases have been reported and two epidemics occurred in poultry causes by avian influenza virus H5 and H7 subtypes. Alongside these massive avian outbreaks, the World Health Organization (WHO) reported more than 300 confirmed human cases of avian influenza A (H5N1), approximately two thirds of whom have subsequently died [4] . Nearly all of these cases are traceable to exposure to infected poultry or birds, but there has not yet been a mutation allowing the H5N1 and H7N7 viruses to spread efficiently in human [5] . However, concern is widespread that the current situation favors the emergence of a highly pathogenic influenza virus with the ability for efficient transmission from person to person, particularly in the presence of a mutation in the viral genome leading eventually another pandemic human influenza. Several public health strategic interventions are required for effective disease prevention and control of the multifaceted issues posed by avian influenza [6] . Of these interventions, it is necessary to create a communication plan to keep the population fully and adequately informed on how to avoid or reduce exposure. A similar approach has been applied during the SARS epidemic [7] and the implementation of appropriate infection control measures was a key aspect for its control. In the past years a limited number of studies have been published investigating knowledge, attitudes, and practice about avian influenza among target groups [8, 9] and general population [10] [11] [12] [13] . This area of investigation seems to be an important one because members of the public often misinterpret their risk of health problems. Therefore, the objectives of the present investigation were to assess the knowledge, attitudes, and behaviors relating to avian influenza and to evaluate the effect of several potential predictors on such outcomes of interest in an adult population in Italy. This cross-sectional survey was conducted from December 2005 to February 2006 in the geographic area of Naples (Italy). A two-stage cluster sampling technique was employed to draw the required sample. In the area surveyed there were 40 schools and each school was considered a cluster. The first stage consisted of selecting four clusters through random sampling. The second stage consisted of randomly select 255 adults from the parents' files of each sampled school that contained 500 students. The questionnaires used for this study were handed out, in sealed envelopes, to the Referent of the health education activities in each school with instructions to distribute one to each family. Families received an information sheet which explained the purpose of the project and requested that the survey be completed by one parent only, a self-administered anonymous questionnaire, and an envelope to facilitate the return of the completed questionnaire. In addition, the letter assured parents about the anonymity and confidentiality of all responses. Participants were asked to return the completed questionnaires to the school personnel anonymously via the envelope enclosed with each questionnaire. Participation was on a voluntary basis and all participants had the right to comply with or refuse participation. The response to questionnaire constituted the participants' informed consent. The study was approved by the Institutional Ethics Review Board. The survey, that was a modification of an instrument previously used [8] , was arranged in different sections enquiring about participants' demographic and socioeconomic characteristics, knowledge of the definition, modes and vehicles of transmission, risk groups, and preventive measures about avian influenza, attitudes towards avian influenza, behaviors regarding use of preventive measures and food-handling practices, whether they had ever received advice and information about avian influenza and, if so, the sources. The response choices for all knowledge questions were given on a three-point Likert-type scale using "yes", "no", "do not know" options for the modes and vehicles of transmission, and risk groups and "agree", "uncertain", and "disagree" for the measures concerning the preventive measures; whereas, the response choice for the question about the knowledge of the definition was open. The responses for all statements relating to attitudes towards avian influenza, to ascertain level of agreement or disagreement were given on a three-point Likert-type scale (from 1 to 3, 1 = agree, 2 = uncertain, 3 = disagree); in two questions respondents were also asked to use a ten-point Likert-type scale to assess perceived degree of risk for contracting avian influenza for him/her and for friends/familiars with responses ranged from 1 (low risk) to 10 (high risk). Five possible categories of responses were allowed, ranging from never to always, to measure compliance with recommendations of the WHO to avoid spread of avian influenza through food [6] . We have also asked whether the respondent and/or member(s) in the household, in the three months preceding the survey, have modified the habits in buying foods or in dietary habits for fear of contracting avian influenza. The original version of the survey instrument underwent a pilot study among a convenient group of 50 subjects to ensure practicability, validity, and interpretation of answers. On the basis of the comments obtained, the questionnaire was revised in item, wording, and format before distribution to the study sample. Multivariate stepwise logistic and linear regression analyses investigated the independent contribution of potential predictors to the following primary outcomes of interest: knowledge about the main modes of transmission (animal-to-animal and animal-to-human) and the animals classified as common vehicles (poultry and birds) for avian influenza (model 1); knowledge that wash hands with soap before and after touching raw poultry meat and use of gloves is a hygienic practices to avoid spreading of the avian influenza virus through food (model 2); wash hands with soap before and after touching raw poultry meat and use of gloves (model 3); perception of risk of contraction avian influenza for him/her (model 4). For the purposes of analysis, the outcome variables originally consisting of multiple categories in the logistic analysis were collapsed into two levels. In Model 1, respondents were divided into those who knew the main modes of transmission (animal-to-animal and animal-to-human) and the animals classified as common vehicles (poultry and birds) for avian influenza versus all others; in Model 2, those who knew that wash hands with soap before and after touching raw poultry meat and use of gloves is a hygienic practices to avoid spreading of the avian influenza virus through food versus all others; in Model 3, those that wash hands with soap before and after touching raw poultry meat and use gloves versus all others. In all models the following independent explanatory variables were included: age, gender, marital status, educational level, number of children, employment status, health professionals and reading scientific journals as sources of information, and need of additional information. Others independent explanatory variables were also included in the different models: perception of risk of contraction avian influenza (models 1-3); knowledge about the modes of transmission of avian influenza (models 2 and 3); correct definition of avian influenza and know that avian influenza could be transmitted by eating and touching raw eggs and poultry foods (model 4); knowledge that wash their hands with soap before and after touching raw poultry meat and use of gloves is a hygienic practices to avoid spreading of the avian influenza virus through food (model 3). Before testing multivariable logistic regression models assessing predictors of the outcomes of interest, we examined correlations to assess collinearity among the independent variables and bivariate relations between the independent variables and the dependent variable. The criterion to be met before any independent variable was considered for entry into an initial multivariable logistic regression model was a p-value ≤ 0.25 obtained for each outcome variable in the univari-ate analysis and noncollinear with other predictors. Furthermore, the significance level for variables entering the logistic regression models was set at 0.2 and for removing from the model at 0.4. Odds ratios (ORs) and their corresponding 95% confidence intervals (CIs) were calculated in the model for the independent variables. When building linear regression model, we have first included only one possible variable at a time. Then, using the variables that were significant at p-value ≤ 0.25, we constructed a stepwise multivariate linear regression model and the significance level for variables entry the model was set at 0.2 and for removal at 0.4. Statistical significance level was defined as a two-tailed p-value ≤ 0.05. Stata version 8.1 software program was used for all statistical analyses [14] . The sample consisted of 683 individuals for a participation rate, defined as the number of completed questionnaires divided by the number of those randomly selected, of 67%. Socio-demographic characteristics of the respondents are reported in Table 1 . The average age was 40.7 years, two thirds were female, almost all were married, the majority had not reached college level education, more than half is inactive or housewife, and one-third has three or more children. Table 3 ). Respondents did not recognize the major risk groups, since a large percentage agreed that poultry workers (88%) were at risk, but lower values were reported for butchers (55.1%), hunters (30.7%), and veterinarians (23.6%). Moreover, 34.6% knew that washing their hands with soap before and after touching raw poultry meat and using gloves is a hygienic practice to avoid spreading of the avian influenza virus through food. Those older (OR = 1.03; 95% CI 1.01-1.05), who knew the modes of transmission and the common vehicles for avian influenza (OR = 1.63; 95% CI 1.11-2.39), and who still need additional information about avian influenza (OR = 1.52; 95% CI 1.09-2.12) were more likely to know this practice (Model 2 in Table 3 ). More than half of the respondents thought that avian influenza was a serious disease (61.9%) and that it was possible to prevent (53.3%). The respondents' level of perceived risk of contracting avian influenza for them and for friends/familiars resulted in a mean total score respectively of 5.9 ± 2.9 and 6.2 ± 2.8, indicating a high risk perception with respectively 19.3% and 20.4% of the respondents having reported feeling "very much" at risk by answering "10". A stepwise multivariate linear regression model was constructed to search for associations with the attitude, aiming at understanding which variable had stronger associations with the perception of risk of contraction avian influenza by the respondent. Respondents considered the risk for them of being infected significantly higher if they were from lower socioeconomic classes, had lower educational level, if they did not know the definition of avian influenza, if they knew that avian influenza could be transmitted by eating and touching raw eggs and poultry foods, and if they believed that they did not need additional information about the disease (Model 4 in Table 3 ). Participants were asked about their behaviors regarding the use of preventive measures and food-handling practices. Approximately two-thirds reported that, in the three months preceding the survey, at least one member in the household had modified the habits in buying foods (63.3%) or in dietary (59.9%) for fear of contracting avian influenza. As regards hygienic practices to avoid spreading of the virus through food, among those who prepared food, only 26.9% reported washing their hands with soap before and after touching raw poultry meat and using gloves. Multivariable logistic regression analysis indicated that compliance with the hygienic practices was more likely by those who perceived a higher risk of contracting avian influenza (OR = 1.13; 95% CI 1.06-1.19), by those Table 3 ). Almost all respondents recalled receiving some information about avian influenza (97.9%) mostly through mass media (85.8%), health professionals (26.5%), and scientific journals (8.4%). A vast majority (65%) reported interest in receiving further information on avian influenza. The results of the present survey depict a mosaic of opinions outlining the stated knowledge, attitudes, and selfreported behavior patterns concerning avian influenza among a large cross-section of a random sample of an adult population in one region of Italy. Guidelines and recommendations have been developed to prevent and control the spread of avian influenza at source and in responding to the pandemic threat [6, [15] [16] [17] [18] [19] . These attempts to provide public health related measures in the community, in workers involved in outbreak disease control and eradication activities, in people involved in producing, marketing, and living with poultry, and in travellers who are visiting countries experiencing outbreaks. The main recommended measures which need to be used in concert, are: 1) intensify collaboration between the animal and public health sectors; 2) appropriate personal protective equipment for medical workers that transport/treat avian flu patients and for workers involved in the culling, transport or disposal of avian influenzainfected poultry; 3) effective disease surveillance for early detection and reporting of outbreaks; 4) food safety of poultry products; 5) control of movement of birds and products that may contain virus; 6) risk communication; 7) rapid, humane destruction of infected poultry and poultry at high risk of infection, and 8) proper use of vaccination. Data gathered showed that a high number of respondents had no detailed understanding of avian influenza. Specifically, less than half answered correctly the questions on the modes and vehicles of transmission and on the recommended hygienic practices to avoid spreading of the avian influenza virus through food. Moreover, it was disturbing to note that detailed questioning revealed gaps in knowledge about the risk groups. That this happened, despite the fact that almost all received information about avian influenza from different sources, is troubling. In rural Thailand a community cluster survey on 200 people has shown widespread knowledge regarding avian influenza and the effective means of protection with, for example, 76% recognizing that people could get the disease from chicken or other poultry [11] . In our study, the investigation of correlates in multivariate comparison, which allowed us to control for different risk factors, yielded several interesting findings such as that older respondents with a higher educational level and from higher socioeconomic class were more likely to be knowledgeable. The inclusion of measures such as participants' level of schooling and employment status greatly reduces the size differences between groups. Failure to account for these socioeconomic factors would result in biased estimates and artificially high differences across groups. Finally, most respondents recognized that their knowledge on preventive measures was fair and indicated the need for increasing that knowledge. This finding is important because it has already been reported that public health education campaigns and general media reports about avian influenza appear to have been effective in reaching those who were at greatest risk of acquiring the disease through contact with backyard poultry [11] . This study revealed a relatively high degree to which respondents themselves perceived themselves to be at risk of avian influenza virus as a health threat and demonstrated a readiness on the part of respondents to be educated. In our sample respectively 19.3% and 20.4% said that they were "very much" concerned about the risk of contracting avian influenza for them and for friends/ familiars in the future. In a previous telephone study, that examined the perceived risk of avian influenza from live chicken sales involving Hong Kong households, it was documented that one third of those surveyed perceived some risks and almost 50% indicated that their friends had expressed anxieties [10] . It has been well established that worried individuals were significantly more likely to have received advice or instruction about the disease by health professionals and by reading scientific journals and to feel the need for more information. Our findings do suggest that the perception of risk was a significant determinant of greater compliance with recommended precautionary procedures. Indeed, respondents worried about their own risk were more likely to use gloves and wash hands with soap before and after touching raw poultry meat. Such findings suggest that the population would both welcome and benefit from tools and strategies that would help them to reduce their fear because it is important in whether or not they adhere to these procedures. The respondents to our questionnaire exhibited higher compliance with recommendations of the WHO to avoid spread of avian influenza through food [6] , such as hand washing and using protective gloves, when compared with the findings reported in two previous surveys about food-borne diseases. In one study only 20.8% of respondents claimed that they used protective gloves, 53.9% reported washing hands before and after touching raw and unwrapped food, and 50.4% reported using soap to wash hands [20] . In the other one respectively 68.7% and 66.2% of food handlers routinely washed their hands before and after handling any food [21] . As we hypothesized, in accordance with a previous study, knowledge influences behavior [8] . Our survey indicates a significant association between those who fail to wash hands and to use gloves and the lack of knowledge that these are standard hygienic practices to avoid spreading of the virus through food. In addition, it is notable that the main source of information was the media and not qualified healthcare representatives. It has been observed that our results confirmed the hypothesis that those who received information from health professionals and scientific journals had higher relevant compliance than those who did not receive information from these sources. These findings support the importance for media campaigns to implement educational and policy strategies and to increase patient education by medical professionals in the context of routine medical care. Despite the novelty and significance level of these findings, some methodological considerations ought to be highlighted when interpreting our results. First, the analyses were based on cross-sectional data and the findings of the analysis of factors associated with the outcomes of interest should be interpreted with caution given the nature of the associations that limited us from drawing definitive causal conclusions or direction of causality about the observed relationships between among independent characteristics and the outcomes measured. However, we feel that our study design was adequate to assess the knowledge, attitudes, and behaviors relating to avian influenza and identify what percentage of these could be explained by several potential predictors. Second, the study was limited to those parents of children in randomly selected schools, which may have implications for the generalizability of the results. However, because in our country the education is mandatory until the age of 16 irrespective of the characteristics of the parents, we believe our results are generalizable to all population. Third, all variables used in this analysis were gathered using entirely respondents' self-reports and self-perceptions, and biases in perception and reporting cannot be ruled out. The problem with self-reporting is that participants' responses may reflect intentionally or unintentionally perceived desirable responses or an attempt to inflate or minimize reports of behaviors. However, to gain a true reflection of knowledge and behavior, respondents were assured that their responses would not be shared, since the survey was delivered with the responses that were anonymous to encourage accurate recording. Fourth, it was not possible to identify characteristics of those who failed to return the questionnaire, so it was not possible to establish whether they were in any way different from those who did return it. However, there is no obvious reason to suspect that non-responders were substantially different from responders. Despite the potential limitations, the main advantage of the current study is that we were able to achieve a relatively large sample and the high response rate excludes one major potential source of bias in the results. We believe that this high response rate was made possible through the extreme importance of the topic surveyed. In conclusion, the results of this study illustrates that, despite being given information, respondents had no detailed understanding of avian influenza, had a great perceived risk of experiencing avian influenza, and had a low compliance with precautions behaviors. These observations raise concerns about a clear need to find the optimal way of correcting these deficiencies by developing and implementing public health policy regarding priorities for tailored educational and promotion strategies and in particular more attention should be given on using preventive approaches in these population. Nevertheless, it is important to consider that dissemination and widespread adoption of preventive measures require education. Encouragingly, respondent's interest in learning more about avian influenza was high in our survey. Therefore, designing and implementing avian influenza educational programs and measuring their effectiveness should be priorities to incentive the population to take a more active role.

The evolution of human influenza A viruses from 1999 to 2006: A complete genome study

BACKGROUND: Knowledge about the complete genome constellation of seasonal influenza A viruses from different countries is valuable for monitoring and understanding of the evolution and migration of strains. Few complete genome sequences of influenza A viruses from Europe are publicly available at the present time and there have been few longitudinal genome studies of human influenza A viruses. We have studied the evolution of circulating human H3N2, H1N1 and H1N2 influenza A viruses from 1999 to 2006, we analysed 234 Danish human influenza A viruses and characterised 24 complete genomes. RESULTS: H3N2 was the prevalent strain in Denmark during the study period, but H1N1 dominated the 2000–2001 season. H1N2 viruses were first observed in Denmark in 2002–2003. After years of little genetic change in the H1N1 viruses the 2005–2006 season presented H1N1 of greater variability than before. This indicates that H1N1 viruses are evolving and that H1N1 soon is likely to be the prevalent strain again. Generally, the influenza A haemagglutinin (HA) of H3N2 viruses formed seasonal phylogenetic clusters. Different lineages co-circulating within the same season were also observed. The evolution has been stochastic, influenced by small "jumps" in genetic distance rather than constant drift, especially with the introduction of the Fujian-like viruses in 2002–2003. Also evolutionary stasis-periods were observed which might indicate well fit viruses. The evolution of H3N2 viruses have also been influenced by gene reassortments between lineages from different seasons. None of the influenza genes were influenced by strong positive selection pressure. The antigenic site B in H3N2 HA was the preferred site for genetic change during the study period probably because the site A has been masked by glycosylations. Substitutions at CTL-epitopes in the genes coding for the neuraminidase (NA), polymerase acidic protein (PA), matrix protein 1 (M1), non-structural protein 1 (NS1) and especially the nucleoprotein (NP) were observed. The N-linked glycosylation pattern varied during the study period and the H3N2 isolates from 2004 to 2006 were highly glycosylated with ten predicted sequons in HA, the highest amount of glycosylations observed in this study period. CONCLUSION: The present study is the first to our knowledge to characterise the evolution of complete genomes of influenza A H3N2, H1N1 and H1N2 isolates from Europe over a time period of seven years from 1999 to 2006. More precise knowledge about the circulating strains may have implications for predicting the following season strains and thereby better matching the vaccine composition.

Every year the influenza A virus causes human infection with varying severity depending on the host acquired immunity against the particular virus strain. Three to five million people experience severe illness and 0.25 to 0.5 million people die of influenza yearly worldwide (WHO EB111/10). The influenza virus evades host immunity by accumulation of point mutations (drift) in the major surface glycoproteins, haemagglutinin (HA) and neuraminidase (NA) or by reassortment of segments from different viruses co-infecting the same cell leading to a new stain with a HA (and NA) not seen in the population before (shift). In the worst case, shifts may cause pandemics. There have been three pandemics the last hundred years, the Spanish flu in 1918 (H1N1), the Asian flu in 1957 (H2N2) and the Hong Kong flu in 1968 (H3N2). It is believed that new pandemics emerge through shifts with strains from the avian reservoir, as was the case of the pandemics of 1957 and 1968, or by direct introduction of an avian strain into the human population as suggested for the 1918 pandemic [1] . At present only two of the 16 possible HA subtypes (H1 and H3), and two of the nine possible NA subtypes (N1 and N2) are circulating in man. H3N2 and H1N1 influenza A viruses have co-circulated in the human population since the re-emergence of H1N1 in 1977, increasing the possibility for genetic reassortments. The prevalence of the different subtype combinations may vary from season to season. The H3N2 has been the predominant influenza A strain during the last 20 years, with the exception of the 1988-1989 and 2000-2001 seasons where H1N1 infections dominated [2] . In the 2000-2001 season a new reassorted human strain, H1N2, emerged in Europe and became established in the autumn 2001 [3, 4] . The new H1N2 subtype was covered by the 2002-2003 H1 and N2 trivalent vaccine components and because both H1 and N2 viruses had circulated the previous years some degree of herd immunity against the new strain was expected. The H1N2 viruses were not associated with severe influenza illness that season. In 2002, a new lineage A/Fujian/411/02(H3N2)-like emerged in Asia and caused significant outbreaks on every continent [5, 6] . For the northern hemisphere the WHO issues the recommendation for strains to be included in the trivalent vaccine for the next season based on epidemiological data and antigenic and genetic analyses of circulating strains. Until the recent release of over 1,800 complete influenza A genome sequences from the Influenza Genome Sequencing Project managed by US National Institute of Allergy and Infectious Diseases [7, 8] The relative prevalence of influenza virus varies from season to season. Influenza A H3N2 was the dominating strain in Denmark during the last seven years, with the exception of the 2000-2001 season where the H1N1 viruses dominated, as can be seen in Figure 1 . Based on phylogenetic analysis of the HA and NA nucleotide sequences from 1999 to 2006 (Figure 2 ), ten isolates representative for the phylogenetic clustering of sequences from each subtype in each season, as far as possible, were included in the final HA and NA tree ( Figure 2 ) and representatives were chosen for complete genome sequencing. Generally the H3N2 HA and NA genes formed seasonal phylogenetic clusters ( Figure 2 ). However, we observed that strains of different lineages and clusters cocirculated within the same season and that viruses had reassorted with viruses from previous seasons ( Figure 2 ). The HA gene of the influenza H3N2 strains from the 1999-2000 season formed a phylogenetic subclade to A/ Moscow/10/99(H3N2) and A/Sydney/5/97(H3N2) (represented by A/Memphis/31/98) (Figure 2 ), located between A/Moscow/10/99 and A/Panama/2007/99 (not shown). The antigenicity of these strains was A/Moscow/ 10/99(H3N2)-like in a haemagglutination inhibition assay, and will therefore be referred to as Moscow-like throughout this report. The NA and the internal genes were all A/Moscow/10/99(H3N2)-like, with the exception of the matrix (M) gene that clustered as a subclade to the A/New York/55/01-like strains ( Figure 3 ). Figure 1 ). Thirteen isolates from this season were available for sequencing, and all were of the H1N1 subtype. These sequences represented two different co-circulating lineages (Figure 4 ). Lineage I is A/Bayern/7/ 95(H1N1)-like and lineage II include the H1N1 strains of today and the A/New Caledonia/20/99(H1N1) vaccine reference strain (Figure 4 ). The phylogenetic trees of NA and the internal genes showed the same topology ( Figure 3 and 4) . The lineage II strains are characterised by a deletion K130 in HA (K134 in H3 numbering) ( Table 1) * Amino acids in brackets indicate less than half but more than two substitutions at the given amino acid position within a season. A single amino acid change in one position is not shown. Amino acids separated by '/' indicate equal substitutions of either amino acid at the given position. Letters in upper case above an amino acid indicate the antigenic site location of the residue. In N1 the upper case letter ' P ' stands for phylogenetically important region (PIR) and the following letters indicate the actual PIR. Based on the ratio of non-synonymous versus synonymous substitutions none of the influenza A genes were directly influenced by positive selection (dN/dS<1) (Table 3) . However, as expected the HA1 region of both H3 and H1 viruses were more influenced by evolutionary pressure (Table 3) . We applied FEL and SLAC maximum-likelihood methods to estimate individual positively selected sites in H3N2 HA and NA and added REL for smaller datasets in all genes (se methods section). The FEL method found one site in the H3 protein (n = 204), position 199 (p = 0.046) to be positively selected, while the more conservative SLAC analysis found none. No positive selected sites were predicted for the N2 genes (n = 166) and none in the internal genes (n = 15) estimated by FEL and SLAC. The REL analysis retrieved four sites in the M1 gene to be selected namely positions 208, 211, 218 and 219. No sites in HA (n = 27) and NA (n = 30) of the H1N1 viruses were directly positively selected with any of the three methods of analysis. [26] and it was anticipated that there would be some extent of herd immunity in the population against this new reassortant. The phylogenetic trees of H3N2 HA and NA showed seasonal clusters but also co-circulating lineages within seasons. The introduction of the A/Fujian/411/02(H3N2) strains in 2002-2003 caused a "jump" in the evolution of both HA and NA genes. Many of the substitutions in HA introduced with the A/Fujian/411/02(H3N2)-like viruses have become fixed, probably reflecting a very fit virus. The genetic variation before the 2003-2004 season may have been more influenced by introduction of new viruses through viral migration than adaptive evolution of the genes. A constant rate of drift was not observed for HA but instead periods of change followed by stasis periods. The low dN/dS ratios (0.232 for HA and 0.247 for NA) also indicated that the influenza genes were not directly influenced by positive selection. Reassortments between co-circulating strains and viruses from previous seasons, and introduction of viruses from other parts of the world might play a larger role than natural selection for some seasons, as also observed by others [27, 28] . [2, 4, 25, 26] . The reassorted H1N2 viruses possessed only the HA from A/New Caldonia/20/ 99(H1N1)-like viruses and the rest of the genome from A/ Moscow/10/99(H3N2)-like viruses as also reported by others using partial sequences [25] . The H1N2 viruses have been introduced to Denmark from elsewhere and are not a local reassortant. (Table 1) suggesting they may be important for viral escape from the host immune system and the overall fit of the virus. The A/Fujian/411/02-like(H3N2) viruses did not antigenically match the A/Moscow/10/99(H3N2) strain included in the 2002-2003 vaccine [29, 36] . The HA substitution D144N was however not responsible for the antigenic drift of the Fujian-like viruses. It has been shown that only two amino acid changes, H155T and Q156H, specified the antigenic difference from Moscow-like to Fujian-like [37] , both are located at antigenic site B. The T155 and H156 amino acids have been maintained in all Danish isolates after the introduction of the Fujian-like viruses. With the A/California/7/04(H3N2)-like viruses in 2004-2005, site 145 has changed from K to S in some isolates and to N in others. S145 has been found at this position before 1999 and N145K was the only cluster-difference substitution between isolates from the seasons 1987/ 1989 and 1992/1995 [38] . It was therefore suggested that substitutions at this site alone could have large antigenic effect. Antigenic site A, located in a loop, makes few contacts with the rest of the structure, therefore 144 and 145 may change drastically without influencing on the overall shape of the HA molecule. Antigenic site A is supposed to be ideal for antibody binding and for amino acid replacements [33] . The preferred antigenic site in the change from A/Moscow/10/ 99(H3N2) to A/Fujian/411/02(H3N2)-like viruses was site B. This observation is in accordance with previously published data [39] . One region in HA, position 225 to 227, that influence antigenic site D, has changed drastically during the study period from GVS → DVS → DIP → NIP. The influence of antigenic site D was first apparent in the antigenic change from Fujian-like viruses to California-like viruses; however, the antigenic site B was still the preferred site for antigenic change. It has been proposed that a minimum of four substitutions in two or more antibody binding sites are required for an epidemically important strain [40] . Gulati et al., [41] [22, [42] [43] [44] [45] . The substitution H274Y is the only neuraminidase resistance mutation identified in H1N1 viruses to date [46, 47] . Danish isolates did not possess any resistance mutations in the NAs. Neuraminidase inhibitory drugs are rarely used for influenza prophylaxis or treatment in Denmark. T-cell epitopes are more conserved than antibody epitopes. Fifteen per cent of the T-cell epitopes are conserved while only 2.7% of the antibody epitopes [48] . The reason for this higher degree of conservation is that 80% of the antibody epitopes are located in the variable glycoproteins HA and NA, while only 40% of the T-cell epitopes are found in these proteins [48] . Recent research has shown some degree of escape from CTL-mediated immunity in addition to escape from neutralizing antibodies [28] . In our dataset we found several substitutions in regions involved in protective T-cell response [48] in NA, PA, M1, NS1 and most in the NP protein. This is not unexpected because most T-cell epitopes are defined for the NP protein and this protein is the main target for the cytotoxic host immune response [49, 50] . The extensive variations in the T-cell epitopes during 1999 to 2006 suggest that these regions and the antibody epitopes are working together for efficient escape from the host defence responses. The M2 proteins from the Danish Wisconsin-like viruses in 2005-2006 possessed the substitution S31N, associated with resistance to matrix inhibitory drugs, like amantadine [22, 44, 45, 51] . These types of drugs are not used for prophylaxis or treatment in Denmark. The S31N substitution is therefore not a local introduced resistance mutation. We cannot exclude that this substitution has arisen by chance, but it is more likely that the mutation has emerged as a resistance mutation in other countries like the USA [52] and Australia [53] where the prevalence of amantadine resistance is high. The resistance may also be related to the increased use of this drug in Asia during the SARS epidemic [21] . The A/Fujian/411/02(H3N2)-like clinical Danish viruses had several substitutions in HA at sites that might influence the virus' capability for egg growth [10, 37] . These include; A131T, I144N, H155T, Q156H, W222R and G225D. The A/California/20/99(H3N2)-like viruses had further changes at positions K145S/N, Y159F, S193F and V226I and A/Wisconsin/67/05(H3N2) possessed in addition S193F and D225N. All isolates after the 1999-2000 A/Moscow/10/99(H3N2)-like viruses possess S186G. In recent years H3N2 viruses have had poor replication efficiency in eggs [54, 55] . It has been shown that positions 186, 226 and 196 are critical determinants for egg growth. The changes G186V and V226I increased egg viral replica-tion of A/Fujian/411/02(H3N2) viruses so did the changes G186V and A196T for A/California/20/ 99(H3N2) viruses [56, 57] . On the contrary, others have stated that the V226I change in combination with T155 and H156 do not result in viral recovery in eggs [37, 54] . This might explain the delay in the 2006-2007 vaccine production for the northern hemisphere due to egg propagation difficulties with the A/Wisconsin/67/05(H3N2) strain. The influence on replication efficiency by the other substitutions observed at receptor binding sites should be investigated further. We did not observe amino acids in the N2 NA protein that would decrease virus replication in eggs. The amino acids known to give good replication in eggs (Q119, K136 and Y347) [56] were all present in this dataset. Oligosaccharides at the surface proteins HA and NA might have greater impact on viral escape from the immune system than single amino acid changes in the antigenic sites. Oligosaccharides which are recognised as "self" by the host immune system may pose conformational changes in the molecule and mask antigenic sites, thereby prevent binding of host antibodies. The number of N-linked glycosylation sites in the H3 HA protein has increased during the years from only three attachment sites in 1968 [58, 59] to ten predicted sites in the Danish isolates after 2004. The A/Fujian/411/02(H3N2)-like stains from the 2002-2003 season gained a potential glycosylation site at position 144, thereby masking the supposed "key" site for antigenic change [33] . Substitutions at antigenic site B and the predicted N-glycosylation at position 144 in HA antigenic site A together with a stronger NA might have contributed to the increased infectivity of the reassorted Fujian-like viruses of the 2003-2004 season, causing an epidemic in Denmark. We have shown that the preferred antibody epitope for genetic change is antigenic site B for the Danish dataset also reflecting that site A is camouflaged by glycosylation. Thus the antigenic changes at a glycosylated site A may not play a major role in escape from the immune system as long as the glycosylation is present. Six potential N-glycosylation sites have been conserved in the N2 NA Danish dataset from 1999 to 2003. The majority of isolates from 2003 to 2006 have lost the site at position 93 which is located in a CTL epitope (HLA-A*0201) region of NA [12] . The recent NAs may therefore be more easily recognised by the cytotoxic immune system. We found two predicted N-glycosylation sites (61 and 70) in the N2 NA stalk region in sequences from 1999-2006. Greater density of carbohydrate in the stalk region of NA might reflect a need for proteolytic protection. The two observed carbohydrates in the stalk have been reported for other isolates in other time periods and other continents [58, 60] . The stalk region has therefore stayed unchanged and the two sequons seem to be conserved. As expected a higher dN/dS ratio was observed for the surface glycoproteins, although none were directly influenced by positive selection. We observed that the HA1 region and the M2 protein have a slightly higher global dN/dS ratio than the other genes (Table 3 ). This is consistent with the findings of others [11, 61] . The M2 protein is a membrane ion channel protein on the surface of the virus molecule, a higher dN/dS ratio for this protein compared to the internal proteins is expected. The ratio might, however, be biased because the M2 protein is spliced from M and the dS is suppressed for overlapping regions giving a higher dN/dS ratio [11] . Earlier findings support positive selection at sites involved in receptor and antigen binding [11, 62] . Five of the 18 codons in HA proposed by Bush et al., [63] to be under positive selection were found to have changed in our HA dataset of H3N2 since 1999, namely: pos 145, 156, 186, 193 and 226. In our H3 dataset (n = 204) site 199 was influenced by positive selection as calculated by FEL analysis, but no positively selected sites were found applying the more conservative SLAC method. Position 199 may be involved in receptor binding and influence on the virus ability to grow in eggs [10] . In a similar study on a slightly larger dataset (n = 284) positions 220 and 229 were found to be positively selected [11] . Another study found positions 13 and 236 [62] ; however, suggested positively selected sites may vary by the dataset applied, method used and the significance level selected for a site to be classified as positively selected. REL analysis identified four sites (208, 211, 218 and 219) in M1 under positive selection pressure. REL analysis tends to give better estimates on small datasets than SLAC and FEL. Sites 211, 218, and 219 were still selected when the bayes factor cut-off was increased from 50 to 200. Further analysis would be needed to determine if these sites actually are positively selected. There is a need for complete genome analysis of European human influenza A viruses in order to gather a comprehensive picture of the evolution and migration of viruses. Our results support the suggestion that the evolution of influenza A viruses is more complex than originally believed [28, 62] . Local short term evolution of influenza virus is a stochastic process, also involving gene reassortments. It will be interesting to further investigate how viruses from other parts of Europe influence on the evolu-tion of Danish isolates when more full length sequences from Europe are made public. Viral RNA was extracted from 140 µl of human nasal swab suspension or nasopharyngeal aspirate by QIAamp ® Viral RNA Mini Kit (QIAGEN, Germany) as described by the manufacturer or by an automated MagNA Pure LC Instrument applying the MagNa Pure LC Total Nucleic Acid Isolation Kit (Roche Diagnostics, Basel, Switzerland). The different gene segments were amplified by OneStep RT-PCR Kit (QIAGEN) as previously described [64] , applying a two minute elongation step for all genes. The primers for RT-PCR were segment specific but subtype universal targeting the highly conserved noncoding RNA regions at the 5'-and 3'-end of each segment [65] . PCR products were purified using the GFX™ PCR DNA and Gel Band Purification Kit (Amersham Biosciences, Germany) prior to sequencing. Purified PCR products were sequenced directly. Primer sequences are available upon request. The sequencing reaction was performed by ABI PRISM ® BigDye™ Terminators v3.1 Cycle Sequencing Kit (Applied Biosystems, USA) as described previously [66] . The sequences were developed on an automatic ABI PRISM ® 3130 genetic analyzer (Applied Biosystems) with 80 cm capillaries. Consensus sequences were generated in SeqScape ® Software v2.5 (Applied Biosystems). Sequence assembly, multiple alignment and alignment trimming were performed with the BioEdit software v.7.0.5 [67] . Distance based neighbor joining (NJ) phylogenetic trees and character based maximum parsimony (MP) trees were generated using the Molecular Evolutionary Genetics Analysis (MEGA) software v.3.1 [68] . Maximum likelihood trees were generated by the Phylogenetic Analysis Using Parsimony (PAUP 4.0) Software (Sinauer Associates, Inc.) [69] applying the HKY85 nucleotide model, allowing transitions and transversions to occur at different rates. The between-seasons nucleotide distance means were computed as the arithmetic mean of all pair wise distances between two seasons in the inter-season comparisons using the MEGA v.3.1 software [68] . The global rate between dN and dS substitutions and the individual sitespecific selection pressure were measured by the likelihood based single likelihood ancestor counting (SLAC) method in Datamonkey (modified Suzuki-Gojobori method) [70, 71] . For datasets over 100 sequences (H3 n = 204 and N2 n = 166) the HyPhy package [72] was applied. The estimations are likelihood-based, employing a codon model cross between HKY85 and MG94. To elucidate single, positively selected amino acids, the HA and NA datasets were analysed with SLAC and a two rate fixed effects likelihood (FEL) [73] using a likelihood approach with neighbor joining phylogenetic trees (HyPhy). Sites with dN>dS with a <0.05 significance level for likelihood ratio test (LRT) were implied as positively selected for the large HA and NA datasets. For the small datasets (genes coding for the internal proteins <15 and the H1N1 viruses <30) we additionally ran a random effects likelihood (REL) analysis using an empirical Bayes approach with NJ phylogenetic trees in Datamonkey [70] . This method is expected to calculate positively selected sites more accurately in small datasets. The accepted significance level for a positively selected site was set at <0.1 (two-tailed binominal distribution) for SLAC and FEL analyses and >50 bayes factor for REL. Potential N-linked glycosylation sites were predicted using nine artificial neural networks with the NetNGlyc 1.0 Server [74] . A threshold value of >0.5 average potential score was set to predict glycosylated sites. The N-Glycosite prediction tool at Los Alamos [75] was used to visualise the fraction of isolates possessing certain glycosylated sites along the sequence alignment. The specific measure of antigenic distance between two strains of influenza were calculated as P epitope values by the method suggested by Muñoz, et al., [39] . The P epitope value was calculated as the number of mutations within an antibody antigenic site divided by the number of amino acids defining the site. It is assumed that an antigenic epitope which has the greatest percentage of mutations is dominant, because that epitope is influenced by the greatest selective pressure from the immune system. The P epitope distance is defined as the fractional change between the dominant antigenic epitopes of one strain compared to another. The P epitope Calculator [76, 77] was applied for H3 sequences. Residues in antigenic epitopes were collected from several references [9, 13, 39, 40, 48, [78] [79] [80] [81] [82] [83] 83, 84] .

Can "presumed consent" justify the duty to treat infectious diseases? An analysis

BACKGROUND: AIDS, SARS, and the recent epidemics of the avian-flu have all served to remind us the debate over the limits of the moral duty to care. It is important to first consider the question of whether or not the "duty to treat" might be subject to contextual constraints. The purpose of this study was to investigate the opinions and beliefs held by both physicians and dentists regarding the occupational risks of infectious diseases, and to analyze the argument that the notion of "presumed consent" on the part of professionals may be grounds for supporting the duty to treat. METHODS: For this cross-sectional survey, the study population was selected from among physicians and dentists in Ankara. All of the 373 participants were given a self-administered questionnaire. RESULTS: In total, 79.6% of the participants said that they either had some degree of knowledge about the risks when they chose their profession or that they learned of the risks later during their education and training. Of the participants, 5.2% said that they would not have chosen this profession if they had been informed of the risks. It was found that 57% of the participants believed that there is a standard level of risk, and 52% of the participants stated that certain diseases would exceed the level of acceptable risk unless specific protective measures were implemented. CONCLUSION: If we use the presumed consent argument to establish the duty of the HCW to provide care, we are confronted with problems ranging over the difficulty of choosing a profession autonomously, the constant level of uncertainty present in the medical profession, the near-impossibility of being able to evaluate retrospectively whether every individual was informed, and the seemingly inescapable problem that this practice would legitimize, and perhaps even foster, discrimination against patients with certain diseases. Our findings suggest that another problem can be added to the list: one-fifth of the participants in this study either lacked adequate knowledge of the occupational risks when they chose the medical profession or were not sufficiently informed of these risks during their faculty education and training. Furthermore, in terms of the moral duty to provide care, it seems that most HCWs are more concerned about the availability of protective measures than about whether they had been informed of a particular risk beforehand. For all these reasons, the presumed consent argument is not persuasive enough, and cannot be used to justify the duty to provide care. It is therefore more useful to emphasize justifications other than presumed consent when defining the duty of HCWs to provide care, such as the social contract between society and the medical profession and the fact that HCWs have a greater ability to provide medical aid.

In the course of providing health care service, health care workers (HCWs) are continually exposed to many workrelated health risks. One of these risks is the exposure to infectious diseases. These diseases can include the flu, AIDS, tuberculosis, and hepatitis, and can be transmitted through physical contact, exposure to contaminated blood, or via the respiratory system. And, needless to say, such risks do indeed at times prove fatal. The consequences of occupational exposure to pathogens are not limited solely to bodily infections. Each year, thousands of HCWs are adversely affected by psychological trauma stemming from months of anxiously awaiting the results of serological tests, tests made necessary due to potential infection incidents. The anxiety experienced by HCWs is related to the perception of risk from the incident and the resulting infection that may occur, and by the worry of what the reactions of others might be, such as colleagues, family, and friends, all who have to be informed. During this uncertain waiting period HCWs will frequently experience intrusive thoughts, problems concentrating, difficulty sleeping, frequent loss of temper, and a decrease in sexual desire, which can act as a catalyst to exacerbate any pre-existing and unresolved emotional issues [1] . And if it turns out that the health care worker has indeed been infected by one of these contagions, the serious personal consequences to that health care worker can include the postponement of childbearing, damaged personal relationships, having to alter sexual practices, experiencing the side effects of prophylactic drugs, chronic disabilities, loss of employment, denial of worker compensation claims, possible need for a liver transplant, and premature death [2] . AIDS, SARS, and the recent epidemics of the avian-flu have all served to remind us of the occupational risks faced everyday by HCWs; the result being that the recent appearance of these diseases has forced this issue onto the common agenda and helped to spark renewed interest in the debate over the limits of the moral duty to treat. In seeking an answer to this question it is useful to have an understanding of the occupational risks faced by HCWs as well as an understanding of the attitudes of HCWs to these risks. For example, studies conducted in various countries have shown that, especially when there was a risk of being infected with AIDS, HCWs may refuse to treat a patient on the grounds that there is a risk of being infected by this patient [3] [4] [5] [6] [7] [8] [9] [10] [11] . And despite the fact that the hepatitis viruses are transmitted more easily than HIV, it is the fear of being infected with HIV that causes many HCWs to experience the greatest amount of stress and anxiety [12] . In a study which compared the relative risks of transmission of both HBV and HIV, the reasons for physicians' underlying fears of particular contagions were also inves-tigated and described. [13] . According to the study, people initially percieve the risk to be greater when there is a high likelihood of death involved with infection (as with HIV) even though there may be less risk of infection, as opposed to when there is a higher risk of infection but a lower risk of death involved with that infection (as with HBV). Additionally, since the likelihood of sexually transmitting HBV between heterosexual partners is less than that of transmitting HIV, the consequences of HBV infection are again percieved to be less severe than the consequences of HIV infection. In this way, the hazards posed by HBV infection conflict less with the obligation to protect family members from harm. It was also found to be important that there is less of a stigma attached to having HBV than there is to having HIV. And, finally, the fact that there is a vaccine for HBV infection, which is more than 90 percent effective (for vaccinated HCWs the risk of death from infection is reduced by a factor of nearly twenty), also was found to greatly influence the perceptions of the physicians. Additionally, factors other than a fear of the contagion can contribute to the reluctance to treat a particular patient. Some physicians and dentists express concern that if it is discovered that they treat patients with AIDS, then those patients who don't have HIV may shun their practice. Still, other physicians insist they do not know enough about HIV infection and are too busy to learn [14] . Another reason for which a physician may refuse to treat HIV-positive patients is that the physician feels they have a duty to protect their other patients, basing their reasoning on the principle of "First do not harm". By treating HIV-positive patients they claim that they may potentially be putting their other patients at risk for infection [15] . Furthermore, as has been reported, there is always the possibility that when a HCW is able to reject the patient based on a more benign excuse, for example if the patient does not have enough money, it is even easier, and all the more likely, for treatment to be refused, even though this refusal was done in the interest of protecting the physical health of the individual health care provider [14] . In the literature, most studies have concentrated primarily on the attitudes and rationale behind the refusal to treat. Before one can set out to effectively explore the attitudes of HCWs however, it is important to first consider the question of whether or not the "duty to treat" might be subject to contextual constraints, such as providing health care to a patient suffering from an infectious disease which may be particularly contagious or for which adequate treatment measures may not yet be available. Clark, in his article about physicians' duty to treat, claimed that there are three reasons in which such a duty is grounded [16] : "... since the ability to render aid is greater, the obligation to assist is (...) elevated. Second, by consideration of Daniels' argument that by freely joining a profession designed to combat disease, one consents to some standard of risk, and third, by realizing that the profession has flourished due to socially negotiated promises to be available in such times of duress." In his article "Duty to treat or right to refuse", Daniels argues that when a person chooses a career in a particular profession, it must be understood by all parties that this individual has both accepted and is willing to take the risks that are inherent to that profession [14] : "Consent is crucial where obligations to take risks exist in various occupations or professions. For example, we assume that in choosing their careers, undergoing the training involved, and agreeing to follow the codes and practices regulating their work, firefighters and police have given consent to facing the significant risks they are obliged to take. There are strong parallels to medicine. People who enter medical fields clearly had alternatives. There is a general understanding that physicians face an increased risk of contagion from disease, an understanding refined during schooling and training." Daniels proposes, however, that some situations can exceed the standard level of risk (SLR) [14] : "For example, it is common to screen new house staff and nurses in medical centers to determine whether any individuals face special risks of contagion, such as immunosuppression or pregnancy. Those at high risk may then be asked to avoid certain treatment situations, materials, or hospital areas.(...) Protecting immunosuppressed providers is reasonable "risk management", a measure taken to reduce bad outcomes. But such special protection supports the claim that only standard risks are included in the duty to treat.(...) Some nosocomial risks clearly take us beyond what duty requires." It is perhaps more illustrative if this argument (from this point on, this statement will simply be referred to as the "presumed consent") is written in classic form: Health care services should be provided to patients who have a contagious disease. Contracting an infectious disease while providing health care services to a patient with a contagious disease is an occupational risk. It is generally assumed that by joining the health care profession physicians have given their consent to be exposed to an increased risk of disease contagion. This assumption is based on the following facts: a. There is a general understanding that physicians face an increased risk of contagion from disease, an understanding refined during schooling and training. b. People who enter medical fields clearly had alternatives. Some nosocomial risks clearly take us beyond what duty requires. There is a moral duty to treat patients who have a contagious disease so long as the risk to the HCW is below the SLR. If we are to accept this argument, then the pressing question becomes how to determine and define the risks which are deemed to be standard and acceptable versus those which are believed to exceed and, indeed, outweigh the duty of the health care provider to treat. In order to begin to answer this question, it will be useful to investigate the nature of the choice (and all that goes along with making it) that an individual makes when they decide to enter a particular profession. For instance, how wise is it to assume that at the time of choosing their future profession the HCW was fully aware of the risks involved with such work? Perhaps they were not made aware of the risks until their education and training. Furthermore, if they were aware of, and fully appreciated, the risks prior to deciding on a particular profession, would they have even chosen that profession in the first place? And, finally, how is the SLR to be determined, and which of the infectious diseases would then exceed this SLR? In order to effectively analyze the presumed consent argument it is necessary to have an awareness of the diverse opinions and beliefs of HCWs and, also, to understand their different motives and backgrounds. Additionally, knowledge of what HCWs feel about the risk concept and of how they feel about their duty to treat patients with contagious diseases can also be of great value to educators as they plan their curricula and it can be used by the authorities in charge of health care systems in order to better organize their services. The purpose of this study was to analyze whether or not the third premise grounds the duty to treat, namely, "it is generally assumed that by joining the health care profession physicians have given their consent to be exposed to an increased risk of disease contagion". In order to carry out this analysis, the opinions and beliefs of physicians and dentists regarding the occupational risks of infectious diseases were investigated; and, by extension, the argument that the notion of "presumed consent" may be grounds for supporting the HCWs' duty to treat was also analyzed. For this cross-sectional survey, the study population was selected from among physicians and dentists in Ankara, the capital of Turkey. A self-administered questionnaire designed to assess the beliefs and opinions of the participants regarding the occupational risks of infectious diseases was used. This questionnaire was also used to obtain the socio-demographic information of the participants. The 17 items on the questionnaire were developed by reviewing previous studies in the literature [1, [3] [4] [5] [6] [7] [8] [9] [10] . A draft of the questionnaire was distributed to experienced health care professionals and later revised based on their criticism and suggestions. In both of the universities in which this study was conducted there are ethics committees which had been established for the purpose of determining the ethical appropriateness of pharmaceutical trials using humans; since our study involved only the use of a questionnaire and not an experimental drug, we did not apply for approval from either of these ethics committees. Instead, written permission to carry out the study was granted by the dean of the faculty of medicine and by the chief manager of university hospitals. In addition, all of the potential participants were fully informed about the aim and structure of the study. Furthermore, potential volunteers were all made aware that participation was strictly voluntary and that all of the answers they provide would be done so anonymously. The questionnaire was administered to a total of 373 health care workers: all of the 236 physicians who work in surgical specialties at the Ankara University Ibn-i Sina Hospital and to all of the 137 dentists in the Gazi University Faculty of Dentistry. Dentists were included in this study because, aside from being HCWs themselves, there are a number of studies in the literature which show that dentists, citing various reasons, may also refuse to treat patients with contagious diseases. And, in order to better assess the fact on which the third premise of presumed consent is based, we decided to include only professional health care workers, instead of students and others who might still be in the process of deciding whether or not to currently enter the field. In total there were 230 participants, 101 physicians and 129 dentists, who completed the questionnaire, for an overall response rate of 61.7%. The questionnaire was later sent back to the non-respondents one month after the first survey, and 28 of these were completed and returned to us. The mean age of the participants was 33.8 ± 9.6 years, while 56.5% were male and 43.5% were female. Additionally, the average amount of time that they had been working in the medical profession was found to be 8.5 years (min. 0, max. 40). All of the data was collected anonymously. The difference between the two groups, physicians and dentists, was compared using the chi-square test, with a p-value of <0.05 accepted as statistically significant. All analyses were carried out using SPSS 11.5. Of the HCWs surveyed in this study, roughly half stated that they understood that by choosing their profession they would be exposing themselves to an increased risk of contracting contagious disease (55.2%). And at the time of entering the faculty, 24.4% of the participants expressed that they were unaware of any increased risks; however, they later learned of these risks during their education and training. In other words, 79.6% of the participants stated that they had known about the risks either at the time they chose their profession or that they had later learned of the risks during their training and education. Additionally, 6.5% of the participants answered that they had only come to realize the kinds of risks they would face after starting to work. The percentage of participants who claimed that if they had been aware of the risks earlier they would not have chosen to enter or continue in the medical profession was 5.2%. Table 1 are statements which physicians and dentists chose as best reflecting their personal opinions regarding the occupational risks of infectious disease. In general, the physicians, prior to their education and training, were significantly more aware of the potential risks associated with their profession than were the dentists (p < 0.05). A significantly higher percentage of the dentists however, stated that they only learned of the occupational risks of dentistry during their education and training (p < 0.05). There was no statistically significant difference between the two groups in terms of the other opinions questioned. The participants were also asked whether or not they agree with the argument "When people choose, and continue to practice, the medical or dentistry profession, they are then required to accept all of the occupational risks resulting from the infectious diseases they might confront". The aim of this question was to determine whether or not the HCWs each have their own individual working-definition for the SLR. Of the participants, 57.4% believed that there is such a level. 52.2% felt that certain diseases would exceed the level of acceptable risk unless specific protective measures were implemented, and 5.2% said that some diseases were always beyond the SLR, no matter what precautions might be taken. No statistically significant difference was found between the physicians and the dentists. Listed in Table 2 are the diseases which, under certain circumstances, were cited as potentially exceeding the SLR. Among the participants who stated that there would be a SLR for providing health care to the patients of specific diseases unless protective measures were implemented, AIDS and Hepatitis C and B were the most frequently cited of these diseases (71.7%, 64.2%, and 56.7%, respectively). The participants who felt that some diseases would always exceed a SLR expressed, that Hepatitis B, Tuberculosis, and Bacterial meningitis always would go beyond the SLR (41.7%, same for all). According to these participants, the occupational risk of potentially being infected with HIV is paramount to all other risks. Percentage-wise, AIDS was the most frequently mentioned disease that would exceed the SLR, more so than SARS. All of the participants who answered that some diseases would be beyond the SLR were then asked what criteria they used to make their determination. The most com-monly expressed criteria, in order, regarding the diseases, were the likelihood of transmission, whether or not protective measures are available, and whether or not immunization is possible (66.7%, 65.2%, and 58.3%, respectively). The distribution of these criteria among the physicians and dentists can be seen in Table 3 . Physicians expressed significantly more often than dentists that if there was no immunization or treatment available for a particular disease, then that disease would exceed the SLR (p < 0.01). In terms of other criteria, there were no significant differences observed between the two groups. The primary aim of this paper is to evaluate the claim that presumed consent may constitute grounds for the moral duty to treat. The presumed consent argument is valid, because its conclusion should logically be accepted if its premises are taken into account. To analyze the soundness of the argument we carried out a survey investigating the opinions of HCWs about the occupational risks of infectious diseases. In total, 79.6% of the participants said that they either had some degree of knowledge about the risks when they chose their profession or that they learned of the risks later during their education and training. In other words, one fifth of the participants either lacked adequate knowledge about the occupational risks when they chose their profession or were not sufficiently informed of these risks during their faculty education and training. This means that the assumption stated in Premise 3 may be wrong for an important proportion of health care workers. It seems reasonable to suggest that the words "there is a general understanding" would be misleading if used to characterize a social concept of which the applicability and, indeed, the very existence, are yet to be established by sociological studies. It is also useful to discuss the other problems associated with presumed consent; in particular, the difficulty of choosing a profession autonomously, the constant level of uncertainty present in the medical profession, the nearimpossibility of being able to evaluate, in retrospect, whether or not every individual was informed, and the seemingly inescapable problem that this practice would legitimize, and perhaps even foster, discrimination against patients with certain diseases. If we are to use the presumed consent argument, then the findings of this study indicate that when a new epidemic of a contagious disease occurs, or when the medical profession is confronted with a disease for which no immunization or treatment options are available, some HCWs are not bound by the duty to treat according to the presumed consent argument. This seems potentially problematic and demands serious consideration. How appropriate is it to describe the healthcare provider's responsibility and duty as stemming from their 'consent'? To address this question, it is helpful to reflect on the conditions required for an individual to be able to give consent that is well-informed. For an HCW's consent to be informed, the following should first be explained to them: (a) the risk posed by each of the contagious diseases known at that given time, (b) commonly agreed criteria and definitions of situations that would surpass the SLR, and (c) the fact that there will always be a degree of uncertainty involved with working in the medical profession, as new risks may emerge at any point during one's professional life. If not necessarily when they choose their profession, then at least after being given the relevant knowledge during education and training, the person's choice should be regarded as informed. It should therefore be ensured that HCWs are acquainted with each new and emerging risk, and with any methods of prevention developed during or after their education and training. If a person's choice is to be confidently regarded as informed, it is imperative that these conditions be met. Of course, the question now becomes: how possible is it to satisfy all these conditions? It is quite easy to imagine more than one answer to this question, but one thing is for sure: any thoughtful answer would acknowledge that choice is determined both by factors that are under the control of the individual and by factors that are not. Personal factors such as educational status, perception of the world and ambitions all influence an individual's choice of profession strongly. Nevertheless, factors outside the individual's control also play a large role in determining that choice. The environment in which the person grew up -their family life, the jobs of their parents, their community, social class and cultureall contribute to forming that individual's background, which (needless to say) has a very large influence on the opportunities and choices available to them. Even though a person may not have been sufficiently informed when they chose their profession, it can be argued that during their education and training they will learn all relevant knowledge about the occupational risks associated with working in the medical profession. If so, it is fair to assume that when this individual begins to work after graduation they will be willing to confront any of those risks. In theory at least, it can be presumed that every student who passes their exams and goes on to graduate from the faculty is informed of the risks; so it can be argued that all HCWs who are currently active in their profession have consented to accept the risks posed by all the contagious diseases known at the time of their graduation. Of course, the diverse factors that determine the quality of education, such as the particular educational methods used, the course content, the abilities and knowledge of the instructors, role-models, and the personal features and motivations of students, are all potential sources of variation. But for the sake of argument, let us assume that a standardized education program is implemented throughout all medical schools. If that were the situation, then the argument that the individual has been made aware of the occupational risks during education would be true to the extent that the education program addressed those risks sufficiently. Nevertheless, it would be hard to claim that presumed consent is valid for every individual. For many people, a degree in medicine is very costly, both financially and in terms of time and energy. Under such circumstances, it is difficult for a person to quit their schooling despite the awareness they gain of the occupational risks involved. Individuals might feel pressured and confused by the two options confronting them: on the one hand, dropping out of medicine and forfeiting all the time, effort and money spent on schooling towards that aim; and on the other, reluctantly accepting the occupational risks, which may look frightening to the individual at that moment. Of course, it is important to remember that the person chose the medical profession in the first place, and numerous positive and beneficial elements are associated with working within it, which may ultimately serve to temper and override the individual's fear of the risks. An additional source of pressure may be that, for whatever reasons, switching educational tracks is too difficult; it may appear too daunting or be financially untenable. It seems likely that in the end the individual will choose to continue with their education and embark on a career in medicine, despite hesitation and fear of the risks. It is difficult to describe a decision made under conditions of such uncertainty and stress as 'informed'. As can be seen, we do not choose our profession from among a wide array of possibilities spread out in front of us by thoroughly researching each one so that we are fully informed of its nature; everybody's options are different and it is a large and difficult task to make oneself sufficiently informed of them. Moreover, as described above, the decision to quit medical school can be quite difficult: on one side of the dilemma there are occupational risks that must be accepted regardless of misgivings on the part of the individual; on the other side, very influential factors pressure the individual to continue their medical education. Thus, the claim that "People who enter medical fields clearly had alternatives" is debatable and sometimes even doubtful. In theory it sounds right; nobody has to be a physician. But in practice, having alternatives does not mean that all our decisions are made freely or autonomously. Theoretically, the fundamental problem with the presumed consent argument is that it cannot explain why there is always some degree of uncertainty about the occupational risks of working in the medical profession, particularly stemming from new and emerging diseases. Quite simply, if there was little or no knowledge of a risk at the time the individual became informed and gave their (tacit) consent, then this individual never accepted the risk, implicitly or otherwise, because it was unknown at the time the individual was informed. From a historical perspective, it is possible to see that while the medical profession was once concerned only with treating diseases, its vocational responsibility came in time to include preventative, promotive and rehabilitative healthcare services. As the notions of human rights and patient rights have developed and become widespread, perceptions about the health profession have changed at the community level. Diseases that once killed millions of people can now be treated with a simple medicament, but today we are faced with new and challenging diseases unheard of in the past. As a result, the continuous cycle of changespurred on by greater knowledge and technological advances and confrontations with new and untreatable diseases -serves to alter the identity and nature of the medical profession, and out of all this arises a constant degree of uncertainty. This characteristic uncertainty is present both when the individual chooses the profession and throughout their education and training periodand, indeed, for the entirety of their professional career. It is therefore not possible for someone to be fully informed when they choose their profession, nor is it possible for them to become fully informed during their education and training; nevertheless, the person should be informed about the uncertainty involved with working in the medical profession. In the light of this uncertainty, answers such as "if I had known I would not have chosen it" are not very meaningful, because there is no way to anticipate all the potential risks one might encounter in the course of a professional life. The only sure thing amid the uncertainty is that diseases such as SARS and avian 'flu will always continue to emerge. So far in the discussion, the difficulties of satisfying the conditions needed to validate the presumed consent argument have been described. It seems virtually impossible to fulfill all these conditions satisfactorily. At this point it is important to discuss two particular problems regarding the argument itself. First, it seems nearly impossible to evaluate individually whether the choice made by every HCW to enter the medical profession was informed. The only way to do that would be laboriously to ask each HCW whether they were informed when they decided on the health care profession. Even then, irrespective of whether the person's answer to this question reflects the truth, the only thing that could be learned from such a broad and extensive interrogation would be the individual's perception, not their actual knowledge. By extension, we could not question the participants' level of knowledge in this study, but rather how they perceive their level of knowledge. Because it is futile to seek objectivity in people's perceptions, it would not be sound to use those perceptions to determine whether the HCW's choice was informed, and thus whether they have taken on the duty to treat. And if these perceptions are regarded as subjective, as they should be, then it would be very difficult to develop a set of standard criteria that could be used to establish whether a HCW has been informed. This would also complicate efforts to reach a consensus on forming criteria by which various levels of risk could be defined universally. Unfortunately, such difficulties can only hamper efforts to protect the right of every patient to receive the best treatment available. The second problem is that there is very likely to be more discrimination against patients with certain diseases, as HCWs use this argument to justify their refusal to treat those diseases. The World Medical Association's Declaration of Geneva, which states the basic moral values of the medical profession, specifies that there should be no discrimination, regardless of the circumstances: "I will not permit considerations of age, disease or disability, creed, ethnic origin, gender, nationality, political affiliation, race, sexual orientation, or social standing to intervene between my duty and my patient" [17] . However, the physicians' right to choose their patients is described by the same organization in another policy called "Twelve Principles of Provision of Health Care in any National Health Care System" [18] : "Any health care system should allow the patient to consult the physician of his choice, and the physician to treat only patients of his choice, without the rights of either being affected in any way." For regulations at the national level, it is generally held that if there is an urgent need for medical care [19, 20] , or if no other physician is available to whom the patient can apply [21, 22] , this right would be negated and the available physician must treat the patient. If we look at these obligations in reverse, we can see that the physician might refuse to provide health services to the patient if there is no urgent need for medical care and/or if another physician is around to whom the patient can be referred. Besides, the physician might also refuse the patient on the grounds that their prejudices may adversely affect the advice or treatment that they provide [20] . In actuality, this flexibility was written into the regulations in order to ensure that the patient receives the highest quality care possible; but as can be seen, it could also be abused. And if the presence or absence of consent is used as a criterion to define the duty to treat, in addition the existing flexibilities, then a mechanism would be created by which HCWs may freely, and perhaps excessively, discriminate among patients. To summarize, considering all the points discussed above, the soundness of the presumed consent argument can be doubted. Therefore, it should not be claimed that there is a duty to treat on the basis of the presumed consent argument, as the argument itself is not persuasive. It is also important to note that 5.2% of our participants said that they would not have chosen this profession if they had been informed of the risks. In other words, most of those HCWs who claimed that they were uninformed of the occupational risks when they entered the faculty, or were not fully informed of them during their education period, stated that they still would have chosen the medical profession even if they had been more aware of the risks. This finding tells us that, generally speaking, HCWs place relatively little importance on being informed beforehand. Further support for these findings comes from the answers given by the participants to the other questions. Nearly half said that there is no SLR, and the other half felt that the diseases they evaluate would not surpass SLR if the appropriate protective measures are available. Also, Table 2 indicates that at least 28.3% of the participants thought that none of the diseases listed in that table would exceed the SLR regardless of circumstances. It can therefore be concluded that a large majority of the HCWs place more emphasis on their working conditions than on being informed beforehand. In addition, the criteria most commonly stated by the participants for determining the SLR were the likelihood of transmission of a disease, whether protective measures are available and whether immunization is possible. Each of these criteria is related to protecting the HCW from infection, not to the treatment or the effects of a particular disease. This means that as long as protective measures are available, the HCWs would regard a given disease as below the SLR, so it has nothing to do with being informed beforehand. Besides, the only disease used as an example in this study that could be claimed to exceed the SLR was SARS; AIDS, hepatitis C, hepatitis B, tuberculosis and bacterial meningitis all fell below the SLR according to these criteria. Nevertheless, only 30.9% of the participants suggested that SARS would surpass the SLR. To put this into perspective, SARS was not even observed until 2003, and the research in the present study was conducted in 2004 and 2005. Therefore, most participants in this study were not aware of SARS when they chose the medical profession, nor were they ever informed of it during education or training. They nonetheless felt that the duty to treat pertained even to patients with SARS. All of this suggests that factors more useful and relevant than presumed consent influence the decision of HCWs to choose and continue in the medical profession; these factors may include the social contract between society and the medical profession, and the greater ability of HCWs to provide medical care [16] . It is these factors that should be investigated and emphasized when defining a moral duty to treat. This study could be limited by several factors. The first limitation could be due to a socially desirable response bias; some participants might have given what they perceived as the 'right' answers to the questions rather than the answers that reflect their opinion or belief. In order to address this concern, future studies could benefit by using qualitative methods, which provide more reliable results about the motives and opinions of participants. Also, this study was not prospective, so recall bias might have affected the responses of the participants. Furthermore, the extent to which the results of this study are applicable to HCWs such as nurses or physicians who work in internal specialties is uncertain. Future studies that include other HCWs as participants may broaden our understanding of the beliefs and opinions of HCWs, thereby allowing us to state our claims and shape our arguments more precisely. Finally, it should be mentioned that the response rate for this study (61.7%) was slightly lower than is generally expected for a survey. Nevertheless, despite all these methodological limitations, we believe that our findings support our conclusion about the persuasiveness of the presumed consent argument. If we use the presumed consent argument to establish the duty of the HCW to provide care, we are confronted with problems ranging over the difficulty of choosing a profession autonomously, the constant level of uncertainty present in the medical profession, the near-impossibility of being able to evaluate retrospectively whether every individual was informed, and the seemingly inescapable problem that this practice would legitimize, and perhaps even foster, discrimination against patients with certain diseases. Our findings suggest that another problem can be added to the list: one-fifth of the participants in this study either lacked adequate knowledge of the occupational risks when they chose the medical profession or were not sufficiently informed of these risks during their faculty education and training. As we stated above, in order for a candidate HCW to be informed literally, three items should be explained to them: (a) the risk posed by each of the contagious diseases known at that given time, (b) commonly agreed criteria and definitions of situations that would surpass the SLR, and (c) the fact that there will always be a degree of uncertainty involved with working in the medical profession, as new risks may emerge at any point during one's professional life. In this study it has been shown that at least some HCWs may not be informed of (a). Also, it is not currently possible to inform HCWs of (b) since there are no widely-agreed criteria and definitions to allow for a universally accepted SLR; and there is currently no standard education for all HCWs to ensure that (c) is satisfied. Considering this in addition to the problems mentioned above, the third premise of the presumed consent argument appears implausible and, consequently, the duty to treat cannot be grounded persuasively on the consent assumption. It is therefore more useful to emphasize justifications other than presumed consent when defining the duty of HCWs to provide care, such as the social contract between society and the medical profession and the fact that HCWs have a greater ability to provide medical aid. Furthermore, in terms of the moral duty to provide care, it seems that most HCWs are more concerned about the availability of protective measures than about whether they had been informed of a particular risk beforehand. It seems important that further research be carried out to improve understanding of the opinions and perceptions of HCWs and the basis of their definitions, as this information could prove very helpful in defining a duty to treat that can be effectively put into practice. It is also important that a well-organized ongoing educational program that is needs-based and easily accessible be provided to HCWs at both the graduate and postgraduate levels. In particular, this program must be continuously updated regarding AIDS and other diseases that may cause the HCWs to behave discriminatively towards patients, even though these diseases are below the SLR. Such continuing medical education is the best answer to the justification "When I chose the profession/when I graduated, this disease did not exist!" for refusing treatment. Emphasizing the social role of HCWs, and educating them about the professional obligations derived from the social contract betweeen the profession and the wider social order, would further reduce that kind of reasoning. In addition, stricter standards for the duty to provide care should established by determining the criteria for a SLR and identifying the situations and conditions that would exceed this SLR. Each of these measures could serve to remind HCWs that they have a moral responsibility, as individual HCWs, to be aware of professional obligations and to act as responsible members of the profession. Moreover, the working environment of HCWs should be provided with preventative measures that can be applied both generally and specifically and should emphasize their use. For a circumstance in which a preventative measure has been developed for a disease but is not available for treating a particular case, it would not be easy to justify the claim that there is an undeniable duty to provide care at that moment.

Improved production of human type II procollagen in the yeast Pichia pastoris in shake flasks by a wireless-controlled fed-batch system

BACKGROUND: Here we describe a new technical solution for optimization of Pichia pastoris shake flask cultures with the example of production of stable human type II collagen. Production of recombinant proteins in P. pastoris is usually performed by controlling gene expression with the strong AOX1 promoter, which is induced by addition of methanol. Optimization of processes using the AOX1 promoter in P. pastoris is generally done in bioreactors by fed-batch fermentation with a controlled continuous addition of methanol for avoiding methanol toxification and carbon/energy starvation. The development of feeding protocols and the study of AOX1-controlled recombinant protein production have been largely made in shake flasks, although shake flasks have very limited possibilities for measurement and control. RESULTS: By applying on-line pO(2 )monitoring we demonstrate that the widely used pulse feeding of methanol results in long phases of methanol exhaustion and consequently low expression of AOX1 controlled genes. Furthermore, we provide a solution to apply the fed-batch strategy in shake flasks. The presented solution applies a wireless feeding unit which can be flexibly positioned and allows the use of computer-controlled feeding profiles. By using the human collagen II as an example we show that a quasi-continuous feeding profile, being the simplest way of a fed-batch fermentation, results in a higher production level of human collagen II. Moreover, the product has a higher proteolytic stability compared to control cultures due to the increased expression of human collagen prolyl 4-hydroxylase as monitored by mRNA and protein levels. CONCLUSION: The recommended standard protocol for methanol addition in shake flasks using pulse feeding is non-optimal and leads to repeated long phases of methanol starvation. The problem can be solved by applying the fed-batch technology. The presented wireless feeding unit, together with an on-line monitoring system offers a flexible, simple, and low-cost solution for initial optimization of the production in shake flasks which can be performed in parallel. By this way the fed-batch strategy can be applied from the early screening steps also in laboratories which do not have access to high-cost and complicated bioreactor systems.

The methylotrophic yeast Pichia pastoris is a favored yeast species as a host for heterologous protein production (for reviews see [1] [2] [3] [4] [5] [6] ). P. pastoris has the potential for high expression levels, efficient secretion of target proteins, posttranslational modifications, and is easily grown to high cell densities on mineral salt medium in bioreactors. It has been demonstrated that P. pastoris is an efficient production system also for very large and complex proteins, such as collagens, which besides the recombinant gene(s) needed for the collagen polypeptide chain(s) needs the parallel expression of two different genes coding for collagen prolyl 4-hydroxylase (C-P4H), an enzyme required for the thermal stability of collagens [7] [8] [9] . Although different promoter systems exist for a controlled or continuous expression of heterologous proteins in P. pastoris, frequently the strong and tightly controlled promoter of alcohol oxidase 1 (AOX1) is applied. The AOX promoter is induced by methanol [6] . Methanol, aside from being the inducer of the promoter is also a carbon/ energy substrate. Generally, in such P. pastoris processes, methanol is the only carbon substrate during the production phase of the AOX1-promoter controlled protein. A drawback is the toxicity of methanol. Generally, methanol concentrations above 3.6% inhibit the yeast growth and lead to death [2] . In previous studies for shake flask cultures, the commonly used protocol recommended by Invitrogen [10] for methanol feeding has been modified individually depending on the product. For the expression of proteins in shake flask cultures it is proposed that methanol is added twice per day [11, 12] . However, due to the small window in which the methanol concentration must be kept for an optimal process, pulse-based methanol addition for optimization of processes in shake flasks seems not to be the best way to provide data for the development of a fermentation process. Optimization work for recombinant processes with P. pastoris is generally performed in bioreactors by the fed-batch strategy. The basic scheme follows in principle the original protocol of Brierley and coworkers [13] . A glycerol batch phase is followed by a transition phase and by a methanol induction phase [14] . Thereafter, methanol is added continuously. Different variations of this strategy have been proposed, mostly evaluated as being advantageous for a specific protein (comprehensively reviewed by [2] ). The requirement of bioreactors is a limitation for parallel fast optimization of processes. Principally parallel bioreactor systems exist on the market, but they are generally expensive and non-scalable. Here we describe a solution, which allows the application of the fed-batch principle in shake flasks. We apply a recently described monitoring system [15] which is connected to a radio-modem controlled feeding device, which can be placed directly on shakers. The recombinant protein produced in the present investigation is human type II collagen which is the major collagenous component in the cartilage [16] . Purified recombinant human type II collagen is regarded very useful in applications for cartilage repair [17] . The formation of stable triple helical collagen at physiological temperature requires C-P4H activity [16, 18] . Consequently, the coexpression of C-P4H, an α 2 β 2 tetramer located within the lumen of the endoplasmic reticulum, has been proven essential for the production of temperature-stable recombinant collagens [7, 8] . Recombinant type II collagen has been produced successfully in the baculovirus expression system [19, 20] and in P. pastoris [8, 20] . In this study, application of semicontinuous feeding of methanol to shake flask cultures improved significantly the production of stable human type II collagen, which is documented by protein and mRNA analyses of the collagen and C-P4H. We have shown previously by applying on-line monitoring sensors in shake flasks that the commonly used methanol feeding protocol for shake flask cultures of P. pastoris (two manual pulses of methanol per day) leads to long starvation phases between feeding pulses [15] . Starvation is obvious by a pO 2 peak after exhaustion of methanol, related to declining respiratory activity ( Figure 1A ). We concluded from this data that the methanol feeding in this shake flask process is not optimal. In the first step of optimization, the rapidly increasing pO 2 level after consumption of methanol was used as a signal to manually add a new methanol pulse. Each single methanol dose was the same as in the standard protocol and consequently much more methanol was added during the course of the cultivation. Although the procedure was laborious and difficult to reproduce manually, a 30% higher cell density was obtained and the product related mRNAs showed higher expression, such as the mRNAs encoding alcohol oxidase 1 (AOX), formaldehyde dehydrogenase, yeast PDI, type II procollagen chain and the C-P4H α(I) and PDI/β subunits ( Figure 1B) . Unexpectedly, the amount of collagen II was not increased, however (not shown). Growth parameters (A) and concentrations of product-related mRNA species (B) during cultivation of a P. pastoris for produc-tion of recombinant human collagen II in shake flasks Figure 1 Growth parameters (A) and concentrations of product-related mRNA species (B) during cultivation of a P. pastoris for production of recombinant human collagen II in shake flasks. Cultivation procedures with two methanol pulses per day (control, blue open circles, interrupted blue line) and pO 2 -dependent manual feeding of methanol (red filled circles, red continuous line) were compared. Although the results from the manual feeding experiment were promising, due to the long cultivation time a real optimization of the feeding procedure would be possible only by a computer-controlled feed, which allows the addition of methanol in a quasi-continuous way, similar as in bioreactor fed-batch cultivations. Therefore, a computer-controlled feeding device was developed, consisting of an encapsulated microcomputer which was connected to feeding valves. The microcomputer is controlled by a normal desktop computer over a radio-module working at 868 MHz as shown schematically in Figure 2 . If needed, the feeding device can be powered by a rechargeable lead battery. Methanol was fed from a sterile 50 ml syringe through a silicon tube connection (1 mm diameter) and a steel needle connected to one of the three side necks of the shake flask. The feeding pulse was given by a miniature LFV valve (The Lee Company, U.S.A.) which was integrated into the wireless feeding device. In the performed experiments a flow through the valve was simply created by gravitation over an approximately 1 m height difference, but we could also generate a constant flow of solutions with low viscosity by applying an air overpressure container or a micropump (data not shown). With the developed device and the programmed backchannel software, different feeding protocols could be applied to shake flasks (not shown). Opening of the feeding valves in short time intervals yielded a quasi-continuous flow. The effect of quasi-continuous methanol feeding into the shake flasks was tested by repeatedly performing series of two parallel 200 mL cultures with different computer controlled feeding schemes. In each experiment for one of the shake flasks (reference culture) the feeding strategy was based on the commonly used pulse-feeding method [10, 12] in which a 5% methanol solution was fed twice a day to a final methanol concentration of 0.5% (v/v). Into the other flasks methanol was fed in a quasi-continuous mode. Within the first 2 hours a methanol pulse was given every 12 min and after that the feeding interval was 30 min. Each methanol pulse was 0.8 ml of the 5% methanol solution. The total feeding volume and total amount of added methanol was the same in both parallel flasks. In the reference cultures the addition of a methanol pulse was followed by an immediate drop of the pO 2 level. However, within 5 to 10 h after a methanol pulse the pO 2 raised, indicating exhaustion of the substrate (methanol). The pO 2 dropped to zero only when the shaker was Schematic presentation of the wireless data collection and control system for shake flask cultivations Figure 2 Schematic presentation of the wireless data collection and control system for shake flask cultivations. P. pastoris cultivation in shake flasks with methanol feeding without (A, B) or with (C, D) manual pH adjustment Figure 3 P. pastoris cultivation in shake flasks with methanol feeding without (A, B) or with (C, D) manual pH adjustment. Cultivations were performed in two phases: initial batch phase in BMG-medium and fed-batch in BMM-medium. Dissolved oxygen (pO 2 ) and pH were measured with a wireless measuring system, cell growth was followed by measurement of the OD 600 . Vertical lines represent methanol feeding points in A and C or start of methanol feed in B and D. stopped for sampling (see pO 2 spikes in Figures 3A and 3B), but remained otherwise above 50%. In the flasks with quasi-continuous feeding the first methanol pulse caused a decrease of the pO 2 . Later, when the feeding intervals were longer, the pO 2 level was stabilized to about 90% ( Figure 3B ). Only after 50 h of cultivation, when the methanol started to accumulate, the pO 2 level increased indicating a higher maintenance or toxification. The methanol concentration in the culture medium of the reference culture was close to the detection limit almost during the whole incubation time. Only towards the end of the cultivation a low concentration of residual methanol could be detected in the medium. It should be remarked that the samples for methanol analysis were always collected before the next methanol pulse. Therefore we concluded that methanol had been fully consumed before the next pulse. This was also obvious from the pO 2 profile. In contrast, small amounts of methanol were present in the culture with quasi-continuous methanol addition. Until 50 h of cultivation the methanol concentration was below 1 g L -1 . Only towards the end of the cultivation an increase of the residual methanol concentration was detected. As the pO 2 level was high in the culture with quasi-continuous feeding it was tested whether also a higher feeding rate would be possible. Feeding with a double amount of methanol was not beneficial, but instead led to the accumulation of the methanol concentration above 5 g L -1 already within 20 h after the start of the feeding (data not shown). Without pH control, the pH value declined during the whole cultivation. The decrease occurred stepwise in the reference culture and was clearly related to the feed pulses (see Figure 3A ). At the end of the cultivation the final pH was very low (close to 2), which surely is not beneficial for the cell growth and production. As the pH in fermenter cultivations is usually kept between 4 and 6, we decided to adjust the pH during the experiment to 5.5 by intermittent addition of 10% (v/v) ammonia. The control of the pH did not result in major changes in the general culture parameters ( Figure 3C,D) . The variations, especially the higher methanol concentration in the culture with quasi-continuous feed ( Figure 3D ) may be mainly attributed to the larger amount of samples collected during the cultivation for analysis of the protein product and mRNA levels. All cultures were followed by analysis of collagen II, C-P4H enzymatic activity and product-related mRNAs. The amount of type II collagen chains (139 kD) derived from correctly assembled protease-resistant triple-helical collagen II was analyzed by reducing SDS-PAGE after HCl extraction with pepsin-treatment according to Myllyharju et al. [8] . Stable and correctly assembled collagens are resistant against pepsin. Collagen II showed a clear band ( Figure 4 ) which was significantly stronger in the cultivations with the quasi-continuous feeding profile, compared to the standard pulse feeding profile. The results were reproducible (see Table 1 ) despite experiment to experiment variations. Generally, more correctly assembled triple-helical collagen was produced in the experiments with quasi-continuous feeding of methanol. These findings were supported also by the higher amount of C-P4H activity (see Figures 3C and 3D , cf. Table 1 ). C-P4H activity was measured from several time points and found to be 2 to 8 times higher in different cultures with quasi-continuous methanol feeding compared to the parallel reference cultures in each experiment (see Figure 5 ). Doubling the methanol concentration in the feeding solution (cf. experiment 4 A in Table 1 , and grey bar in Figure 5 ) led to higher activities compared to the reference, but the C-P4H activity was slightly lower than in the experiment with the lower feed rate. Surprisingly, also the cell density was not increased in this cultivation compared to the culture with quasi-continuous feed of more diluted feed. SDS-PAGE analysis of expression of human collagen II in P. pastoris shake flask cultivation (experiment 3, Table 1 ) after 21, 46 and 72 hours cultivation Figure 4 SDS-PAGE analysis of expression of human collagen II in P. pastoris shake flask cultivation (experiment 3, Table 1 ) after 21, 46 and 72 hours cultivation. R represents a reference culture with pulse feeding of methanol and F predetermined quasi-continuous feed. Collagen chains were derived from correctly folded collagen II molecules by HCl-extraction and pepsin digestion. The collagen II chains are marked with an arrow. The effect of the feeding procedure on the formation of the product was investigated in more detail at the mRNA level. A quantitative sandwich hybridization assay which was developed earlier in our laboratory was applied [21, 22] . Figure 6 shows the levels of mRNAs encoding the procollagen II chain, AOX1, EF3 (translational factor), and C-P4H. The level of procollagen II mRNA was already high during the first 12 h when the cultures were grown in a batch mode on methanol in both types of cultivations. In the reference culture with pulse feeding, the procollagen II mRNA level decreased to a very low level after the start of the feeding. In contrast, the mRNA level stayed high during the whole cultivation in the culture with quasi-continuous feeding of methanol, being approximately 10 times higher when compared to the pulse feed method. The level of AOX1 mRNA, encoding the alcohol oxidase enzyme and therefore indicating the induction of methanol as well as an active methanol metabolism, was much lower than the level of procollagen II mRNA. Interestingly, the procollagen II mRNA decreased fast in the reference culture with pulse feeding strongly during the shift from the methanol batch to the pulse feeding, similarly to all other mRNAs analyzed. Possibly this is a result of growth inhibition due to carbon/energy source starvation as shown above. In contrast, the AOX1 mRNA level was approximately the same over the whole cultivation when the quasi continuous feeding mode was applied. The same behavior was detected for the EF3 mRNA, a translation factor, and for the level of the C-P4Hα(I) mRNA. The higher level of the C-P4Hα(I) mRNA correlates well with the higher amount of C-P4H activity obtained. Experiments at the scale of shake flasks are generally batch processes. This restricts the applicability of the results for the further development of fermentation processes, which normally apply the fed-batch principle. Despite of this limitation shake flask cultivations are commonly used due to their simplicity and high flexibility. In many cases it might be useful to apply the fed-batch principle already in the earliest process optimization phase, not primarily for obtaining higher cell densities, but for applying similar physiological conditions as in the later process. As such tools are not commonly available we developed a new feeding system which works wireless and therefore can be easily applied on any shaker. If powered by a lead accumulator, the feeding unit can be C-P4H activity in three different experiments, each with a quasi-continuous feeding culture (black bars) and a reference culture (manual feeding twice a day, white bars) Figure 5 C-P4H activity in three different experiments, each with a quasi-continuous feeding culture (black bars) and a reference culture (manual feeding twice a day, white bars). Additionally, in experiment no. 4 double concentrated methanol was fed (grey bar). mRNA levels of a) procollagen II chain, b) AOX1, c) translation factor EF3, and d) C-P4H α(I) subunit during shake flask culti-vation of P. pastoris Figure 6 mRNA levels of a) procollagen II chain, b) AOX1, c) translation factor EF3, and d) C-P4H α(I) subunit during shake flask cultivation of P. pastoris. Predetermined quasi-continuous feed of methanol was investigated. The cells were first grown in BMG medium and changed to BMM medium at 0 h. The first sample represents the time when the methanol feeding was started (13 h) . Predetermined constant feed (red filled circles); reference culture with pulse Feeding (open blue circles). The data are from experiment 4 in Table 1 . directly placed on the shaking table without the need for any direct electrical contact. This unit works together with the earlier described SENBIT wireless measurement system [15] . Control functions and set-points can be easily programmed in the controlling software (BackChannel) which communicates with the measuring software (SEN-BIT). Recombinant P. pastoris cultures were an interesting test case to apply the feeding system in shake flasks. Especially with the widely used AOX1 promoter system, the results from shake flask experiments may be of low practical value, because the methanol concentration has to be controlled within a narrow concentration window which is not possible in a batch system. The starting point for our optimization approach was the generally applied recommended protocol for pulse feeding of methanol twice a day [11, 12] . This procedure is very different from the various cultivation techniques applied in fermentation processes which all are based on continuous addition of methanol. As model system we used the expression of recombinant human collagen II in P. pastoris, which has been described earlier [8] . The assembly of the triple-helical collagen is directly correlated with the hydroxylation of prolyl residues by C-P4H, which is coexpressed together with the collagen polypeptide chains. It is very obvious that the proposed pulse feeding protocol is unfavorable for collagen production. During cultivation, long phases of carbon/energy starvation between the feeding pulses were observed by measurement of oxygen levels in shake flasks. Presence of these phases was confirmed by analysis of process-related marker mRNAs and analysis of the products. Overall, lower expression of these mRNAs was observed compared to a constant feeding protocol. We could prove that a quasi-continuous feeding in the shake flask scale can significantly improve the conditions for collagen production and leads to increased expression of triple-helical stable collagen by simultaneous enhancement of C-P4H expression and activity. The corresponding mRNAs were also expressed at higher level than in the culture with pulse feeding. Almost 10-fold higher amount of procollagen II mRNA was detected in the cultures with quasi-continuous feeding of methanol compared to the pulse method. Also the expression of the translation factor EF3, which is an indicator of protein synthesis, was higher with constant feeding. Generally, in the case of such optimization approaches, and not limited to shake flasks studies, the quantitative analysis of mRNAs, especially for C-P4H and collagen in our case, seems to be a good indicator for the quality of the cultivation. Although we applied here mainly a predetermined feeding strategy, also other feeding schemes and control principles are applicable with the newly developed feeding system, such as the pO 2 -dependent feeding. Although the DO-Stat control principle did not improve the product yield in our case (data not shown), such extended studies are principally interesting and will widen the usefulness of shake flask cultures. As discussed above, it is generally not the aim of shake flask studies to reach the high cell densities which are commonly reached in bioreactors. In the experiments described here cell concentrations were rather moderate which is due to the lower oxygen transfer rate in shake flasks compared to bioreactors. It was the intention in our experiments to keep the oxygen level high as it is a necessary cofactor in the C-P4H reaction, although for other proteins good production may be achieved also by running the culture into oxygen limitation which has been recently shown for the production of an scFv antibody fragment [23] . Furthermore, the moderate cell densities in our experiments are also due to the use of diluted methanol feed solutions. To avoid long time intervals without any feed and to ensure a feed which is metabolically seen as continuous methanol was fed as a 5% solution, which correspondingly dilutes the medium and results in a lower final cell yield. The use of low-power micropumps instead of microvalves is currently tested, and may provide an improved continuous supply of small liquid amounts. Simple controlled bioreactors can be obtained by applying feeding to shake flask cultivations. Although this will not generally change specific characteristics for shake flask cultures such as the problem to establish specific pO 2 levels and poorly controllable aeration of the cultures, the proposed solution is a way to control the physiological state of the cultures in early screening phases in a similar way as during later process development. Furthermore, the described system is advantageous in terms of flexibility and investment costs compared to parallel bioreactor systems. The comparison of the commonly used pulse feeding of methanol to P. pastoris cultures in shake flasks with the quasi continuous feeding clearly showed clearly that the commonly applied standard protocol is not favorable to optimize recombinant processes, mainly due to the longterm starvation phases. Commonly such optimization is performed in bioreactors, which however are not available for many of the molecular biology laboratories. We show here a simple solution by applying a quasi-continuous feed profile, which is a method commonly used in bioreactors. However, the developed feeding device is not limited to such simple applications, but pre-determined feed functions and even feed-back control of the feed rate on the basis of measured parameters can be easily programmed and principally turn the shake flask into a bioreactor. In our case the quasi-continuous feeding profile increased both the amount of C-P4H and stable collagen II. This clearly demonstrates in the case of one complicated large recombinant protein the power of monitoring and control in shaken cultures. The tools applied here provide a valuable system for parallel optimization at small scale also for other kind of cultivations. The P. pastoris strain sCII7 (Mut + phenotype) contained the pICZB expression vector with the cDNAs coding for the human procollagen II chain and the two C-P4H subunits (kindly provided from Fibrogen Europe Ltd., Helsinki, Finland). The cells were stored as glycerol stocks at -70°C. Cultivation medium and cultivation conditions P. pastoris was cultivated in BMG or BMM culture medium containing 0.1 M phosphate buffer (pH 6.0), 1.34% YNB, 0.4 mg L -1 biotin, and 1% glycerol (BMG) or 0.5% methanol (BMM) respectively [10] . Cultivations were performed in 1 L Erlenmeyer flasks with three baffles and three side necks for the sensors and sampling (Glasgerätebau Ochs GmbH, Bovenden, Germany). The flasks contained 200 mL of BMG and were incubated at 30°C on a rotary shaker (Certomat H, B. Braun Int.) at 250 rpm. When the OD 600 reached the value between 2 and 6, the cells were centrifuged at 3500 rpm for 5 min, and washed with BMM by further centrifugation. The supernatant was decanted and the cell pellet was resuspended in BMM to obtain an initial OD 600 of about 1.5. The baffled shake flasks, containing 200 mL of BMM and the pH-and pO 2 -electrodes, were incubated at 30°C on a rotary shaker at 250 rpm. To maintain the expression of the product, 100% of methanol was added to a final concentration of 0.5% (v/v) twice a day or manually at the time when the pO 2 level increased (shake flask experiment with optimized feeding, cf. Figure 1) . Alternatively, the feeding was performed with the wireless feeding device with a 5% or 10% methanol solution as explained in the results section by a quasi-continuous predetermined feed rate, or on the base of the pO 2 signal (pO-stat principle). In some experiments ammonia (10%) was added manually to keep the pH between 4 and 6, as described in detail in the results section. The SENBIT wireless system (teleBITcom GmbH, Teltow, Germany) was used in the shake flask cultures to follow the pO 2 and pH as described earlier [15] . The feeding module was constructed as described in the results section. The samples were taken with a sterile needle (0.9 × 120 mm) connected to three-way sterile plastic valve with Luer-lock (Oriplast GmbH) inserted into a side neck of the flask and tightened with a rubber seal into 10 ml syringes. After immediate cooling on ice the samples were portioned for the further analyses, with exception for the mRNA samples which were directly chilled in the inhibition solution (see below). Growth was monitored by measuring the turbidity (OD 600 ) at 600 nm. Samples were centrifuged (2 min, 13000 rpm, 4°C) and the cell-free supernatant was analyzed by gas chromatography. The analysis was performed as recently described [24] with the following modifications for P. pastoris. For immediate chilling of metabolic activities P. pastoris samples (4 × 2 mL) were immediately mixed with 200 μL inhibition solution (95:5 v/v ethanol/phenol, pre-cooled at -20°C). After centrifugation (2 min, 13000 rpm, 4°C) the supernatant was removed and the pellet was dispensed in 100 μl of RNA Later (Ambion) and stored at -70°C until analysis. Total RNA extraction was performed with the RNeasy Mini Kit and mechanical cell disruption according to the manufacturers instructions (Qiagen). Oligonucleotide probes ( Table 2) were designed using the CloneManager5 program with following submission of the sequences a NCBI BLAST search [25] to exclude alignments with other genes. HPLC-purified unlabelled and biotin-labeled oligonucleotide capture probes were purchased from Metabion GmbH (Martinsried, Germany). Dig-tail labeling of the detection probes was performed according to manufacturer's instructions using the Roche Dig-tailing kit (2 nd generation, Roche Diagnostics GmbH, Mannheim, Germany). In vitro RNA standards were designed for the quantitative analysis of each gene as described before [22, 24] . Therefore, primers as indicated in Table 3 were used for in vitro transcription (purchased from Sigma-Genosys, Cambridge, UK). Protein analysis P. pastoris cell pellets were broken with zirconia beads both in acidic conditions followed by pepsin digestion for collagen II analysis. Protein samples were analyzed on SDS-page. C-P4H activity was analyzed from P. pastoris cell pellets that were broken with zirconia beads in basic conditions. The activity assay is based on the hydroxylation-coupled decarboxylation of 2-oxo [1-14 C] glutarate with (Pro-Pro-Gly) 10 as the peptide substrate [26] . The total protein concentrations were determined with the RC DC Protein assay Kit (Biorad Laboratories). probes and primers and performed together with MK the mRNA analyses. MR, AN and ERH performed the protein analyses. JM and ERH contributed with their experiences in the collagen production process. MR wrote the initial manuscript, which was read and approved by all authors. PN initiated and practically supervised the scientific work.

HMGB1: Endogenous Danger Signaling

While foreign pathogens and their products have long been known to activate the innate immune system, the recent recognition of a group of endogenous molecules that serve a similar function has provided a framework for understanding the overlap between the inflammatory responses activated by pathogens and injury. These endogenous molecules, termed alarmins, are normal cell constituents that can be released into the extracellular milieu during states of cellular stress or damage and subsequently activate the immune system. One nuclear protein, High mobility group box-1 (HMGB1), has received particular attention as fulfilling the functions of an alarmin by being involved in both infectious and non-infectious inflammatory conditions. Once released, HMGB1 signals through various receptors to activate immune cells involved in the immune process. Although initial studies demonstrated HMGB1 as a late mediator of sepsis, recent findings indicate HMGB1 to have an important role in models of non-infectious inflammation, such as autoimmunity, cancer, trauma, and ischemia reperfusion injury. Furthermore, in contrast to its pro-inflammatory functions, there is evidence that HMGB1 also has restorative effects leading to tissue repair and regeneration. The complex functions of HMGB1 as an archetypical alarmin are outlined here to review our current understanding of a molecule that holds the potential for treatment in many important human conditions.

Activation of the innate immune system is a necessary step in mounting an anti-microbial response to pathogens. Clinicians have long appreciated that infectious insults and tissue injury from sterile inflammatory states produce a similar inflammatory response. Recent advances in understanding the mechanisms of innate immune system activation have pointed to certain pattern recognition receptors, such as the family of Toll-like receptors, as a common pathway for immune recognition of both microbial invasion and tissue injury. By recognizing either pathogens or endogenous danger signals released upon cellular stress or damage, these pattern recognition receptors are capable of alerting the host of danger by activating the innate immune system. In this review, we will describe one such endogenous danger signal, High mobility group box-1 (HMGB1), and its role in the pathogenesis of many disease states. Pathogens have long been known to cause both local and systemic inflammation via activation of the immune system. Classically, descriptions of immune regulation arose from the idea that immune cells can discriminate self from non-self, and activate innate immunity when there is foreign invasion. This model, which stemmed from work done in the 1960s by Burnet and Medawar, has been modified through the years to account for new discoveries in immune cell function (1, 2) . Although the self/nonself model is suitable in explaining immune activation and inflammation that occurs as a consequence of clinical scenarios, such as invasion by foreign pathogens or transplant rejection, it fails to account for the inflammation which occurs in settings such as trauma or autoimmunity, which are void of non-self stimuli. In an attempt to better describe these phenomena, Matzinger has proposed the concept of danger signaling (3) . In this model, initiation of the inflammatory response occurs in response to molecular patterns that are associated with both pathogens and some normal cellular components that are released by damaged cells during both infectious and sterile processes (4) . The concept of cellular communication by "danger signals," whether exogenous or endogenous, reconciled the paradox of immune activation in both infection and sterile injury. In the danger model, immune activation is the result of recognition of molecular patterns by cellular receptors. Molecular elements from pathogens that elicit an immune response (i.e. lipopolysaccharide [LPS], bacterial DNA, viral RNA) are specific patterns termed Pathogen Associated Molecular Patterns (PAMPs). These PAMPs are generally invariant and it is their recognition by the immune system that triggers the inflammatory response (5, 6) . The cellular receptors that recognize these patterns are evolutionarily conserved and called Pattern Recognition Receptors (PRRs) (6) . Inflammatory responses following sterile injury closely mimic those seen during infection, with similar cytokine and chemokine production patterns (7, 8) . Several endogenous molecules that are released during both infectious and sterile inflammation, such as HMGB1, heat shock proteins (HSPs), S100s, and hyaluronan, have been implicated as possessing the capacity to trigger the immune system through PRRs, much like PAMPs (9) (10) (11) . These signals are normal cell constituents and are released either passively by necrotic cells or actively secreted by stressed cells in response to cellular injury. While the term PAMP is restricted to patterns located on pathogens, these endogenous analogues are termed alarmins (12) . The exogenous PAMPs and endogenous alarmins are subgroups of the larger category of danger signals termed Damage Associated Molecular Patterns (DAMPs) (13, 14) . Alarmins are of particular interest because of their role in both infectious and sterile inflammation. They are present either locally or systemically in severe sepsis, burns, infection, arthritis, and cancer (10, (15) (16) (17) . Alarmins are found in a variety of organelles in all cell types studied and maintain functions in normal cellular homeostasis. They are found in the nucleus as transcription factors (HMGB1), in the cytoplasm as calcium regulators (S100s), in exosomes as chaperones (HSPs), or as components of the cell matrix (hyaluronan). Although diverse in their locations during homeostatic conditions, alarmins have many common functional characteristics. In addition to immune activation, an alarmin is released rapidly during necrosis, sequestered in apoptosis, has potential for active secretion by immune cells, and ultimately promotes homeostasis (13) . Because alarmins are a diverse group of ubiquitous molecules impli-cated in nearly all inflammatory states, understanding and ultimately modulating their activity may allow us to control the inflammatory processes. The high mobility group (HMG) nuclear proteins were discovered in 1973 in an effort to better define the specific regulators of gene expression (18) . This group of non-histone, chromatin-associated proteins has since been characterized to be involved in DNA organization and regulation of transcription. This family of proteins shares structural characteristics which make them unique from other chromosomal proteins. These characteristics include transcripts with long AT-rich 3′ untranslated regions as well as carboxy-terminal regions which are highly negatively charged (19) . HMG proteins are constitutively expressed in the nucleus of eukaryotic cells. Collectively, they share functional motifs that bind specific DNA structures and induce conformational changes without specificity for target sequences. HMGB1 is a member of a subfamily of the HMG proteins. On average, HMGB1 is found at concentrations of 10 6 molecules per cell and non-specifically binds to the minor grooves in DNA (19) . This binding is regulated by two 80 amino acid HMG-1 domains (or boxes), each structurally represented as three α-helices in a characteristic L-shaped fold (20) . The nuclear localization of HMGB1 and its affinity for DNA is regulated through phosphorylation and acetylation, and has been found to have a dynamic relationship with chromatin (21) . Like the other members of this protein family, HMGB1 plays an important role in DNA architecture and transcriptional regulation. HMGB1 first was implicated as an important endogenous signaling molecule in 1999 when Wang et al. described the cytokine activity of HMGB1 by identifying it as a late mediator of endotoxinrelated lethality in mice (10) . This study reported increased serum levels of HMGB1 from 8 to 32 h after endotoxin exposure, attenuation of lethality with delayed administration of an HMGB1 neutralizing antibody, and lethality with direct administration of HMGB1 (10) . This groundbreaking work sparked renewed interest in HMGB1 as an important component of the immune response. The specific interactions and functions of the HMGB1 DNA binding domains have been described. As noted above, the structure of HMGB1 contains two separate "boxes," the A-and B-boxes, each containing ~80 amino acids in an L-shaped fold, along with an acidic C-terminal tail (20) . These separate structural motifs appear to function differently when isolated from HMGB1. Several studies have identified the B-box domain as important for many of the proinflammatory properties of HMGB1 including cytokine release (22, 23) . In comparison, the A-box does not possess the pro-inflammatory properties of the B-box and instead competes with HMGB1 for binding sites leading to attenuation of the inflammatory cascade (24) . HMGB1 is released passively during cellular necrosis by almost all cells which have a nucleus and signals neighboring cells of ongoing damage (25) . However, HMGB1 also is secreted actively by immune cells such as monocytes, macrophages, and dendritic cells (10, 16, 26) . Active secretion of HMGB1 is generally through non-traditional, leaderless pathways which are not routed through the endoplasmic reticulum or Golgi apparatus, similar to IL-1 (27) . During normal cellular homeostasis the dynamic relationship of HMGB1 with the nucleus and cytoplasm heavily favors the nucleus. However, HMGB1 localizes in secretory lysosomes when hyper-acetylated on lysine residues and then is released upon appropriate signaling stimuli (21, 28) . When HMGB1 is not acetylated, it remains localized to the nucleus and is not secreted or released, such as during apoptosis (25) . How the acetylation of HMGB1 is regulated is yet to be determined but is surmised to in-volve deacetylases in the nucleus (16, 21, 29) . In other studies, specifically with TNFα stimulated macrophages, the secretion of HMGB1 is dependent upon phosphorylation (30) . While these modifications are clearly important in the release of HMGB1, it currently is unclear how these specific modifications differ. Stimuli for secretion of HMGB1 from immune cells are diverse and include PAMPs, cytokines, and certain states of cellular stress. Macrophages release HMGB1 starting approximately 8 h following exposure to LPS (10,31-33) but do not demonstrate the same response to CpG DNA (34) . It is notable that release due to LPS is dependent, at least partially, on TNFα (31) . HMGB1 also can be released in response to polyinosinicpolycytidylic acid (a model of viral infection) (33, 35) , while in vitro viral stimulation has demonstrated only passive release of HMGB1 due to cytokines or necrotic cells (36) . In addition to these PAMPs that result in the release of HMGB1, endogenous molecules such as cytokines released during other states of injury can result in secretion of HMGB1. While first demonstrated with IFNγ, macrophages also release HMGB1 in response to stimulation with TNFα, IL-1, and oxidative stress (35, (37) (38) (39) (40) . Interestingly, the PAMPs and cytokine stimuli for HMGB1 secretion from macrophages occurs through distinct pathways (35) . For example, while LPS regulates HMGB1 release by hyper-acetylation, TNFα-induced secretion is mediated through phosphorylation, which directs it to the cytoplasm for release (30) . While it was thought initially that HMGB1 was released only from cells of the immune system, other cells have since been demonstrated to actively secrete alarmins. The first non-immune cell to be studied for active HMGB1 secretion was the pituicyte, which was found to release HMGB1 in response to IL-1 or TNFα stimulation (41) . Enterocytes release HMGB1 following stimulation with a cytokine mixture (42) . Hepatocytes also can release HMGB1 in response to hypoxic conditions or oxida-tive stress; this release appears to be mediated by calcium signaling changes within the cell (43) . The inhibition of HMGB1 also has been an important topic for those seeking to ameliorate injury by decreasing the level of HMGB1 release. A variety of pharmacologic agents have been studied for their potential to inhibit release of HMGB1; however, a full discussion of pharmacologic inhibition of HMGB1 is beyond the scope of this paper. It is worth noting, though, that HSP72, an endogenous molecule that has itself been indicted as an alarmin, has been demonstrated to inhibit HMGB1 release. Originally, it was demonstrated that brief heat shock resulted in decreased HMGB1 release from LPS-stimulated macrophages, but no specific pathway was identified (32) . Recently, HSP72 overexpression has been shown to inhibit HMGB1 release from macrophages in response to LPS, TNFα, or oxidative stress (hydrogen peroxide). This inhibition appears to be due to intranuclear interactions between HSP72 and HMGB1 (44, 45) . As noted above, PRRs are a group of molecules that recognize the molecular patterns of DAMPs. These receptors may be activated by PAMPs, alarmins, or both, to activate the immune system. Several important receptors have been implicated in HMGB1 signaling, including the receptor for advanced glycation end products (RAGE) and members of the Toll-like family of receptors (TLRs). RAGE is a transmembrane protein expressed at low levels in normal tissues that is upregulated at sites where its ligands accumulate (46) . The receptor first was identified to bind advanced glycation end products in diabetes, but has since been identified to bind other ligands and is involved in multiple inflammatory states (46, 47) . RAGE was also the first receptor demonstrated to bind HMGB1 (48) . At the time, the consequences of HMGB1 interaction with RAGE were unknown, but it was discovered later that HMGB-1 signaling through RAGE promotes chemotaxis and the production of cytokines in a process that involves the activation of the transcription factor nuclear factor-κB (NF-κB) (49, 50) . Other RAGE-dependent effects of HMGB1 appear to involve the maturation (23, (51) (52) (53) and migration (38, (53) (54) (55) (56) of immune cells as well as the upregulation of cell surface receptors (57) (58) (59) . In addition to RAGE, the Toll-like family of receptors has been demonstrated to be important in HMGB1 signaling. Members of the TLR family share many structural similarities, both extracellularly and intracellularly, but they differ from each other in ligand specificity, expression patterns, and, in some instances, the signaling pathways they activate. Generally, TLRs can recognize both DAMPs and PAMPs and hence are involved in immune response to both infection and injury. Members of the Toll family have specific ligands, and TLR4 might be the most well known and for its role in the bacterial endotoxin (LPS) recognition complex (60) . TLR4, TLR2, and TLR9 have all been implicated as HMGB1 receptors. HMGB1 binding of TLR2 and TLR4 results in NF-κB upregulation (61) (62) (63) , thus making it likely that HMGB1 stimulation of these receptors can lead to cytokine release. Interestingly, HMGB1-mediated TLR4 activation is different from that resulting from LPS stimulation (50, 62) . For example, when applied to cell cultures, HMGB1 activates both IKKα and IKKβ, whereas LPS only activates IKKβ. Additionally, the MAPK protein activation and cytokine production profiles differ between HMGB1-and LPS-treated cells. There is much speculation that HMGB1 does not act alone in the triggering of receptor activation. Proof of this concept was provided by Tian et al. who demonstrated that HMGB1-DNA complexes activate TLR9 signaling (64) . HMGB1 involvement in TLR9 activation appears to be mediated by HMGB1-DNA complexes rather than HMGB1 alone. While HMGB1-DNA complex stimulation of TLR9 is involved in maturation of immune cells and cytokine secretion (64, 65) , there is some evidence that these complexes may also suppress the immune response in some cell types (66) . In addition to the HMGB1-DNA interactions that activate TLR9, HMGB1 interactions with other cytokines such as IL-1β, IFNγ, and TNFα lead to an increased pro-inflammatory response compared with HMGB1 stimulation alone (67) . Ultimately, whether other interactions or modifications of HMGB1 are required for pattern recognition receptor binding or activation is uncertain. As HMGB1 has multiple downstream signaling responses due to activation of different receptors, it also induces cell specific responses when it stimulates cells of the immune system ( Figure 1 ). HMGB1 induces DC maturation as measured by the increased expression of many cell surface markers as well as the secretion of inflammatory cytokines (23, (51) (52) (53) . Monocytes stimulated with HMGB1 have an increased capacity for adhesion (38) and release numerous cytokines and inflammatory mediators (61, (68) (69) (70) . This effect is augmented when administered concomitantly with other cytokines (67, 71) . Neutrophil stimulation with HMGB1 increases the interaction of MAC-1 and RAGE, thus activating the adhesive and migratory function of these cells (56) . Furthermore, HMGB1 stimulates the production of reactive oxygen species by neutrophils through a TLR4 dependent activation of NAD(P)H oxidase (72) , as well as increases the activation of NF-κB which results in increased production and release of cytokines (50, 62) . T cells stimulated with HMGB1 release cytokines and appear to have increased proliferation, survival, and Th1 functional polarization (23, 51) . The response of endothelial cells to exogenous HMGB1 has been studied in relation to its possible pro-inflammatory role in vascular disease. Endothelial cells release TNFα, IL-8, and MCP-1, all of which enhance the local inflammatory en-vironment (57, 58) . Additionally, HMGB1 stimulation appears to increase the expression of ICAM-1 and VCAM-1 on the surface of endothelial cells, thus increasing the adhesion of inflammatory cells (57, 58) . These effects appear to be at least partially mediated by RAGE, which also appears to be upregulated in HMGB1stimulated endothelial cells (57, 58) . In contrast to their role in promoting inflammation, the ability of alarmins to promote tissue repair and regeneration is of increasing interest (73) . Importantly, HMGB1 induces migration of stem cells toward inflamed regions to promote repair and regeneration (74) . Furthermore, it results in increased mesangioblast and endothelial proliferation and migration to sites of inflammation and induces transit of these cells across the endothelial layer (75, 76) . Myoblasts stimulated by HMGB1 migrate toward damaged regions and stimulate repair (77, 78) . These effects on specific cell types are reflected in increased regeneration at the tissue level. In smooth muscle, HMGB1 induces proliferation and rapid changes in cellular architec-ture leading to cell migration (79, 80) . In skeletal muscle, HMGB1 promotes increased myogenesis and angiogenesis (49, 76, 77) . Further confirmation of this is reflected in the finding that inhibition of HMGB1 signaling leads to decreased vessel density and tissue regeneration (77) . Additionally, direct injection of exogenous HMGB1 to peri-infarcted cardiomyocytes in mice results in increased myocytes within the infarcted area and ultimately improved outcomes by both structural and functional measures (81) . In an animal model of wound healing, topically applied HMGB1 accelerates the process in diabetic mice while inhibition of HMGB1 signaling slows it in normal mice, thus implicating functional HMGB1 signaling as an important component of diabetic wound healing (82) . The ability of HMGB1 to stimulate angiogenesis has been demonstrated independently in studies demonstrating that exogenous HMGB1 stimulates endothelial proliferation and migration (75, 83) . Interestingly, many of these restorative effects are mediated through the same receptors (i.e. RAGE) that mediate the pro-inflammatory properties of the molecule (75, (77) (78) (79) 84) . Taken together, the regenerative properties of HMGB1 represent a potentially important manner by which the functions of an alarmin can be manipulated to promote healing as opposed to injury. While early work on HMGB1 demonstrated its role as a late mediator of sepsis (10), more recently HMGB1 has been implicated as a putative danger signal involved in the pathogenesis of a variety of non-infectious inflammatory conditions including autoimmunity, cancer, trauma, and hemorrhagic shock, and ischemia-reperfusion injury. Furthermore, it has been studied in a number of organ systems including liver, heart, pancreas, brain, bone, and kidney. In addition to these studies highlighting the pro-inflammatory effects of HMGB1 in in vivo models of multiple diseases, there is emerging evidence to suggest that HMGB-1 can participate in tissue repair and remodeling as well as preconditioning; all of which are increasingly being recognized as important capacities of danger signals. Considerable evidence exists implicating extracellular HMGB1 in the pathogenesis of a variety of autoimmune diseases. Anti-HMGB1 antibodies are present in the serum of patients with rheumatoid arthritis and drug-induced systemic lupus erythematosus (85, 86) . Furthermore, HMGB1 has been found to be overexpressed in the extracellular milieu of synovial biopsy specimens in rheumatoid and experimental arthritis (87, 88) . When given via direct intraarticular injection, HMGB1 induces arthritis in mice, while treatment with HMGB1 antagonists ameliorates collagen-induced arthritis in both rats and mice (89, 90) . Increased extracellular HMGB1 also is detectable in the dermis and epidermis of skin lesions in patients with cutaneous lupus erythematosus and in biopsy specimens of the minor salivary glands of patients with Sjogren's Syndrome (91, 92) . HMGB1 was initially identified as a nuclear DNA-binding protein. As such, it plays a role in the transcription of several genes, some of which include those that have been implicated in cancer development including E-selectin, TNFα, insulin receptor, and BRCA (93) (94) (95) (96) . Furthermore, cancer cells that have undergone necrotic cell death can release HMGB1 into the local microenvironment. Extracellular HMGB1 can lead to chronic inflammatory/reparative responses that, in the setting of cancer, may lead to tumor cell survival, expansion, and metastases (97) . Interestingly, numerous studies suggest that HMGB-1 plays a role in metastasis development, and thus links it to poor prognosis in a variety of cancers including prostate, breast, pancreas, and colon (98) (99) (100) (101) (102) (103) (104) (105) (106) . In contrast to its potential role in tumor growth and spread, it also has been shown that HMGB1-induced TLR4 signaling is required for effective responses to chemo-radiation in established tumors in animal models (107) . These findings are validated further by the observation that node positive breast cancer patients carrying a loss of function TLR4 allele tend to relapse more quickly after treatment with chemo-radiation as compared to patients with wild-type TLR4 alleles (107, 108) . More recently, HMGB1 has been recognized as a tumor derived DAMP capable of recruiting and activating eosinophils. The role of tumor infiltrating eosinophils is still being elucidated as they can either limit tumor growth through destructive effector functions or promote tumor growth through immunoregulation and tissue repair/remodeling (109, 110) . In addition to the evidence linking HMGB1 to established cancers, it has also been linked to the pathogenesis of premalignant conditions. Elevated serum levels of HMGB1 are present in mice with chemically induced colitis. Furthermore, neutralizing antibodies to HMGB1 decrease tumor incidence and size in colitis-associated cancer models (111) . The sum of these findings suggests that HMGB1 plays a role in tumor development, growth, and spread. Additionally, immune responses to established cancers prior to, as well as during, systemic treatment appear dependent upon HMGB1/TLR-4 mediated signaling cascades. Consequently, HMGB1 warrants further investigation as a possible therapeutic target in malignant processes. End organ dysfunction in trauma and hemorrhagic shock results from systemic inflammatory responses (112) . Barsness et al. provided indirect evidence that HMGB1 may be involved in the pathogenesis of end organ injury following hemorrhagic shock by linking TLR4 activation to the development of acute lung injury (113) . Subsequently, it has been shown that pulmonary HMGB1 levels are increased as early as 4 h after the initiation of hemorrhagic shock (114) . Furthermore, HMGB1 neutralizing antibodies ameliorate hemorrhage-induced acute lung injury (114) as well as gut barrier dysfunction and ultimately lead to improved survival (115) . These findings are even more relevant given the clinical observation that circulating HMGB1 levels are elevated in trauma patients with hemorrhagic shock as compared with normal volunteers (115) . Additionally, HMGB1 has been shown to play a central role in the initiation and propagation of the inflammatory response following traumatic injury. In a mouse femur fracture model, administration of neutralizing HMGB1 antibodies results in decreased serum IL-6 and IL-10 levels. This blunted systemic response corresponds with decreased hepatic and gut barrier dysfunction as measured by serum transaminase levels and NF-κB activation, respectively. Interestingly, the contribution of HMGB1 to end organ injury in this model appears once again to be TLR4 dependent, as mutants are protected when compared with wild-types. No additional protec-tion is seen when antibodies are administered to TLR4 mutants (116) . The role of HMGB1 in the pathogenesis of acute or chronic viral illnesses remains elusive. As mentioned, HMGB1 can escape extracellularly upon necrotic cell death. West Nile encephalitis and acute hepatitis are viral-induced pathologies associated with necrotic cell death either via direct cytotoxic effects of the virus itself or the inflammatory response to viral infection (36) . In a recent study utilizing hepatitis B virus (HBV) transgenic mice injected with virus specific cytotoxic T lymphocytes (CTLs), HMGB1 translocates from the nucleus to the cytoplasm of hepatocytes surrounding CTLcontaining necroinflammatory foci. Furthermore, treatment with HMGB1 inhibitors in this model significantly decreases the recruitment of inflammatory cells (117) . In addition to passive release from necrotic cells, HMGB1 can be secreted actively by inflammatory cells (monocytes, macrophages, dendritic cells, etc.). A number of viral illnesses including SARS and influenza result in the elevation of proinflammatory cytokines (TNF, IL-1, IL-6, types I and II interferon) that could be responsible for inducing release of HMGB1 from inflammatory cells, thus leading to the propagation of the immune response (36) . A preponderance of evidence exists implicating HMGB1 in the pathogenesis of ischemia/reperfusion injury (IRI) in multiple organ systems including kidney, brain, heart, and liver. Much of the seminal work identifying extracellular HMGB1 as an alarmin that initiates the inflammatory response resulting from IRI was performed using a model of partial warm hepatic IRI in mice. In this model, tissue levels of HMGB1 are elevated as early as 1 h following reperfusion (118) and continue to increase for up to 24 h following the insult. Furthermore, in this same model, neutralizing antibodies to HMGB1 ameliorate the damage resulting from IRI in a TLR4-dependent manner. The importance of HMGB1 signaling through TLR4 in hepatic IRI was further clarified in subsequent in vivo experiments by the generation of chimeric mice in which TLR4-derived bone marrow cells were shown to be largely responsible for the initiation and propagation of the inflammatory response. In this set of experiments, mice expressing TLR4 mutant bone marrow-derived cells were protected from liver damage as compared to those expressing TLR4 wild type bone marrow-derived cells. This occurred independently of the TLR4 phenotype expressed on hepatocytes (119). Subsequently, it was shown that increasing the number of hepatic dendritic cells worsened liver damage in TLR4 wildtype but not TLR4 mutant animals (59) , thus suggesting that dendritic cells are involved in the recognition of and response to DAMPs (HMGB1 included) released after acute injury in this model. The molecular mechanisms of HMGB1/ TLR4-dependent IRI involve the production of reactive oxygen species in a process that appears to be dependent on calcium signaling via the calcium/ calmodulin-dependent kinases (CaMK) whose activation is ultimately involved in the release of HMGB1 (43) . While the significant body of literature described above elucidates the role of HMGB1 in hepatic IRI, similar work in other organ systems including the heart (120,121), kidney (122) , and brain (123) (124) (125) has provided confirmatory evidence that HMGB1 is involved in the initiation of the inflammatory response following IRI. A well characterized effect of pathogen-associated molecular patterns (PAMPs), such as LPS, is their ability to confer protection when administered in small doses prior to more significant insults in a process that has come to be known known as preconditioning. Therefore, alarmins ought to be able to precondition against later insults as well. Proof of principle of this concept has been demonstrated by HMGB1's ability to provide protection when administered as a precondition agent in hepatic IRI (126) . HMGB1 administration 1 h prior to the onset of injury results in a dose-dependent protection as evidenced by decreased circulating biochemical markers of liver damage as well as decreased serum TNFα and IL-6 levels. This preconditioning effect appears to be mediated through an inhibition of TLR4 signaling (126) . There is a growing body of literature surrounding the specific functions of HMGB1 on cells of the immune system as well as its role in important disease states. As the mechanisms promoting the release of HMGB1 and the signaling pathways it activates remain to be completely elucidated, evidence that suggests its potential as a therapeutic target/agent in various models of inflammation continues to accumulate. Importantly, human-subject studies which demonstrate its elevation in several disease states further indicate its importance. Interestingly, in addition to its proinflammatory roles, HMGB1 also appears to have several regenerative effects that lead to tissue repair. Understanding HMGB1 and its complex effects on the immune system may lead to the development of novel strategies to attenuate inflammation and/or promote tissue repair and regeneration in various clinical states.

Quantitative measurement of thyroglobulin mRNA in peripheral blood of patients after total thyroidectomy

Previous studies have reported the clinical usefulness of reverse transcription-polymerase chain reaction (RT-PCR) detection of thyroglobulin (TG) mRNA in the peripheral blood of patients with differentiated thyroid carcinoma. To evaluate this usefulness, we measured TG mRNA in the peripheral blood of patients diagnosed with thyroid carcinoma after total thyroidectomy by real-time quantitative RT-PCR using glyceraldehyde-3-phosphate dehydrogenase (GAPDH) mRNA as an internal control. Surprisingly, we detected TG mRNA in all samples obtained after total thyroidectomy, including those from 4 medullary carcinomas. Further, there was no statistical difference in expression levels of TG mRNA in the patients with or without metastasis, and no significant correlation was found between serum TG concentrations and the expression levels of TG mRNA. These results give rise to a question regarding the clinical applications of not only RT-PCR detection but also quantitative measurement of TG mRNA in peripheral blood. © 2001 Cancer Research Campaign http://www.bjcancer.com

Thyroid carcinomas often recur many years after surgery (Loh et al, 1997) . The monitoring of serum thyroglobulin (TG) by immunoassay can be used in detecting residual or recurrent differentiated thyroid carcinoma (DTC) after total thyroidectomy. However, the usefulness of this method is limited by both the requirement for thyroid hormone withdrawal to attain optimal sensitivity and interference by antithyroglobulin antibodies (Singer et al, 1996) . Recent reports have demonstrated that the reverse transcription-polymerase chain reaction (RT-PCR) can be used to detect circulating cancer cells in the peripheral blood of patients with malignancies such as prostate cancer (Ghossein et al, 1995) and neuroblastoma (Mattano et al, 1992) . A sensitive RT-PCR assay amplifying thyroid-specific mRNAs such as TG or thyroid peroxidase (TPO) may be utilized for the early detection of DTC recurrence and thus may have important therapeutic and prognostic implications. In fact, 2 recent reports have shown the clinical usefulness of the RT-PCR detection of TG, TPO, and ret /PTC in the follow-up of DTC (Ringel et al, 1998; Tallini et al, 1998) . For example, TG mRNA in peripheral blood became detectable earlier than serum TG in a case of recurrent thyroid papillary carcinoma. Recently, Ringel et al have reported the clinical usefulness of the real-time quantitative measurement of TG mRNA in the peripheral blood of the patients with DTC. In their report, however, they did not use an appropriated internal reference such as glyceraldehyde-3-phosphate dehydrogenase (GAPDH) mRNA . Encouraged by these optimistic reports, we applied this method to our patients for the follow-up of DTC. However, we found that TG and TPO mRNAs were detectable in the peripheral blood of all the patients tested, a result that differs from those reported previously, and we therefore felt a need to re-evaluate TG mRNA detection. One of the problems of previous studies has been that the expressions levels of TG mRNA have not been calculated in conjunction with the use of an appropriate internal control, as the results of simple RT-PCR detection can vary in response to subtle changes in conditions. Patients who have undergone total thyroidectomy constitute good model cases for evaluating this problem. As such, we measured TG mRNA in the peripheral blood of patients after total thyroidectomy and determined the expression levels of TG mRNA by real time-quantitative RT-PCR using GAPDH mRNA as an internal control. We evaluated 57 patients (3 males and 54 females, between 15 and 77 years of age, 49 papillary, 4 follicular, and 4 medullary carcinomas) who underwent total thyroidectomy. Blood was drawn at least 6 months after surgery or the administration of radioactive iodine. Of the 53 DTC patients, 21 had evident metastases (16 lung, 1 bone, 2 mediastinum, and 2 lung and bone) detected by either computed tomography or 131 I scintigram. All the patients with metastasis and 21 of 32 patients without metastasis received 5 to 150 (mean: 55.7 mCi) and 5 to 100 (mean: 17.0 mCi) 131 I, respectively. All the patients after total thyroidectomy, except those with medullary carcinoma, received a suppressive dose of thyroxine so that their serum thyroid stimulating hormone (TSH) levels were undetectable. 17 healthy subjects (5 males and 12 females, between 29 and 59 years of age) with no evidence of thyroid disease were also examined as controls. The protocol was approved by the institutional review boards at the participating institutions, and informed consent was obtained. Peripheral blood was collected in heparinized tubes in 10 ml samples and placed immediately on ice. Next, 10 ml of 3% dextran 200 000 (WAKO, Osaka, Japan) was added to the tubes, which were than placed in ice for 30 min. The supernatant was collected and centrifuged. The precipitant was dispersed in distilled water to lyse the remaining erythrocytes, and an equal volume of 1.8% NaCl was then immediately added. The isolated cells were washed, then centrifuged. Total RNA was extracted following standard procedures (Chomczynski and Sacchi, 1987) and resolved in 20 µl of distilled water, then stored at -70˚C. 1 µl of total RNA was reverse transcribed in an RT mixture containing 50 mM Tris-HCl (pH 8.3), 75 mM KCl, 10 mM dithiothreitol, 3 mM MgCl 2 , 0.5 mM deoxynucleotide triphosphates (dNTPs), 200 U M-MLV reverse transcriptase (Gibco, Gaithersburg, MD), 2 U µl -1 RNase inhibitor (Takara, Shiga, Japan), and 2.5 µM random hexamer (Takara) in a total volume of 20 µl at 42˚C for 60 min. 1 µl of first-strand cDNA was used as a template for the PCR reaction with specific primers for TG cDNA as previously described (Tallini et al, 1998) . Each reaction mixture consisted of 1 µl of cDNA, 0.5 µM of each primer, 1 µl of 10 × Ex Taq Buffer, 0.8 µl of 2.5 mM dNTP mix, 0.5 U of Ex Taq polymerase, and nucleasefree water to a final volume of 10 µl. 10 × Ex Taq Buffer, dNTP mix, and Ex Taq polymerase were obtained from Takara. The reaction mixture was then subjected to 35 cycles of denaturation (94˚C, 1 min), annealing (55˚C, 1 min) and extension (72˚C, 1 min). After PCR amplification, the reaction mixture was run on a 2% SeaKem GTG agarose gel (Takara). The gel was then stained with ethidium bromide. Samples without reverse transcription were used as negative controls. Real-time quantitative RT-PCR (TaqMan PCR) of TG and GAPDH mRNAs using an ABI PRISM 7700 Sequence Detection System and a TaqMan PCR Core Reagent Kit (PE Biosystems, Foster City, CA) was performed as previously described (Takano et al, 1999) . The primers used for amplification of TG cDNA were changed to the same set of primers used in the report by Tallini et al. The probe for TG cDNA used for the TaqMan PCR was: (5′-FAM-CACTTCGAGTTCCAGGAATGGCCTGACCCT-TAMRA-3′). 1 µl of each cDNA was used for measurement of the copy number. The conditions for the TaqMan PCR were as follows: 95˚C for 10 min, followed by 40 cycles of 95˚C for 15 seconds and 60˚C for 1 min. Recombinant pGEM T-vectors (Promega, Tokyo, Japan) containing TG or GAPDH cDNA were constructed by PCR cloning with the same set of primers used in the TaqMan PCR and were used as standard samples. The expression levels of TG mRNA were calculated by dividing the copy number of TG mRNA by that of GAPDH mRNA. Intraassay variabilities were 35% and 24% at 10 and 100 ng l -1 thyroid total RNA, respectively. Interassay variabilities across a 1-month period were 38% and 28% at 10 and 100 ng/L thyroid total RNA, respectively. A portion of normal thyroid tissue from the opposite lobe of a thyroid papillary carcinoma was obtained by surgery. Total RNA was extracted following standard procedures. Serum TG was measured using a commercial radioimmunoassay kit (Ab-Beads Thyroglobulin, Eiken, Tokyo, Japan). Intrassay variability was 8.6% at 5 µg l -1 . Interassay variabilities across a 1month period were 11.8%, 7.4%, 3.1% at 15, 15, 50 µg l -1 , respectively. Assay sensitivity as determined from the 20% interassay coefficient of variation (CV) was 0.5 µg l -1 . The cut off level of this kit was 28.4 µg l -1 . Serum anti-TG antibody was detected with a semi-quantitative microtitre particle agglutination test for the in vitro diagnostic detection and titration of anti-TG antibodies in human serum, SERODIA-ATG (Fujirebio, Tokyo, Japan). Statistical analysis of differences between the groups was carried out using the Mann-Whitney U test. Linear correlation analysis was used to examine the correlation between serum TG and expression levels of TG mRNA in the peripheral blood. Ps of <0.05 were considered significant. TG mRNA was detected in all the samples tested, including the samples from 4 patients diagnosed with medullary carcinoma. Representative data are shown in Figure 1 . The reliability of the quantitative RT-PCR was estimated by adding total RNA from a normal thyroid to that from 10 ml of blood of a healthy subject. The TG/GAPDH mRNA ratio in the sample without the addition of thyroid RNA was 6.75 × 10 -5 . As the thyroid RNA was increased from 10 ng per 1 l blood to 10 µg per 1 l blood, a linear increase in the TG/GAPDH mRNA ratio was observed, indicating the reliability of this quantitative measurement ( Figure 2) . These results were used in the construction of the standard curve in the following experiments. Using this method, the expression levels of TG mRNA in the peripheral blood of 57 patients and 17 healthy subjects were calculated. All the samples were successfully measured by real-time quantitative RT-PCR, which confirmed the expression of TG mRNA in all the samples. There was no statistical difference between the patients with and without metastasis (Figure 3 ). The serum TG was measured in the 49 DTC patients without serum anti-TG antibody, and the set of results were compared (Figure 4) . No statistical differences in the TG mRNA levels were observed. While serum TG was detectable in all the patients with metastasis, it was undetectable in 14 of 29 patients without metastasis. Further, serum TG was undetectable in all 4 patients diagnosed with medullary carcinoma, while quantitative measurement of TG mRNA was possible (Figure 3) . These results suggest that serum TG is a superior marker of distant metastasis of DTC than TG mRNA in peripheral blood. Further, the correlation between TG mRNA and serum TG was estimated in these 49 patients and no correlation was observed between the two ( Figure 5 ). The vast majority of thyroid carcinomas are differentiated tumours that originate from the follicular epithelium and show Total RNA from a normal thyroid was added to that from 10 ml of blood from a healthy subject, after which real-time quantitative RT-PCR, as described in Methods was carried out. The copy number of GAPDH mRNA was simultaneously measured, and the expression levels of TG mRNA were calculated as the ratio of TG and GAPDH mRNA. The results are shown with mean ±SD for triplicate determinations either follicular or papillary structure. Serum TG immunodetection measurements are currently the best available indicator of tumour metastases, even though their diagnostic usefulness has several limitations. Circulating tumour cells can be detected by sensitive PCR amplification of tumour-specific abnormal gene sequences or through the detection of tissue-specific abnormal gene transcripts normally absent in the peripheral blood. Thyroid tumours represent the ideal target for such studies because they actively express tissue-specific markers such as TG and TPO. Despite promising reports in 1998 on the detection of TG mRNA in peripheral blood, our preliminary study showed that TG transcripts were detectable in all samples tested, including the ones from healthy subjects. Recently, 2 studies suggesting limitations in the clinical applications of this method have been reported. Wingo et al have detected TG transcripts in the peripheral blood of healthy subjects that can be measured by real-time quantitative RT-PCR (Wingo et al, 1999) . Bojunga et al have found that TG mRNA expression is not specific to thyroid tissue and is not correlated with a diagnosis of thyroid cancer in patients, and they have recommended further examination of this method by quantitative measurement (Bojunga et al, 2000) . We therefore decided to re-evaluate this method using samples obtained from patients after total thyroidectomy, in which TG transcripts in the circulating thyroid cells from the normal thyroid may be negligible. We detected TG transcripts in all the samples tested, even in that of the patient diagnosed with medullary carcinomas, in whom the circulating of thyroid follicular cells was not likely to occur. Real-time quantitative RT-PCR assay confirmed the expression of TG mRNA in these samples, in which an increase in the fluorescent signal from hybridized TG cDNA-specific probe was observed. By real-time quantitative RT-PCR, high levels of TG mRNA were expressed in some patients without evident metastasis. For example, in one patient, the copy number of TG mRNA in 1 ml of peripheral blood was comparable to that in over 300 pg total RNA from normal thyroid tissue, which is approximately equal to 30 thyroid cells. It is hard to believe that this many thyroid cells still circulated in the peripheral blood of a patient after thyroidectomy without any evidence of a distant metastasis, which could be a source of circulating cells. TSH receptor mRNA, which has previously been considered to be expressed only in the thyroid, was found to be expressed in adipocytes and lymphocytes (Francis et al, 1991; Endo et al, 1995) . The TG mRNA detected by our study thus might not have been derived from thyroid cells; instead, it is most likely that lymphocytes express small quantities of TG mRNA, as they are known to express various kinds of mRNAs such as alpha-fetoprotein (AFP) or carcinoembryonic antigen (CEA) (Lafarge-Frayssinet et al, 1989; Coutelier et al, 1994) , which used to be considered tumour-specific. Further, the expression of TG mRNA is not likely to be limited to thyroid follicular cells, as we have recently found it to be expressed in cell lines derived from lung or gastric carcinomas (data not shown). Possibly, the regulation of TG mRNA expression in lymphocytes is controlled by some unknown factors. If so, the detection of TG mRNA is not likely to become an alternative means of achieving early detection of recurrent thyroid carcinomas because sensitive detection of recurrence would be interfered with by basal expression of TG mRNA in peripheral blood. It therefore seems likely that the diagnostic methods presented by Tallini and Ringel may be useful only in limited populations of patients in whom a considerable number of thyroid tumour cells expressing a high copy number of TG mRNA are circulating. In fact, our results showed no statistical difference between the samples from patients with or without distant metastasis. Further, we obtained no evidence that this method is superior to the conventional measurement of serum TG, as it was hard to utilize the TG mRNA data to differentiate the patients with metastasis from those without, whereas serum TG measurements showed more than 5 µg l -1 in all patients with distant metastasis, and the high values of interassay CV in the quantitative measurement of TG mRNA can be a disadvantage in the follow-up of a patient for a long period. Moreover, TG mRNA and serum TG measured in the same samples showed no significant correlation. These results suggest that the sources of TG mRNA and serum TG might be different, and, as discussed above, the former might not be derived from thyroid follicular cells. The use of TPO mRNA in the detection of thyroid carcinomas has also been reported in conjunction with some studies. However, our preliminary data showed that TPO mRNA was detectable by RT-PCR in all the samples tested, including those from healthy subjects and patients after total thyroidectomy without distant metastasis (data not shown). Considering that TPO mRNA are usually less abundant than TG mRNA in thyroid cells, it is not likely that the use of TPO mRNA instead of TG would improve the clinical usefulness of this method. In summary, we have found no advantages in diagnosing DTC by the quantitative measurement of TG mRNA in peripheral blood. We consider that an intensive re-evaluation, including a determination of what percentage of TG mRNA derives from thyroid follicular or cancer cells, is necessary with regard to this method before considering the clinical applications. This work was supported by a Grant-in-Aid for Encouragement of Young Scientists (to TT; No. 12771474) from the Ministry of Education, Science, Sports and Culture of Japan.

Functional Analysis of the 5′ Genomic Sequence of a Bovine Norovirus

BACKGROUND: Jena Virus (JV), a bovine Norovirus, causes enteric disease in cattle and represents a potential model for the study of enteric norovirus infection and pathogenesis. The positive sense RNA genome of JV is organised into ORF1 (non-structural proteins), ORF2 (major capsid protein) and ORF3 (minor capsid protein). The lack of a cell culture system for studying JV replication has meant that work to date has relied upon in vitro systems to study non-structural protein synthesis and processing. PRINCIPAL FINDINGS: Only two of the three major ORF1 proteins were identified (p110 and 2C) following in vitro translation of JV RNA, the N-term protein was not detected. The N-term encoding genomic sequence (5′GS) was tested for IRES-like function in a bi-cistronic system and displayed no evidence of IRES-like activity. The site of translation initiation in JV was determined to be at the predicted nucleotide 22. Following the insertion of an epitope within the 5′GS the JV N-term protein was identified in vitro and within RNA transfected cells. CONCLUSIONS: The in vitro transcription/translation system is currently the best system for analysing protein synthesis and processing in JV. Unlike similarly studied human noroviruses JV initially did not appear to express the N-terminal protein, presenting the possibility that the encoding RNA sequence had a regulatory function, most likely involved in translation initiation in an IRES-like manner. This was not the case and, following determination of the site of translation initiation the N-term protein was detected using an epitope tag, both in vitro and in vivo. Although slightly larger than predicted the N-term protein was detected in a processed form in vivo, thus not only demonstrating initial translation of the ORF1 polyprotein but also activity of the viral protease. These findings indicate that the block to noroviral replication in cultured cells lies elsewhere.

Jena virus, a bovine norovirus, is a member of the Caliciviridae family of positive sense RNA viruses and was first isolated from the diarrhoeic stools of newborn calves [1, 2] . JV is a type I genogroup III (GIII) norovirus which is closely related to the type II GIII bovine noroviruses Newbury agent 2 and Dumfries [3, 4] . The GIII noroviruses are responsible for causing enteric disease in cattle [2, 5] and, thus, likely share a similar tissue tropism to the human-associated enteric noroviruses. Like human noroviruses [6] bovine noroviruses have a high seroprevalence [4] . JV is therefore a potentially useful model for studying the molecular biology of enteric norovirus pathogenesis and replication. The 7.3 kb polyadenylated RNA genome of JV has been characterised previously [7] and, like other noroviruses, is organised into 3 open reading frames (ORFs). ORF1 encodes the non-structural proteins in the form of a large 185 kDa polyprotein, which is subsequently cleaved into functional replication proteins by the viral encoded 3C-like protease. ORF2 encodes the structural capsid protein (56 kDa) and ORF3 encodes a small basic protein, which has been shown to function as a minor capsid component [8] . JV ORF1 is consistent with other caliciviruses in that it encodes a 39 kDa 2C-like nucleoside triphosphatase (NTPase), a 3C-like protease and a 56 kDa 3D-like RNA-dependent RNA polymerase [7, [9] [10] [11] [12] [13] . However, the genomic sequence within the 59 region of JV ORF1 (59GS) displays a high level of divergence. This divergence is mainly attributed to the presence of several proline-encoding polypyrimidine tracts within the region predicted to encode a 35 kDa Nterminal protein [7] . The predicted size of N-terminal proteins relative to the size of the respective 2C proteins differs within the norovirus genus. Within the GI noroviruses, such as Southampton virus, the N-terminal protein (44.8 kDa) is larger in size compared to the 2C protein (39.6 kDa). This is in contrast to the GII noroviruses, such as Lordsdale virus and Camberwell virus, in that the N-terminal protein is smaller in size compared to the 2C protein [11, 14] . This is also the case for Jena virus in which the predicted JV N-terminal protein (35 kDa) is smaller than the JV 2C protein (39 kDa) [7] . The norovirus N-terminal protein varies in relative size across the genus, and the encoding sequence bears no similarity to other cellular or viral proteins. Alignment of the N-term protein sequences of various noroviruses indicates little similarity between genogroups within the first 180 residues, however towards the Cterminal end of the protein similarity between the amino acid residues increases. Recent studies investigating the functions of the Norwalk virus N-terminal protein have successfully demonstrated association with the Golgi apparatus in transfected cells [15] . In addition this study also identified a picornaviral 2B like region within the N-terminal protein, suggesting that the protein is involved with host cell membrane interactions, reinforcing other findings that have suggested that the Norwalk virus N-terminal protein disrupts intracellular protein trafficking, including proteins destined for the host cell membrane [16] . A 3C protease-mediated cleavage event within the N-terminal protein (37 kDa) was described for Camberwell virus, a genogroup 2 norovirus, yielding proteins of 22 kDa and 15 kDa [17] . Based on these observations and location within the genome it was hypothesised that the Nterminal protein of noroviruses corresponds to the 2AB region in picornaviruses. Another possibility is that the N-term encoding RNA itself serves to function as a translational enhancer by interacting with cellular proteins involved in translation. Indeed, this phenomenon has been previously reported for Norwalk virus, within which a double stem loop structure has been predicted at the 59 end of the genomic RNA [18] . It was subsequently demonstrated that elements within the 59 end of Norwalk virus bind specifically with cellular proteins such as La, PTB and PCBP2 [19] which have all been implicated in IRES-mediated cap-independent translation in the closely related picornaviruses [20] [21] [22] [23] . In this study the role of the JV 59GS was investigated, including its potential to direct capindependent translation initiation. The precise location of translation initiation in JV was also investigated. Previous studies of norovirus polyprotein processing have yielded three major products following in vitro transcription and translation, representing the uncleaved 3ABCD, N-term and 2C proteins. However, initial analysis of JV polyprotein processing indicated that only two major proteins are synthesised initially which, based on molecular weight predictions, are the 3ABCD (110 kDa) and the 2C (39 kDa). The lack of an N-terminal protein encoded by the JV 59GS, predicted to be 35.3 kDa, is unique among the noroviruses that have been studied in this way. The in vitro transcription and translation profile for JV was therefore studied in more detail. As initial experiments had analysed TNTH reactions following a 1 hr incubation, reaction aliquots were harvested at time points before and after the recommended 1 hr incubation. The results in Figure 1 show that there are no major reaction products synthesised prior to the 1 hr time point, at which time the 3ABCD/p110 and 2C/p39 proteins are clearly visible. Extended incubation past the 1 hr point resulted in further proteolytic cleavage of p110 that coincided with the appearance of proteins of the following sizes: 86 kDa, 55 kDa, and 51 kDa. In addition proteins of 29 kDa, 22 kDa and 20 kDa were also visible at the 24 hr time point (Figure 1, lane 7) . The only protein that was consistently visible following the 1 hr time point was the 2C/p39 protein. Despite prolonged incubation there was no indication that the N-terminal/p35 protein was synthesised. A comprehensive study of polyprotein processing within the murine norovirus (MNV) suggests likely identities for the equivalent proteins in the similar profile for JV [12] . Using region specific antisera the authors were able to identify p110 as the 3ABCD uncleaved precursor, p90 as the 3BCD, p57.5 as the 3D-like polymerase, p52 as a 3ABC precursor and p40 as the 2Clike NTPase, which was determined by mutagenesis and microsequencing experiments. The 19 kDa protein was identified as the 3C-like protease. The antisera used to detect the MNV Nterm protein recognised 3 products; one was the predicted molecular weight at 39 kDa and the other two bands migrated as a 45 kDa doublet. The 59GS region of JV is highly divergent compared to other noroviruses, mainly due to the relatively high cytosine content (32%), which contributes to an overall G/C content of 58%. There are many polypyrimidine tracts within the sequence, potentially yielding a relatively high degree of RNA secondary structure. Previous studies have described potential secondary RNA structure and interaction with proteins involved with IRESmediated translation within the 59 genomic region of Norwalk virus [18, 24] . It was of interest therefore, based on these findings, to ascertain whether or not the 59GS of JV possessed IRES-like properties within the context of a 'Bi-cistronic' expression system, independently of other viral proteins, including the VPg which, in other caliciviruses, has been shown to be associated with translation initiation factors [25, 26] . Traditionally the bi-cistronic vector system has been used to define potential IRES-like sequences from a variety of viral and cellular mRNAs, and is recognized as being the standard test for this function [27] . A bi-cistronic vector is comprised of a 59 and 39 cistron; translation of the 59 cistron being cap-dependent and translation of the 39 cistron regulated by the putative IRES-like sequence. Thus, if the 39 cistron is translated in addition to the 59 cistron then the sequence of interest is said to have IRES-like properties, as translation is initiating internally. To test for IRES-like function in JV, bi-cistronic constructs were made with a cap dependent 59 EGFP cistron and a 39 lacZ cistron under the translational control of either the JV 59GS (pEGFP-C1/ JV59GS/lacZ) or an authentic EMCV IRES (pEGFP-C1/IRES/ lacZ). CRFK cells were transfected with the bi-cistronic constructs and, following incubation, were assayed for EGFP and lacZ expression. Both constructs were able to direct translation of the EGFP cistron effectively as expected (Figure 2a and Figure2d). The use of an authentic EMCV IRES to direct translation of the lacZ cistron was also effective (Figure 2e), with levels of b-galactosidase activity comparable to those of the b-galactosidase reporter ( Figure 2f ). However, no b-galactosidase activity was detected from cells transfected with the pEGFP-C1/JV59GS/lacZ construct (Figure 2b ), demonstrating that the JV 59GS was unable to initiate translation, and therefore, in this context, did not possess any IRES-like functions. As it was clear that the JV 59GS did not posses any IRES-like functions it was necessary to determine the location of translation initiation within ORF1. This was predicted be the ATG encoding methionine at nucleotide position 22, as it is situated in a favourable context for translation initiation [7] . To investigate this multiple translation termination codons (polySTOP) were inserted into the JV genome within the 3B-encoding region, downstream of the 59GS, to halt translation at a defined point. In vitro transcription and translation of this construct would, in theory, yield a product whose size would relate to the initiation codon used within the 59GS (Figure 3 ). To address the unlikely event of translation read-through or re-initiation downstream of the polySTOP, which would result in subsequent translation of the 3C protease and cleavage of the truncated ORF1 polyprotein, a mutation was made within the active site encoding region of the 3C protease within JV ORF1, to prevent any viral mediated cleavage of ORF1 translation products (JV 3C mut /polySTOP). A point mutation of the critical cysteine residue within the highly conserved GDCG motif to a glycine residue was performed, and this approach has been described for the successful inactivation of other norovirus' 3C activity [28] . In vitro transcription and translation analysis was performed on JV wild type ( Figure 4 within the 3C region of JV successfully inactivated the 3C protease, thus a large, .200kDa uncleaved polyprotein is yielded following TNTH. The major product generated by JV 3C mut / polySTOP was calculated to be 103kDa in size. Based on computer predictions this is in agreement with the initiation of translation occurring at nucleotide 22, which demonstrates that the JV N-term protein is translated in full in vitro. At this time it is not possible to determine whether translation of intracellular VPgbound viral RNA initiates at nucleotide 22, although it is likely given the favourable context in which the initiation codon is situated. As the JV N-term was found to be translated in vitro attempts were made to express and purify the protein in bacteria for immunisation so that the protein could be identified by radioimmune precipitation assay (RIPA), as it was possible that the Nterm protein was migrating on gels aberrantly and possibly comigrating with 2C. Attempts to express the protein in bacteria were unsuccessful due to toxicity. Therefore, the 14aa V5 epitope encoding sequence was cloned in frame into the JV cDNA construct at nucleotide position 123 (JV V5). The V5 epitope originates from the P and V proteins of the SV5 paramyxovirus [29] , for which a commercially available monoclonal antibody is used for detection. Following in vitro transcription and translation of JV V5 a new product, approximately 42 kDa in size, was visible ( Figure 5 , lane 2). This product was not observed in any prior analyses of JV. To confirm that this protein was V5/N-term associated the TNTH reaction was subjected to RIPA using the anti-V5 antibody ( Figure 5, lane 3) . This confirmed expression of the N-term protein in vitro. To confirm expression of the V5/N-term protein in cell culture capped RNA was synthesised from the JV V5 T7 cDNA construct, which was used to transfect CRFK cells. As there is currently no host cell line in which to propagate JV the CRFK cell line was used as it has been shown to support the replication of feline calicivirus [30] . Confocal immunofluorescence of transfected cells using the anti-V5 antibody demonstrated expression of the V5/Nterm protein in cultured cells ( Figure 6 ). Expression of the V5/Nterm protein was diffuse and did not co-localise with the Golgi/ ER/plasma membrane marker wheat germ agglutinin (WGA) and therefore displays a different pattern of cellular expression compared to Norwalk virus [15] . Cells transfected with the wild type full length JV RNA were negative for fluorescence (data not shown). Lysates of cells transfected with wild type JV and JV/V5 RNA were subjected to western blot using the anti-V5 antibody ( Figure 7) . No product was present for cells transfected with wild type JV RNA, but a protein of approximately 42 kDa in size was visible in cells that had been transfected with JV/V5 RNA, confirming N-term expression and size as seen in the in vitro system. In addition, this important observation also confirms for the first time that the JV 3C protease was active in cells transfected with capped RNA as the size of the V5/N-term indicated successful cleavage of the protein from the ORF1 polyprotein. To address the issue of potential rapid degradation of the JV Nterm protein CRFK cells were transfected with JV V5 RNA and were harvested at designated time points following the addition of the protein synthesis inhibitor cycloheximide. Cell lysates were analysed by Western blot using the anti-V5 antibody (Figure 8 ). The consistent appearance of the N-term/V5 protein suggested that it is stable and insensitive to degradation by viral and host cell proteases. The predicted molecular weight of the JV N-term is 35.3 kDa, based on the site of initiation of translation and location of conserved cleavage sites. The appearance, therefore, of a previously unseen 42 kDa protein in the in vitro transcription and translation profile was unexpected but this protein does represent a translation product for the JV 59GS. To date, it has not been possible to explain the difference in the predicted and observed sizes for the JV N-term, and the addition of the 14 amino acid V5 epitope within JV N-term does not account for this apparent large shift in molecular weight. However, a recent study described a similar anomaly when investigating proteolytic processing in the murine norovirus MNV-1 [12] . The predicted molecular weight for the MNV-1 N-term protein was 38.3 kDa. The authors successfully generated antisera against the MNV-1 N-term and used it to immunoprecipitate the protein from in vitro transcription and translation reactions and observed that the N-term existed as a 45 kDa doublet, in addition to the predicted size of 38 kDa. However, when MNV-1 N-term antisera was used to probe MNV-1-infected cell lysates only the 43-45 kDa doublet and a large 115 kDa precursor could be detected, suggesting that the predicted 38 kDa form of the N-term is not generated in cell culture. Again, it was not possible to conclusively determine the cause of this discrepancy, but it was speculated that the N-term protein may migrate abnormally in SDS-PAGE, or may be proteolytically processed at a previously unknown cleavage site downstream of the protein's predicted C-terminus. It is also possible that the N-term protein might be modified in some way leading to a shift in observed molecular weight. At this time the same conclusions would seem appropriate for the JV N-term. In addition, it is not known why the JV N-term was previously not detected in in vitro transcription and translation studies prior to the insertion of the V5 epitope. It cannot be ruled out, however, that the wild type JV N-term aberrantly co-migrates with the 39 kDa JV 2C protein in SDS-PAGE. Indeed, the appearance of the V5/ N-term product from transfected cell lysates would appear to be one of a doublet (Figure 8) , also analogous to the observed appearance of the MNV N-term protein in infected cells, suggesting the likelihood of a further cleavage site within the JV N-term protein which has yet to be elucidated. Nevertheless, these studies clearly demonstrate that a protein representative of the 59GS of JV is translated both in vitro and in vivo and is proteolytically processed from the ORF1 polyprotein following translation initiation at nucleotide 22. Human norovirus infection has been shown to be the leading cause of non-bacterial gastroenteritis [31] , however there is currently no cell culture system available to facilitate viral replication and ethical considerations have hindered progress in establishing a permissive human organ culture system. The study of Jena virus offers a potential animal model of enteric noroviral infection. However, until a permissive bovine cell and/or organ culture systems is established analysis of the molecular mechanisms underpinning viral replication and pathogenesis rely upon in vitro systems, most notably polyprotein synthesis and processing. Unlike similarly studied human noroviruses JV initially did not appear to express the N-terminal protein, presenting the possibility that the encoding RNA sequence had a regulatory function itself, most likely involved in translation initiation in an IRES-like manner. This was shown not to be the case and, following determination of the site of translation initiation at the predicted nucleotide 22 the N-term protein was detected following the insertion of an epitope tag, both in vitro and in vivo. Although slightly larger than predicted the N-term protein was detected in a processed form in vivo, thus not only demonstrating initial translation of the ORF1 polyprotein but also activity of the viral encoded protease. These important findings indicate that the block to replication of enteric norovirus in cultured cells cannot be attributed to a failure to synthesise and process the non-structural proteins. The detection of processed and active ORF1 proteins in transfected cultured cells, however, highlights the potential for the development of cell and bovine organ based systems to facilitate the replication of Jena virus. The pEGFP-C1 vector (Clontech) comprises of an EGFP coding sequence under the control of a CMV promoter and a Kozak translation initiation site. Downstream of the EGFP sequence is the multiple cloning site containing unique BglII, SacI, HindIII and ApaI restriction sites. Contruction of pEGFP-C1/JV 59 GS/lacZ was as follows; The JV 59 GS sequence was amplified from the JV full length cDNA clone [7] using Bio-X-Act DNA polymerase (Bioline) with the primers 59 GS F (59-AACTGCA-GATCTTAATAAGTGAATGAAGACTTTGACGAT-39), containing the BglII restriction site (bold) and two in-frame translation termination codons (underlined) to ensure that translation of the EGFP sequence did not carry over to the 59 GS, and 59 GS R (59-AACTGCAAGCTTCTGCAGGACACAATGAGG-39), containing theHindIII restriction site. The JV 59 GS amplicon was ligated to the pEGFP-C1 vector, following restriction enzyme digestion of both amplicon and vector with BglII and HindIII restriction enzymes, and the ligated DNA used to transform E.coli Top10 (Invitrogen). This intermediate construct was named pEGFP-C1/ JV 59 GS. The lacZ coding sequence was amplified from the pSVb-Gal reporter vector (Promega) using Bio-X-Act DNA polymerase and the primers lacZ F (59-AACTGCAAGCTTGA-TATGGGGGATCCCGTCGTTTTACAACG-39), containing the HindIII restriction site (bold) and a kozak translation initiation site (underlined), and lacZ R (59-AACTGCGGGCCCTTAT-TATTTTTGACACCAGACCA-39) containing the ApaI restriction site (bold) and translation termination codons (underlined). The lacZ amplicon was ligated to the pEGFP-C1/JV 59 GS vector following restriction enzyme digest of both amplicon and vector with HindIII and ApaI restriction enzymes, and the ligated DNA used to transform E.coli Top10. The construct was verified by sequencing. Construction of pEGFP-C1/IRES/lacZ was as follows; The EMCV IRES sequence was amplified from the pIRES2-EGFP vector (Clontech) using Bio-X-Act DNA polymerase and the primers IRES Bgl F (59-ACTCGAAGATCTTAA-TAGAGCTTCGAATTCTGCAGTCGA-39), containing the BglII restriction site (bold) and translation termination codons (underlined) to prevent carry over translation as before, and IRES Sac R (59-ACTCGAGAGCTCTGTGGCCATATTATCATC-GTG-39), containing the SacI restriction site (bold). The IRES amplicon was ligated to the pEGFP-C1 vector follwing restriction Whole cell lysate was collected at the following time points following CHX treatment: 0 hr, 1 hr, 3 hr, 6 hr, 12 hr, 24 hr. Bradford analysis was performed on the lysates to ensure equal loading. Following Western analysis the ECL treated membrane was exposed to film for 1 min. Molecular weight marker is represented in lane 1. doi:10.1371/journal.pone.0002169.g008 enzyme digest of both amplicon and vector with BgIII and SacI restriction enzymes, and the ligated DNA used to transform E.coli Top10. The lacZ amplicon described previously was ligated to the intermediate pEGFP-C1/IRES vector following restriction enzyme digest of both amplicon and vector with HindIII and ApaI restriction enzymes, and the ligated DNA used to transform E.coli Top10. The JV 3C protease mutant was created by point mutation of the critical TGT encoded cysteine residue, within the GDCG active site motif, to a GGT encoded glycine residue by mutagenic overlap PCR using Bio-X-Act DNA polymerase. Three rounds of amplification using the JV full length cDNA clone as template were used to generate the final mutant protease cassette. Round 1 used the primers JV F1 (59-CGTCTCAGGGTTGATACT-39) and JV Mut 1 (59-GCAACCACCGTCACCAG-39), yielding a 222 bp amplicon (point mutation nucleotide shown in bold). Round 2 used the primers JV Mut 2 (59-CTGGTGACGGT-GGTTGC-39) and JV R2 (59-TTCCTGGGAGGAACAAGTT-39), yielding a 651 bp amplicon. Amplicons generated in rounds 1 and 2 were pooled to serve as template for round 3 using the primers JV NF (59-ATGTCAACCACCACCAGC -39) and JV NR (59-AAGGGCTCCGGTGAAGG-39). This cassette contained two BclI restriction sites flanking the 3Cprotease active site, as also found in the wild-type full length clone. Restriction digest using BclI was used to remove the appropriate wild-type cassette from the JV full length clone. The mutant cassette was also digested with BclI prior to ligation to the BclI-digested JV full length clone. The ligated DNA was used to transform E.coli Top10, and was designated JV 3C mut . Construction of JV 3C mut /polySTOP was as follows: complementary oligonucleotides with three translation termination codons (underlined) in each reading frame in sense and anti-sense orientations were desgined in such a way that upon annealing the duplex would contain blunt termini. The oligonucleotides were termed pSTOP Top (59-CTAGGTAAGTAAACGCGTCTACT-CACTCAC-39) and pSTOP Comp (59-GTGAGTGAGTA-GACGCGTTTACTTCAATAG-39). Each oligo (1 mg) was incubated with T4 polynucleotide kinase and ATP to phosphorylate the 59 termini, pooled and heated to 75uC for 15 min, and left to cool to room temperature to anneal the oligos. Following purification the polySTOP duplex was ligated to Eco47III digested JV 3C mut , and ligated DNA was used to transform E.coli Top10. The duplex contained the unique restriction site MluI (shown in bold) to assist screening of recombinant clones. The V5 epitope (N-Gly-Lys-Pro-Ile-Pro-Asn-Pro-Leu-Leu-Gly-Leu-Asp-Ser-Thr-C) is recognized by the anti-V5 monoclonal antibody (Invitrogen). Complementary oligonucleotides encoding the V5 epitope were designed in such a way as to generate SacII compatible termini following annealing (bold), and to preserve the reading frame when inserted into the SacII restriction site at nucleotide 123 within the 59 GS of the JV genome (underlined). The oligos were termed V5 Top (59-GGTAAGCCTATCCC-TAACCCTCTCCTCGGTCTCGATTCTACGAGC-39) and V5 Comp (59-TCGTAGAATCGAGACCGAGGAGAGGGT-TAGGGATAGGCTTACCGC-39). The oligos were phosphorylated and annealed as described previously and the duplex ligated to the SacII digested JV full length clone. Ligated DNA was used to transform E.coli Top10. In vitro coupled transcription and translation was performed using the TNTH Coupled Reticulocyte Lysate System (Promega) as per the manufacturer's instructions. Reactions were incubated at 30uC for 1-2 hr. For non-radiolabelled reactions the 35 S-Methionine was replaced with 1 mM unlabelled Methionine (2 ml). Reaction products (1-2 ml) were analysed by SDS-PAGE. Gels were stained and prepared for autoradiography by incubating for 30 min in a solution containing 32 g sodium salicylate, 100 ml methanol and 100 ml dH 2 O. Gels were dried under vacuum and the reaction products were detected by exposure to Kodak X-Omat scientific imaging film (Sigma) at 270uC for 16 hr followed by developing using a Kodak automated developer. Specific V5-tagged proteins synthesised by TNTH were precipitated from 5-10 ml of reaction product using the anti-V5 monoclonal antibody (Invitrogen) at the recommended dilution in 600 ml of 16 RIPA buffer (diluted from 10x stock: 10 mM Tris-HCl (pH 7.5), 1 mM EDTA, 0.15 mM NaCl, 0.1% SDS, 0.5% Empigen BB, 0.1 mM phenylmethylsulphonylfluoride) for 1 hr at 37uC. This was followed by a second incubation of tube for 2 hr rotating at room temperature with goat anti-mouse immunoglobulin G agarose beads (Sigma) to absorb the immune complexes. The beads were washed three times with 500 ml 16 RIPA buffer and once with 500 ml PBS. The beads were resuspended in sample buffer for analysis by SDS-PAGE and autoradiography as before. Endotoxin-free preparations of plasmid DNA were prepared using the GenElute TM Endotoxin free plasmid midi prep kit (Sigma). Crandall-Reese Feline Kidney cells (CRFKs) were seeded into a 12 well tray at approximately 40-50% confluence. CRFK cells were transfected with no DNA (negative control), pSV-b-Gal (control for b-galactosidase activity), pEGFP-C1/JV 59 GS/lacZ and pEGFP-C1/IRES/lacZ (control for IRES activity) using the Superfect TM transfection reagent (Qiagen) as per the manufacturer's recommendations. Following a 16 hour incubation the cells were observed for EGFP expression using a Leica Leitz DMRB fluorescence microscope. The cells were washed in PBS and fixed using a 0.5% solution of glutaraldehyde for 30 min at room temperature. The cells were incubated with an X-Gal stain solution: 5 mM K 3 Fe(CN) 6 , 5 mM K 4 Fe(CN) 6 , 2 mM MgCl 2 , 1x X-Gal (Sigma) for 4 hours at 37uC and were observed for b-glactosidase activity by light microscopy. The experiment was performed more than once to confirm the results. JV V5 and JV FLC T7 cDNA plasmid constructs were linearised using NdeI (Invitrogen). Capped RNA was synthesised using the mMessage mMachineH Capped RNA Transcription kit (Ambion) according to the manufacturer's instructions. CRFK cells were seeded into 6 well trays at approximately 50% confluence and were transfected with 2 mg purified RNA per well using Transmessenger transfection reagent (Qiagen) according to the manufacturer's instructions. For immunofluorescence CRFK cells were seeded onto 19 mm coverslips in 6 well trays and were transfected with RNA as described. Following a 24 hr incubation the coverslips were washed with PBS and fixed in 4% formaldehyde for 15 min at room temperature. Cells were permeabilised and blocked in saponin buffer, also used as staining buffer, (0.1% saponin, 10% foetal calf serum, 0.1% sodium azide) for 1 hr at 4uC. Cells were stained using an anti-V5 monoclonal antibody (Invitrogen) followed by an anti-mouse Alexafluor 488 conjugated secondary antibody (Molecular Probes) at the recommended dilution in staining buffer for 30 min in the dark. Cells were then stained for 30 min in the dark with a Wheat Germ Agglutinin Alexafluor 594 nm conjugate (Molcular Probes) to allow identification of plasma and Golgi membranes. Coverslips were washed and mounted onto slides using Vectashield containing DAPI (Vector Labs). Microscopy was performed using an inverted Leica TCS-NT confocal laser scanning microscope. The anti-V5 antibody was also used to detect V5-tagged protein by Western blot. Cell lysates were prepared following transfection using lysis buffer (0.15 M sodium chloride, 0.5% (v/v) sodium deoxycholate, 0.1% (w/v) SDS, 50 mM Tris-Cl pH 8.0) and protease inhibitor cocktail (Sigma). Lysates were incubated for 15 min on ice followed by sonication to shear genomic DNA. Following Bradford analysis equal protein content from JV V5 and JV FLC lysates were run on a 10% SDS-PAGE gel and subsequently transferred onto Immobilon-P PVDF membrane (Millipore) according to the manufacturer's recommendations. The membrane was probed using the anti-V5 monoclonal antibody at the manufacturer's recommended dilution, followed by an anti-mouse HRP-copnjugated secondary antibody (Santa Cruz) at the recommended diltution. The ECL Western blotting reagents kit (G.E. Healthcare) was used to detect antibody bound protein, which was visualised by exposure to BioMax Light film (Kodak). CRFK cells were seeded into 6 well trays and transfected with capped JV V5 RNA as described. Following a 24 hr incubation cycloheximide (Sigma) was added to the cells at a final concentration of 50 mg/ml. Cells were harvested at indicated times for the preparation of lysates for V5 Western analysis as described above. Bradford reagent (Sigma) was used to ensure equal loading of lysates according to the manufacturer's recommendations.

Preliminary Findings of a Randomized Trial of Non-Pharmaceutical Interventions to Prevent Influenza Transmission in Households

BACKGROUND: There are sparse data on whether non-pharmaceutical interventions can reduce the spread of influenza. We implemented a study of the feasibility and efficacy of face masks and hand hygiene to reduce influenza transmission among Hong Kong household members. METHODOLOGY/PRINCIPAL FINDINGS: We conducted a cluster randomized controlled trial of households (composed of at least 3 members) where an index subject presented with influenza-like-illness of <48 hours duration. After influenza was confirmed in an index case by the QuickVue Influenza A+B rapid test, the household of the index subject was randomized to 1) control or 2) surgical face masks or 3) hand hygiene. Households were visited within 36 hours, and 3, 6 and 9 days later. Nose and throat swabs were collected from index subjects and all household contacts at each home visit and tested by viral culture. The primary outcome measure was laboratory culture confirmed influenza in a household contact; the secondary outcome was clinically diagnosed influenza (by self-reported symptoms). We randomized 198 households and completed follow up home visits in 128; the index cases in 122 of those households had laboratory-confirmed influenza. There were 21 household contacts with laboratory confirmed influenza corresponding to a secondary attack ratio of 6%. Clinical secondary attack ratios varied from 5% to 18% depending on case definitions. The laboratory-based or clinical secondary attack ratios did not significantly differ across the intervention arms. Adherence to interventions was variable. CONCLUSIONS/SIGNIFICANCE: The secondary attack ratios were lower than anticipated, and lower than reported in other countries, perhaps due to differing patterns of susceptibility, lack of significant antigenic drift in circulating influenza virus strains recently, and/or issues related to the symptomatic recruitment design. Lessons learnt from this pilot have informed changes for the main study in 2008. TRIAL REGISTRATION: ClinicalTrials.gov NCT00425893 HKClinicalTrials.com HKCTR-365

A randomised controlled trial of face masks and hand hygiene in reducing influenza transmission in households. Relevant definitions: Index case: the first subject to be infected with influenza in a household. Household contact: any person living in the same household as the index case. Secondary attack ratio (SAR): the proportion of household contacts of an index case who subsequently become infected with influenza. Hand washing: a process for the removal of soil and transient microorganisms from the hands [1] . Hand antisepsis: a process for the removal or destruction of transient microorganisms [1] . The investigational products include surgical face masks, liquid hand soap and hypoallergenic waterless alcohol-based hand cleanser with emollient. The interventions will incorporate distribution of these investigational products and education on their proper use. There is currently concern about the possibility of an impending emerging influenza pandemic. In the event of such, a limited number of interventions would be available to reduce and control the spread of the disease and thus the resulting morbidity and mortality [2, 3] . Antiviral drugs could be used subject to availability, although their effectiveness against the novel pandemic strain is uncertain and resistance could develop quickly with large-scale use. Furthermore a vaccine specific to the pandemic strain would take optimistically, given current technology and production capacity, at least 6 months to develop and mass produce [4] . In addition to vaccination and targeted antiviral prophylaxis, other population-level social distancing measures such as school and workplace closures and travel restrictions are likely to be somewhat effective in reducing influenza transmission in the community [5, 6] , but implementation on a prolonged basis and with repeated waves of the pandemic could be difficult. Household-based quarantine and isolation will likely be effective in mitigating the impact of a pandemic [5] [6] [7] . There is however considerable uncertainty about the efficacy of some non-pharmaceutical interventions at the personal level including face masks and hand hygiene. Our proposed study, to assess the efficacy of masks and hand-hygiene for influenza control, is a direct response to the World Health Organization's recent call for urgent research on the efficacy of nonpharmaceutical public health interventions [3] . In addition to pandemic preparedness, knowledge about the efficacy of masks and hand hygiene would also be important for inter-pandemic influenza control. In western temperate and regional subtropical countries, influenza is a major source of morbidity and mortality during the seasonal periods of epidemic circulation [8, 9] , and a number of measures are typically taken to try to reduce transmission in hospitals and elderly care homes [1] . However there are few data on the efficacy of such measures in the household, although household transmission is thought to be one of the most important settings for the community transmission of influenza [10] . Previous studies have tentatively suggested that prophylaxis with amantadine, rimantadine, and the neuraminidase inhibitors oseltamivir (Tamiflu ® ) and zanamivir (Relenza ® ) are effective in reducing influenza transmission in households [11, 12] . Vaccination is known to be an effective preventative measure [13] provided that the at least one of the vaccine strains closely matches the circulating strain [14] . Protective interventions at the personal level to reduce influenza transmission such as wearing masks and improving hand hygiene are often recommended [15] but few studies have investigated the efficacy of these measures outside nosocomial settings in a rigorous manner. Influenza is thought to be mainly transmitted through airborne droplet nuclei [16] but to a significant extent also spread by hand and surface transfer [17] . Some studies have suggested that transmission of upper respiratory tract infections was reduced after household-based hygiene interventions [18] [19] [20] , while a recent case-control study has suggested that masks and hand-washing may have been effective in reducing the transmission of SARS in a hospital setting [21] . A recent population study has suggested that improved hygiene measures and decreased community mixing during the SARS outbreak in Hong Kong resulted in reduced incidence of respiratory viral infections [22] . The results of this study will have important implications for influenza prevention both in a pandemic and in interpandemic periods. Quantitative estimates of the efficacy of non-pharmaceutical interventions will inform resource allocation under pandemic preparedness plans. Wearing masks may diminish the rate of influenza transmission by reducing the amount of virus-containing droplet nuclei entering the surrounding area after leaving the mouth and nose of an infected subject. When worn by a non-infected subject in the presence of infected airborne droplets, surgical masks may reduce the amount of infected droplets inhaled and thereby reduce the chances of infection. However the degree to which a surgical mask can reduce airborne transmission is difficult to quantify given the lack of prior research in this area, and the expected benefit is uncertain. There are few apparent risks of wearing a mask, perhaps the greatest risk being that the mask engenders a feeling of overconfidence in the ability of the mask to prevent infection, leading to riskier activity (e.g. sitting closer to family members at mealtimes) than might have taken place if the mask were not worn, and thus an increased rather than decreased risk of influenza transmission. Proper hand hygiene is thought to reduce community transmission of some viral infections including rhinoviruses and RSV [10] , but the specific effect of hand hygiene on influenza has not been quantified. Again there are few apparent risks of proper hand hygiene, perhaps the greatest being the detrimental effects on skin of frequent hand washing, particularly with alcohol-based products although most of these contain emollients to buffer against excessive drying. To try to mitigate these effects as much as possible we will only supply an alcohol-based hand cleanser which includes emollients, and encourage study participants to be wary of skin irritation. The other very rare potential adverse consequence, as with all topical applications, is allergic reaction. We will use a hypoallergenic product and exclude participants who have a known allergy to alcohol or additive components of the alcohol handrub deployed. The face masks and hand washing interventions will include an intensive counselling session to describe and demonstrate proper use of the respective hygiene aids. Given that the index case could be symptomatic for a further 5 days, and asymptomatic incubation of the disease in household contact could take 1-2 days before symptoms appear, we propose that the hygiene measures should be maintained for at least 7 days. The trial will be conducted in compliance with this protocol, GCP, and the applicable regulatory requirements. The population studied are households in Hong Kong containing three or more individuals where at least one individual is suspected to be infected with influenza (ascertained either by meeting specific symptom criteria or by a positive result on a rapid diagnostic test, see 4(d)) and where no other household members have experienced symptoms of influenza-like-illness in the preceding two weeks. To quantify the efficacy of face masks and/or hand hygiene in reducing household transmission of influenza. The primary outcome measure is the SAR i.e. the proportion of household contacts with laboratory-confirmed influenza during the study period (follow-up period of 10/7 days in pilot/main study). Inference will be made at the individual rather than household level, adjusting for the potential within-household correlation. The main study will follow a parallel design with three intervention arms (i.e. routine health education only, hand hygiene only, masks and hand hygiene). Households will be randomly assigned to one of the three interventions although intention-to-treat analysis will be at the individual level; therefore this is a cluster-randomised trial design. The pilot study will have three arms (mask only, hand hygiene only, routine health education only). Subject recruitment will take place at selected government general outpatient clinics, group/managed practices, public hospital emergency rooms, private hospital outpatient departments, and private primary care clinics throughout Hong Kong, Kowloon and the New Territories. For the pilot study we propose to recruit 500 individuals with ILI symptoms and apply the QuickVue rapid diagnostic test, so that we can follow-up a maximum of 200 index cases with a positive test result, and their households. We will stop recruiting as soon as we have 200 influenza-positive index cases, and we will stop recruiting after we have used 1,000 rapid diagnostic tests even if we have fewer than 200 influenzapositive index cases. Each household will be randomized to receive one of the three interventions, and all household contacts will be followed up. Details of the power calculation to justify this sample size are given below. Given an average household size of 3.8, a study of 200 households will involve the enrolment of a total of 760 individuals (200 index cases and 560 household contacts). For the main study we propose to recruit approximately 6,000 individuals with symptoms of influenza-like illness (ILI) and follow-up an anticipated 1,200 index cases who meet specific criteria (in most cases a positive rapid diagnostic test result, or in some cases those meeting symptom based criteria -further described in 4(d)) and their households. Each household will be randomized to receive one of three interventions, and all household contacts will be followed up. Given an average household size of 3.8, a study of 900 households (from 1,200 randomized households after an anticipated 25% dropout) will involve the enrolment of a total of 3,420 individuals (900 index cases and 2,520 household contacts). Details of the power calculation to justify this sample size are given in 9(c) below. A cluster-randomisation design has been chosen because simple individual randomisation of household contacts would almost certainly result in crosscontamination between family members who are assigned to different intervention arms. This cluster randomisation design requires a larger sample size to allow for potential within-household variability in SAR. Given the nature of the interventions, participants will be un-blinded to the intervention received although they will be blinded to the nature of the other interventions. The same will apply to our research nurses who will be educating participating households on uses of assigned interventions and prevention of influenza transmission. Nurses will initially be randomly assigned to administering one of the three interventions and will receive training relevant to their assigned intervention only, hence avoiding potential cross-contamination by nurses. When the details of a new index case are uploaded to the online database by fax to the trial manager, a unique identifier will be assigned. A pre-specified table of random numbers will be used to assign one of the three interventions to the household of the index case. Therefore the randomised intervention will be unknown to the doctor at or after the time of recruitment to minimise allocation and ascertainment biases respectively. In the majority of recruiting sites, the criteria for further household follow-up will be a positive result for influenza A or B using the QuickVue rapid diagnostic test on a nose and throat swab (Quidel Corp, San Diego, CA) which has a reported sensitivity of 79% and specificity of 92% for influenza A or B [23] . If the QuickVue test is positive and informed consent is obtained, the index case and their household will be further followed up with home visits as described above. We will calculate the sensitivity and specificity of the rapid diagnostic test in our study setting, using viral culture or PCR of a nose and throat swab as the gold standard. In a small number of recruiting sites at defined periods during the pilot and main studies we will use a symptom-based criteria to determine eligibility for further follow-up; specifically we will enrol subjects and their households for further followup if they present with at least two of the following symptoms: fever≥37.8°C (≥38°C in the pilot study); cough; headache; sore throat; pain in muscles or joints [24] . In a small number of recruiting sites at defined periods during the main study we will use an alternative rapid diagnostic test, namely the HX Diagnostics Influenza A+B test (HX Diagnostics, San Francisco, CA) which is based on 3rd generation lateral flow immunoassay technology and may be more sensitive and specific than the QuickVue test. In this case, subjects with a positive result on the HX Diagnostics Influenza A+B test would be eligible for further follow up as described above. We will calculate the sensitivity and specificity of the rapid diagnostic test in our study setting, using viral culture or PCR of a nose and throat swab as the gold standard. See also 7(c). See 6(a). Following randomization, an immediate home visit will be scheduled (to take place within at most 36 hours, and ideally within 12 hours) to implement the intervention. During the home visit by a trained nurse, the purpose of the study will be explained to all household contacts and their consent obtained. Consent for children aged 17 years or younger will be obtained from their parents. Assent will also be obtained for children aged between 7 through 17. The nurse will then collect details on household composition, risk perceptions, attitudes and beliefs on influenza (questionnaire Q2, appendix), and take a nose swab and a throat swab from each household contact, except for asymptomatic children under the age of 2. Due to concerns about difficulties of taking nose and throat swabs from infants, for household members who are under 2 years of age only information on clinical symptoms will therefore be collected unless they are symptomatic. For participants who are symptomatic, a nose and a throat swab will be collected regardless of their age. The swabs will later be tested to confirm the absence of influenza in any household contact at baseline. The nurse will provide and describe proper use of a free tympanic thermometer, and the daily symptom diaries. Finally, the nurse will administer the standardized intervention through intensive counselling, and demonstration of proper wearing of masks or hand washing. Three and six days (±1 day) after the initial home visit, a trained nurse will revisit each household. During the visit, the nurse will take nose swabs and throat swabs from the index case and all household contacts, except for children under 2 years of age who are asymptomatic, and collect the symptom diaries (questionnaire Q3, appendix) from each household contact. During the final visit (day 6), the nurse will ask the household members to complete a final questionnaire (Q4, appendix) and assess their adherence to the interventions (questionnaire Q5, appendix). The 8-week pilot study will take place from January to April 2007. The exact starting and stopping dates of patient recruitment will not be fixed in advance. Recruitment will begin after the start of the annual influenza peak season (typically Feb/Mar) has been confirmed by the Department of Microbiology, HKU, and will continue until the prespecified sample size is reached. The 39-week main study will take place from January to September 2008. Individuals may at any time decide to stop participating if they wish, without prejudice or any adverse consequences. There are no formal rules for stopping the trial early. When the details of a new index case are uploaded to the online database, a unique identifier will be assigned. A table of random numbers will be generated by the trial statistician prior to the start of the trial, and this will be used to assign one of the three interventions to the household of the index case. Randomisation codes will not be used since for the first home visit it is necessary for the nurse to know which of the interventions has been assigned. Randomisation codes will be masked from those assessing the outcomes. k) The identification of any data to be recorded directly on the case record forms (i.e. no prior written or electronic record of data) and to be considered source data. See questionnaires Q1-Q5 (appendix) for the source data that will be recorded in this study. Note that we will not access subjects' medical records. Nurses will be offered influenza vaccination prior to the study. We will offer each participating household a HK$200 (US$25) supermarket voucher (HK$150 / US$20 in the pilot study). Specimens collected in recruiting clinics will be stored in a 2-8°C refrigerator (overnight, if required). Specimens collected during home visits will be stored at room temperature or if necessary (i.e. in hot weather) in a cool box with at least 2 icepacks immediately after collection. Before the end of the day of a home visit, the study nurse will take any collected samples to the nearest recruitment clinic for storage in a 4°C refrigerator (overnight, if required) or directly to the Department of Microbiology, HKU. Samples stored at 2-8°C in recruiting clinics will be delivered to the Department of Microbiology, HKU as soon as possible, via courier and maintained at 2-8°C en route. Samples will be eluted and cryopreserved at minus 70°C at the destination. The criteria for further follow-up are discussed in 4(d), in the majority of recruiting sites the criteria will be a positive result on the QuickVue Influenza A+B rapid diagnostic test. During the final home visit in the main study we will request buccal swabs from all household members. The samples will be anonymised and then processed by a biotech company specialising in DNA extraction from human samples. These data will be studied at a later date to investigate possible genetic factors in influenza susceptibility or transmission. Inclusion criteria for index cases are as follows: (1) a Hong Kong resident; (2) reporting ILI symptoms including at least two of fever (recorded fever ≥38°C), cough; nasal congestion; sore throat; headache; runny nose and pains in muscles or joints [24] (3) onset of symptoms within the preceding 48 hours. If these conditions are satisfied, the subject will be approached to determine household eligibility to enrol in the study as below prior to further follow-up as described in 4(d) and 4(f). Inclusion criteria for households are as follows: The household must contain at least three people including the index case and any domestic helpers, and where no household contacts have had ILI symptoms in the preceding two weeks. In the main study (January to September 2008), if we are ahead of our recruitment target at any time in March or later, we will consider revising the inclusion criteria to include only subjects with onset of symptoms in the preceding 36 (or 24) hours, since we believe the interventions will be most effective if applied sooner after symptom onset. If implemented, this protocol change would be taken into account in the final analyses. There are no exclusion criteria. The entire household will be withdrawn from the study if there is failure to obtain proper informed consent from any one or more individual household contacts for whatever reason. Informed consent will be sought from each individual household contact, or proxy consent from the parents of any individual under the age of 18. Assent will also be obtained for children age between 7 through 17. Households will be withdrawn from the trial if no home visit can be scheduled within 36 hours. Withdrawn households will be replaced to maintain the specified sample size. If one or more household contacts are not present during one of the home visits, we will attempt to reschedule a supplementary home visit to collect clinical specimens, and in the case where the initial home visit was missed we will request signed informed consent and apply interventions as necessary during the supplementary home visit. In the main study, if a randomized household refuses to allow any home visits, we will request permission to contact them by telephone after 7 days to enquire how long the index case experience symptoms for, and whether any household members reported clinical influenza; if available, these data will allow a simple comparison of households who dropout with those who are successfully followed up. We will randomize households among three study arms, each of which will include intensive counselling during the first household visit. All recruited households will receive educational pamphlets specific to their assigned intervention arms in English or Chinese or both, as appropriate. The intervention arms are as follows: This group will receive education about the importance of a healthy diet and lifestyle for boosting the immune system against influenza and other directly transmitted respiratory pathogens leading to ILI symptoms, both in terms of illness prevention and symptom alleviation. This group will receive the control intervention (health education) plus education about the potential reduction in transmission of respiratory infections to household contacts if all parties maintain proper hand hygiene, and demonstration of proper hand washing and hand antisepsis. Each household member (including the index case) will receive a uniquely labelled 100ml bottle of hypoallergenic waterless alcohol-based hand rub for individual use only, and households will be provided with one 220ml bottle of antimicrobial Ivory liquid hand soap (Proctor & Gamble, Cincinnati, OH) for each washroom and kitchen sink. This group will receive the control intervention plus the hand-hygiene intervention plus education about the potential reduction in transmission of acute directly transmitted respiratory infections to household contacts if all parties wear masks, distribution of 50 (75) surgical masks for each adult (child aged 3-7) household member, and demonstration of proper face-mask wearing. The pilot study will include the first two arms and a third mask-only arm (i.e. the mask intervention but not the hand hygiene intervention). For the first 100 households, we will apply an unbalanced randomisation of 2:1:1 among arms (1), (2) and (3). For the subsequent households (up to a further 100 households) we will apply an unbalanced randomisation of 8:1:1 among arms (1), (2) and (3) to allow us to extract maximum information about the transmission dynamics of influenza in the absence of non-pharmaceutical control measures. If the maximum sample size of 200 households is reached, there will be approximately 130, 35 and 35 households in arms (1), (2) and (3) respectively. Following completion of the pilot study, information derived about the characteristics of influenza transmission will be invaluable not least in validating our sample size calculation for the main study. The main study will randomise households equally among the three arms above, using a block randomisation structure with randomly permuted block sizes of 18, 24 and 30. We will use separate randomisation tables for subjects recruited with different criteria (as described in 4(d)) to ensure the intervention groups are balanced; this is because the QuickVue test is likely to capture subjects with on average higher viral shedding than a symptom-based criteria. There are no restrictions on the use of other medications during the trial period. However the use of antivirals, antibiotics and other Western and Chinese medicine to relieve ILI symptoms during the study period will be recorded by the visiting nurse in the nurse's assessment sheet at the first and last home visits (ie. Questionnaire Q2 and Q4 respectively, appendix). Self-reported use of hygiene measures including mask wearing and hand washing will be recorded in the symptom diaries (Q3) by each household contact and the index case. Overall use will be reported at the final home visit, and quantities of masks, alcohol and liquid hand soap used will be measured by the visiting nurse during the final household visit. The primary outcome measure is the secondary attack ratio (SAR) which is the proportion of household contacts with laboratory-confirmed influenza during the study period (10/7 days after recruitment in the pilot/main studies). We will preferentially use the laboratory definition of influenza rather than the clinical definition, when available. Household contacts will be confirmed to be influenza-free at the first household visit, within 36 hours of recruitment of the index case. In the main study, clinical influenza is defined as the presence of at least two of the following symptoms: feverishness (we will strongly encourage household members to use the supplied thermometer to assess whether a fever is ≥38°C); cough; headache; sore throat; pain in muscles or joints (following [24] ). In the pilot study (as per the initial protocol), clinical influenza is defined as the presence of feverishness, or at least two of the following symptoms: cough; sore throat; nasal congestion, rhinorrhoea, or sneezing; fatigue; headache; stiffness; myalgias. In the main study during the follow-up period of 7 days after recruitment of the index case, the index case and all household contacts will be asked to maintain a daily record of their symptoms (questionnaire Q3, appendix) and their tympanic temperature. A nurse will visit the household on three occasions during follow-up, namely 3 and 6 days after the initial home visit (day 0) with a window period of ±1 day. During each visit the nurse will collect nose swabs and throat swabs from the index case and all household contacts. These will be cryopreserved at the Department of Microbiology HKU as soon as possible as decribed in 4(n). The nose swabs and throat swabs taken during the follow-up visits will provide independent confirmation of the presence or absence of influenza virus in all household contacts, and the duration of viral shedding in the index case. In the pilot study the follow-up period will be 10 days, with visits on days 0, 3, 6 and 9. The primary endpoint of our study will compare the SAR in each of the intervention groups with the control intervention. We will use χ 2 tests and odds ratios adjusting for potential within-household correlation, with a 5% type I error rate. The criteria for further follow-up of the index subject and their household were described in 4(d) above. In the majority of cases, the criteria for further follow-up will be a positive result on a rapid diagnostic case. In some recruiting clinics the criteria will be presentation with at least two of the following symptoms: fever≥37.8°C (≥38°C in the pilot study); cough; headache; sore throat; pain in muscles or joints. However we will only include households in the final analyses if the index subject is laboratory confirmed to have influenza infection. This laboratory confirmation will require a positive result for influenza A or B by viral culture or standard PCR of a nose and throat swab collected from the index case at the recruitment site, and/or during the first home visit. During the first household visit, a responsible adult (usually the household head or a parent) will be asked to provide an overview of the composition of the household, and details on past illness history and influenza vaccinations (Q2, appendix). At the final household visit, the nurse will collect information (questionnaire Q4, appendix) on the overall self/proxy-reported compliance with the intervention, and on any medication taken during the follow-up period, by asking household members and also by personally checking how many masks remain unused, or how much soap or alcohol is left in the bottles and dispensers. There are no safety parameters in this trial. The characteristics of households, index cases and household contacts in the three intervention groups will be compared and assessed for similarity with χ 2 tests, adjusting the comparison of household contacts for potential within-household correlation. In the primary analyses, households will only be included if the index case has laboratory-confirmed influenza infection -all other households will be dropped from the primary analysis. Furthermore, in the main study households will be dropped from the primary analysis if any of the household contacts are found to have laboratory-confirmed influenza infection at baseline although this will not be incorporated in analyses for the pilot study as per the original protocol. Therefore our results will not be biased by the potentially different transmission dynamics of other respiratory diseases compared to influenza A or B, or by the potential for more than one index case in a household when the interventions are applied. The primary endpoint of our study will compare the SAR in each of the intervention groups with the placebo intervention (1). We will use χ 2 tests and odds ratios adjusting for potential within-household correlation [25] , with a 5% type I error rate. We will investigate the efficacy of the interventions on the SAR in multivariable logistic regression models with a generalized estimating equations approach to allow for potential within-household correlation [26] . Analysis will first be performed including only the effects of masks and hand hygiene. Further analyses will allow for the effects of potential confounders on the SAR. Confounders of the SAR that we will assess for each household contact include the age, gender, smoking status, chronic disease status, prior vaccination status, and additionally the age and gender of the corresponding index case. We will investigate the intervention effect in age/gender subgroups, although the statistical power for these analyses is unlikely to be high. We will further investigate the intervention effect in households where the intervention was applied sooner (with 36 hours) after symptom onset in the index case. We will further investigate the intervention effects for influenza A and influenza B separately, although with likely lower incidence the statistical power for the latter may be low. We will also assess the adherence of the index case and household contacts to the interventions, and conduct as-treated analyses of the primary outcome measure. We will conduct sensitivity analyses excluding households where the index case was prescribed antiviral medication, since onward transmission may be less likely in this scenario. In further analyses, we will investigate the effect of the interventions on secondary outcomes listed below, and further adjust for the effect of possible confounders in multivariable logistic and proportional hazards regression models where appropriate. 1. The proportion of household contacts with clinical influenza, adjusting for the potential within-household correlation. 2. The proportion of households with one or more secondary case of influenza (laboratory definition used in preference to clinical definition where available). 3. The proportion of households with one or more secondary case of clinical influenza. We will investigate the predictors of influenza infection and the factors affecting duration of symptoms. We will further examine the effect of environmental and lifestyle factors, and measures of risk perception on the disease course and onward transmission using regression models. We will examine the factors affecting adherence to interventions using regression models. We will develop and apply novel modelling approaches to the analysis of the household infection data to estimate specific transmission parameters, building on previous research [27] . We will investigate the performance of the QuickVue Influenza A+B rapid diagnostic test (and the HX Diagnostics Influenza A+B test) by comparison with the gold standard of laboratory confirmed influenza by viral culture or PCR. We will further investigate the factors potentially affecting rapid diagnostic test performance, including age, gender, and time since symptom onset. If funding is available, we will conduct further laboratory tests of collected samples to allow us to investigate the incidence and transmission dynamics of non-influenza respiratory viruses. Subjects' consent for these additional respiratory virus tests are provided on the current version and all previous versions of the informed consent forms. If funding is available, we will apply quantitative PCR tests to nose and throat swabs collected from home visits, to investigate potential correlation between the degree of viral shedding and onward transmission, as well as the degree of viral shedding in any resulting secondary cases. Finally, if funding is available, we will sequence the genome of influenza viruses detected in index cases and secondary cases to investigate genetic variability in the virus as well as the evolution rate between successive cases (and indeed whether secondary cases were truly infected by their corresponding household index, or from some other source). Pilot study: A simple calculation of the SAR is given by dividing the number of household contacts with influenza by the total number of household contacts. To estimate the anticipated SAR of 0.241 to within ±5% would require 283 household contacts. Given an average household size of 3.8 (i.e. 2.8 contacts per household) we would require at least 101 households in the placebo group; we propose to recruit 130 households in this arm to allow for some households being lost to follow-up. To test the randomization and intervention procedures we will recruit a further 35 households to the second and third study arms (masks and hand-hygiene). This will also enable a preliminary estimate of the efficacy of hand hygiene and masks although the power of the pilot study will be low to detect small or medium effect sizes with statistical significance. More importantly, the pilot study will allow an idea of the feasibility of these interventions, and coherence to them. Main study: For the sample size calculation we require an estimate of the anticipated SAR (P), the degree of within-household correlation (ρ) in the SAR, the relative risk (r) that we would like to detect, and the relevant critical values of the standard normal distribution Z for a specified power (1-β) and type I error rate (α). For average household size m, the required number of individuals n in each intervention arm is given approximately by and thus the number of required households is given by n/m [25] . A recent study of influenza transmission in household contacts in France found a SAR of 24.1%, and a within-household correlation of ρ=0.29 [28] . We will assume a reduced SAR of 20% to allow for the possibility of some transmission occurring prior to randomization and the likely inclusion of some index cases without influenza. A relative risk reduction of at least 30% is generally accepted to be clinically and epidemiologically important. We note that the efficacy of masks and hand-washing were estimated to give relative risk reductions of 90% and 75% respectively during a nosocomial outbreak of SARS, and while we doubt such high efficacy in the household setting we anticipate relative risk reductions of perhaps around 30%-50%, although there is no literature to guide us on such estimates (and hence the need for this trial!). The average household size in Hong Kong excluding houses with single or double occupancy is 3.8 (source: Hong Kong Thematic Household Survey 2002), therefore the average number of household contacts per index case would be m=2.8. We would like to have at least 80% power to detect a 30% reduction (i.e. r=0.7) in the relative risk between intervention 2 (or intervention 3) and intervention 1 (anticipated P=0.2), with a 5% type I error rate. Using the formula given above we calculate that we would require the randomization of 840 household contacts into each arm of the study. Allowing for a 25% dropout rate following randomization, we would require the randomization of 840 household contacts into each study arm corresponding to a total study requirement of 2,520 household contacts in 900 households. Thus we will recruit a total of 3,420 individuals including 900 index cases with positive results on the QuickVue rapid diagnostic test, and 2,520 household contacts. The specified sample size would also have high statistical power to detect larger relative risk reductions if the observed secondary attack ratio were lower (Table 1) . To achieve this sample size, we would need to recruit approximately 6,000 suspected influenza cases and follow-up those who meet the specific criteria (in the majority of recruited subjects this would be a positive result on the QuickVue Influenza A+B rapid diagnostic test), and our reasoning is as follows. During the peak season we would conservatively anticipate that 50% of subjects with ILI symptoms, as we have defined these above, are infected with influenza rather than another virus (3,000 of the 6,000 tested). Recent international oseltamivir trials found that during the influenza peak season the proportion of subjects with influenza-like-illness who had confirmed influenza was 60% [29] and 66% [30] . Allowing for a sensitivity of 79% for the QuickVue test, only 2,370 index cases would be correctly identified (the remainder would be misclassified as false negatives). Given the estimated specificity of the QuickVue test of 92%, testing the 3,000 non-influenza index cases would result in misclassification of 240 subjects without influenza (false positives), and their households would also be visited. While the specificity of the symptom-based definition in 4(d) is likely to be lower, the sensitivity may be higher and with symptom-based recruitment only occurring in a small number of sites the approximate calculation above is appropriate. Thus we anticipate that we would need to recruit approximately 6,000 index cases with ILI symptoms and apply the specific criteria (typically the rapid diagnostic test), of who 1,477 (2,370+240) would meet our criteria and be subject to randomization and follow-up. Given the intention-to-treat approach, randomization and follow-up of some non-influenza cases is an expected consequence of the speed required by this study. d) The level of significance to be used We will use a significance level of α=0.05. Given the short duration of the trial, we do not plan to conduct any interim analyses or specify any early-stopping rules. The laboratory definition of influenza will be preferentially used as the primary outcome measure, but when this is unavailable we will use the clinical definition. In the pilot study, index cases and household contacts with missing data on important predictors will be excluded from analyses. In the main study, we will use multiple imputation [31] with 10 imputed datasets to replace missing values on outcome and predictor variables. If 10 imputed datasets are not sufficient to ensure stability of estimates we will use 20 imputed datasets. Multiple imputation makes maximum use of available data and maximises statistical power while requiring less strict theoretical assumptions than to a complete case analysis, or single imputation of mean values. We note that this is now one of the preferred (and standard) methods for analysing clinical trials data [32] . Any deviations from the original statistical plan will be described and justified in the final report. We will follow an intention-to-treat approach in the analyses. Households without a laboratory confirmed index case, and additionally in the main study those households where a household contact is laboratory confirmed to have influenza infection at baseline, will be excluded from the primary analysis. We will permit direct access to source data and documents for the purposes of trialrelated monitoring, audits, IRB/IEC review and regulatory inspections. As required by the NIH/CDC funding agency, the anonymised individual participant data will be made publicly available after publication of our results in peer-reviewed journals and no later than 24 months after the conclusion of our study. An important ethical consideration is that households randomised to the control intervention might be considered to have less benefit from the trial than those assigned to the mask or hand washing interventions. However we note that there is little evidence that masks or hand washing can reduce influenza transmission, whereas this study will provide that evidence. Further, the participation of a control arm is essential to allow estimation of the effect of the interventions, given the lack of localspecific data on the SAR in typical circumstances. Thus we believe that those in the control arm, as the whole of society, will still benefit indirectly from this research. Socio-demographic and epidemiological data; confidential patient details (name address) will be collected from all subjects via the four questionnaires included in the appendix. All data will be anonymized when entered into an electronic database (double entry) and stored in the Department of Community Medicine, HKU. Original identities will be kept in a separate file accessible only to the trial manager. Original documents will be destroyed at the conclusion of the pilot study. This study is financed by a grant from the Centers for Disease Control and Prevention (Appendix A). The results will be published in international peer-reviewed journals.

Real Time Bayesian Estimation of the Epidemic Potential of Emerging Infectious Diseases

BACKGROUND: Fast changes in human demographics worldwide, coupled with increased mobility, and modified land uses make the threat of emerging infectious diseases increasingly important. Currently there is worldwide alert for H5N1 avian influenza becoming as transmissible in humans as seasonal influenza, and potentially causing a pandemic of unprecedented proportions. Here we show how epidemiological surveillance data for emerging infectious diseases can be interpreted in real time to assess changes in transmissibility with quantified uncertainty, and to perform running time predictions of new cases and guide logistics allocations. METHODOLOGY/PRINCIPAL FINDINGS: We develop an extension of standard epidemiological models, appropriate for emerging infectious diseases, that describes the probabilistic progression of case numbers due to the concurrent effects of (incipient) human transmission and multiple introductions from a reservoir. The model is cast in terms of surveillance observables and immediately suggests a simple graphical estimation procedure for the effective reproductive number R (mean number of cases generated by an infectious individual) of standard epidemics. For emerging infectious diseases, which typically show large relative case number fluctuations over time, we develop a Bayesian scheme for real time estimation of the probability distribution of the effective reproduction number and show how to use such inferences to formulate significance tests on future epidemiological observations. CONCLUSIONS/SIGNIFICANCE: Violations of these significance tests define statistical anomalies that may signal changes in the epidemiology of emerging diseases and should trigger further field investigation. We apply the methodology to case data from World Health Organization reports to place bounds on the current transmissibility of H5N1 influenza in humans and establish a statistical basis for monitoring its evolution in real time.

A pandemic of H5N1 influenza in birds is presently unfolding, with over 50 countries around the world affected, resulting in hundreds of millions of dead animals through infection or culling [1] [2] [3] . This emergency and the associated risk of a devastating new human pandemic [4] [5] [6] stress the need for new approaches targeted specifically at detecting and monitoring the evolution of emerging infectious diseases [7] [8] [9] . Assessing the risk of emergence of a human epidemic at the genetic level requires accounting for rare stochastic events, associated with genetic mutation and recombination, over vast pathogen and host populations [4, 8, 10] . This makes prediction of pathogenic evolution at the molecular level typically still very difficult. Consequently, the first indications of disease emergence are usually observed as infected cases in human and animal populations. Thus, for early assessments of the epidemic potential of a new outbreak, it is essential to assign quantitative meaning to existing epidemiological surveillance data in real time, with quantified uncertainty, and to use this knowledge to enable primary prevention strategies targeted at reducing chances of pathogenic evolution. The quantity that measures the epidemic potential of a pathogen is the basic reproduction number R 0 [11, 12] . R 0 is defined as the average number of new infections created by an infectious individual in an entirely susceptible population. For established human pathogens, leading to standard epidemics, R 0 .1, as is the case of seasonal or pandemic influenza [13] [14] [15] [16] [17] [18] [19] . In practice, epidemiological data typically permit only the estimation of the effective reproduction number R, which may differ from R 0 due to acquired immunity and other factors. For an emerging infectious disease, when transmission is only incipient [20] and the pathogen is adapting to the population, it becomes crucial to monitor quantitative changes of the effective reproduction number over time. Thus, the detection and tracking of an emerging disease can be formalized in terms of monitoring R, as it evolves and approaches the critical threshold RR1. This is likely the current state of H5N1 avian influenza in humans, where complete absence of human to human transmission would imply R = 0, but likely R is very small, as a few cases of possible human contagion suggest [21] [22] [23] . Notwithstanding a marked recent increase in systematic surveillance by national and international organizations, and the advent of real time reporting of many public health indicators (syndromics) [24] , the epidemiological regime of incipient but evolving transmission has received little attention in terms of quantitative modelling [22, [25] [26] [27] [28] . The main difficulty is that data in these circumstances tend to be very stochastic, involve small case numbers and may be plagued by uncertainties and inconsistent reporting. As an example, we contrast in Figure 1 the time series of confirmed new human cases of H5N1 avian influenza in Vietnam, reported by the World Health Organization (WHO), with weekly isolate numbers for seasonal H3N2 influenza in the USA during 2004-2005 (see Methods for ''Data Sources''). The ultimate objective of this paper is to propose a methodology to extract quantitative inferences and generate epidemiological outlook in real time from time series like that of Figure 1a . Recently the problem of real time monitoring of (emerging) communicable diseases has gained growing attention, with a few new methods proposed to estimate R. One method proposes the analysis of the distribution of the sizes of case clusters to provide indications of changes in R. Specifically, increases in R(,1) translate on average into larger case cluster sizes [21, [27] [28] [29] . Another approach [30] relies on the inference of probable chains of transmission among observed cases from knowledge of the statistical distribution of the infectious period. From an ensemble of such chains and their associate compounded probability, R can be estimated. This method has recently been applied to ''real time'' monitoring of SARS [31, 32] , via a Bayesian inference scheme. The strength of this class of methods is that they allow insights into heterogeneities in the population. This demands the consideration of all pairs of possible transmissions, which may become computationally intense as case numbers rise and can be sensitive to under reporting, competing risk and to the details of the distribution of infectious periods. Moreover those studies considered the efficacy of control measures for a disease with an initial R.1 and no new cases introduced during the epidemic, whereas it is typical of emerging communicable diseases that adaptation of the pathogen's tropism to the host population is the result of numerous such introductions [22] . Here we propose an alternative approach, which addresses the issue of new introductions, requires in general smaller computational overhead and results in the estimation of the full probability distribution for R. The method is based on the probabilistic formulation of standard SIR disease transmission models analogous to the time-series SIR (TSIR) approach [33] , which simplifies the need to reconstruct transmission chains by aggregating all infectious and susceptible individuals into classes that are assumed to mix homogeneously. A Bayesian procedure is then employed to translate the time series of case numbers into a probability distribution for epidemiological parameters. The method adopts the standard assumptions made in epidemiological compartment models with homogeneously mixing classes, and benefits from their simpler computational structure allowing efficient estimation with available sparse empirical data. The estimation method developed here has been applied once before [34] to 1918 influenza pandemic death notifications time series for San Francisco, with the purpose of comparing its performance with other conventional methods for estimating R. Here we present its full derivation, provide more details and examples, include introductions from an animal reservoir and show how the method can be used to provide statistical expectations for new case predictions. We also show how case predictions with quantified uncertainty do, in turn, define possible statistical anomalies for future case numbers, which can be used to inform surveillance and logistical management in the event of a new or continuing outbreak. As an example, we apply the method to human case data of H5N1 influenza in Vietnam and Indonesia, to produce bounds on its effective reproduction number, R, and establish a basis for its continued monitoring in real time. Time series of H5N1 influenza cases in humans were assembled from World Health Organization (WHO) reports of confirmed cases (http://www.who.int/csr/don/en/), from January 2004 to June 1, 2006 (see Supplementary Material S1, including Figure S1 , for more information). Data for H3N2 seasonal influenza was obtained from the Centre for Disease Control (CDC) Surveillance Weekly Reports in the United States (http://www.cdc.gov/flu/ weekly/fluactivity.htm). New human cases of avian influenza may result from two alternative processes: i) infection of humans from animal sources [4] , or ii) human to human transmission [23] . For a standard epidemic, explicit consideration of multiple introductions is not important as each case produces many secondary infections. For emerging infectious diseases multiple introductions from a reservoir [22, 25] may constitute an important fraction of all observed cases, and the progression of secondary cases must be carefully assessed and monitored. Our objective is to cast standard SIR-class models in a form that directly relates to time series data of emerging infectious diseases by i) accounting for cases from reservoir sources, ii) casting the model variables in terms of observable quantities reported from field surveillance, iii) formulating the model in a discrete probabilistic form, and iv) quantifying uncertainty in the estimation of epidemiological parameters and future cases, and assimilate new data to reduce it. We consider a standard epidemic susceptible-infected (SIR) model where S(t) is the average number of susceptibles at time t, I(t) is the average number of infectious, N is the size of the population, which decreases due to disease-induced deaths (taken as a fraction a of progressing infections), b is the contact rate, and c 21 is the infectious period. After an average residence time c 21 , infectious individuals recover or die (not shown in [1] ). The total number of cases up to time t, T(t) obeys the equation dT/dt = b S/N I. Epidemic reports most commonly state the occurrence of new infected cases, which over the period t, are given by To find the expression accounting for the evolution of new cases DT(t+t) we integrate Eq. [1] , for I(t) between t and t+t, to obtain reproduction number) is a function of time; the last expression is exact if S(t)/N(t) is constant in the period [t, t+t] . This simplifying assumption is generally excellent for emerging infectious diseases, which result in few cases within a much larger population. Generally the validity of the assumption can be assessed through consideration, from [1] , of its evolution equation This condition is usually satisfied as the fraction of infectious at a given time, I(t)/N(t), is typically less than a few percent (even for seasonal influenza), while other quantities in the product are of order unity. The quantity, in expression , accounting for the number of new cases resulting from infections over time t [35] . To obtain the disease progression in terms of epidemiological observables, we discretize the differential equation for the change in total number of cases between t and t+t as where we used [2] and the assumption that S(t)/N(t) is piecewise constant over [t, t+t], but does vary between intervals contributing to changes in R t . At time t, the total number of cases is also Substituting expression [5] into [4] , we obtain: We see that the well known multiplicative progression between new cases at successive times due to contagion appears, on average, as a linear relation between DT(t+t) and DT(t) in an epidemic time delay diagram, Figures 2a-d. Expression [6] generalizes similar relations in the TSIR literature by casting them in terms of new cases over arbitrarily chosen observation intervals t, not necessarily coinciding with the average generation time c 21 . Expression [6] also shows how the initial R t can be estimated geometrically (without the need for parameter search or numerical optimization) from an epidemic time delay plot of surveillance data: b(R t ) is the slope of the tangent at the origin of case trajectories (dashed line in Fig. 2a, b) . For emerging infectious diseases relative fluctuations in case numbers are large, see e.g. Figure 2d , and this simple geometric approach is not valid, thus making more robust estimation methods, as the one presented here, necessary. For emerging infectious diseases, many introductions from a reservoir may occur before the pathogen adapts its tropism to the new host population and produces epidemic outbreaks [22, 25] . As a result epidemiological models for the time evolution of new cases must account for two processes: (incipient) human transmission and infections from the reservoir. We introduce a new source of infected individuals, through contact with the reservoir (birds). The evolution of I is now given by dI dt~b The first term on the right accounts for the human-to-human infectious process. The last term is a source, creating new I through contact with a reservoir of infectious agents of size K(t), with contact rate b bh . We denote the number of new infections from the reservoir per unit time as dB/dt = b bh S(t) K(t). As a result the number of humans infectious, I, and the total number of cases evolve as The evolution of I(t) between t and t+t, accounting for the effects of the inhomogeneous source term, is where Y(t, t ,B) denotes the integral. We use this expression to solve for the number of new cases (Eq. [8] ), giving Probabilistic models of contagion A probabilistic description is crucial for realistic modelling of new cases of emerging infectious diseases, which are typically characterized by large coefficients of variation. This probabilistic description is achieved, as in [33] , by defining the number of new cases, DT(t+t) as a stochastic discrete variable generated by a probability distribution with average number of cases given by [10] , i.e. where P{l} denotes a discrete probability distribution with mean l. If the only information on case evolution is their average number then the maximum entropy distribution for P{} is Poisson, which we adopt throughout this paper. If additional information on the magnitude of fluctuations were also known, a generalized distribution should be employed, such as a negative binomial [33] , which can account for clumping effects (we reproduce all estimations from the main paper using a negative binomial in Supplementary Material S1, Table S2 ). To use expression [11] in practice, we need to evaluate the integral Y. Assuming the introductions from the reservoir per unit time to be approximately constant between t and t+t, the integral can be calculated to first order as Y(t,t,B) = t dB/dt and, [11] is written as where we replaced tdB/dt by its discrete approximation DB(t). This is the expression used in practice in all quantitative estimations presented. Parameter estimation with quantified uncertainty can be achieved using a Bayesian approach in the context of probabilistic epidemiological models. Bayes' theorem expresses the full probability distribution for model parameters, such as the effective reproduction number, R, in terms of the probabilistic epidemiological model [12] , given the time series for new cases. Specifically, the probability distribution of R, compatible with the observed temporal data stream is given by P[R] is the prior, which captures given knowledge of the distribution of R. The distribution P[DT(t+t)rDT(t)] is independent of R, and corresponds to a trivial normalization. From successive applications of Bayes' theorem, a sequential estimation scheme, that uses streaming epidemiological observations performed in real time, can be constructed using the posterior distribution for R, at time t as the prior in the next estimation step at time t+t, leading to an update scheme via iteration of Eq. [13] . The resulting probability distribution for R includes information on all observations up to time t, and contrasts with the ''instantaneous'' R t , used above, which only considers the data at times t and t+t. Thus, R is a robust estimator of the effective reproduction number assumed to be constant for the whole epidemic up to time t. Any changes in R over time result from the assimilation of each new data point, leading to an updated estimate of R. This in turn allows the use of our estimation procedure as an anomaly detection tool (see below). Ideally, field case tracing should provide a measure of the likelihood that a new case resulted from contact with the (animal) reservoir, DB in Eq. [12] , or was due instead to human contagion. Another possibility is to explicitly model the introductions from the reservoir, K(t) in Eq. [7] . Although some empirical studies start to address this possibility [27, [36] [37] [38] , it is still difficult to calibrate such models and uncertainties remain large. Thus, for the calculations in this study, we choose to use a minimal statistical approach. Formally, assuming statistical independence between different cases, we model each new introduction as a Bernoulli trial with probability h. h is defined as the average probability that a case is attributed to human to human contagion, and 1-h that it is the result of an infection from the reservoir. Note that for emerging infectious diseases this probabilistic model is more appropriate and generalizes (see Supplementary Material S1, Figure S2 ) the more typical modelling of introductions as homogeneous Poisson processes [33] . We can provide an upper bound for h by considering observed clusters of cases. If we take all confirmed cases in clusters, except the index case, to be due to human contagion, then an estimate for h is given by the proportion of such cases divided by the total number observed over the same period. This gives an upper bound on h, because it is unlikely that all cluster cases arise from human infection, rather some could have a common reservoir source. Two epidemiological studies of H5N1 influenza, one from January 2004 to July 2005 [39] and another from July 2005 to June 2006 [40] , found that 26 of 109 cases and 15 of 54 cases, respectively, occurred in family clusters Attributing those cases to human infection gives h = 0.24-0.28. This estimate of h is consistent with an independent statistical analysis of a case cluster in Indonesia, which found that the secondary attack rate for household transmission of H5N1 influenza was 0.29 [28] . In the remaining of this paper we treat h as a constant parameter common to all reported cases. The sensitivity of R estimates is then assessed as a function of h (Table 1 ). In addition to tests of the estimation procedure on epidemic data [34] , we also verified the precision of our methodology on an extensive number of simulated case time series, based on a standard SIR model, with introductions from a reservoir and different R 0 . We used the observed time series of new cases of human H5N1 influenza DT(t) to compute the probability distribution for R using programs implemented both in Matlab and Fortran. We used an unbiased uniform distribution for R between 0 and 3 as the initial prior. For each subsequent weekly iteration, we computed the full posterior distribution from [13] using the posterior at the previous week as the new prior. The product of the two probabilities on the right-hand side of [13] was evaluated as a non-parametric function defined in terms of 1000 discrete bins in R between 0 and 3, as shown in Figure S3 in the Supplementary Material S1. Parameters choices used in the calculations in the main text are: t = 1 week, c = 1 week 21 , h variable as in Table 1 . We also explored other parameter choices reported in the literature for seasonal influenza and corresponding results are given in Table S1 of Supplementary Material S1. Here we show how the method performs at estimating R from single realization time series, produced by simulation with a known value of R 0 . In all instances the time series for human H5N1 cases in Vietnam (Figure 1a ) was used as introductions into the human population. For a choice of R 0 .1 in the simulation, any introduction readily develops into an epidemic. For R 0 ,1 each introduction leads to small outbreaks that eventually become extinguished. The effective reproduction number, R, calculated by our method changes over time, because of decreases in the fraction of susceptibles, S(t)/N(t), and the availability of more information, as more cases are observed. Thus, we use the values obtained for R at early times, when S(t)/N(t) approximates its initial value, to estimate R 0 of the simulation by assuming that max(R) = R 0 . As shown in Figure 3 , in all circumstances, the method gives an excellent estimation of R 0 as outbreaks unfold, usually making accurate predictions when supplied with a mere two or three observation points. Uncertainty, measured by the width of the 95% credible interval, is reduced by larger case numbers, so that it typically remains higher the smaller the R 0 . In all instances uncertainty is reduced as more cases are reported over time. We next applied our method to the time series of cases of H5N1 influenza in humans. We produce estimates and credible bounds for R, under different scenarios for the expected fraction of new observed human cases that is attributable to human contagion h. Summary results for Vietnam (and Indonesia) are given in Table 1 . Even in the worst case scenario, where all observed cases are attributed to human transmission (h = 1), the most likely (as of June 2006) R is 0.53 (0.56), with an upper 95% bound of R,0.77 (0.89). For the estimated h = 0.29 (see Methods), the most likely R for both Vietnam and Indonesia is 0, although the estimated upper 95% bound in Vietnam gives the bound R,0.42. For Indonesia, the corresponding estimate gives an R entirely consistent with zero at the 95% credible level. For less than 20% of the cases attributable to human transmission, R is entirely consistent with zero, even when accounting for the uncertainty in the duration of the infectious period (c 21 ). Reported information does not allow at present a precise determination of c 21 for H5N1 influenza in humans, so that different scenarios are possible, which we explore in detail in Supplementary Material S1 ( Figure S4 , Table S1 ). Data permitting, a hierarchical Bayesian estimation method for the distribution of c can also be envisaged [32] . Figure 4 shows the evolution of R and of its corresponding 95% credible interval. The computation of successive probability distributions for R gives a basis for assessing the evolution of transmissibility over time, including the approach to the epidemic threshold RR1. At present we conclude that, even in the unrealistic worst case scenario, where cases are aggregated at the national level and all cases are attributed to human transmission, R remains below unity. The emergence of a new epidemic in humans often requires shifts in pathogen biology and/or changes in the human population structure. The methodology developed here can signal these events as anomalies in the expected number of new cases. Assuming no change of epidemiological conditions, knowledge of the distribution of R, accumulated until time t, provides expectations for future case numbers DT(t+t) , with quantified credible intervals, via where P[R] is taken as the posterior in [13] at time t, and P[DT(t+t)rDT(t)|R] is the statistical epidemic model. Failure to predict future observed cases at time t+t, can then be formulated as a p-value significance test at any chosen level of credibility. A statistical anomaly, i.e., future cases falling outside the credible interval defined by previous observations, may signal changes in epidemiological parameters, specifically in transmissibility (either by pathogen evolution or host population changes) as measured by R. We provide an example in Figure 5 , for simulated data with R 0 = 0.8 changing to R 0 = 1.3, where we show the predicted 95% credible interval for new cases vs. the number of cases actually observed (see also Figure S7 in Supplementary Material S1). Emerging and re-emerging infectious diseases pose some of the greatest health risks to human populations worldwide. Increasingly they are a feature of our time, stoked by changes in human demographics, mobility, land use and climate, and compounded by poor standards of public health in parts of the world [26, 41] . Importantly, new surveillance and intervention strategies are now becoming possible, guided by quantitative interpretation of epidemiological data, potentially strengthening the hand of primary prevention efforts. The modelling and prediction approaches developed here (see also Chowell et al. [34] for a comparison to other methods) provide tools for real time estimation of epidemiological parameters that are appropriate for emerging infectious diseases. The method is intentionally simple, relying on standard epidemiological population models, in order to be commensurate with the paucity of epidemiological data typically available for emerging infectious diseases. These features are illustrated by the application of the method to H5N1 influenza infection time series in humans. Clearly, the SIR class of models, even when cast in probabilistic terms, relies on several general assumptions, which are simplistic in specific situations. First, these models do not account for contact heterogeneities, resulting from spatial effects, age, and/or the structure of social networks. These effects can be partially addressed by structuring the population compartments in terms of spatial and risk classes [42] . For example, different regions in an affected country may be taken as separate compartments, provided that there are data pertaining to each region. Indeed the method would then allow estimation of correlations between R at different spatial locations. Second, in its present form, the model does not include independent estimation of infectious or incubation periods. It is straightforward to include an incubation period [43] and, given data on the duration of these periods on a case by case basis, these issues can be addressed by including additional Bayesian estimation steps (see [32] ). Notwithstanding these limiting features, the SIR structure allows reliable real time parameter estimation with quantified uncertainty at very low computational overhead, as verified extensively via simulations at varying known input R, and applications to past pandemic outbreaks (Supplementary Material S1 - Fig. S5 and S6 -and Ref [34] ). One feature of the bounds on R derived here is their dependency on the fraction of cases attributed to human transmission, h. Although h is judged to be small from present surveillance [23, 27, 28] , it remains hard to quantify with certainty. Given the paucity of data, we chose in practice to assume independence and use a binomial probability to attribute cases to human transmission vs. infection from the reservoir. However, other procedures to determine h are possible. If enough data were available, an explicit dynamical model of the (animal) reservoir could be built, or an empirical function correlating introductions through time could be used. Indeed, in the optimal scenario, the actual cases of introduction from the reservoir would be known from field work, and the method proposed here could incorporate that information directly. We note that in the most relevant case, when R becomes larger than 1, the effect of h,1 quickly vanishes, as cases multiply exponentially. For R,1, even a choice of h = 1 will lead to estimates of R,1, but different values of h may lead to credible intervals that include the critical threshold. In general, we believe that a suspicion of a possible R<1 should be followed up with careful field investigations. We presented a general methodology capable of interpreting quantitatively emerging disease surveillance data in real time with quantified uncertainty that complements other methods proposed recently [21, 29, 30, 32] . Although we illustrated the method with data for H5N1 influenza, these inference strategies are general and can be applied to time series from other communicable diseases. We verified that the model developed here also applies to standard epidemics, yielding agreement with previous estimates of R for the 1918 influenza pandemic in US cities [14, 44] (Figures S5, S6 in Supporting Information) and with other estimation methods [34] . We have also shown how to construct p-value statistical significance tests suitable for automatically monitoring changes in transmissibility of (emerging) communicable diseases [43] . While still in their infancy, we believe that the current emerging trend in mathematical epidemiology towards real time predictive methods will enable a shift towards more quantitative surveillance and primary prevention, resulting in more consistent and extensive monitoring of emerging infectious diseases and improved designs for health interventions and logistic allocations as epidemics unfold. Material S1 In Supplementary Material S1, we present further details of the method, including extensions and additional examples. Found at: doi:10.1371/journal.pone.0002185.s001 (0.28 MB PDF) We thank Miles Davenport, Mac Hyman, Alan Perelson, Timothy Reluga, our PLoS One Editor and anonymous referees for comments that substantially improved the manuscript. Indonesia, under the pessimistic assumption that 29% of reported cases are due to human-to-human transmission (see Table 1 ); and (c) for seasonal H3N2 human influenza isolates in the USA during the 2004-2005 season. (Note that isolates represent only a small fraction of total cases, and may contain reporting biases.) The estimate of the effective reproduction number for an epidemic outbreak asymptotes to unity at late times because initial growth and long-term decay in new case numbers (due to depletion of susceptibles) average out over the history of the outbreak. doi:10.1371/journal.pone.0002185.g004 Between weeks 74 and 75, the reproduction number is shifted R 0 = 0.8R1.3 to create an epidemic. Although we continued to iterate the R distributions via the Bayesian procedure described in the text, note that the shift in R upwards leads to many statistical anomalies (indicated by black arrows). The anomaly is detected immediately, on weeks 75 and 76. Anomalies here are defined as observed numbers of new cases that fall outside the expected 95% credible interval. These anomalies indicate a violation of the hypothesis that R is unchanged, and could be used to trigger alerts in surveillance. doi:10.1371/journal.pone.0002185.g005

DetectiV: visualization, normalization and significance testing for pathogen-detection microarray data

DNA microarrays offer the possibility of testing for the presence of thousands of micro-organisms in a single experiment. However, there is a lack of reliable bioinformatics tools for the analysis of such data. We have developed DetectiV, a package for the statistical software R. DetectiV offers powerful yet simple visualization, normalization and significance testing tools. We show that DetectiV performs better than previously published software on a large, publicly available dataset.

One of the key applications of metagenomics is the identification and quantification of species within a clinical or environmental sample. Microarrays are particularly attractive for the recognition of pathogens in clinical material since current diagnostic assays are typically restricted to the detection of single targets by real-time PCR or immunological assays. Furthermore, molecular characterization and phylogenetic analysis of these signatures can require downstream sequencing of genomic regions. Many microarrays have already been produced with the aim of characterizing the spectrum of microorganisms present in a sample, including detection of known viruses [1] [2] [3] [4] [5] , assessment of bioterrorism [6, 7] and monitoring food quality [8] . However, the use of DNA microarrays for routine applications produces many challenges for bioinformatics. Firstly, probe selection is a difficult and time consuming process. There are a huge number of diverse species in nature, of which we have sequence information for only a tiny fraction. This makes it difficult to find oligonucleotides, either alone or in combination, that uniquely identify species of interest. Oligos may have homology to multiple species, which results in a complex and noisy hybridization pattern. Secondly, each nucleic acid sample tested will typically contain a mixture of DNA and RNA from the organism of interest, the host and from a variety of contaminants, which may all contribute to the resulting microarray profile. Furthermore, this may be complicated by the presence of multiple, possibly related, pathogen species, making it difficult to separate patterns due to cross-hybridization from a true positive result. Urisman et al. [9] have previously reported E-Predict, a computational strategy for species identification based on observed microarray hybridization patterns. E-Predict uses a matrix of theoretical hybridization energy profiles calculated by BLAST-ing completely sequenced viral genomes against the oligos on their array, and calculating a free energy of hybridization. Observed hybridization profiles are then compared to the theoretical profiles using a similarity metric, and a p value calculated using a set of experimentally obtained null probability distributions. E-Predict has been shown to produce useful results in a number of situations. However, at present, E-Predict does not contain any tools for visualization, and requires extensive customization and calculation before it is applicable to new arrays. Also, E-Predict is only available as a CGI script for Unix/Linux platforms. We present DetectiV, a package for R [10] containing functions for visualization, normalization and significance testing of pathogen detection microarray data. R is a freely available statistical software package available for Windows, Unix/ Linux and MacOS, meaning DetectiV is a platform independent solution. DetectiV uses simple and established methods for visualization, normalization and significance testing. When applied to a publicly available microarray dataset, DetectiV produces the correct result in 55 out of 56 arrays tested, an improvement on previously published methods. When applied to a second dataset, DetectiV produces the correct result in 12 out of 12 arrays. DetectiV is implemented as a package for R, a powerful, opensource software package for statistical programming [10] . Many packages for R already exist for the analysis of biological datasets, including microarray data, and the bioconductor project [11] is just one example of a group of such packages. As it is implemented in R, DetectiV easily integrates with many of the packages available for microarray analysis, including limma [12] , marray [11] and affy [13] . DetectiV is written in the native R language and uses standard functions within R. As R is available on Microsoft Windows, Unix (including linux) and MacOS, DetectiV represents a platform independent solution for the analysis of pathogendetection microarray data. The flow of information through DetectiV is shown in Figure 1 . The basic dataset required is a matrix of data, with rows representing probes on the array, and columns representing measurements from individual microarrays. This dataset is easily produced from data structures created by limma [12] , which includes functions for reading in many common microarray scanner output formats, and affy [13] , which provides functions for reading in affymetrix data. Commonly, researchers will have an additional file of information giving details about each probe. In the case of pathogen detection arrays, this file will most often contain the type, species, genus and other classification data for the pathogen to which each probe is designed. It should be noted that there may be more than one entry in this file for each probe; for example, if a given probe is thought to hybridize to multiple pathogens. In text format, these may be read in using the native read.table command, or in excel format using the RODBC library. Once these two datasets are in R, DetectiV prepares them for analysis using the prepare.data function. This function joins the array data to the probe information data based on a unique ID. The researcher may choose to subtract local background if appropriate. The default at this stage is to average over replicate probes, again based on a unique ID. This will result in a single value for each unique probe for each array. The data will have one or more columns of extra information from the annotation file, and these columns will be used to group the data for further analysis. Researchers will wish to visualize their data in order to compare the hybridization signals for the probes recognizing the different pathogen signatures. DetectiV provides a function called show.barplot for this. The output from prepare.data is passed to the function, along with the name of the column containing the variable by which the data will be grouped, referred to here as group. An example in pathogen detection data may be species, genus, family, and so on. The data are sorted into unique groups as defined by the unique values of group. A barplot is drawn, with one bar per unique probe. Probes from the same group are drawn together. Each group is represented by a unique background color, enabling the user to easily visualize the different groups. An example output is shown in Figure 2 . This sample comes from Urisman et al [9] and represents data from a virus detection microarray hybridized with amplified RNA from nasal lavage, positive for respiratory syncytial virus by direct fluorescent antibody (DFA) test. The group chosen here is virus family. It is quite clear from this image that there is a virus from the family Paramyxoviridae present in the sample, demonstrated by the high bars associated with that family. These images are often very large, and so DetectiV offers the ability to subset the data before plotting by using the get.subset function. Figure 3 shows a similar barplot using a subset of the data: only those oligos representing species that belong to the Paramyxoviridae family. It is clear from this image that those oligos representing different groups/species of respiratory syncytial virus have the highest intensity, as we would expect, although there is cross-hybridization with oligos for human metapneumovirus (another paramyxovirus in the same sub-family: Pneumovirinae). DetectiV may also carry out normalization and significance testing. For this, there is the function normalise. Here, the aim of normalization is to represent the data in relation to a negative control. The idea is that if the values for each probe are divided by the negative control and then the log2 taken, then the data should be normally distributed, and each group should have a mean of zero (providing a pathogen is not present). Traditional statistical tests can then be used to test if any group of probes is significantly different from zero. DetectiV offers three methods of normalization, each using a different 'type' of negative control, and these are summarized in Table 1 . Flow of information, and steps taken, when analyzing pathogen detection microarray data using DetectiV Figure 1 Flow of information, and steps taken, when analyzing pathogen detection microarray data using DetectiV. The median method calculates the global median value for each array. It should be noted that this method assumes that most probes will not hybridize to anything. If this assumption is false then this method should not be used. However, if the assumption holds, then the median is a good representation of that value we would expect to see from probes that have not hybridized to anything. The control method relies on specific negative controls having been spotted on the array. The researcher may then choose one of these controls, and the mean value is calculated for that control for each of the arrays. The mean control value for each array is then used as a divisor for each probe on their respective arrays. Finally, the array method utilizes an entire control array or channel. In this instance, an entire array is chosen to be the negative control, and all probe values are divided by their respective elements from the control array. An obvious example for a control array may be RNA from a known uninfected animal. The control array therefore has a value for each specific probe representing that value we would expect to see if that specific probe has not hybridized to anything. In all instances, after taking the log2, groups of probes that have not hybridized to anything should be normally distributed and have mean zero. We can therefore split the probes into groups and perform a t-test for each one. DetectiV does this using the do.t.test function. The normalized (or raw) data are split into groups as defined by the unique values of a user defined annotation column. Providing each group has more than two probes, a t-test is performed to test the difference of the observations from zero. The average value is also calculated. The output is a table, sorted by p value. The data used were downloaded from the Gene Expression Omnibus (GEO) [14] , accession number GSE2228. The array platform for this data is GEO accession GPL1834, and includes over 11,000 oligos representing over 1,000 viral and bacterial species [4] . The dataset itself consists of 56 arrays including 15 independent HeLa RNA hybridizations, 10 independent nasal lavage samples positive for respiratory syncytial virus, 7 independent nasal lavage samples positive for influenza A virus, a serum sample positive for hepatitis B virus, a nasal lavage sample positive for both influenza A virus and respiratory syncytial virus, and culture samples of 11 distinct human rhinovirus serotypes. Both DetectiV and E-Predict [9] have been used to analyze the data. For DetectiV, the data were not corrected for local background. Missing, negative and zero values were set to a nominal value of 0.5. Intensities were averaged across replicate probes. Median normalization was then carried out, followed by a t-test grouping the data by virus species. Probes representing actin, GAPDH and Line_Sine were filtered from Table 1 DetectiV normalization methods Median Where is the value for probe i on array j and is the median value for all probes on array j Where is the value for probe i on array j and is the mean value for control oligo c on array j Array Where is the value for probe i on array j and is the value for probe i on control array/channel c Explanation of the three normalized statistics offered by DetectiV. the results. Results were first filtered such that groups had a normalized log2 ratio greater than or equal to 1 (a ratio of two to the control) and then sorted by p value. This method will be referred to as DetectiV. For E-Predict, default values for all parameters were used, and are shown in Table 2 . Data points were corrected for local background, as per the examples in Urisman et al. [9] . E-Predict filters out 266 oligos by default, and this setting was kept. In all cases, E-Predict carried out two iterations, although only results from the first iteration are shown here. The best performing method of interpreting the results was to take those species with a p value ≤ 0.05 and sort by distance (termed E-Predict.dist). Note that this is the method cited in [9] , example 3, used to demonstrate E-Predict's ability to detect SARS. Pathogen detection arrays have also been implicated in the discovery of SARS. Urisman et al. [9] reported that although their original platform did not contain oligos designed to SARS, once the SARS genome had been published, it was possible to recalculate the energy matrix for E-Predict and find that the energy profile for SARS was the top hit (after taking those viruses with low p values and sorting by distance). We have applied DetectiV to the same dataset (GEO accession GSE546). To include oligos for SARS, we searched a database of oligo sequences on the array with sequence NC_004718 from RefSeq using NCBI blast. There were 61 oligos on the array that hit the SARS genome with greater than 80% identity across an alignment of 20 bp or more. In the analysis, these oligos were assigned as representative of two viruses: their original virus and SARS. The data were median normalized and a t-test carried out using DetectiV. Finally, having established that DetectiV compares favorably with previously published software, we have validated the DetectiV software by applying it to a second dataset. The data used were downloaded from the GEO [14] , accession number GSE8746. The array platform for this data is GEO accession GPL5725, and consists of 5,824 oligos representing over 100 viral families, species and subtypes. The dataset itself consists of 12 arrays, 4 hybridized with RNA from cell cultured footand-mouth disease virus (FMDV) type O, 3 hybridized with RNA from FMDV type A, 1 hybridized with RNA from a sheep infected with FMDV type O, and 4 hybridized with cell-cultured avian infectious bronchitis virus (IBV). Analysis using DetectiV was carried out as described above. We present here results from two methods of analysis, termed DetectiV and E-Predict.dist, as described above. There are 56 arrays in the dataset, the expected results of which are known. Each array was hybridized with RNA containing a single virus, except GSM40845, which was infected with both influenza A and respiratory syncytial virus. We assigned a correct result for each method if the top hit from the analysis was the same as the known infectious agent or, if that agent was not represented on the array, the top hit was a very closely related virus. In the case of GSM40845, we report a correct result if both viruses were at the top of the reported hits, to the exclusion of other virus species (but not closely related strains). Additional data file 1 gives the top hit for both analysis methods in all 56 arrays. As can be seen, DetectiV generated a correct result in 55 out of the 56 arrays. In comparison, the E-Predict.dist method gave a correct result in 53 out of the 56 arrays. These results are discussed in greater detail below. Full results for each of the arrays can be found on the DetectiV website [15] . Within the 55 correct results, there are three classes that require slightly different interpretation, examples of which are GSM40806, GSM40810 and GSM40817. Results for these arrays are given in Table 3 . Array GSM40806 was hybridized with amplified HeLa RNA, and the top hit from DetectiV is human papillomavirus type 18, as expected. This virus has both the smallest p value and largest mean normalized log ratio. There is also clear GSM40810 was hybridized with RNA containing human rhinovirus 28. There are 24 distinct groups of human rhinoviruses represented on the array, including a group of oligos for all members ('human rhinovirus sp.), one each for human rhinovirus A and B, and several groups for distinct serotypes. Human rhinovirus 28 is not one of those serotypes specifically targeted by the array; however, as a serotype of the human rhinovirus A species, we would expect the groups for human rhinovirus sp. and human rhinovirus A to be prevalent amongst the results. As can be seen from Table 3 , the top hit from DetectiV is human rhinovirus sp., closely followed by human rhinovirus A, the expected result. The reason we have highlighted this array, however, is that the result for Enterobacteria phage M13 shows a higher mean normalized intensity than any of the rhinovirus groups. This is representative of a class of result from DetectiV whereby a virus group has a higher mean normalized log ratio, but a larger p value, than the top hit. Here, as in GSM40806, we see orders of magnitude between the p value for the top hit and that for Enterobacteria phage M13, which identifies human rhinovirus as being the infectious agent, but in this case we cannot rely on the mean normalized intensity. In this particular instance, Enterobacteria phage M13 is represented by 10 oligos, all of which have intensities far greater than the global median, but which vary considerably between 982 and 18,864. These high values may be due to hybridization with a cloning vector. Finally, array GSM40817 was hybridized with respiratory syncytial virus. The results are again shown in Table 3 , but for this array only, they have not been filtered on mean normalized intensity. Human herpesvirus 5 has by far the smallest p value of any of the virus groups; however, it also has a very small mean normalized log ratio. The correct hit, respiratory syncytial virus, has the second smallest p value, but has a much larger mean normalized log ratio. This represents the final class of result seen by DetectiV, where the correct virus group does not have the smallest p value, but does have a much larger mean normalized log ratio than those groups that have smaller p values. The small p value of respiratory syncytial virus combined with the large mean normalized log ratio identifies respiratory syncytial virus as the only infectious agent. In this instance, human herpesvirus 5 is represented by 241 oligos, 167 of which are greater than the global median, but all of which have intensities less than 1,000. This could be due to the oligos for human herpesvirus 5 having distant homology with the infectious agent or host cell. These three types of result are typical of DetectiV, and explain why both the p value and the mean normalized log ratio must be taken into account when interpreting the results. Thus, if the results from DetectiV are filtered such that only viruses whose mean normalized log ratio is ≥ 1, and then sorted by p value, the three scenarios described here are accounted for, and we obtain the correct result in 55 out of the 56 arrays. The single incorrect result for DetectiV comes from GSM40816, which reports human herpesvirus 7 as the top hit, whereas the infectious agent was in fact respiratory syncytial virus. The top five hits for this array using the DetectiV method are shown in Table 4 . As can be seen, bovine respiratory syncytial virus and respiratory syncytial virus are second and third, respectively. Both respiratory syncytial virus and bovine respiratory syncytial virus have higher mean values than human herpesvirus 7, although the latter has a smaller p value and a mean value that is above the cut-off of 1. Had the results been filtered for p value ≤ 0.5 and then ordered by average value, then the top hit would have been respiratory syncytial virus; similarly, if a cut-off of 2 had been applied instead of 1, a correct result would have been reported. However, across the entire dataset these methods of interpreting the results perform worse than the DetectiV method described above. It is worth noting here that for this array, E-Predict gives the correct top hit. The results from E-Predict follow similar patterns to those of DetectiV. In most cases it is obvious which virus is the infectious agent, either by examining the p value, the similarity or both together. Full results can be seen on the DetectiV website [15] . However, there are certain results reported by E-Predict where it is impossible to obtain the correct result no matter which combination of p value and similarity is used. These arrays are arrays are GSM40809, GSM40821 and GSM40847, and the top five results for these arrays can be seen in Table 5 . GSM40809 was hybridized with RNA containing human rhinovirus 26. Again, this is a serotype not specifically targeted by the array; however, as a serotype of human rhinovirus B we Array GSM40821 was infected with hepatitis B virus but E-Predict reports orangutan hednavirus as having both a smaller p value and a higher similarity. This is not that surprising given that hepatitis B and orangutan hepadnavirus are closely related; however, the fact remains that with no a priori knowledge, the only logical conclusion from this result would be that the infectious agent was orangutan hepadnavirus. Again, DetectiV calls this array correctly. Finally, array GSM40847 was hybridized with RNA containing human rhinovirus 87. Again, this is a serotype not specifically targeted by the array, and is not present in the NCBI taxonomy database [16] at the time of writing. We can therefore expect the 'human rhinovirus sp.' group to be high amongst the results (in fact, it is the top result for DetectiV). E-Predict reports human enterovirus B as having the smallest p value and human echovirus 1 as having the largest similarity. In fact, E-Predict does not report any rhinovirus oligos in the first iteration at all, and it is only in the second iteration that the group human rhinovirus A is reported as significant. In the three cases outlined above, there is no clear way of distinguishing the incorrect virus from the correct one. There is also no consistent method of sorting or filtering the results that would give the correct results. In these three cases, E-Predict is unable to distinguish closely related virus species and serotypes. We have reported here the best performing method of interpreting E-Predict results, whereby virus groups with a p value ≤ 0.05 are sorted by distance. This results in a success rate of 53 out of 56 arrays. The top five hits from the analysis of the SARS dataset can be found in Table 6 . As can be seen, the top hit is SARS, with the lowest p value and the highest mean normalized log ratio. SARS is distinct from the other viruses, having a p value three orders of magnitude lower than the second top hit. Full results can be found on the DetectiV website [17] . The top hit from DetectiV for each of the 12 arrays from GSE8746 can be found in Table 7 . As can be seen, DetectiV clearly identifies the infectious agent in all 12 cases. DetectiV works for both the cell-cultured samples and the infected sheep, and shows the ability of the array to distinguish between different subtypes of FMDV. Developing a quick and reliable test for the presence/absence of thousands of bacterial and viral species in a single experiment is an attractive proposition, and a function that DNA microarrays are ideally suited to. Microarrays are extremely high-throughput and relatively cheap. In the case of pathogen detection, the aim must be to quickly and clearly identify those pathogens present in a sample with high confidence, keeping false positives and false negatives to a minimum. However, the data from such microarrays pose many problems. Firstly, oligos may not be unique to the species they are designed to. For certain species it is impossible to find a large number of oligos that are unique only to that virus that meet the criteria for oligo selection. This is particularly problematic for closely related species and strains. In such cases, the 'best' oligos are added to the array, in the knowledge that multiple viruses may hybridize to them. This leads to noisy signals across multiple virus families, species and serotypes. Secondly, infected biological samples may contain many different virus species and strains, making interpretation difficult. Thirdly, it is known that certain oligos simply do not work, even when the array is hybridized with the species that those oligos were designed to. Without testing the array with each virus, we are incapable at present of predicting which oligos will work and which will not. With thousands of species per array, many of which cannot be cultured in vitro, it is unfeasible to challenge arrays with every species. Finally, we of course do not know, nor can we ever know, the complete genome sequence of every virus we may encounter. Therefore, though we think we have oligos unique to a species or strain, that is only ever in the context of our knowledge at the time of design, and they may not in fact be unique. Despite these problems, many species detection arrays have been developed [1] [2] [3] [4] [5] . However, reliable methods of data analysis have been rare. Initial methods included visual inspection of the array [4] and clustering [18] , both of which are subjective and time-consuming. To combat this, Urisman et al. [9] have proposed a more robust method, E-Predict. E-Predict utilizes a pre-calculated energy matrix for each oligo on the array and uses a variety of normalization and similarity metrics to calculate a p value and similarity for each virus. The advantages of E-Predict are that it is quantitative, produces good results and is extensible, through the extension of the energy matrix. The disadvantages of the software are a lack of visualization tools, the need to customize parameters for different array platforms and hybridization conditions, and the availability of the software only as a CGI script on the Unix/Linux platform. We have developed DetectiV, a package for R containing visualization, normalization and significance testing functions for pathogen detection data. DetectiV uses simple and well established visualization and statistical techniques to analyze data from pathogen detection microarrays. DetectiV offers a powerful visualization option in the form of a barplot, enabling researchers to quickly and easily identify possible infectious agents. Data can then normalized to a negative control (be that a specific probe, array or the global median), transformed by taking the log2 and then subjected to a t-test for each species on the array. Oligos are allowed to represent any number of viruses, and thus any analysis is easily extensible by simply updating the list of which oligos represent which species. DetectiV requires minimal set up and configuration, requiring only an additional file detailing which species each oligo represents. In the majority of cases, these files will already exist. It is then possible to apply DetectiV 'out of the box' to any array data that is readable by R or bioconductor. DetectiV requires no training, configuration or customization specific to each array. DetectiV is available as a package for R on both Windows and Linux/Unix, and as such may be considered platform-independent. In this study, DetectiV produced the correct result in 55 out of 56 arrays, by filtering for viruses with a mean normalized log ratio greater than 1 and then sorting by p value. We make the distinction here between biological and statistical significance. A statistically significant result may be obtained by a group of oligos that display intensities only marginally larger than the negative control (in this case the global median intensity). This is demonstrated by human herpesvirus 5 on array GSM40820 (Table 3) . However, we know that from a biological perspective, we would expect to see intensities far higher than the negative control, and that intensities only marginally higher result from low homology between the probe and the sample. We can therefore use the statistical significance (p value) in combination with our idea of biological significance (the mean normalized log ratio) to successfully call the correct result in over 98% of the arrays. In the majority of cases there is a clear difference in the p value, the mean normalized log ratio, or both, between the correct hit and subsequent hits, allowing for both automatic and manual detection of true and false positives. However, this does require careful interpretation. Both DetectiV and E-Predict predict multiple, significant matches on all of the arrays. When using DetectiV, it is only when looking for major changes between the top hit and subsequent hits, in terms of p value or mean log ratio, that it is possible to separate the true positives from the false positives. In many cases, using automatic rules will result in the correct result; however, there will inevitably be borderline cases where human inspection of the results is required. This is all the more important when considering the possible economic impacts of a false positive for certain species. At present, the safest way to employ such arrays, and their analysis methods, may be simply as a first step towards identifying infectious agents, informing researchers about which viruses they should test for using more conventional methods. The results from the application of DetectiV to the SARS dataset are encouraging. Here, oligos designed to SARS were not present on the array. However, using a simple NCBI blast search, it was possible to extend the range of viruses covered by the array to include SARS -61 existing oligos showing significant homology to the SARS genome. On application of DetectiV to the updated data, SARS was the top hit. Not only does this offer the promise of being able to extend the coverage of the array without adding further oligos, it also suggests that it is possible to detect viruses without having any unique oligos. This may inform the oligo selection process -it may be equally desirable to have multiple, non-unique oligos to represent a species as it is to have a few that are unique. The results from the application of DetectiV to a second dataset are also encouraging, with the correct result being the top hit in all 12 cases. Of particular interest is the ability of the array, and DetectiV, to distinguish not only between separate viral species, but also between different subtypes of FMDV. It should be noted that in order to apply DetectiV to a second dataset from a completely different array to the first dataset, the user only has to change the GEO accession number and the number of arrays within that dataset. This compares favorably with E-Predict, which would require a separate training dataset from the second array, the calculation of a large and complex sequence similarity matrix and the optimization of several parameters. There are a number of ways in which DetectiV may be developed. In terms of visualization, better browsing capabilities of the barplots would be desirable, perhaps using a web-interface. In terms of the analysis, we may borrow ideas from gene expression arrays. For example, limma uses an empirical Bayes method to shrink each gene's standard error towards a common value, and has been shown to perform better than standard statistical methods [12] . It may be that we can apply a similar method here to shrink the standard error for each virus species towards a common value, thus increasing sensitivity. It may also be possible to apply multiple-testing procedures to the resulting p values. The Bonferroni correction may be appropriate, in which the p values are multiplied by the number of comparisons, or a more conservative approach may be needed, such as that suggested by Benjamini and Hochberg [19] , in order to control the false discovery rate. In conclusion, DetectiV is a highly accurate tool for the analysis of pathogen detection microarray data, offering simple but powerful visualization, normalization and significance testing functions. DetectiV performs better than previously published software on a publicly available microarray dataset. DetectiV is available as a package for R, a platform-independent statistical software package, and requires little configuration or customization. It is released under the GNU General Public License and may be downloaded from the DetectiV website [20] .

Expression of Foot-and-Mouth Disease Virus Capsid Proteins in Silkworm-Baculovirus Expression System and Its Utilization as a Subunit Vaccine

BACKGROUND: Foot-and-mouth disease (FMD) is a highly contagious disease of livestock that causes severe economic loss in susceptible cloven-hoofed animals. Although the traditional inactivated vaccine has been proved effective, it may lead to a new outbreak of FMD because of either incomplete inactivation of FMDV or the escape of live virus from vaccine production workshop. Thus, it is urgent to develop a novel FMDV vaccine that is safer, more effective and more economical than traditional vaccines. METHODOLOGY AND PRINCIPAL FINDINGS: A recombinant silkworm baculovirus Bm-P12A3C which contained the intact P1-2A and 3C protease coding regions of FMDV Asia 1/HNK/CHA/05 was developed. Indirect immunofluorescence test and sandwich-ELISA were used to verify that Bm-P12A3C could express the target cassette. Expression products from silkworm were diluted to 30 folds and used as antigen to immunize cattle. Specific antibody was induced in all vaccinated animals. After challenge with virulent homologous virus, four of the five animals were completely protected, and clinical symptoms were alleviated and delayed in the remaining one. Furthermore, a PD(50) (50% bovine protective dose) test was performed to assess the bovine potency of the subunit vaccine. The result showed the subunit vaccine could achieve 6.34 PD(50) per dose. CONCLUSION: The results suggest that this strategy might be used to develop the new subunit FMDV vaccine.

Foot-and-mouth disease (FMD) is an economically important disease of domestic and wild cloven-hoof animals including cattle, swine, goat, sheep and buffalo. It can result in great reduction of productivity in adult animals and death in young animals. FMD is endemic in parts of Asia, Africa, the Middle East and South America (with sporadic outbreaks in other ''free areas''). In countries affected by the disease, livestock trade and animal products have been impacted. Even in developed countries and areas, outbreak of FMD would greatly affect the economy. In 2001, the outbreak of FMD in England brought a loss of 8 billion dollars [1] , and the consequent occurrence in Holland killed 0.2 million animals [2] . In addition, the 1997 outbreak in Taiwan brought a loss of about 3.6 billion dollars to its exports [3] . Therefore, controlling bath endemic and epidemic FMD has become a global concern in livestock raising. At present, vaccination is a major means of FMD control in most endemic areas. The present method of FMDV vaccine production is this: The virus is propagated in BHK-21 cell line, concentrated, and chemically inactivated. Although the inactivated vaccine has been shown to be effective, it may lead to new outbreaks of FMD because of either the incomplete inactivation of FMDV in large-scale production or the escape of the live virus from vaccine production workshops [4] . Therefore several expression systems such as E.coli [5] , transgenic plant [6] , yeast [7] , adenovirus vector [8] [9] [10] [11] [12] , vaccinia virus vector [13] , and DNA vaccine [14] , have been used for expression of FMDV antigen to prepare subunit vaccines. But such methods have problems such as poor immunogenic capability or low efficiency. Adenovirus based vaccine known for its best protective effects can protect 5 of 5 vaccinated cattle, but this vaccine is still unacceptable because of safety problem and preservation difficulty. The baculovirus expression system, a valuable expression system to produce virus-like particles, has successfully produced many kinds of empty viral capsids [15, 16] , such as rabbit hemorrhagic disease virus, Norwalk-like viruses,SARS and so on [17, 18, 19] . Compared to the baculovirus expression system (AcNPV-Sf cell), silkworm-baculovirus expression system has distinct advantages [20, 21] . First, expression levels in silkworm are 50-1000 times higher than that in insect cell lines. Second, silkworm don not have any pathogens that can cross infect with vertebrates and animal serum is not needed to produce foreign proteins in this expression system, so that the expressed antigens are safer to vertebrates. In view of all these advantages, the silkworm-baculovirus expression system was employed for expression of intact P1-2A, 3C coding regions of FMDV Asia I/HNK/ CHA/05. All five cattle that were vaccinated with diluted expression antigen were induced specific antibody, four of which were considered completely protected. Furthermore, the PD 50 (50% bovine protective dose) value of the subunit vaccine was 6.34 in bovine potency test. Intact P1-2A and 3C protease coding regions from FMDV Asia I/HNK/CHA/05 strain were amplified by RT-PCR and inserted into the transfer vector pVL1393 to generate plasmid pVL-P12A3C. The plasmid was digested with BamH I/EcoR I to generate fragments of 2.3 kb and 10 kb, while fragments of 0.7 kb and12 kb were generated with EcoR I/Bgl II. Based on the length of these fragments, it was confirmed that the full target fragment (P12A3C) was correctly incorporated into the transfer vector and that the expression cassette was located downstream of polyhedrin promoter. Sequence analysis indicated that the P12A3C was 3,024 bp containing full P1-2A and 3C genes and partial 2B,3B genes. Bm-N cell line was co-transfected with baculoviral transfer plasmid pVL-P12A3C and linearized BmBacPAK-6 DNA. The supernatant was collected 4 days post transfection as the viral stock for screening of recombinant virus. Twenty four isolated viral plaques from the plaque assays were cultivated in a 24-well plate and were inoculated into silkworms. The plaque of recombinant virus expressed at maximal activity was selected to purify. The pure recombinant virus from the last round was used as stock virus and confirmed to contain the full expression cassette. The recombinant virus Bm-P12A3C was used to express FMDV protein in cells or silkworm. The expression of polyprotein in Bm-N cells was analyzed by IFAT and sandwich-ELISA. IFAT pictures demonstrated that Bm-N cells infected with Bm-P12A3C produced specific fluorescence, while only very weak background fluorescence appeared in the control cells (Fig. 1 ). This indicated that polyprotein was indeed expressed in Bm-N cell. The sandwich-ELISA results indicated that the FMDV antigen in Bm-P12A3C infected cells was expressed at levels about equivalent to the positive control, but was not detected in BmBacPAK-6 infected cell lysate. The Bm-P12A3C virus was inoculated 1,000 silkworms. The dying silkworm's haemolymph was collected (about 4-5 d post infection, corresponding with rearing mean temperature about 25-27uC). Sandwich-ELISA was conducted to evaluate the expressed antigen in silkworm. The results indicated that OD value of the harvested haemolymph from silkworm infected by Bm-P12A3C decreased as the dilution rate increased, which was in good agreement with variation of positive control of FMDV antigen. The expression yield was about 100 fold (positive antigen: 1/32,0.964; expressed antigen in haemolymph: 1/4096, 0.995) more than the positive control (BHK-21 cell vaccine which had a PD 50 value of 3.6), but was not detectable in the negative control (BmBacPAK-6 infected silkworm's haemolymph) (Fig. 2) . In order to determine the time course of expressed antigen in silkworm and the optimum acquisition time for large scale's production, haemolymph from 10 silkworms was harvested every 12 h beginning at 60 hpi and stored at 220uC. Subsequently, the haemolymph was diluted to 1000 folds for detection of expression products (Fig. 3 ). There was a little at 60 hpi, and the accumulation of recombinant products were dramatically increased from 84 hpi and kept at the high levels during the late phase of infection. So, expressed antigen could be harvested at 108-120 hpi (at the condition about the mean rearing temperature of 25uC). Cattle were vaccinated by the vaccine prepared from 1/30 diluted expressed antigens (the quantity of antigen was about 3 folds of the BHK-21 cell vaccine). LPBE-antibody titer was determined at 7, 14, 21 and 28 dpv following the LPBE method. It was found that all five cattle vaccinated with Bm-P12A3C antigen developed a detectable FMDV-antibody response at 7 dpv, and dramatically reached to high level at 14 dpv. By 21and 28 dpv, the antibody level was maintained at the same level or higher, and reached to a titer of 360 in cattle No33 and No50. In contrast, antibody level in the two control cattle (vaccinated with vaccine prepared from BmBacPAK-6) was not boosted (Table 1) . Furthermore, sera were analyzed for neutralizing antibodies against FMDV (Table 2) . While the two control cattle could not develop any neutralization antibodies against FMDV, all five cattle vaccinated with Bm-P12A3C antigen developed a FMDV-specific neutralizing antibody response in 14pv and maintained at the same level or higher in 28 dpv. The result was in agreement with the LPBE-antibody titer. All vaccinated cattle were challenged with 10,000 BID 50 of Asia I/HNK/CHA/05 at 28 dpv. Body temperature, mouth and feet were observed consecutively for ten days to evaluate the incidence of disease (Table 3) . Four of the five cattle were considered completely protected. Only one vaccinated cattle, No45, developed lesions. The lesions were detectable by 6 dpc and the clinical signs were less severe compared to control group. By contrast, vesicles developed in the control animals by 2 dpc at the sites of all feet and mouth. This indicated that antigen produced in silkworm could be effectively protective. The PD 50 test was performed to assess the subunit vaccine potency by following the bovine potency test protocol described by the OIE to test the traditional inactivated FMD vaccines. In this research, the result showed the vaccine potency of the batch immunized with the expressed antigens reached 6.34 PD 50 per dose (Table 4 ). Discussion FMD still threatens world livestock production. Seven distinct serotypes of FMDV have been identified, named A, O, C, Asia I, SAT I, SAT and SAT with none cross immune protection occurring between them. The FMDV serotype Asia which was first isolated in Pakistan is epidemic within Southeast Asia and Indian peninsula, disseminating among Near East, Middle East and Far East [22] . In March 2005, FMDV serotype Asia was found in HongKong (Asia I/HNK/CHA/05 strain). Subsequently, this type of the virus was reported from mainland of China in April 2005 [23] . The P1 sequence of Asia 1/HNK/CHA/05 isolate was aligned and compared with 9 reference sequences. The result confirmed that Asia 1/HNK/CHA/05 has a high identity with nine Asia I reference sequences from 85.9 to 92.6% [24] . Expression products of baculovirus expressing system are generally considered to be well immunogenic and possess the ability to assemble empty viral capsid. When the same FMDV expression cassette were expressed in E.coli and baculovirus expression system, the expression products from baculovirus excels that from E.coli in terms of the immunogenic [25] and protective effects [26] . Empty capsid comes into being only when capsid precursor P1-2A, protease L and 3C coding region from FMDV O1K serotype were all expressed in baculovirus expressing system(AcMNPV-Sf cell). Truncated protease L can not be selfcleaved from VP0. But, the expressed protease L is harmful to host cell growth, reducing the expression efficiency [27] . In adenovirus expression system, P12A3C expression cassette, including full structure of P1-2A,3C and portion of 2B and 3B, can be expressed and assembled into empty capsids [8] . Myristoylation of the animo terminus of P1-2A is of great importance to the assembly of viral particles [28] . It has been reported that the expression products of AcMNPV-Sf cell can be myristoylated well [29] . Based on the above studies, and using the design previously published by Mayr et al [8] , the P12A3C expression cassette of FMDV serotype Asia I was constructed. After two sorting rounds of recombinant virus and measurements of expression efficiency for more than 20 viral clones, the over-expressed recombinant virus Bm-P12A3C was obtained. It can express with very high efficiency in the hyperexpression variety of silkworm (JY1). The specific antigen produced per milliliter in silkworm haemolymph at least 100 folds more than the BHK-21 cell vaccine which had a PD 50 value of 3.6. Because cattle are the most important economic and susceptible cloven-hoof animal, we designed an experiment to verify whether this produced antigen can be used for preparing a cattle FMD vaccine. We followed the bovine potency test protocol described by the OIE to test this subunit vaccine potency. We used 1/30 diluted dosage to vaccinate five cattle and two controls were vaccinated with vaccine prepared from BmBacPAK-6's. By two weeks post vaccination, the antibody level of the five vaccinated cattle reached a high titer. The antibody level has some ascension but maintained thereafter two weeks, while the control group maintained lower than titer 8. After virulent homologous virus challenge, four of the five were considered protected, and one delayed the disease and ease the clinical symptom, but two unvaccinated cattle developed lesions on all the feet and in the inside of mouth on the second day. Cell-mediated -immune response was probably involved in the protection: that would explain why animal 45 has the same neutralizing antibody titers as 122 but is not protected. This demonstrated that the expression products from silkworm-baculovirus expression system were immunogenic as well. Based on above result, we did the PD50 test to assess the bovine potency of the subunit vaccine. When employed for routine prophylactic use, the vaccine should contain at least 3 PD 50 per dose for cattle by OIE recommended. The result showed the subunit vaccine potency could get 6.34 PD 50 a dose for cattle. This leads to a conclusion that it is feasible to use the silkworm-baculovirus expression system for FMD vaccine production. Genomic RNA was extracted from the viral supernatant with RNeasy (Qiagen) and used immediately for cDNA synthesis. cDNA synthesis was performed with Superscript II reverse transcriptase (Invitrogen). PCR was used to amplify the Intact P1-2A and 3C protease coding regions from the cDNA using two pairs of specific primer: Forward primer for P1-2A: 59-ATAGGATCCACCATGG-GAGCCGGGCAATCCAGCC-39 (BamH I site was introduced) Reverse primer for P1-2A 59-CGCGAATTCTGACATGT CCTCCTGCATCTGGTTG-39 (EcoR I site was introduced) Forward primer for 3C:59-GCGGAATTCAAGAAACCT GTCGCTTTGAAAGT-39 (EcoR I site was introduced) Reverse primer for 3C:59ATAAGATCTCTACTCGTGG TGTGGTTCGGGAT-39 (Bgl II site was introduced) The PCR products were separated by 1% Agarose gel electrophoresis. The target fragments of P1-2A and 3C were inserted into baculoviral transfer vector pVL1393 using BamH I/ EcoR I and EcoR I/Bgl II sites following the routine protocols to generate a transfer plasmid. The whole P1-2A and 3C coding regions in transfer plasmid was sequenced and named pVL-P12A3C. The baculoviral transfer plasmid pVL-P12A3C was co-transfected with linearized Bm-BacPAK6 DNA into Bm-N cells by liposome-mediated method using transfection reagent lipofectamin 2000 (Invitrogen) [30] . The co-transfection supernatant was subject to plaque assays to screen the individual viral plaques. PCR amplification was conducted to confirm that the P1-2A and 3C genes had been incorporated into the baculoviral genome. Primers were designed as follows: Sense: 59-ACTGTTTTCG-TAACAGTTTTGTAA-39 and Reverse: 59-CTACTCGTGGT GTGGTTCGGGAT-39. Another two rounds of screen were performed and the pure recombinant virus was used to generate high titer viral stocks for expression. The expression of FMDV polyprotein in Bm-N cells infected with Bm-P12A3C was analyzed by immunofluorescence test (IFAT) and sandwich-ELISA. The Bm-P12A3C was multiplied in Bm-N cells. Bm-N cells (2.0610 5 ) were cultured on cover slips and inoculated at a MOI of 10 pfu with Bm-P12A3C. After 48 hours post infection (hpi), IFAT was conducted to analyze the expression of FMDV proteins. Cells were then rinsed with PBS for 1 or 2 times and fixed in 100% cold acetone (220uC for 30 min). Samples were incubated with rabbit serum against FMDV (37uC for 30 min) in humid box, washed with PBS for five times, and then stained with fluorescein-conjugated goat anti-rabbit serum at 37uC for 30 min. The cover slips were coated with glycerin and observed on an Olympus fluorescence microscope. Bm-N cells infected with BmBacPAK-6 were used as control. When Bm-N cells infected with Bm-P12A3C were partial floating, they were detached and collected (about 72 hpi), the cells pellet was freezed and thawed at -70/37uC in PBS for three times and centrifuged at 10,000g for 5 min at 4uC. The supernatant was tested using the sandwich-ELISA method. 96-well plat-bottomed plates (Costar) were coated with the rabbit serum against FMDV overnight at 4uC and blocked with defatted milk powder for 1h. Then the plates were washed five times. FMDV antigen (from the vaccine which had a PD 50 value of 3.6), lysates of Bm-N cells with Bm-P12A3C and BmBacPAK-6 infected were diluted in a twofold series and incubated at 37uC for 1h. Subsequently, the plates were washed thoroughly and guinea pig sera against FMDV was added to each well. The plates were incubated at 37uC for 1 h, and then rabbit anti-guinea pig IgG peroxidase conjugate (Sigma) at 1:10000 dilution was added and reacted at 37uC for 1 h. Substrate (0.05% H 2 O 2 plus orthophenylene diamine) was added, reacted for 15 minutes and stopped by the addition of 1M sulphuric acid. Absorbance was determined at 492 nm. Early fifth-instar silkworms were infected with the recombinant virus at about 10 5 pfu per larva. The dying silkworm's haemolymph was collected on ice and stored at 220uC for sandwich-ELISA. The infected silkworm's haemolymph was collected every 12 h starting at 60 hpi for determining the expression course of target antigen. Silkworm haemolymph was lysed ultrasonically and cell debris was removed by centrifugation. The diluted supernatant was used to produce vaccine. Liquid-phase blocking ELISA(LPBE)(http:// www.oie.int/eng/normes/MMANUAL/A_00024.htm) was performed to determine the antibody titer for screening of candidate cattle according to the standard method of World Organization for Animal Health, Office International desEpizooties (OIE) before vaccination. Candidates with a potency lower than 8 were housed in disease-secure isolation facilities in Lanzhou Veterinary Research Institute. Detection kit was prepared by Lanzhou Veterinary Research Institute. Seven cattle (6-8 months old) were immunized by intramuscular inoculation at the site in the neck. Five cattle were vaccinated with 3ml/animal of vaccine with Bm-P12A3C's, while two control cattle were vaccinated with the same dose of vaccine with BmBacPAK-6's. Cattle serum were collected at 7, 14, 21and 28 days postvaccination (dpv). Antibody against FMDV was detected by LPBE method (as above) and SNT. For SNT, the sera was diluted in DMEM as two-fold in a 96well flat-bottomed tissue culture plates (Castor, USA). Virus suspension with a titer of 100 TCID 50 in 50 ul was added to each sera well and the mixture was incubated for 1 h at 37uC and 5% CO2. 50 ul of BHK-21 cell suspension (1.5610 6 ml 21 ) was added to each well and incubated for 4-5 days. Appropriate serum, virus and cell control were included in this test. The plates were observed via microscope for cytopathic effect. According to the descriptions by standard protocol of OIE (http://www.oie.int/eng/normes/MMANUAL/A_00024.htm), all animals were challenged by intradermal inoculation at two sites in the tongue with 10,000 bovine infectious doses (BID 50 ) of Asia I/HNK/CHA/05 at 28 dpv. The body temperature of the animals was monitored daily. The restrained animals were carefully examined in the mouth, and feet every day for the first 10 days after challenge. We followed the bovine potency test protocol described by the OIE to test this subunit vaccine potency. Three groups of five cattle per group and a control group of two non-vaccinated animals were vaccinated. The vaccinated groups were administered different doses (1 dose, 1/3 dose, 1/9 dose)of the subunit vaccine prepared from diluted expression products and all animals were challenged 3 weeks after vaccination with 10,000 BID 50 of Asia I/HNK/CHA/05 to the vaccine under test by intradermal inoculation into two sites on the upper surface of the tongue. The animals were observed daily for 10 days after challenge for clinical signs of FMD. Control animals developed lesions on at least three feet. Unprotected animals showed lesions at sites other than the tongue. From each animal protected in each group, the PD 50 (50% bovine protective dose) content of the vaccine was calculated based on the Reed-Muench method.

The effect of network mixing patterns on epidemic dynamics and the efficacy of disease contact tracing

In networks, nodes may preferentially contact other nodes with similar (assortatively mixed) or dissimilar (disassortatively mixed) numbers of contacts. Different patterns of contact support different epidemic dynamics, potentially affecting the efficacy of control measures such as contact tracing, which aims to identify and isolate nodes with infectious contacts. We used stochastic simulations to investigate the effects of mixing patterns on epidemic dynamics and contact-tracing efficacy. For uncontrolled epidemics, outbreaks occur at lower infection rates for more assortatively mixed networks, with faster initial epidemic growth rate and shorter epidemic duration than for disassortatively mixed networks. Contact tracing performs better for assortative mixing where epidemic size is large and tracing rate low, but it performs better for disassortative mixing at higher contact rates. For assortatively mixed networks, disease spreads first to highly connected nodes, but this is balanced by contact tracing quickly identifying these same nodes. The converse is true for disassortative mixing, where both disease and tracing are less likely to target highly connected nodes. For small epidemics, contact tracing is more effective on disassortative networks due to the greater resilience of assortative networks to link removal. Multi-step contact tracing is more effective than single-step tracing for assortative mixing, but this effect is smaller for disassortatively mixed networks.

For a wide range of epidemic and epizootic diseases, individuals can be usefully modelled as nodes in a network, where the network links represent potentially infectious contacts between individuals. This network representation applies to many complex systems such as the Internet, the World Wide Web (Albert & Barabási 2002) , social and transportation networks (Liljeros et al. 2001; Jones & Handcock 2003; Hufnagel et al. 2004) and livestock movement networks Robinson et al. 2007) . Models that use explicit contact structures between individuals, households, cities, regions, countries or farms have been used to analyse the spread of human diseases, such as SARS (Hufnagel et al. 2004; Meyers et al. 2005) and pandemic influenza (Eubank et al. 2004; Ferguson et al. 2005) , and animal diseases, such as foot-and-mouth disease (FMD; Green et al. 2006a; Kao et al. 2006; Kiss et al. 2006a ) and avian influenza (Le Menach et al. 2006) . Much attention is now focused on analysing network properties and the dynamic processes they support (for reviews, see Newman (2003b) and Keeling & Eames (2005) ), as well as how to effectively control these processes. Contact tracing is commonly used to identify individuals that have been in contact with infectious individuals, to reduce the number of potential onward infections from traced individuals. Contact tracing was successful in the 2003 SARS epidemic (Lipsitch et al. 2003) , but unsuccessful in the 2001 FMD epidemic in the UK (Ferguson et al. 2001; Keeling et al. 2001; Kao 2003) . For simple unstructured populations, contact tracing has been modelled using a combination of branching process analysis complemented by stochastic simulation models (Müller et al. 2000) . However, in more highly structured contact networks, the properties of the network markedly influence the efficacy of contact tracing. For example, contact tracing is more effective on clustered networks than on random networks (Eames & Keeling 2003; Kiss et al. 2005) , and on scale-free networks a higher tracing effort is needed to control an epidemic than on random Poisson networks . In this paper, we examine the impact of departure from the common assumption of proportionate (random) mixing upon the efficacy of contact tracing. Proportionate mixing assumes that the probability of any two nodes being connected is proportional to the product of their number of contacts. This is unlikely in many cases where 'like attaches to like, assortatively mixed' (e.g. STI transmission; Anderson et al. 1990; Catania et al. 1992) or where 'opposites attract, disassortatively mixed' (e.g. livestock trading among farms and markets in the UK ; correlation properties of the Internet (Pastor-Satorras et al. 2001) ). The empirical evidence and importance of connectivity correlations have led to numerous network-and differential equation-based models (Boguñá et al. 2003a; Newman 2003a; Barthélemy et al. 2005) . Many of these models focus on the effect of connectivity correlations on the epidemic threshold, initial growth rate and hierarchical spread. For example, it has been shown that epidemics on networks characterized by high node degree variance grow rapidly, and in the limiting case of infinite variance, instantaneously, independently of the mixing pattern (Boguñá et al. 2003b) . The initial growth of the epidemic changes from exponential to power law if non-random mixing is combined with small-world properties, with the powerlaw exponent determined by the average node distance on the network ( Vazquez 2006) . While in the limit of infinite populations, the implications of preferential mixing for disease invasion are well known, here we focus on disease control and consider the efficacy of contact tracing for different mixing patterns. For networks that are poorly characterized, the efficacy of contact tracing is difficult to determine without exact knowledge of the underlying contact network, i.e. who is connected to whom? By analysing disease transmission on theoretical networks with different mixing patterns, we aim to identify the implications of non-random mixing for epidemic dynamics and control strategies. The network-based disease transmission and contacttracing model is based on models considered by Huerta & Tsimring (2002) , Eames & Keeling (2003) and Kiss et al. (2005 Kiss et al. ( , 2006a . Each node in the network is classified according to one of four states of disease progression: susceptible nodes (S ); nodes at once infected and infectious (I ); nodes 'triggering' contact tracing (T ), which are identified as being infectious, are immediately isolated and initiate tracing; and, finally, removed nodes (R), which are no longer infectious and do not initiate tracing. The transitions between states are depicted in figure 1. Infection S/I. The epidemic is seeded with one or more infected nodes. Thereafter, infection progresses via a contact network; the probability of a node becoming infected depends on the state of the nodes directly linked to it. The probability p of a susceptible node with k infectious contacts becoming infected in a small time interval Dt is pZ1Kexp(KtkDt). Here, t is the rate of infection spreading through a single contact between an infectious and a susceptible node. Contact-independent identification of infectious nodes I/T. Disease is detected at an infectious node (e.g. via clinical signs or screening) at rate a. This results in the isolation of the node and triggers the tracing of its contacts. The removal of triggering nodes T/R. Triggering nodes are removed at rate d. Multi-step contact tracing I/T. The infectious neighbours of triggering nodes (T ) can themselves become triggering nodes through contact tracing if they are found to be infected, creating a multi-step contacttracing chain that tracks the paths of disease transmission. Traced I nodes are not directly removed, but instead enter the T state at rate 4. Unless otherwise stated, we consider multi-step tracing. Single-step contact tracing of infectious nodes I/R. Diagnostic tests are often necessary to determine the status of traced nodes (individuals); these may be imprecise or slow and the isolation and observation of traced individuals may not be a viable option. In such cases, multiple-step contact tracing is less likely. In an alternative single-step contact-tracing model, the infectious neighbours of triggering nodes are traced at rate 4 per contact. These are then directly removed and do not initiate further tracing. Control is always modelled through either multi-step or single-step tracing. Undirected networks with different mixing patterns (assortative and disassortative) are generated using a method proposed by Newman (2003a) . The mixing pattern here is based on node degree (i.e. the number of links of each node). The level of mixing is given by the correlation coefficient of the 'excess' degrees (see below) calculated on all pairs of connected nodes. The generation of networks with different mixing properties is based on a Monte Carlo sampling scheme with repeated link switching at a probability determined by the values of the connectivity matrix EZe ij . Here, e ij is the probability that a randomly chosen link connects a node with i connections to a node with j connections where the link under consideration is itself not counted. Defining the distribution of the 'excess' degree (i.e. degree minus one) of vertices at the end of links as q k Z P j e jk , the level of mixing by vertex S T R I Figure 1 . Transitions among the four disease progression states. Contact tracing is represented by either the I/T transition (multi-step, dashed) or the I/R transition (singlestep, dotted), both with rate 4. Tracing occurs through either multi-step or single-step tracing. degree is given by where s 2 q is the variance of the distribution q k . For disassortatively mixed networks K1%r!0, for random networks rz0 and for assortatively mixed networks 0!r%1. Many real networks display a wide distribution of node degrees (Newman et al. 2001; Albert & Barabási 2002) . To reflect such systems, we consider networks that are generated according to an exponentially truncated power-law degree distribution The function Li n ðxÞ is the nth polylogarithm of x and acts as a normalizing constant. The exponential cut-off of the scale-free distribution is determined by K. Following Newman (2003a), a value of gZ2.5 is used for the power-law exponent. The network analysis that illustrates the effect of preferential mixing is performed for KZ100. For this value of K, the average number of connections per node in the networks is hkiz1.7. To explore a range of parameter values that support meaningful epidemics (i.e. where the epidemic is large enough to be of a significant public health concern), hki must be sufficiently large to allow for a range of new infections in the first generation well above 1. Therefore, the networks are generated by only accepting nodes with kR3 (see Kiss et al. 2006b ). Moreno et al. (2003) proposed a numerical method for epidemic models that can account for connectivity correlations. While this method works without explicitly generating the network, it offers less flexibility when considering disassortatively mixed networks. Here, we use epidemic simulations on networks with NZ10 000 nodes and consider values of rZK0.2, K0.10, K0.05 for disassortatively mixed networks and rZ0.05, 0.10, 0.2 for assortatively mixed networks. The limits are chosen as the values over which the algorithm appears to be robust (M. E. J. Newman 2006, personal communication) ; this range of r values is sufficient to give marked differences in the epidemic threshold for the transmission rate, number of nodes traced, final epidemic size and to illustrate important trends in contact-tracing efficacy. A further check on the mixing pattern is illustrated in the electronic supplementary material by a plot of the average connectivity of a neighbour as a function of node connectivity for different values of r. All simulated epidemics were seeded with 10 index cases chosen at random, in order to avoid early stochastic extinction. Averages of 10 000 simulations are presented, consisting of 100 epidemic runs on each of 100 different network realizations. Proportions across all plots are relative to the total network size NZ10 000. The simulation time step used is DtZ0.04. Smaller time steps produced effectively identical results (not shown). The s.e. values of the averages of simulation outputs are at least three to four orders of magnitude smaller than the measurement itself and therefore are not discussed further. The giant component (GC) is the largest subset of nodes such that any two nodes from this subset can be connected by a series of links. For undirected networks, the GC size represents the upper limit for the potential size of an epidemic. In figure 2 , the structural differences of preferentially mixed networks are demonstrated by the size of the GC as a function of the cut-off parameter K for three different values of the assortativity coefficient r. As the cut-off parameter increases, the number of links in the network increases and the network becomes denser. A K-value exists where the GC size is equal on disassortatively and assortatively mixed networks. Consistent with previous results ( Newman 2002 ( Newman , 2003a , below this value, GC size is larger on assortatively mixed networks and above it, the GC is larger on disassortatively mixed networks, with GC size generally intermediate on random networks. The two different regimes are a direct consequence of the mixing pattern. In assortatively mixed networks, nodes of high degree preferentially connect to each other and form a highly connected core group. Therefore, for low link density, a larger GC size is found than for either disassortatively mixed or random networks. Link density within the GC is higher than in the network as a whole. In contrast, in disassortatively mixed networks at low link density, the links are more dispersed, forming many isolated components of small size (figure 3). As link density increases, the probability that a link with one node in the GC will only connect to another node already in the GC also increases; this is a finite size effect that is exacerbated by assortative mixing since high-degree nodes are relatively few. Thus, at higher link densities, assortatively mixed networks have a smaller GC size compared with disassortatively mixed networks, as in the latter, added links are more likely to result in smaller components being absorbed into the GC (figure 2). The analysis of the GC growth and component distribution was performed on networks generated using gZ2.5 and kR1. For KZ100, the average number of connections per node is hkiz1.7. For random networks, the percolation transition occurs at hkiZ1, above which the network will support large epidemics, i.e. epidemics that scale with total population size (e.g. Moore & Newman 2000; Kao et al. 2006) . Thus, hkiz1.7 supports only a narrow range of epidemiological parameters over which large epidemics can occur. To circumvent this problem, the networks used below were generated using the same parameter values, but only nodes with kR3 were accepted during the network generation process , resulting in networks with hkiz6. We define two variables T p and F for ease of reporting and comparing results. With per-contact transmission rate t and detection rate a, the transmission probability per link over the period before the detection of the infectious node is given by T p Zt/(tCa) (Keeling & Grenfell 2000) . Constant hkiT p provides a constant number of secondary infections caused by the introduction of infection at a node when the remainder of the network is susceptible (Keeling & Grenfell 2000) , at the cost of having different epidemic dynamics as hki varies ). In a parallel manner, with per-contact-tracing rate 4 and removal rate of triggering nodes d, the tracing probability per traceable link over the whole tracing triggering period is given by FZ4/(4Cd). In figure 4 , in the absence of tracing, the final epidemic size R(N) is plotted against the transmission probability for r2{K0.2, K0.1, K0.05, 0, 0.05, 0.1, 0.2}. In this case, compartments T and R are equivalent and the current SITR model is equivalent to the well-known SIR model with an effective infectious period of 1/a. The relationship between R(N) and T p is qualitatively similar to the relationship between GC size and link density in figure 2. The epidemic threshold for assortatively mixed networks occurs at a lower transmission probability than for the other two mixing patterns. However, the final epidemic size approaches its asymptote, total network size, faster for disassortatively mixed networks. The two network types with rZK0.2 and 0.2 produce approximately equal epidemic sizes at tz0.0528 (transmission probability T p z0.15), when R(N)z0.225 on both types. Epidemics on assortatively mixed networks have a faster initial growth rate and a shorter duration than those on disassortatively mixed networks (figure 5a). This is mainly due to the GC containing a 'core' group of high-degree nodes that are highly connected. The differences in the initial epidemic growth rate across the different network types are directly related to the basic reproduction number R 0 (Anderson & May 1991) . The value of R 0 can be estimated as the lead eigenvalue of the next-generation matrix CZ(c ij ) (Diekmann & Heesterbeek 2000) , and, in this case, it can be approximated by the product of the contact matrix and the per-link probability of transmission c ij ZT p M ij . A non-zero entry M ij Z1 denotes that an infectious node j can transmit the infection to a susceptible node i; the magnitude of c ij is given by the per-link transmission probability T p . Estimates of R 0 averaged over 100 generated networks of each type are as follows: R 0 Z12.12T p (s.e. 0.11) for disassortatively mixed networks; R 0 Z15.0T p (s.e. 0.11) for random networks; and R 0 Z17.81T p (s.e. 0.14) for assortatively mixed networks. As expected, the threshold for epidemic outbreaks occurs at lower infection probabilities for assortatively mixed networks than for either disassortatively mixed or random networks. The differences in the initial epidemic growth rate and epidemic duration are likely to have consequences for the efficacy of contact tracing on the different networks. For example, on assortatively mixed networks with fast epidemic turnover, efficient contact tracing has to be comparably fast. The prevalence of traced nodes (dashed line) for assortatively and disassortatively mixed networks is illustrated in figure 5a . The average degree of newly infected (Barthélemy et al. 2004 ) and contact-traced nodes is plotted in figure 5b . While assortatively mixed networks sustain epidemics with fast turnover and quick initial growth rate, they also allow contact tracing to remove a larger number of highly connected nodes early on in the epidemic. In contrast, on disassortatively mixed networks, disease spread is slower, but contact tracing is also less efficient. Contact tracing can be viewed as an exploration of the local network structure (Cohen et al. 2003) , and thus we would expect its efficacy to depend on the mixing patterns of the network. This is investigated by varying 4, while keeping d fixed. For comparison purposes, we contrast the cases where R(N) is the same on both assortatively and disassortatively mixed networks, but possibly with differing transmission rates, later concentrating on the unique transmission rate that results in the same R(N) on both networks. The final epidemic size R(N), the proportion of nodes that become triggering nodes via clinical signs or screening and the proportion of nodes that have been contact traced during the epidemic are plotted in figure 6 as a function of tracing probability F for disassortatively (rZK0.2) and assortatively (rZ0.2) mixed networks. First, transmission rates are chosen such that R(N)z0.73 on both networks (tZ0.125 and 0.175 on disassortatively and assortatively mixed networks, respectively; figure 6a ) and, second, the transmission rates are the same (tZ0.0528) with R(N)z0.225 on both networks (figure 6b). For the first case, the effect of contact tracing is similar on both networks with comparable R(N); however, we identify two distinct regimes above and below FZ0.61, above which contact tracing becomes more effective on disassortatively mixed networks (figure 6a). In figure 6a , if F%0.61, R(N) is smaller on assortatively mixed networks. Contact tracing on assortatively mixed networks removes nodes of higher degree (figure 5b) than on disassortatively mixed networks. For small F values, the final epidemic size is still high and the population of highly connected nodes is depleted well before the epidemic ends, as shown by the crossover between the average degree of newly infected nodes on the two network types (figure 5b). Over the epidemic, susceptible nodes are on average of lower degree and are more difficult to reach. On disassortatively mixed networks, the depletion of highly connected nodes is less marked and epidemics persist, with disease spread alternating between highly and less well-connected nodes producing an average degree of newly infected nodes which is more even in time (figure 5b). In the latter stages of the epidemic, on disassortatively mixed networks, highly connected nodes are not completely depleted and can become infected. Thus, on the assortatively mixed networks, the epidemic ends earlier with a smaller final epidemic size. On assortatively mixed networks, when global depletion of susceptible nodes is important, contact tracing acts to enhance the early depletion of nodes of high degree (figure 5b). At higher tracing probability, the proportion of nodes removed through contact tracing (figure 6a) decreases with increasing tracing probability. This is indicative of effective control with a limited proportion of nodes becoming infectious and, hence, fewer targets for tracing. In the regime of more effective control (FR0.61), the epidemics die out early on and susceptible depletion becomes less important. This is illustrated in figure 7 , showing that the depletion of susceptibles is only found at low F values. This is also reflected in the very rapid reduction of R g (i.e. average reproduction ratio in generation g defined as the ratio between the number of nodes infected in consecutive generations) below 1 when F is high. This is a more general effect corroborated by examining contacttracing efficacy in a different parameter regime. In figure 6b , the scenario where tZ0.0528 on both networks is considered. For this value of t and in the absence of tracing, the final epidemic size is the same on both networks (R(N)z0.225). Here, the final epidemic size is similar to the point at which contact tracing starts to perform better on disassortatively mixed networks in figure 6a. In this case, for all F values, contact tracing always performs better on disassortatively mixed networks. The behaviour in this regime can be explained by interpreting contact tracing as a mechanism acting to reduce the effective transmission probability T p past the first generation of infection. Contact tracing achieves this by shortening the average infectious period of traced nodes. In figure 4 , lower transmission probabilities correspond to lower values of the transmission rate t. While we do not have exact analytic relationships between the implications of reducing the transmission probability through the different routes, the effect of contact tracing can be approximated by following the trend of the final epidemic size (figure 4) as the transmission probability decreases. The steeper curve of the final epidemic size on disassortatively mixed networks suggests that a small decrease in the transmission probability has a more marked effect on final epidemic size than on assortatively mixed networks. This effect is especially dominant since in the absence of tracing the same R(N) is observed on both networks. This supports the higher contact-tracing efficacy observed on disassortatively mixed networks. Multiple-step contact tracing is not always logistically feasible. It is therefore important to determine the benefit provided by it. For both network types (figure 8), there are considerable differences between the final epidemic sizes for single-and multiple-step contact tracing at high tracing rates, particularly for assortatively mixed networks (figure 8b). In multiple-step tracing, a triggering node can generate other triggering nodes, potentially creating a cascade of triggering nodes throughout the infected portion of the network. This leads to tracing a higher number of infectious nodes and reducing the number of links that successfully transmit the disease. For high tracing rates, the epidemics are short-lived, and, on assortatively mixed networks, as a result of the faster initial epidemic growth rate, the extra proportion of untraced infectious nodes in the case of single-step tracing generates a higher number of infections and hence there is a marked difference between single-and multistep tracing. In figure 8 , the faster decrease in the proportion of traced nodes with increasing tracing probability indicates earlier and more effective control in the case of multiple-step contact tracing compared with the single-step case. Contact tracing performs comparably well on both assortatively and disassortatively mixed networks. This is mainly explained by a balance that is reached between the epidemic time scale (i.e. slow for disassortatively mixed and fast for assortatively mixed; figure 5a) and the hierarchy of spread (figure 5b) on assortatively mixed networks (and the lack of it on disassortatively mixed networks) on the one hand, and the contact-tracing mechanism. On assortatively mixed networks, the epidemic spread is faster and the disease typically spreads to nodes with high degree. This is counterbalanced by contact tracing that removes infectious nodes with high degree. On disassortatively mixed networks, the epidemic spread is slower; however, owing to the connectivity pattern, contact tracing also alternates between removing poorly and highly connected nodes and therefore is comparably less effective. The higher average degree of traced nodes is a reflection of the average degree of nodes becoming infected earlier on in the epidemic when highly connected nodes are more abundant. This combined with the depletion of highly connected nodes, accentuated by finite size effects, leads to producing a higher average degree of traced nodes when compared with the average degree of infected nodes. In the case of large epidemics and small values of the contact-tracing rate, contact tracing is more effective on assortatively mixed networks than on disassortative mixed networks, although the difference is small. Here, on the assortatively mixed networks, the early, global depletion of highly connected nodes results in a rapid increase in the proportion of susceptible nodes that are poorly connected nodes, and thus are less likely to become infected. For smaller values of the final epidemic size, the epidemics die out earlier and the depletion of susceptible nodes is less important. The efficacy of contact tracing in this case is determined by the more resilient nature of the assortatively mixed networks to the removal of potentially infectious links through tracing. In the case of singlestep versus multi-step contact tracing, the differences are more marked for assortatively mixed networks and small epidemics. The algorithm used to generate networks with different mixing patterns is robust for mixing values in the range of K0.2 to 0.2. This range covers many of the values measured from networks based on real data (Newman 2002) . For more marked differences in the mixing pattern, we expect similar qualitative conclusions with possibly more significant quantitative differences. For the models presented above, differences in contact-tracing efficacy were investigated for various parameter values in addition to those presented. The results agreed qualitatively across the range of parameter values studied. However, further investigation is needed to determine the relative contributions of the different determinants of contact-tracing efficacy. The model presented here does not incorporate time or resource constraints for contact tracing. The implementation of epidemic control strategies often involves qualified personnel and costly or time-consuming diagnostic tests. Although contact tracing performs comparably well on both network types, the faster time course of the epidemic on assortatively mixed networks is more likely to stretch resources in real situations, since it requires a greater and more timely concentration of resources. As in previous studies investigating the effects of contact clustering (Kiss et al. 2005 ) and degree distribution ) on contact tracing, we show that, unless contact tracing is very good, the mixing patterns have little effect on the course of the epidemic and the number of nodes removed. These results would suggest that it is difficult to exploit network structure to achieve better control via tracing. This may seem somewhat surprising, as previous studies have shown that control strategies such as acquaintance sampling (Cohen et al. 2003 ) that is based on the local exploration of the population contact structure provide an efficient epidemic control strategy, compared with random removal of nodes. However, such studies considered only networks that are randomly mixed and compare random removal with targeted removal (a form of contact tracing). Here, however, we compare targeted removal but where the networks themselves differ. Like the disease itself, contact tracing exploits the local network structure and, for assortatively mixed networks, identifies and removes early the highly connected nodes. This beneficial effect, however, is counterbalanced by the fast initial disease spread to such highly important nodes. Therefore, in the present case, the properties of the network can at the same time enhance disease spread and also increase control efficacy, highlighting the non-trivial interactions between the network structure and the dynamics on networks, showing that added attention is needed when evaluating the efficacy of epidemic control strategies.

General Practice and Pandemic Influenza: A Framework for Planning and Comparison of Plans in Five Countries

BACKGROUND: Although primary health care, and in particular, general practice will be at the frontline in the response to pandemic influenza, there are no frameworks to guide systematic planning for this task or to appraise available plans for their relevance to general practice. We aimed to develop a framework that will facilitate planning for general practice, and used it to appraise pandemic plans from Australia, England, USA, New Zealand and Canada. METHODOLOGY/PRINCIPAL FINDINGS: We adapted the Haddon matrix to develop the framework, populating its cells through a multi-method study that incorporated the peer-reviewed and grey literature, interviews with general practitioners, practice nurses and senior decision-makers, and desktop simulation exercises. We used the framework to analyse 89 publicly-available jurisdictional plans at similar managerial levels in the five countries. The framework identifies four functional domains: clinical care for influenza and other needs, public health responsibilities, the internal environment and the macro-environment of general practice. No plan addressed all four domains. Most plans either ignored or were sketchy about non-influenza clinical needs, and about the contribution of general practice to public health beyond surveillance. Collaborations between general practices were addressed in few plans, and inter-relationships with the broader health system, even less frequently. CONCLUSIONS: This is the first study to provide a framework to guide general practice planning for pandemic influenza. The framework helped identify critical shortcomings in available plans. Engaging general practice effectively in planning is challenging, particularly where governance structures for primary health care are weak. We identify implications for practice and for research.

Primary health care, and in particular general practice, will be at the frontline in the response to pandemic influenza. Preparedness planning for this sector has lagged behind public health planning, despite evidence from SARS [1, 2] and influenza epidemics [3] of the important role played by general practice. Preparedness may be defined as the capacity to respond to a range of public health threats including natural disasters and infectious disease outbreaks, human-caused accidents and intentional attacks [4] . There is an increasing recognition of the need for an 'allhazards' approach to planning that integrates acute clinical care, public health, and emergency management systems [4] . Since September 2001, the US government has invested about $5 billion to upgrade preparedness plans for emergency management systems [5, 6] . There are three challenges for pandemic planning by general practice. First, there is no systematic framework for planning this sector's response. Preparing for health threats and emergencies is an essential function of public health, but is not core business for general practice. Second, the way in which ambulatory health services will interact with each other and with the broader health system response to a pandemic is unclear. General practitioners (GPs) in Canada [7] , Australia [8] and the UK [9] have expressed uncertainty about how to participate in such a response. Third, planning and implementing changes for pandemic influenza across the health system is complex. Although there is little evidence linking specific preparedness activities to effective system-wide responses to pandemic influenza [5, 6] , change management theories point to a need for dynamic partnerships between general practices and other ambulatory care services, hospitals and public health departments [10] . The strength and structure of these linkages vary around the world, depending on decentralisation processes, the regulatory and legal system, and financing within health systems [11, 12] . Although general practice, or family medicine, is organised differently in different countries, there is considerable potential for transferable learning at the meso-level of management planning [11] . We aimed to develop a framework that will facilitate systematic planning for the general practice response to pandemic influenza and used it to appraise coverage of key elements in publicly available pandemic plans from Australia, England, USA, New Zealand and Canada. To guide planning and to appraise available plans, we adapted the Haddon Matrix, a planning tool developed in the field of injury research and intervention [13] , and more recently applied to the public health response to bioterrorism, SARS [14] , and pandemic influenza [15] . The matrix consists of a grid of columns of four factors (human, agent, and physical and organisational environment) impacting upon the event [15] . Pandemic influenza may be perceived as a form of injury on a mass scale and the matrix helps us understand the multi-dimensional nature of epidemics and of the associated challenges that could be expected by general practice. The framework can be readily shared with public health units and other parts of the health system, as it identifies the general practice contributions to primary health care services and to public health surveillance and control. Because all disasters are local, the matrix is flexible enough to allow a focused analysis of the smallest unit of study, such as an individual, or group of general practitioners. The methods used to construct the cells of the modified Haddon matrix have been detailed elsewhere [16] . In brief, a team with expertise in social science, public health and general practice reviewed objectives and strategies in WHO guidelines for preparing and responding to a pandemic [17] to define the context and potential contributions of general practice. Next, we undertook a narrative review of the peer-reviewed and grey literature on pandemic influenza to identify papers that elaborated strategies relevant for general practice. A search of the peerreviewed literature through PubMed using the terms 'general practice', 'family physician', 'family medicine' and various combinations of the terms 'influenza', 'epidemic', 'preparedness' and 'pandemic' yielded 24 eligible papers from 157 search results . The process of constructing the framework and populating the cells was informed by organisational theories that emphasise multilevel approaches to change from the individual to the broader health system [10, 18] , and by methods for measuring [5] and improving the quality [6] of public health emergency preparedness. We tested our framework through interviews with a purposive sample of health professionals engaged in pandemic planning. Nineteen general practitioners and practice nurses with expertise in pandemic planning were nominated by the two participating Divisions of General Practice, each of which was a national leader in disaster preparedness and response. Eight general practice policy leaders were identified by representative organisations (Australian Medical Association, Royal Australian College of General Practitioners, Australian General Practice Network). Group interviews were held with 14 state and territory public health leaders attending a national pandemic preparedness meeting. We held two workshops, attended by representatives of state and territory health services, Commonwealth policymakers, non-government organisations, and general practice organisations. In addition, we conducted two focus groups of GPs and nurses working in aged care in two cities. Finally, we undertook four desktop exercises [19] attended by 25 GPs, 11 practice nurses and 10 administrative staff. The five countries in this study had national response plans. Contextualised detail about health-sector responses is contained in plans at the level of administrative decentralisation where decisions are made about patient-service groupings including general practice. In practice, this level was the state or provincial health departments in Federal systems where those jurisdictions have responsibility for health service management and planning (USA, Canada, and Australia). In England, the managerial level for health services is located at the Primary Care Trust (PCT), while in New Zealand it occurs at the level of the District Health Board. Although these are not identical loci of health service governance, they were sufficiently similar in the planning aims for comparisons to be drawn. Plans were obtained from websites of health departments of states or provinces (USA, Australia, Canada), District Health Boards (New Zealand) and PCTs (England) ( Figure S1 ). For New Zealand and England, publicly available records of Board Meetings were also examined. Consumer information and isolated sub plans (e.g. for infection control) were excluded. Plans for 95 jurisdictions were identified; six were excluded as they addressed isolated aspects such as only the distribution of medications, or communication with the public, leaving 89 plans suitable for analysis. Of the five countries, Canada exhibits the most variation between provinces in health system coordination. We examined the websites of Canada's 84 provincial regional health authorities (RHAS, 14 plans identified) and Ontario's 36 public health units (26 plans identified) and 14 Local Health Integration Networks (no pandemic plans identified). We excluded the RHA and public health unit plans from inter-country quantitative analysis, as their level of devolution and/or responsibilities for health management differed from those examined in the other four countries, but have included descriptive details from some of the RHA plans where they illustrate innovative approaches. All plans were examined by two clinicians, and searched for the following terms: primary care, primary health, ambulatory, general practice, general practitioner, GP, family practice, family physician. The roles of general practice/family practice in the plans were assessed across the four domains of general practice identified in the first part of this project. No attempt was made to quantify the extent of coverage of general practice in the plans as this rarely extended beyond a few sentences. Where there was detailed coverage of an issue, we analysed the text and the health system context. The study was approved by the Australian National University Human Research Ethics Committee and the National Research and Evaluation Ethics Committee of the Royal Australian College of General Practitioners. Written informed consent was obtained from participants. A conceptual framework of the general practice response to pandemic influenza is shown in Table 1 . The framework identifies four domains of practice: clinical services, public health responsibilities of general practice, internal (physical and organisational) environment of the general practice unit, and the macro-environment of general practice. In each domain, we list the key challenges to be anticipated by general practice during an influenza pandemic, and the type of responses that need to be addressed in the plan. Table 2 summarises the organisational levels in the five countries, the proportion of jurisdictions with accessible pandemic plans, and coverage of general practice in these plans. While almost all plans from US jurisdictions were accessible, three quarters of Australian states/territories and one third of New Zealand's District Health Boards had accessible plans. Only 13% (20/152) of England's PCTs had pandemic plans available in the public domain. Figure S1 shows the jurisdictions and health management systems whose plans were included in this study; they comprise 49 jurisdictions from the USA, 20 from England, 8 from Canada, and 6 each from Australia and New Zealand. Table 3 shows the number and rates of coverage of each of the four domains of the general practice response in jurisdictional plans of the five countries. The domain covered most frequently was influenza-related clinical care (in all plans from England and Canada). Overall less than half the plans mentioned non-influenza clinical care, with the exception being England, where 90% of PCT plans mentioned non-influenza clinical care. Public health surveillance was addressed in all plans from Canada and New Zealand and infection control in general practice in almost all plans from England and Canada. Functional linkages of general practice with other parts of the health system were addressed in almost all the English plans, but a smaller proportion of other plans. Clinical care Essential planning elements. This domain includes two sets of clinical care needs. The first, prevention and treatment of influenza, includes care for the surge in patients with acute respiratory illness, and for people at high risk of exposure to, or complications from, influenza. These aspects are discussed extensively in the literature [20] [21] [22] [23] . Most people with influenza can be managed in the community, protecting hospitals by delaying or avoiding admission and facilitating early discharge. The second clinical care need is for non-influenza-related care. General practitioners provide most chronic disease care, though there is inter-country variation in their capacities to do this efficiently [24, 25] . While activities like cervical screening may cease in a pandemic, chronic illnesses like diabetes or cardiac disease will still need management. Some acute care usually undertaken in hospitals, like acute asthma or injuries, may be transferred to the community. In an earlier paper, we advanced a range of models of practice to balance clinical services for influenza and non-influenza care [16] . In the recovery phase, the clinical needs of patients are for psychological care and chronic illness management. If the pandemic occurs in waves, as in 1918-19, recovery activities may need to be tempered by preparations for the next wave. Coverage of essential elements in plans. All Canadian and English plans outlined a role for general practice in clinical care for influenza. While only 41% of plans from the USA addressed clinical care for influenza by primary care practitioners (Table 3) , every US plan included guidelines on influenza management by hospital physicians. Some plans articulated a surge in demand for influenza care as a threat to general practice's survival, and proposed assessment and treatment clinics as a way of protecting them [26, 27] . In other plans [28] [29] [30] the response to a surge was to support general practices to become more resilient by collaborating and changing their work practices. In two US state plans, the failure of the ambulatory care sector in the face of a surge was assumed. The planning challenge became to find ways to redeploy workers into other health care sectors [31, 32] . Most plans were sketchy on systems to maintain non-influenzarelated clinical care, with the exception of some PCT plans, which included activities like triage, extended prescribing, identifying deferrable reasons for presentation, and management of more acute problems to protect hospitals [29, [33] [34] [35] [36] . The main non-influenza clinical area was mental health care, mentioned in six plans from the USA [37] [38] [39] [40] [41] [42] (reflecting a focus in the national plan [43] ) and one Canadian plan [44] . Coverage of the needs of vulnerable populations-the elderly, homeless, prisoners and the psychologically unwell -was most detailed in plans from Canada and England. Essential planning elements. This domain includes surveillance of influenza-like illness and influenza virology, and control of influenza in the general practice and the community. Surveillance includes early diagnosis and notification, and specimen collection to confirm clinical diagnosis and to monitor viral characteristics and resistance to antiviral drugs. GPs and private specialists are currently central to surveillance activities [45] [46] [47] [48] . In the early stages of the pandemic, it is likely that public health authorities will undertake contact tracing to facilitate containment, but their capacity to sustain this approach as the epidemic continues will be limited. General practice may then be expected to include contact tracing, and monitoring and support of people in quarantine or home isolation. Other responsibilities may include prescribing and dispensing antiviral drugs and participating in mass immunisations against the pandemic strain of the virus. Coverage of essential elements in plans. Surveillance in general practice was mentioned in 53% of US plans and in only 33% of English plans, in all Canadian and New Zealand plans, and all but one Australian plan ( Table 3 ). The low rates of coverage of surveillance in PCT plans are not in accord with the UK plan which imputes to general practice a role in surveillance, and recommends that PCTs operationalise this recommendation [49] . The College of Family Physicians in Canada is a partner in FluWatch, recruiting sentinel physicians to undertake surveillance, so this role is well understood within the Canadian health sector. The role of general practice in contact tracing, in monitoring people in home isolation, and in distributing antiviral drugs is unclear in most plans. Home care by GPs for people in quarantine is mentioned in two US Plans [50, 51] , and one English plan [36] , though the recently released guidelines for PCTs anticipate a role for general practices in home care [52] . In all country plans, dispensing antiviral medications was generally performed by public health units. Only 22% of PCT plans and 40% of US plans mention a role for primary care in dispensing antiviral medications. None of the Canadian plans, and only one NZ and two Australian state plans, mentioned antiviral dispensing by primary care. The only plan to set out contingencies when decisions about dispensing may change was one Canadian RHA plan [27] . Although immunisation was mentioned most frequently after surveillance as a public health activity by general practices, in most plans the immunisations were against pneumococcal disease and seasonal influenza, but not mass immunisations against pandemic influenza. Essential planning elements. This domain includes the physical environment of the general practice and its practice-level organisation. The risk of transmission of infections within the surgery could be minimised through separate waiting rooms and entrances, triage and personal protective equipment and handwashing facilities. Hogg has outlined infections control procedures in the practice and the associated financial costs [53] . Some general practices (for example, those with small waiting rooms, or only one consulting room) may be deemed too much of a transmission risk to continue providing face-to-face services. The practice needs to develop strategies to maintain reliable and efficient access to essential drugs and equipment and influenza and pneumococcal vaccines. It also needs to strengthen the capacity of its communication technologies with patients and the broader health system, including telephones, faxes, internet, work-from-home technologies for staff, compatible software for sharing electronic medical records, and recall and reminder systems for patients. Preparation at the organisational level relates mainly to business continuity plans. These plans should include leadership delegations, staffing contingencies, safe and flexible working hours and family care plans for staff, criteria for considering clinic closure, recruiting and training ancillary staff, early psycho-social support, support for making difficult clinical decisions, record keeping to ensure accountability for actions and 'inactions', use of antiviral medications, and plans for simulation exercises to complement training, and to evaluate and refine local practice plans. Tools [54, 55] and desktop simulation exercises [19] are available to help GPs plan for continuity. Coverage of essential elements in plans. Infection control strategies were well covered in plans from Canada and England, but were mentioned in only 39% of US plans ( Table 3) . None of the plans provided an inventory of fixed features, such as size and layout of waiting room, or a single entrance, which could compromise infection control. Business continuity was a focus of the English plans, which frequently referenced resources available on the UK Resilience website [56] . This aspect of preparedness was enhanced after the Exercise Winter Willow simulation in February 2007, and new PCT guidelines addressing workforce planning [52] . Some PCT plans addressed the need for general practice resilience in the face of workforce sicknesses [33] , increased aggression from patients, and threatened loss of capacity in single doctor practices [57] . Few plans from other countries discussed business continuity for primary care in such detail. This may be because such issues are felt to be outside the normal purview of state or provinces, and to be the responsibilities of the businesses themselves or corporate interests. Essential planning elements. This domain includes the overall organisation of, and interactions with, the health system that will facilitate or impede effective functioning of general practice services during a pandemic, including adaptation of relevant regulatory and financing systems. The health system requires a plan that adopts the 'all-hazards approach' and integrates roles, responsibilities and actions for acute clinical care, public health, and emergency management systems [4] . This calls for coordination across general practices and other ambulatory care services to ensure primary health care needs within the community are effectively monitored and addressed; with hospitals to avoid/delay hospitalisation and facilitate early discharge; and with public health units to share responsibilities for contact tracing, monitoring and treating people in home isolation or quarantine, dispensing of anti-viral medications, and participation in mass immunisations against pandemic strains of the virus (when these become available). Neighbouring general practices and other ambulatory care services will need local leadership with strategic approaches to collaborate and maintain services through a pandemic. England's PCTs and New Zealand's Primary Health Organisations (PHOs) represent two ways of linking general practices under the governance of regional boards. These networks are consolidated by financial relationships between the PCT or the PHO and general practices. The links between Australia's Divisions of General Practices and GPs are purely voluntary. In the USA, managed care systems function as another way of linking ambulatory and hospital services. Communication infrastructure between Canada's family practitioners, 25% of whom are solo practitioners [58], is still being developed, as is the incorporation of general practice into Canada's Pan-Canadian Public Health Network [59] . The regulatory environment includes accreditation of retired medical practitioners and allied health professionals, laws and regulations which support or hinder the flow of qualified personnel across a jurisdiction's health facilities [48] , and ensuring an appropriate medicolegal framework to support clinical decisions on prioritising medical care during a pandemic, for example, modifying clinical standards, deferring treatment, and restricting access to certain treatments. Funding mechanisms for general practice may impact upon the capacity to provide extra services. In countries with fee-for-service payment systems, general practices may profit from a surge in attendances, but may equally run into business difficulties if they are short-staffed for prolonged periods. GPs funded through a capitated system may have more freedom to alter their practice to provide different service mixes. In the post-event phase, patients and GPs may require support for psychological recovery. It may be necessary to provide some formal relief through a system of locum GPs from areas less affected by the pandemic. Organisational partnerships at this stage may need to be with social services and mental health support services. Coverage of essential elements in plans. Countries with mechanisms for linking general practices with other sectors were more likely to address networking in their plans. Ninety five per cent of English plans addressed systems to support collaboration between general practices (Table 3) . These plans addressed buddy systems, practice networks, and contingency plans for communities of practice. Four of the six New Zealand plans also addressed collaboration, though only one in significant detail; this plan outlined a distinction between key practices, and other practices which might decide to partner one another [55] . Of the three Canadian provincial plans that addressed collaboration, the most comprehensive was from Quebec, which identified a need to bridge the gap between salaried practitioners and independent physicians. The plan of the Montreal Regional Authority [60] operationalises this by setting up a system of active and sustained outreach by the public health department to independent physicians. The absence of plans for networking between general practice and public health is most marked in the USA. With the exception of Louisiana [61] , US plans which mentioned networking did so in one line, generally advocating partnership between private and public services without indicating how this might occur. Louisiana's strategic approach built a participatory structure for rural practitioners through a partnership between the state public health department and the Bureau of Primary Rural Health Care. The Canadian national pandemic plan [62] is framed around a set of ethical precepts incorporated into pandemic planning at the provincial and regional health level. The UK has recently released an ethical framework for policy and planning, though this has not yet been incorporated into planning documents [63] . The regulatory framework most mentioned was in relation to credentialing for retired GPs and other volunteers [33, 64, 65] , and less frequently, indemnity [36] . Although most plans include coverage of the relevant public health legislation, no country's plan included an inventory of legislation relevant to general practice that might need to be amended. Only one plan [66] and the PCT guidelines [52] , canvas the potential of recompense for financial loss to a general practice. The only country in which the planning level coincided with the level that made decisions about funding of health care was Canada. One regional health authority plan provided an outline of specific issues likely to affect physicians, and raised the possibility of reviewing funding mechanisms in a pandemic [67] . There appear to be no ancillary plans addressing principles of altered funding for private physicians in a pandemic. This is the first study to provide a framework that brings together multiple functions, structural relationships and the responsiveness of general practice to prepare for pandemic influenza. The framework provides clarity of purpose and a structure to guide planning through four functional domains: clinical care, public health responsibilities, and the internal and macro environments of general practice. The domains have been structured as integral components of a complex system that can respond to uncertainty [68] and be adapted for a given local setting and health system context. We draw three conclusions regarding general practice from our analysis. First, none of the 89 jurisdictional plans addressed all domains of the general practice response during a pandemic. Second, while many aspects of the first three domains are included in plans for general practice, there are critical gaps and inconsistencies in the fourth domain (macro-environment) that render some elements of the jurisdictional plan ungrounded or unrealistic. Third, few plans addressed the broader ambulatory care context, including the need to engage private specialists and other allied health professionals [48] . Planning and implementing change across the health system is complex. Targeting individual sectors for change (e.g. public health departments, hospitals or general practices) without securing reciprocal changes and strengthening inter-relationships across the health system, is unlikely to succeed [10, 18] . Planners must consider how connectivity across the health system might be strengthened to enable optimal use of general practice resources for planning [68] . While this may be challenging, particularly in countries with weak governance structures for primary health care, omitting general practice input into the planning process may be considered unethical [69] and counterproductive. Limitations of the study: Our findings are exploratory rather than definitive, and indicate directions for further planning and research. Like any new tool, the framework and its application in a given context needs testing and refinement through simulation exercises targeting ambulatory care services as well as the broader health system. Planning is an evolving activity that reflects a 'map' rather than a 'destination', and our findings provide a snapshot of the plans accessible in late 2007. The scope and content of the plans will change over time, as seen in two countries that adjusted their plans after simulation exercises, Exercise Cumpston in Australia [70] and Winter Willow in the UK [71] . Interestingly, the former identified specific weaknesses in the involvement of the primary health care sector and made recommendations to better integrate primary health care providers into planning at the national and jurisdictional levels [70] . National and sub-national pandemic plans may be intended to provide a strategic focus and not to elaborate on operational activities; it is possible the latter may have been addressed, but were not accessible at the time of our study. Another potential limitation of our study is that the gaps we identified in many plans were grounded in theories about the ways to enhance the quality and outcomes of clinical care [10, 18] or of public health preparedness planning [6] . The science of preparedness planning is still maturing [4] [5] [6] and there is relatively little systematic evidence for linking specific preparedness structures to the ability to implement efficient and effective responses [5, 6] . Two important limitations to the implementation of preparedness activities are uncertainties in knowing how much preparedness is enough [5] and in having a measurable assessment of the outcomes of preparedness activities. It may be more meaningful to perceive of the activities as a 'preparedness production system' in which a variety of processes and activities have been completed to prepare for an optimal response [6] . We are unable to comment on the extent to which these preparedness plans have been implemented, except in the case of those jurisdictions which have held pandemic exercises [70, 71] . General practice response is rarely tested in pandemic exercises, which tend to focus on hospital and public health responses. A notable exception is Operation Sparrowhawk in Singapore, where the feasibility of general practice influenza clinics was tested [72] The Haddon matrix is not a final check-list for preparedness planning but a problem-solving tool used as a starting framework for planning. The contents of each cell of the matrix help identify a particular problem or challenge that needs to be addressed. We recognise that the challenges will be neither static over time, nor uniform across general practices; responses will have to be modified in the context of the general practice setting as the pandemic evolves and as other parts of health system, particularly hospitals and public health units respond to the epidemic. Implications of our study for primary health care in developing countries: Endemic and epidemic infectious diseases inflict high levels of morbidity and mortality in developing countries because of a combination of poor living conditions, effects of multiple concurrent illnesses particularly in children, fragile national health systems, overburdened and overstressed health workers, and negative work environments [73] . Although our study targeted general practice in developed countries, the conceptual framework we developed (Table 1) can be used by primary health care services in developing countries to deconstruct the multidimensional challenges posed by pandemic influenza. Identifying possible solutions and apportioning responsibilities across components of the health system is more complex. Operational guidelines have been developed for the detection and rapid containment of a potentially pandemic strain of influenza to the epicentre of the outbreak [74] , for example, if this were to occur in a South East Asian country. However, because of the immense global implications of such an event, this intensive strategy will need to be supported by extraordinary resources from the global community, an action not sustainable once the pandemic strain spreads beyond the initial epicentre. In an analysis of pandemic influenza plans in Asia-Pacific countries in 2006, Coker found that although all countries recognised the importance of pandemic planning, operational responsibility particularly at the local level, remained unclear; most plans relied on specialised flu hospitals, while few developed the possibility of caring for patients at home [75] . (The study made no reference to primary health care or the private practice sector). In his analysis of public health emergencies in developing countries, Quarantelli identified relatively poor adaptive capabilities to be the key barrier to effective responses at the central and local levels [76] . Possible reasons included poorer public health infrastructures and human and financial resources, organisational structures that functioned mainly in a top-down manner with a strong emphasis on structures more than functions, and lack of planning initiatives the further away one moved from central level [76] . Many poor countries already have a health crisis, and need massive international investments, including mobilisation and strengthening of human resources to build sustainable health systems, strong leadership and political commitment [73] . In the face of the pandemic threat, primary health care in developing countries will need resources to develop a suite of policies, including: clarification of what essential primary health care will continue through a pandemic, developing health workforce plans that may entail diverting clinicians from other areas of the health workforce, establishing non-hierarchical links between primary health care, hospitals and public health, and injecting funds into hospital and primary care preparedness simultaneously. It may be argued that the absence of general practice elements from pandemic plans is not problematic, that it is outside the responsibility of public health departments that do not have a governance role for general practice. We argue instead that the general practice sector, which is characterised by loose networks between ambulatory care services, and often lacks the appropriate organisational structure and mandate, cannot spearhead many elements of planning for primary care. This calls for actions by health departments as well as by general practices. Actions by health departments. Ensuring that the community receives appropriate health care during public health emergencies is a government responsibility. Consequently, health departments must emphasise in national and sub-national plans, the critical need for all levels of the health system to integrate the general practice sector in the planning process. This should include appropriate general practice representation in high level planning and decision-making committees, in incident-commandcontrol structures and in the management of community-based specialised clinics such as 'fever clinics' or 'community information and assessment centres'. Good planning must focus on the planning process rather than the production of a written document [76] . The process includes collaborative activities such as meetings, drills, exercises, simulations, developing techniques for training, knowledge transfer, identifying and obtaining resource materials, and continually updating materials and strategies. These planning activities are important not only because they inform, but because they also foster collaborative learning and problem-solving, and generate an atmosphere of mutual trust and solidarity among people who will be affected by a pandemic and whose collaboration will be essential in the response. The willing general practitioner sector [7, 8] is an essential resource for extending the surge capacity of health departments. Health departments should harness and support interactions and networking among general practices, and between them and ambulatory health care providers, hospitals and public health units. The role of general practice in contact tracing, monitoring and treating people in home isolation or quarantine, dispensing antiviral drugs and participating in mass vaccinations -omitted in most plans -needs to be clarified. In addition, health departments should modify or adopt where appropriate, legislation and financing mechanisms to enable general practices to function optimally during the pandemic. Action to support planning by general practice. While the diversity of the general practice sector means that there will not be guidelines to cover all scenarios and contexts, a coherent approach would enable multi-actor accountability and more efficient, contextual planning by jurisdictions. The guidelines for PCTs [52] are an example of such an approach, designed for a particular health system. They could act as a useful point of departure for planning integrated general practice plans by other health systems. There is a need for a system of sharing innovations and exemplary solutions to challenges for pandemic planning by general practice, analogous to those targeting mainly hospitals and public health departments [77] . Given the diversity in organisation of general practice systems, a web presence comparing exemplary approaches from different health systems would be a useful resource for planners. An important challenge will be ensuring collaboration and coordination across the health sector during a pandemic. Research is needed to identify the prevailing barriers and facilitators to effective collaboration across the health sector, how these may change under the stressor of a pandemic, and how this information could be used to optimise the response. The regulatory environment is founded on a set of ethical principles, often unarticulated. Since there is likely to be some dispute between utilitarian philosophical approaches used in public health and deontological or virtue ethical approaches used in clinical medicine [78] , there is a need for some preparatory work with general practitioners clarifying ethics of clinical behaviour, restriction of liberty under quarantine orders, and resource allocation and distribution. In an established pandemic, it is likely that there will be shortfalls in the GP workforce, due to illness among GPs, caring duties or closure of small practices. Non-hospital clinical specialists, retired general practitioners, allied health professionals and medical students could be trained to fill the gap in services. Research is needed to define the clinical work that can be done by other health personnel in general practice, eligibility criteria and accreditation processes for this cadre of workers, and optimal training processes. All public health problems have a clinical dimension, and all clinical problems have a public health dimension. At present, the plans in the five countries provide more detail on the public health dimension of the pandemic. There are intercountry differences in the emphases provided to different domains of the general practice response. Some of this reflects the emphasis on particular elements contained within the relevant national plan. Some of the differences are due to the ways in which general practice is structured in a country, and the strengths of its linkages to other components of the health sector. There is an urgent need to incorporate general practice and the broader primary care sector into pandemic planning activities, and to undertake the preparedness activities that would make this sector, which provides the majority of health care work, a true partner in pandemic response. Figure S1 Jurisdictions or health management organizations whose plans were included in the study. Found at: doi:10.1371/journal.pone.0002269.s001 (0.04 MB DOC)

Apoptotic signals induce specific degradation of ribosomal RNA in yeast

Organisms exposed to reactive oxygen species, generated endogenously during respiration or by environmental conditions, undergo oxidative stress. Stress response can either repair the damage or activate one of the programmed cell death (PCD) mechanisms, for example apoptosis, and finally end in cell death. One striking characteristic, which accompanies apoptosis in both vertebrates and yeast, is a fragmentation of cellular DNA and mammalian apoptosis is often associated with degradation of different RNAs. We show that in yeast exposed to stimuli known to induce apoptosis, such as hydrogen peroxide, acetic acid, hyperosmotic stress and ageing, two large subunit ribosomal RNAs, 25S and 5.8S, became extensively degraded with accumulation of specific intermediates that differ slightly depending on cell death conditions. This process is most likely endonucleolytic, is correlated with stress response, and depends on the mitochondrial respiratory status: rRNA is less susceptible to degradation in respiring cells with functional defence against oxidative stress. In addition, RNA fragmentation is independent of two yeast apoptotic factors, metacaspase Yca1 and apoptosis-inducing factor Aif1, but it relies on the apoptotic chromatin condensation induced by histone H2B modifications. These data describe a novel phenotype for certain stress- and ageing-related PCD pathways in yeast.

Gene expression in all organisms is regulated at multiple levels, including transcription initiation, mRNA stability and turnover, translation and protein degradation. Not surprisingly, rapid changes in cell metabolism and most responses to environmental stimuli involve significant flux of many cellular RNAs, of which mRNA transcriptome profiles have been most extensively studied to date (1, 2) . However, also stable RNAs, such as ribosomal, transfer, nuclear and nucleolar RNAs (rRNAs, tRNAs, snRNAs and snoRNAs), are likely to undergo specific transformations in altered conditions. It has been demonstrated that certain pathways of cell death are accompanied by the destruction of nucleic acids. For example, in metazoans the programmed cell death (PCD) called apoptosis, in addition to irreversible DNA damage, which is considered an apoptotic hallmark (3) , also involves specific cleavage of several RNA species, including 28S rRNA, U1 snRNA or Ro RNP-associated Y RNAs (4) . It was proposed that rRNA degradation could contribute to cell autodestruction, whereas degradation of anti-apoptotic factors mRNAs would accelerate apoptosis. In higher eukaryotes, RNA cleavage is probably carried out by RNase L, a 2 0 5 0 -oligoadenylate-dependent endoribonuclease, which functions in RNA decay during the interferon-induced response to viral infection, and whose activation in animal cells causes apoptosis (5) (6) (7) . However, RNase L-independent cleavage of 28S rRNA in virus infected cells has also been reported (8) . The occurrence of apoptosis was assumed to be limited to metazoans, where elimination of single cells does not kill the whole organism. Nevertheless, recent studies revealed the existence of cell death pathway in yeast, Saccharomyces cerevisiae, and other unicellular eukaryotes, with typical hallmarks of apoptosis: DNA fragmentation, externalization of phosphatidyl serine and chromatin condensation. PCD in yeast is triggered by several different stimuli, including ageing, expression of mammalian pro-apoptotic proteins, exposure to low doses of H 2 O 2 , acetic acid, hyperosmotic stress and mating-type a-factor pheromone (9) (10) (11) . Unicellular organisms are believed to undergo PCD for a variety of reasons, including elimination of old, infected and damaged cells in growth-limiting conditions for better survival of the remaining population, and adaptation of the more fit subpopulation to the ever changing and challenging environment (12) . Orthologues of core regulators of mammalian apoptosis, such as the caspase-related protease Yca1, a homologue of mammalian pro-apoptotic mitochondrial serine protease HtrA2 (Nma111), a yeast EndoG nuclease Nuc1, apoptosis inducing factor (Aif1) involved in chromatin condensation, AIF-homologous mitochondrionassociated inducer of death (AMID) Ndi1 and an inhibitor of apoptosis (IAP) Bir1, are conserved in yeast (13) (14) (15) (16) (17) (18) . In addition, in both human and yeast cells, histone modifications (histone H2B phosphorylation at serine 14 in human and at serine 10 in yeast and H2B deacetylation at lysine 11 in yeast) play an important role in apoptotic chromatin condensation and cell death (19) (20) (21) . However, several apoptotic factors are missing in yeast, including the Bcl-2/Bax family and the apoptosis protease activator factor Apaf-1. Also, there is no good homologue of RNase L to execute possible RNA degradation. Nevertheless, yeast apoptosis has recently been shown to be activated in mRNA decay mutants (dcp and lsm) (22, 23) , supporting the notion that RNA metabolism and apoptosis are linked. PCD occurs via a number of different mechanisms, e.g. caspase-dependent or independent, however, in all eukaryotes it is thought to be correlated with high levels of reactive oxygen species (ROS). ROS can either be generated exogenously through respiration or originate from exogenous sources such as exposure to hydrogen peroxide (H 2 O 2 ), superoxide anions or hydroxyl radicals. Excessive ROS results in damage of cellular components (DNA, lipids and proteins), cell cycle arrest, ageing and finally cell death (24) . Cells have developed a complex network of defence mechanisms, both enzymatic and non-enzymatic, against adverse consequences of oxidative stress (25) . Non-enzymatic system comprises a set of small molecules acting as ROS scavengers (e.g. glutathione, thioredoxin, glutaredoxin and ascorbic acid), whereas enzymatic system eliminates oxygen radicals by the action of specialized cytosolic or mitochondrial enzymes (e.g. catalases, superoxide dismutases, glutathione peroxidases and thioredoxin peroxidases) (25) . Most genes encoding components of these systems are induced in response to oxidative stress and are under transcriptional control of specific factors, for example Yap1, Msn2/Msn4 and Skn7 in budding yeast (26) . However, it appears that there is no general oxidative stress response. In S. cerevisiae, different response pathways are triggered by specific oxidants and different genes are involved in maintaining efficient cellular resistance to various sources of ROS (1, 2) . Interestingly, recent genomic approaches to identify these genes showed that strains lacking proteins which function in RNA metabolism were oversensitive to oxidative stress (1,2). These included genes encoding rRNA helicases (Dbp3, Dbp7), rRNA processing factors (Nop12, Nsr1), mRNA deadenylases (Ccr4, Pop2) and several mitochondrial RNA splicing components (2, 27) . This indicates that RNA processing and degradation may have a role in cellular response to ROS. In this study, we examined the effects of elevated ROS levels generated by oxidative stress, ageing and other apoptotic-inducing treatments on the status of ribosomal RNA in yeast S. cerevisiae and we have shown that mature rRNAs become specifically fragmented as a result of the cell response to these conditions. RNA degradation coincides with fragmentation of chromosomal DNA but occurs considerably earlier and most likely upstream of the activation of major apoptotic regulators, Yca1 and Aif1. The existence of this mechanism underscores the role of gene expression, namely rRNA turnover, in regulating certain pathways of cell death, in this case most likely through destruction of ribosomes and subsequent inhibition of translation in the early stages of apoptosis. Yeast strains and plasmids used in this work are listed in Supplementary Table S1 . The transformation procedure was as described (28) . Strains were grown at 308C either in YPD, YPGal or YPGly medium (1% yeast extract, 2% Bacto-peptone, 2% glucose or 2% galactose or 2% glycerol, respectively) or in synthetic complete medium (SC, 0.67% yeast nitrogen base, 2% glucose or 2% galactose, supplemented with required amount of amino acids and nucleotide bases). Strains W303-1A-Bax, W303-1A-Bcl-x L , W303-rDNA were grown in SC media without leucine, tryptophan or uracil, respectively. Yeast cultures in early logarithmic phase (OD 600 $ 0.2) were stressed with H 2 O 2 (0.12-180 mM), menadione (0.05-0.6 mM), cumene hydroperoxide (CHP, 0.1-0.2 mM), tert-butyl hydroperoxide (t-BHP, 1-3 mM), paraquat (0.1-10 mM), diamide (0.2-2 mM) and linoleic acid hydroperoxide (LoaOOH, 0.05-0.2 mM) for 200 min (all reagents from Sigma). Treatment with acetic acid stress was performed as described (29, 30) . Cells were grown in SC media to early exponential stage, shifted to SC media, pH = 3, and treated with 175 mM acetic acid (Sigma) for 200 min. Hyperosmotic shock was achieved by growth of exponential cells in SC complete media containing 60% (wt/wt) glucose (Fluka) or 30% (wt/wt) sorbitol and 2% glucose (Sigma) for 2-8 h (31). Chronological ageing was performed by constant growth of yeast cultures in SC complete medium for 2-16 days (32) . For expression of murine Bax or Bcl-x L proteins, cells were grown to early exponential phase in SC-Leu or SC-Trp, respectively. Expression of murine Bax was induced by shifting the cells grown to early exponential phase in SC-Leu medium containing glucose to SC-Leu medium containing galactose for 6 h. Pre-treatment with ascorbic acid (20 mM, Fluka) and respiratory chain inhibitors oligomycin A (7.5 mg/ml, Sigma) and sodium azide (1.5 mM, Sigma) was performed for 30 or 60 min prior to treatment with H 2 O 2 . For inhibition of protein synthesis, cells were treated with 50 mg/ml or 200 mg/ml of cycloheximide (Sigma) for 30 min. Cell fixation was achieved by addition of formaldehyde (final concentration 1%) or EtOH (final concentration 70%) to exponentially growing yeasts and incubation in room temperature for 10 min or 60 min, respectively. Formaldehyde was quenched by addition of glycine to the final concentration of 0.4 M for 5 min. Following the removal of fixation agents, cells were exposed to H 2 O 2 for 200 min as described earlier. Preparation of samples and analysis of chromosomal DNA fragmentation by pulsed field gel electrophoresis (PFGE) was performed exactly as described (30) . PFGE was conducted in a CHEF-DRIII Chiller System (Bio-Rad). One percent agarose gels were run in 0.5% Tris borate-EDTA buffer at 148C with an angle of 1208 with a voltage of 6 V/cm and switch times of 60-90 s for 24 h. Gels stained in ethidium bromide were analysed after destaining using Syngene Gene Genius Bioimaging System. RNA extraction, northern hybridization and primer extension were essentially as described (33, 34) . Lowmolecular weight RNAs were separated on 6% acrylamide gels containing 7M urea and transferred to a Hybond N+ membrane by electrotransfer. High-molecular-weight RNAs were analysed on 1.2% agarose gels and transferred by capillary elution. Oligonucletides used for RNA hybridization and primer extension (W235 and W237) are listed in Supplementary Table S2 . Quantification of northern blots was performed using a Storm 860 PhosohorImager and ImageQuant software (Molecular Dynamics). Dideoxy-DNA sequencing was performed on PCR-templates prepared from genomic yeast DNA using the same primers as for primer extension (W235 and W237) and a fMolSeq kit (Promega) according to manufacturer's instructions. The 3 0 RACE assay was carried out on total RNA (20 mg) isolated from untreated cells and treated with 1mM H 2 O 2 . DNA 'adaptor' oligonucleotide (W242) carrying aminolinker at the 3 0 -end was ligated with the 3 0 -end of total RNA using 20U of T4 RNA ligase (NEB). Ligation was performed in the presence of 25% PEG1000 (Sigma) at 378C. RNA was extracted with phenol: chloroform: isoamyl alcohol (v/v 25:24:1), precipitated and used as a template for cDNA synthesis using W243 primer complementary to the anchor sequence and the Enhanced Avian HS RT-PCR Kit (Sigma) according to manufacturer's instructions. cDNA was amplified using primers W241 and W243, the resulting PCR product was gel purified, cloned into pGEM-T Easy vector and sequenced using primer W241. RNase H cleavage was performed essentially as described (35) . Samples of 10 mg of total RNA were annealed with 40 ng of oligonucleotide complementary to the specific regions within rRNA at 688C for 10 min and digested with 1.5 U RNase H at 308C for 1 h. For detection, samples were separated on polyacrylamide gels and analysed by northern hybridization using probes located upstream of RHase H cleavage. To examine the existence of the RNA degradation pathway in yeast under oxidative stress, we have performed treatments with low doses of oxidative agents generating different ROS. These included the inorganic H 2 O 2 (concentrations 0.4-3 mM), superoxide-generating menadione (concentrations 0.05-0.6 mM), paraquat (concentrations 0.1-10 mM), thiol oxidant diamide (concentrations 0.2-2 mM), organic CHP (concentrations 0.1-0.2 mM), t-BHP (concentrations 1-3 mM) and a LoaOOH (concentrations 0.05-0.2 mM). Such concentrations of oxidants result in 30-90% of cell death (2, 36, 37) . Total RNA from wild-type W303 or BY4741 cells grown to early exponential phase (OD 600 = 0.25) in YPD media and treated with chemical compounds for 200 min was separated on 1.2% denaturing agarose/formaldehyde gels and analysed by northern hybridization using a probe complementary to the 5 0 -end of mature 25S rRNA (starting at position +40). This RNA species was chosen in the first place, since effects on 28S rRNA have been observed in apoptotic mammalian cells (8, 38, 39) . On treatment with two oxidants, H 2 O 2 and menadione, extensive decay of the mature 25S and accumulation of specific degradation products was observed, whereas little or no degradation occurred for other chemicals tested ( Figure 1A , data shown only for W303 strain and treatment with H 2 O 2 , menadione, CHP and t-BHP). This indicates that RNA cleavage accompanies some oxidative stress pathways, as it is known that different oxidants elicit specific cellular responses that, though partly overlapping, induce different groups of genes and require individual sets of specialized defence functions to maintain resistance (1, 2, 24) . In addition to the 25S rRNA, other rRNA species were also probed for undergoing specific decay. After treatment with H 2 O 2 , RNA damage with accumulation of characteristic breakdown products occurred for 5.8S and, to a much lesser extent, for 18S, but not for 5S, (Figure 1B and C; data not shown). In the case of 18S, hardly any degradation intermediates were detected, there was some decay of the mature RNA, however, it was approximately 2.5 to 3-fold weaker than for the mature 25S. Therefore, we conclude that mainly the two components of the large ribosomal subunit, 25S and 5.8S, undergo specific apoptotic degradation. The oxidants utilized, except for H 2 O 2 , had not been tested for apoptotic effects in yeast. One of the most recognized apoptotic markers is fragmentation of chromosomal DNA. Internucleosomal DNA laddering, typical for mammalian apoptosis, has not been detected during PCD in yeast; nevertheless, a higher order chromatin fragmentation to segments of several hundred kilobases also occurs in yeast (30, 40) . This DNA breakdown can be monitored either by the terminal deoxynucleotidyl transferase dUTP nick-end labelling (TUNEL) assay or by using PFGE of genomic DNA. The latter approach was applied to verify which oxidative agents lead to apoptotic phenotypes. Chromosomal DNA from cells treated with H 2 O 2 (1 mM), menadione (0.5 mM), CHP (0.2 mM), paraquat (5 mM), diamid (2 mM) and t-BHP (1 mM) for 200 min was analysed using PFGE ( Figure 1D ). Clear DNA degradation was observed only for cells exposed to H 2 O 2 and menadione, other treatments did not result in a visible apoptotic fragmentation. This is in a striking agreement with rRNA degradation that occurred only in H 2 O 2 -and menadione-treated cells. This strongly indicates that rRNA decay phenotype can be related to apoptosis. To confirm this, we have examined other conditions known to provoke apoptosis in yeast, i.e. acetic acid, ageing and hyperosmotic shock ( Figure 1E -H). For treatment with acetic acid, cells were grown in SC complete medium (pH 3) to exponential phase and exposed to 175 mM acetic acid for up to 200 min (29,30). Hybridizations with probes against 25S (probe 007, position +40, lanes 1-6; probe W234, position +344, lanes 7-12; probe W236, position +600, lanes 13-18; probe W238, position +843, lanes 19-24; probe W239, position +2168, lanes 25-30 and probe W240, position +3323, lanes 31-36). Asterisks above the arrows indicate the products that were further analysed. Arrow marked with a hatch shows a band matching the potential 3 0 product of the major cleavage, 5 0 product is marked with one asterisk. (B-C) Primer extension analysis for two main cleavage sites in the 25S rRNA in W303 cells treated with 1 mM H 2 O 2 (A) and in 16-day old chronologically aged rho0 W303 cells (B). Primer extensions were performed using primers W235 for sites around positions +400 and +470 and W237 for sites around position +600 relative to the 5 0 end of the mature 25S. DNA sequencing on a PCR product encompassing the 5 0 end of the 25S from +40 to +701, using the same primers was run in parallel on 6% sequencing polyacrylamide gels (lanes 1-4). The sequences with primer extension stops are shown on the right. Secondary structures of the regions in the vicinity of the cleavages, indicated by arrowheads and shown beside corresponding primer extension reactions, were adapted from the website http://rna.icmb.utexas.edu/. (D-E) 3 0 ends of cleaved-off products for the major cleavage at positions +610-611 were mapped by 3 0 RACE. (D) PCR reactions on cDNA prepared using total RNA from untreated control (lane 1, C) and cells treated Hyperosmotic shock was achieved by growth of exponential cells in SC complete media supplemented with 60% (wt/wt) glucose or 30% (wt/wt) sorbitol (31) . And finally, chronological ageing was performed by constant growth of yeast cultures in SC complete medium for 2-16 days (32) . This analysis revealed that all apoptotic stimuli tested resulted in the 25S and 5.8 rRNA fragmentation with the degradation pattern specific for each condition (shown in Figure 1E -G for 25S in all apoptotic conditions and in Figure 1H for 5.8S during ageing). The accumulating intermediates generated by some factors were comparable (see for example, cleavages mediated by 60% glucose and 30% sorbitol, acetic acid and H 2 O 2 , Figure 1E and F); however, the general outcome of each treatment indicated differences in the course of events during each response. The occurrence of RNA degradation triggered by H 2 O 2 was monitored during a time course between 5 and 60 min and over a broad range of concentrations (0.12-180mM for 200 min) ( Figure 1E , lanes 6-9 and Figure 2A ). Degradation was initiated relatively fast, since it was apparent at 5 min for 0.6 mM H 2 O 2 ( Figure 1E , lanes 6-9), 15-30 min for 175 mM acetic acid ( Figure 1E , lanes 1-5), 2 h for 60% glucose and 30% sorbitol ( Figure 1F ) and 3 days for ageing ( Figure 1G ) following the treatment. This onset of rRNA degradation distinctly precedes the timing of DNA damage characterized in apoptotic yeast exposed to the same stimuli (30) , indicating that RNA decay process is activated early during the response. Also, in the case of H 2 O 2 , low doses of the oxidant, starting with 0.12 mM and optimal at 0.4-3 mM, were sufficient to initiate rRNA degradation with the appearance of specific bands. When high concentration of H 2 O 2 (180 mM), believed to result in cell necrosis, was used, these specific degradation products were absent; however, the level of mature 25S and 18S rRNAs was also significantly reduced ( Figure 2A ; data not shown). These data show that different ROS-generating treatments that lead to yeast apoptosis, namely H 2 O 2 and acetic acid, ageing and hyperosmotic shock, induce RNA fragmentation that most likely precedes the DNA damage and, as in higher eukaryotes, can be considered a hallmark of the induction of PCD in yeast. Cleavages in the 25S rRNA are endonucleolytic and require cellular machinery Specific cleavages within the 25S rRNA generated in the presence of hydrogen peroxide were monitored by northern hybridization with probes located along the molecule to narrow down the regions to be further analysed ( Figure 2A ). This analysis showed the accumulation of diverse degradation products, some of which extended from the 5 0 -end of the molecule (Figure 2A, lanes 1-6) , whereas others were also truncated at their 5 0 ends ( Figure 2A, lanes 25-30) . The striking decrease in the level of the mature 25S rRNA at higher doses of the oxidant (1-3 mM) indicates that following specific cleavages the majority of rRNA becomes degraded, possibly by the exosome complex of 3 0 !5 0 exonucleases that participates in the decay of rRNA precursors and excised transcribed spacers (41) . The emergence of the characteristic cut-off in the signal at the fragment size corresponding to the position of the probe indicated that major cleavage sites are located around positions +400, +600 and +900 with respect to the 5 0 -end of the molecule. Two of these cleavages, at positions +610-611 and +398-403, were mapped for treatment with 1 mM H 2 O 2 by primer extension using primers W237 and W235 situated downstream of the expected cleavage sites ( Figure 2B ). Similarly, major cleavage sites were analysed in 16-day old chronologically aged cells using the same primers ( Figure 2C ) and mapped at positions +601-602 and +478-501. According to the secondary structure of the 25S rRNA taken from (42), the regions where mapped cleavages occur (shown in Figure 2B and C besides corresponding primer extension reactions) are located at unpaired nucleotides in loops or bulges. This points to the action of single-stranded RNA nucleases. To establish the nature of the observed RNA fragmentation, 3 0 -ends of the products generated by the H 2 O 2mediated cleavage in the 25S at positions +610-611 and +398-403 were determined by the 3 0 RACE. To this end, DNA 'anchor' oligonucleotide (W242) was ligated with T4 RNA ligase to total RNA from untreated and treated W303 cells to prepare cDNA using a primer specific for the anchor (W243). This served as a template to amplify products containing required fragments using the same 3 0 primer and a 5 0 primer that covers the 5 0 -end of 25S RNA starting at position +50 (W241). The ensuing PCR fragments ( Figure 2D ) were cloned into pGEM-Teasy and sequenced. The results of 10 sequenced clones for the cleavage at +610-611 and 19 clones for the cleavage +398-403 are shown in Figure 2E . In the case of the major site (cuts at positions +610-611), this analysis confirms that the 5 0 and 3 0 ends of this degradation product overlap ( Figure 2E , lower panel), which is consistent with the endonucleolytic mechanism of the cleavage. Mapping the 5 0 and 3 0 ends at site +398-403 by primer extension and 3 0 RACE produced a different pattern: these ends do not match ideally but the products with 1 mM H 2 O 2 (lane 2). To generate cDNA, total RNA that had been ligated to an 'anchor' oligonucleotide (W242) with T4 RNA ligase, was reverse transcribed using a primer specific for the anchor (W243). This was followed by PCR reaction using the same 3 0 primer and the 5 0 primer starting at position +50 in the 25S rRNA (W241). Arrows indicate products corresponding to fragments cleaved at +398-404 (lower) and +610-611 (upper). PCR fragments were cloned into pGEM-Teasy and sequenced. (E) Sequences obtained by the 3 0 RACE analysis for fragments cleaved at site +398-403 (19 independent clones) and site +610-611 (10 independent clones). The corresponding regions of the 25S with cleavage sites mapped by primer extension and indicated with empty arrowheads are shown above in grey. Figures in parentheses show the number of identical clones. (F) Mapping 3 0 ends of two major cleavages sites using RNase H cleavage on total RNA extracted from wild-type, rrp41-1 and ski7D cells treated with 1mM H 2 O 2 (lanes, 2-4) and from wild-type untreated control (lane 1, C). RNase H treatment was performed on RNA samples annealed to DNA oligonucleotides W244 and W263 complementary to positions +271 and +510, respectively. Samples were separated on a 8% acrylamide gel and hybridized with probe W234 (F-I) and probe W264 (F-II) to detect 3 0 ends of fragments cleaved at +398-403 (F-I) and at +610-611 (F-II), respectively. Arrows show more defined 3 0 ends of products cleaved at +610-611 for all strains and at +398-403 in the mutants; vertical bar in F-I indicates heterogenous 3 0 ends of products cleaved at +398-403 in wild-type cells. get progressively shorter pointing at the action of 3 0 !5 0 exonucleases ( Figure 2E, upper panel) . The most likely candidate is the exosome, a large complex with a 3 0 !5 0 exonucleolytic activity involved in the processing and degradation of mRNA, rRNA and other RNA substrates (43) . Mutants in the exosome core component Rrp41, the nuclear subunit Rrp6 and the cytoplasmic cofactor Ski7, were used to assess the status of the product 3 0 ends by a specific RNase H cleavage. This cleavage, directed by a DNA-RNA hybrid between oligonucleotide W244 and a complementary region in the 25S starting at residue +271, allows higher resolution of analysed RNAs. This analysis shows that products generated at site +398-403 in the exosome mutants rrp41-1 and ski7D, but not in rrp6D, are 3 0 extended and less heterogenous than corresponding fragments in the wild-type strain (Figure 2F-I; data not shown). In contrast, positions of cleavages at site +610-611 are not affected by mutations in the exosome (Fig. 2F-II) . This suggests that the cytoplasmic exosome may contribute to rRNA decay by digesting 3 0 ends of at least some cleavage products. To ascertain that RNA degradation process is enzymatic and not chemically induced by various reactive compounds, yeast cells were fixed with 1% formaldehyde for 10 min or with 70% ethanol for 30 min prior to exposure to increasing concentrations of hydrogen peroxide ( Figure 3A and B) . Both fixation procedures preserve cellular structures, however, it is known that most fixatives have harmful consequences, e.g. cause some loss of cellular components, including ribosomes. Nevertheless, a similar approach had been used to demonstrate that DNA damage in apoptotic yeast cells was an enzymatic process (30) . Also, in the case of RNA, the appearance of specific H 2 O 2 -induced degradation products was prevented by fixation, although the overall level of rRNA was reduced. Some faster migrating RNA species were detected in formaldehyde or ethanol fixed cells, however, these were generated also in the absence of the oxidant and did not intensify after treatment ( Figure 3A and B, lane 4) . Finally, to check whether rRNA destruction during apoptotic response is not due to cessation of translation in dying cells, cells were treated for 200 min with the translation elongation inhibitor cycloheximide (200 mg/ml) and this did not lead to an apparent rRNA degradation (Supplementary Figure S1A) . This is also supported by our earlier observations that exposure of yeast cells to many oxidative agents that cause cell death does not result in rRNA decay ( Figure 1A) . Together, this strongly suggests that rRNA degradation observed in apoptotic and oxidative stress conditions is not simply a result of cell death but is produced in the process that requires enzymatic activity and functional cellular machinery. rRNA is most likely cleaved endonucleolytically and in some cases, dictated probably by the RNA structure, this is followed by exonucleolytic digestion by the exosome. rRNA degradation is strongly correlated with ROS levels and is connected with oxidative stress response and apoptosis pathways Treatment with oxidative agents and apoptotic stimuli generate elevated levels of ROS in the cell. To test whether there is a direct link between RNA fragmentation and ROS, a potent ROS scavenger, L-ascorbic acid (vitamin C), was used (44) . The presence of 10mM ascorbic acid prior to treatment with standard doses of H 2 O 2 almost totally abrogated degradation of the 25S rRNA ( Figure 4A ). In addition, the ectopic expression of murine Bcl-x L protein of the anti-apoptotic mammalian Bcl-2 family, known to have a protective effect against ROS in yeast (45, 46) , also strongly safeguarded the 25S rRNA from rapid degradation by H 2 O 2 ( Figure 4B ). In contrast, expression of the mammalian pro-apoptotic Bax protein that increases ROS level (36, 47) , additionally enhanced the degradation phenotype ( Figure 4C ). This confirms the direct correlation between the production of ROS and the fate of cellular nucleic acids, leading not only to DNA but also rRNA damage and destruction of ribosomes. From the data presented so far, it appears that the observed rRNA fragmentation may possibly represent a part of the cellular oxidative stress and apoptotic responses. This was assessed by testing the extent of the 25S rRNA degradation in different mutants defective in these pathways. In the first place, yca1D strain, lacking the only identified apoptotic metacaspase Yca1 in yeast, and aif1D cells not expressing the yeast apoptosis inducing factor Aif1 (13,15), were assayed for H 2 O 2 -induced RNA decay, however, no significant differences were observed (Supplementary Figure S1B and C) . Similarly, addition of a broad-range caspase inhibitor z-VAD-FMK (20 mM) that prevents Yca1-dependent cell death in yeast (13, 48) had no effect on rRNA fragmentation (data not shown). This indicates that rRNA degradation detected in all apoptotic conditions tested is independent of the two major apoptosis mediators, Yca1 and Aif1, and of other potential yeast caspases. Likewise, treatment with translation inhibitor cycloheximide, that has been shown to prevent apoptotic cell death induced by H 2 O 2 and acetic acid (29, 36) , had little or no effect on H 2 O 2 -mediated rRNA degradation (Supplementary Figure S1A) . However, it appears that events in the course of apoptosis that require protein synthesis are rather late, for example DNA fragmentation and chromatin condensation, whereas rRNA decay is initiated relatively fast. In contrast, different outcome was observed for mutants inhibiting chromatin condensation during H 2 O 2 -induced apoptosis. Phosphorylation of Serine 10 and deacetylation of Lysine 11, both in histone H2B, were reported to have an essential role for the progress of cell death in yeast (20, 21) . In agreement, S10A or K11Q mutations in histone H2B that prevent these modifications and abrogate apoptosis resulted in the significant reduction of the 25S rRNA degradation, both the decay of the mature rRNA and the amount of degradation products ( Figure 5A ). Together, these data show that the destruction of ribosomal RNA in cells treated with H 2 O 2 , and possibly with other apoptotic stimuli, is a part of a yeast cell death pathway that involves histone modification and not the caspase-dependent pathway. Remarkably, Nuc1-mediated apoptosis resulting from over-expression of yeast EndoG homologue Nuc1, a major mitochondrial nuclease, was also reported to be Yca1-and Aif1-independent and related to histone modifications (18) . Alternatively, it can be envisaged that damage of ribosomes triggered by apoptotic stimuli is an early event during the response, does not require protein synthesis, precedes caspase activation and acts as an upstream signal in the apoptotic pathway. This scenario is consistent with most observations so far. As the oxidative stress in yeast proceeds through multiple pathways that involve different response mechanisms (1,2), we tested several known enzymes and factors that regulate these responses. These included two major transcription factors, Yap1 and Skn7, that control expression of several genes induced by oxidative stress and in this way participate in ROS sensing (1, 26, (49) (50) (51) , as well as components of antioxidant pathways, e.g. superoxide dismutases Sod1-2, glutathione peroxidases Gpx1-3, glutaredoxins Grx1-2, peroxiredoxins Tsa1-2, Prx1, Dot5 and Ahp1, thioredoxins Trx1-2 and thioredoxin reductases Trr1-2 (24,52). These enzymes are required for protection against ROS either by catalysing the breakdown of oxidative compounds or by restoring natural intracellular redox equilibrium. In addition, as glutathion protects cells against ROS, we also used a gsh1D strain lacking a g-glutamylcysteine synthetase, which, when grown on glutathion-free synthetic medium, leads to glutathion depletion and cell death (36, 53) . Strains lacking these proteins are more sensitive to several oxidants than their isogenic wild-types (2), and following treatment with H 2 O 2 they showed a marked increase in the 25S rRNA degradation, however, to different degrees depending on the mutant. In Figure 5B -D, yap1D, skn7D, gpx1/2/3D, sod1/2D, grx1/2-D, prxD (tsa1/2D/prx1D/ ahp1D/dot5D) and gsh1D strains are shown, which gave the most evident effects in comparison with their respective isogenic wild-types, particularly when considering the decay rate of the mature 25S rRNA. These data indicate that properly functioning oxidative stress response also protects cellular components such as nucleic acids from the attack by ROS and that defects at any step of this defence result in a more severe and faster breakdown. It is noteworthy that multiple anti-oxidant mutants lacking all components of each enzymatic pathway exhibit a stronger effect on the 25S degradation than single mutants, pointing to the additive protection actions of these systems. Striking effects on rRNA stability in strains lacking stress response transcription factors Yap1 and Skn7 indicate that the synthesis of new anti-oxidant proteins that are induced by oxidative stress might be required for protection of ribosomes. Consistently, blocking protein synthesis by pre-treatment with cycloheximide resulted in somehow stronger rRNA degradation (Supplementary Figure S1A, lanes 1-5 and 11-15 ). Taken together, this indicates that targeting rRNA degradation during oxidative stress may directly contribute to cell death. To test whether the level of rRNA, which reflects the amount of cellular ribosomes, may be somehow linked with cell survival under oxidative stress, we attempted to create a situation where the steady-state level of mature rRNAs will be increased or decreased. However, additional copies of rDNA present on a multicopy pNOY102 plasmid under control of the inducible GAL7 promoter (54) did not affect the amount of any mature rRNA species, possibly due to mechanisms that regulate ribosome abundance (data not shown). In contrast, the level of total genomic and plasmid-derived 18S and 25S rRNAs was, to our surprise, reduced to 70% in a strain transformed with the multicopy pJV12 plasmid expressing a tagged rDNA gene under the control of the constitutive PGK1 promoter (55) , when compared to a strain transformed with vector alone ( Figure 6A ). The basis of this effect is unclear, particularly that it was seen even though the tagged rRNA versions were expressed as confirmed by northern blots using probes specific for plasmid borne 25S rRNA ( Figure 6A ). Nevertheless, the strain carrying pJV12 showed a decreased viability already in the absence of oxidant (1.6-fold) and even more strikingly reduced following treatment with different concentrations of H 2 O 2 (4.6-fold for 0.6 mM and 6.24 for for 1 mM, respectively) relative to the strain with vector alone ( Figure 6B ). In another approach, we used the NOY504 strain, which carries a temperature sensitive (ts) RNA polymerase I (56) . At 378C, this strain ceases to grow but it sustains slow growth at 308C due to reduced levels of mature rRNA. Expression of additional copies of rDNA from pNOY102 or pJV12 plasmids improves growth at all temperatures and rescues the ts-lethal phenotype (55, 56) . Growth of NOY504 expressing additional rDNA under the control of GAL7 (pNOY102) or PGK1 (pJV12) promoters resulted in total rRNA levels lower by 15% in the latter case ( Figure 6C ). This relatively modest difference in rRNA abundance led to a 10% decrease in survival of cells exposed to oxidative stress ( Figure 6D ). This indicates that there may exist a correlation between the quantity of ribosomal subunits and the capacity of the cell to elicit functional defence mechanisms and prevent cell death. It is possible that there is a feedback mechanism that controls this relationship: a healthy cell that contains an adequate number of ribosomes is able to respond more efficiently to stress stimuli to protect cell components from damage, including ribosomes themselves. Therefore, provided that the level of ribosomal RNA monitors cell fitness, its sudden reduction may act as one of the signals to initiate cell death mechanisms, including apoptosis. rRNA degradation depends on the mitochondrial activity in the cell Mitochondria are the major source of endogenous ROS generated by oxidative phosphorylation. The extent to which mitochondria are involved in mammalian or yeast apoptosis is still questionable, although it appears that mitochondrial ROS could be important in some signalling pathways (57,58) and have a central role in some apoptotic pathways and less crucial in others (59) . For example, apoptotic cell death in yeast induced by acetic acid, pheromone and Bax expression was shown to be mediated by mitochondria (31, 60) . The correlation between mitochondria and rRNA stability was assessed, in the first place, by checking rRNA level in respiratory-deficient rho0 cells lacking mtDNA in conditions inducing apoptosis, i.e. exposed to H 2 O 2 ( Figure 7A ), 175 mM acetic acid ( Figure 7B ), hyperosmotic stress (60% glucose, Figure 7C ) and during chronological ageing ( Figure 7D ). All treatments resulted in a remarkably robust degradation of the 25S and 5.8S rRNAs in rho0 strains when compared to the parental W303 and BY4741 strains (Figure 7 , Supplementary Figure S1E ; and data not shown). This points to the importance of the functioning mitochondria in the stressinduced rRNA degradation. To test the contribution of the oxidative phosphorylation, two mutants in these pathways were used, op1 with a point mutation in a major ADP/ATP carrier AAC2 (Arg96!His96), and a triple aac1/2/3Á deletion mutant (61) ( Figure 7E ). Both mutants behaved in a similar manner as rho0 cells and exhibited more pronounced rRNA degradation in the presence of H 2 O 2 than the isogenic wild-type; however, the phenotype was stronger for op1 than upon deletion of the three carrier proteins, possibly due to a dominant negative effect in the point mutant. In addition, treatment with F 0 -F 1 ATPase proton-pump inhibitors, oligomycin A (7.5 mg/ml) and sodium azide (NaN 3 , 1.5 mM) that block electron transfer and the synthesis of mitochondrial ATP (62-64), resulted in a moderate increase in the rRNA cleavage ( Figure 7F and Supplementary Figure S1D ). All these experiments indicate that the process of respiration, though generating the endogenous ROS, is also vital for counteracting its adverse effects. This was further supported by the degree of rRNA protection against oxidative damage caused by H 2 O 2 observed for yeast cells grown on different carbon sources, which are known to affect the level of respiration (65) . The most extensive RNA decay was observed in glucose, where a process called glucose repression discourages respiration. It was less pronounced in galactose and least of all in the nonfermentable source, glycerol, where mitochondrial respiration is forced ( Figure 7G) . These experiments directly correlate functional mitochondria and the process of respiration with the defence against oxidative stress triggered by H 2 O 2 , acetic acid, hyperosmosis and ageing that, among others, prevents destruction of cellular components, including rRNA. Several cellular responses are regulated at the translational level, particularly by selective translation of specific mRNAs or by inhibition of the ribosome and protein synthesis, as these processes consume a large amount of energy. Such inhibition can follow various stimuli, including endoplasmic reticulum stress and unfolded protein response (UPR), transition into quiescence and different stress-related and cell death-related signals (66). The most straightforward and fastest way to achieve translation inhibition is to target ribosomal RNA. Interestingly, it has been proposed that repression of protein synthesis during UPR in human cells is due to the cleavage of 28S rRNA by hIRE1b, a second homologue of IRE1 (67) . Also, during apoptosis in some mammalian cells degradation of 28S rRNA by RNaseL, but also of other RNAs such as Y RNA or some mRNAs, has been suggested to block protein synthesis that contributes to, but could even initiate, cell death. Furthermore, damage to the 28S rRNA may act as a ribotoxic stress and induce an early death-committing signal through activation of SAP and MAP kinases (68, 69) . These possibilities were not examined in yeast PCD pathways. We have analysed the behaviour of ribosomal RNAs during oxidative stress and in apoptotic pathways that are induced by different stimuli. We have observed that cells exposed to all apoptotic conditions tested, such as H 2 O 2 , acetic acid, hyperosmotic stress (60% glucose) and ageing, reveal a significant degradation of the 25S and some of 5.8S rRNAs, with a much lesser effect on the 18S rRNA. The decay of mature rRNAs was accompanied by the accumulation of treatment-specific, yet partly overlapping degradation intermediates. Although there is no evidence so far that rRNA damage during apoptotic conditions in yeast can directly initiate cell death by activating signalling pathways, it is tempting to speculate that this might be the case. Such signalling could be conveyed either by a critical decrease of the mature 25S rRNA or, alternatively, by the accumulation of degradation intermediates/products, which can function as signal molecules triggering a specific PCD pathway. In most cases, when ribosomal RNAs are depleted (e.g. in pre-rRNA processing mutants), rRNA degradation is conducted rapidly with no or little rRNA fragments detectable; however, lack of Lsm proteins has been reported to result in degradation of ribosomal RNAs with accumulation of unusual intermediates (70) . It is noteworthy that one apoptotic pathway that is linked with RNA metabolism is triggered by the defect in mRNA turnover caused by mutations in enzymes involved in 5 0 decapping, including components of Dcp1-2 and Lsm1-7 complexes (22) . Each of the applied apoptosis-inducing conditions resulted in a clear-cut pattern of the 25S degradation intermediates or products. This points to the endonucleolytic nature of the reactions, though exonucleolytic destruction, with certain RNA fragments temporarily protected by compact structures or tight interactions with proteins, cannot be excluded. However, mapping 5 0 and 3 0 boundaries of the major H 2 O 2 -induced cleavage product by primer extension and 3 0 RACE, respectively, confirmed that both ends strictly overlap and thus result from the endonucleolytic cut. It has been shown that apoptotic stimuli in yeast generate ROS that are closely correlated with the onset or progression of PCD (71) . We saw that, also the degree of rRNA decay corresponded to the cellular level of ROS, which was modified by using ROS scavenger (ascorbic acid) or ectopic expression of pro-or anti-apoptotic proteins (Bax and Bcl-x L , respectively) known to affect ROS generation. Also, rRNA degradation was more robust in oxidative stress defence mutants, both enzymatic and non-enzymatic, where intracellular ROS is not properly neutralized. All these observations argue that there is a direct link between ROS production and rRNA fragmentation. It can be envisaged that various reactive species themselves are able to produce endonucleolytic nicks in RNA molecules that will lead to breakdown, particularly as all mapped cleavages occur in singlestranded regions that constitute loops and bulges and are more accessible to chemical compounds in the solvent. However, this is not the case, given that specific RNA fragmentation did not occur in cells fixed with formaldehyde or ethanol prior to treatment with H 2 O 2 . In addition, oxidative agents used in this work generate different forms of ROS such as H 2 O 2 , hydroperoxide (LoaOOH, CHP and t-BHP), superoxide anion (menadione and paraquat) and hydroxyl free radical (produced from H 2 O 2 or superoxide anion). Although all were applied at toxic doses, only two of them, H 2 O 2 and menadione, mediated rRNA degradation, supporting the notion that oxidative compounds as such do not provoke cuts in RNA molecules within the cell. Taken together, this strongly suggests that active cellular machinery, such as signalling factors and RNA degrading enzyme(s), is required for this process. Nevertheless, these enzymatic activities are still to be identified. The closest yeast homologue of mammalian RNase L, Ire1, a sensor of the unfolded protein response, functions in the unconventional splicing of Hac1 pre-mRNA and contains protein kinase and endoribonuclease domains similar to those in RNase L (72) . Although our unpublished data show that deletion of Ire1 has no effect on the H 2 O 2 -induced rRNA degradation (M.S. and J.K.), it does not exclude its participation in other apoptotic pathways, for example those related to endoplasmic reticulum stress and unfolded protein response. We are currently testing several known yeast endo-an exonucleases for participation in apoptosis-related rRNA degradation. Preliminary observations suggest that, contrary to expectations, this function may involve not one but a number of unspecific nucleases, including mitochondrial Nuc1p, that act in a redundant fashion (M.S. and J.K., unpublished data). The reason why treatment with some oxidative agents but not others produce fragmented RNA is not entirely clear, however, it is known that different chemicals induce specific responses leading to expression of distinct sets of genes (1, 2) , and possibly only a few activate pathways that involve RNA destruction. Our results suggest that rRNA degradation phenotype most likely accompanies apoptosis, and only apoptosis-inducing oxidants result in rRNA decay, whereas cell death caused by others probably occurs via a different mechanism. Another question is why each apoptotic stimuli resulted in slightly different set of cleavage products. Those which were mapped for treatment with H 2 O 2 and in aging cells illustrate that cleavages often occur in loops and bulges in closely located regions within the 25S rRNA, or rather within the accessible RNA elements in the compact RNP structures. The difference in cleavage patterns could be due to stress-induced subtle or severe alterations in ribosome particles that change the local accessibility of the rRNA components presented for cleavage. Alternatively, if rRNA decay is indeed carried out by more than one nuclease, these differences may reflect varying enzyme specificities in each death-inducing condition. It is also possible that somehow different set of nucleases is activated or recruited to rRNA substrates during apoptosis triggered by oxidative stress, acetic acid treatment and ageing. Mitochondrial respiration protects rRNA against deleterious consequences of ROS Certain aspects of apoptosis, such as the change in mitochondrial membrane potential, fragmentation of mitochondria and the requirement of cyt c and AIF release to the cytoplasm, are strongly conserved among different organisms and point to the pivotal role of mitochondria. In yeast, PCD pathways triggered by acetic acid, Bax expression and pheromone, are strictly correlated with these events and do not proceed in cells devoid of mtDNA (rho0) (31, 60) . During chronological ageing and hyperosmotic shock, rho0 strains were reported to have somehow higher survival, which can be attributed to their long doubling time (1.5 to 2-fold), but they still die apoptotically (23, 73) . However, mitochondrial function is required for resistance to oxidative stress by way of detoxification or repair of the oxidative damage and, consequently, rho0 cells are more sensitive to several oxidants, have higher level of endogenous ROS and undergo apoptosis caused by H 2 O 2 or amino-acid starvation (61, 71, (74) (75) (76) (77) . Also, mammalian rho0 cell lines undergo apoptosis in response to some but not all cell death-activating stimuli. It has been postulated that these differences may arise from a distinct mechanism by which rho0 cells maintain membrane potential by way of ATP consumption (78) . Our results show that, in all conditions tested, rRNA degradation is tightly connected with mitochondrial function and the active process of respiration. Rho0 cells that are less resistant to oxidative stress suffer severe rRNA degradation in the presence of H 2 O 2 , acetic acid and 60% glucose and during chronological ageing. Also, impediment of oxidative phosphorylation by protonpump inhibitors or by mutations in ATP/ADP carriers gives a similar outcome. In contrast, rRNA is markedly more stable when mitochondrial respiration is enhanced (e.g. by growth on glycerol versus glucose). This is consistent with the protective role of active mitochondria against the damaging effects of ROS on cellular components, rRNA destruction included. To begin with, several anti-oxidant enzymes localize to mitochondria and are more abundant in respiring cells. In addition, it has been proposed that some anti-oxidant activities may require energy (75) . As some apoptotic pathways depend on the release of mitochondrial factors to the cytoplasm and are not induced in rho0 cells, this poses an important question regarding the link between rRNA degradation and apoptosis. rRNA degradation-apoptotic or not? Although several apoptotic mediators (Yca1, Aif1, Nma111, Bir1 and Ndi1) have been identified in yeast, their networking in regulation of apoptosis is not yet fully understood. They may interact and function in a similar fashion as their mammalian counterparts, however, in contrast to mitochondrial mammalian proteins, Nma111 and Bir1 are located in the nucleus, so in yeast there might be some deviations from the mammalian model (13, 14, 71) . Nevertheless, apoptotic cell-death pathways in yeast induced by H 2 O 2 , acetic acid, hyperosmotic shock, ageing and increased mRNA stability were reported to require metacaspase Yca1, nominating them as caspasedependent pathways (13, 23, 31) . Still, deletion of Yca1 in the mRNA turnover mutant lsm1D does not attenuate mRNA decay, placing Yca1 action downstream of the signal rising from mRNA level (23) . In contrast, histone phosphorylation at Ser10, which requires prior deacetylation at Lys11, has been shown to mediate H 2 O 2 -induced apoptosis independently of Yca1 (20, 21) . Also, the activity of the major mitochondrial nuclease, Nuc1, in the celldeath pathway does not require either apoptotic mediators, Yca1 or Aif1, but is affected by H2B modifications (18) . Moreover, two PCD pathways, namely induced by defects in protein N-glycosylation and triggered by ammonia in multicellular yeast colonies, do not rely on Yca1 but on as yet unknown caspase-like activity (79, 80) . Degradation of the 25S/5.8S rRNAs is observed in all conditions inducing apoptosis, however, this process is not dependent on Yca1 and Aif1. On the other hand, mutations in histone H2B that inhibit phosphorylation at Ser10 and block the progress of H 2 O 2 -induced apoptosis also severely affect rRNA degradation. More importantly, rRNA degradation coincides with apoptotic DNA fragmentation; from several compounds that lead to oxidative stress and cell death, only those that provoked DNA destruction also triggered rRNA decay. Furthermore, at least some apoptotic pathways strictly require the involvement of mitochondria and do not occur in yeast lacking mtDNA, whereas rRNA degradation in all stress conditions tested is more powerful in rho0 cells. To sum up: rRNA decay induced by apoptotic stimuli occurs during cell death pathway that involves ROS generation, mitochondrial activity, fragmentation of chromosomes and histone modification but not apoptotic regulators, Yca1 and Aif1. This could be due to the existence of numerous different but partly overlapping PCD mechanisms, whether caspase-and mitochondria-dependent or independent. Ever increasing numbers of such pathways has been described in the literature in recent years. However, we favour a different model, where all these elements function in concert at different steps of the whole scenario. Stress stimuli induce signals, possibly via ROS, affecting different levels of gene expression, namely chromatin modifications, transcription and translation, which as a result activate defence response. At this level, mitochondrial respiration helps to protect cellular components via adaptive mechanisms with anti-oxidant functions. Even so, when the attack is not successfully pacified, generated ROS molecules initiate the destruction of cellular machineries, targeting in the first place crucial elements such as protein synthesis (i.e. ribosomes). Now, the progression of the response comes to the crossroadsif the conditions are appropriate, the cascade of events leading to apoptosis is triggered (e.g. release of CytC and other mitochondrial factors to the cytoplasm) resulting in cell death with characteristic apoptotic markers. Alternatively, when apoptotic prerequisites are not met, cells do not enter this pathway. The more fit cells escape death, whereas others die anyway, maybe less rapidly and through a different pathway. In this scenario, certain events, including generation of ROS, histone modifications and possibly also RNA fragmentation, occur upstream of subsequent steps, such as activation of caspases and other apoptotic regulators with resulting apoptotic phenotypes.

Virus Adaptation by Manipulation of Host's Gene Expression

Viruses adapt to their hosts by evading defense mechanisms and taking over cellular metabolism for their own benefit. Alterations in cell metabolism as well as side-effects of antiviral responses contribute to symptoms development and virulence. Sometimes, a virus may spill over from its usual host species into a novel one, where usually will fail to successfully infect and further transmit to new host. However, in some cases, the virus transmits and persists after fixing beneficial mutations that allow for a better exploitation of the new host. This situation would represent a case for a new emerging virus. Here we report results from an evolution experiment in which a plant virus was allowed to infect and evolve on a naïve host. After 17 serial passages, the viral genome has accumulated only five changes, three of which were non-synonymous. An amino acid substitution in the viral VPg protein was responsible for the appearance of symptoms, whereas one substitution in the viral P3 protein the epistatically contributed to exacerbate severity. DNA microarray analyses show that the evolved and ancestral viruses affect the global patterns of host gene expression in radically different ways. A major difference is that genes involved in stress and pathogen response are not activated upon infection with the evolved virus, suggesting that selection has favored viral strategies to escape from host defenses.

One of the first consequences of organisms' adaptation to new environments is the manipulation of resources [1] [2] [3] [4] . In this sense, the interaction between intracellular parasites and their hosts represents a paradigm of resource manipulation. In general, a virulent relationship results in the alteration of many aspects of cellular metabolism and development, which are taken over in the parasite's own benefit [5] [6] [7] . Whether the relationship between a host and a parasite evolves towards a more or less virulent or benign situation depends on several genetic and ecological factors that may affect virus accumulation and transmission between hosts [5] . Of particular interest in the context of emerging infectious diseases is the characterization of changes in the pathogen's genome that are responsible for adaptation to a new host after spilling over from the original one and to understand how these changes may alter host's metabolic and regulatory interactions. High-density DNA microarrays offer an unparalleled view of the transcriptional events that underlie the host response to pathogens, providing a quantitative description of the behavior of tens of thousands of genes. In recent years, microarrays have been widely used to analyze the alteration of gene expression in host cells after infection with both animal [e.g., [8] [9] [10] [11] [12] [13] and plant [e.g., [14] [15] [16] [17] [18] viruses. However, a common drawback of these previous studies is that experiments were either done in cell cultures [8] [9] [10] [11] [12] [13] , which always represent an artificial and oversimplified environment, or using host-virus pairs whose previous evolutionary history of association is unknown and the degree of impact of abiotic environmental factor uncontrolled [14, 17] . Therefore, the relevance of these studies and, more importantly, their evolutionary implications for the problem of emergent infectious diseases, are rather limited. In the following, the results from an experiment simulating the emergence of a plant virus that crossed the species barrier and is horizontally spreading on a population of partially-susceptible hosts are reported. Evolutionary changes in viral genome and phenotypic properties and, more importantly, in the way it interacts with its host's transcriptome are the focus of the study. The pathosystem Tobacco etch potyvirus (TEV)-Arabidopsis thaliana ecotype Ler has been chosen for the present study. TEV genome is composed of a 9.5 kb positive polarity single-strand RNA that encodes a large ORF whose translation generates a polyprotein that is subsequently self-processed by virus-encoded proteases into 10 mature peptides [19, 20] . TEV has a moderately wide host range infecting around 149 species from 19 families [21] , although most of its natural hosts belong to the family Solanaceae. In these plants TEV induces stunting and mottling, necrotic etching and malformation in leafs [21] . A. thaliana ecotypes vary in their susceptibility to TEV. Some ecotypes (e.g., C24 and Ler) are fully susceptible [22, 23] whereas many other (e.g., Col-0 and Ws-2) do not allow for systemic movement but support replication and cellto-cell spread in inoculated leafs [22, 23] . Arabidopsis is a member of the family Brassicaceae, which belongs to a different order than the Solanaceae within the class Magnoliopsida [24] . Therefore, adaptation of TEV to A. thaliana represents a jump in host species at the taxonomic level of orders. TEV adaptation to A. thaliana: phenotypic changes The ancestral TEV was poorly adapted to A. thaliana Ler and infection concurred with the development of very mild symptoms ( Figure 1 ). Furthermore, 21 days post-inoculation (dpi), the viral load in infected plants, measured as the number of lesion-forming units (LFU) produced per milligram of tissue, was low, 48.3362.95 LFU/mg (6SEM), and the infectivity of the newly produced viral particles (i.e., the efficiency of initiating a new infection using a normalized amount of viral particles) was as low as 17.95% [95% confidence interval (CI): 7.54-33.53%]. Viral particles obtained from a single tobacco plant were used to initiate an evolution experiment in A. thaliana Ler plants. Seven independent lineages were founded. Each lineage consisted on 10 plants. Twenty-one dpi, positive infections were confirmed by Western blot hybridization using an anti-coat protein antibody (data not shown). One of the infected plants from each lineage was randomly chosen to be the source of viral particles for infecting the next batch of plants. This basic transfer protocol was serially repeated every three weeks. In six out of seven cases, lineages went to extinction as a consequence of decreases in viral loads beyond the threshold value that ensures efficient transmission. The only surviving lineage was maintained for 17 serial passages (hereafter TEV-At17). The viral load reached by TEV-At17 21 dpi, was 2138.386134.08 LFU/mg. In other words, TEV-At17 accumulation was ,44-fold larger than the value estimated for the ancestral TEV (two-sample t-test, t 43 = 15.58, P,0.0001). Not only more viral particles were produced per gram of infected tissue, but also the infectivity of TEV-At17 was 100% (95% CI: 77.91-100%) and significantly larger than for the ancestral TEV (Binomial test, 1-tailed P,0.0001). Furthermore, symptoms induced by TEV-At17 were more severe (Figure 1 ), including stunting, vein clearing and leaf deformation. TEV adaptation to A. thaliana: genotypic changes The above phenotypic changes have a correlate at the genetic level. Full-genome sequencing of TEV-At17 indicates that six changes have occurred during adaptation (first six rows in Table 1 ); three of them were non-synonymous. The first non-synonymous change, A1047V, affected the P3 protein. P3 localizes in nucleus and nucleoli in association with the NIa protein and participates in virus amplification through its interaction with the CI protein [20] . In other potyviruses, P3 is also involved in systemic movement [25, 26] . The second mutation is a T1210M replacement in the 6K1 peptide. This short peptide has been implicated in plant pathogenicity since its deletion results in symptomless infections [20] . Finally, the third amino acid replacement observed is L2013F in the VPg domain of the NIa protein. VPg is covalently attached to the 59 end of the viral RNA and has essential functions in the viral replication and, relevant for the problem in hand, it has been reported as a key determinant in host-genotype specificity for systemic movement or replication [20] and it has been recently demonstrated that the proper interaction between the translation initiation factor eI4B and VPg is necessary for TEV infection [27] . In conclusion, these three mutations may explain the observed improvement in virus amplification and pathogenicity. The relevance of the three synonymous substitutions observed is not as clear, although their adaptive value cannot be ruled out. To further characterize the relationship between these changes and symptoms severity, we introduced them by site-directed mutagenesis in the ancestral TEV clone. In addition, all three possible pairs of non-synonymous mutations and the triple non- synonymous mutant were also created. A. thaliana Ler plants were inoculated with these nine mutant clones and maintained in the same growth conditions for three weeks. The results of this experiment are summarized in Table 1 . All mutant genotypes were viable and replicated and accumulated in the plants, as confirmed by Western blot analysis (data not shown). Among the three single mutants, the only clone that produced visible symptoms was the one containing the L2013F allele in VPg. These symptoms were, nonetheless, qualitatively milder than those produced by TEV-At17 ( Figure S1 ). Concerning the three double mutants, only the combination of VPg and P3 substitutions induced symptoms that were qualitatively more severe than those produced by the single VPg L2013F mutant (Table 1 ) and almost as severe as those observed for TEV-At17. By contrast, mutation 6K1 T1210M does not have any effect on aggravating the symptoms associated with VPg L2013F. The combination of substitutions in P3 and 6K1 did not produce any symptom. Finally, the triple mutant recreated the strong symptoms characteristic of TEV-At17 (Table 1 and Figure S1 ). All together, these results suggest that the presence of substitution L2013F in the VPg protein is enough for triggering symptoms and that the severity of these symptoms is enhanced by the presence of substitution A1047V in P3, suggesting an epistatic interaction between these two mutations. The role of substitution T1210M in the 6K1 peptide remains unclear. It has been recently reported that the correct interaction between potyvirus' VPg and host's eIF4E is required to initiate a successful infection [27] . Recessive resistance of peppers to potyvirus depends on the substitution of relevant amino acid residues in eIF4E that disrupt the normal interaction between this translation factor and VPg. Resistance-breaking viral strains restore the normal interaction [27] . Therefore, we can hypothesize that TEV-AT17 has enhanced its ability to infect A. thaliana Ler by improving the interaction of its VPg with the host's translation initiation factor eIF4E. Differential effect of evolved and ancestral viruses on the overall pattern of host gene expression Next, we sought to unravel what component of the plant gene interaction networks and metabolic pathways have been targeted by the virus during its adaptation to A. thaliana Ler. Our goal is not to identify single genes but rather global transcriptomic changes. Long-oligonucleotide microarrays representing almost all genes in A. thaliana genome have been used to this end. Five replicates were analyzed per experimental treatment (control mock-inoculated plants, and plants infected with TEV and TEV-At17) using a global reference experimental design. After quality analysis, a total of 13,722 spots, corresponding to 12,180 genes, were considered as valid for further analyses (Table S1 ). Data were normalized to the median expression of non-infected plants, and thus they reflect biological differences in gene expression in each sample analyzed. Statistical analysis allowed identification of genes whose expression responded differentially upon infection with either TEV or TEV-At17 ( Figure 2 ). When comparing global patterns of gene expression in plants infected with ancestral and adapted viruses, 496 genes showed higher expression and 1,322 genes lower expression in TEV-At17 infections than in TEV infections ( Figure 2 and Tables S2 and S3); which represents 2.7 times more down-regulated than up-regulated genes (Binomial test, P,0.0001). Differentially expressed genes were grouped according to selforganizing maps (SOM) (Figure 3 and Table S4 ). Three global patterns of gene expression were observed among genes that were up-regulated by TEV-At17 infection ( Figure 3A ). The first pattern (SOMs A1 plus A2) corresponds to 130 genes whose expression was activated by both viruses but the magnitude of expression was magnified by TEV-At17. Genes belonging to this category include the pathogenesis-related protein PR1, which is well known to be a marker for the activation of salicylic acid-dependent defenses, such as the systemic acquired resistance (SAR) pathway [28, 29] . The second pattern (SOM A3) corresponds with 141 genes that were down-regulated after infection with TEV but showed expression levels similar to uninfected plants when infected with TEV-At17. The third pattern (SOM A4) corresponds to 234 genes whose expression was not significantly affected by TEV infection but show increased expression after infection with TEV-At17. Three distinct patterns were also observed among genes downregulated after infection with TEV-At17 relative to the infection with TEV ( Figure 3B ). The first pattern (SOMs B1 plus B2) represents 683 genes that were over-expressed by plants infected with TEV but infection with TEV-At17 resulted in expression levels similar to those observed in uninfected plants. Interestingly, proteins related with disease response such as PR5 and several other PR-like proteins as well as four proteins of the TIR-NBS-LRR class [29, 30] belong to this category. The second pattern (SOM B3) includes 196 genes that were down-regulated after infection with both ancestral and evolved viruses, although the magnitude of down-regulation was larger for TEV-At17. Finally, the third pattern (SOM B4) corresponds to 456 genes whose expression was not affected by TEV but showed lower expression when TEV-At17 infected the plants. The expression of transcription factors (TF) was also differentially affected by TEV and TEV-At17. Table S5 shows the list of differentially up-and down-regulated TF in plants infected with each type of virus. Fifty-one TFs, belonging to 20 families, were upregulated whereas 84 TFs, from 27 families, were down-regulated, including 13 ethylene-responsive binding factors (ERF), after infection with TEV-At. ERFs are linked to stress responses [31] and delays in ERF induction had been described in A. thaliana plants infected with virulent strains of the bacteria Pseudomonas syringae when compared with avirulent strains of the same bacteria [31] . Viral adaptation by avoidance of plant defenses Next, we examined the distribution of genes involved in related biological processes that are differentially affected by TEV and TEV-At17 (i.e., gene ontologies (GO) categories [32] ). The algorithm implemented in FatiGO [33] was applied to the nonredundant gene list grouped in each SOM (results are shown in Table S6 ). Only a significant differential category, response to salt stress, was identified for the SOM A3 of up-regulated genes shown in Figure 3A . By contrast, a large number of GO terms show significant over-and under-representation in the differentially down-regulated genes ( Figure 3B ). Table 2 shows the nonredundant functional categories that correspond to SOMs B1 plus B2 (i.e., genes over-expressed after infection by TEV but not differing from uninfected plants when infected with TEV-At17). Interestingly, significantly over-represented genes belong to functional categories which are related to plant responses to different abiotic (wounding, light intensity, temperature, salinity) and biotic stresses. Furthermore, genes involved in the SAR and in the activation of innate immune responses [29] were not expressed on plants infected with TEV-At17 while they were over-expressed on plants infected with the ancestral TEV, suggesting that the evolved virus acquired the ability to evade certain plant defense mechanisms, perhaps explaining the observed increase of viral load. Genes involved in basic cellular processes such as nucleic acid metabolism, translation and proteolysis were under-represented among down-regulated genes in SOMs B1 plus B2 ( Table 2 ), suggesting that the plant may be compensating for the consumption of these resources by an increased viral replication. A single significant GO category was also found in SOM B3 of down-regulated genes ( Figure 3B ), that is, gene expression was repressed in presence of both viruses but to a larger extent when TEV-At17 was infecting plants. Genes involved in response to auxin were under-expressed to a larger extent by plants infected with TEV-At17 than with TEV. We have shown that adaptation of a virus to a new host occurs by few changes in viral genome. The increase in viral fitness correlates with deep changes in the patterns of host's gene expression, illustrating that the subtle but dynamic interplay between the pathogen and the plant shifts as the virus adapts to its host. Under the experimental conditions imposed, it may be speculated that natural selection may had favored viral genomes that avoided plant defense mechanisms as suggested by the observation of stress-related genes not being activated after infection with the evolved virus (Table 2 ). Therefore, perhaps as a consequence, increases in the strength of symptoms, virus accumulation and transmissibility have been observed. These phenotypic changes are associated to a few genomic changes fixed in the viral genome. In particular, the development of symptoms is associated to a single substitution in the viral VPg protein, whereas ulterior mutations in other viral components simply magnify symptoms. Our starting hypothesis was that viral adaptation occurs throughout the integration of viral replication processes within host physiology and circuitry of genetic and metabolic interactions. Necessarily, this integration has to affect the patterns of host's gene expression. Our experiments directly test this hypothesis, supporting its validity and, furthermore, pinpointing some physiological processes that may be targeted by the virus as it improves its fitness. The obvious follow-up of this study is to dissect the physiological processes and identify, whenever possible, the precise steps and proteins that are getting targeted by the virus during its adaptation. Serial-passage experiments simulating horizontal transmission are well known to produce increases on parasite's virulence due to enhanced within-host competition among pathogenic strains, the decoupling between intra-host growth rate and transmission rate, and the lack of evolutionary innovation in the host [34] . The outcome of a different experimental design in which transmission would be vertical, and hence making high virulence detrimental, or in which virus and host are engaged in a coevolutionary armsrace may produce different results; perhaps with the evolution of a less severe virus and different alterations in plant gene expression. Finally, the findings here reported call for extra precaution when analyzing data from microarray experiments seeking for the effect of pathogen's infection on host gene expression: the pathogen effect on host's transcriptomic profiles would depend on the degree of adaptation of the pathogen to the host and to environmental conditions. Therefore, the only fully meaningful studies would be those in which pathogens and their experimental hosts would have an evolutionary history of association in the experimental growth conditions, whereas results from studies in which hosts are infected with naïve pathogens or the effect of environmental variables on pathogen's growth remain uncontrolled would be of very limited interest. An infectious clone pTEV-7DA [35] (GeneBank DQ986288), kindly provided by Prof. J.C. Carrington (Oregon State Univ.) was used as ancestor virus. This infectious clone contains a full-length cDNA of TEV and a 44 nt long poly-T tail followed by a BglII restriction site cloned into the pGEM-4 vector downstream of the SP6 promoter. 59 capped infectious RNA was obtained upon transcription of BglII-digested pTEV-7DA using SP6 mMES-SAGE mMACHINE kit (Ambion). A stock of ancestral TEV viral particles was generated as follows. Five mg of RNA transcripts were rub-inoculated into the third true leaf of four-week old Nicotiana tabacum var Xanthi plants. Afterwards, plants were maintained in the green house at 25uC and 16 h light photoperiod. Seven dpi, virions were purified as described elsewhere [36] , aliquoted and stored at 280uC. The viral load reached by replicating TEV populations in A. thaliana was estimated by the dilution-inoculation assay method on the local-lesion host Chenopodium quinoa [37] . In short, 2 g of tissue was ground in 1 mL of 0.5 M phosphate buffer. Four different leafs from each one of three different 4-week-old C. quinoa plants were rub-inoculated with 5 mL of undiluted, 5-and 10-fold diluted virus, respectively; 100 mg/mL carborundum were added to facilitate inoculation. Nine dpi, the number of local lesions was recorded and transformed into viral infectious loads (LFU/mg) by estimating the intercept of the regression line of the observed number of lesions on the dilution factor. A. thaliana Ler seeds were obtained from Lehle Seeds (cat. # WT-04 18 01). Seven independent evolution lineages of TEV were maintained by serial passages until extinction or up to 17 passages. All evolving lineages were initiated from the ancestral TEV stock population. Therefore, initial viral genetic variation among inoculated A. thaliana plants was minimal. To maximize transmission success, 10 plants were inoculated per lineage. Plants were inoculated between growth stages 3.5 and 3.7 [38] . Plants were maintained at 25uC and 16 h light photoperiod. Successful infections were confirmed by Western blot hybridization analysis 21 dpi using commercial antibodies anti-coat protein conjugated with horseradish peroxi-dase (Agdia). One gram of leaf tissue from a randomly-chosen infected plant per lineage were carefully ground in 1 mL 0.5 M phosphate buffer (pH = 8.0) and used to inoculate the next batch of 10 plants. Plants were always inoculated with similar viral doses. The consensus full-genome sequence of TEV-At17 was obtained following standard methods. In short, RNA was extracted using the RNeasyH Plant Mini kit (Quiagen), it was reverse-transcribed using MMuLV polymerase (Fermentas) and PCR amplified with Taq polymerase (Roche). The ABI Prism Big Dye Terminator Cycle Sequencing Kit 3.1 (Applied Biosystems) was used for cycle sequencing with fluorescently labeled dideoxynucleotides. Cycle sequencing reactions were carried out on a GeneAmp PCR System 9700 thermal cycler (Applied Biosystems). Labeled products were resolved in an ABI 3100 Genetic Analyzer (Applied Biosystems). Seven pairs of specific primers were used to amplify the 9.5 kb of TEV genome. The resulting fragments were overlapping, facilitating the task of fragment sequence assembly. Sequences were processed and analyzed with the STADEN 1.4b1. The 59-and 39-ends were sequenced by the RACE-PCR method [39] . The seven mutant genotypes created in this study were generated by site-directed mutagenesis using the QuikchangeH II XL kit (Stratagene) and following the indications of the manufacturer. Mutagenic primers were also designed according to Stratagene recommendations. To minimize unwanted errors during the mutagenesis process, the kit incorporates the PfuUltra TM high fidelity DNA polymerase. The presence of the desired mutation was confirmed by sequencing. To assess the presence of undesired mutations on each clone, the Surveyor TM Mutation Detection Kit Standard Gel Electrophoresis (Transgenomic) was employed. All six mutant genotypes presented the expected genome-wide band pattern. Total RNA was extracted from control and infected plants and used in an amplification reaction with the MessageAmp II aRNA Amplification kit (Ambion) following manufacturer's instructions. Five replicates for each sample category were generated, and compared with a global reference, generated from an equimolar mix of amplified RNAs from each of the 15 plants. RNA from each individual sample, plus the reference, were amplified, and used for labeling. For each category, three samples were labeled with Cy5 and two with Cy3, and compared with the corresponding reversed-labeled reference mix. Long 70-mers oligonucleotide microarrays, provided by Dr. D. Galbraith (Univ. Arizona), contain 29,110 probes from the Qiagen-Operon Arabidopsis Genome Array Ready Oligo Set (AROS) Version 3.0. This oligo set represents 26,173 protein-coding genes, 28,964 protein-coding gene transcripts and 87 miRNAs and is based on the ATH1 release 5.0 of the TIGR Arabidopsis genome annotation database (www.tigr.org/tdb/e2k1/ath1/) and release 4.0 of the miRNA Registry at the Sanger Institute (www.sanger.ac.uk/Software/ Rfam/mirna/index.shtml). Further information can be found at the Operon website (omad.operon.com/download/index.php). Oligos were rehydrated and immobilized by UV irradiation. Slides were then washed twice in 0.1% SDS, 4 times in water, dipped in 96% ethanol for 1 min, and dried by centrifugation. Slides were prehybridized 30 min at 42uC with 100 mL of 66 SSC, 1% BSA and 0.5% SDS, under a 60622 mm coverslip LifterSlip (Erie Scientific) in an ArrayIt microarray hybridization cassette (TeleChem). Slides were then rinsed five times in H 2 O and dried by centrifugation. Slides were hybridized immediately. Labeled RNA was used to hybridize the slides basically as described in [40] . After hybridization and wash, slides were scanned at 532 nm for the Cy3 and 635 nm for the Cy5 dyes, with a GenePix 4000B scanner (Axon Molecular Devices), at 10 nm resolution and 100% laser power. Photomultiplier tube voltages were adjusted to equal the overall signal intensity for each channel, to increase signal-to-noise ratio, and to reduce number of spots with saturated pixels. Spot intensities were quantified using GenePix Pro 6.0 (Axon Molecular Devices). Microarray raw data were deposited in the NCBI's GEO database under accession GSE11088. Spots with a net intensity in both channels lower than the median signal background plus twice standard deviations were removed as low signal spots. Data were normalized by median global intensity and with LOWESS correction [41] using the Acuity 4.0 software (Axon Molecular Devices). Finally, only probes for which a valid data was obtained in at least 13 out of the 15 slides were considered for further analysis (13,722 spots; Table S1 ). Median, mean and standard deviations were calculated from each treatment (control, TEV-and TEV-At17-infected plants), and all data were normalized to the median of the expression in control samples. To detect differentially expressed genes in plants infected with TEV-At17 compared to TEV, data were analyzed with the SAM package [42] , using two-class comparison (TEV versus TEV-At17) with a false discovery rate (FDR) of 5.38% with no fold-change cut-off. Differentially over-and under-expressed genes were grouped in 262 self-organizing maps (SOMs) [43] using Acuity with Euclidean squared similarity metrics. Gene lists were further analyzed with FatiGO [33] to find differential distributions of gene ontology (GO) terms between statistically differential genes in each SOM and the rest of genes in the microarray, with P values adjusted after correcting for multiple testing [33] . SAM analysis at 1% FDR gave qualitatively identical results, confirming their robustness to changes in arbitrarily-chosen statistical thresholds. Figure S1 Representative plants showing the symptoms induced by several of the viral genotypes described in Table 1 Table S4 SOM clustering of significant genes, both up-and down-regulated, between TEV-and TEV-At17-infected plants. Genes belonging to each of the eight SOMs in Figure 3 are listed on different spreadsheets, along with their annotation and mean expression data in control and in TEV-and TEV-At17-infected plants. Found at: doi:10.1371/journal.pone.0002397.s005 (0.48 MB XLS) Table S5 Transcription factors differentially expressed after infection with TEV and TEV-At17. A. thaliana transcription factors and other transcription regulators were mainly downloaded from arabidopsis.med.ohio-state.edu/AtTFDB/index.jsp, and collapsed with the significantly up-regulated (Table S2 ) and downregulated (Table S3) Table S6 Differential GO categories among differential genes grouped by SOM. FatiGO analysis was carried out for each SOM in Figure 3 . Differential categories were identified for downregulated genes in SOMs B1 plus B2 and B3 ( Figure 3B ) and in up-regulated genes in SOM A3 ( Figure 3A ). List1 includes the differential genes (gene name, number and percentage) belonging to each GO category, while List2 include the rest of genes in the same GO category represented in the microarray. Unadjusted and adjusted P values after correcting for multiple-tests are also indicated. Found at: doi:10.1371/journal.pone.0002397.s007 (0.07 MB XLS)

Deletion of human metapneumovirus M2-2 increases mutation frequency and attenuates growth in hamsters

BACKGROUND: Human metapneumovirus (hMPV) infection can cause acute lower respiratory tract illness in infants, the immunocompromised, and the elderly. Currently there are no licensed preventative measures for hMPV infections. Using a variant of hMPV/NL/1/00 that does not require trypsin supplementation for growth in tissue culture, we deleted the M2-2 gene and evaluated the replication of rhMPV/ΔM2-2 virus in vitro and in vivo. RESULTS: In vitro studies showed that the ablation of M2-2 increased the propensity for insertion of U nucleotides in poly-U tracts of the genomic RNA. In addition, viral transcription was up-regulated although the level of genomic RNA remained comparable to rhMPV. Thus, deletion of M2-2 alters the ratio between hMPV genome copies and transcripts. In vivo, rhMPV/ΔM2-2 was attenuated compared to rhMPV in the lungs and nasal turbinates of hamsters. Hamsters immunized with one dose of rhMPV/ΔM2-2 were protected from challenge with 10(6 )PFU of wild type (wt) hMPV/NL/1/00. CONCLUSION: Our results suggest that hMPV M2-2 alters regulation of transcription and influences the fidelity of the polymerase complex during viral genome replication. In the hamster model, rhMPVΔM2-2 is attenuated and protective suggesting that deletion of M2-2 may result in a potential live vaccine candidate. A more thorough knowledge of the hMPV polymerase complex and the role of M2-2 during hMPV replication are being studied as we develop a potential live hMPV vaccine candidate that lacks M2-2 expression.

Human metapneumovirus (hMPV) infection can cause acute respiratory illness in young infants, the immunocompromised, and the elderly [1] [2] [3] . HMPV infection has been detected in 4 to 15% of pediatric patients hospitalized with acute lower respiratory infections [4] [5] [6] [7] [8] [9] [10] . Cur-rently there are no licensed measures to prevent hMPV disease. Based on analyses of genomic sequences hMPV has been assigned to the metapneumovirus genus of the pneumovirus subfamily within the paramyxovirus family [11, 12] . The genome contains 8 transcription units with at least 9 open reading frames (ORFs) that encode a nucleocapsid protein (N), a matrix protein (M), a phosphoprotein (P) that likely associates with the polymerase complex, a fusion glycoprotein (F), an attachment glycoprotein (G), a large polymerase protein (L), a small hydrophobic protein (SH), and two proteins (M2-1 and M2-2) encoded by overlapping ORFs in the M2 gene. Among paramyxoviruses, SH is found in rubulaviruses and pneumoviruses, while M2 is found only in pneumoviruses. The functions of M2 proteins have not been studied extensively. Mutants of hMPV have been constructed by deleting M2-1, M2-2, SH or G, either individually or in combination, using the CAN97-83 isolate of hMPV, which requires trypsin for growth in cell culture [13, 14] . Recombinant hMPV lacking either M2-2 or G were attenuated and immunogenic in African green monkeys and have been proposed as promising vaccine candidates [15] . Such a suitably attenuated live hMPV is desirable because it would deliver the nearly complete set of viral antigens and closely mimic a natural hMPV infection. To construct a rhMPVΔM2-2 virus that can replicate efficiently in Vero cells without trypsin supplementation, we engineered the M2-2 deletion in a different subtype A hMPV strain. This recombinant strain is derived from hMPV/NL/1/00, and contains F 2 /F 1 cleavage-enhancing mutations in the F gene, a property that could facilitate the testing and manufacture of potential live hMPV vaccine candidates [16, 17] . The impact of the physical deletion of M2-2 on hMPV replication, and genetic stability in tissue culture were evaluated. rhMPV/ΔM2-2 exhibited somewhat restricted replication in Vero cells, but was significantly attenuated in hamsters. Hamsters immunized with rhMPV/ΔM2-2 were protected from experimental challenge with wthMPV/NL/1/00. The deletion of M2-2 resulted in higher levels of viral mRNA transcripts in tissue culture, giving rise to aberrant ratios of genomic RNA to viral transcripts. In addition, previously unreported genetic instability was observed, resulting in a higher frequency of point mutations and random insertions of U nucleotides in poly-U tracts of the rhMPV/ΔM2-2 genomic RNA. Recombinant hMPV harboring a deletion in M2-2 gene was recovered from rhMPV/ΔM2-2 cDNA. The M2-2 deletion was designed to preserve the native ORF of M2-1, which overlaps the M2-2 ORF by 51 nucleotides. The first 21 amino acids of the putative M2-2 protein and the entire M2/SH non-coding region (NCR) were maintained ( Figure 1A ). Recombinant rhMPV/ΔM2-2 was efficiently recovered. RT-PCR was performed on the recovered rhMVP/ΔM2-2 virus to confirm the presence of the M2-2 deletion. In Vero cells, rhMPV/ΔM2-2 plaques were less than 50% the size of rhMPV plaques (Figure 2A) . A 4-day multi-cycle growth curve was performed in Vero cells, a cell-line used for production of live vaccines, to compare the replication kinetics of rhMPV/ΔM2-2 and rhMPV. Data for the replication curves of these viruses were collected from three independently performed infections. The peak titer of rhMPV/ΔM2-2 in Vero cells was 7.22 +/-0.16 log 10 PFU/ ml, which was not significantly lower than the 7.52 +/-0.29 log 10 PFU/ml titer achieved by rhMPV ( Figure 2B ). However, the plaque size of rhMPV/ΔM2-2 was markedly diminished compared to rhMPV. Thus, hMPV M2-2 is dispensable for replication in Vero cells. During the preparation of viral stocks, we noted several mutations in rhMPV/ΔM2-2. To further assess the genetic stability of rhMPV/ΔM2-2, one-step RT-PCR was performed on a virus stock that was serially passaged 4 times in Vero cells. Sequence analysis of an RT-PCR product spanning the M2 and SH genes (nt4536 to nt6205) revealed nucleotide polymorphisms in several poly A tracts (sense direction) in the M2-1 and SH genes. Figure 3A shows a representative chromatogram of the sequence of an RT-PCR product generated from a rhMPV/ΔM2-2 virus stock. The wild-type sequence AGAGAAACTGA 6 TT is shown overlapping another sequence containing an inserted A in the poly A 6 tract. Three independently derived virus stocks of rhMPV/ΔM2-2 had major subpopulations with inserted A nucleotides (nts) at nt5060, nt5166 or nt5222 in M2-1, each of which would cause a premature translation termination in the M2-1 ORF. (See figure 3D for numbering of A insertions). Subpopulations with inserted A's were also detected at nt5551 or nt5572 in SH that would result in premature translation termination in the SH ORF. To compare the frequency of inserted A nucleotides in rhMPV/ΔM2-2 to that in rhMPV, RT-PCR products spanning nt4536 in F to nt5623 in SH were obtained from a passage 4 virus stock of rhMPV/ΔM2-2 or rhMPV. For this study, both positive sense and negative sense RNA were amplified using a one-step RT-PCR reaction. 1 kb RT-PCR fragments were inserted into pCR2.1 plasmids and 15 independent plasmids were sequenced. Surprisingly, 14 of the 15 (93%) cloned RT-PCR products of rhMPV/ΔM2-2 had an inserted A nucleotide at nt5060, nt5213 or nt5222 in the M2-1 gene ( Figure 3B ): there were 6 clones with insertion of A at nt5060, 2 with insertion at nt5213 and 6 with insertion at nt5222. Insertions of U, C or G Construction of cDNA for rhMPV/ΔM2-2, rhMPV/GFPpolyA and rhMPV/ΔM2-2/GFPpolyA Figure 1 Construction of cDNA for rhMPV/ΔM2-2, rhMPV/GFPpolyA and rhMPV/ΔM2-2/GFPpolyA. A) rhMPV/ΔM2-2 has a deletion in the M2-2 gene adjacent to a SwaI site. Nucleotides that were modified to introduce the SwaI site are underlined. Translational stop codons are bold and the intergenic (IG) sequence is bold italics. B) To construct rhMPV/GFPpolyA, an NheI site was introduced at the M2-2 stop codon of rhMPV and an NheI-N/P-GFP-polyA-NheI cassette was inserted. The modified nucleotides are underlined, the stop codon is bold and the IG sequence is bold italics. C) To construct rhMPV/ΔM2-2/GFPpolyA, an NheI site was introduced between the stop codon of M2-1 (bold) and the SwaI site (italics) in rhMPV/ΔM2-2 and an NheI-N/ P-GFP-polyA-NheI cassette was inserted. The modified nucleotides are underlined and the IG sequence is in bold italics. D) The reading frame of GFP is aligned with that of GFPpolyA to show the stop codon and frame shift resulting from the 11 nt insertion. were not observed. In sharp contrast, no A nucleotide insertions were detected in 15 cloned RT-PCR products derived from the identical region of rhMPV. Transitions, transversions, and deletions were also observed for rhMPV/ΔM2-2 in addition to insertions of A. For rhMPV/ ΔM2-2, 14 of 15 cloned RT-PCR sequences exhibited a total of 79 transition mutations, 2 transversions, and 1 deletion. Only 1 cloned sequence from rhMPV/ΔM2-2 had no nucleotide changes. In comparison, 7 of 15 cloned RT-PCR products of rhMPV showed a total of 17 transitions, 2 transversions, and 1 deletion. Eight cloned RT-PCR sequences from rhMPV had no nucleotide changes ( Figure 3B ). Thus, by passage 4, both rhMPV and rhMPV/ ΔM2-2 contained heterogeneous subpopulations and rhMPV/ΔM2-2 had a higher frequency of transition mutations and a propensity for insertion of A nucleotides in poly A tracts, compared to rhMPV. To determine whether U insertions were present in the antisense genome, a two step RT-PCR was performed to specifically amplify only genomic RNA. Again, total RNA was extracted from a passage 4 stock of rhMPV/ΔM2-2 and the region from nt4536 in F to nt5623 in SH was amplified. 1 kb RT-PCR products were inserted in pCR2.1 and 15 individual plasmids were sequenced. All 15 cloned RT-PCR products contained an insertion of 1 or 2 T nts (antisense) in either the F gene (non-coding region), the M2-1 gene or the SH gene ( Figure 3C ). One sequence had an A inserted at nt4964 in the M2-1 gene. However, no insertions of C or G were observed. The 15 cloned RT-PCR products also contained 18 transitions and 4 transversions. Thus, there is a high frequency of U insertions in the genomic RNA suggesting that insertions were propagated in the viral genome. Whether the insertion events occurred during synthesis of the genomic or antigenomic RNA cannot be determined from these data. We next examined the frequency of poly A and poly U tracts in the hMPV sequence spanning nt4536 to nt5623, to determine whether there is a bias between insertions of A or U. This region contains 14 poly A tracts and 3 poly U tracts with 4 or more contiguous A or U residues, respectively. Among the 15 cloned RT-PCR products amplified from the genomic RNA, 26 incidences of inserted A and 1 of inserted U were observed ( Figure 3C ). Thus, the data suggest a strong bias for insertions of A. We also looked for insertions outside of the region that encoded the non-essential genes M2-1 and SH. RT-PCR was performed on rhMPV/ΔM2-2 and rhMPV total RNA to amplify the N/P, P/M, F/M2, SH/G and G/L non-coding sequences. There was a total of 23 poly A tracts and 2 poly U tracts with 4 or more contiguous A or U residues, respectively, among these sequences. However, no insertions of A were observed in the any of these non-coding sequences, showing that the high frequency of A insertions was predominantly confined to the region encoding the non-essential genes M2-1 and SH. To investigate whether these insertions also occur in non-hMPV sequences, a GFP gene was inserted into the sixth gene position between the M2 and SH transcription units of rhMPV and rhMPV/ΔM2-2. An assay was developed to detect insertions, by designing a GFP ORF with an 11-nt sequence, CGA 6 TTA, positioned downstream of the first two GFP codons. This resulted in a frame shift in the downstream reading frame and a premature translational stop codon at the 6 th GFP codon, abrogating expression of GFP ( Figure 1B ,C and 1D). The modified GFP ORF is labeled GFPpolyA ( Figure 1D ). Insertion of a single nucleotide (or 4, 7, 10, etc.) in the A 6 tract of the CGA 6 TTA sequence would restore the translationally silenced GFP ORF, resulting in a fluorescent hMPV infectious focus. Four full-length cDNA's were engineered to recover Growth of rhMPV and rhMPV/ΔM2 in Vero cells Titer (log 10 PFU/ml) rhMPV/GFP, rhMPV/GFPpolyA, rhMPV/ΔM2-2/GFP, and rhMPV/ΔM2-2/GFPpolyA viruses. Titers ranged from 6.6 log 10 PFU/mL for rhMPV/ΔM2-2/GFP to 7.3 log 10 PFU/ml for rhMPV/GFPpolyA and plaque sizes between all four viruses were similar ( Figure 4A ). However, rhMPV/ΔM2-2 and rhMPV/GFP plaques were both smaller than rhMPV plaques. Vero cells were inoculated at MOI of 0.1 with rhMPV/GFP, rhMPV/GFPpolyA, rhMPV/ΔM2-2/GFP or rhMPV/ΔM2-2/ GFPpolyA as well as the control viruses rhMPV and rhMPV/ΔM2-2. Viruses were harvested on day 4 for Western blot analysis. The Western blot was probed for expression of hMPV F and GFP. Actin was also probed as a loading control ( Figure 4C ). The levels of hMPV F as detected by Western blot were considered equivalent among the GFP-viruses ( Figure 4B ). As expected GFP protein was detected by Western blot only in rhMPV/GFP and rhMPV/ΔM2-2/GFP, and not in rhMPV/GFPpolyA and rhMPV/ΔM2-2/GFPpolyA ( Figure 4D ). These data indi-Chromatogram and frequency of A insertions and point mutations in rhMPV/ΔM2-2 compared to rhMPV Figure 3 Chromatogram and frequency of A insertions and point mutations in rhMPV/ΔM2-2 compared to rhMPV. A) A chromatogram of the RT-PCR product derived from P4 of rhMPV/ΔM2-2, spanning nt4536 in F to nt6205 in NCR of SH, contained this sequence showing two subpopulations. One population is the correctly cloned sequence; the second population has one inserted A nt (sense direction) at nt5222 in the M2-1 gene. B) To assess the relative frequency of mutations, RT-PCR fragments spanning nt4536 in F to nt5623 in SH were obtained from rhMPV/ΔM2-2 or rhMPV using one-step RT-PCR, and were cloned into pCR2.1 plasmids. Among 15 independent plasmids the number of inserted As, single nt deletions, and point mutations (transition or transversion) for each virus were tabulated. 14 of the 15 (93%) rhMPV/ΔM2-2RT-PCR products had an inserted A (sense direction) nucleotide. No fragments containing A nucleotide insertions were detected in any of the 15 RT-PCR fragments spanning the identical region in P4 of rhMPV. C) To study frequency of mutations in genomic RNA, RT-PCR fragments spanning nt4536 to nt5623 were obtained from rhMPV/ΔM2-2 using two-step RT-PCR, and were cloned into pCR2.1 plasmids. Nucleotide insertions were predominantly T (genomic antisense direction), with one A, and were distributed among 8 locations in the fragments. D) To describe the position where insertion of an A was observed, the nt number of the last A in the poly A tract is used, though it is not known which A residue in the poly A tract is the inserted residue. The example shown is A inserted at nt5166. Gene: cate that insertion of the GFP cassette at this genome position was well tolerated by hMPV in vitro and insertion of the CGA 6 TTA sequences in the N terminus of the GFP ORF effectively silenced GFP expression. To indirectly monitor A nucleotide insertions in GFP-polyA, Vero cells were inoculated with rhMPV/ΔM2-2/ GFPpolyA or one of the control viruses rhMPV/GFP, rhMPV/GFPpolyA or rhMPV/ΔM2-2/GFP, at MOI of 0.1, and viewed by fluorescence microscopy for 6 days. Fluorescence was readily observed throughout the monolayers Functional GFP expression in rhMPV/ΔM2-2/GFP6 poly A by A nucleotide insertion Figure 4 Functional GFP expression in rhMPV/ΔM2-2/GFP6 poly A by A nucleotide insertion. A) rhMPV and rhMPV/ΔM2-2 viruses containing the native GFP ORF or GFP-polyA sequences, harboring an engineered poly A tract that silenced the translation of GFP, formed comparable plaques in Vero cells. B) Western blots indicated F expression was comparable between viruses. C) A duplicate Western blot was probed with antibody directed to actin to serve as a loading control. D) GFP was detected by Western blot in viruses that contained native GFP ORFs. E) Fluorescence was robustly detected in viruses that contained native GFP ORFs, was readily detectable in some fluorescent foci in rhMPV/ΔM2-2/GFPpolyA, and was not detected in rhMPV/GFPpolyA. F) Nucleotide insertion of one A restored function of GFPpolyA ORF. Nucleotide insertion of 3 As would not restore functional GFPpolyA, but indicated heterogeneity at this polyA locus. of Vero cells infected with rhMPV/GFP or rhMPV/ΔM2-2/ GFP, but not in cells infected with rhMPV/GFPpolyA (Figure 4E ). Initially, no fluorescence was observed in cells infected with rhMPV/ΔM2-2/GFPpolyA. However, after two days, a few foci of fluorescent cells were observed in monolayers infected with rhMPV/ΔM2-2/GFPpolyA, suggesting that some cells were infected with GFP-expressing hMPV. One focus containing approximately a hundred infected fluorescent cells is shown ( Figure 4E ). The expression of GFP indicated that the reading frame of the GFP gene had been restored in some virions, and cell-to-cell spread within the focus of infection suggested that the restored GFP gene sequences were present in progeny virions. The low level of GFP expressed was only observable by fluorescence microscopy and not by Western blotting ( Figure 4D ). To assess the frequency of insertions that restored expression of GFP, Vero cells in 96-well plates were inoculated with P2 stocks of rhMPV/ΔM2-2/GFPpolyA or rhMPV/ GFPpoly A. GFP expression was monitored by fluorescence microscopy 4 days post infection. Plates were inoculated with 1, 10, 100 or 1000 PFU/well (Table 1) . No GFP-expressing foci were observed in wells inoculated with either 100 or 1000 PFU/well of rhMPV/GFPpoly A (Table 1 ). In contrast, cells inoculated with 10, 100, or 1000 PFU/well of rhMPV/ΔM2-2/GFPpoly A developed fluorescent foci. Fluorescent multicellular foci were observed in 25 out of 384 wells (6%) inoculated with 10 PFU/well of rhMPV/ΔM2-2/GFPpolyA (Table 1) . At 100 PFU/well of rhMPV/ΔM2-2/GFPpolyA, fluorescence was observed in 65% of the infected wells (Table 1) . Finally, fluorescent multicellular foci were observed in 100% of wells inoculated with 1000 PFU/well of rhMPV/ΔM2-2/ GFPpolyA. Thus, this assay shows that at least one insertion occurs out of approximately every 17 infections at 10 PFU/infection and the frequency of insertions was significantly elevated in the absence of M2-2. Viruses from 24 of the wells that exhibited fluorescence and that had been inoculated at a MOI of 0.1 were passaged once in Vero cells and each of the 24 viruses retained GFP expression. Total RNA was extracted from a mixture of cells plus supernatant and RT-PCR was performed to amplify a 1.5 kb fragment encompassing the GFPpoly A gene. The RT-PCR product was cloned into pCR2.1 and 8 individual clones were sequenced. 4 cloned GFP fragments contained the 11-nt CGA 6 TTA insert as constructed, 3 contained 1 inserted A that restored the reading frame of GFP, and 1 contained 3 inserted A nucleotides in the A 6 tract ( Figure 4F ). Thus, insertions of A nucleotides occurred frequently in non-hMPV sequences as well during rhMPV/ΔM2-2/GFPpolyA replication, suggesting that misincorporation of A nucleotides is not hMPV sequence-specific. To further investigate the role of hMPV M2-2, we compared the transcription and genome replication of rhMPV/ΔM2-2 with rhMPV in Vero cells. First, we compared the amounts of rhMPV/ΔM2-2 viral transcripts with that of rhMPV by Northern blotting. Northern blot analysis was performed using hMPV-specific anti-sense DIGlabeled riboprobes to detect M2, SH, N, F, or G mRNA. At 24-hr intervals, RNA was extracted from Vero cells inoculated with rhMPV or rhMPV/ΔM2-2 at an MOI of 0.1, and Northern blot analysis was performed in 6 replicates. The M2 and SH riboprobes each detected two RNA species from rhMPV-infected cells (Lanes 1, 3, 5 and 7 of Figure 5A and 5B). The size of the minor species is consistent with the monocistronic transcript while the size of the major species coincided with the predicted size of the M2/ SH read-through product. No monocistronic M2 transcripts were observed at 24 or 48 hours post rhMPV/ΔM2-2 infection in Vero cells. The predicted M2/SH readthrough product showed a reduction in size in rhMPV/ ΔM2-2 infected cells consistent with the deletion of M2-2 (compare lanes 1 and 2 of Figure 5A ). The levels of bicistronic compared to monocistronic SH transcripts were higher in both rhMPV and rhMVP/ΔM2-2 infected cells, but the difference was more pronounced in rhMPV/ΔM2-2 infected cells ( Figure 5B ). This increased level of readthrough was unexpected since we had sought to preserve the native M2/SH noncoding sequences. One explanation could be that transcription termination at the M2 gene end sequences required nucleotides in the coding region of M2-2 that had been inadvertently removed and/or the M2 termination signal was altered by the introduction of the Swa I site. Next we compared the amounts of M2 transcripts in cells infected with rhMPV or rhMPV/ΔM2-2 at days 1 to 4 postinfection (p.i.). At day 1 p.i., the levels of transcripts were equivalent between both viruses (lanes 1 and 2 in Figure 5A ). By day 2 p.i., the relative levels had changed markedly. The amount of transcripts in cells infected with rhMPV/ΔM2-2 was several-fold higher compared to cells infected with rhMPV (lanes 3 and 4 in Figure 5A ). The upregulation was maintained up to day 4, when peak titers were observed (lanes 7 and 8 in Figure 5 ). More SH, N, F, and G transcripts were also observed in cells infected with rhMPV/ΔM2-2 compared to rhMPV ( Figure 5B ,C,D and 5E). Therefore, M2-2 deletion resulted in up-regulation of viral transcripts of genes upstream (N, F) and downstream (SH, G) of the M2 gene. However, the increased levels of viral transcripts produced by the rhMPV/ΔM2-2 mutant were not accompanied by an increase in virus titer. On days 2 and 3, rhMPV/ΔM2-2 had higher levels of transcripts but equivalent or lower titers compared to rhMPV ( Figure 5H) . Neither was there a concomitant increase in protein expression, at least for the F gene ( Figure 4B , lanes 1 and 2). Thus, the higher levels of viral transcripts produced by the M2-2 deletion mutant did not yield a greater number of infectious rhMPV/ΔM2-2 virions compared to rhMPV. We noted that the levels of rhMPV transcripts peaked at day 3 (lanes 1, 3, 5 and 7 in Figure 5 ), while the levels of rhMPV/ΔM2-2 transcripts remained the same on days 3 and 4 (lanes 6 and 8 of Figure 5 ). RNA samples from day 4 were also probed for genomic (anti-sense) RNA using a mixture of three riboprobes directed to P, M and F genes. No significant differences were observed between the amount of genomic RNA in cells infected with rhMPV/ΔM2-2 and rhMPV (lanes 7 and 8, Figure 5F ). Thus, deletion of M2-2 altered the ratio between hMPV genomic RNA and mRNA. Syrian Golden hamsters are highly permissive for hMPV replication and were used to assess the attenuation of rhMPV/ΔM2-2 [14, 18] . Groups of 8 hamsters were inocu-lated on day 0 with 10 6 PFU of wthMPV/NL/1/00, rhMPV or rhMPV/ΔM2-2. Both the recombinant viruses were P3 stocks. On day 4, titers of virus in the nasal turbinates and lungs were compared. As expected, the titers of wthMPV/ NL/1/00 and rhMPV in nasal turbinates and lungs were comparable (Table 2) . However, the titers of rhMPV/ΔM2-2 were 3.7 log 10 PFU/gm lower in the URT and 1.8 log 10 PFU/gm lower in the LRT, relative to rhMPV titers. Therefore, rhMPV/ΔM2-2 was approximately 10,000-fold and 100-fold more restricted in the URT and LRT, respectively, compared to rhMPV. To determine if the lower level of replication in lungs and nasal turbinates of hamsters was sufficient to protect the hamsters from subsequent infection with hMPV, 4 hamsters were challenged with 10 6 PFU of wthMPV/NL/1/00 4 weeks post immunization. Four days post administration of the challenge, no virus was detected in either lungs or nasal turbinates of the immunized hamsters while unvaccinated animals had 5.6 +/-0.6 PFU/gm in URT and 4.5 +/-1.5 PFU/gm in the LRT (Table 2) . Therefore, replication of rhMPV/ΔM2-2 was restricted in hamsters and animals were protected from challenge with wthMPV/NL/1/ 00. Using reverse genetics, we engineered rhMPV lacking the M2-2 gene with the aim of generating a potential vaccine candidate. rhMPV/ΔM2-2 grew to high titer in Vero cells, was attenuated in the respiratory tract of hamsters, and protected immunized hamsters from challenge with wthMPV/NL/1/00. These results agree with a similar study reported by Buchholz et al. in which a different subtype A hMPV strain, CAN97-83, with a deletion of M2-2 was proposed as a potential vaccine candidate [14, 15] . Our studies utilized the rhMPV/NL/1/00/E93K/S101P backbone which contained engineered mutations in the hMPV F gene that allows this virus to replicate efficiently in Vero cells without trypsin supplementation [17] . This property is expected to facilitate the testing and manufacture of potential live hMPV vaccine candidates. To assess the genetic stability of our M2-2 deletion mutant, sequence analyses were performed on P4 stocks of rhMPV/ΔM2-2. These analyses revealed major subpopulations (as high as 50%) that contained insertions of A nucleotides (sense direction) in the M2-1 and SH ORFs. These insertions appeared predominantly in A tracts and were also observed in non-hMPV sequences. Nucleotide insertions were also readily detected in an A tract introduced in the GFP ORF. Interestingly, insertions of A were not observed outside the region encompassing the nonessential genes M2 and SH. Transcriptional editing, whereby alternative reading frames of viral genes are accessed, has been observed in the P gene of several paramyxoviruses [19] [20] [21] [22] [23] . Therefore it is possible that an inserted A could occur frequently during transcriptional editing of paramyxovirus RNA. The nucleotide insertions observed in rhMPV M2-2 deletion mutants differ somewhat from transcriptional editing in that (i) the positions of inserted A nucleotides did not appear to be sequence biased beyond selecting for A tracts and is not hMPV sequence specific, and (ii) the nucleotide insertions were incorporated into the viral genome and could be propagated, as shown by passaging of fluorescent rhMPV/ΔM2-2/GFPpolyA viruses. Interestingly, these insertions did not appear to confer growth advantages in Vero cells because further passaging of rhMPV/ΔM2-2 promoted new A insertions and did not increase the subpopulations of earlier insertions. Many of the A nucleotide insertions caused premature translation terminations in the non-essential M2-1 and/or SH ORFs. These observations argue mechanistically against transcriptional editing and suggest that the insertions observed when M2-2 was deleted may be caused by an alteration in the fidelity of the replication complex directly or indirectly. Removal of the hMPV M2-2 gene resulted in up-regulation of viral transcription, although there was no alteration in the level of genomic RNA. This had been observed previously for the respiratory syncytial virus (RSV) M2-2 gene as well as for hMPV [14] . Deletion of RSV M2-2 resulted in higher levels of viral transcripts compared to wt RSV. Based on these observations it was postulated that the RSV M2-2 is involved in regulating the balance between transcription and genome replication [24, 25] . Our observation that the levels of rhMPV transcripts peaked at day 3 p.i., while the levels of rhMPV/ΔM2-2 transcripts remained high through day 4 p.i. is also consistent with a higher level of viral transcripts in rhMPV/ ΔM2-2 infected cells. Thus, deletion of the hMPV M2-2, like its RSV counterpart, appears to cause aberrant regulation of viral transcription. Comparison of monocistronic and polycistronic viral transcripts showed differences in the frequency of readthrough transcription at the M2 gene end sequences between rhMPV and rhMPV/ΔM2-2 infected cells. In RNA from cells infected with rhMPV, the M2 riboprobes detected a minor monocistronic M2 transcript and a major polycistronic M2/SH readthrough transcript. While transcription readthrough is not unique to the M2/SH intergenic region, the polycistronic readthrough transcripts at other noncoding regions such as N/P and F/M2 were less pronounced and monocistronic transcripts predominated. The genes immediately upstream and downstream of the M2 and SH transcription units also existed predominantly as monocistronic transcripts indicating that the M2 gene-end sequences are particularly prone to high frequency of readthrough transcription. The frequency of readthrough transcription at the M2 gene stop sequences appeared to be accentuated by the removal of the M2-2 ORF. This may in part be attributed to the sequences that were removed and/or altered by the introduction of a Swa I site at the proximity of the M2 gene end sequences. Nonetheless, the increased frequency of readthrough at this gene junction may perturb the expression of downstream genes such as SH, G and L. In rhMPV/ ΔM2-2 infected cells, there are major populations of M2-1 transcripts that contained premature termination codons introduced by the high point mutation frequency. Therefore, it is possible that M2-1 expression was significantly reduced during rhMPV/ΔM2-2 infection and this reduction in M2-1 expression may also contribute to the up-regulation of transcription and increased frequency of read-through observed. Our results differ somewhat from that reported for the recombinant CAN97-83 strain of hMPV. Growth of recombinant rΔM2-2 CAN97-83 is trypsin-dependent and peak titer was not observed until 11 days post infection [14] . In contrast, our rhMPV/ΔM2-2 achieved peak titers at 4 days post-infection, a significant savings in production time. Interestingly, both ΔM2-2 viruses showed dramatic up-regulation of transcription at 48 hours post infection despite very different growth kinetics. No increase in the frequency of read-through transcription was observed for rΔM2-2 CAN97-83 whereas we observed increased polycistronic M2/SH transcripts in rhMPV/ ΔM2-2 infected cells. This may stem from differences in the construction of the M2-2 deletion. rΔM2-2CAN97-83 had a deletion of 152 nt in the M2-2 ORF whereas our construct had a deletion of 142 nt and a SwaI site introduced adjacent to the polyA tract of the M2 gene stop sequences. However, the ratio of polycistronic M2/SH transcripts to monocistronic M2 transcripts was significantly different even between the two wild-type hMPV strains, with the Netherlands strain exhibiting a higher frequency of readthrough at the M2/SH noncoding region than the Canadian strain. Sequence analysis of rhMPV CAN 97-83, showed that mutations do develop in SH, G, L and non-coding regions (NCR), with a particularly high frequency in the SH gene [26] . While mutations and insertions were reported for rhMPV and rhMPVΔ G of the CAN 97-83 strain, the sequence analysis did not include viruses that lacked the M2-2 gene [26] . In a separate evaluation of rΔM2-2 CAN97-83, no increase in the frequency of point mutations was reported [14] . While the M2-2 proteins of both strains are completely identical, the SH protein of the CAN97-83 strain is only 83% identical to the NL/1/00 strain. There are also 26 amino acid differences in the L gene between the two strains. Finally, although deletion of the hMPV M2-2 ORF succeeded in attenuating both hMPV strains, the CAN97-83ΔM2-2 virus was more attenuated in hamsters than rhMPV/ΔM2-2 [14] . Clearly there are differences between the CAN97-83 and hMPV/NL/1/ 00 strains. Further study will be required to elucidate the differences in phenotype. In summary, M2-2 plays an important role in the genetic stability of the hMPV genome. Silencing of M2-2 expression resulted in a greater frequency of hMPV subpopulations harboring insertions and point mutations. Stabilizing the sequence of the rhMPV/ΔM2-2 genome by re-engineering all poly A tracts will only be partially effective because this does not address the increased frequency of point mutations. More studies are needed to gain detailed knowledge of the hMPV polymerase complex and the role of M2-2 during hMPV replication. Ablation of the M2-2 ORF also resulted in up-regulation of viral transcription but not genomic RNA and increased the frequency of readthrough transcription at the M2/SH noncoding region. Aberrant transcription regulation and increased genetic instability could both contribute to the attenuation phenotype rhMPV/ΔM2-2. Further studies are being conducted to develop a potential live hMPV vaccine candidate. [2, 17, 27] . The following recombinant viruses were generated by reverse genetics from full-length cDNA plasmids using the rhMPV/NL/1/00/ E93K/S101P backbone: rhMPV/ΔM2-2, rhMPV/GFP, rhMPV/GFPpolyA, rhMPV/ΔM2-2/GFP, and rhMPV/ ΔM2-2/GFPpolyA. The cDNA for rhMPV/NL/1/00/E93K/S101P was constructed as previously described [17] . To generate rhMPV/ ΔM2-2 cDNA, two Swa I restriction sites were introduced into the SphI/ClaI subclone, which contained the SphI (nt 100) to Cla I (nt8678) fragment derived from rhMPV/NL/ 1/00/E93K/S101P, using a Quik change mutagenesis kit (Stratagene). The first SwaI site was positioned at nt5326, down stream of the M2-1 stop codon, and was generated with the primer 5' CAGTGAGCATGGTCCAATTTAAAT-TACTATAGAGG and its complement. The stop codon of M2-1 is in bold and the Swa I restriction site is underlined. The second SwaI site was positioned at nt5468, upstream of the native stop codon of M2-2, and was generated using the primer 5' CATAGAAATTATATATGTCAAGGCTTATT-TAAATTAG and its complement. Digestion with Swa I resulted in the removal of 142 nts of the M2-2 gene. The Stu I (nt4495) to Cla I (nt8693) fragment in the subclone, containing M2-1, SH, and G genes, was transferred into the full-length rhMPV/NL/1/00/E93K/S101P cDNA to form rhMPV/ΔM2-2 as depicted in Figure 1A . rhMPV/ ΔM2-2 cDNA plasmid used to recover virus was sequenced from nt4495 to nt8693 to confirm that no unexpected changes were generated by PCR during cloning. To construct the cDNA for rhMPV/GFP, a NheI-N/P-GFP-NheI cassette was constructed with the N/P noncoding region (NCR) upstream of the GFP gene. An NheI site was introduced into the hMPV SphI/ClaI subclone at nt 5474, located at the stop codon of M2-2, using the primer 5'GCTTACTTAAGCTAGCTAAAAACACATCAGAGTGG (NheI site underlined) and its complement. Following NheI digestion, the NheI-N/P-GFP-NheI cassette was ligated into the subclone. A StuI (nt4495) to ClaI (nt8693) fragment, containing M2, GFP, SH and G genes, was isolated from the hMPV SphI/ClaI subclone and inserted into the full-length cDNA to form rhMPV/GFP. To construct rhMPV/GFPpolyA, an 11 nt Poly A insert was cloned into the NheI-N/P-GFP-NheI cassette using the primer 5' TGAGCTAGCTTAAAAAAGTGGGACAAGTCAAAATGGT-GCGAAAAAATTAAGCAAGGGCGAGG (hMPV P gene start sequences is in bold, the GFP sequences are italicized, and the 11 nt poly A insert is underlined) to generate the cassette NheI-N/P-GFP-PolyA-NheI. The NheI-N/P-GFP-PolyA-NheI cassette was ligated into the hMPV SphI/Cla I subclone at the Nhe I site located at nt5474. Again the StuI (nt4495) to ClaI (nt8693) fragment of the hMPV SphI/ Cla I subclone was inserted into the full-length cDNA to generate rhMPV/GFPpolyA ( Figure 1B) . To construct rhMPV/ΔM2-2/GFP and rhMPV/ΔM2-2/ GFPpolyA, an Nhe I site was introduced into the SwaIΔM2-2 subclone at nt5316, using the primer 5'GCACTAATCAAGTGCAGTGAGCTAGCATTTAAATTAG and its complement (the stop codon of M2-1 is in bold and the Nhe I site is underlined). The NheI-digested GFPcontaining NheI-N/P-GFP-NheI or NheI-N/P-GFP-PolyA-NheI cassette was inserted at nt5316 to generate rhMPV/ ΔM2-2/GFP or rhMPV/ΔM2-2/GFPpolyA respectively ( Figure 1C ). Recombinant viruses were generated from cDNA as described previously [27] . Briefly, 1.2 ug of pCITE hMPV N, 1.2 ug of pCITE hMPV P, 0.9 ug of pCITE hMPV M2, 0.6 ug pCITE hMPV L, and 5 ug of full-length hMPV cDNA plasmid in 500 uL optiMEM containing 10 uL lipofectamine 2000 (Invitrogen), was applied to a monolayer of 10 6 BSR/T7 cells. The medium was replaced with fresh optiMEM 15 hr post transfection and incubated at 35°C for 2 to 3 days. Cells and supernatant from the transfection were harvested together and used to infect Vero cells. Virus recovery was verified by positive immunostaining 6 days post inoculation with ferret polyclonal Ab directed to hMPV. Recovered viruses were further amplified in Vero cells by inoculating at a multiplicity of infection (MOI) of 0.1 PFU/cell. Virus titers (plaque forming units (PFU)/ml) were determined by plaque assay in Vero cells. Monolayers of Vero cells in TC6-well plates were inoculated with 10-fold serial dilutions of virus. After 1 hour adsorption, the inoculum was aspirated, the monolayer was overlaid with optiMEM diluted 1:1 with 2% methylcellulose, and the plates were incubated for 7 days at 35°C. Plaques were immunostained with ferret polyclonal antisera directed to hMPV diluted 1:500 in PBS containing 5% powdered milk (w/v) (PBS-milk). The cells were then incubated with horseradish peroxidase-conjugated goat anti-ferret antibody (Dako) diluted 1:1000, followed by incubation with 3-amino-9-ethylcarbazole (AEC) (Dako) to visualize plaques. Ferret polyclonal antisera were generated by collecting blood 4 weeks after infection of 8-10 week old ferrets with 6.0 log 10 wthMPV/NL/1/00 administered intranasally. Subconfluent monolayers of Vero cells in TC6-well plates were inoculated at a MOI of 0.1 PFU/cell with rhMPV or rhMPV/ΔM2-2 diluted in optiMEM. After 1 hr adsorption at 35°C, the virus inoculum was replaced with 2 ml opti-MEM. Combined cells and supernatant were collected at 24 hr intervals for 4 or 6 days, stabilized with 1× SPG and frozen at -70°C. Virus titers were determined by plaque assay in Vero cells. Five-week-old Syrian golden hamsters (8 animals per group) were infected intranasally with 10 6 PFU/animal of virus or placebo medium in 100 uL. Four days post infection, the nasal turbinates and lungs were harvested from 4 animals per group, homogenized and titered by plaque assay in Vero cells. On day 28 post infection, immunized hamsters were challenged with an intranasal dose of 10 6 PFU/animal of wthMPV/NL/1/00. Four days post-challenge, the nasal turbinates and lungs were harvested and assayed for challenge virus replication by plaque assay. Total RNA was extracted from hMPV-infected cells using TRizol (Invitrogen) reagent according to the manufacturer's instructions followed by phenol/chloroform (Amresco) extraction and ethanol precipitation. RT-PCR products of total hMPV RNAs were generated using a one step Ultrasense RT-PCR kit (Invitrogen) with sense and anti-sense primers designed to generate 1 to 2 kb fragments from total RNA. Genomic RNA was amplified with a two-step Super Script III platinum RT-PCR kit (Invitrogen). To ensure that only genomic RNA was amplified in the two-step process, only the sense primer was present during the RT step, the RT was deactivated by treatment of the product at 94°C for 15 minutes, and phenol/chloroform extraction was performed on the RT product to remove any residual RT enzyme prior to the PCR reaction. Sequence analysis was performed only on DNA fragments of the expected size isolated from agarose gels using a gel extraction kit (Qiagen). Vero cells were inoculated at MOI = 0.1 and incubated at 35°C for 4 days. Cells and supernatant were scraped together and total RNA was extracted using Trizol reagent (Invitrogen) followed by an additional phenol-chloro-form extraction and ethanol precipitation. RNA was separated on 1% agarose gel in the presence of 0.44 M formaldehyde and transferred to a positively charged nylon membrane (Amersham Biosciences). The RNA was hybridized with gene-specific riboprobes labeled with digoxigenin using a DIG RNA labeling kit (Roche). The hybridized bands were visualized with a DIG luminescent detection kit (Roche). Western blot analysis was performed as described previously [17] . Briefly, in duplicate, cell lysates of hMPVinfected Vero cells were separated on a 12% polyacrylamide Tris-HCl Gel (Bio-Rad), transferred to a Hybond-P polyvinylidene difluoride membrane (Amersham Biosciences) and immunostained with either hamster Ab # 121-1017-133 (MedImmune) directed to hMPV F, mouse Ab directed to GFP (Roche Molecular Biochemicals), or mouse Ab directed to actin (Chemicon MAB #1501). Bands were visualized by incubation with horseradish peroxidase-conjugated anti-hamster Mab or anti-mouse Mab, developed with chemiluminescence substrate (Amersham Biosciences), and exposed to Biomax MR film (Kodak).

No Longer an Innocent Bystander: Epithelial Toll-Like Receptor Signaling in the Development of Mucosal Inflammation

Diseases of mucosal inflammation represent important causes of morbidity and mortality, and have led to intense research efforts to understand the factors that lead to their development. It is well accepted that a breakdown of the normally impermeant epithelial barrier of the intestine, the lung, and the kidney is associated with the development of inflammatory disease in these organs, yet significant controversy exists as to how this breakdown actually occurs, and how such a breakdown may lead to inflammation. In this regard, much work has focused upon the role of the epithelium as an “innocent bystander,” a target of a leukocyte-mediated inflammatory cascade that leads to its destruction in the mucosal inflammatory process. However, recent evidence from a variety of laboratories indicates that the epithelium is not merely a passive component in the steps that lead to mucosal inflammation, but is a central participant in the process. In addressing this controversy, we and others have determined that epithelial cells express Toll-like receptors (TLRs) of the innate immune system, and that activation of TLRs by endogenous and exogenous ligands may play a central role in determining the balance between a state of “mucosal homeostasis,” as is required for optimal organ function, and “mucosal injury,” leading to mucosal inflammation and barrier breakdown. In particular, activation of TLRs within intestinal epithelial cells leads to the development of cellular injury and impairment in mucosal repair in the pathogenesis of intestinal inflammation, while activation of TLRs in the lung and kidney may participate in the development of pneumonitis and nephritis respectively. Recent work in support of these concepts is extensively reviewed, while essential areas of further study that are required to determine the significance of epithelial TLR signaling during states of health and disease are outlined.

Standing at the interface between the host and the environment, mucosal-lined surfaces represent the first line of defense against potential pathogens. This defensive role is particularly relevant to the mucosa of the gastrointestinal tract, the pulmonary system, and the urinary tract, each of which is particularly susceptible to the development of inflammatory diseases due to their role as a barrier that must not only protect, but also serve the physiological function of each of the organ systems. In the case of the gastrointestinal tract, mucosal inflammation is manifest as inflammatory bowel disease (including Crohn's disease and ulcerative colitis) (1) (2) (3) or necrotizing enterocolitis (NEC), a leading cause of death in preterm infants (4) . In the case of the pulmonary system, mucosal inflammation may be manifest as pneumonitis, pneumonia, or asthma (5-7), acute and chronic pulmonary conditions that have a high degree of morbidity and potential mortality. And in the case of the urinary tract, mucosal inflammation may be manifest as interstitial nephritis, cystitis, and urethritis (8) (9) (10) , causes of significant morbidity in patients of all ages. To elucidate the pathogenesis of mucosal inflammatory diseases, research over the past several decades has focused on the role of the immune system in their developmentin particular the relationship between mucosal lymphocytes, macrophages, and neutrophils, and the effects of their cellular by-products on mucosal integrity and function (11) (12) (13) . However, recent work has shed light upon the important role that the epithelia itself may play as a primary regulator of the immune response in the development of mucosal inflammation. No longer an innocent bystander, the epithelial-lined mucosa at each of these sites has been shown to possess all of the required armamentarium to allow an effective response to invading challenges, and to lead the battle to neutralize potential microbial threats (14) (15) (16) . Not only is the epithelium able to respond to potentially dangerous microbial products, it also may sense endogenous molecules that are released during conditions of stress, hypoxia, or injury-so-called danger molecules that may play a critical role in the development of mucosal inflammation (17) (18) (19) . In order, therefore, to understand the pathogenesis of mucosal inflammation and to assist in the rational design of anti-inflammatory strategies, it is necessary to define the receptors and signaling pathways that mediate the inflam-matory response with respect to the epithelium itself. The innate immune system consists of a series of receptors and their associated signaling molecules that is present both on leukocytes and epithelial cells through the body, and which initiates an immune response by responding directly to pre-formed ligands ( [20] provides a recent review). The innate immune system lies in contradistinction to the adaptive immune system, a set of cellular and molecular interactions that must first "learn" how to deal with a potential pathogen, and then respond through the release of antibodies or other cellular derived products. The innate immune system includes pattern recognition receptors such as the TLRs, the NOD-like receptors (NLRs), the RIG-like receptors (RLRs), and C-type lectins, and their role in inflammatory signaling in leukocytes has been extensively reviewed (21) (22) (23) 186) . Relatively few reports have focused on the ability of the innate immune system to signal within the epithelium, although emerging evidence from a variety of laboratories including our own indicates that innate immune signaling within the epithelium plays a critical role in the pathogenesis of mucosal inflammation. The current review will focus on the TLR family, which has been shown to play a critical role in the response of epithelial cells to bacterial and endogenous ligands in the pathogenesis of various mucosal inflammatory diseases (Table 1) . A central controversy in the field of mucosal inflammation may be stated as follows (Figure 1 provides a pictorial representation of this): Is mucosal inflammation a reflection of a leukocyte-driven immune response that has gone awry, resulting in tissue injury and the loss of mucosal barrier function ( Figure 1A )? Or is it the mucosa itself, long known to play a role as a primary immune organ that is capable of producing a large number of pro-inflammatory molecules, that somehow has developed an exaggerated response that then leads to mu- Flagellin (58) 6 Diacyl Lipopeptides (144) Lipoteichoic acid (131) Zymosan (131) 7 Single-stranded RNA (145, 146) Imidazoquinoline (59) 8 Single-stranded RNA (146) Imidazoquinoline (147) 9 Bacterial (demethylated CpG) DNA (148) 10 Unknown 11 Profilin (149) Uropathogenic Bacteria (114) Figure 1 . Mechanisms by which TLR signaling leads to mucosal inflammation. As stated in the text, there are two potential mechanisms by which TLR signaling can lead to the development of mucosal inflammation. (A) TLR activation in response to DAMPs (damage associated molecular patterns) and PAMPs (pathogen associated molecular patterns) on leukocytes leads to the release of pro-inflammatory cytokines, resulting in epithelial destruction. (B) TLR activation by DAMPs and PAMPs on the epithelium itself directly impairs epithelial function and initiates the release of cytokines, leading to the development of mucosal inflammation. The development of mucosal inflammation likely arises from a combination of these mechanisms. cosal injury ( Figure 1B) ? Mucosal surfaces, such as the intestinal mucosa and the upper respiratory tract, are constantly exposed to environmental stimuli, such as commensal luminal bacteria in the intestine, as well as endogenous stimuli, or, as is the case in the lower respiratory tract and urinary tract mucosa, may encounter and respond to endogenous and exogenous stimuli in disease states. In either case, the mucosa must be able to mount an effective immune response, resist barrier failure, and coordinate this response with both mucosal and sub-mucosal leukocytes, while avoiding initiation and propagation of an exaggerated inflammatory response. But where does the answer lie in terms of what is initiating the mucosal inflammatory response? In seeking to answer this question, we and others have focused on the innate immune receptors that are present on the mucosa, and have examined the response of these receptors to known ligands in the development of mucosal inflammatory disorders. Such ligands may be broadly grouped into two categories: the so called "danger signals," a term used by Matzinger (24) , also called damage-associated molecular patterns (DAMPs) ( [25, 26] provide recent reviews); and those ligands on the surface or interior of pathogen-associated molecular patterns (PAMPs) ( Table 2) . Not surprisingly, there is a great deal of interest in identifying the important receptors for DAMPs and PAMPs expressed by various cell types, so as to accurately define their relative role in the development of inflammation. In this regard, several investigators have established that DAMPs and PAMPs are recognized by TLRs in many cells, including epithelial cells: a list of the TLRs and their cognate ligands appears in Table 1 . Although current dogma suggests that circulating leukocytes play a central role in the coordination of the immune response, emerging evidence suggests that the epithelium also plays a key role in the recognition and response to various "danger molecules" (27) (28) (29) (30) (31) . This review will examine in detail the various roles of epithelial signaling via TLRs in the development of common, and often devastating, mucosal inflammatory conditions. Much of the focus will be on the TLR-initiated signaling in response to PAMPs, as the majority of work has been performed in this area. Several recent investigators have shed light upon the important role of TLR signaling in the development of mucosal inflammation (31) (32) (33) (34) . To understand how TLRs may signal within the epithelium, information may be gained by analyzing TLR signaling in other systems, primarily within leukocytes. A full description of the molecular mechanisms by which TLR signaling occurs is beyond the scope of this review; currently accepted concepts with respect to TLR signaling are described below (21, 35 have recent reviews). The structure of each member of the TLR family of receptors provides important clues to how they function. All currently recognized TLRs are homologous with the interleukin-1 (IL-1) receptor, sharing an intracellular signaling domain, known as the Toll/IL-1R (TIR) domain (36) . A model that depicts the currently accepted mode of TLR signaling in leuko-cytes is shown in Figure 2 , in which the interaction with TLR4 and its cognate ligand lipopolysaccharide (LPS, endotoxin) is shown. The interaction of TLR4 with LPS leads to the activation of myeloid differentiation primary response protein 88 (MyD88)-dependent signaling, resulting in the induction of pro-inflammatory genes such as TNF-α, IL-1β, IL-6, and IL-10 (37-39) and MyD88-independent signaling cascades leading to activation of type-1 interferon (40 has a recent review) (38, 41) . MyD88-depedent signaling occurs as TIR domain-containing adapter protein (TIRAP/Mal) (42, 43) and MyD88 (44) (45) (46) interact with TLR4 and recruit IL-1 receptor-associated kinase (IRAK) family members IRAK1 and IRAK4 to the signaling heterocomplex consisting of TLR4, MyD88, and TIRAP (46) . Subsequent signaling occurs through tumor necrosis factor receptor-associated factor 6 (TRAF6)mediated (47) activation of transforming growth factor-β-activated protein kinase 1 (TAK1) (48) . TAK1 forms a complex with TAK1 binding proteins (TAB), TAB1 (49), TAB2 (50) , and TAB3 (51) . The TAK1/ TAB1/TAB2/TAB3 complex formation leads to the phosphorylation of IκB by IκB kinase (IκK) (52), initiating nuclear factorkappa B (NF-κB) signaling pathways and parallel activation of several mitogen-activated protein kinases (MAP-kinases) including c-Jun N-terminal kinase (JNK) (154) Viral proteins (106,143) Self DNA (155) Flagellin (58) Uric Acid (156) Bacterial DNA (148) Profilin (149) (53), p38 MAP-kinase (54) , and extracellular signal-regulated kinase (ERK) (55), ultimately resulting in pro-inflammatory gene induction (56, 57) . It is noteworthy that MyD88-dependent signaling is thought to be the predominant signaling pathway for TLR2 (41) , TLR5 (58), TLR7 (59) , and TLR9 (60,61). Figure 2 , TLR4 also may signal in the absence of MyD88. Evidence for this was demonstrated by Kawai et al., who described that the activation of NF-κB and MAP-kinases was reduced significantly, but not abolished, in MyD88-deficient mice (38) . The MyD88-independent signaling pathway proximally involves the activation of TRAM, a TIR-domain containing adapter molecule. TRAM associates with and activates TRIF, another TIR-domain containing adapter protein (62, 63) . TRIF then interacts with and activates TANKbinding kinase 1 (TBK1) and IKKε, two IκK homologs, which leads to the phosphorylation of IRF3 (64, 65) and translocation of IRF3 to the nucleus where it regulates the expression of various genes, including the type I IFN family of genes (66) . TRIF also interacts with TRAF6 and receptor interacting protein 1 (RIP), leading to the activation of NF-κB (67, 68) . Importantly, TLR3 signals mainly through the MyD88-independent, TRIFdependent pathway (62) . There is a wide and diverse spectrum of diseases that involves the development of inflammation of the intestinal mucosa. Such diseases include inflammatory bowel disease that is, Crohn's disease and ulcerative colitis, necrotizing enterocolitis (which is a leading cause of death and disability in newborn infants) and a variety of infectious causes of intestinal dysfunction, due to enteroinvasive organisms such as Salmonella and Escherichia coli. As shown in Figure 1 , signaling via TLRs could lead to the development of intestinal inflammation through direct interaction of TLRs with the intestinal epithelium, or through effects on sub-epithelial and circulating leukocytes whose activation then leads to the initiation and propagation of mucosal inflammation. Although evidence exists to support this latter possibility, the expression of various TLRs in enterocytes (Table 3) suggests the possibility that direct interaction of intestinal TLRs with cognate ligands (see Table 1 ) may occur. Enteric bacteria in general, and LPS in particular, have been shown to play a critical role in the development of many diseases of intestinal inflammation (69) (70) (71) (72) , further suggesting the possibility that enterocyte TLR signaling may contribute directly to the development of these diseases. To address the role(s), if any, of intestinal epithelial TLR signaling in the pathogenesis of intestinal inflammation, we, and others, have focused on TLR4, the receptor for LPS. Multiple enterocyte cell lines, including (IEC-6) rat enterocytes (27, 73) , primary and cultured (HT-29 and T84) colonocytes (74) (75) (76) and (CMT93) mouse rectal cells (75) express TLR4, the TLR4 adapter protein MD-2, and MyD88. In these cell lines, activation by LPS leads to pro-inflammatory signaling (74-76) as well as changes in cellular processes including proliferation (73) and intracellular TLR4 trafficking (77) . These findings provide evidence that enterocytes may respond directly to LPS via TLR4, yet by no means prove the physiological relevance of such a response. However, clinical significance for TLR signaling in the pathogenesis of intestinal mucosa is suggested as patients with inflammatory bowel disease demonstrate an increase in the expression of TLR4 and TLR2 in the intestinal mucosa (75, 78) , and we have found that TLR4 expression is increased in experimental and human necrotizing enterocolitis (32) . Sensitization of the intestinal mucosa through upregulation of TLRs also occurs in other diseases of intestinal inflammation, including inflammatory bowel disease and intestinal celiac disease (75, (79) (80) (81) , suggesting a potential role in the injury response. In seeking to further understand the role of enterocyte TLR4 in the pathogenesis of intestinal inflammation, our laboratory recently has examined the role of enterocyte TLR4 activation in the pathogenesis of necrotizing enterocolitis (NEC) (32) . NEC is the leading cause of death from gastrointestinal disease in preterm infants (71) , and, currently, is one of the leading causes of death of newborns in the United States overall with a mortality rate of nearly 15% (82) . We have established recently that enterocyte TLR4 activation plays a critical role in the pathogenesis of NEC (32) . Specifically, we found that NEC in both mice and humans is associated with increased expression of TLR4 in the intestinal mucosa, and that physiological stressors associated with NEC development, namely exposure to LPS and hypoxia, sensitize the murine intestinal epithe-lium to LPS through upregulation of TLR4 (32) . In support of a critical role for TLR4 in the development of NEC, TLR4-mutant C3H/HeJ mice were protected from the development of NEC compared with wild-type C3H/HeOUJ littermates (32), a finding consistent with previous work by Caplan et al. (33) . TLR4 activation in vitro led to increased enterocyte injury by induction of enterocyte apoptosis and reduced epithelial healing, due to an inhibition of enterocyte migration and proliferation. This latter finding suggests a role for enterocyte TLR4 in the regulation of intestinal mucosal repair. In support of this possibility, increased NEC severity in wildtype C3H/HeOUJ mice resulted from increased enterocyte apoptosis and reduced enterocyte restitution and proliferation compared with TLR4-mutant mice. TLR4 signaling also led to increased serine-phosphorylation of intestinal focal adhesion kinase (FAK), a molecule necessary for efficient enterocyte migration. Surprisingly, TLR4 coimmunoprecipitated with FAK in enterocytes, and siRNA-mediated FAK inhibition restored enterocyte migration after TLR4 activation, demonstrating that the FAK-TLR4 association regulates intestinal healing. Taken together, these findings demonstrate a critical role for TLR4 signaling in the intestinal epithelium in the development of NEC through effects on enterocyte injury and repair (32) . In addition to the effects of enterocyte TLR4 activation on the regulation of intestinal injury and repair, our group also has demonstrated a surprising role for enterocyte TLR4 in the regulation of bacterial translocation across the intestinal barrier (see Figure 1A ). Translocation of bacteria across the intestinal barrier is important in the pathogenesis of not only intestinal inflammation, but also systemic sepsis, and may be a critical determinant of the development of multisystem organ dysfunction. We recently have shown that enterocyte TLR4 plays a key role in regulating the ability of enterocytes to internalize Gram-negative bacteria into membrane-bound phagosomes. Further evidence that TLR4 signaling is both necessary and sufficient for phagocytosis by epithelial cells was found as cultured enterocytes were able to internalize LPS-coated but not uncoated latex particles, and MD2/TLR4-transfected HEK-293 cells acquired the capacity to internalize E. coli, whereas non-transfected HEK-293 and HEK-293 transfected with dominant negative TLR4 bearing a P712H mutation did not. Strikingly, the internalization of Gram-negative bacteria into enterocytes in vivo and the translocation of bacteria across the intestinal epithelium to mesenteric lymph nodes were significantly greater in wild-type mice as compared with mice with mutations in TLR4 (27) . These data suggest a novel mechanism by which bacterial translocation occurs, and suggest a critical role for TLR4 in the phagocytosis of bacteria by enterocytes in this process. The work reviewed above indicates that activation of TLR4 within the intestine is deleterious to the host, through effects on intestinal barrier injury, repair, and bacterial translocation. The overriding concept that enterocyte TLR4 activation has negative effects on intestinal homeostasis is supported by work demonstrating that TLR4 plays an important role in protecting the host from the development of chemical-induced colonic inflammation through the maintenance of intestinal homeostasis and the production of cytoprotective factors (83) (84) (85) . However, subsequent studies have demonstrated that TLR4 may play a permissive role in the development of spontaneous colonic inflammation (86) , suggesting either that the net effects of TLR4 on intestinal inflammation are dependent on the specific disease process examined, the anatomic location of the disease process, or that the interaction with various downstream effectors influences the extent of intestinal inflammation that develops. It is noteworthy that the inflammation observed in NEC is predominantly localized to the small intestine as opposed to the colon (4,87), implying that the effects of TLR4 activation within small intestinal epithelial cells may lead to different effects than its role on the colonic epithelia. In support of this concept, it has been demonstrated previously that small intestinal enterocytes are more responsive to LPS than colonic enterocytes, due in part to differences in TLR4 expression and/or activity (88, 89) . Moreover, the increase in expression of TLR4 within the ileum that we have observed after exposure to hypoxia and endotoxin suggests that TLR4dependent signaling within the small bowel mucosa may be increased after exposure to these stressors. The combined effects of the enhanced baseline sensitivity of the small intestine to LPS, and the upregulation of TLR4 expression in the intestine may partially explain the observed effects of enterocyte TLR4 in the induction of NEC. In support of this possibility, Caplan et al. have recently demonstrated that TLR4 expressing mice are more susceptible to the development of NEC in a model of formula feeding and cold asphyxia through a mechanism involving the enhanced interaction with luminal bacteria (33) . In addition to TLR4, other TLRs have been shown to play a role in the pathogenesis of intestinal inflammation, poten-tially via TLR-dependent signaling of the enterocytes themselves. For instance, both TLR2-/-and TLR9-/-mice were found recently to develop more severe intestinal inflammation compared with wild-type counterparts (90, 91) . Moreover, TLR5-/-mice have been found to develop spontaneous colitis (92) and the TLR5 ligand flagellin has been found to protect against enterocyte apoptosis (93) . These findings indicate that TLR2, TLR5, and TLR9 may exert protective roles in the pathogenesis of intestinal inflammation, or indeed may provide support for the maintenance of intestinal homeostasis. Since TLR2, TLR5, and TLR9 share the downstream mediator MyD88, it is possible that these studies provide mechanistic insights into the protective role of MyD88 in the maintenance of intestinal homeostasis as identified by Medzhitov et al. (83) . Once again, though the story is more complicated than appears on first glance, as activation of TLR3, the only TLR family member that does not require MyD88 to signal, with the specific ligand polyinosinic:polycytidylic acid (poly I:C) protected against the severity of DSS-induced colitis (94) . How do we reconcile the apparently contradictory roles of TLRs in the development of intestinal inflammation? It is possible that there may be cross talk between various TLR family members in the maintenance of intestinal inflammation, and the balance between intestinal homeostasis versus intestinal injury may be a reflection of the relative balance between TLRs and their associated signaling molecules. Alternatively, different TLRs within the intestine may be more or less susceptible to upregulation by different physiological stressors. We and others also have shown that intestinal mucosal TLR expression varies in different parts of the GI tract (SC Gribar and DJ Hackam, unpublished report) (95), which could explain in part the regional effects of TLR signaling on intestinal inflammation that is observed. It also may be possible that unique, epithelialspecific, intracellular signaling networks are activated by specific TLR ligation in enterocytes. Additional studies designed to delineate the precise interaction between the various enterocyte TLRs and their downstream receptors are required to resolve these possibilities. Further insights regarding a potential role for TLR signaling in the pathogenesis of intestinal inflammation may be learned from studying genetic polymorphisms in humans with diseases of intestinal inflammation and sepsis. The TLR4 Asp299Gly mutation is known to render TLR4 hyporesponsive to endotoxin (99) . This mutation has been associated with an increased incidence of inflammatory bowel disease (ulcerative colitis and Crohn's disease) (100, 101) . Furthermore, pancolitis, the most severe manifestation of ulcerative colitis, is more common in patients with the TLR1 Arg80Thr polymorphism and the TLR2 Arg753Gly polymorphism (102) . In patients with Crohn's disease, the TLR1 Ser602Ile polymorphism is associated with a reduced risk of developing ileal disease (102) . While no genetic polymorphisms have been associated with NEC, further study is necessary as the current observations were made on small cohorts of patients (103) . Pulmonary inflammatory diseases represent a broad spectrum of conditions that include allergic asthma, acute lung injury and acute respiratory distress syndrome (ARDS), chronic obstructive pulmonary disease (COPD), and infectious pneumonia (96 has a recent review). Although these diseases traditionally have been considered to reflect the combined effects of activation of the adaptive immune system with the release of antibodies and mobilization of host immune cells, recent evidence has demonstrated an important role for the TLR family members of the innate immune system in their pathogenesis. And, in parallel with the mechanisms leading to the development of mucosal inflammation in the intestine, an emerging body of literature now provides evidence that epithelial TLR signaling plays a central role (15) . Previous authors have shown that the pulmonary epithelium expresses a variety of TLRs (see Table 3 ), suggesting their role in the pathogenesis of pulmonary inflammation. In support of a role for TLR signaling in the pulmonary epithelium in the development of pulmonary inflammation, Noulin et al. have demonstrated that TLR4 and MyD88dependent signaling are required for the bronchoconstriction, cytokine response, protein leak, and neutrophil recruitment observed in response to inhaled endotoxin (28) . Furthermore, using MyD88-/bone marrow chimeras, Noulin et al. demonstrated that both resident and hematopoietic cells are necessary for the mucosal inflammatory response to inhaled endotoxin (28). Hajjar et al. demonstrated that MyD88-deficient mice transplanted with bone marrow from MyD88-expressing mice showed reduced chemokine production compared with MyD88 expressing mice that were transplanted with MyD88-expressing bone marrow in a model of experimental Pseudomonas aerogeninosa pneumonia, indicating a requirement for resident pulmonary parenchymal cells in the response to experimental pneumonia. The local pulmonary cytokine response was predominately dependent on competent MyD88 signaling in bone-marrow derived cells, suggesting that collaboration between local parenchymal cells, including epithelial cells, and bone-marrow derived cells is required (29) . In a model of bacterial pneumonia that utilizes inhaled LPS, the uptake of LPS was observed in bronchial epithelial cells and was associated with increased TLR2 and TLR4 expression in the bronchial epithelium (97) . Similarly, in an equine model of recurrent airway obstruction associated with inhaled endotoxin-rich stable dust, increased epithelial expression of TLR4 was observed and was associated with increased IL-8 expression by the airway epithelium (98) . Taken together, these reports provide supportive evidence for an important role for epithelial TLR signaling in the pathogenesis of mucosal inflammation in the pulmonary system. Several groups have shown that airway epithelial cells (AEC) express TLRs and secrete cytokines in response to TLR activation. AECs have been shown to express TLR2 and TLR4 and release IL-8 in response to Streptococcus pneumoniae, lipoteichoic acid, and lipopolysaccharide (99, 100) . Furthermore, TLR9 activation in bronchial epithelial cells has been shown to potentiate IL-8 release from bronchial epithelial cells (101) . Although it has been shown that multiple TLRs may signal in the airway epithelium (15) , microarray analysis of the lung has revealed that TLR4 signaling accounts for 74% of the pulmonary response to experimental Klebsiella pneumoniae pneumonia by comparing the pulmonary response in wildtype mice to C3H/HeJ TLR4 mutant mice. The particular TLR4-dependent responses included genes that are involved in cytokine and chemokine induction, neutrophil activation and recruitment, growth factor receptors, and TLR adaptor molecules (102) . In addition to the evidence for TLR signaling in pulmonary epithelial cells in vitro, a variety of studies have shown that TLR activation may lead to the development of pulmonary inflammation in vivo. For instance, the TLR4 mutant strains C3H/HeJ and C57BL/10ScCr showed reduced clearance of pulmonary H. influenzae and E. coli (103, 104) . In an experimental Chlamydia pneumoniae pulmonary infection model, both TLR4 and TLR2 were found to be required for survival (105), while TLR4 and CD14 were found to play an important role in the response to respiratory syncytial virus (RSV) infection (106) . In response to pulmonary Streptococcus pneumoniae infection, TLR2-deficient mice revealed only modestly reduced inflammatory response and unchanged bacterial clearance (107) compared with wild-type counterparts. TLR3-deficient mice developed a survival advantage compared with wild-type mice, as well as reduced expression of IL-6 in the bronchoalveolar fluid in a murine model of influenza A virus infection (108) . Additional evidence implicating a role for TLR signaling in the development of pulmonary inflammation may be found in studies examining the development of pulmonary inflammation in human patients with TLR polymorphisms. For instance, polymorphisms in TLR4 (Ala299Gly and Thr399Ile), which are known to lead to hyporesponsiveness to LPS (187), lead to a marked resistance to infection with Legionella pneumophila (188) . These TLR4 mutations have been correlated with the development of severe RSV infection in infants (120) . An inactivating polymorphism in TLR5 (TLR5392STOP) that encodes a stop codon in the ligand binding domain of TLR5 is associated with an increased susceptibility to infection with Legionella pneumophila causing Legionnaire's disease (188) . Taken in aggregate, the results of these in vitro and in vivo studies provide evidence for a role for TLR signaling in the pathogenesis of pulmonary inflammation. Additional studies are required utilizing pulmonary-specific TLR deletions to further delineate the relative contributions of pulmonary epithelial cell versus infiltrating leukocytes in the development of mucosal inflammation in the lung. Akin to the gastrointestinal and pulmonary tracts, dysregulated epithelial signaling in the genitourinary system may lead to marked organ dysfunction. The expression of multiple TLRs within the urinary epithelium has now been established, suggesting the possibility that TLR signaling may regulate the interaction of the urinary epithelium with potential pathogens (see Table 3 ). TLR signaling within urinary tract epithelial cells leads to pro-inflammatory signaling in response to uropathogenic E. coli (UPEC) and LPS (109, 110) . Furthermore, modulation of uroepithelial inflamma-tion may be mediated by sensitization of the uroepithelium through regulation of epithelial TLR expression in response to infection or injury. An increase in TLR4 expression in the urinary epithelium has been observed during systemic sepsis in a murine model of cecal ligation and puncture (111) , and an increase in renal epithelial TLR2 and TLR4 expression has been observed in a murine model of local renal inflammation induced by ischemia (112) . Further demonstrating a role for TLR activation in uroepithelial inflammation, TLR4 mutant C3H/Hej mice failed to clear uropathogenic E. coli (UPEC) and showed reduced inflammatory mediator production compared with wild-type controls (109) . TLR4 mutant C3H/Hej mice were resistant to LPSinduced renal failure, had less renal neutrophilic infiltrate, and less renal cell apoptosis compared with wild-type controls (113) . In addition to TLR4, other TLRs may participate in the development of uroepithelial inflammation. For instance, TLR5-deficient mice were found to be more susceptible to experimental UPEC urinary tract infection compared with wild-type counterparts (16) , while mice with null mutations in TLR11, which is normally found to be strongly expressed in the bladder and kidney epithelium, developed markedly less severe kidney inflammation compared with wild-type counterparts (114) . TLR2-deficient mice were protected from tubular injury and renal function deterioration in a model of kidney ischemiareperfusion (115) . The clinical significance of a role for TLR signaling in the pathogenesis of genitourinary inflammation is found in clinical studies in which the incidence of acute rejection after kidney transplantation is reduced in patients who received a graft heterozygous for either the TLR4 Asp299Gly or Thr399Ile polymorphism compared with grafts without these mutations (189) , although conflicting results have been reported (190) . Taken together, these studies suggest an important role for TLR signaling in the development of urinary tract inflammation in a variety of models. Which cells are required for the development of TLR-induced inflammation in the urinary tract? Evidence suggests that both epithelial and non-epithelial cell types may play a role. For instance, when TLR4 mutant C3H/Hej mice were transplanted with wild-type hematopoietic cells, the mice were unable to mount the necessary response to UPEC (30) . By contrast, in a model of cisplatin-induced renal injury, the development of inflammation was dependent on competent TLR4 signaling in resident renal parenchymal cells, as demonstrated in the study of TLR4-/-bone marrow chimeras (31) . Additional work is required to define more accurately the relative roles of TLR signaling within the epithelium versus the leukocytes in the development of mucosal inflammation in the epithelial tract. The information reviewed above highlights the important roles that TLRs play in the regulation of the inflammatory response at mucosal surfaces. However, it is well known that these mucosal surfaces are constantly bathed in bacteria, and yet appear to mount little, if any, inflammatory response. These observations lead to the question, "What controls the activation of TLRs during basal states, and what leads to their activation during inflammatory conditions?" While a complete answer to this question remains lacking, current evidence suggests that the regulation of TLR activity occurs through altering TLR or co-receptor expression, TLR localization, TLR polarity, or signaling intermediate or negative regulatory protein expression, as described below. Regulation of epithelial TLR expression has been suggested as a mechanism for the regulation of epithelial cell responsiveness in the setting of commensal bacterial exposure and during disease states. For instance, low expression of TLR4, as has been observed in colonic biopsies from humans (75) , has been suggested as a mechanism for colonocyte LPS hyporesponsiveness. Similar findings of low TLR4 expression have been observed in colonocyte cell lines (HT-29, SW480, colo205) and increased TLR4 expression after IFN-γ or TNF-α priming, as may occur during inflammatory states, has been shown to enable LPS responsiveness (88) . Similarly, low expression of TLR2 has been implicated in bronchial epithelial cell hyporesponsiveness to Gram-positive bacteria (15) . Expression of TLR co-receptors in the epithelial cells also may play a role in epithelial TLR responsiveness. Hyporesponsiveness to LPS in colonic epithelial cell lines (Caco-2, T84, SW837, and in the basal state is associated with low or absent expression of the TLR4 coreceptor MD-2 (88, 191) and priming of cultured colonocytes (HT-29) with IFN-γ or TNF-α enabled LPS responsiveness in a mechanism that involved increased MD-2 expression (192) . Recently, we also have demonstrated a transient increase in the expression of the LPS co-receptor CD14 in enterocytes after exposure to LPS (193) . In the pulmonary system, absent expression of the TLR2 coreceptor CD36 has been implicated in the hyporesponsiveness of bronchial epithelial cell to Gram-positive bacteria (194) . Changes in the subcellular localization of TLRs also may play a role in their responsiveness. Dissimilar to plasma membrane localized TLR4 in macrophages, TLR4 has been shown to be localized predominately in the Golgi apparatus in enterocytes (174) , and TLR4 activation in enterocytes has been shown to require intracellular recognition of LPS in the Golgi apparatus and recruitment of IRAK-1 and MyD88 to the Golgi apparatus (195) . Furthermore, colonic HT-29 and colo205 cells express TLR4 predominately in cytoplasmic fractions and are hyporesponsive to LPS in basal states (89) . Intracellular TLR redistribution has been suggested as a mechanism for flagellin tolerance, as prolonged flagellin exposure resulted in redistribution of TLR5 to an intracellular location in T84 colonocytes. Increased cell surface TLR expression also has been suggested as a mechanism of increased TLR sensitivity. Colonic SW480 cells are LPS-responsive and express TLR4 on the cell surface, where LPS internalization is not necessary for TLR4-LPS interaction (89) . Also, increased TLR9 surface expression was noted in response to DNA from pathogenic bacteria in HT-29 colonocytes (196) . Epithelial cell polarity and differential localization of TLRs on the apical and basolateral cell surface also has been shown to play a role in TLR sensitivity as the apical surface of epithelial cells is more likely to encounter bacteria in the normal state, whereas the basolateral surface of epithelial cells may be more likely to encounter TLR ligands only in states of disease. In support of this concept, TLR4 and TLR2 are expressed at the apical pole of T84 cells and redistribute to a cytoplasmic compartment near the basal pole with activation (77) . Furthermore, differential TLR9 signaling has been shown in colonic HCA-7 epithelial cells. Apical TLR9 activation leads to attenuation of activation of NF-κB pathways, whereas basolateral TLR9 activation leads to activation of NF-κB signaling (91) . In the pulmonary system, TLR2 was located at the apical pole of airway epithelial cells, and increased surface expression was observed in response to bacteria, whereas TLR4 was noted predominately in a basolateral location (160) . In a recent finding by Soong et al., TLR2 became enriched in lipid rafts on the apical surface after bacterial infection in airway epithelial cells, suggesting a role in the regulation of TLR sensitivity (197) . The regulation of signaling intermediate molecules also may affect TLR sensitivity. For instance, although fetal intestinal cells are known to be responsive to LPS, postnatal endotoxin hyporesponsiveness of enterocytes has been observed, and recently shown to be due to a decrease in the expression of the TLR4 signaling intermediate, IRAK1 (198) . Negative regulatory molecules may play a role in regulating epithelial TLR signaling, including peroxisome proliferatoractivated receptor-γ; the cytoplasmic zinc finger protein, A20; and the negative regulator of TLR signaling, IRAK-M, as has been reviewed recently (199) . The relevance of these molecules to signaling within epithelial cells remains to be definitely demonstrated. Given the importance of TLR signaling to the development of mucosal inflammation, it is understandable that a great deal of interest exists in the development of agents that can interfere with TLRsignaling pathways. Such an antiinflammatory approach may have particular relevance in the case of epithelial inflammation, due to ready access of the gastrointestinal, pulmonary, and urinary mucosa through ingestion, inhalation, or instillation via catheter delivery methods. Considerable attention has been placed on developing agents that are capable of modulation of the TLR4mediated response, in particular through manipulation of the lipid A moiety of LPS. Such lipid A mimetics, termed aminoalkyl glucosaminide phosphates (AGPs), have been demonstrated to reduce inflammation in experimental models of systemic sepsis induced by intravenous injection of Listeria monocytogenes (116) , pulmonary infection after intranasal administration of influenza virus (116) , murine models of colitis including DSS-induced colitis (117) , and spontaneous colitis in multidrug resistance gene 1a-deficient mice (117) . In parallel studies, soluble TLRs may reduce TLR signaling by binding to circulating ligands, rendering them unable to initiate pro-inflammatory signaling. Brandl et al. demonstrated that the synthetic molecule "LPS-Trap" was capable of blocking LPS-mediated macrophage activation in vitro by fusing MD-2 to the C-terminus of a soluble form of TLR4 (118) . Iwami et al. cloned an alternatively spliced soluble murine TLR4 (smTLR4) that, when transfected into murine macrophages, was secreted and inhibited LPS-mediated macrophage NF-κB activation and TNF-α release in vitro (119) . In addition, TLR-signaling intermediates have been targeted chemically to minimize the host inflammatory response. The synthetic peptide-mimetic compound ST2825 prevents MyD88 homodimer formation leading to an inhibition in MyD88dependent signaling, and prevents TLR9dependent inflammation in vivo (120) . A cyclohexene derivative, TAK-242, prevents TLR4 activation and was found to reduce the cytokine response to endotoxemic shock in mice (121, 122) . The Vaccinia virus protein A52R also reduces the TLR-mediated response by interacting with IRAK2 and TRAF6, and was found to reduce the cytokine response in an animal model of infectious otitis media (123, 124) and to increase survival in models of endotoxemia (125) . As the effects of TLR signaling in hematopoietic cells, as well as epithelial cells, are more clearly defined, manipulation of TLR signaling may play a larger role in the treatment of patients with diseases of inflammation, and, as evidence continues to mount suggesting a protective role for particular TLR signaling, directed and specific TLR activation may hold therapeutic promise. Mucosal surfaces and the epithelial cells that line them are constantly exposed to potential pathogens. The evidence reviewed above suggests that the innate immune system, comprised of TLRs and their associated molecules, plays a pivotal role in the regulation of mucosal inflammation in response to invading pathogens. However, the very fact that these mucosal surfaces are bathed in potential pathogens as part of their daily existence and yet don't develop inflammation under normal conditions raises an important scientific ques-tion: When does epithelial TLR signaling within mucosal surfaces become pathological? Or stated differently, when is the balance tipped between "physiological" signaling and "pathological" signaling in favor of a pathological response? A definitive answer to this important question not only is necessary to fully elucidate the steps required for the development of mucosal inflammatory diseases, but is central for the design of effective anti-inflammatory strategies. Our current thinking in this area based upon our work and the work of others is shown in Figure 3 , in which the intensity of TLR signaling within the epithelium varies depending upon the prevailing degree of systemic stress. Under basal conditions, epithelial-bacterial interactions that may occur via TLRs are likely to play roles in the regulation of processes that regulate barrier integrity, such as epithelial migration, proliferation, and apoptosis ( Figure 3A ). However, during states of systemic stress, such as hypoxia or remote infection, we submit that the extent of TLR signaling within the epithelium becomes exaggerated in response to PAMPS and DAMPS that are encountered. This "tips the balance" in favor of mucosal barrier disruption, and adversely affects mucosal repair while worsening mucosal injury ( Figure 3B ). The extent of inflammation that develops within the local microenvironment likely is compounded further by the contribution of TLR activation on leukocytes, and the release of proinflammatory molecules. Under conditions in which the balance of TLR signaling within the epithelium can be "tipped back" to a homeostatic state, mucosal inflammation may not develop. By contrast, when the extent of TLR signaling is persistent, we propose that a "feedforward" loop develops within the mucosa, resulting in persistent TLR signaling, cytokine release, and mucosal inflammation. The evaluation of the factors that maintain the degree of TLR signaling within the mucosa in the maintenance of homeostasis and the pathogenesis of disease is a topic of intensive investigation. The importance of mucosal inflammation as a clinical problem is well accepted; however, the molecular and cellular signaling pathways that lead to its development remain incompletely understood. Although much attention has been placed on the role of the epithelium as a target in the mucosal inflammatory cascade, recent evidence has shed light upon the critical role that the epithelium itself, signaling in part through Toll-like receptors, may play in the initiation of a pro-inflammatory cascade in response to external stimuli. The field of mucosal inflammation research is likely to be advanced significantly through success in the following areas of study: 1) What are the relative roles of TLR signaling within the epithelium versus circulating leukocytes in the pathogenesis of mucosal inflammation? 2) What is the precise trigger for TLR signaling within the epithelium that adversely affects the host, and what are the essential roles played by mucosal TLRs in the maintenance of mucosal homeostasis? 3) Are there TLR-signaling molecular intermediates that differ between epithelial cells and leukocytes, and do such molecules confer epithelial-specific responses in the development of mucosal inflammation? 4) What regulates the interplay between the epithelium and the other cellular constituents of the mucosa, including neurons, endothelial cells, and endocrine cells during TLR activation? It is our belief that by addressing these important questions, one can be optimistic for the development of novel classes of anti-inflammatory strategies aimed specifically at the treatment of these devastating diseases of mucosal inflammation. DJH is supported by grant R01GM078238-01 from the National Institutes of Health and the State of Pennsylvania Tobacco Settlement Fund. SCG is supported in part by the Loan Repayment Program for Pediatric Research of the National Institutes of Health and a Resident Research Award from the American College of Surgeons. WMR is supported in part by the Loan Repayment Program for Pediatric Research of the National Institutes of Health.

Ubiquitination Is Required for Effective Replication of Coxsackievirus B3

BACKGROUND: Protein ubiquitination and/or degradation by the ubiquitin/proteasome system (UPS) have been recognized as critical mechanisms in the regulation of numerous essential cellular functions. The importance of the UPS in viral pathogenesis has become increasingly apparent. Using murine cardiomyocytes, we have previously demonstrated that the UPS plays a key role in the replication of coxsackievirus B3 (CVB3), an important human pathogen associated with various diseases. To further elucidate the underlying mechanisms, we examined the interplay between the UPS and CVB3, focusing on the role of ubiquitination in viral lifecycle. METHODOLOGY/PRINCIPAL FINDINGS: As assessed by in situ hybridization, Western blot, and plaque assay, we showed that proteasome inhibition decreased CVB3 RNA replication, protein synthesis, and viral titers in HeLa cells. There were no apparent changes in 20S proteasome activities following CVB3 infection. However, we found viral infection led to an accumulation of protein-ubiquitin conjugates, accompanied by a decreased protein expression of free ubiquitin, implicating an important role of ubiquitination in the UPS-mediated viral replication. Using small-interfering RNA, we demonstrated that gene-silencing of ubiquitin significantly reduced viral titers, possibly through downregulation of protein ubiquitination and subsequent alteration of protein function and/or degradation. Inhibition of deubiquitinating enzymes apparently enhances the inhibitory effects of proteasome inhibitors on CVB3 replication. Finally, by immunoprecipitation, we showed that coxsackieviral polymerase 3D was post-translationally modified by ubiquitination and such modification might be a prerequisite for its function in transcriptional regulation of viral genome. CONCLUSION: Coxsackievirus infection promotes protein ubiquitination, contributing to effective viral replication, probably through ubiquitin modification of viral polymerase.

Coxsackievirus B3 (CVB3), a small RNA virus in the picornaviridae family, is an important human pathogen associated with various diseases, including myocarditis, aseptic meningitis, pancreatitis and possibly insulin-dependent diabetes. We and others have shown that CVB3 infection leads to activation of several intracellular signaling pathways [1] [2] [3] [4] [5] [6] [7] , and downregulation of host proteins likely through the ubiquitin/proteasome system (UPS) [7] [8] [9] . It is well-established that the UPS is the major intracellular proteolytic system of all eukaryotic cells [10, 11] . The ATPdependent system begins with covalent attachment of ubiquitin to the ubiquitin-activating enzyme (E1). Then the ubiquitin is transferred to a ubiquitin-conjugating enzyme (E2). Finally, ubiquitin ligase (E3) transfers the ubiquitin to the substrate protein. After several cycles of ubiquitination, multiple ubiquitin molecules are attached to the substrate which is then quickly recognized and subsequently degraded by the 26S proteasome. Ubiquitin is recycled through the actions of deubiquitinating enzymes (DUBs) [12, 13] . There are at least two classes of deubiquitinating enzymes, the ubiquitin C-terminal hydrolases (UCHs) and ubiquitin-specific processing proteases family. In addition to the degradation of mutant, damaged and misfolded proteins, this system is responsible for the modulation of many regulatory proteins such as cyclins [14] , inhibitors of cyclin-dependent kinases (p21, p27) [15] , tumor suppressors (p53) [16] , and inhibitor of NFkB (IkB) [17] , which are essential for a variety of cellular functions, including cell-cycle regulation, apoptosis and host immune responses [18] . Unlike polyubiquitination in the regulation of protein degradation, monoubiquitination of cellular proteins, such as histones, calmodulins, actin, proliferating cell nuclear antigen and receptor tyrosine kinases, plays more diversified roles involving in the regulation of chromatin remodeling, DNA repair, transcriptional regulation and endocytosis [19] . Since the first discovery that human papillomavirus protein E6 targets the cellular tumor suppressor protein p53 for the UPSmediated degradation [16] , increasing studies, including those from our laboratory, have suggested that various viruses evolve different mechanisms to utilize or manipulate the host UPS for their own benefits [9, [20] [21] [22] [23] [24] [25] . We have previously shown that CVB3 infection results in downregulation of several host proteins [7, 9] , such as cell-cycle protein cyclin D1, tumor suppressor p53, and transcription activator b-catenin in infected HeLa cells. The downregulation of host proteins following CVB3 infection is most likely through the UPS. Specific inhibitors to 26S proteasome reverse the degradation of proteins in HeLa cells [7, 9] and reduce CVB3 replication in murine cardiomyocytes [26] . In this study, we investigated the possible underlying mechanisms by which the UPS regulates CVB3 replication. We demonstrated that protein ubiquitination was enhanced after coxsackievirus infection. We further showed that knockdown of ubiquitin expression by small-interfering RNA (siRNA) decreased CVB3 infection, likely through the downregulation of ubiquitination and subsequent alteration of protein function and/or degradation. In addition, we showed that inhibition of deubiquitinating enzyme increased the inhibitory effects of proteasome inhibitors on CVB3 replication. We also found that CVB3 RNA-dependent RNA polymerase 3D (3D pol ) was modified by ubiquitination. Taken together, our study suggests an important role of ubiquitination in the regulation of coxsackieviral replication. To uncover the underlying mechanisms of the antiviral activities of proteasome inhibitors, we chose to use the well-characterized HeLa cells to further our study. We first examined the role of proteasome inhibition in CVB3 replication. As shown in Fig. 1 , we found that proteasome inhibitor, MG132, significantly reduced CVB3 viral RNA synthesis (Fig. 1A) . Both proteasome inhibitors used in the study, MG132 and lactacystin, decreased the synthesis of CVB3 capsid protein, VP1, in a dose-dependent manner (Fig. 1B ). In addition, two inhibitors inhibited CVB3 viral titers by up to fifteen folds (Fig. 1C) . Although MG132 and lactacystin significantly inhibited cellular 20S proteasome activities, we have previously demonstrated there was no apparent difference in proteasome activities between CVB3-infected and sham-infected HeLa cells [9] . Together, these results suggest that efficient replication of CVB3 requires the intact UPS function rather than the core proteasome activity alone. We also performed cell viability assay and morphological examination to determine whether inhibiting viral replication by proteasome inhibitors is due to the toxicity. We found that there was no measurable cell death throughout the incubation period for all doses of proteasome inhibitors used in this study (Fig. 1D) . On the contrary, virus-induced cell death was markedly inhibited after the treatment of proteasome inhibitors as a result of decreased viral replication (Fig. 1D ). As alluded to earlier, two successive steps are involved in protein degradation: (1) covalent attachment of ubiquitins to the target protein substrate, and (2) degradation of the polyubiquitinated protein by the 26S proteasome with the release of ubiquitin for recycling. To dissect out the role of ubiquitination and degradation in CVB3 infection, we next decided to investigate the protein ubiquitination after CVB3 infection. As shown in Fig. 2A , protein ubiquitination was gradually increased along the time-course of CVB3 infection, which was accompanied by a decrease of free ubiquitin levels. Densitometric analysis further demonstrated that the increases in protein ubiquitination at 3 h, 5 h, and 7 h post-infection were statistically significant as compared to sham infection (Fig. 2B) . We have previously demonstrated that 26S proteasome activities were unchanged during CVB3 infection [9] . Thus, the finding of increased accumulation of ubiquitinated proteins is likely due to enhanced protein ubiquitination as opposed to reduced proteasome activity. Decreased levels of free ubiquitin could be a direct consequence of the increased protein ubiquitination. These results suggest that enhanced ubiquitin conjugation may be a prerequisite for efficient synthesis of CVB3 viral RNA and continuation of its lifecycle. In addition to blocking proteasome proteolytic activities, proteasome inhibitors are known to reduce free ubiquitin levels in treated cells [27] . It has been suggested that proteasome inhibition negatively affects the budding of retroviruses through reducing free ubiquitin level and subsequently interfering with ubiquitination of viral Gag proteins [22, 24, 25] . Ubiquitin is generated in the cell by proteolysis of polyubiquitinated proteins or ubiquitin fused to carboxyl extension proteins (CEPs) [28] . To investigate whether protein ubiquitination is beneficial to CVB3 replication in HeLa cells, we used the ubiquitin-specific siRNA to gene-silence the expression of human ubiquitin-CEP Uba80, which codes for ubiquitin fused to ribosomal protein S27a [29] . As shown in Fig. 3A , both ubiquitin conjugates and free ubiquitin levels were markedly knocked down after the treatment of ubiquitin siRNA. We further showed that viral titers were significantly reduced in the ubiquitin siRNA-transfected cells as compared to scramble siRNA control (Fig. 3B ), suggesting that protein ubiquitination is a critical process adopted by coxsackievirus for the successful completion of its lifecycle. It has been demonstrated that protein ubiquitination can also be regulated by deubiquitinating enzymes that specifically cleave ubiquitin from ubiquitin-conjugated protein substrates [12, 13, 30] . To further explore the role of protein ubiquitination in viral replication, we examined the influence of DUB inhibition on viral protein expression. Two commercially available ubiquitin cterminal hydrolase inhibitors, UCH L1 and UCH L3 inhibitors, were used for this study. As shown in Fig. 4 , specific inhibition of UCH L1 or UCH L3 further reduced CVB3 protein expression and virus titers in proteasome inhibitor-treated cells, suggesting that these enzymes may be involved in the lifecycle of CVB3. Nevertheless, it was found that inhibition of the UCH L1 and L3 activities alone was not sufficient to block coxsackievirus replication since no significant changes in viral protein expression and CVB3 titers were observed in cells treated with two UCH inhibitors either separately or in combination (Fig. 4) . As discussed earlier, stabilization of short-lived host proteins and prevention of protein ubiquitination by reducing recycled ubiquitin likely contribute to the inhibitory effect of proteasome inhibition on viral replication. Thus, it is speculated that DUB inhibition by UCHL1/L3 inhibitors alone, in the absence of apparent inhibition of protein degradation, is not sufficient enough to block viral replication. However, additional reduction of recycled free ubiquitin by DUB inhibition can further enhance the inhibitory effect of proteasome inhibitor. CVB3 RNA-dependent RNA polymerase 3D is ubiquitinated Some virus RNA-dependent RNA polymerases including the sindbis virus and the turnip yellow mosaic virus RNA polymerases have been demonstrated to be phosphorylated and ubiquitinated [31] . Although the role of ubiquitination of these RNA polymerases in the regulation of virus replication remains to be determined, such observation raises the interesting possibility that the ubiquitin/proteasome system may regulate CVB3 replication through ubiquitinating viral polymerase 3D, which is essential for initiating viral RNA replication. To examine whether coxsackieviral proteins are subjected to ubiquitination during viral infection, we performed immunoprecipitation with anti-ubiquitin antibody, followed by immunoblots using antibodies against 3D pol and viral capsid protein VP1, respectively. As shown in Fig. 5 , immunoreactive bands of around 60 kDa were detected in CVB3infected cells. Non-modified 3D pol has a molecular weight of about 53 kDa, thus this observation suggests that 3D pol likely undergoes post-translational modification by monoubiquitination. No protein ubiquitination was observed for VP1 (data not shown). Our results implicate that the ubiquitination process of CVB3 viral proteins might be required for successful replication of the virus. In trying to understand the mechanisms by which CVB3 manipulates the UPS, we examined the protein expression of several key enzymes involved in the process of protein ubiquitination and deubiquitination. We measured expression levels of ubiquitin-activating enzyme E1A/E1B, ubiquitin-conjugating enzyme Ubc H7, ubiquitin C-terminal hydrolase and two p53related E3 ligases, human papillomavirus E6-associated protein and mouse double minute 2 homolog. However, no apparent changes were observed during the time-course of CVB3 infection (data not shown). These results indicate that the manipulation of the UPS by CVB3 is unlikely regulated by the above-examined ubiquitin-related key enzymes or molecules. Future studies will determine whether CVB3 infection targets on specific ubiquitin ligases or deubiquitinating enzymes. In the present study, we have provided further evidence that CVB3 manipulates the UPS for its infection. CVB3 infection results in increased protein polyubiquitination and a subsequent decrease in free ubiquitin levels. Knockdown of ubiquitin and ubiquitinmediated protein modification and/or degradation by siRNA It is increasingly apparent that viruses can evolve various strategies to utilize the host UPS for their own benefits. The UPS has been suggested to play a critical role in the different steps of viral lifecycle, including viral entry, viral replication, maturation, viral progeny release, and latent virus reactivation [32] [33] [34] . The mechanisms that the UPS regulates viral infection involve degrading intracellular proteins or excessive viral proteins that are against efficient viral replication and modulating viral protein function through ubiquitin-mediated modification or by directly encoding ubiquitin-related enzymes [35] . The finding in this study that CVB3 infection stimulates protein ubiquitination without inhibition of the core 20S proteasome activity highlights the possibility that CVB3 manipulates the UPS to destabilize or modulate the host and viral proteins. Polyubiquitination and degradation of host antiviral proteins has been suggested to be a mechanism of HIV-1 replication [36] . We have previously identified several proteins, such as cyclin D1, p53 and b-catenin, which are downregulated through the UPS after CVB3 infection [7, 9] . Destabilization of these short-lived host proteins is likely required for CVB3 viral RNA and protein synthesis in its lifecycle. Moreover, it is speculated that nonstructural viral proteins of CVB3 could also be potential targets of the UPS for degradation. Previous studies on picornavirus have shown that several viral proteins, such as encephalomyocarditis virus (EMCV) 3C protease and hepatitis A virus (HAV) 3C protease, are ubiquitinated and present in low concentrations in infected cells [37] [38] [39] . Several E3 ubiquitin ligases, such as human E3a ubiquitin ligase, have been shown to catalyze the ubiquitination of these viral proteins [38, 39] . Although the exact role of ubiquitination and subsequent degradation of nonstructural viral proteins of EMCV and HAV in infected cells remains elusive, such rapid turnover may be required for efficient viral RNA replication, viral protein synthesis and virus maturation. As alluded to earlier, DUBs are a large family of cysteine protease responsible for the removal of ubiquitin from substrate proteins [40] . It is estimated that the human genome encodes more than 100 DUBs. Although UCHL1 is identified as an important DUB, inhibition of UCHL1 alone has been shown to only partially block the activities of DUBs [41] . Thus, the finding in this study that UCHL1/L3 inhibition is not as efficient in blocking viral replication as general inhibition of proteasome function or knockdown of ubiquitin is likely attributed to incomplete inhibition of DUBs by UCHL1/L3 inhibitors. In addition to protein degradation, ubiquitin-modification has been suggested to be involved in the regulation of protein function. It was reported that monoubiquitination of the Gag protein of retroviruses is required for virus budding [22, 24, 25] . Depletion of free ubiquitin by proteasome inhibitors prevents Gag ubiquitination, subsequently blocks virus progeny release/budding. In addition, ubiquitination of human immunodeficiency virus type 1 Tat protein and human T-cell leukemia virus type 1 Tax protein has been shown to modulate their transactivation activities [20, 23] . We speculate that monoubiquitination is also an important machinery for post-translational modification and activation of CVB3 viral proteins. In the current study, we have shown that CVB3 RNA-dependent RNA polymerase 3D is posttranslationally modified by ubiquitination, suggesting a critical role of protein ubiquitination in the regulation of viral protein functions. Based on the results in the manuscript, in combination of our previous findings that CVB3 infection promotes host protein degradation, including cyclin D1, p53 and b-catenin, a model system on the role of the UPS in CVB3 replication is proposed in Fig. 6 . Coxsackievirus infection facilitates host protein polyubiquitination, which subsequently increases intracellular protein degradation by the proteasome and/or viral protein modification, such as 3D pol , by monoubiquitination. Degradation of host antiviral proteins provides a favorable environment for virus to achieve successful replication. Knockdown of ubiquitin decreases host protein degradation and viral protein ubiquitination. Proteasome inhibition blocks host protein degradation and viral protein ubiquitination by reducing recycled ubiquitin. DUB inhibitors further decreases the viral replication when used together with proteasome inhibitors through the additional reduction of recycled free ubiquitin. In conclusion, we have demonstrated for the first time that CVB3 infection results in increased protein ubiquitination and consequent decreases in free ubiquitin levels. We further demonstrate that protein ubiquitination is required for the completion of viral lifecycle, likely through ubiquitin modification of viral polymerase. HeLa cells (American Type Culture Collection) were grown and maintained in complete medium [Dulbecco's modified Eagle's media (DMEM) supplemented with 10% heat-inactivated newborn calf serum (NCS) (Invitrogen)]. CVB3 (Kandolf strain) was propagated in HeLa cells and stored at 280uC. Virus titer was routinely determined by a plaque assay prior to infection as described below. The monoclonal anti-b-actin and anti-ubiquitin antibodies were purchased from Sigma-Aldrich. The monoclonal anti-VP1 antibody was obtained from DakoCytomation. The ubiquitin siRNA, scramble control siRNA, and horseradish peroxidaseconjugated secondary antibodies were obtained from Santa Cruz Biotechnology. The proteasome inhibitors, MG132 and lactacystin, the UCH L1 inhibitor (LDN-57444) and the UCH L3 inhibitor (4,5,6,7-Tetrachloroindan-1,3-dione), and the polyclonal anti-ubiquitin antibody were obtained from Calbiochem. The polyclonal anti-3D pol antibody was a generous gift from Dr. Karin Klingel (University Hospital Tuebingen, Germany). HeLa cells were grown in complete medium to 70-80% confluence, and then infected at a multiplicity of infection (MOI) of 10 with CVB3 or sham-infected with phosphate-buffered saline (PBS) for 1 h in serum-free DMEM. Cells were then washed with PBS and cultured in serum-free medium. For inhibition experiments, HeLa cells were infected with CVB3 for 1 h, washed with PBS, and then incubated with DMEM containing various concentrations of inhibitors. Immunoprecipitation and immunoblot analysis Cell lysates were prepared using lysis buffer (50 mM pyrophosphate, 50 mM NaF, 50 mM NaCl, 5 mM EDTA, 5 mM EGTA, 100 mM Na 3 VO 4 , 10 mM HEPES (pH 7.4), 0.1% Triton X-100, and the protease inhibitor cocktail) as described previously [2] . For immunoblot analysis, equal amounts of protein were subjected to sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE) and then transferred to nitrocellulose membranes (GE Healthcare). Membranes were blocked for 1 h with nonfat dry milk solution (5% in PBS) containing 0.1% Tween 20. Blots were then incubated for 1 h with the primary antibody followed by incubation for 1 h with the secondary antibody. Immunoreactive bands were visualized by enhanced chemiluminescence (GE Healthcare). When protein ubiquitination was examined, membrane was heat-activated by autoclaving at 121uC for 35 min prior to blocking with nonfat dry milk solution to enhance antigenic site recognition. For immunoprecipitation, cells were lysed using the abovedescribed lysis buffer with freshly added 20 mM iodoacetamide. A total of 500 mg of cell lysates were incubated with a monoclonal anti-ubiquitin antibody (1:100) at 4uC overnight, followed by 2 h incubation with protein G-agorose beads (Amersham). Immunocomplexes were washed five times with the lysis buffer containing 20 mM iodoacetamide, and then boiled for 5 min in the 26 nonreducing sample buffer which lacks both b-mercaptoethanol and DTT, but with addition of 20 mM iodoacetamide. After centrifugation, the precipitated proteins were separated by SDS-PAGE. Ubiquitin conjugates were analyzed by immunoblot using polyclonal anti-3D pol antibody. HeLa cells were grown and maintained on two-chamber culture slides (Becton Dickinson Labware). Subconfluent cells were infected with either PBS or CVB3 (MOI = 10). Following 1 h of incubation at 37uC, cells were washed with PBS and replenished with complete medium in the absence and presence of MG132. HeLa cells were incubated for an additional 6 h. The culture slides were then washed gently with PBS, fixed with formalin buffer for 15 min, and then air-dried at room temperature. Culture slides were then subjected to in situ hybridization assays to detect the sense-strand of CVB3 genomic RNA as previously described [26] . Plaque assay CVB3 titer in cell supernatant was determined on monolayers of HeLa cells by an agar overlay plaque assay in triplicate as described previously [2] . Briefly, samples were serially diluted and overlaid on monolayer of HeLa cells. After 1 h incubation, medium was replaced with complete medium containing 0.75% agar. Cells were incubated for 72 h, then fixed with Carnoy's fixative (75% ethanol-25% acetic acid), and stained with 1% crystal violet. Plaques were counted and viral titer was calculated as plaque forming unit (PFU) per milliliter. MTS (3, 4-(5-dimethylthiazol-2-yl)-5-(3-carboxymethoxy phenyl)-2-(4-sulfophenyl)-2H-tetrazolium salt, Promega) assay was performed to determine cell viability as previously described [7] . Briefly, cells were incubated with MTS solution for 2 h prior to collection. Absorbance was measured at a wave length of 490 nm using an ELISA reader. HeLa cells were grown to 50% confluency and then transiently transfected with ubiquitin-specific siRNA (200 nM) using oligofectamine according to the manufacturer's suggestion (Invitrogen). A scramble siRNA (200 nM) was used as a control. The silencing efficiency was detected by immunoblot analyses using the antiubiquitin antibody. After 24 h of transfection, cells were infected with CVB3 as indicated. Statistical analysis was performed using the paired Student's t test. A p value of less than or equal to 0.05 was considered statistically significant.

Professional and Home-Made Face Masks Reduce Exposure to Respiratory Infections among the General Population

BACKGROUND: Governments are preparing for a potential influenza pandemic. Therefore they need data to assess the possible impact of interventions. Face-masks worn by the general population could be an accessible and affordable intervention, if effective when worn under routine circumstances. METHODOLOGY: We assessed transmission reduction potential provided by personal respirators, surgical masks and home-made masks when worn during a variety of activities by healthy volunteers and a simulated patient. PRINCIPAL FINDINGS: All types of masks reduced aerosol exposure, relatively stable over time, unaffected by duration of wear or type of activity, but with a high degree of individual variation. Personal respirators were more efficient than surgical masks, which were more efficient than home-made masks. Regardless of mask type, children were less well protected. Outward protection (mask wearing by a mechanical head) was less effective than inward protection (mask wearing by healthy volunteers). CONCLUSIONS/SIGNIFICANCE: Any type of general mask use is likely to decrease viral exposure and infection risk on a population level, in spite of imperfect fit and imperfect adherence, personal respirators providing most protection. Masks worn by patients may not offer as great a degree of protection against aerosol transmission.

With a potential influenza pandemic looming, governments need to decide how they can best use available resources to protect their people against severe illness and death, and to mitigate health and social effects for society as a whole. Much research is being devoted to develop optimal strategies for the use of (pre)pandemic vaccines and of anti-virals. There are only limited data to assess the potential effectiveness of non-pharmaceutical interventions to reduce the risk of transmission, including the effect of different kinds of face-masks worn by the general public or by patients. Respiratory infections such as influenza are transmitted through infectious particles, small enough to be suspended in air [1] . Influenza transmission can occur via large droplets, which only remain suspended in the air for a short period of time thus requiring close contact, and can occur via small airborne particles, which remain suspended in air for considerable longer periods of time, and can thus be transmitted over larger distances [2] . Furthermore, some transmission may occur via direct contact with respiratory secretions such as on hands and surfaces [2] . Interruption of transmission may allow containment of major outbreaks, like pandemic influenza. Opportunistic data collected during the SARS epidemic in Asia suggested that population-wide use of face masks may significantly decrease transmission of not only SARS but also influenza [3, 4, 5, 6, 7] . As part of pandemic preparedness, many are contemplating the contribution wide-spread use of masks could have [8, 9] . As this has major implications for resource allocation and for communication, there is great need for data to guide such decisions and make them evidence-based. Protective effects of face masks have been studied extensively, but usually this involved personal respirators for professionals under idealized conditions, because of specific applications, for instance in military or occupational uses, involving protection of specifically trained personnel. This is different from deployment of masks in the general population during an outbreak of an infectious disease, where anyone may encounter the infectious micro-organism, implying much greater heterogeneity, in training levels (experience and understanding), goodness of fit of a mask, and activities interfering with mask use and thus reducing potential reduction of transmission. The protective effect of masks is created through a combined effect of the transmission blocking potential of the material, the fit and related air leakage of the mask, and the degree of adherence to proper wearing and disposal of masks. Personal respirators such as those worn by staff attending TB patients, are used primarily to protect the wearer, and are designed to fit to the face with as tight a seal as possible. Their efficiency is graded on the degree of protection the material offers, assuming a perfect fit and optimal compliance. In contrast, surgical masks, as commonly worn in the operating theatre, are primarily used to protect the environment from the respiratory droplets produced by the wearer. With these masks, facial fit is much looser. The fit of home made masks, which could be e.g. made of a tea cloth or other comparable material available in the home, is likely to be even looser. Thus personal respirators confer a higher degree of protection than surgical masks, and these are again likely to give a higher degree of protection than home-made masks. In professional situations, ample time might be available prior to use to ensure a perfect fit and to give extensive counselling on adherence, but it is unlikely this will apply to the general population in case of a pandemic. It is possible that the discomfort in wearing associated with a certain type of masks will lead to reduced adherence and thus to a loss in overall protectiveness [10, 11] . Indeed a review among health care workers could not determine whether personal respirators conferred better protection for the health care workers than surgical masks [10] . To investigate the levels of protection, and their variation, wearing of face masks could convey to untrained subjects we designed a study in which healthy volunteers would be wearing different types of professional and home-made masks during a selection of activities, in different conditions (inward protection). We also assessed the protection different types of masks could convey when worn by a simulated infectious patient (outward protection). Resulting quantitative descriptions of distributions of protection factors may be used for assessing the importance of mask use in respiratory disease transmission. Three different experiments were undertaken to assess 1) shortterm protection for different types of masks worn during 10-15 minutes by the same volunteer following a standardized protocol, 2) long-term protection of a specific mask worn continuously by a volunteer for 3 hours during regular activities, and 3) effectiveness of different types of mask in preventing outgoing transmission by a simulated infectious subject. Inward protection was defined as the effect of mask wearing to protect the wearer from the environment; outward protection was defined as the effect of a mask on protecting the environment from the generation of airborne particles by a patient (or in this case a mechanical head). In the first short-term experiment, 28 healthy adult volunteers were recruited, as well as 11 children between 5 and 11 years of age. Each volunteer followed the same protocol wearing a Filtering Facepiece against Particles (FFP)-2 mask 1872VH (3M); which is the European equivalent of a N95 mask, a surgical mask (1818 Tie-OnH, 3M; with a filtering efficiency of around 95% for particles of sizes between 0.02 mm to 1 mm; http://jada.ada.org/ cgi/content/full/136/7/877) and a home-made mask (made of TD Cerise MultiH teacloths, Blokker). In this standard protocol, the volunteer was asked to perform five successive tasks in a fixed sequence 1.5 minute of duration each: no activity-sit still, nod head (''yes''), shake head (''no''), read aloud a standard text, stationary walk. In this sequence of activities, the respiratory rate is gradually increased. Throughout this exercise, the concentration of particles was measured on both sides of the mask through a receptor fixed on the facial and on the external side. These were connected to a portable counter of all free floating particles in the air via an electrostatic particle classifier and counter, the PortacountH. The PortacountH can register particles floating in the air with sizes between 0.02 mm to 1 mm, covering most of the size range of infectious respiratory aerosols [12] . Total inward leakage (TIL) percentage was calculated by dividing the concentrations on the outside and on the inside (TIL = (concentration inside/concentration outside)6100); the calculated quantitative protection factor was the inverse of the leakage (PF = (TIL/ 100) 21 ). To ensure small numbers of particles produced by the volunteers would not affect measurements, we checked that at least 10,000 particles per cm 3 particles of this size class (0.02 mm-1 mm) were present in the room which were produced by a number of lit candles. (Figure 1 ) In the second long-term experiment, 22 volunteers, all adults, 10 men, 12 women, were divided into 3 groups. Each group wore a single type of mask for a period of three hours, being either a FFP2 mask (4 males, 4 females), a surgical mask (3 males, 4 females) or a home-made mask (3 males, 4 females), similar to the masks used in the short-term experiment described above. At the beginning and end of each three-hour period, full series of measurements were taken using the standardised protocol as described for the short-term experiment, and during the three hour period while wearing the masks, participants reported back at regular intervals for a short measurement during rest (absence of activity). For the remainder of the period, participants carried on with their usual daily activities. During regular activities in between measurements, the probes of the masks were plugged which did not involve dislodging of the masks. In the final experiment, we assessed the effectiveness of different types of masks in reducing outgoing transmission from an infectious subject shedding aerosolised particles. This was simulated by fitting the different types of masks to an artificial test head, which was connected to PC-driven respirator (BacouH LAMA AMP, Modelref 1520307). Breathing frequency was varied to mimic different respiratory rates (15, 25 and 40/minute). Only expiration was simulated; twice for each mask at each respiratory rate. The breathing flow was defined as (respiratory rate/minute x volume per breath (2 litres)) resulting in a breathing flow of 30, 50 and 80 litres per minute, which correlates with light (walking), medium (marching with backpack) and strenuous (running) activities [13] . Concentrations of particles were measured as described above by a TSI Portacount Respirator Fit tester, model 8020, measuring outward protection, rather than inward protection. All volunteers received written information prior to the experiments and gave oral informed consent. For the children also a parent gave oral informed consent, and a parent remained present during the experiments. The Dutch Central Committee on Research Involving Human Subjects (CCMO) informed us in writing that this project did not need to be assessed by an Ethics committee. Protection factors (PF) calculated from measurements of particle concentration by PortacountH devices were reported as the ratio of particle concentrations outside and inside the mask. This is a similar concept to the fit factor as used by the US Occupational Safety and Health Administration (http://www.osha.gov/pls/ oshaweb/owadisp.show_document). Therefore, a higher PF is better and PF = 1 means complete absence of protection. For statistical analysis, the following transformation was used: The inverse of the PF (1/PF) can be interpreted as a probability (that any particle succeeds in moving through the barrier the mask provides). The logit transformation is a standard transformation to transform the probability scale (0,1) to the real axis (-infinity, +infinity) to allow standard regression techniques (including ANOVA) to test the effects of co-variables (mask type, age class, sex, activity, duration of use) on transformed PFs in a linear model, using the statistical application R (version 2.5.0). The p-values are based on testing the ratio of mean squares for a factor (like 'mask') and the mean square of errors (random fluctuations), assuming that ratio is F-distributed. Whenever the p-value (the probability of a greater value of the tested ratio) is greater than 0.05, the ratio is considered significantly different from 1 ( = indifference) at the 95% level. All masks provided protection against transmission by reducing exposure during all types of activities, for both children and adults (Table 1) . Within each category of masks, the degree of protection varied by age category and to a lesser extent by activity. We observed no difference between men and women. Surgical masks provided about twice as much protection as home made masks, the difference a bit more marked among adults. FFP2 masks provided adults with about 50 times as much protection as home made masks, and 25 times as much protection as surgical masks. The increase in protection for children was less marked, about 10 times as much protection by FFP2 versus home-made masks and 6 times as much protection as surgical masks. In these short term experiments, adjusting for covariates, face mask type had a strongly significant independent effect on protection (p,0.001). Children were significantly less protected than adults (p,0.001). There was no significant impact of activity on protection. As in the short term experiment, mask type was a strong determinant of protection (Table 2 ). Protection factors for each type of mask were similar to the protection factors measured in the short term experiments for adults. There was considerable variability between volunteers. The median protection factors measured over a 3 hour period increased for those wearing homemade masks, decreased for those wearing FFP2 masks, and did not show a consistent pattern for those wearing a surgical mask (Figure 2 ), but overall protection factors calculated per type of mask were stable over time, and did not change statistically significant with prolonged wearing. Overall, protection factors were relatively stable over time for each individual (ANOVA p = 0.4). Males and females did not have significantly different protection factors (ANOVA p = 0.9). As in the short term experiment, protection conferred by surgical masks was higher than protection given by a home-made mask, and protection provided by a FFP2 masks was again markedly higher than protection provided by a surgical mask. As in the short term experiment, more strenuous activities (reading and walking) tended to increase the protection of the home-made mask and to a lesser extent of the surgical mask, and decreased the protection by the FFP2 mask, but there was no overall significant effect of type of activity on PF (ANOVA p = 0.1). Outward protection experiment In a final experiment, retention of particles expelled inside the masks was studied. Here again, mask type was strongly correlated with (transformed) protection factors. Protection factors for all type of masks were considerably lower than those observed for inward protection. The home-made masks only provided marginal protection, while protection offered by a surgical mask and an FFP2 mask did not differ ( figure 3) . The simulated breathing frequency did not significantly affect the measured protection factors. Adjusting for covariates, mask type and particle concentration, but not flow rate, were significant factors for protection in the reverse flow experiment. In our experiments, the main determinant of the magnitude of protection factors measured by masks was the type of mask, which can be seen as a proxy for potential reduction in infectious disease transmission. The duration of wear and the type of activity did not have a significant impact on exposure reduction. Thus, the expected superior protection conferred by a professional FFP2 mask compared to a surgical mask or a home-made mask was maintained when these FFP2 masks were worn by healthy lay people in spite of the increased risk of a poor fit and significant behavioural leakage. Children were significantly less protected from exposure than adults, which might be related to an inferior fit of the masks on their smaller faces. Although we observed a high degree of individual variability in the degree of protection conferred as reflected in the wide interquartile ranges of the measured PFs, no systematic difference was found between men and women, suggesting a poorer fit only has a noticeable impact on protection when the mismatch between face and mask is considerable. All types of masks provided a much higher degree of exposure protection against inward transmission of particles, then in preventing outward transmission by a mechanical head as a proxy for an infected patient exposing the environment. Data from professional users suggest a decrease in protection over time due to a reduction in fibre charges [13] . In our data, this effect was not significantly present, although a tendency towards reduced protection over time was seen for the FFP2 masks. Also, our study showed a high degree of individual variation in exposure protection. This is important as it reflects the presence of many different sources of variation, behavioural as well as anatomical, which can also be expected to be present if the general population would be requested to wear face masks in case of a pandemic. Furthermore, we do not know from these experiments whether reduced exposure has a linear or non-linear relationship to the reduction of infection risk. Although this could imply that individual subjects may not always be optimally protected, from a public health point of view, any type of general face mask usage can still decrease viral transmission. Also, it is important not to focus on a single intervention in case of a pandemic, but to integrate all effective interventions for optimal protection. Surprisingly, the protection conferred by each of the masks appeared stable over time and was not dependent on activity. This suggests that leakage associated with suboptimal fit and compliance was stable over time. The tendency towards improved protection of the poorer fitting masks with increased activities such as reading, might be attributable to reduced leakage when breathing through the mouth rather than the nose, which could give some overpressure and thus reduce inward leakage. We had assumed that compliance would decrease during the three hours of continuous wearing, in particular with more strenuous activities. Indeed, among professionals like cullers, there have been some anecdotal reports that FFP3 masks were associated with poorer compliance than FFP2 masks in wearing. Where a reduction in protection was found with the FFP2 mask, the reverse was seen for the home-made mask. It is possible that the experimental situation, sufficient motivation to endure a relatively limited time of discomfort, and the absence of physically challenging activities, has provided more stable protection than might be found in reallife situations. However, overall these experiments show that significant protection against influenza transmission upon exposure can be conveyed also for lay people, including children, in spite of imperfect fit and imperfect adherence. It is also clear that home-made masks such as teacloths may still confer a significant degree of protection, albeit less strong than surgical masks or FFP2 masks. Home made masks however would not suffer from limited supplies, and would not need additional resources to provide at large scale. Home made masks, and to a lesser degree surgical masks, are unlikely to confer much protection against transmission of small particles like droplet nuclei, but as the reproduction number of influenza may not be very high [14] a small reduction in transmissibility of the virus may be sufficient for reducing the reproduction number to a value smaller than 1 and thus extinguishing the epidemic [15] . Greater reduction in transmissibility may be achieved if transmission is predominantly carried by larger droplets. In a typical human cough half of the droplets may be small (,10 mm), but these comprise only a small fraction (2.5*10 26 ) of the expelled volume [12] . Smaller droplets may however more easily penetrate the smaller bronchi and be more effective in transmission [1] . A more detailed analysis of aerosol and droplet inoculation and infectivity may provide better insight into the impact of either transmission mode on population spread. The difference in measured protection against inward and outward protection is remarkable, and cannot be explained from the available data as we only measured the overall effect. A differential effect on the amount of leakage seems most plausible. At the same time, we cannot exclude that wearing of face masks, even FFP2 or surgical masks by patients might still significantly reduce transmission. However, the observed limited particle retention in our experiments may still be an overestimate of protection, as it may for instance be challenging to enforce adherence to mask wearing by a patient who is short of breath. Wearing of masks by caregivers might be more feasible and more effective, in particular where additional preventive measures are in place as well for caregivers. Furthermore, we should bear in mind that this is an experimental study, with relatively small numbers of volunteers, which limits the generalisability of some of our findings. E.g., for masks to have any impact during an actual pandemic, people may need to be wearing masks during several weeks with many shorter or longer mask-free periods. Furthermore, the PFs may be an over-or underestimation of the actual protection conferred. And although our simulated patient varied its breathing frequency, we have not assessed the impact of e.g. coughing or sneezing on outward transmission through a mask. A recent analysis of the 1918 epidemic, noted that cities where strict interventions were implemented early on to prevent transmission, were overall worse-off than cities where some degree of transmission occurred early on [16] . Given the need for the population to acquire sufficient natural immunity over time, it can not be excluded that the amount of protection conferred by home made masks might sufficiently reduce viral exposure to impact on transmission during the early waves, while allowing people enough exposure to start mounting an efficient immune response. Further field studies are needed to assess acceptability and effectiveness of masks worn by people from the general population. Also, experimental data are needed to develop dose-response models which may improve understanding of determinants of transmission. A cost-effectiveness analysis might give further insights in the relative benefits of home made masks.

Oct-4 Expression Maintained Cancer Stem-Like Properties in Lung Cancer-Derived CD133-Positive Cells

CD133 (prominin-1), a 5-transmembrane glycoprotein, has recently been considered to be an important marker that represents the subset population of cancer stem-like cells. Herein we report the isolation of CD133-positive cells (LC-CD133(+)) and CD133-negative cells (LC-CD133(−)) from tissue samples of ten patients with non-small cell lung cancer (LC) and five LC cell lines. LC-CD133(+) displayed higher Oct-4 expressions with the ability to self-renew and may represent a reservoir with proliferative potential for generating lung cancer cells. Furthermore, LC-CD133(+), unlike LC-CD133(−), highly co-expressed the multiple drug-resistant marker ABCG2 and showed significant resistance to chemotherapy agents (i.e., cisplatin, etoposide, doxorubicin, and paclitaxel) and radiotherapy. The treatment of Oct-4 siRNA with lentiviral vector can specifically block the capability of LC-CD133(+) to form spheres and can further facilitate LC-CD133(+) to differentiate into LC-CD133(−). In addition, knock-down of Oct-4 expression in LC-CD133(+) can significantly inhibit the abilities of tumor invasion and colony formation, and increase apoptotic activities of caspase 3 and poly (ADP-ribose) polymerase (PARP). Finally, in vitro and in vivo studies further confirm that the treatment effect of chemoradiotherapy for LC-CD133(+) can be improved by the treatment of Oct-4 siRNA. In conclusion, we demonstrated that Oct-4 expression plays a crucial role in maintaining the self-renewing, cancer stem-like, and chemoradioresistant properties of LC-CD133(+). Future research is warranted regarding the up-regulated expression of Oct-4 in LC-CD133(+) and malignant lung cancer.

Lung cancer is one of the leading causes of cancer-related deaths in industrialized countries [1, 2] . Radiotherapy and chemotherapy play significant and crucial roles in clinical antilung cancer treatment to achieve prolonged patient survival [3, 4] . However, a high failure rate and low median survival rate are observed in patients undergoing chemoradiotherapy with recurrent, intractable lung cancer [5] . To improve the patient survival rate, investigation to elucidate the mechanism of tumorigenesis of lung cancer is needed [5] . Recent data have demonstrated that tumors contain a small subpopulation of cells, i.e., cancer stem-like cells (CSCs) or cancer-initiating cells (CICs), which exhibit a selfrenewing capacity and are responsible for tumor maintenance and metastasis [6] . Stem cells have been isolated by their ability to efflux Hoechst 33342 dye and are referred to as the ''side population (SP)'' [7] . Ho and colleagues isolated and characterized SP cells from six human lung cancer cell lines and showed that an elevated expression of ABCG2 as well as other ATP-binding cassette transporters were positively correlated with resistance to multiple chemotherapeutic drugs [8] . In addition, Gutova and colleagues have purified uPAR-positive CSCs from three lung cancer cell lines. These uPAR-positive cells co-expressed with CD44 and MDR1, and had the ability to promote advanced malignancy and chemoresistance [9] . CD133 (prominin-1), a 5-transmembrane glycoprotein, was first recognized in CD34 + progenitor populations from adult blood, bone marrow, and fetal liver cells [10] . Recently, CD133 has been considered an important marker to represent the subset population of CSCs in leukemia, brain tumors, retinoblastoma, renal tumors, pancreatic tumors, colon carcinoma, prostate carcinoma, and hepatocellular carcinoma [11] [12] [13] [14] [15] [16] [17] [18] [19] . Based on immunohistochemical findings, Hilbe and colleagues suggested that CD133-positive (CD133 + ) progenitor cells play a role in the development of tumor vasculature in patients with non-small-cell lung cancer (NSCLC) [20] . More recently, a well-designed study by Eramo and colleagues showed that lung cancer contains a population of CD133 + CSCs able to self-renew and generate an unlimited progeny of non-tumorigenic cells. These CD133 + cells are also resistant to conventional chemotherapy [21] . However, the gene regulation mechanisms in maintaining the self-renewal and drugresistant properties in putative cancer stem-like cells of lung tumors are still unclear. Oct-4, a member of the family of POU-domain transcription factors, is expressed in pluripotent embryonic stem (ES) and germ cells [22] [23] . Oct-4 mRNA is normally found in totipotent and pluripotent stem cells of pregastrulation embryos [24] . Knocking out the Oct-4 gene in mice causes early lethality due to the lack of ICM formation, indicating that Oct-4 has a critical function for self-renewal of ES cells [25] . Oct-4 activates transcription via octamer motifs, and Oct-4 binding sites have been found in various genes, including fgf 4 (fibroblast growth factor 4) and pdgfar (platelet-derived growth factor a receptor) [26, 27] . This suggests that Oct-4 functions as a master switch during differentiation by regulating the pluripotent potentials of the stem cell, and Oct-4 plays a pivotal role in mammalian development [24, 25] . In this study, the CD133-positive cells (LC-CD133 + ) and CD133-negative cells (LC-CD133 2 ) were isolated from tissue samples of lung cancer (LC) patients and LC cell lines. These LC-CD133 + cells possessed both the characteristics of stem-like cells and malignant tumors. Our data further demonstrated that Oct-4 expression in LC-CD133 + is involved in tumor malignancy of lung cancers and exhibits refractory properties for chemoradiotherapy in cancer stem-like cells. These results suggested that the expression of Oct-4 plays a crucial role in maintaining cancer stem-like and chemoradioresistant properties in lung cancerderived CD133 + cells. This research followed the tenets of the Declaration of Helsinki and all samples were obtained after patients provided informed consent. The study was approved by the Institutional Ethics Committee/Institutional Review Board of Taipei Veterans General Hospital. The dissociated cells from the samples of nonsmall cell lung cancer patients (Table 1 ) and the lung cancer (LC) cell lines were labeled with 1 mL CD133/l micromagnetic beads per 1 million cells using the CD133 cell isolation kit (Miltenyi Biotech, Auburn, CA). CD133 + cells were cultured in a medium consisting of serum-free DMEM/F12 (Gibco-BRL, Gaithersburg, MD), N2 supplement (R&D Systems Inc., Minneapolis), 10 ng/ml human recombinant bFGF (R&D Systems) and 10 ng/ml EGF (R&D Systems) [28] . The isolated CD133 + and CD133 2 cells were cultured in a 96well cell culture cluster (Corning Costar, Acton, MA) at a density of 3610 3 cells/well with 100 ml culture medium. At specific time points during cultivation, the medium was discarded and replaced with an equal volume (100 ml) of fresh medium containing 0.2 mg/ml of 3-(4,5-dimethylthiazol-2-yl)-5-(3-carboxymethoxyphenyl)-2-(4-sulfophenyl)-2H-tetrazolium (MTS, Promega, Madison, WI) and 0.038 mg/ml of phenazine methosulfate (PMS; Promega) and incubated for additional 1.5 hours in 37uC 5% CO 2 . Cell viability proportionate to optical density (OD) was measured by colorimetric assay in terms of mitochondria activity to convert tetrazolium salt into a colored soluble formazan product in the medium. The OD values were measured at the wavelength of 490 nm with a 1420 multilabel counter VICTOR from Wallac (PerkinElmer Wallac, Turku, Finland). For real-time RT-PCR analysis, the total RNA of cells was extracted by using the RNA easy kit (Qiagen, Valencia, CA). Briefly, the total RNA (1 mg) of each sample was reversely transcribed in 20 mL using 0.5 mg of oligo dT and 200 U Superscript II RT (Invitrogen, Carlsbad, CA). The amplification was carried out in a total volume of 20 ml containing 0.5 mM of each primer, 4 mM An avidin-biotin complex method was used for the immunofluorescence staining in the differentiated spheroid and neuronallike cell. In brief, cells were plated onto poly-L-ornithine-coated glass coverslips and fixed with 4% paraformaldehyde for 15 to 20 minutes at room temperature, and then were washed twice (10 minutes each) with 16 PBS. Cells were permeabilized with 0.1% Triton X-100/PBS for 10 minutes at room temperature, and then twice (10 minutes each) with 16 PBS. The cells were then blocked with blocking solution for 30 minutes and were incubated with primary antibodies (Oct-4, Chemicon, Temecula, CA) for 1 hour at room temperature. We then washed the cells three times (10 minutes each) with 16 PBS. Immunoreactive signals were detected with a mixture of biotinylated rabbit antimouse IgG and Fluorsave (Calbiochem, San Diego, CA). Cells were further probed with fluorescein isothiocyanate (FITC)tagged secondary antibodies. Fluorescence images were visualized with a fluorescence microscope. To quantitatively analyze the fluorescence intensity, we recorded images with an inverted fluorescence microscope equipped with a CCD camera. The percentage of signal fluorescence per photographed field was analyzed by image processing software (Image Pro-Plus, Media-Cybernetics, Inc., Silver Spring, MD). For cell surface marker identification, a single cell suspension of sixth-to eighth-passage cells from trypsinized spheres was stained with anti-CD133, CD117 (c-Kit), or ABCG2 and secondary fluorescein (FITC)-or phycoerythrin (PE)-coupled antibodies (Dako, Carpinteria). Cells were fixed with 2% paraformaldehyde and were analyzed with a BD FACSCalibur apparatus (Becton, Dickinson and Company, Franklin Lakes, NJ). The gamma radiation (ionizing irradiation; IR) was delivered by a Theratronic cobalt unit T-1000 (Theratronic International, Inc., Ottawa, Canada) at a dose rate of 1.1Gy/min (SSD = 57.5cm). To evaluate the cell proliferation rate we seeded cells on 24-well plates at a density of 2610 4 cells/well. Cells were seeded 24 hours after IR and then they were analyzed by methyle thiazol tetrazolium assay (MTT assay, Sigma-Aldrich, St. Louis, MN). The amount of MTT formazon product was determined by using a microplate reader and the absorbance was measured at 560 nm (SpectraMax 250, Molecular Devices, Sunnyvale, CA). Cisplatin, etoposide (VP16), and paclitaxel were obtained from Sigma-Aldrich and were dissolved in DMSO (Sigma-Aldrich) at 100 mM of stock solution. The 24-well plate Transwell system with an 8-mm pore size polycarbonate filter membrane (Corning Costar, Corning, NY) was used. The filter membrane was coated with Matrigel (BD Biosciences, San Diego) diluted with serum-free medium and incubated overnight at 37uC. The cell suspensions were seeded to the upper compartment of the Transwell chamber at the cell density of 1610 5 in 100 ml serum free medium. After 24 hours, the medium was removed and the filter membrane was fixed with 4% formalin for 1 hour. The opposite surface of the filter membrane facing the lower chamber was stained with Hoechst 33342 (Sigma-Aldrich) for 3 minutes and the migrated cells were then visualized under an inverted microscope. The protocol of soft agar colony assay is described as follows. Each well (35 mm) of a six-well culture dish was coated with 2 ml bottom agar mixture (DMEM, 10% (v/v) FCS, 0.6% (w/v) agar). After the bottom layer solidified, 2 ml top agarmedium mixture (DMEM, 10% (v/v) FCS, 0.3% (w/v) agar) containing 2610 4 cells was added, and the dishes were incubated at 37uC for 4 weeks. Plates were stained with 0.5 ml of 0.005% crystal violet for 1 hour and then a dissecting microscope was used to count the number of colonies [29] . The pLVRNAi vector and pCDH-MCS1-EF1-copGFP vector were purchased from Biosettia Inc. (Biosettia, San Diego, CA). The method of cloning the double-stranded shRNA sequence is described in the manufacturer's protocol. The siRNA oligonucleotide 59-CCGGCCCTCACTTCACTGCACTGTACTCGAGTA-CAGTGC AGTGAAGTGAGGGTTTTT-39 targeting human Oct-4 (NM_002701, nt 1035-1055) was synthesized and cloned into pLVRNAi to generate a lentiviral expression vector. The Oct-4 cDNA was amplified and purified by RT-PCR and cloned into a pCDH-MCS1-EF1-copGFP vector. Lentiviral production was done by transfection of 293T cells using Lipofectamine 2000 (LF2000, Invitrogen). Supernatants were collected 48 hours after transfection and then were filtered; the viral titers were then determined by FACS at 48 hours post-transduction. Subconfluent cells were infected with lentivirus at a multiplicity of infection of 5 in the presence of 8 ìg/ml polybrene (Sigma-Aldrich). All procedures involving animals were in accordance with the institutional animal welfare guidelines of Taipei Veterans General Hospital. 1000, 3000, and 10 4 cells were injected into the tail vein of SCID mice and/or nude mice (BALB/c strain) each aged 8 weeks. In vivo GFP imaging was visualized and measured by an illuminating device (LT-9500 Illumatool TLS equipped with excitation illuminating source [470 nm] and filter plate [515 nm]). The tumor size was measured with calipers and the tumor volume was calculated according to the formula (Length6Width 2 )/2. The integrated optical density of green fluorescence intensity was captured and then analyzed by Image Pro-plus software [29] . Statistical Package of Social Sciences software (version 13.0) (SPSS, Inc., Chicago, IL) was used for statistical analysis. The independent Student's t-test or ANOVA was used to compare the continuous variables between groups, whereas the x 2 test was applied for comparison of dichotomous variables. The Kaplan-Meier estimate was used for survival analysis, and the log-rank test was used to compare the cumulative survival durations in different patient groups. The level of statistical significance was set at 0.05 for all tests. Using the magnetic bead method, we isolated CD133 + cells (Fig. 1A) from tissue samples of ten non-small cell lung cancer (NSCLC) patients (Table 1 ) and five lung cancer (LC) cell lines ( Table S1 ). The high percentage (97%) of CD133 + (LC-CD133 + ) subset was isolated in the LC tissues and parental LC cell line (Fig. 1A) . It has been reported that cancer stem-like cells can be cultured in suspension to generate floating spheroid-like bodies (SB) under serum-free medium with bFGF & EGF [30] . We found that LC-CD133 + isolated from these ten patients (Table 1 ) and five LC cell lines (Table S1 ) can form SB in DF-12 serum-free medium with bFGF and EGF ( Fig. 1B 1C ) and proliferation rate (Fig. 1D ) in LC-CD133 + were significantly higher than that in LC-CD133 2 (p,0.05). In addition, to determine the in vivo tumorigenic activity of LC-CD133 + and LC-CD133 2 , we injected respective amounts of 1000, 3000, and 10 4 cells into the tail veins of SCID mice. The results showed that 10 4 LC-CD133 2 did not induce tumor formation but 3,000 LC-CD133 + from the lung cancer tissues of ten patients and five LC cell lines in xenotransplanted mice can all generate visible tumors 4 weeks after injection (Table 1 and Table S1 ). To characterize our isolated LC-CD133 + , FACS analysis was used to detect the expression profile of cells surface markers. As shown in Figure 2A , the majority of isolated LC-CD133 + were stained with higher expression levels of CD133, CD117 (c-Kit), and ABCG2 compared with LC-CD133 2 . This result demonstrated that isolated LC-CD133 + were almost ABCG2-positive cells (Fig. 2B) . To further evaluate the enhancement of tumorigenicity of LC-CD133 + , we examined in vitro Matrigel-combined Transwell invasion and soft agar colony formation assays. Compared with LC-CD133 2 , LC-CD133 + derived from NSCLC Patients No.1 (PLC) and No. 2 (LLC) showed higher invasion activity through Matrigel Transwell invasion assay (p,0.001; Fig. 2C ). Similarly, the foci formation ability of LC-CD133 + from PLC (No.1) and LLC (No.2) was enhanced when compared with the LC-CD133 2 of those two patients (p,0.001; Fig. 2D ). Increased In Vivo Tumor-restoration and Proliferative Ability in LC-CD133 + We further evaluated the in vivo tumor-restoration and proliferative ability of LC-CD133 + and LC-CD133 2 by xenotransplanted tumorigenicity analysis (Fig. 3A) . Four weeks after 10 4 cells were injected into the tail veins of SCID mice, a significant increase in the multiple nodules of tumor formation on lung surface was noted in the LC-CD133 + -injected group (Figs. 3A4 & 3A7) but not in the LC-CD133 2 group (Fig. 3A1) . Diffuse infiltrations of LC-CD133 + from the lung parenchyma to the alveolar cavity were observed (Figs. 3A5, 3A6, and 3A8) . The histological examination demonstrated that the prominent neovascularization and thrombus formation were detected in the pulmonary parenchyma of LC-CD133 + -injected SCID mice (Fig. 3A9) . In contrast, no significant tumor foci or neovascular formation was found in the lungs of LC-CD133 2 -injected SCID mice (Figs. 3A2 & 3A3) . We further investigated the in vivo tumor growth rate in 10 4 LC-CD133 + cells, 10 6 LC-CD133 2 cells, and 5610 6 parent tumor cells from the same patient. The finding demonstrated that the tumor growth rate of the 10 4 LC-CD133 + group (from Patients No. 1, 2, 4, and 7; Table 1 ) was significantly higher than that of the 10 6 LC-CD133 2 group and 5610 6 parental tumor cell group (Fig. 3B) . Furthermore, 10 4 LC-CD133 + isolated from secondary tumors can further generate new (second) tumors from transplanted SCID mice. Results of flow cytometry showed that a high percentage (60%) of CD133-positive cells was detected in the second tumor (Fig. 3C ). In addition, one thousand LC-CD133 + isolated from the second tumor can also generate a new (third) tumor in transplanted SCID mice (Fig. 3C) . To sum, our data indicated that LC-CD133 + present self-renewing and repopulation capabilities both in vitro and in vivo. Enhanced Chemo-and Radiation-resistance in LC-CD133 + We evaluated the multidrug (chemotherapy)-resistant abilities of LC-CD133 + and LC-CD133 2 . We further tested four common chemotherapeutic agents including cisplatin, VP16 (etoposide), doxorubicin, paclitaxel. Compared with LC-CD133 2, LC-CD133 + are significantly resistant to the four tested chemotherapeutic agents (p,0.01; Fig. 4A ). To further determine the radiation effect on the rate of tumor growth, we used an ionizing radiation (IR) dose from 0 to 10 Gy to treat both LC-CD133 + and LC-CD133 2 . As shown in Fig. 3B , after IR treatment, the survival rate and number of LC-CD133 + were significantly higher than those of LC-CD133 2 (p,0.01). We further found that the LC-CD133 + cells possess a higher degree of radioresistance (p,0.01; Fig. 4B ). Moreover, we investigated the combined treatment effect of radiochemotherapy in LC-CD133 + . Experiments were conducted with cisplatin (10 mM) alone, VP-16 (10 mM) alone, or combined cisplatin and VP-16 on IR (2 Gy)-treated LC-CD133 + . As shown in Figure 3C , the data revealed that the cell survival rate in IR-treated LC-CD133 + was not significantly decreased by the IR treatment combined with cisplatin, with or without VP-16 (p.0.05). On the contrary, cell survival significantly declined after chemotherapy with cisplatin combined with VP-16 in IR-treated LC-CD133 2 (p,0.01; Fig. 4C ). These results suggest that LC-CD133 + may play a vital role in the tumor's ability to resist radiation and chemotherapies. To investigate whether Oct-4 expression plays a role in maintaining self-renewal or cancer stem-like properties in LC-CD133 + , we used the siRNA method with lentiviral vector for knockdown of Oct-4 expression in LC-CD133 + . We found it important that the treatment of Oct-4 siRNA in LC-CD133 + can significantly interfere with the capabilities of spheroid-like bodies (SB) formation (p,0.001; Fig. 5C ). After 7 days of the Oct-4 siRNA treatment, the SB of LC-CD133 + could not maintain floating spheres but differentiated into attached epithelial-like cells (Fig. 5C ). In contrast, the treatment of scramble control siRNA did not influence the SB formation capability in LC-CD133 + (Fig. 5C) . The SB of LC-CD133 + endogenously expressed strong positive signals for Oct-4 and CD133 (Fig. 5D) . Furthermore, the immunofluorescent results demonstrated that both the CD133 and Oct-4 expressions in LC-CD133 + were significantly blocked after 7 days of Oct-4 siRNA treatment (Fig. 5D) . FASC assay confirmed that the amount of CD133 was dramatically decreased in Oct-4 siRNA-treated LC-CD133 + and the percentages of LC-CD133 2 were significantly increased in LC-CD133 + after 7 days of Oct-4 siRNA treatment (p,0.001; Fig. 5E ). These data suggested that Oct-4 may maintain the properties of primitive stem cells and inhibit the tendency for differentiation in LC-CD133 + . To further study the role of Oct-4 in tumor malignancy of LC-CD133 + in vitro, the migration/invasive and soft agar colony assay were used. The results showed that the abilities of the in vitro migratory invasion and colony formation in LC-CD133 + treated by Oct-4 siRNA were significantly decreased compared with nontreated LC-CD133 + , or LC-CD133 + treated with scramble-siRNA (control; p,0.001; Fig. 6A ). Furthermore, the treatment effect of chemoradiotherapy for the LC-CD133 + group can be significantly improved by the treatment of Oct-4 siRNA compared with nontreated LC-CD133 + or LC-CD133 + treated by scramble-siRNA ( Fig. 6B; p,0.001 ). In addition, we found that the apoptotic activities of annexin V (Fig. 6C) and caspase 3 (Fig. 6D; upper part) were quickly and effectively induced in LC-CD133 + treated by Oct-4 siRNA after 72 hours. In accordance with the result of cell survival and treatment effects in Oct-4 siRNA-treated LC-CD133 + (Fig. 6B) , the western blot data further demonstrated that the amounts of activated (cleaved) form of PARP were consistently elevated in LC-CD133 + treated by Oct-4 siRNA with IR alone or combined with chemotherapy ( Fig. 6D; lower part) . Thus, knockdown of Oct-4 expression in LC-CD133 + can effectively enhance chemoradiosensitivities and apoptotic activities in response to IR and chemotherapy, suggesting that Oct-4 could be a key factor enables LC-CD133 + to resist radiochemotherapeutic stress. To investigate the treatment effects of chemoradiotherapy on Oct-4 siRNA-treated LC-CD133 + , LC-CD133 + was first transfected by lentivector combined with green fluorescent protein gene (GFP), and then in vivo GFP imaging and histological study were used to monitor the tumor-growth effect. We first injected 10 4 LC-CD133 + -GFP cells into the subcutaneous sites of nude mice with different treatment protocols. The tumor volumes were significantly decreased in Oct-4 siRNA-treated LC-CD133 + when exposed to IR alone, cisplatin alone, or IR combined with cisplatin (p,0.01; Fig. 7A ). To further evaluate the capabilities of tumor invasion and metastasis of LC-CD133 + treated by different regimens, we injected 10 4 LC-CD133 + -GFP cells from each treatment groups into the tail vein of SCID mice. The results of in vivo GFP imaging showed that the tumor foci of lung and metastatic lesions in the Oct-4-siRNA-treated LC-CD133 + groups were significantly lower than those of the LC-CD133 + without Oct-4-siRNA-treated groups (p,0.01; Fig. 7B ). In addition, to investigate the treatment effects of Oct-4 expression in LC-CD133 + in vivo, we injected the seven groups with different regimens individually into the tail vein of SCID mice for xenotransplanted tumorigenicity analysis (Fig. 7C ). Immunohistochemistry (IHC) showed that the expression levels of Oct-4 in the lung tumors of LC-CD133 + -injected SCID mice were highly expressed in comparison with the other treated-groups (Fig. 7C) . Oct-4-IHC levels were significantly decreased in Oct-4 siRNAtreated LC-CD133 + when exposed to IR alone, cisplatin alone, or IR combined with cisplatin (p,0.01; Fig. 7C ). Moreover, using combined chemoradiotherapy with the treatment of Oct-4 siRNA, the mean survival rate of the LC-CD133 + group was significantly prolonged compared with the control or other treated groups (p,0.05; Fig. 7D ). This in vivo study also confirmed that the treatment effect of chemoradiotherapy for the LC-CD133 + group can be improved by the treatment of Oct-4 siRNA. Self-renewal and pluripotency are the central features in the definition of embryonic stem cells (ESC), and Oct-4 is a key regulator in this process [24] [25] [26] . Oct-4 has been suggested as one of four major factors that render the reprogramming capability of adult cells into germline-competent-induced pluripotent stem cells [31] [32] [33] . Previous studies also showed that mouse pulmonary stem cells endogenously express Oct-4 [34] . Recently, Oct-4 transcript can be consistently detected in human embryonal carcinomas, testicular germ cell tumors, seminomas, and bladder carcinomas [35] [36] [37] [38] . The expression of Oct-4 has further been shown in human breast cancer stem-like cells, suggesting that its expression may be implicated in self-renewal and tumorigenesis via activating its downstream target genes [39] . Herein we reported the isolation of CD133-positive cells (LC-CD133 + ) from clinical tissue samples and lung cancer cell lines. LC-CD133 + showed strong proliferative and invasive capabilities in vitro and in vivo (Figs. 1, 2, and 3) . LC-CD133 + also displayed significant resistance to chemotherapeutic agents (Fig. 4) . We also demonstrated that Oct-4 expression was transcriptionally and translationally up-regulated in LC-CD133 + (Fig. 5) . Indeed, Oct-4 functions as a master switch during differentiation by regulating the pluripotent potential in stem cells [31] [32] [33] . Using the siRNA method with lentiviral vector for knockdown of Oct-4 expression in LC-CD133 + , our data showed that the treatment of Oct-4 siRNA can block the sphere formation of LC-CD133 + and further facilitate LC-CD133 + to differentiate into LC-CD133 2 (Fig. 5) . Furthermore, in vivo animal studies demonstrated the IHC of Oct-4 in the lung tumors of LC-CD133 + -injected SCID mice were prominently up-regulated, and the total lung tumor volume as well as Oct-4 IHC levels can be significantly decreased in LC-CD133 + -injected mice by the treatment of Oct-4 siRNA with or without chemoradiotherapy (Fig. 7) . In addition, we showed that increased incidence of Oct-4 expression correlated positively with the advanced stage of 78 lung cancers (Figs. S1A-C). To our knowledge, this is the first study to show that Oct-4 expression plays a crucial role in maintaining selfrenewal and cancer stem-like properties in LC-CD133 + . The property of resistance to chemotherapy and irradiation treatment is the major clinical criterion to characterize ''cancer stem-like cells (CSCs)'' [6] . The existence of cancer stem-like cells may explain why conventional anti-cancer therapies are able only to suppress or shrink a tumor but often cannot completely eradicate it, resulting in eventual recurrence [6, 40, 41] . Consistent with these hypotheses, LC-CD133 + were significantly resistant to cisplatin, VP16 (eptoposide), doxorubicin, and paclitaxel than LC-CD133 2 (p,0.001; Fig. 4 ). Even IR alone or a single chemodrug can effectively inhibit cell growth of LC-CD133 2 (Fig. 4) ; however, IR treatment combined with cisplatin and VP-16 still failed to cause cell death in treated LC-CD133 + (Fig. 4D) . To overcome resistance to radiotherapy and chemotherapy in LC-CD133 + , treatment of Oct-4 siRNA was used and results showed that the knockdown Oct-4 in LC-CD133 + can significantly improve the anti-cancer effect in single-or combination-treated LC-CD133 + in vitro and in vivo (Figs. 6 and 7) . Moreover, the mean survival rate of the LC-CD133 + group can be significantly prolonged after the treatment of Oct-4 siRNA with IR and chemotherapy (Fig. 7) . Recently, Oct-4 has been suggested to be a protector for survival of ES cells from apoptosis induced by etoposide, UV, or heat shock through the Stat3/Survivin pathway [42] . Consistent with this important finding, our results suggest that knockdown of Oct-4 expression can effectively enhance the chemoradiosensitivity of LC-CD133 + through activating the apoptotic activities of caspase 3 and PARP (Fig. 6) . Importantly, our in vivo animal study and clinical data provide the evidence that the amount of Oct-4 in LC-CD133 + (Fig. 7C ) and in patients with high-grade lung cancer (Fig. S1 ) is positively correlated with the degree of resistance to chemoradiation therapy. Taken together, these results indicate that the up-regulated expression of Oct-4 in LC-CD133 + may contribute to the development of chemoradioresistance in patients with lung cancer. Recent studies have revealed that the human ABCG2 transporter is a P-glycoprotein that causes multidrug resistance (MDR) including mitoxantrone, doxorubicin, and topoisomerase I inhibitors of irinotecan, topotecan, and 7-ethyl-10-hydroxycamptothecin (topoisomerase inhibitor) and gefitinib (an inhibitor of EGF receptor) in patients with lung cancer [41, 43] . Hirschmann-Jax and colleagues were the first to observe that ''side population'' cancer stem-like cells isolated from cell lines and patients with neuroblastoma expressed high levels of ABCG2 and ABCG3 transporter genes as well as a greater capacity to expel cytotoxic drugs [44] . Monzani and colleagues further showed that cancer stem-like cells derived from the melanoma cell line highly coexpressed CD133 and ABCG2 markers with enhanced tumorigenic potential [45] . In this study, we found that LC-CD133 + are highly co-expressed with ABCG2 transporter and are significantly resistant to conventional treatment methods compared with LC-CD133 2 (Figs. 2 & 4) . Interestingly, a significant down-regulating of ABCG2 expression and an increase in the chemosensitivity of LC-CD133 + were observed when the Oct-4 siRNA treatment was given (Data not shown). Thus, more studies are needed to investigate whether over-expression of Oct-4, CD133, and/or ABCG2 play a role in the development of MDR in LC-CD133 + or surrogate markers of therapeutic response in patients with lung cancer. In conclusion, we demonstrated that LC-CD133 + display a higher Oct-4 expression with the ability to self-renew and may represent a reservoir with unlimited proliferative potentials for generating lung cancer cells. The resistance of LC-CD133 + to in vitro and in vivo chemoradiotherapy is partially due to preferential activation of Oct-4 gene expression. In addition, these data support that the up-regulated expressions of the Oct-4, selfrenewing gene of embryonic stem cells, play an important role in the tumorigenicity of patients with lung cancer. Table S1 Found at: doi:10.1371/journal.pone.0002637.s001 (0.05 MB DOC)

Early transcriptional response in the jejunum of germ-free piglets after oral infection with virulent rotavirus

Germ-free piglets were orally infected with virulent rotavirus to collect jejunal mucosal scrapings at 12 and 18 hours post infection (two piglets per time point). IFN-gamma mRNA expression was stimulated in the mucosa of all four infected piglets, indicating that they all responded to the rotavirus infection. RNA pools prepared from two infected piglets were used to compare whole mucosal gene expression at 12 and 18 hpi to expression in uninfected germ-free piglets (n = 3) using a porcine intestinal cDNA microarray. Microarray analysis identified 13 down-regulated and 17 up-regulated genes. Northern blot analysis of a selected group of genes confirmed the data of the microarray. Genes were functionally clustered in interferon-regulated genes, proliferation/differentiation genes, apoptosis genes, cytoskeleton genes, signal transduction genes, and enterocyte digestive, absorptive, and transport genes. Down-regulation of the transport gene cluster reflected in part the loss of rotavirus-infected enterocytes from the villous tips. Data mining suggested that several genes were regulated in lower- or mid-villus immature enterocytes and goblet cells, probably to support repair of the damaged epithelial cell layer at the villous tips. Furthermore, up-regulation was observed for IFN-γ induced guanylate binding protein 2, a protein that effectively inhibited VSV and EMCV replication in vitro (Arch Virol 150:1213–1220, 2005). This protein may play a role in the small intestine’s innate defense against enteric viruses like rotavirus. ELECTRONIC SUPPLEMENTARY MATERIAL: The online version of this article (doi:10.1007/s00705-008-0118-6) contains supplementary material, which is available to authorized users.

With an estimated death rate of more than 400,000 per year, mainly affecting children less than 5 years of age in developing countries, rotavirus is recognized as one of the major infectious diseases of the gastrointestinal tract [38] . Rotaviruses are members of the family Reoviridae, viruses with segmented double-stranded RNA genomes [17] . In the small intestine, mature enterocytes near the top of the villi are the primary target cells for virus replication [29] . Replication in these cells provokes numerous intraand extracellular pathological changes that inevitably lead to disruption of the absorptive and digestive functions of the small intestine, and consequently, to malabsorption and diarrhea. These changes include destruction of enterocyte brush borders, enterocyte vacuolization, loss and destruction of enterocytes, villus blunting and atrophy, thinning of the intestinal wall, and crypt hyperplasia (for comprehensive reviews, see [29, 39] ). However, the nature and severity of histopathological alterations in vivo can be quite different depending on the species and virulence of the rotavirus strain. There is no clear correlation between these alterations and manifestation of clinical symptoms. A systemic inflammatory response can be absent, and rotavirus infections can be asymptomatic [29, 39] . This suggests that the interplay between host and viral factors is important for determining the course of this disease. For Electronic supplementary material The online version of this article (doi:10.1007/s00705-008-0118-6) contains supplementary material, which is available to authorized users. instance, rotavirus NSP4 acts as an enterotoxin that induces diarrhea in mice in the absence of rotavirus replication [5] . NSP4 affects Ca 2+ and electrolyte homeostasis in an auto-and paracrine fashion in both rotavirus-infected and uninfected intestinal cells [41] . NSP4 increases Ca 2+ permeability of the ER and plasma membrane, resulting in an increased Ca 2+ concentration in the cytosol ([Ca 2+ ] cyt ), causing derailment of numerous Ca 2+ -dependent cellular processes [40] . In uninfected enterocytes and crypt cells this rise in [Ca 2+ ] cyt is induced by binding of exogenous NSP4 to an apical receptor that modulates the PLC-IP 3 pathway [16] . The higher [Ca 2+ ] cyt triggers laminal secretion of peptides and amines by uninfected enterocytes, and luminal Cland H 2 O secretion by crypt cells [28, 39] . In infected enterocytes, the rise in [Ca 2+ ] cyt is believed to be independent of PLC modulation [41] , and this rise perturbs cytoskeleton and tight junction integrity, which ultimately leads to cell lysis [9, 10, 26, 37] . In vitro studies with cell lines, mainly derived from colon, have contributed significantly toward understanding the pathogenesis of rotavirus on a molecular level. However, the intestinal mucosa consists of a diversity of specialized cell types in different states of differentiation. Presumably, all these different types of cells respond differently to environmental changes, and accordingly to changes in their neighboring cells. Therefore, the regulation of genes responsible for these complex phenotypic responses in vivo may not be detected by challenging single types of cultured cells with rotavirus. To address this issue, we studied the early transcriptional response in jejunal mucosa of 3-week-old, just-weaned piglets after oral infection with virulent group A rotavirus. To assign measured responses exclusively to rotavirus, we performed these experiments in germ-free piglets. Differential expression patterns of uninfected versus infected jejunum were recorded 12 and 6 h before severe diarrhea was expected, using a homemade pig intestinal cDNA microarray [34] . The biological significance of elevated or reduced expression of these genes for rotavirus pathogenesis is discussed. Seven germ-free piglets (Groot Yorkshire 9 [Cofok 9 Large White]) were obtained by caesarean section and housed in isolators, fed with sterilized condensed milk till the age of 14 days and thereafter with pelleted feed (sterilized by X-ray radiation) and water ad lib. On day 21, three of the seven piglets were transported to the necropsy room and served as uninfected control piglets. The four remaining pigs were orally infected with virus suspension diluted in a total volume of 5 ml PBS and containing 2 9 10 7 rotavirus particles (as determined by negative-stain semi-quantitative electron microscopy) of strain RV277 [45] . The virus suspension was prepared from the contents of the small and large intestine of a rotavirus-infected gnotobiotic piglet [32] . The above applied oral dose caused severe diarrhea from 24 hpi (hours post infection) in 3-week-old gnotobiotic piglets [32] . Infected piglets were housed in their isolators under the same conditions as described above for another period of 12 (two piglets) or 18 h (two piglets) before they were transported to the necropsy room. Immediately after arrival in the necropsy room, 10 ml of EDTA blood for hematological analysis was collected from the jugular vein. Subsequently, animals were killed by barbiturate overdose and their intestines were taken out. The jejunum was opened and rinsed with cold saline, and 10 cm of mucosa in the middle of the jejunum was scraped off with a glass slide, frozen in liquid nitrogen, and kept at -70°C until RNA and DNA extraction. An adjacent part of the collected jejunum was fixed in 4% formaldehyde and used to determine the villus height and crypt depth. Villus and crypt dimensions were determined on hematoxylin-eosinstained 5-lm tissue sections [34] . During the experiment, fecal samples were collected at 0, 12 and 18 hpi from the rectum for determination of the percent dry matter [18] . Fecal samples were tested for the presence or absence of rotavirus by ELISA [33] . The germ-free status of each piglet was confirmed by analyzing throat saliva and feces samples, collected on days 6, 12 and 19, and on the day of slaughter, for the presence of microorganisms. Isolation of RNA and DNA From 1 g of frozen mucosal scrapings, total RNA (DNasefree) was isolated using TRIzol Ò reagent (Invitrogen) as described recently [34] . The yield per gram of tissue and the purity of the RNA were calculated from measurement of the extinction at 260 and 280 nm. The integrity of all RNA samples was checked by analyzing 5 lg of RNA on a denaturizing 1% (w/v) agarose gel. After ethidium bromide staining, the gel was scanned to calculate the 28S/18S peak ratio (volume 28S over volume 18S) for each preparation. RNA with a ratio [2 was considered of adequate quality to be used for real-time PCR and microarray analysis. A part of the isolated RNA was used to prepare RNA pools for microarray analysis. A control pool was prepared by mixing equal amounts of RNA isolated from the jejunum of the three uninfected piglets (n = 3). The same was done for the two infected piglets slaughtered at 12 h and for the two piglets slaughtered at 18 h. After gentle homogenization in lysis buffer, DNA was extracted from 0.5 g of frozen mucosal scrapings, and 4 lg of purified DNA was analyzed on a 0.9% agarose gel [22] . The relative concentrations of interferon-gamma (IFN-c) and ornithine decarboxylase antizyme 1 (OAZ1) mRNA in all RNA samples was determined by real-time PCR. Two hundred ng of total RNA was reverse transcribed in a standard RT reaction using Superscript II reverse transcriptase (Invitrogen) and pd(N) 6 primers. IFN-c cDNA in these RT reactions was quantified using labeled Light-Cycler probes (Roche Diagnostics) as described [15] and expressed as pg/ll control plasmid. A 20-mer forward primer (5 0 -GACCCGACGCTTGCTTCATG-3 0 ) and a 19mer reverse primer (5 0 -GAGTGAGCGTTTATTTGCAC-3 0 ), generating a cDNA fragment homolog to nucleotide 702-895 of the human OAZ1 mRNA reference sequence (gi:34486089), were used to quantify OAZ1 cDNA using Cybergreen as label in a standard LightCycler reaction. The relative concentration of OAZ1 mRNA was calculated by extrapolation on a standard curve prepared from dilutions of an RT reaction prepared from a reference RNA sample [34] . The quantity of 18S rRNA in each RNA sample was determined using the above described RT reactions by real-time PCR [15] and used to normalize the IFN-c and OAZ1 concentrations. The quantity of 18S rRNA showed no essential differences among all individual RNA samples (average concentration ± SD; 4.95 ± 0.82 lg/ll of control plasmid). The same collection of pig probes (ESTs) used in earlier studies [34, 35] were spotted in triplicate on Corning Ul-traGAPS slides. Briefly, this collection consisted of 2,928 probes prepared from jejunal mucosal scrapings collected from 4-week-(672) and 12-week-(2,256) old pigs, probes coding for porcine cytokines (IFN-c, TNF-a, GMCSF, IL-2, 4, 6, 8, and 10) and lung surfactant proteins SFTPA and SFTPD, and 110 Marc1 and Marc2 probes (porcine ESTs) homolog to trefoils, collectins, defensins, and glycosyltransferases [34] . A list of the probes already sequenced/ annotated is accessible on the website of Arch Virol (gene list Hulst et al. pdf). Dual-color (Cy3-Cy5) hybridization of slides was performed using the RNA MICROMAX TSA labeling and detection kit (PerkinElmer) as described earlier [34] . Messenger RNA levels in both infected pools (12 and 18 hpi) were independently compared to the expression levels in the control (uninfected) pool. For each comparison, a dye swap was performed. In addition, a control hybridization experiment was performed in which a microarray slide was simultaneously hybridized with Cy3labeled control RNA and Cy5-labeled control RNA. Scanning of slides, processing of raw images, creation of data reports, data-normalization and statistical analysis were performed as described by Niewold et al. [34] with minor modifications. Briefly, probes were considered to be differentially expressed when at least four of the six data points (spots) on both dye swap slides hybridized with a ratio of 3.6-fold (M = [log 2 (Cy3/Cy5)] \ -1.85 or [1.85) or more (3.6 is considered significant according to the manufacturer of the TSA kit) and were identified by significant analysis of microarrays (SAM) [43] with a median false discovery rate (FDR or q value) of \5%. Equal amounts of total RNA (5 lg) were separated on a denaturizing 1% (w/v) agarose gel. After several washes with RNase-free water, the gel was blotted on Hybond-N membranes (Amersham), and blots were hybridized with 32 P-labeled DNA fragments homolog to the mRNA in question, in the same manner as was described in an earlier study [34] . After post-hybridization washes, the blots were scanned using a Storm phosphor-imager (Molecular Dynamics, Sunnyvale, California, USA). Four 3-week-old germ-free piglets were orally infected with a dose of rotavirus that caused severe diarrhea from 24 hpi in 3-week-old gnotobiotic piglets [32] . For practical reasons, three uninfected germ-free piglets were slaughtered at the zero time point (mock, see Table 1 ). In order to isolate highquality RNA from jejunal mucosal scrapings, infected piglets were slaughtered 12 and 18 hpi. Thus, 12 and 6 h before severe diarrhea would have been induced. In three of the four infected animals, rotavirus was detected in their feces. Determination of the percent dry matter showed that only the fecal samples collected 18 hpi (piglets 65 and 67) had a significantly lower (pasty) consistency [18] . This indicated that not all of the piglets developed diarrhea before 12 hpi, and that the two piglets slaughtered 18 hpi did not develop the severe form of diarrhea normally observed at 24 hpi [32] . In jejunal tissue sections prepared from these two 18-h piglets, villus length was decreased to two-thirds of the average length measured in corresponding sections prepared from the three control piglets. No significant differences in crypt depths were observed between infected and control animals. For both piglets that were slaughtered at 18 h, these results indicated that the orally applied rotavirus reached the Transcriptional response to rotavirus 1313 jejunum and induced the desired limited (not severe) pathological symptoms. In addition, the lower concentration of lymphocytes in the blood indicated that the animals were effectively infected with rotavirus [47] . In the feces of piglet 63, slaughtered at 12 h, no rotavirus could be detected. Moreover, only a small decrease in villus length (20%) was observed for this piglet and its 12-h replicate. To find additional evidence whether the jejunum of piglet 63 was effectively challenged with rotavirus, the level of IFN-c mRNA in jejunal mucosal scrapings was measured by real-time PCR (Fig. 1 ). In RNA samples isolated from the three uninfected pigs, hardly any IFN-c mRNA could be detected. In contrast, the IFN-c mRNA levels in scrapings of all infected piglets were up to 50-fold higher than in uninfected piglets, whereas OAZ1 mRNA levels were nearly equal for all RNA samples. This indicated that the jejunum of all piglets, including piglet 63, responded to the orally applied rotavirus challenge. Despite the loss of cells from the tip of the villi, the amount of RNA (Table 1 ) isolated from all infected piglets was comparable to that of uninfected piglets. On an ethidium-bromide-stained agarose gel, no degradation of RNA was visible for any of the extracted RNA samples (see also Fig. 2a ). In addition, 28S/18S peak ratios were [2 for all these samples (Table 1 ). These results showed that scrapings collected from all piglets yielded high-quality RNA suitable for real-time PCR and microarray analysis. No random (necrotic) and/or fragmentized (apoptotic) DNA was visible after gel electrophoreses of DNA samples extracted from any of the piglets, indicating that the majority of cells imbedded in the epithelial layers of all the piglets were not apoptotic or necrotic (results not shown). Mid-jejunal mucosal gene expression analysis was performed using a homemade pig cDNA small intestinal microarray [34] . In two separate hybridization experiments, mRNA expression levels in an uninfected RNA pool (n = 3) were compared to expression levels in RNA pools prepared from 12-and 18-h infected piglets (both n = 2). For both comparisons, dye swaps were performed. Probes that hybridized differentially with a ratio (FC; infected over uninfected) of \0.28 or [3.6 in both slides of the dye-swap and that were identified with the lowest possible false-discovery rate (FDR; based on SAM, [43] ), i.e., 0.35% for the 18-h comparison and 4.6% for the 12-h comparison, were selected for further analysis. Raising of the FDR to a maximum of 10% did not identify additional probes with a FC \ 0.28 or [3.6 in either comparison. Selected probes were sequenced and annotated after blastn or blastx analysis (when not yet annotated). For each differential expressed probe, the mean FC calculated from the two dye-swap slides is presented in Table 2 . When Cy3-and Cy5-labeled cDNA was prepared from the same uninfected RNA pool and simultaneously hybridized on the array, none of the probes that hybridized differentially in the 12-and 18-h comparisons hybridized differentially (results not shown). Six out of the nine probes that hybridized with a FC of 3.6-fold or more in the 12-h comparison (panel ''higher in infected'') also hybridized significantly more strongly in the 18-h comparison. For three of these probes (R14-R16), the ratio of differential expression further increased with time. In contrast, only one probe (R1) hybridized significantly less strongly at both time points. Based on literature search and data mining, a tentative function was assigned for the genes identified by blast analysis (see Table 2 ). In addition, the FC (infected over uninfected) of genes which were also found to be regulated in a previous study by Cuadras et al. [13] , i.e., 16 h after infection of human intestinal epithelial Caco-2 cells with rotavirus live virus vaccine RVV, is provided in parentheses in Table 2 after the annotations. Probes coding for IFN-c, TNF-a, GM-CSF, and IL-2, 4, 6, 8, and 10 were spotted on the array. However, the fluorescence intensity was not higher than the background threshold for any of these probes, indicating that mRNA concentrations of these cytokines (including IFN-c) in infected and uninfected mucosal scrapings were too low to detect by microarray analysis, probably due to the relatively low percentage of cytokine-producing (immune) cells present in intestinal mucosa [6] . Northern blots (NB) loaded with equal amounts of RNA from each of the piglets were hybridized with P 32 -labeled cDNA probes homologous to six differentially expressed mRNAs and to three mRNAs that were not identified as differentially expressed. For all nine mRNAs, the length of the transcript(s) detected on blots were comparable to the length of porcine or human mRNA reference sequences posted in the NCBI databank or reported in the literature. In accordance with array data, NB analysis showed that expression levels of GBP-2, KRT20, SAT, MGAM, and CASP3 mRNAs were significantly higher in both 18-hinfected piglets than in uninfected piglets. For all of these mRNAs, hybridization intensities for piglet 65 were nearly equal to that of piglet 67. GBP-2, KRT20, SAT, and MGAM mRNA expression was also up-regulated in infected piglet 63, slaughtered at 12 hpi. This indicated that the response to rotavirus infection in this 12-h piglet was comparable to the response observed in both 18-h piglets. However, no significant up-regulation of these mRNAs was observed in the other piglet slaughtered at 12 h (64), indicating that this piglet responded differently to the rotavirus infection than its 12-h replicate and the two piglets slaughtered at 18 h. In accordance with array data, NB analysis showed that the expression level of IFABp2 mRNA was significantly lower in 18-h-infected piglets than in uninfected piglets. For mRNAs that showed no significant differential expression on the arrays (glutathione-S-transferase, calbindin-D, and aldolase-B), no large differences in hybridization intensities were observed between uninfected piglets and the two piglets slaughtered 12 hpi and one of the piglets slaughtered 18 hpi (65). However, significantly lower hybridization intensities were observed for calbindin-D and aldolase-B mRNAs for piglet 67 than for its 18-h replicate. Using microarray analysis, we detected a set of genes that are differently expressed in rotavirus-infected jejunal mucosa compared to uninfected mucosa. For nine mRNAs, expression levels in individual piglets were analyzed by NB. These analysis confirmed the array data. In addition, NB analysis showed that one piglet slaughtered at 12 hpi responded quite similarly to rotavirus infection as both 18h piglets did, whereas its 12-h replicate did not, despite the fact that this latter piglet also showed an IFN-c mRNA response. In addition, 7 out of the 12 genes differentially expressed at 12 hpi also reacted at 18 hpi. These results indicated that three out of four infected piglets responded quite analogously. Because we used a limited number of germ-free piglets per time point and measured responses in a mixed population of cells, we imposed stringent criteria for selection of genes ([3.6-fold up-or down-regulation and a false-discovery ratio of less than 5%). Using this approach, we minimized the chance of selecting genes hybridizing differentially solely due to inter-animal variation in gene expression and/or cell composition. However, such stringent selection criteria could have excluded the detection of more rotavirus-regulated genes, especially of genes regulated exclusively in specific types of cells that are present in low quantities in the jejunal mucosa. The different responsiveness of one of the 12-h piglets, however, obliged us to interpret our overall results carefully, especially, concerning the five genes that reacted solely at 12 hpi (TXN, FRK, DppC-2, IF144, and ACTB; see Table 2 ). Nevertheless, data mining revealed relevant relationships between these five genes, 18 h response genes, and processes known to be important for rotavirus pathogenesis. Cuardras et al. [13] measured the transcriptional response in the human enterocyte cell line Caco-2, 1, 6, 12 and 16 hpi with Rhesus rotavirus live vaccine. Four genes up-regulated in our experiments (GBP-2, SAT, MGAM, PPA1) were also up-regulated 16 hpi in Caco-2 cells. Recently, Aich et al. [1] profiled the transcriptional response in surgically prepared jejunal loops from 1-dayold colostrum-deprived calves after 18 h of perfusion with bovine rotavirus (BRV). Several genes for which we detected more than 3.6-fold up-or down-regulation (TXN, NADH5, SGLT1, ACTB, SAT, CASP3, and PPA1) were also present on the cDNA array they used (NCBI GEO acc. number GPL325). None of these genes showed a differential expression of twofold or more in their study. The different route of administration and virulence of the strain used, the digestive differences between the jejunum of omnivores and herbivores, and, in the case of the study of Cuardras et al., the various specialized cell-types present in the jejunum of living animals versus cultured colon-derived Caco-2 cells are probably responsible for the poor correspondence between these three studies. Based on relevant literature and functional information in databanks, we assigned a function and a possible type(s) of cell(s) responsible for expression for most of the genes on our list (Fig. 3) . In this hypothetical model, information from existing models dealing with the pathogenesis of rotavirus infection [29, 39, 40] and the development and maintenance of the small intestinal epithelium [20] were used to fit in our data. Possible functions of these genes in relation to processes and pathways known to be important for rotavirus pathogenesis are discussed below. Measurements of villus length indicated that considerable numbers of epithelial cells were lost from the tip of the villi, including (infected) mature enterocytes. In part, down-regulation of genes involved in transport of ions and nutrients over the membranes of mature enterocytes, like meprin A, SGLT1, and IFABp2, may be a direct result of this loss. In another part, replication of rotavirus in enterocytes imbedded in the epithelial layer could have down-regulated transcription of these genes. This may also be the case for other down-regulated genes from our list, especially for genes detected only at 18 hpi (R4-R13, Table 2 ). In addition to down-regulation of SGLT1, IFABp2, and meprin A, we observed up-regulation of two other genes that may affect the absorptive and digestive function of the intestine: a gene coding for a protein carrying a Ca 2+ -permeable cation channel CD20/IgE Fc receptor subunit b domain (MS4A2) and a gene homolog to a bacterial oligo/dipeptide permease (DppC-2). It is tempting to link up-regulation of MS4A2 directly to NSP4induced enhancement of Ca 2+ permeability of the plasma and ER membranes in intestinal epithelial cells [29, 40] . Likewise, up-regulation of the DppC-2 homolog may be related to enhanced laminal secretion of peptides and amines by uninfected epithelial cells, a process believed to be triggered by raised [Ca 2+ ] cyt [40] . Characterization of these porcine transcripts/proteins is needed to provide further insight in the role of these genes in rotavirus pathogenesis. The same applies for the TMEM106B gene. This gene showed the highest level of up-regulation. So far, only a DUF1356 protein domain with unknown function has been predicted in TMEM106B. Recently, it was reported that rotavirus infection in infant mice induced apoptosis in vivo [7] . Although DNA analysis showed that the majority of cells present in infected mucosal scrapings were not apoptotic, we observed upregulation of the apoptosis effector protein CASP3. This suggests that programmed cell death in the epithelial layer was stimulated by rotavirus infection. NB analysis detected a considerable level of CASP3 mRNA expression in uninfected mucosal scrapings. This constitutive expression of CASP3 is, most likely, related to the process of maintenance of the absorptive status of the intestinal epithelial layer. A process in which mature enterocytes continually die due to apoptosis and are replaced by differentiating cells migrating from surrounding crypts to the tip of the villi [20] . In contrast to mature enterocytes, an in vivo study in mice showed that goblet cells are largely spared from apoptosis in rotavirus-infected mice [8] . Moreover, migration of goblet cells from the crypt to the tips of the villi was stimulated in these mice. We found up-regulation of the goblet cell marker gene KRT20 [51] and the lower and mid-villus immature enterocyte marker gene MGAM [42] . This could indicate that transcriptional activity in both of these cell types was promoted. Stimulation of apoptosis in rotavirus-infected enterocytes and higher proliferation/differentiating activity in goblet cells and immature enterocytes could be a coordinated response of the jejunum to remove infected enterocytes and overlay villus tips with fresh enterocytes, goblet cells, and mucus layer. We did not detect genes on our array that were directly associated with cell-cycle progression/arrest. However, down-regulation of the nuclear kinase FRK (an antagonist of cell proliferation) and TMS4F20 may be associated indirectly with this process. In humans, TMS4F20 is strongly homologous to TM4SF4, a protein that reduced the ability of the crypt cell line HT29 to proliferate [48] . Several other genes that may play a role in cell death and repair were regulated. SAT was up-regulated, and TXN and PRR13 were down-regulated. For this later protein, reduced expression in cells was correlated with increased sensitivity to taxane-induced cell death, a caspase-independent process characterized by the polymerization of tubulins to extraordinarily stable microtubules [27, 30] . TXN is the major carrier of redox potential in cells, and it is crucial for the defence of cells against oxidative-stressmediated apoptosis. TXN also regulates gene expression by increasing binding of redox-sensitive transcription factors like p53 [44] , NF-jB [25] , and the Nrf-2/polyamine- Table 2 . Information about the Paneth cell marker THOC4 is provided in reference [24] Transcriptional response to rotavirus 1319 inhibited SAT expression [23] . Therefore, up-regulation of SAT gene expression at 18 hpi may be directly related to down-regulation of TXN at 12 hpi. SAT is a rate-limiting enzyme in spermine/spermidine metabolism. Acetylation of these polyamines by SAT promoted their degradation and excretion [46] . Recently, it was reported that depletion of polyamines suppresses apoptosis in normal intestinal epithelial cells by AKT-kinase-mediated inhibition of CASP3 activity [50] . NB analysis showed that the increase in SAT mRNA expression coincided with the goblet cell marker KRT20. Therefore, it would be interesting to determine whether SAT expression in goblet cells can be stimulated by rotavirus infection and whether it plays a role in protecting these cells from apoptosis [8] . Interestingly, it was recently demonstrated that RNA viruses can directly modulate polyamine metabolism by regulation of SAT transcription and splicing [36] . A most interesting gene that we found more than tenfold up-regulated codes for an as yet not-well-characterized hypothetical human protein (LOC646627) carrying a phospholipase A2 inhibitor domain (PLA2-inh). The magnitude and kinetics of up-regulation of this gene corresponded exactly with GCNT3, suggesting that this gene was expressed in the same types of cells as GCNT3, most likely mucus-producing goblet cells and/or differentiating (immature) enterocytes. The amino acid sequence translated from our PLA2-inh EST showed an overall amino acid identity of 56%, and all cysteines aligned perfectly with cysteines of the human LOC646627 protein and of other proteins that bear a typical PLA2-inh domain. PLA2s comprise a diverse family of cytosolic and secreted enzymes that hydrolyze membrane phospholipids to free fatty acids. They play an important role in many exogenous and intracellular processes, ranging from fatty acid metabolism and lysis of membranes to the synthesis of arachidonic acid (A-acid), an essential precursor for the production of inflammatory mediators such as eicosanoids. Secreted PLA2s are calcium-dependent enzymes. Cytosolic PLA2s (cPLA2) can also be calcium-independent. A moderate increase in [Ca 2+ ] cyt mediated translocation of calcium-dependent cPLA2 to intracellular membranes where it hydrolyses phospholipids to A-acid [12] . A similar effect was observed after activation of the MS4A2 calcium-permeable cation channel (up-regulated here at 18 hpi, see above) on the surface of mast cells [14] . Perhaps, enhanced expression of a PLA2-inh in our study may be a countermeasure of specific intestinal epithelial cells to normalize and/or inhibit PLA2 enzyme activity in response to extra-and intracellular changes in [Ca 2+ ] evoked by rotavirus replication, either to protect specific cells from extra-and intracellular membrane-damage or to negatively regulate A-acid production. With respect to the latter process, we observed down-regulation of the enzymes CYP2C39 and ACSL3, which both utilize A-acid acid as substrate. Interestingly, the capsid protein of parvovirus possesses PLA2 activity [49] , and HMCV particles carry a cell-derived PLA2 activity [2] . For both viruses, PLA2 activity appeared to be essential for infectivity. Our results showed that IFN-c mRNA expression in the jejunum of infected piglets peaked around 12 hpi and tended to decline beyond this time point (see Fig. 1 ). This suggests that IFN-c was produced for a short period. Recently, Aich et al. [1] measured the mRNA expression levels of several cytokines in jejunal loops perfused for 18 h with BRV. However, they observed no IFN-c mRNA response. Because our orally administered rotavirus needed time to reach the jejunum, our 18 h infection period represents a shorter period than 18 h of perfusion. After 18 h of perfusion, expression of IFN-c mRNA may have dropped to normal levels. Interestingly, they did detect a rotavirus-induced IL-6 (alias IFN-b 2) mRNA response at 18 hpi [1] . Recent studies showed that the interplay between IFN-gamma and IL-6 controls the influx and clearance of neutrophils and, subsequently, the transition to a more sustainable influx of mononuclear cells during acute inflammation [21, 31] . Therefore, it would be interesting to study which immune cells produce IFN-c and IL-6 and whether an orchestrated action of these cells regulates an influx of vital immune cells in the jejunum after rotavirus infection. The IFN-c-inducible GBP-2 gene was up-regulated 18 hpi. Overexpression of GBP-1 and GBP-2 in HeLa cells and NIH 3T3 cells abrogated the cytopathogenic effect mediated by VSV and EMCV, respectively, by an unknown mechanism [3, 11] . Furthermore, it was shown that expression of murine GBP-2 in NIH 3T3 cells neutralized the cytotoxic effect of the taxane drug Paclitaxel [4] . This drug specifically stimulates polymerization of tubulins to extraordinarily stable microtubules. These stable microtubules interfere with the function of normal microtubule filaments and inevitably induce cell death [4] . The reduced expression of PRR13 we observed (discussed above) may be an indication that the formation of extraordinarily stable microtubules in intestinal epithelial cells actually takes place in response to rotavirus infection. In fact, several studies have shown that rotavirus infection induces disorganization of the cytoskeleton network and microtubule filaments in enterocytes [9, 10, 26] . Moreover, we also observed the up-regulation of the IFN-a/b-inducible gene IFI44, a cytosolic protein associated with microtubular structures, and the cytoskeleton gene ACTB. Therefore, enhanced expression of GBP-2 in specific intestinal epithelial cells could contribute to a cellular mechanism(s) that impairs and/or prevents disorganization of the microtubule filaments. Further in vivo studies are needed to determine whether the genes identified in this study are representative for an intestine with a normal microflora. If so, more focussed studies involving in situ hybridization and immuno-histology may specify where along the crypt-villus axis and in which type of epithelial (or immune) cells elevated or reduced expression of these genes is induced.

Gatekeepers of health: A qualitative assessment of child care centre staff's perspectives, practices and challenges to enteric illness prevention and management in child care centres

BACKGROUND: Enteric outbreaks associated with child care centres (CCC) have been well documented internationally and in Canada. The current literature focuses on identifying potential risk factors for introduction and transmission of enteric disease, but does not examine why these risk factors happen, how the risk is understood and managed by the staff of CCCs, or what challenges they experience responding to enteric illness. The purpose of this study was to explore the understanding, knowledge and actions of CCC staff regarding enteric illness and outbreaks, and to identify challenges that staff encounter while managing them. METHODS: Focus groups were conducted with staff of regulated CCCs in Southern Ontario. Five focus groups were held with 40 participants. An open ended style of interviewing was used. Data were analyzed using content analysis. RESULTS: CCC staff play an important role in preventing and managing enteric illness. Staff used in-depth knowledge of the children, the centre and their personal experiences to assist in making decisions related to enteric illness. The decisions and actions may differ from guidance provided by public health officials, particularly when faced with challenges related to time, money, staffing and parents. CONCLUSION: CCC staff relied on experience and judgment in coordination with public health information to assist decision-making in the management of enteric illness and outbreaks. Advice and guidance from public health officials to CCC staff needs to be consistent yet flexible so that it may be adapted in a variety of situations and meet regulatory and public health requirements.

In 2003, over 2 million Canadian children from the ages of 0-5 years were in some form of non-parental care and 28% of these children were in a daycare [3] . The cost of child care ranged from $300 to over $700/week and varied between provinces, territories and age in 2005 [4] . Two types of child care, regulated and unregulated, are present in Canada. There are approximately 11, 000 regulated child care facilities in Ontario which include nursery schools, preschools, centre-based full-day child care and regulated home child care. These centres must meet legislated requirements for operation of services as set out by provincial/territorial regulations. Unregulated arrangements may consist of care provided by a relative, by an unregulated family child care provider, or in-home caregiver [5] . Canadian data from 1998 showed that over 98% of child care centre (CCC) staff were female and 45% were under the age of 30. Staff had a high level of experience; 60% had worked in the field over 5 years and 71% of teaching staff held an early childhood education (ECE) credential. Hourly wages varied between provinces and territories. The national average hourly rate was $11.62 for a teacher and $9.59 for an assistant. Only 74% of staff reported paid sick days, which averaged 7.6 days/year. Even with these challenges, a high proportion of staff reported experiencing enjoyment and reward from their work. Twenty percent reported feeling disrespected by some other professionals [5] . Numerous enteric illness outbreaks in CCCs have been reported, although investigations rarely identified a common source. Reports commonly cited person-to-person transmission or risk factors that increased the potential for transmission and illness [6] [7] [8] . These factors included inadequate hygiene related to diapering, toileting and hand-washing [9, 10] ; poor cleaning of environmental surfaces [6] ; improper food preparation [6] ; inadequate exclusion of ill children and staff [9, 10] ; and working or being cared for in a younger age group [7, 11] . Acts and regulations related to child care and health are issued by the provinces and territories in Canada. Additionally, public health officials at the local level have a high level of responsibility for performing inspections, responding to outbreaks and providing guidance or developing resources for CCCs according to these Acts and regulations. Staff of CCCs are most likely made aware of these Acts and regulations through their formal training, but also through regular interaction with public health officials. The variation between provincial/territorial or local health regions can make consistent public health response to enteric outbreaks challenging. Currently, knowledge is limited on how staff working in CCCs interpret and implement the guidance they are provided from public health officials. A previous study identified that staff may have inconsistent definitions of what constitutes diarrhea, and that their actions during an outbreak may differ from those recommended by public health officials [11] . Furthermore, reports from three recent Canadian investigations recommended the development of consistent guidance for CCCs related to outbreak management, prevention and control of illness [9, 11, 12] . Resources or programs can be more successful when developed in collaboration with the groups or people who may be impacted by them [13] . Staff of CCCs were identified as an important group to consult with regarding the development of consistent guidance. The exploration of perspectives, understanding and opinions were important aspects, and thus qualitative research approaches were selected as the most suitable tool to gather this information. The qualitative approach allowed for data to be collected in the form of the participant's own words which provided a more detailed and deeper understanding than would be possible through quantitative studies. Use of qualitative methods required an inductive approach to explore the participant's experiences and perspectives [14] . Focus groups are a mechanism for collecting data. This form of group interview sets up an environment for the participants to interact, and interaction can lead to a wide range of information on experiences and perspectives [15] . The purpose of this study was to explore the perspectives, knowledge, self-described practices and challenges CCC staff working in regulated centres in southern Ontario have related to enteric illness and outbreaks. It was also intended to investigate how staff use current public health guidance. The final objective was to gather information to feed into the development of consistent recommendations for the prevention and management of enteric illness in child care settings. Between November 2005 and January 2006, five focus groups were conducted with staff from regulated CCCs in Southern Ontario. Four groups were conducted at different CCCs. Each of the four groups consisted of staff from the respective CCCs. The fifth group was mixed and consisted of staff from different centres and who did not work together. Ethics approval was obtained from the Research Ethics Board of the University of Guelph. All participants provided written, informed consent A purposive sampling approach was used to recruit participants from regulated CCCs in southern Ontario. Centres were selected by the researchers and contacted by phone. Study information and requirements were sent to each centre for their further consideration and permission to conduct the focus groups. This sampling approach sought participants who worked in a variety of settings and who could contribute their experiences on managing enteric outbreaks. Centres and staff members were defined as those who provided daytime care for an entire day to a group of children (e.g. did not include those who provided only after-school care), and whose staff's first language was English. Participating staff had daily direct contact with the children and had worked in a centre for more than six months. Preliminary research was conducted with public health officials and CCC staff to develop appropriate questions for use in the focus groups. One-on-one interviews were conducted with public health officials to gather information related to the types of centres that they worked with, how enteric illness outbreaks were typically managed, what support public health officials provided, and their experiences while working with CCC staff. Observations of daily operations were also conducted in five centres to better understand the working environment and how staff interacted with their coworkers and supervisors. Time spent observing in the centres ranged from two hours to a full day. Open ended inquiry was used in each focus group. Sixteen questions were developed and used per session. Questions related to staff priorities, knowledge of enteric illness in the CCC, definitions and practices, how they used information from public health officials, challenges they faced when managing enteric illness and recommendations for improvement of these situations. Public health officials assisted in the verification of the question set prior to the focus groups to ensure clarity of the order, timing and flow of discussion. The inductive process used during and between groups with the CCC staff allowed the questions to be altered or refined. For example, follow-up or probing questions were added to allow for experiences and details not in the original question set to be captured as directed by the participants. This process of modifying questions ensured that new situations and recurring trends were explored thoroughly. Focus groups were conducted in a CCC or public location and lasted between 45 and 60 minutes. The principal investigator moderated all groups and a co-facilitator assisted by taking detailed notes. All groups were audiotaped using a digital recorder and transcribed verbatim by the principal investigator. Additionally, detailed memos were written by the researcher following each group. Notes pertained to group participation, how questions were responded to, interesting features of the discussion, and considerations for future groups. Content analysis was used to analyze the data. This involved examining the transcripts and use of descriptive codes to compare and look for common patterns and themes among the groups [14] . All transcripts were examined by the principal investigator for common patterns within groups. Next, transcripts were reviewed and coded using descriptive codes such as: "communication with parents", "description of diarrhea" and "acknowledgement". Further content and theme analysis was done using NVivo 7. NVivo is a software program often used by qualitative researchers to handle text based data. It assists with the categorization, analysis and displaying of data [16] . Descriptive codes were later grouped into larger categories. Through continuous comparison of the transcripts, categories that were common among the groups were refined. This constant comparison allowed for the identification of themes based on common categories among the groups to be identified. A total of 40 staff participated in one of five focus groups. The number of participants per group ranged from 3-12. All participants were female and the median age was 35 years (range 20-60 years). Eighty-three percent of the participants had an ECE Diploma (college), and 15 % had a University degree. The median number of years working in a child care centre was 11 (range 2-30 years). The centres represented a variety of settings from large, public centres with multiple classes for infants, toddlers, preschool and school-aged children to smaller centres which only had two rooms. The centres also differed on geographic location and the type of population they served. Five major themes emerged from the data. The following is a description of the themes and text based examples to illustrate them. Among the groups, the health and safety of the children was the primary concern and responsibility of the staff. Staff ensured health and safety through observation and the creation of an informal surveillance system for any illness among the children. Staff reported an intimate knowledge of the behaviour and health status of each child they cared for, including indicators of enteric illness such as bowel movements. Their intimate understanding of the children was used to monitor and look for cues that indicated a change in the child's health. These cues included changes in behavior, eating or sleeping patterns, additional symptoms as well as a change in the frequency, colour and consistency of a child's bowel movement. Staff: You get to know the smell, like the regular smell of somebody's bowel movement, the colour, you get to know the timing, like something is different or not right. Staff recognized that bowel movements, and the causes of irregular bowel movements, vary between children and took these factors into account when making decisions and taking actions. This knowledge allowed them to consider additional factors such as the use of antibiotics and medication, teething, nutritional intake as well as the period of time that would most likely indicate an infectious disease as the cause of illness. Staff:...and that's the hardest thing..., is this really diarrhea from a virus or is it just they're on antibiotics or they're teething. Staff worked in a dynamic and continually changing environment. Based on previous experience, staff knew when enteric illness appeared infectious, for example, if they observed multiple children ill throughout the centre with similar symptoms in a short period of time. They also noted that illness was often first identified in younger ages and spread throughout the centre rapidly. Staff: The number of cases of anything, you know when you start to see so many children with similar symptoms-quickly in succession that's when you start wondering if something is happening. Staff monitored the health and safety of the children by collecting information on various forms (attendance, toileting, health, etc.) and in personal journals. Staff described collecting a large amount of information, however, the number and types of forms being used, the detail collected and how the information was recorded and shared varied between centres. There was no standardardization of forms, although in some jurisdictions standard forms may have been recommended by public health officials. Information collected by staff often remained specific to the room or the group of children that they worked with. The supervisor or director of the entire centre was responsible to collect it centrally and track the information. Staff relied on the supervisor to identify centre-wide health concerns and inform all staff, as well as act as a liaison with public health. Monitoring and observation was reduced when a child was at home or in school, wherein the staff were unable to monitor hand-washing or health status and had to rely on parents and teachers for information. The staff had less ability to monitor older children (closer to kindergartenage), as the child's independence was much higher. For older children, staff were less involved in diapering and toileting and less in depth records were kept. The staff role as "eyes and ears" created an informal surveillance system for enteric illness within the centre and was also the first stage in a decision-making process. Staff demonstrated knowledge of the children they worked with and the potential causes and symptoms of enteric illness through various cues. The system relied on the staff's recognition of symptoms and their ability to report this information for tracking and communication purposes. In order to maintain the health and safety of the children, staff acted practically and responses were action-oriented. When a child was ill, their first actions were to care for the child. Staff 2: Get out the gloves, the pads, the wet cloths, the concentrated cleaning, clean that off. These practical actions ensured a safe environment for the children through thorough cleaning, hand-washing, and restricting play areas (communal water and sand play) where the potential for transmission of illness from child to child existed. During an outbreak, staff remained action-oriented but began a more intensive cleaning procedure. Responses to outbreaks were described: Staff spoke easily and frequently about they type of cleaning they did during an outbreak, how often they did it, and how it became a habit that they integrated into their day. Staff: I mean you end up putting it into your program, you end up putting it into the transitions of your day. Like in the infants [group] for example, once one staff goes outside with the children that are awake, the other staff stays inside with the children that are sleeping and they're washing toys, it just becomes a daily part, an everyday practice. Staff had a high level of comfort with this response and it was equivalent to practices that would be recommended by public health officials. From their experience working in the centre, staff felt that cleaning stopped further transmission of illness. This assisted them in meeting their goals of a healthy and safe environment for the children. When staff defined diarrhea, they often used visual and sensory based descriptions like: "uncontained", "out of the diaper", "running down their legs". Staff used these cues to take action. However, staff definitions or descriptions of diarrhea were not the same; there was a level of ambiguity and uncertainty and they experienced a dilemma in determining whether a child was ill or not. The staff described two reasons for this uncertainty: 1) Diarrhea could be child specific. One definition may not apply to everyone. 2) The definition of diarrhea could vary between staff members and often was considered a "judgment call". Staff experienced a similar difficulty in clearly defining an "outbreak". In public health, an outbreak is often defined as a sudden or unexpected increase of disease within a population. [17] Staff discussed that there were a number of ways in which they determined when an outbreak was occurring. These included: increased number of children who were ill, comparison to a set baseline, more illness when compared to previous years, or situations where there were multiple illnesses. Staff: But there's centres that are confused about that, you know I thought it was always 10% and you might talk to one public health person and she may say, oh, it's 10%, you might talk to another one who might say go on past history. Staff stated very clearly their need for standard definitions for both diarrhea and outbreak to assist them when managing enteric illness. They felt that if the definitions were clear, they would know which actions to take and when. Staff: It's diarrhea, even to recognize diarrhea versus a loose bowel movement, cause they're two different things, that would help everybody, this is where we need experts to tell us what is the true definition of diarrhea. In these situations staff used their experience working in the centre, knowledge of the children and personal judgment to enable them to develop definitions that assisted them, but expressed that further guidance from public health officials would be welcomed. Staff expressed that at times they felt overloaded with demands and/or had limited resources to respond appropriately to unpredictable situations of enteric illness. In these situations, they required flexibility in how they provided care and what actions they took. Staff discussed their use of judgment and previous experience in consideration with guidance provided by public health officials to make appropriate decisions. Staff had different forms of education and varying experiences. This attribute influenced how staff worked with children and managed enteric illness and outbreaks. Although experience could assist in making judgments, experienced staff indicated their personal judgments of diarrhea and how they responded to situations varied. This was in part due to the uncertainty associated with how to define diarrhea. New or inexperienced staff required training by more experienced staff to assist with decision-making. Although policies and guidance related to management of enteric illness are provided to the centre as a whole, staff reported that modification took place on a situation specific basis. This was reported by staff when they felt that the procedures and guidelines were inflexible or differed from a situation that they experienced. In some cases, staff actions varied from what they knew, leaving the individual staff member with a great deal of responsibility in deciding what to respond to and when. Staff: We have to say that this is our policy, like there are certain things that we can bend on, like if the fever is 99 degrees and our policy is whatever, then yes-bring your child in. But when it's something like that [excluding a child with diarrhea], it's "this is our 24 hour policy". Guidance provided to staff from public health officials and CCC management in this dynamic environment needed to be flexible and pragmatic. Staff used the recommended information provided by public health officials, but also felt that they had adequate ability to modify it to unique situations. This created new practices at the frontline level that were often more specific to the situation and the centre. This responsibility, and the decisions staff made, could significantly influence the management of enteric illness and the potential for outbreaks. In certain situations, staff experienced conflict in the care they provided due to a specific challenge. These challenges prevented them from taking required actions, and fell into four areas: money, time, staffing and parents. Most centres identified that on a regular basis they used bleach and water to clean and disinfect. During outbreaks or increased illness one of the first precautions was to change to a more powerful cleaning product. Use of these products was considered important for making the environment safe for the children. Staff: We tend to be a little proactive, well just yesterday we had one case of diarrhea and today we sent a child home with vomiting, so that room has now been totally disinfected with Virox. Many staff reported that when the outbreak was considered to be over, centres returned to normal bleach and water solution due to the higher cost associated with using alternate cleaning products on a regular basis even though they felt the use of the product was beneficial. Staff: I was going to say that's truly the reason why, the cost, bleach is a lot cheaper than Virox. Staff described that they were always pressed for time, and in certain circumstances, additional responsibilities could require changes to routines in child care that may not be feasible. The expectations to keep the environment clean, especially during an outbreak, are very high and staff felt increased pressure to find ways to incorporate this demand into an already busy schedule and maintain a high level of care. Staff: I mean it's a lot, on top of all the stuff with parents and the kids and keeping everyone healthy ... during outbreaks we are disinfecting every surface, every toy, like every half day to every day ... cause not only are you trying to run your usual program and everything else but you're also trying to take time within those hours to care for sick kids and do this mass disinfecting...it's just added stress and workload. Staff worked hard to provide the cleanest and safest environment possible, but recognized they were limited by time and resources. Cleaning products which are fast acting and can be used around the children throughout the day were deemed as necessary. Staff also often suggested that having additional staff support just to clean would be high on their wish list. Record keeping was cited as an important component of routine monitoring, and especially so during an outbreak. Staff used several forms related to the group they cared for (e.g. attendance) and each individual child (e.g. bowel movement records). This record-keeping required a large amount of time and may prevent staff from performing other actions such as cleaning, speaking with a parent, planning other programs and activities, or looking after themselves. Forms related to enteric illness were only some of the records staff kept. Staff were also responsible for documenting information related to diet, allergens, what activities the child participated in, or behaviour they exhibited. Staff said that it was difficult for them to take time off from work when they were sick. Reasons included not having compensated sick days, concern over loss of pay, and the inability to find a substitute staff member. Staff knew that they should, and were expected to take time off when experiencing symptoms of enteric illness. For most of the staff, this could be a challenge due to a limited incentive to remain at home while ill, and they admitted not always excluding themselves. When staff were ill, they relied on a substitute staff member to replace them. However, if a substitute staff member worked in a centre during an outbreak they were not permitted to work in another centre concurrently. This policy is in place to prevent disease transmission. Staff reported that when they are unable to find a substitute staff, they may choose to work. Staff: If she happens to come to my centre and then the next day we were officially in an outbreak, I would have to call that supply teacher and say "you know what, you were here yesterday and we're in an outbreak". She can come back to my centre but she can't go to [another] centre, so that's the frustrating part. Appropriate ratios between staff and children are regulated and must be maintained in a centre at all times. When staff became ill and were unable to work, or when an ill child needed to be segregated from the larger group until a parent could come and pick them up, the maintenance of ratios between staff and children could be compromised. In these circumstances, there may not be enough staff to segregate the child from the rest of the group. Staff balanced variables such as how ill the child was, how long they had already been in the centre, and the number of available staff, before proceeding to make an informed decision regarding how to appropriately exclude children while maintaining required ratios. Staff 2: sometimes we just don't have the staff or ability to segregate a child. The relationship between staff and parents was described as positive. During an outbreak, or when a child needed to be excluded, it could become strained. Lack of understanding and level of parental awareness about the transmission of enteric illness created a barrier with staff. Some parents did not understand the importance of keeping their children at home when ill. Staff felt that experience working in the CCC gave them a greater understanding of the severity of symptoms, potential causes for illness, and when a child was truly ill due to infectious diseases. Staff: That is probably one of the biggest reasons as well as being inconvenient, perhaps inconvenience is the first one...just the financial end of it is another reason why parents are perhaps tentative to withdraw their child from daycare..."we paid for that". Staff recognized their inability to monitor the child and ensure proper hygiene when they were sent home. They relied on the parent's honesty regarding their child's health status. Staff 1: I mean the hard part with it is we are relying on parents to be completely honest. A number of reasons why they felt parents found it difficult to come and pick up an ill child or keep them at home was discussed. These reasons included a lack of access to alternate care providers or the inability to take time off for financial reasons or the nature of the work place. During the focus groups, a number of the staff expressed empathy for the parents' situations, and staff often felt conflicted between balancing the needs of the parents and what they deemed best for the centre. They expressed that overcoming the challenges associated with parents would be the most helpful to assist them when dealing with enteric illness. This study illustrated that CCC staff had an intimate knowledge of symptoms and potential causes of illness in the children they cared for. This knowledge was used in addition to experience related to enteric illness to guide their decisions and actions. An informal surveillance system to identify illness and take appropriate actions was created using the knowledge gained from daily interaction with the children and the observations and records staff kept. Staff described a high level of comfort in their ability to thoroughly clean, and thereby to help prevent and manage enteric illness within the centre. When staff felt uncertain, they relied on their own judgment and experience to assist them. This was apparent when staff described their difficulty in defining both "diarrhea" and "outbreak". In situations where challenges related to money, time, staffing and parents were identified, or staff required flexibility in their response, staff adapted their actions to ensure appropriate care, even if it meant modifying recommendations provided by public health officials. Staff gave examples including: adapting cleaning schedules, exclusion guidance, and record-keeping. Public health officials provide guidance based on legislated regulations. The purpose of this guidance is to ensure that staff consistently achieve outcomes which protect the safety and health of the children. This includes the prompt identification of cases and outbreaks of enteric illness so that appropriate public health preventative measures can be put in place. Findings demonstrated the health and safety of the children was a priority for CCC staff but how objectives related to this priority were achieved and how guidance from public health officials was used may vary by staff and facility. To the best of our knowledge, a study of this nature has not been done. However, the findings support other work that demonstrated staff may have inconsistent definitions and actions related to enteric illness [11] . The challenges identified, such as time, understanding and financial or logistical needs have also been documented in other studies that have examined health care practitioners and their challenges with hygiene [18] [19] [20] . Additionally, the findings help to support the need for consistent forms of management for enteric illness and outbreaks in CCCs. Consistent guidance was supported by the staff who discussed their needs related to standard definitions and actions to take based on the definitions. The use of focus groups with staff of CCCs allowed for the collection of in-depth data about a wide range of experiences and opinions related to enteric illness and outbreaks. Development of the original question set using information collected from public health officials and staff of CCCs ensured thorough exploration of themes and experiences related to the management of enteric illness. The inductive process allowed for new questions and details to be explored, and led to a range of discussion about the perspectives, experiences and challenges of CCC staff that would not have been possible through the use of a standard, closed questionnaire. Using a purposive sampling approach ensured that participating centres represented a range of care settings available in Southern Ontario. Centres varied in the number and age of children cared for, and setting type (teaching facility, private, public, etc). Although there were some differences between staff and centres, the major themes, experiences and perspectives that the staff spoke of were the same regardless of what type of centre they worked in. Non-regulated facilities were not approached due to the challenges associated with identifying them, size and physical setting. The insight and understanding of CCC staff can be used in further development and implementation of practical guidance. The process of developing, following and adapting policies at the frontline level has been examined in other public service workers. "Street Level Bureaucrats" are defined as public service workers who interact directly with citizens in the course of their job, and who have substantial discretion in the application of policies in the execution of their work [21] . This level of discretion and ability of the frontline worker to modify set government policies for the individual situation has also been described among police officers, social workers and others who work within the public sector [21] . Additionally, studies of nurses have demonstrated similar reliance on previous experience and visual cues such as touching, observing, listening, feeling or sensing, and "knowing" in decision-making [22, 23] . Conflict in care has also been reported in health care workers who demonstrated balancing the risks and demands of caring for patients [20] . The results of our study demonstrated that the decision making process and demands placed on CCC staff are similar to those experienced by other professionals, who work in high stress environments and care for others on a daily basis. The decision-making process that CCC staff used could best be described as Naturalistic. Naturalistic decisionmaking is most often associated with proficient decisionmakers who have extensive experience [24] . The process is informal and relies on the intuition and judgment of the decision-maker. The environments where this decisionmaking process is most frequently used are those with shifting goals due to dynamic and changing conditions, time constraints and high stakes. Typically identified in firefighters or military personnel, this framework also appears to apply to CCC staff. Their decision-making relies heavily on using their judgment and experience to make decisions and modify plans to meet the needs of the situation in a workable and timely fashion [24, 25] . Previous research has concluded that control measures in the form of standard guidance, education and hygiene are necessary to assist with the prevention and control of infectious diseases [26] . Other studies in areas of infection control and hygiene have highlighted issues related to compliance with guidance. To minimize these issues and increase compliance, they recommend guidance should be easy to follow, accessible and that identified challenges should be considered during their development. Strategies for changing practices should address needs at the individual and group level [27] [28] [29] . The staff in this study also highlighted these as considerations, and based on the findings in this study a number of factors were identified that could strengthen and be considered when developing further guidance to ensure optimal compliance. The process of identifying, managing and preventing enteric illness in children in CCC settings is inherently variable due to the number of factors, but continuing to enhance consistent decision-making tools and resources for all CCC staff is important. For example, consistently updating public health manuals and onsite visits from public health officials would be beneficial. A visual framework to assist decision-making, which could be used by CCC staff and parents, could be designed to include the variables identified by CCC staff in this study, as well as additional factors considered important by public health officials (e.g., blood in the stool). This framework could be designed to allow CCC staff to incorporate their experience and knowledge into it. Decisions and actions rely on a clear and consistent understanding of "diarrhea" and "outbreak" and CCC staff indicated a need for this guidance. Developing one definition may not be possible when there are a number of factors to consider, but inclusion of these factors in a clear decision making framework could be of assistance to staff. It is important for public health officials and management of CCCs to work together while continuing to develop and strengthen definitions of diarrhea and outbreak. As well, it remains an important task to continue to clarify procedures and activities with all CCC staff to ensure consistency. In CCCs, staff require resources that they can use and adapt as needed. For example, improved record-keeping forms could be developed that are visual and easy for all staff to complete and which could improve challenges associated with time. The information collected on the forms is invaluable in illness surveillance for the entire centre. Staff were very aware of the children they work with but enhancing awareness and communication among staff could ensure the rapid implementation of preventative measures such as increased cleaning throughout the entire centre. This could be accomplished by exchanging information on a regular basis with staff in a consistent manner, through a regularly scheduled briefing session or standard communication log. Staff overwhelmingly expressed the desire for tools that would make the process of cleaning easier. This information is useful for public health officials and CCC management to consider when developing guidelines or providing guidance on cleaning methods and products that staff could use which are of high quality and efficiency but not seen as a challenge due to cost or time. Consultation with public health officials indicated that public health training for CCC staff would be the most important tool to assist staff. In contrast, although continuing education offered by public health officials was important, CCC staff felt that education and information should also be made available to parents. This training would increase the parent's understanding of enteric disease and provide information on topics such as symptoms, the importance of exclusion and proper prevention and control. Outbreaks may provide educational opportunities to bring staff and parents together for education and information by public health officials. Therefore, it is suggested that educational material be directed to parents, as well as CCC staff. Staff providing care to children on a daily basis were proud of the impact they had on children's development and education. The intrinsic value placed on child care is significant, but staff often felt that their work was undervalued. Staff need to be acknowledged for the work they do. Basic personal needs such as salaries, and paid sick days, should reflect the level of work and responsibility. Staff should not be penalized financially for taking time off when sick. Many of the staff relied on their wages to support themselves and their families and if not paid while ill, they might need to continue to work. Likewise, parents who keep their children home while ill receive no compensation for doing so, and in most cases they still pay for the days of care, even when their child is not there. Accommodations and incentives also need to be considered for families and parents, especially those in situations of financial need. In addition to staff specific needs, centres may require assistance to ensure additional staff for proper ratios, enhanced cleaning and ensuring substitute staff are available. The relationship between CCC staff in this study and local public health officials was very positive. During an outbreak, staff looked to public health officials to provide assistance and to reassure them that they took appropriate actions, particularly when dealing with parents. Staff stated a number of times that they referred parents to public health for further support and regarded public health officials as an authority figure when further assistance was required. Guidance should be supported by all groups involved. The relationship between staff, parents and public health is essential to ensuring proper response and management of enteric illness. As with all qualitative research, questions regarding representativeness and generalizability must be addressed. This study was restricted to a small geographic area in one province in Canada. Although the data and recommendations appear applicable to other jurisdictions, it would be useful to hold similar groups with staff in other jurisdictions to explore and confirm these themes further, taking potential regional differences into account. The one group that was mixed with staff from different centres demonstrated the highest level of interaction from the participants and also gave the greatest degree of contrast. Further research should consider maximizing the number of mixed groups to gain further interaction and insight which would allow for a deeper comparison between groups in analysis due to the potential for contrasting discussion. The relationship between staff and parents is important as it relates to monitoring and preventing enteric illness in CCCs. Recommendations from this study will have impact on parents and therefore further research with parents is required before any recommendations should be implemented. An intervention study could be conducted to test the effectiveness of the recommendations in reducing the amount of illness or improving the response and understanding of staff. Further work with public health officials to gain their perspectives regarding strategies to implement recommendations would also be of value. This qualitative assessment provides an enhanced understanding and appreciation of the perspective, practices and challenges that staff of CCCs experience in responding to enteric illness and outbreaks. In general, it was found that CCC staff are dedicated to and well informed about the children they work with and have a tremendous responsibility. Results from this study will be useful to public health officials responsible for developing tools and resources to further support or better inform current knowledge and practices for preventing and managing enteric disease. The recommendations from this study were made based on data directly from staff of CCCs and are designed to be practical and developed in further collaboration with them. The experience and knowledge CCC staff use to identify and take action for prevention and management of enteric illness clearly demonstrates their responsibility as gatekeepers of health among the children they care for.

Dating the time of viral subtype divergence

Precise dating of viral subtype divergence enables researchers to correlate divergence with geographic and demographic occurrences. When historical data are absent (that is, the overwhelming majority), viral sequence sampling on a time scale commensurate with the rate of substitution permits the inference of the times of subtype divergence. Currently, researchers use two strategies to approach this task, both requiring strong conditions on the molecular clock assumption of substitution rate. As the underlying structure of the substitution rate process at the time of subtype divergence is not understood and likely highly variable, we present a simple method that estimates rates of substitution, and from there, times of divergence, without use of an assumed molecular clock. We accomplish this by blending estimates of the substitution rate for triplets of dated sequences where each sequence draws from a distinct viral subtype, providing a zeroth-order approximation for the rate between subtypes. As an example, we calculate the time of divergence for three genes among influenza subtypes A-H3N2 and B using subtype C as an outgroup. We show a time of divergence approximately 100 years ago, substantially more recent than previous estimates which range from 250 to 3800 years ago.

Precise estimates are sorely lacking for dating the emergence and divergence of viral subtypes. Improved estimates equip epidemiologists and virologists to begin to correlate these important establishing events with historical demographic changes, geographical invasions and zoonoses, the transferring of a virus from one host species to another [7, 1, 25] . For example, archeological sequence data can furnish accurate dates and show that substantial genomic changes associate with geographical invasion and zoonosis [14, 17] . Further, the recent availability of viral gene sequences sampled at a pace commensurate with their rate of nucleotide substitution vastly augments the ability to rigorously infer the time scale of phylogenies and hence determine the time of the most recent common ancestor (TMRCA) for different viral types [18, 26, 6] . Systematic studies characterize the substitution process and substitution rate process of several classes of viral subtypes in, for example, Dengue, influenza subtype A, human immunodeficiency virus (HIV) and the virus responsible for sudden acute respiratory syndrome (SARS). For the last three viruses, a unique zoonotic transfer appears to co-occur with substantial changes in both the composition of nucleotides and amino acids as well as alterations in the rate of nucleotide substitution [15, 14, 1] . In Dengue, where a single subtype simultaneously inhabits two hosts (humans and Aedes aegypti) in a persistent zoonotic process, the introduction of the virus to new geographical environments associates with a dramatic increase in sequence diversity [25] . Unfortunately, no studies thus far analyze the rate of nucleotide substitution during either geographical invasion or zoonosis. Consequently, studies of the date of origins of viral subtypes must use strong a priori assumptions on the rate structure of nucleotide substitution. Two primary methods find use to date the time of viral subtype divergence. The most commonly employed approach determines the divergence time of subtypes using a molecular clock assumption (MCA) over an entire phylogeny [18, 21, 5, 26] . In its strict formulation, the MCA posits a proportional relation between the number of substitutions and the intervening time period over the entire phylogeny. Looser forms of MCAs require only that the proportionality hold along individual branches, with the rates across branches drawn from a pre-specified distribution [5] . Committed to some variant of the MCA, current algorithms then estimate the rate of nucleotide substitution over all taxa in a given set. Consequently, these methods provide inference most suitable for situations where sequence evolution follows a MCA (e.g. influenza A-H3N2 in human hosts, as in [9] ) or deviates from the MCA homogenously in time (e.g. perhaps influenza A in wild fowl, see [3] ). In considering divergence events between viral subtypes, even when the MCA well-approximates nucleotide substitution within a given subtype, the above methods may incorrectly infer the time of divergence across subtypes. By either assuming that a single rate of nucleotide substitution holds for the region preceding the common ancestor of each subtype or by smoothing the rate of nucleotide substitution over clades with different numbers of taxa, the adherence to a MCA prevents direct inference of the rate during subtype divergence. Suzuki and Nei (2002) propose an alternative, more heuristic method of estimation to counteract the problem of differing rates of substitution before and after zoonotic events [23, 25] . In these studies, the evolutionary models draw a distinction between the rate of substitution within a given subtype and the rate of substitution between subtypes. However, trouble arises since there are no methods for estimating the latter quantity. Consequently, the models assume that the rate of substitution for portions of the phylogeny between the subtypes equals the mean rate in the initial host species population. For instance, in dating the time of divergence between influenza B hemagglutinin and influenza C hemagglutinin-esterase, Suzuki and Nei use the rate of amino acid substitution for water fowl for the portions of the phylogeny previous to the TMRCA of these two proteins [23] . While this method may accurately reflect the rate within avian and human hosts, it neglects whatever additional changes in the rate of substitution are due to the process of zoonotic adaptation, likely leading to a substantial underestimation of the date of the TMRCA. The study here focuses on influenza, although the techniques are readily applied to other rapidly evolving organisms. Influenza has three types, A, B and C, classified based on serological analysis. To date, only type A sequences have been demonstrably associated with global pandemics [4] . Since modern surveillance began in the 1930s, type B has only been responsible for mild epidemics while type C has been nearly asymptomatic in human infection. Several subtypes of A, notably H1N1 and H3N2, are currently co-circulating in the human population. As the H1N1 and H3N2 subtypes may be as divergent from each other as they are from types B and C, we will refer to all types and subtypes simply as subtypes for the remainder of this paper. We select for this study three genes, coding for hemagglutinin (HA), the matrix protein (MP) and the non-structural protein (NS) responsible for interfering with host immune response. Subtype C has a hemagglutinin-esterase gene that is analogous to the hemagglutin gene in other subtypes [1] . We hence refer to the hemagglutinin gene generally and the hemagglutininesterase gene when referring specifically to the subtype C sequences. We present a simple estimation tool to determine the date of divergence among viral subtypes that overcomes the difficulties encountered with use of the MCA by measuring the pairwise rate of substitution between taxa. Our estimator derives from the triplet statistic developed in [26, 22, 13] , where each sequence member of the triplet draws from a different subtype. In this manner, we generate from each triplet an estimate of the rate of nucleotide substitution between the most recently diverged subtypes, and consequently provide an estimate of the TMRCA. This circumvents the problems posed by earlier methods by directly estimating the pairwise rate of nucleotide substitution over the set of pairs of sequences straddling the subtype divergence without any further rate assumptions other than the existence of a mean. However, this method is only capable of determining the rate between two subtypes where a third, more distantly related, subtype functions as an outgroup. This method thus trades the ad hoc rate assumptions of the previous methods with two implicit conditions: (i) that subtypes have a unique divergence and (ii) a third, comparable subtype is available to serve as an outgroup. In exchange, we arrive at a precise statistical measure of the TMRCA that converges as the number of taxa increases and is robust to the balancing of the numbers of taxa between different subtypes. We show that applying this method to dating the divergence of influenza subtypes A-H3N2 and B gives a time of diver-gence approximately 100 years before present, substantially more recent than previous estimates. To calculate the rate of nucleotide substitution, we require a measurement of the number of nucleotide substitutions occurring in a given time interval. Starting from a given set of aligned sequences {s 1 , ..., s n } for n taxa, we define the pairwise distance in number of substitutions to be the estimates {K ij } under a given model of nucleotide substitution. Naturally the unobservable true values {D ij } of the pairwise distances differ from their estimates {K ij }. To understand this difference, we associate each D ij with an error ε ij and assume that ε ij tends to zero as sequence lengths increase without bound. We further assume that the covariance between errors, cov(ε ij ; ε mn ), is bounded and known. For time measurements, we assume that each sequence is labeled by a sampling time t i given in consistent units. Since we know only the sampling time of a given sample up to the unit of time reported (day, month, year) we posit an uniform error ν i ~ U [0, 1] underlying each t i over the unit sampling interval. To complete the error structure specification we force the two forms of error (ν i and ε ij ) to be independent. Finally, for a set of three sequences (s i , s j , s k ) and their associated pairwise distances, we enforce a fixed topology among sequences, as shown in Figure 1 , via methods outlined in [26] . We augment the topology with the observed sampling times of the three sequences, α, the divergence time between the two sequences of interest and β, the divergence time of all sequences. When necessary for clarity, we write α ij to indicate the true time of divergence between sequences i and j. Under our triplet method, we aim to estimate the true rate of nucleotide substitution, p ij , between sequences s i and s j with an unobserved error δ ij . With respect to outgroup sequence k, an unbiased estimate is where the factor corrects for bias resulting from the time sampling error structure (see Appendix for derivation). We superscript to denote its weak dependence on outgroup sequence k. Dependence is weak as the path of evolution from t k to α is shared between the paths from sequence k to both sequence i and sequence j and hence largely cancels out in Equation 1. We make this transparent in the following derivation. For brevity, we consider only unobservable true values, ignoring error terms. Let u be the location on the triplet in Figure 1 corresponding to time α and let p xy be the true rate along the path connecting locations x and y. Then, as distance is rate multiplied by time, we have Subtracting the first equation from the second equation , which is equivalent to Equation 1. This derivation makes clear that the estimator (1) measures the rate along the path from sequence i to sequence j, with only incidental dependence on sequence k. The variance for the estimator (1) is well approximated by Further, we can estimate the time of subtype divergence α ( Figure 1 ) between sequences via We note that the term t i + t j -1 is used rather than t i + t j to account for the expected error coming from the uniformly The phylogenetic relationships between three sequences s i , s j and s k , sampled on dates t i , t j and t k respectively As nucleotide data increases without bound, K ij → D ij and → p ij , ensuring that → α ij . For finite sequence lengths, this relation ensures that . To gain an understanding of this estimator, we note that with a standard model of substitution (e.g. JC69, HKY85), a rate of substitution of 10 -4 (s/s/yr) and a sequence of 2000 nucleotides, the above estimator yields a standard error of approximately 23 years [20] . The above derivations express our rate and time estimates for a single triplet of sequences. We now consider estimates that combine information across multiple representative sequences from each subtype. For discussion, we label subtypes A, B and C (which are only incidentally the same as the labels for influenza) and we assume the topology in Figure 1 for these groups. We let n r , where r ∈ {A, where P α is the sum of the inverse variance of each estimate, . Having found , we estimate its variance by a bootstrap resampling of sequences from each subtype [8] . The computational efficiency of this estimator is on the order O(n 3 ) for a tree of n taxa. This is natural as each of the initial rate estimates is composed of information concerning three taxa. While the growth of computational expense in the number of taxa may appear unpleasant, in practice this algorithm is both fast and stable, owing to the absence of costly optimization procedures for parameter inference, and is able to handle data sets of thousands of taxa. The authors detail the computational efficiency of a similar statistic in [26]. As an example, for the data presented below all computations required only a few seconds on a desktop computer. We demonstrate the advantage of our triplet estimator through analysis of influenza A-H3N2/B subtype divergence using the hemagglutinin (HA), matrix protein (MP) and non-structural (NS) genes. Each analysis is performed on 60 gene sequences constructed from 20 genomes each drawn from influenza subtypes A-H3N2, B and C. We download these data along with their dates of sampling from the Los Alamos Influenza Database [16] . We perform sequence alignment using ClustalX [24, version 1.8]. For consistency with previous studies of A-H3N2 HA evolution, we use the HKY model of nucleotide substitution [10] . We use the TREBLE algorithm, which implements a MCA, on sets of sequences solely drawn from a single subtype to derive within-subtype rates. The phylogenetic tree, generated by TREBLE, for the HA gene is depicted in Figure 2 (± 0.48) × 10 -3 s/s/yr and the subtype C rate is 1.31 (± 0.33) × 10 -3 s/s/yr. Lastly, for the NS gene, the rates are similar to those of the MP gene. The subtype A-H3N2 rate is 2.14 (± 0.25) × 10 -3 s/s/yr, the subtype B rate is 1.92 (± 0.20) × 10 -3 s/s/yr, and the subtype C rate is 1.68 (± 0.51) × 10 -3 s/s/yr. Table 1 presents these results. Figure 3 provides histograms of the bootstrap distributions for all three genes and subtypes. Assuming a molecular clock within a subtype and with the rates above, we generated the corresponding dates of the TMCRA. Figure 3 shows histograms of the TMRCA estimates for different genes and subtypes. All genes are similar in dating the TMRCA for A-H3N2 to approximately 1965 (1964, 1965, and 1962 for HA, MP and NS genes, respectively). These dates are consistent with the emergence of the A-H3N2 subtype into global circulation dur-ing the 1968 pandemic [1] . Both the MP and NS genes date the TMRCA of subtype B to 1943, while the HA rate places the TMRCA at 1953. This latter value is inconsistent with the influenza B sub-epidemics of 1950-51 but is consistent with the emergence of the more lethal Victoria strain of influenza B in 1953 [11] . Each of these estimates has a standard error of approximately 2 years and so these discrepancies may be accounted by measurement uncertainty. The 10 year gap between the TMCRA suggested by the different genes can be explained by a reassortment event. Finally, the TMRCA of subtype C is calculated as 1952 and 1953 by the MP and NS genes, respectively, while the HA gene places the TMCRA at 1906. This nearly half century discrepancy suggests that the subtype C HA gene experienced a markedly different evolutionary history than either the MP or the NS gene. A biologically plausible explanation would be a reassortment event. Another possible explanation is that non-MCA rate behavior has lead to substantial bias in dating the TMRCA. We now compare the results from pairwise rate estimates across subtypes A-H3N2 and B with those from application of the MCA to the same data. These results are summarized in Table 2 and Histograms of the time of most recent common ancestor for subtypes A-H3N2, B and C, respectively, derived from molecular clock estimates on hemagglutinin (HA), matrix (MP) and nonstructural (NS) gene sequences This discrepancy between the two sets of estimates of the TMRCA likely owes to the inability of the MCA to integrate information from the period of evolution between the two subtypes, leading to a substantial underestimate of the rate of substitution, and consequent underestimation of the date of the TMRCA. We present a new method for ascertaining the rate of nucleotide substitution between subtypes and apply this method together with traditional MCA methods to date the divergence of influenza subtypes A-H3N2, B, and C. We use three genes, HA, MP and NS, to date two types of divergence events: the time of the most recent common of each subtype and the time of divergence between two subtypes, A-H3N2 and B. For the former event type, we show that the three genes are loosely consistent in their dating of the TMRCA of the subtypes, with the notable exception of the HA-derived estimate of subtype C's TMRCA approximately 50 years before the MP-and NS-derived estimates. This discrepancy may indicate either that subtype C's hemagglutinin-esterase gene engaged in a biologically significant event, such as reassortment, or that MCA estimation does not adequately model the evolution of the gene. For the divergence between subtypes A-H3N2 and B, previous studies using the MCA generally place a time of divergence of several hundred years ago, ranging from the 16th to early 19th centuries. Other analysis have yielded estimates of 3600 years ago [23] . In the current study, application of the MCA yielded estimates in the last half of the 18th century. However, applying the pairwise rate estimate developed above we find uniformly, across genes, that the divergence likely occurred in the very early 20th century. The discrepancy between these two measures is likely due to the increased modeling flexibility of the pairwise rate estimate relative to the MCA. This discrepancy between the rates and corresponding TMCRA estimates has important biological consequence. The phylogenetic divergence between subtype A-H3N2 and B corresponds to a subspeciation event for the virus. The results in this study indicate that the process of speciation is not neutral but instead a period of rapid and intense genetic change. The three genes studied here consistently show large acceleration in the rate of nucleotide substitution for the divergence period relative to the rates observed within a stable subtype. This study gives strong evidence that, at least for influenza viral subtype divergence, the process of subspeciation is associated not just with large genomic changes but also with an accelerated, finite process of adaption. Histograms of the time of most recent common ancestor of subtypes A-H3N2 and B, derived from molecular clock estimates (light grey) and pairwise estimates (dark grey) on hemagglutinin (HA), matrix (MP) and nonstructural (NS) gene sequences Assuming that the more recent estimate is correct, a subsequent question is whether or not a pandemic or epidemic associates with subtype A-H3N2/B divergence. In the twentieth century, all influenza pandemics associate with the emergence or reemergence of subtypes (A-H1N1 in 1918, A-H2N2 in 1957 and A-H3N2 in 1968). Serological analysis indicates that the 1897 pandemic was likely due to subtype A-H2N2. However, the pandemic of 1900 is of uncertain type, although it is commonly reported in the literature as being due to A-H3N2 [4] . The above analysis suggests that it is possible to postulate that the cause of this pandemic is due to the emergence of subtype A-H3N2 or B. As noted above, we condition the results presented here on a specific sequence alignment. As the question under consideration concerns the divergence of specific genes and proteins over a (presumably) long time scale, the capacity to generate reasonable alignments diminishes with increasing time of divergence between types, conditional on the rate of substitution. We find that for the hemagglutinin gene, a proportion of sequence alignments support the split of subtype B from subtype C after the split between subtypes A-H3N2 and B, in opposition to the topology enforced in our analysis. Hence, to some unknown degree, our analysis is necessarily biased by the choice of alignment. This suggests that improved dating can be found by integrating estimation procedures over an ensemble of alignments [19] . The pairwise estimate method presented above is accurate in the scale where is the total time over the phylogeny and p is mean rate over the phylogeny [26] . This relation dictates that as divergence events become more remote the ability of the triplet method to resolve the time of divergence diminishes. While this limit prohibits the calculation of remote divergence events, the example presented above lies within the appropriate scale. In place of a specific MCA, the estimates presented here directly calculate the rate of substitutions between taxa from different viral subtypes. As such estimates span paths between subtypes, they simultaneously capture the rate evolution along branches both within and between subtypes. From these estimates, we are able to directly infer the time of divergence between subtypes. As a trade-off for limited MCAs, the method requires an outgroup subtype to function as an origin relative to the subtypes under consideration. We feel that the triplet method provides a simple and widely applicable way to calculate the dates of divergence of rapidly evolving organisms without the pitfalls of the MCA. We present a simple method for calculating the time of viral subtype divergence that does not assume a molecular clock over the entire phylogeny. Additionally, the estimator of this method, a weighted sum of pairwise estimates, furnishes a defined variance for the time of the most common ancestor between subtypes. As a tradeoff for this increased precision, the structure of the triplet statistic requires an outgroup set of sequences, usually a closely related subtype. We apply this estimator to the case of influenza subtype divergence, considering three genes. We show that the estimated divergence time of subtypes A-H3N2 and B is more than a century later than those calculated with a molecular clock. Since we assume that the ν and ε structures are independent, the right side of the equation can be further reduced, yielding Let Δt = t i -t j . The final expectation on the right hand side resolves by direct integration, We note that as the sampling time is independent of the rate of nucleotide substitution, the error increases in proportion to the magnitude of the initial statistic. We can then create a new, unbiased statistic by counterbalancing the original statistic with this factor, making a new statistic

Discovery and Development of Toll-Like Receptor 4 (TLR4) Antagonists: A New Paradigm for Treating Sepsis and Other Diseases

Sepsis remains the most common cause of death in intensive care units in the USA, with a current estimate of at least 750,000 cases per year, and 215,000 deaths annually. Despite extensive research still we do not quite understand the cellular and molecular mechanisms that are involved in triggering and propagation of septic injury. Endotoxin (lipopolysaccharide from Gram-negative bacteria, or LPS) has been implicated as a major cause of this syndrome. Inflammatory shock as a consequence of LPS release remains a serious clinical concern. In humans, inflammatory responses to LPS result in the release of cytokines and other cell mediators from monocytes and macrophages, which can cause fever, shock, organ failure and death. A number of different approaches have been investigated to try to treat and/or prevent the septic shock associated with infections caused by Gram-negative bacteria, including blockage of one or more of the cytokines induced by LPS. Recently several novel amphipathic compounds have been developed as direct LPS antagonists at the LPS receptor, TLR4. This review article will outline the current knowledge on the TLR4-LPS synthesis and discuss the signaling, in vitro pre-clinical and in vivo clinical evaluation of TLR4 antagonists and their potential use in sepsis and a variety of diseases such as atherosclerosis as well as hepatic and renal malfunction.

Bacteria are classified into two groups based on a staining procedure (1) . This staining response is a consequence of the composition of their membranes. Gram-positive bacteria present a multi-layered, cross-linked polymer of peptidoglycan surrounding their plasma membrane, whereas Gramnegative bacteria have essentially a monolayer (1) . The Gram-negative outer membrane is an asymmetric lipid bilayer interspersed with proteins. The lipid of this outer leaflet is almost exclusively constituted by LPS molecules. Bacterial infection can be life threatening, requiring the host organism to develop a system to respond to this insult. The innate immune response is the first line of defense against infectious agents and is devoted to recognize highly conserved pathogen motifs in lipopeptides, DNA, dsRNA, ssRNA, specific proteins and LPS. These motifs are known as pathogen-associated molecular patterns (PAMPs) (2) . Lipopolysaccharide is composed of three distinct domains, lipid A, a short core of oligosaccharide and the O-antigen polysaccharide (Fig. 1) . The lipid A domain is the bioactive component and is recognized during human infection. The composition of the O-antigen varies between different Gram-negative bacterial strains. The presence or absence of O chains determines whether LPS is considered rough or smooth (3) . Full length O chains would render the LPS smooth while the absence or reduction of O-chains would make the LPS rough (3, 4) . Lipopolysaccharide is a potential drug target since its presence is critical in membrane stability and also it plays a prominent role in raising an immune response (2) . LPS triggers the release of many inflammatory cytokines, in particular TNFα, interleukin-1β and IL-6, and it has been implicated as the etiological agent of a variety of pathologies ranging from mild (fever) to lethal (septic shock, organ failure and death) (5) . Thus the structure, function and biosynthesis of LPS have been areas of intense research in the last decade (6) . The receptors capable of recognizing the pathogenassociated molecular patterns are Toll-like receptors (TLR) and scavenger receptors. Ten members of the TLR family have been identified in humans (7) . The Toll was originally described as a type I transmembrane receptor that controls the embryonic dorsal-ventral pattern of Drosophila (8) . In fact this pioneering work identified a group of ten different genes which when deleted produced qualitatively similar phenotypes. Null mutations on any of these genes lead to a failure to differentiate patterns on the dorsoventral axis and resulted on embryonic lethality. The identification of the sequence of Toll led to the recognition that its carboxyl terminal domain was significantly related to that of the vertebrate interleukin-1 receptor (IL-1R) (8) . IL-1R activation is part of a cascade of events linked to an acute phase response to infection. This suggested that TLRs could not only be involved in development but also in the initial responses to infection in vertebrates. This hypothesis received further support from the work of Lemaitre et al., who found that Toll and other genes from the dorsal group played a role in innate immune responses to pathogenic fungi and bacteria (9) . The TLRs belong to a cluster of molecules called the IL-1R/TLR super-family characterized by the presence of cytoplasmic Toll/IL-1R (TIR) domains (10) . The three subgroups are: the IL-1R (which present extracellular immunoglobulin domains), the adapter subgroup (cytoplasmic proteins without extracellular region) and the TLRs (9) . TLRs are type I transmembrane proteins with extracellular amino terminus and a carboxy terminal intracellular domain. The extracellular domain of the TLR4 contains over 600 amino acids and is highly polymorphic compared with the transmembrane and cytosolic domains (6) . The TIR domain, composed of three highly conserved regions, contains 150 amino acids and modulates protein-protein interactions between the TLRs and the adaptor proteins involved in the signal transduction cascade (10) . Unlike other receptors, TLRs do not have an enzymatic activity (6) . Researchers have identified at least fifteen different negative regulators of the TLRs, including MyD88s (a short form of MyD88), IRAKM, suppressor of cell signaling-1 (SOCS1), nucleotidebinding oligomerization domain 2 (NOD2), phosphatidylinositol-3-kinase (PI3K) and Toll-interacting protein (TOLLIP) (11) . The TLR activation leads to responses that involve the induction of new genes via transduction pathways such as NFκB and AP-1 (9) . The discovery of TLR lead to the understanding that an adaptive response mediated by antibody responses and T cell activation is tightly coupled to a second unknown process that requires the presence of microbial extracts (2, 7) . Toll-like receptor 4 (TLR4) is the central signaling receptor for LPS in mammals (12) . The current knowledge on the structure and function of the TLR4 has opened the possibility to develop new drug targets to fight sepsis and other diseases associated with this signaling molecule. TLR4 was identified as the first human homologue of the Drosophila Toll (13) . TLR4 not only engages LPS but it recognizes an envelope glycoprotein encoded by mouse mammary tumor virus (MMTV) (14) . In addition, TLR4 recognizes ligands such as heat shock proteins and EDA (extracellular domain A) in fibronectin (15, 16) . TLRs activate a potent immunostimulatory response which needs to be tightly controlled. TLRs homo o heterodimerize upon ligand binding whereas TLR4 and TLR9 homodimerize (6) . TLR signaling involves a family of adaptor proteins which recruit downstream protein kinases which activate transcription factors such as nuclear factor-kB (NF-κB) and members of the interferon (IFN)-regulatory factor (IRF) family (10) . LPS signaling involves the binding of the LPS-binding protein (LBP) to LPS; this interaction leads to a disruption of LPS aggregates (10) (Fig. 2 LPS signaling, modified from (10) with permission). Upon ligand binding there is the formation of a TLR4 complex with CD14. CD14 was the first molecule shown to enhance LPS signals (17) . Interestingly TLR4 does not require CD14 to trigger epithelial signaling to uropathogenic E. coli since bladder cells do not express CD14 (18) . In addition a small molecule, myeloid differentiation 2 receptor (MD-2), participates in this complex by associating with the TLR4 extracellular domain (19) . MD-2 binds to the LPS monomer and is sensitive to the acylation pattern of the lipid A moiety. Association of the MD-2:LPS complex to the ectodomain of the TLR4 finally transduces the signal through the association of intracellular TIR domain, recruiting the adapter proteins triggering the signaling cascade (20) . In a similar way to TLR2, TLR4 uses the myeloid differentiation primary-response gene 88 adapter like protein (MAL) as a bridging adaptor to recruit the myeloid differentiation primary-response gene 88 (MyD88) to activate the NF-κB, p38 and JNK/MAPK pathways via TRAF6 (9) . MAL is recruited to plasma membrane microdomains containing the phospholipid PtdIns (4,5)P 2 (phosphatidylinositol-4,5-bisphosphate). MAL subsequently recruits MyD88 (20) . Another pathway activated by TLR4 involves TRIF-related adaptor molecule (TRAM). Similar to MAL, TRAM is also membrane proximal and requires myristoylation to lodge into the membrane. TRAM recruits the Toll/interleukin-1 receptor (TIR)-domain-containing adaptor protein inducing interferonβ (TRIF) which activates the tumor-necrosis factor-receptorassociated factor 3 (TRAF3), TRAF6 and receptor interacting protein 1(RIP1). Recent work with CD14 knockout mice suggested that TRL4 can function in two ways: one where full signaling occurs in the presence of CD14 and one limited to MyD88-dependent signaling (21) . In addition to blocking the intracellular LPS signaling there are other means to modulate the endotoxin response. Approaches to alleviate the morbidity and mortality of patients associated with severe sepsis and septic shock include: (a) neutralizing LPS or blocking initial LPS-signaling events by preventing the generation of cell-surface signals, (b) blocking the intracellular signals induced by endotoxin or the synthesis of cytokines and other cellular mediators, (c) inhibiting the release of cytokines (Il-1, IL-6 and IL-8) and cellular mediators, (d) blocking the TNF-α and IL-1 receptors to the cellular mediators on their responsive target cell and (e) inhibiting downstream pathophysiological events such as acute respiratory distress or aberrant blood clotting (22) . TLR mediated innate and/or adaptive immune responses play an important role in a variety of diseases, including sepsis, infectious disease, atherosclerosis, kidney failure, liver disease, pulmonary disease and myocardial ischemia/reperfusion injury (5, (23) (24) (25) (26) (27) (28) . TLRs are expressed in a variety of cell types including immune and non-immune cells. In addition, the capability of these receptors to recognize PAMPs is indicative of their distinct roles in infection, inflammation and tissue damage (29) (Fig. 3 ). According to the definition made by ACCP/SCCM Consensus Conference in 1992, sepsis is known to be an early syndrome that may progress to a pathologic state manifested by hypotension and hypoperfusion known as septic shock. LPS has been associated with sepsis and the high mortality rate seen in septic shock (5) . However, it is the exaggerated host response to the systemic release of endotoxin that accounts for septic shock from Gram-negative bacteria (23) . TLR4 up-regulation in non-immune cells after initial TLR mediated immune response may trigger secondary responses such as activation of endothelial cells that promotes the production of adhesion molecules, followed by macrophage infiltration and vascular permeability during infection (30) . This cascade may result in a systemic septic syndrome including tissue perfusions, an imbalanced coagulation cascade and organ failure (31) . Atherosclerosis is an inflammatory disease where activated cells are involved in its initiation and progression. Guha and Mackman (32) have shown that activated TLR4 elicits the production of inflammatory cytokines and chemokines. Edfeldt et al. (33) have also found that TLR4 is prominently expressed in endothelial cells of human atherosclerotic lesion, but poorly expressed in normal human arteries. In the early atherosclerotic lesion, LPS and other ligands can stimulate the TLR4 expression on macrophages. The activated receptors can then initiate the signaling cascade that induces the expression of inflammatory cytokines, proteases, and cytotoxic oxygen and nitrogen radicals. These entities further speed up the progression of the atherosclerotic lesion (34) . In advanced atherosclerotic lesion, LPS can induce the proliferation of vascular smooth muscle cells, as well as the expression of elastin-degrading enzyme via TLR4 (35) . Besides that, in response to chemokines, more smooth muscle cells will also migrate to the sites of the lesions (36). These predominant changes cause the accumulation of cells, extracellular matrix components, thickening of the intima, as well as the deformity of the arterial wall. Furthermore, TLR4 signaling might also be involved in atherosclerotic plaque destabilization. Grenier and Grignon have demonstrated that LPS induces the expression of matrix metalloproteinase-9 (MMP-9) by TLR4 in macrophages; MMP-9 has been shown to degrade collagen fibrous cap, thus predisposing plaque to rupture (37) . In many forms of liver diseases such as alcoholic or nonalcoholic liver disease, liver failure and inflammation are the result of a cascade of insults which result in hyper-activation of inflammatory pathways and liver injury (26, 27) . Velayudham and colleagues (26) have shown that there is an up-regulation of TLR4 receptors in liver granulomas and LPS induced liver injury. Pathogen-induced TLR4 activation also activates reac- tive oxygen species (ROS), which is a major source of acute hepatocyte injury and death in the liver. Up-regulation of peripheral blood monocyte expression of TLR4 also occurs in patients with chronic hepatitis C (38) . In addition, endogenous gut-derived bacterial LPS have also been implicated as important cofactors in the pathogenesis of liver injury. Within the liver, LPS binds to LPS-binding protein (LBP), which then facilitates its transfer to membrane CD14 on the surface of Kupffer cells in the liver (39) . Moreover, TLR4 can also interact with a protein ligand released from damaged hepatocytes to extend an existing injury in the liver (40) . In other studies, there is evidence that high-mobility group box 1 (HMGB1) can interact with both TLR2 and TLR4 to induce an inflammatory response during liver ischemia/reperfusion (IR) injury similar to that initiated by LPS (41, 42) . HMGB1 is an intracellular protein present in many species that functions in regulation and modulation of gene transcription (42) . HMGB1 is released readily from necrotic or damaged cells, which may signal through TLR4 the presence of advancing tissue injury, initiating an inflammatory response that further damages viable cells (42) . The restoration of blood flow to the ischemic heart has often caused myocardial ischemic/reperfusion (MI/R) injury. An inflammatory response triggered by MI/R injury can irreversibly cause damage to the viable tissue surrounding the infarct, thereby further extending the injury. It is still unclear how innate immune signaling pathways are initiated during MI/R injury (43) . However, TLR4, which is also present in cardiomyocytes, has been thought to play a role in mediating MI/R injury. Schuster and Nelson (28) have shown that TLR4 receptor is up-regulated in response to myocardial injury. Furthermore, Shimamoto and colleagues have shown that TLR4 activates NF-κB-dependent transcription of inflammatory cytokine genes in MI/R injury. The TLR4-mediated injury appears to occur through activation of c-Jun NH 2terminal kinase (JNK) and translocation of NF-κB (41) . It is also believed that TLR4 recognizes endogenous molecules that are exposed during cellular injury and extracellular matrix remodeling, independent of pathogen invasion (44) . Thus, inhibition of TLR4 signaling pathway may be a potential therapeutic target to treat the myocardium damage in the ischemia/reperfusion setting. Acute renal failure (ARF) occurs in close to 5% of hospital admissions, and is a leading cause of morbidity and mortality. A common cause of ARF is sepsis, which results from overwhelming infection (25) . Cunningham and colleagues (25) have shown that LPS insult leads to renal cell apoptosis and renal neutrophil infiltration. Tubular epithelial cells of the kidney are among the nonimmune cells that express TLR1, TLR1-2, TLR1-3, TLR1-4, and TLR1-6, suggesting that these TLR might contribute to the activation of immune responses in tubulointerstitial injury (45) . In addition, receptors such as TLR4, TLR2, CD91 and the receptor for the advanced glycation end-products (RAGE), allow leukocytes and renal cells to recognize molecules released by injured cells. These receptors are sentinels for tissue necrosis (46) . Upon stimulation with LPS in renal infection or other endogenous ligands from necrotic tubular cells, the activated TLR4 has been shown to specifically stimulate the NF-κB pathway in response to oxidative stress (47) . Furthermore, TLR4 activation on tubular epithelial cells and circulating immune cells leads to secretion of cytokines and chemokines that either directly or indirectly contributes to renal injury. Inflammatory bowel disease (IBD) is a medical condition that predominantly affects the gastrointestinal tract (48) . De Jager et al. showed that TLR4 and its signaling molecule TIRAP affect susceptibility to IBD (49) . Recent studies have shown that TLR4 −/− and MyD88 −/− knockout mice tend to be more prone to severe dextran sulfate sodium-induced colitis than their wild-type littermates (50) . Interestingly, CRX-526 a TLR4 antagonist has been shown to prevent an inflammatory disease in the dextran sulfate sodium and mdr1a −/− /1b −/− deficient mice models (51) . To explain these contradictory results we have to consider that constitutive signaling through TLR4 may result in the production of tissue protective factors such as IL-6 and TNF-α (49) . This is the scenario in the MyD88 −/− knockout mice, while in the case of the CRX-526 we may have selective downregulation of one of the TLR-4/ LPS signaling pathways. Simpson and colleagues observed an increased expression of TLR2, TLR4 and CD14, as well as the proinflammatory cytokines IL-8 and IL-1β in neutrophilic asthma and bronchiectasis patients compared to controls. These groups also had higher airway endotoxin levels than the control group (23) . In another study, there was also an increased pulmonary expression of inflammatory cytokines occurring in the lung during experimental endotoxemia. The cytokine production further contributes to acute lung injury (ALI) and acute respiratory distress syndrome (ARDS) (52, 53) . Baumgarten and colleagues showed that LPS induces pro-inflammatory cytokines in the lung via the TLR4/CD14 signaling cascade, suggesting a role of the innate immune response in the pathogenesis of ALI/ARDS (52) . LPS biosynthesis occurs by two distinct, yet convergent pathways: one for the lipid A core and another for the polysaccharide O antigen. After independent synthesis, the two parts are ligated together to complete the LPS molecule (4) . Amongst the most important TLR4 antagonists developed so far we have CRX-526, E5531 and E5564. Aminoalkyl glucosaminide 4-phosphates or AGPs are a class of lipid A mimetics in which the reducing sugar of lipid A has been replaced with an N-acylated aminoalkyl aglycon unit (54) . The AGPs contain an L-serine-based aglycon unit as well as three (R)-3-n-alkanoyloxytetradecanoyl residues comprised of even-numbered normal fatty acyl chains between 6 and 14 carbon atoms in length. All members of this family have 14 carbon atoms in the "primary" fatty acid, which is the -hydroxy fatty acid attached directly to the AGP backbone. -Hydroxymyristic acid is the most common of the primary fatty acids present in lipid A. These compounds were used in a variety of cell-based and an in vivo model to determine structure-activity relationships related to AGP acyl chain length and stimulation via TLR4 (54) . Figure 4 shows the chemical structure of two AGPs: MPL and CRX-526. The structure of CRX-526 differs significantly from monophosphoryl lipid A (MPL) and other TLR4-agonist AGP in the length of its secondary fatty acyl chains (SAC): for instance CRX-526 contains 3 SAC of 6 carbons in length, whereas MPL and other AGP, which signal through the TLR4 complex, contain SAC of >10 carbons in length (55) . The synthesis of AGPs has been described elsewhere (55) . Briefly, the AGPs were prepared by a highly convergent method, which allowed chemical differentiation of the hydroxyl and amino groups and sequential introduction of the (R)-3-n-alkanoyloxytetrahexanoyl residues. The AGPs were purified by flash chromatography on silica gel (to >95% purity) and analyzed as a triethylammonium salt by standard analytical methods. For stimulation in vitro, the AGPs were formulated in water containing 216 μg/ml dipalmitoylphosphatidyl choline [aqueous formulation (AF)], 0.2% triethanolamine (pH 7.4), or in 2% glycerol (i.v. formulation). (53) . The structure of E5531, an analog of the lipid A of Rhodobacter capsulatus, is presented elsewhere (56) . The structure of E5564 is depicted in Fig. 5 . The crystal structure of the TLR4-MD2 complex with bound E5564 was recently published (57) suggesting that the mechanism of action of E5564 is binding through a large internal pocket in MD-2. An important consideration in the development of assays is to determine the efficacy and toxicity of the compounds and the potency required to obtain a biological effect. A high potency is a key factor that would limit the high cost of synthesis and purification of such a compound. To illustrate the challenges in developing a synthetic endotoxin receptor antagonist we present the case of E5564. In vitro, E5564 dose-dependently inhibited LPS-mediated activation of primary cultures of human myeloid cells and mouse tissue culture macrophage cell lines as well as human or animal whole blood at nanomolar concentrations as measured by production of tumor TNF-and other cytokines. E5564 also blocked the ability of Gram negative bacteria to stimulate human cytokine production in whole blood. In vivo, E5564 blocked induction of LPS-induced cytokines and LPS or bacterial-induced lethality in primed mice. E5564 was devoid of agonistic activity when tested both in vitro and in vivo and had no antagonistic activity against Gram positive-mediated cellular activation at concentrations up to 1 μM. E5564 blocked LPS-mediated activation of nuclear factor-B in TLR4/MD-2transfected cells. In a mouse macrophage cell line, activity of E5564 was independent of serum, suggesting that E5564 exerts its activity through the cell surface receptor(s) for LPS, without the need for serum LPS transfer proteins. Similar to E5531, another lipid A-like antagonist, E5564 associates with plasma lipoproteins, causing low concentrations of E5564 to be quantitatively inactivated in a dose-and time-dependent manner. However, compared with E5531, E5564 is a more potent inhibitor of cytokine generation, and higher doses retain activity for a period of time likely sufficient to permit clinical application. These results indicate that E5564 is a potent LPS antagonist and lacks agonistic activity in human and animal model systems, making it a potentially effective therapeutic agent for treatment of disease states caused by LPS. Compared with E5531, E5564 is structurally and synthetically less complex, yet seems to possess superior activity and pharmacological characteristics. Although E5531 demonstrated potent inhibition of LPS when added to blood in vitro and in vivo, activity decreased as a function of time. This reaction has been shown to be due to interaction of E5531 with plasma lipoproteins (58) (59) (60) . E5564 is an inhibitor of LPS-mediated stimulation of responsive cells in vitro and in vivo as measured by production of cytokines, as well as morbidity and mortality associated with LPS poisoning in animal models. Because E5564 is a structural analog of the lipid A portion of LPS, it is logical to hypothesize that the antagonist interacts with the same signaling components that bind to LPS such as the soluble serum proteins LBP and sCD14, as well as membrane-associated CD14 and perhaps the TLR4/MD-2 receptor complex. E5564 blocked LPS/ sCD14-induced reporter activity in TLR4/MD-2-expressing HEK293 (61), but not TLR2-mediated signaling by heatkilled S. aureus. These findings indicate that E5564 selectively inhibits LPS signaling via TLR4/MD-2. However, a limitation to this model system is that LPS requires the presence of sCD14 for cellular activation, making it difficult to determine whether E5564 blocks LPS binding to sCD14 or TLR4/MD-2. Results from experiments indicated that serum components did not affect the potency of E5564, indicating that they are not critical to E5564 antagonistic activity (61) . Further support of the hypothesis that interaction of E5564 at CD14 does not play a key role in its activity comes from a previous study by Lien et al. describing the activity of novel synthetic acyclic lipid A-like agonists that activate TLR4/MD-2 in the absence of CD14 (62) . E5564 inhibited the actions of these agonists under serum-free conditions. Taken together, these lines of evidence make it tempting to speculate that E5564 binds to TLR4/MD-2 complex, thereby blocking LPS binding or transmembrane signaling. The downstream effect of inhibiting the initial signaling by LPS seems to be an inhibition of all LPS-induced cytokines measured, including TNF-, IL-1, IL-6, IL-8, IL-10, and nitric oxide, which was measured in cultured cells, whole blood, and in vivo. Recently the crystal structure of the TLR4/MD-2 and E5564 has been described confirming the physical interaction of these molecules (57) . Comparisons of antagonistic potency in cells cultured in 10% serum versus whole blood allow us to determine whether the high concentration of proteins/lipoproteins present in serum inhibit E5564 activity. In all systems but the rat, antagonistic activity of E5564 in cultured cells was within fourfold that measured in high serum (blood) compared with assays done in low-serum conditions (cultured cells or monocytes) (60, 63) . This indicates that serum has little or no inhibitory effect on antagonistic activity under these in vitro conditions. However, extended incubations in whole blood demonstrated that activity of E5564 was measurably reduced. Other studies indicate that like E5531, E5564 is not rapidly metabolized, but binds to lipoproteins, and time dependently loses antagonistic activity (64) . The observation that lipoproteins reduce drug activity may explain the poor activity of E5564 in rat blood that has relatively high lipoprotein content. During extended incubation in whole blood, E5564 retained activity better than similar concentrations of the first-generation antagonist E5531. Based on the proposed mechanism of action as a cell surface antagonist, it is likely that E5564 can completely block cellular activation by LPS. This block is achieved by concentrations of E5564 as low as 10 nM (14 ng/ml) in vitro, and at doses of 1 mg/kg or less in animal models challenged with lethal LPS doses. Both LPS-challenge model and infection model use animals that have been sensitized or primed to LPS by previous infection with BCG, increasing cytokine response and lowering the threshold lethal dose of endotoxin (61) . All animal models of sepsis and infection have been criticized for their inability to closely mimic human sepsis. The primed model is the most relevant to the study of endotoxin antagonists such as E5564. It is well known that compared with humans, unprimed rodents such as rats and mice and primates demonstrate a profound insensitivity to endotoxin, requiring endotoxin doses as high as milligrams per kilogram, whereas humans demonstrate reproducible response to endotoxin at doses as low as 2 ng/kg. This argues that either LPS contributes only weakly to the inflammatory process in animal models, or that response to infection occurs only after the level of infection is very high, representing a process different from that in more LPSsensitive species such as humans (61) . Even in primed animal models, lethal doses of LPS are high, approximately 100 μg/kg, generating estimated plasma concentrations of ∼1 μg/ml. These plasma levels are still >100-fold that found in even the most extreme cases of human sepsis (65) . Because the dose of E5564 required to protect against LPS is proportional to the LPS challenge dose, studying E5564 in these animal models indicates that E5564 can be a safe and effective antagonist even under these extraordinary conditions. E5564 is approximately tenfold better in human blood than mouse blood (IC 50 =1.6 nM in human whole blood; Table I versus ∼20 nM in mouse whole blood; Table II) . Complete block of cytokine response by 10 nM E5564 in blood extrapolates to a human dose of approximately 100 μg in a 70-kg individual. Recent studies have supported this extrapolation by finding that a dose of 100 μg of E5564 given to normal volunteers over 30 min completely blocks response (signs, symptoms, and cytokines) to a dose of 4 ng/kg endotoxin administered at the midpoint of the E5564 infusion (66) . In vitro and ex vivo assays have found that low concentrations of E5564 time dependently lose ability to inhibit LPS response. In light of these observations, it is perhaps not surprising that low doses of E5564 demonstrate a time-dependent loss of activity after administration into normal volunteers. This loss in activity is overcome when E5564 doses are increased (67) . Phase I clinical safety and tolerability assays indicate that E5564 is safe and except for the occurrence of phlebitis, well tolerated at doses up to 252 mg administered over 72 h. At this dose, in vivo antagonistic activity is retained for at least an additional 72 h after discontinuing infusion. This leads us to believe that sufficient therapeutic activity can readily be administered to patients (67) . The safety and efficacy of E5564 are currently been analyzed in a phase III randomized controlled study. TAK-242 has been demonstrated to suppress LPSinduced inflammation (68, 69) . Recently, TAK-242 has been shown to almost completely suppress production of nitric oxide or TNFα induced by LPS in mouse RAW264.7, human U937 and P31/FUJ cells (70) . In a HEK293 cell model where TLR4, MD-2 and CD14 were co-expressed, this antagonist showed specificity to TLR4 as other TLRs, TLR1/2, TLR2/6, TLR3, TLR5, TLR7 and TLT9 were not affected by this drug (70) . Sepsis is a major cause of high mortality rate in intensive care units in the USA (71) . Severe sepsis usually leads to organ failure. Currently, over 30 pharmaceutical products have been in the development stage to treat this condition, yet only few have reached the market (72) . Many of these target specific inflammatory mediators have been unsuccessful because of the complex nature of sepsis. For the treatment of sepsis, there are a few products that are being investigated in clinical studies via blocking different mechanisms of the body's innate immune system. Eli Lilly's Xigris ® was one of the few drugs currently available on the market to treat sepsis. Xigris ® is a recombinant human activated protein C that has anti-inflammatory, anti-thrombotic and pro-fibrinolytic properties to block the coagulation cascade which plays a critical role in the development of organ failure due to sepsis (73) . In addition, simvastatin and atorvastatin had also shown to have some non-specific anti-inflammatory effects contributing to their clinical benefits in treating sepsis (63) . However, statins are currently not been considered as a treatment for sepsis. To find a more specific target, scientists have identified TLR4 as one of the candidates in blocking the innate immune system. Only two TLR4 antagonists, E5564 and TAK-242, have made far into the clinical phase (Table III) . In Wong, et al., determined the safety and tolerability of E5564 following a 30-min intravenous infusion in healthy male volunteers (74) . This was a single-center, randomized, double-blind, placebo-controlled, sequential-group, singledose study of E5564. The drug dose levels used were 350, 1,000, 2,000 or 3,500 μg. All doses of E5564 presented a long pharmacokinetic half-life and short in vivo pharmacodynamic half life which generally less than several hours when it is coadministered with LPS in healthy volunteers (74) . The C max and AUC (area under the curve) of E5564 increased in a dose-dependent manner. E5564 pharmacokinetics was characterized by a slow clearance (0.67-0.95 ml h −1 kg −1 ), a small volume of distribution (41-54 ml/kg), and a relatively long elimination half life (42-51 h) in healthy male volunteers. Thus, to overcome this low PD, the doses of E5564 given to the volunteers needed to be adjusted. In summary, all doses were demonstrated to be safe and well tolerated. Safety and tolerability assessments included monitoring and questioning of the subjects about adverse events, physical examinations, clinical laboratory tests (including hematology, blood chemistry, and urinalysis), and vital sign measurements (including supine and standing pulse rate and blood pressure), and 12-lead electrocardiograms (ECGs) and cytokine concentration testing. In this study, E5564 inhibited LPS-induced tumor necrosis factor-α in a dose-dependent manner, and at the higher doses (2 and 3.5 mg), antagonistic activity was measurable up to 8 h post-infusion. E5564 lacked LPS-like agonist activity at doses up to 3.5 mg (74). In another study of healthy volunteers with experimental endotoxemia, Lynn et al. (66) found phlebitis was only associated with 72 h continuous intravenous infusion of E5564 but not with four hour infusion of E5564 into a peripheral vein. In this study the authors explored the possibility of extended pharmacokinetic activity of E5564. The infusion period was changed from bulk dosing to a 4-and 72-h infusions of E5564 into normal volunteers. They observed that at 4 h infusion of E5564, 3 mg/h completely blocked endotoxin administered 8 h post-dosing. Additionally, they observed that administration of 3.4 mg of E5564/hX72 h completely blocked the effects of endotoxin challenge at the end of dosing (72 h), and at 48 and 72 h post dosing. A lower dose of E5564 of 2 mg was also studied, and they found that 0.5 mg/h×4 h, ameliorated but did not block most effects of endotoxin 8 h post-dosing. This work also studied the effect of varying plasma lipoprotein content on E5564 activity in subjects who have high or low cholesterol levels (>180 or <140 mg/dl) after a 72 h infusion of 252 mg of E5564. The distribution of E5564 into the lipoprotein fractions was not significantly different between the low-and highcholesterol groups (66) . In another study by Rossignol et al., a 72 h intravenous infusion and higher doses (500, 2,000 or 3,500 μg/h) of E5564 were administered into healthy volunteers (67) . E5564 has a slow plasma clearance (0.679 to 0.930 ml h −1 kg −1 of body weight), a small volume of distribution (45.6 to 49.8 ml/kg), and a relatively long half-life (50.4 to 62.7 h). All these pharmacokinetic parameters obtained are comparable to the study done by Wong et al. (74) . The association of E5564 with plasma lipoproteins was also investigated and it was found that the majority (∼55%) of the drug was bound specifically to high-density lipoprotein (HDL), but not low-density lipoproteins, very-low-density lipoproteins, or albumin (67) . A Phase II multi-site, double-blind, randomized, ascendingdose, placebo-controlled safety study on E5564 was conducted in cardiac surgery patients (75) . Patients undergoing coronary artery bypass graft and/or cardiac valvular surgery with cardiopulmonary bypass were enrolled. Patients received a four hour infusion of 2, 12 or 28 mg of E5564 before cardiopulmonary bypass. No significant safety concerns were identified. No significant difference was observed in most variables related to systemic inflammation or organ dysfunction/injury. This phase II safety study suggests that the administration of E5564 is not associated with toxicity in cardiac surgical patients. However, the relatively small sample size used in this study limits the conclusion regarding rare adverse events or the potential clinical benefits of this drug (75) . The potential of E5564 as a sepsis treatment was addressed by Kaneko et al. (76) , Surface Plasmon resonance (SPR) analysis indicated that E5564 binds to LPS binding protein (LBP), in a manner similar to LPS. Blood withdrawn from healthy volunteers was treated with heparin to prevent clotting. At doses of E5564 relevant to its clinical use (i.e. 6 μg/ml), antibodies against LBP did not influence either the distribution of E5564 to non-HDL lipoprotein fractions or the transfer of E5564 from non-HDLs to HDL. LBP binds E5564 in a manner similar to LPS, but does not play a role in E5564 redistribution/binding to lipoprotein and plasma clearance. Czeslick et al. (77) carried out an ex vivo study on the effect of E5564 on production of LPS-induced pro-inflammatory cytokines, particularly IL-6 and TNF-α, in LPS-induced human monocytes. In this study, they recruited 10 healthy volunteers and obtained their whole blood samples and preincubated with 0.001, 0.003, 0.01, 0.03, 0.1, 1 and 10 ng/ml E5564 for 45 min and after stimulated with 0.2 ng/ml of LPS. They found that E5564 (0.003 up to 10 ng/ml) caused a dosedependent inhibitory effect on IL-6 and TNF-α production in LPS-stimulated human monocytes. They concluded that E5564 has a significant LPS inhibitory effect via down regulation of the intracellular generation of pro-inflammatory cytokines IL-6 and TNF-α in human monocytes (77) . The association of E5564 with plasma protein and lipoprotein was studied in plasma obtained from fasted human subjects with various lipid concentrations (64) . It was reported that the majority of E5564 was recovered in the high-density lipoprotein (HDL) fraction. Additionally, they had shown increasing levels of TG-rich lipoprotein (TRL) lipid (TC and TG) concentrations resulted in a significant increase in the percentage of E5564 recovered in the TRL fraction. Furthermore, their findings had suggested that E5564 does not influence CETP-mediated transfer activity (64) . Human peripheral blood mononuclear cells (PBMCs) were isolated from peripheral blood obtained from healthy human volunteers by density gradient centrifugation (68) . TAK-242 was effective in human cells and inhibited the production of TNF-α, IL-6, and IL-1b from PBMCs stimulated with LPS and IFN-gamma, with IC 50 values of TAK-242 ranging from 5.3 to 58 nM. There were four donors used for this study. There was no marked difference in the IC 50 values of TAK-242 amongst them. TAK-242 showed suppressive effects on the production of various inflammatory mediators from human monocytes and macrophages stimulated with LPS. TAK-242 also suppressed the production of these cytokines from LPS-stimulated human peripheral blood mononuclear cells (PBMCs) at IC 50 values from 11 to 33 nM. In addition, the inhibitory effects on the LPS-induced IL-6 and IL-12 production were similar in human PBMCs, monocytes, and macrophages. TAK-242 suppressed the cytokine production induced by Toll-like receptor (TLR) 4 ligands, but not by ligands for TLR2, TLR3, and TLR9. TAK-242 suppresses the production of multiple cytokines by selectively inhibiting TLR4 intracellular signaling (68) . The manipulation or intervention of TLR-mediated immune responses is a potential approach to treat and prevent the septic shock and variety of associated diseases. However, blocking TLR may lead to 'inappropriate' immune responses such allergic Th2 responses, or immunological tolerance (78) . Thus, it seems clear that the risks and benefits of manipulation of TLR mediated immune responses need to be balanced and require further investigation.

Multiplex cytokine profile from dengue patients: MIP-1beta and IFN-gamma as predictive factors for severity

BACKGROUND: Dengue virus pathogenesis is not yet fully understood and the identification of patients at high risk for developing severe disease forms is still a great challenge in dengue patient care. During the present study, we evaluated prospectively the potential of cytokines present in plasma from patients with dengue in stratifying disease severity. METHODS: Seventeen-cytokine multiplex fluorescent microbead immunoassay was used for the simultaneous detection in 59 dengue patients. GLM models using bimodal or Gaussian family were determined in order to associate cytokines with clinical manifestations and laboratory diagnosis. RESULTS: IL-1β, IFN-γ, IL-4, IL-6, IL-13, IL-7 and GM-CSF were significantly increased in patients with severe clinical manifestations (severe dengue) when compared to mild disease forms (mild dengue). In contrast, increased MIP-1β levels were observed in patients with mild dengue. MIP-1β was also associated with CD56+NK cell circulating rates. IL-1β, IL-8, TNF-α and MCP-1 were associated with marked thrombocytopenia. Increased MCP-1 and GM-CSF levels correlated with hypotension. Moreover, MIP-1β and IFN-γ were independently associated with both dengue severity and disease outcome. CONCLUSION: Our data demonstrated that the use of a multiple cytokine assay platform was suitable for identifying distinct cytokine profiles associated with the dengue clinical manifestations and severity. MIP-β is indicated for the first time as a good prognostic marker in contrast to IFN-γ that was associated with disease severity.

During the last decades dengue became the most important arthropod-borne emerging viral disease in tropical countries [1] . It is estimated that about 2.5% notified cases are classified as dengue haemorrhagic fever (DHF) and about 2.5-20% of DHF cases are lethal [2] [3] [4] . In the last two decades, Latin America saw a dramatic increase in frequency and in geographic extension of dengue fever. Specifically, the situation in Brazil has worsened during the last decade since the introduction of the Dengue-3 serotype. In the past years Brazil had dengue outbreaks with at least 1 million cases (2001) (2002) and within the last 18 months 900 thousand cases were reported [5] . In addition, severe disease forms are occurring with increased frequency and mortality rates. Dengue pathogenesis is not completely understood, and the main determinants of the development of severe forms are not yet well established. Increase in capillary permeability associated with endothelial activation and haemorrhagic phenomena are landmarks of severe clinical manifestations, strongly suggesting an alteration in immunoregulation [6] . Cytokines are proteins secreted during innate and adaptive immunological responses, acting as inflammatory mediators or modulatory molecules during several haemorrhagic fevers [7] . Clinical studies support a key role for cytokines in the DHF pathogenesis [8] [9] [10] [11] [12] [13] . During Dengue virus infections, cytokines are involved in the disease onset and homeostatic regulation. Specifically, TNF-α, IL-1β and IL-6 have been associated with both coagulation (F1+2 and TATc) and fibrinolysis (t-PA, PAPc, and D-dimmer) activation markers [14] . This activation is more striking in patients with severe clinical manifestations, although it can be found at lower degrees in patients with mild disease [15, 16] . Despite the fact that cytokine network and their multiple regulatory pathways are highly complex and not fully elucidated during dengue fever, these molecules seem to represent interesting markers for patient stratification or prognosis. An emerging interest has appeared in order to define biomarkers that may have pathophysiological roles during disease and that may be used as future therapeutic targets. New technologies have been developed in order to detect multiple biomarkers within a single and small blood sample. Such approaches may lead to the development of specific marker panels for dengue fever. Accordingly, cytokine patterns have been indicated as serum biomarkers during infectious diseases such as Hepatitis C [17] , ARDS [18] and sepsis [19] . In this study, we prospectively evaluated the potential use of plasma cytokine concentrations for severity stratifica-tion of patients with dengue, using a 17 cytokine-multiplex assay. Among tested cytokines, we were able to recognize ten significantly altered circulating factors and to characterise cytokine patterns related to determined clinical manifestations and disease severity. The Ethics Committee of the Oswaldo Cruz Foundation approved this study protocol and written informed consent was obtained from all patients or their guardians prior to blood collection. We included prospectively 59 dengue-infected patients ( A detailed history and physical examination was performed to detect hemorrhagic manifestations (positive tourniquet test for capillary fragility, skin haemorrhages, epistaxis, gingival, gastrointestinal, or urinary tract haemorrhage), signs of plasma leakage (pleural or pericardial effusion, ascites), signs of circulatory failure (cold extremities, cyanosis, hypotension, tachycardia, shock), and hepatomegaly. In addition to the suggestive clinical diagnosis, all patients had the Dengue virus infection confirmed either by antidengue enzyme-linked immunoabsorbent assay (ELISA)-IgM, serotype specific reverse transcription-polymerase chain reaction (RT-PCR) or by virus isolation [20] [21] [22] . Dengue immune response was considered as primary or secondary by IgG ELISA according to previously established criteria [23] . As previously reported [24] [25] [26] , we also were often unable to characterize the severe disease forms based on WHO criteria [3]. In Nicaragua, Harris et al. [24, 27] described four key severe clinical manifestations associated with dengue -shock, plasma leakage, marked thrombocytopenia or internal haemorrhage -that do not fit DHF/DSS classification as single parameters. According to these criteria, we considered: • Severe dengue -Dengue confirmed cases plus severe thrombocytopenia (<50,000 platelets/mm 3 ) and/or hypotension (postural hypotension with decrease in systolic arterial pressure in 20 mm Hg in supine position or systolic arterial pressure < 90 mm Hg) and/or plasma leakage (either haemoconcentration fluctuation of packed cell volume ≥ 20% during illness course and recovery or clinical signs of plasma leakage, such as pleural effusion) and/or severe haemorrhagic manifestations. • Mild dengue -Dengue confirmed cases in absence of severe thrombocytopenia, hypotension, plasma leakage signs or haemorrhagic manifestations. Blood samples were collected from a peripheral vein and kept on ice. Plasma was collected by centrifugation at 800 g for 15 min at 4°C, aliquoted, and stored at -70°C until the analysis day. A multiplex biometric immunoassay, containing fluorescent dyed microspheres conjugated with a monoclonal antibody specific for a target protein, was used for cytokine measurement according to the manufacturer's instructions (Bio-Plex Human Cytokine Assay; Bio-Rad Inc., Hercules, CA, USA). Cytokines measured were: IL-1β, IL-2, IL-4, IL-5, IL-6, IL-7, CXCL8 (IL-8), IL-10, IL-12 (p70), IL-13, IL-17, granulocyte colony stimulating factor (G-CSF), granulocyte-monocyte colony stimulating factor (GM-CSF), monocyte chemoattractive protein (MCP-1/CCL2), macrophage inflammatory protein (MIP-1β/CCL4), and TNF-α. Briefly, 20 μl plasma samples were diluted 1:4 and incubated with antibodycoupled beads. Complexes were washed, then incubated with biotinylated detection antibody and, finally, with streptavidin-phycoerythrin prior to assessing cytokine concentration titres. Concentrated human recombinant cytokine was provided by the vendor (Bio-Rad Laboratories). A range of 1.95-32,000 pg/ml recombinant cytokines was used to establish standard curves and to maximize the sensitivity and the assay dynamic range. Cytokine levels were determined using a multiplex array reader from Luminex™ Instrumentation System (Bio-Plex Workstation from Bio-Rad Laboratories). The analyte concentration was calculated using software provided by the manufacturer (Bio-Plex Manager Software). Liquid nitrogen cryopreserved peripheral blood mononuclear leukocytes were isolated by Histopaque-1077 (Sigma Chemical Co., Saint Louis, MO, USA) from 35 out of 59 dengue patients. Cells were stained for CD56 surface marker using anti-CD56-Cy5 (IgG1, clone B159) from Pharmingen (San Diego, CA, USA) and positive cells were detected by flow cytometry as described before [22] using FACScalibur (Becton-Dickinson). Events (10,000-20,000) were acquired and analyses were carried out with FlowJo (TreeStar, version 4.3) software. The nonparametric Mann-Whitney U test was used to evaluate differences between cytokine ratios from severe and mild dengue patients. GLM models were used to evaluate factors independently associated with quantitative variables. Analysis of factors independently associated with dengue severity and other clinical manifestations was performed with GLM with logistic regression or Gaussian family. Results from the logistic regressions are given as odds ratio (OR). The confidence interval (CI) was established at 95%. Alternatively, for a GLM Gaussian family t values were recorded. A probability value of P<0.05 was considered to be significant. The statistical programs R [28] and Prism 4 (Graph-Pad Software, San Diego, CA, USA) were employed. The Fisher's exact test was applied to determine the significance of positive samples from patients when comparing different virus serotypes or sequential infections. Correlation between platelet counts and cytokine production in blood samples was estimated by Spearman's correlation. From the 59 patients included, 39 were classified as severe dengue and 20 as mild dengue. Detailed demographic, clinical, and laboratorial data from dengue patients are summarized in Table 1 . Blood collection was performed between 3 and 10 days after disease onset. In order to avoid effects due to differences in the blood collection time, we compared mild and severe dengue patient groups using Mann-Whitney U Test, which showed no differences in the disease onset time at the moment of sample collection [see Additional file 1]. The original data used to perform this analysis is shown at Figure 1 . Patients with mild and severe dengue were investigated for prior incidence of infection, detected by serologic immune response (IgG antibodies for DENV). Patients with severe dengue (42%; 14 out 33) were more likely to be experiencing a secondary Dengue virus infection than patients with mild dengue (28%; 6 out 20), although no statistical significance was found in Fisher's exact test (P = 0.3989). Among 22 patients with DENV-1, 5 were classified as secondary infection, whereas among 35 patients with DENV-3, 12 were classified as secondary infection (P = 0.3912). Dengue fever is characterised by a high fever phase and an abrupt drop in body temperature that has been called defervescence phase. Characteristically the disease outcome is defined during this phase, when patients can either recover rapidly or progress to a severe life-threatening stage. Cytokines and immunoactivation markers such as IFN-γ, IL-2, soluble CD8 and receptors for TNF-α [12, 29] are associated with the defervescence phase and with disease severity. IFN-α levels are higher in DHF than in DF during defervescence [30] . During the febrile phase significant increase in cytokine circulating levels was detected including IL-4, IL-6, IL-10, MCP-1 and MIP-1β levels, which were maintained also elevated in defervescence (data not shown, analysed by non parametric Kruskal-Wallis test and Dunn's Multiple Comparison Test when compared with controls, P < 0.05); no significantly altered febrile levels were found when compared to defervescence. During the febrile Cytokine levels in plasma from patients with mild and severe dengue Figure 1 Cytokine levels in plasma from patients with mild and severe dengue. Box-and-whiskers graph. The box extends from the 25 th to the 75 th percentile and the line at the middle is the median. The error bars, or whiskers extend down to the lowest value and up to the highest. Factors were sorted according to their functional groups. Mann-Whitney U test was used to evaluate differences between cytokine concentration from severe and mild dengue patients. * P < 0.05, ** P < 0.01 and ** P < 0.001. phase IL-7 was significantly higher than in defervescence. IL-1β, IL-13, IFN-γ were significantly increased during defervescence as compared to control samples. Significant levels of IL-5, IL-12, and IL-17 were not detected during dengue disease in our patients. IL-2 was detected both in healthy individuals and in dengue patients but no difference between these two groups was detected [see Additional file 2]. We studied the cytokine profile from Brazilian patients in order to compare severe and mild dengue cases during the acute phase of the disease. Figure 1 shows data from patients with regard to their plasma cytokine contents, which were sorted in four groups according to their reported function. We observed that cytokine concentrations of IL-1β, IFN-γ, IL-4, IL-6, IL-7, IL-13 and GM-CSF were significantly increased during severe dengue as compared to mild dengue, while MIP-1β levels are higher in mild dengue. MIP-1β and IFN-γ were independent variables associated disease outcome as determined by a logistic regression model (Table 2 and Figure 2 ). While MIP-1β was increased during mild dengue with odds ratio (OR) of 0.181 and confidence interval (CI) 0.045-0.72, IFN-γ was associated with severity with OR of 1.138 (CI, 1.0541-1.245). To assess relationships between cytokine levels and several clinical manifestations, the patient cohort with severe dengue was divided into distinct subgroups: those with hypotension, thrombocytopenia (≤50.000 counts/mm 3 ) and/or haemorrhagic manifestations. A logistic regression model was used for binomial response subgroups and GLM models using Gaussian family were employed for subgroups with continuous response in order to determine cytokine profiles. IL-1β was associated with marked thrombocytopenia with OR = 1.058 (CI, 1.012-1.106) in dengue patients. TNF-α was inversely related to thrombocytopenia with OR = 0.978 (CI, 0.962-0.995) (Table 2, Figure 2 ). Considering platelet counts as a continuous variable for statistical analysis with a Gaussian family, it was possible to determine that IL-8 (P = 0.0434) and MCP-1 (P = 0.0146) levels are inversely related to their counts, displaying therefore an association with thrombocytopenia, while MIP-1β (P = 0.0114) confirms its association with higher counts -normal or tending to normal (Table 3) . GM-CSF (OR = 1.004; CI, 1.001-1.007) was related with hypotension, whereas IL-1β had a negative predictive Table 2 and Figure 2 . Natural Killer (NK) cells have been earlier related to mild cases of dengue [22] . Forty-eight PBMC samples from thirty-five patients had their CD56+ rates determined by flow cytometry and a good correlation was observed with Cytokines detected in plasma as independent factors in predicting severity Table 2 . their respective MIP-1β plasmatic levels (r = 0.4668; P = 0.0008). Considering that different cytokines act in the immunological network as stimulating/up regulating factors and also in a feedback loop as down regulators, the cytokine balance might play a role in the immune response outcome. Therefore we calculated MIP-1β/IFN-γ ratios for every patient and compared those from mild dengue with those from severe dengue. Ratio averages ± SEM were respectively 3032 ± 514 and 864 ± 240 (P = 0.0003; Mann Whitney U test), confirming our earlier data that these cytokines are acting as opposing factors. The different models built here using clinical manifestations as independent variables each exhibit specific cytokine profiles. The cytokine profile identified in patients with dengue may represent a valuable tool for the characterisation of immunological response patterns and may assist the identification of patient groups at risk for developing severe disease. In the present study, the use of a multiplex analysis for cytokine plasma detection in patients with dengue could identify cytokine profiles associated with the disease severity. Early identification and management of severe dengue disease are essential to prevent death. It has been increasingly recognized that the inflammatory response and deregulated cytokine production play key roles in the development of severe clinical manifestations [31] . However, cytokine profiles associated with dengue evolution and prognosis are not well established. New technologies for cytokine quantification were developed including the multiplex immunofluorescent bead array analysis system, allowing multiple biomarkers to be tested simultaneously in a small volume from one single plasma aliquot. Recently, this methodology has been used for cytokine profile evaluation during several infectious diseases including viral infections [17, 32, 33] and sepsis [19] , among others. We were able to identify models for cytokine circulation during dengue acute phase that may vary with clinical manifestations. MIP-1β was for the first time associated with a good prognostic and was identified in the different disease models presented here. MIP-1β has been earlier detected after Dengue virus cell stimuli in vitro [34, 35] but preliminary studies in vivo did not report their role in severity. In accordance with a protective role for MIP-1β, changes in MIP-1β levels were significantly correlated with decreases in viral titre after Hepatitis C treatment [17] . In addition, MIP-1β was up regulated in acute infection in chimpanzees only when viral clearance took place, but not in those animals which failed to eradicate the virus [36] . MIP-1β is produced by human monocytes and dendritic cells upon different stimuli [37] as well as by activated NK cells [38] and lymphocytes [39, 40] . Activated NK cells release granzyme A, which displays cytolytic functions and MIP-1β is chemoattractant for NK cells, recruiting them to inflammatory sites. NK cells have been associated with mild dengue [22] . Here a good correlation between MIP-1β plasma levels and NK cells was observed, reinforcing the relevance of these pathways and strongly suggest- ing their role in dengue protective mechanisms. An early and efficient virus clearance by direct or indirect NK functions is likely controlling virus replication, restricting intense immunological activation and the dengue immunopathology and therefore favouring a mild dengue disease. In previous studies, TNF-α has been reported to be associated with severity, mainly during DHF in Brazilian patients [11, 13, 41] . In the present study, however, this cytokine was not strongly associated with severity. Indeed, other authors also found inconsistency or no difference in TNF-α levels in severe vs. mild disease forms [10, 42] . We may hypothesize that differences in Dengue virus serotypes or in host immune response such as different TNF-α genetic polymorphisms may explain the disease outcome. In our study from 2001 (Braga et al., 2001) , patients were Dengue-2 infected, while in the present study, patients were Dengue-1 and -3 infected. A recent report [43] describes non-significant TNF-α serum levels in adult patients and suggests that the discrepancy may have been caused by a transient TNF-α peak which was not detected at a later time point. In the present study we observed an association of IFN-γ with disease severity. Indeed, increased IFN-γ concentrations have been detected in dengue patients in a variety of studies [29, [44] [45] [46] [47] . DHF induced by Dengue-3 was associated with higher viremia early in illness and earlier peak plasma IFN-γ levels; maximum plasma viremia levels correlated with the degree of plasma leakage and thrombocytopenia [45] . However, in a previous study from our group we failed to observe association of IFN-γ with disease severity [44] , probably due to the small number of severe patients analyzed or to the Dengue-1 incidence. IFN-γ is produced during a T-lymphocyte helper response type 1 and may reflect CD8+ T cell activation with production of inflammatory cytokines. High levels of IFN-γ were observed in patients with dengue from Asian and Latin America and were associated with severity [9] . IFN-γ produced by T cells may also activate mononuclear phagocytes (monocytes and dendritic cells), which would produce factors such as TNF-α, tissue factor, and plateletactivating factor, among other mediators. These factors may all participate in platelet and endothelial cell activation, leading to platelet consumption, increase in endothelial permeability, hypotension and ultimately to shock. IFN-γ has also been associated with secondary heterologous Dengue virus infections [47, 48] inducing a strong antigenically cross-reactive inflammatory response, probably inefficient in terms of antibody and T-cell specific response. Indeed, we observed earlier in several patients a CD8+T cell activation with HLA-DR+ subset increase that was associated with severity [9] . GM-CSF acts at early differentiation processes at myeloid progenitors or resting monocytes [49] . An additional stimulus may be required to activate monocytes or dendritic cells in order to produce proinflammatory cytokines [50] . GM-CSF was associated with hypotension as well as MCP-1, likely acting both in concert, contrasting with MIP-1β, once more associated with good prognostics. MCP-1 was earlier detected in DHF patients [51] but for the first time we reported clinical and laboratory findings associated with severity. IL-8 and MCP-1, here associated with thrombocytopenia, are chemokines and may contribute to platelet activation, either by their chemoattractant properties or by their effect on endothelial permeability. Both factors were detected in patients with DHF [51, 52] . These cytokines are produced by monocytes after various activation stimuli, such as virus infection, and increase the endothelial permeability by disrupting tight junctions among cells [53] . Despite the fact that our study could identify cytokines with good accuracy for the stratification and/or prognosis of dengue, it has potential limitations. Here we identified cytokines related to dengue severity, but the small sample size represents a shortcoming regarding the generalization of our results. In addition, only one time point was used for the measurement of cytokines, not allowing further insights provided by sequential measurements. Moreover, classification of disease severity has been a matter of debate, especially for adult patients' management and classification. Indeed, the WHO criteria for DHF has failed to identify severe disease, including fatal cases, in adult Latin America population [24, 54] (S.M.O. Zagne, R.M.O. Nogueira, unpublished) and clearly do not satisfy the stratification of our studied population. Accordingly, in the present study severe disease forms were classified following other proposed criteria [24] . While a direct correlation of cytokine concentrations and the pathophysiology of severe dengue is tempting, we believe that the full burden of disease severity cannot be attributed to a single cytokine. Cytokines may be increased simply as one of the several steps in the network loops without necessarily playing a direct harmful role and most likely more than one factor may be involved, including others not tested here such as IL-18, TGF-β, and MIF among others [9] . We can suggest a mechanism explaining our cytokine models for dengue fever (Figure 3) . MIP-1β would be associated with a protective pathway for its chemoattractive and activating effect on NK cells, which in turn are efficient cells in early virus clearance, by their antiviral cytokine production and cytotoxic activity against infected target cells. IFN-γ has a deleterious effect for the host regarding its action in activating T cells for virus anti-genic cross-reactive response and monocyte/dendritic cell activation. Mononuclear monocytes are activated by IFNγ and GM-CSF among other cytokines and in turn produce several factors such as IL-1β and MCP-1 that may act on vascular permeability leading to plasma leakage and haemoconcentration. As suggested by other authors, it is likely that viral replication in antigen presenting cells, cytokine liberation and circulation, and T cell activation may not be a linear process [55] , but in fact a complex interaction network, with positive and negative feedbacks, where viral clearance and pathologic events take place, such as increased vascular permeability and circulatory collapse, and their balance may determine the disease outcome. Our study demonstrated the plasma cytokine profile in dengue fever from a Brazilian population detected by a multiplex bead immunoassay. MIP-β is indicated for the first time as a good prognostic marker and in contrast to IFN-γ that was associated with the disease severity. Both cytokines can discriminate mild from severe cases. Moreover, we show here for the first time that during the dengue course different cytokine profiles may be present and vary according to determined clinical manifestations. The cytokine profiles identified herein by bead array multiplex system may favour an early identification of patients with the worst prognosis and may contribute to the establishment of more directed therapeutic procedures than the present ones. Hypothetic mechanism to explain cytokine models during dengue fever Figure 3 Hypothetic mechanism to explain cytokine models during dengue fever. MIP-1β is associated with a good prognostic and IFN-γ has a predictive value for severity. GM-CSF, MCP-1, IL-1β, IL-6, IL-8, IL-12, IL-13 are also playing important roles during dengue pathogenesis (see text for detailed description).

Temporal trends in the discovery of human viruses

On average, more than two new species of human virus are reported every year. We constructed the cumulative species discovery curve for human viruses going back to 1901. We fitted a statistical model to these data; the shape of the curve strongly suggests that the process of virus discovery is far from complete. We generated a 95% credible interval for the pool of as yet undiscovered virus species of 38–562. We extrapolated the curve and generated an estimate of 10–40 new species to be discovered by 2020. Although we cannot predict the level of health threat that these new viruses will present, we conclude that novel virus species must be anticipated in public health planning. More systematic virus discovery programmes, covering both humans and potential animal reservoirs of human viruses, should be considered.

Despite long-standing interest in global biodiversity (May 1988) , only recently has the diversity of human pathogens been catalogued . Approximately 1400 pathogen species are currently recognized ( Woolhouse & Gaunt 2007) . Fewer than 200 of these are viruses, but novel virus species are being reported in humans at a rate of over two per year, much faster than for other kinds of pathogen ( Woolhouse & Gaunt 2007) . Novel viruses are a major public health concern, whether causing disease on the massive scale of HIV/AIDS, more transient events such as the SARS epidemic or potential future threats such as pandemic influenza. An analysis of temporal patterns of virus discovery is therefore of considerable interest. Our analysis is based on the rate of accumulation of new human virus species: the 'discovery curve'. Discovery curves have previously been used to estimate the total diversity of various plant and animal taxa (Dove & Cribb 2006; Bebber et al. 2007) . However, to our knowledge, the discovery curves have not previously been compiled for any category of human pathogen. Having compiled the discovery curve, we proceed to develop a simple statistical model which we use to estimate the size of the pool of human virus species, N, and the expected rate of discovery of new species to 2020. A standard method for estimating numbers of species is to extrapolate the cumulative species discovery curve (Bebber et al. 2007) . We gathered data for this curve by systematically searching the primary literature for first reports of human infection with each of the currently recognized virus species, using species as defined by the International Committee on Taxonomy of Viruses (ICTV; http://www.ictvonline.org/). We note that the set of viruses we are interested in-those that can infect humans-is a small subset of the total (over 1500 species according to ICTV ) and, as is discussed below, not a closed set because many of these viruses can also infect other hosts . We regard this as analogous to constructing species discovery curves for any subdivision of geographical range or habitat. As we demonstrate below, this approach yields an excellent description of the discovery curve. We used piecewise linear regression to test for changes in the slope of the discovery curve. The results suggested upswings in 1930 (95% CI, 1929 -1933 ) and 1954 (1953 -1956 . We therefore restricted detailed analysis to the period 1954-2006. We modelled discovery since 1954 assuming a total number of species available to be discovered (the species pool) of N virus species, each discovered in any given year with probability p. The model was fitted to the data and assessed using Markov chain Monte Carlo (MCMC)-based Bayesian inference, generating distributions and credible intervals for the parameters. The model defines the expected number of discovered viruses in year t as l t ðN; pÞ Z Npð1K pÞ tK1 ; ð2:1Þ where year tZ1 corresponds to 1954. The binomial distribution B(N, p) can be accurately approximated by a Poisson distribution with parameter Np for the range of values of N and p of interest. We considered fitting a distribution for values of p; however, provided individual p-values are low there is minimal improvement in model fit. Thus, for a set of model parameters, the likelihood of observing data, XZ{x i }, the number of viruses discovered for years 1 to k, is given by Parameter distributions for N and p were calculated using MCMC simulation using a standard Metropolis algorithm with flat prior information. It was necessary to compute a correlation matrix to define a joint proposal since N and p are closely correlated. We monitored convergence using two chains. Once they had converged, we had a burn in period of 10 5 samples. We compared the model with the observed data by calculating the mean, trend in the mean and variance for the number of virus species discovered per year (based on five million simulations using best-fit parameter values). The model was extrapolated to year 2020 by calculating the expected number of viruses discovered using the best-fit model. The 95% posterior prediction intervals were calculated using two million model simulations taking into account parameter uncertainty (as given by data from 1954 to 2006) and natural model simulation stochasticity. As a validation exercise, the model was also fitted to the curve for accumulated virus families from 1954 using the same methods, except that the Poisson approximation no longer holds, so a binomial distribution was used. A family (based on current ICTV classifications) was added to the total when the first post-1954 species was allocated to that family. We tested the assumption that species can be randomly assigned to families (weighted by the size of the families) by noting the number of years in which 0, 1, 2, etc. virus families were discovered. This was done one million times to obtain a distribution for comparison with the observed values. From a comprehensive search of the primary literature, we found 188 virus species that have been reported to infect humans, going back to yellow fever virus in 1901 (table 1) . Since then, the number of human virus species discovered in any given year has ranged from zero to six. As is typical (Bebber et al. 2007) , the cumulative species discovery curve increases slowly initially and then more rapidly (figure 1). Piecewise linear regression suggests no further upswings since 1954, roughly corresponding to the advent of tissue culture techniques for virus detection (figure 1). We confirmed that our model reproduced the observed slight downward trend in the rate of discovery since 1954 (figure 1) and the observed variance in the data from 1954 to 2006 (figure 2). The distribution of the number of virus species discovered per year shows slight overdispersion (meanZ2.69; varianceZ3.07; varianceto-mean ratio greater than 1) which falls within the predicted range (meanZ2.70 with 95% credible interval 2.41-3.00; varianceZ3.03 with interval 1.99-4.49). Together, these results support our choice of model, even though we do not explicitly consider heterogeneity in the probability of discovering a given species in any one year ( p) or temporal variation in sampling effort, detection techniques and reporting. Noting that p and N are highly correlated (figure 3), our best estimate for p is 0.015 (95% credible interval, 0.004-0.026) with 117 (38-562) so far undiscovered virus species. Extrapolating the discovery curve, allowing for parameter uncertainty and stochastic discovery, we obtain a best estimate of 22 new species (10-40) by 2020 ( figure 1) . Data on the cumulative discovery of new virus families are also reproducible (figure 4). The predicted distribution of the number of virus families discovered per year (assuming random allocation of species to families) compares favourably with the observed distribution ( figure 5 ). This provides further support for the appropriateness of our model. Virus discovery curve M. E. J. Woolhouse et al. 2113 We conclude that it is extremely probable that new human viruses will continue to be discovered in the immediate future; we are not yet close to the end of the virus discovery curve. As a direct result of this, it is not possible to estimate the size of the species pool for human viruses with precision. However, in contrast to the negative assessment by Bebber et al. (2007) of the use of incomplete species accumulation curves, we consider that the upper and lower limits to our estimate of the size of the species pool are of interest and also have practical implications. Current trends are consistent with a pool of at least 38 undiscovered species that will be reported at an average rate of at least approximately one per year to 2020. In this context, it is worth noting that three new species were reported in 2007: two polyoma viruses, Ki and Wu, and a reovirus, Melaka (Allander et al. 2007; Chua et al. 2007; Gaynor et al. 2007) . Other viruses may have been reported but not yet classified. In practice, future rates of discovery will, of course, be affected by any major advances in virus detection technology or by any major shifts (upwards or downwards) in the effort expended on virus discovery programmes. Tissue culture was regarded as the 'gold standard' for virus detection up until a few years ago when molecular methods came to the fore (Storch 2007) , although there has not been a detectable increase in discovery rates as a result. Indeed, it is striking that there have been no dramatic changes in the pattern of virus discovery for over 50 years; extrapolations from our data should therefore provide a useful benchmark for probable future discovery rates. The upper limit for N is finite but large; we cannot rule out hundreds of novel human viruses to be reported in the future. There are two (not mutually exclusive) possible explanations for such a high level of diversity. First, it could reflect the largely unknown extant diversity of viruses in the non-human animal reservoirs that constitute the major source of emerging human pathogens Woolhouse & Gaunt 2007) . The majority of human viruses are known to be capable of infecting nonhuman hosts (almost exclusively mammals and birds), and the animal origin of many apparently novel human viruses (e.g. HIV1 and HIV2, SARS CoV, Nipah virus) has been frequently remarked upon (Morse 1995; Woolhouse & Gowtage-Sequeria 2005; Wolfe et al. 2007 ); indeed, recently discovered viruses are even more likely to be associated with a non-human reservoir ( Woolhouse & Gaunt 2007) . All these observations are consistent with the idea that a significant fraction of viruses discovered in the last few decades is ecological 'spillover' from animal populations rather than newly evolved specialist human viruses. We have very limited knowledge of the diversity of viruses present in most mammal and bird species (with most attention having been paid to viruses of domestic animals; Cleaveland et al. 2001) , so it is unclear for how long this process might continue. An alternative explanation for a large pool of human viruses is that this reflects a high rate of evolution (within a reservoir population) of truly novel species capable of infecting humans. This hypothesis is difficult to test directly without much more comprehensive sequence data from both human and non-human virus populations. We note that the finite upper limit for the current estimate of N does not necessarily imply that the process of virus discovery is not open-ended (as a result of the evolution of new species) since there could be a low background rate of virus evolution, which will remain once extant diversity has been fully revealed. The balance between revealing extant diversity and the continual evolution of new species could be explored using a more complex model than equation (2.1); however, the available data are insufficient to yield useful estimates of the additional parameters required. Although we cannot know in advance how big a threat they will pose, novel human viruses must be anticipated in public health planning and surveillance programmes for emerging infectious diseases (King et al. 2006; Jones et al. 2008 ). However, current approaches to virus discovery are largely passive, usually relying on investigation of reports of human disease with unfamiliar clinical symptoms and uncertain aetiology. Recently, there have been calls for more active discovery programmes for viruses and other pathogens involving 'systematic sampling and phylogeographic analysis of related pathogens in diverse animal species' ( Wolfe et al. 2007) . We consider that such calls are supported by the results reported here. Virus discovery curve M. E. J. Woolhouse et al. 2115

Investigating selection on viruses: a statistical alignment approach

BACKGROUND: Two problems complicate the study of selection in viral genomes: Firstly, the presence of genes in overlapping reading frames implies that selection in one reading frame can bias our estimates of neutral mutation rates in another reading frame. Secondly, the high mutation rates we are likely to encounter complicate the inference of a reliable alignment of genomes. To address these issues, we develop a model that explicitly models selection in overlapping reading frames. We then integrate this model into a statistical alignment framework, enabling us to estimate selection while explicitly dealing with the uncertainty of individual alignments. We show that in this way we obtain un-biased selection parameters for different genomic regions of interest, and can improve in accuracy compared to using a fixed alignment. RESULTS: We run a series of simulation studies to gauge how well we do in selection estimation, especially in comparison to the use of a fixed alignment. We show that the standard practice of using a ClustalW alignment can lead to considerable biases and that estimation accuracy increases substantially when explicitly integrating over the uncertainty in inferred alignments. We even manage to compete favourably for general evolutionary distances with an alignment produced by GenAl. We subsequently run our method on HIV2 and Hepatitis B sequences. CONCLUSION: We propose that marginalizing over all alignments, as opposed to using a fixed one, should be considered in any parametric inference from divergent sequence data for which the alignments are not known with certainty. Moreover, we discover in HIV2 that double coding regions appear to be under less stringent selection than single coding ones. Additionally, there appears to be evidence for differential selection, where one overlapping reading frame is under positive and the other under negative selection.

In the past few years we have witnessed an explosion in the viral genomic data available. GenBank alone holds over 80,000 close to complete viral genomes, and numbers are rising fast. For example, since the submission of the first SARS genome in May 2003, over 140 more have been published. With this genomic data at hand we hope to finally be able to tackle our understanding of viruses. Mechanisms of selection, that is to say the rate f at which a mutation resulting in a change in amino acid is accepted, and evolution on viruses are still strongly debated, and a methodology which is trimmed towards answering these questions is required. A step towards this is our attempt to develop a method which can deal with the vast amount of viral data, as well as the complexity of viral genomes and their high divergence and subsequent unreliability of alignment. Several papers [1, 4, 5, 7, [13] [14] [15] [16] 18] have been dedicated towards the study of selection on viral genomes, in particular focusing attention on the evolutionary behaviour of overlapping reading frames. These are a feature common to viruses, where due to the three periodicity of the genetic code, up to three genes may be encoded simultaneously. The constraints placed on a nucleotide involved in such a multiple coding region will naturally have an effect on its mutational pattern, and as a result the concept of selection is complicated further. Another complication is the uncertainty of alignments when dealing with genomes of reasonable evolutionary distance. Recent papers have shown that parameter estimation can be greatly biased by the use of a fixed alignment [10] . It is often thought that overlapping regions tend to be more constrained in their evolution than single coding ones, since a mutation may cause a non-synonymous substitution in up to three genes simultaneously. Some methods rely on these assumptions for the de novo detection of overlapping genes [19, 22] . Various researchers have attempted to measure selection acting on overlapping reading frames, by investigating the K a /K s ratio within these regions for seperate reading frames [4, 7, [14] [15] [16] . Comparing non-synonymous to synonymous substitution rates only makes sense when the synonymous substitutions are unconstrained. In the case of coding for multiple genes, however, a synonymous substitution in one gene may well be non-synonymous in the other and thus constrained. This biases the analysis towards an under-estimation of the 'true' synonymous substitution rate and thus can lead to the false inference of positive selection. An attempt to resolve this problem has been made, for example by focusing on synonymous substitutions in one reading frame which indeed are unconstrained in the other [18] . Hein & Støvlbaek [5] developed an evolutionary model particular to multiple coding regions, and used this for a study of selection on these. de Groot et al. [1] used this model of varying selection to comparatively annotate two viral genomes with evolved gene structure. McCauley et al. [13] incorporated a slightly extended version into their multiple sequence annotation method, which additionally provides a selection annotation of the genome. However, their method looks at selection on an individual nucleotide level, and does not make assumptions about the modelling of selection on specific regions. Our method presented here looks at selection on genomic segments as opposed to nucleotides, and thus in overlapping coding regions can discern selection for different reading frames. We may therefore attempt to draw conclusions about the nature of not only selection but also the interaction of selection on two different genes. Also, to study the imprint of evolution on viral genomes, it is necessary for the samples to have a reasonably high level of divergence. A benchmark herefore in our experience would be an evolutionary distance a + 2b of at least 0.4. Since more divergent genomes are harder to align, this brings uncertainty about the alignment into the inference. We decide to circumvent this problem by considering the set of all possible alignments -and their corresponding likelihood under our model -, as opposed to a fixed 'optimal' alignment. This method has previously been used for similar purposes, to minimize variability in parameter estimation due to uncertain alignments [10, 12] . We work with a simple indel model, together with our evolutionary model, to generate a pairwise statistical alignment. For two sequences x and y, a set of seed parameters then gives us the probability p ij of each i th position x i being aligned with each j th position y j . We subsequently work with expected observations as opposed to actual ones. We iteratively calculate the alignment probabilities and the maximum likelihood estimates of evolutionary parameters, until we reach a given level of convergence. We also extend our methodology to a multiple pairwise method. The work presented in this paper thus improves on both the above methods [1,13] by our ability to pry apart selection for two genes on overlapping segments and us not having to rely on a fixed alignment anymore. We run a simulation study to gauge the improvement made by considering all possible alignments as opposed to a single fixed one. Even though viruses containing a large number of multiple coding sites might be expected to be easy to align, our simulation results suggest that this is not necessarily the case. The improvement in parameter estimation made by getting rid of uncertainty in alignments appears to be non-negligible, even for viruses with overlapping reading frames. We run our method on a set of 5 HIV2 sequences, as well as a set of 3 Hepatitis B genomes. These are good candidates for analysis of overlapping reading frames, with 11% of the HIV2 genome being double coding and an average overlapping segment being of length 171 nucleotides. Hepatitis B is even more compact with 49% of the genome being double coding and an average overlapping segment length of 532 nucleotides. We subsequently investigate various questions relating to overlapping read-ing frames and the selectional mechanism underlying these. We test our method on simulated data, to see whether summing over all alignments does actually improve results notably. All the results in this section, unless stated otherwise, are obtained using the 'worst-case-scenario' of only two sequences. By taking a 600 nucleotide sequence chunk out of a double coding region of the Hepatitis B NC00397 sequence, we construct a long double coding region, flanked by 300 nucleotides on either side of background sequence. We let this evolve according to the TKF91 model [21] into a descendent sequence, where the Match-Match state emits a descendant according to the Hein & Støvlbaek [5] model with specified evolutionary parameters. We use a gap opening probability of 0.02 and a gap extension probability of 0.4 -these being values similar to the ones encountered in the real sequences we wish to analyse. We also only allow gaps of length 3 within coding regions, so as not to cause a frame shift in coding. We fix all selection parameters to 0.5 and test a variety of evolutionary distances, with transition rate a ranging from 0.2 to 0.7 and transversion rate b = a/2. We annotate using our statistical alignment method described above, as well as performing parameter optimization on a fixed alignment produced by both GenAl [6] and ClustalW [20] . As we can see from Figure 1 , ClustalW gives consistently rather bad results, since it is not designed to deal with overlapping coding regions. Our method achieves better results than GenAl on sequences of evolutionary distance less than 0.8, but cannot quite compete with GenAl on sequences further apart. Here our estimation error is shown as the fraction between the average absolute deviation of our estimated parameters to the true parameter value and the true value itself. The statistical alignment method performs, when applied to evolutionary distances we are realistically going to encounter, within 10% of the true value. Similar results hold for a number of other tested scenarios, including cases where one reading frame is under much stronger selection than the other and both are under positive or both under strong negative selection. We wish to find out what effect the length of a double coding region has on our estimation accuracy. Letting the length of the double coding region in our above simulation vary from 600 down to 25, with transition and transversion rate 0.4 and 0.2 respectively, we obtain Figure 2 . As to be expected, the shorter the region, the worse our prediction results, since our data set decreases. However, above a length of 50 nucleotides we start picking up selection within a distance of ± 0.15, and above 200 nucleotides we are within the ± 0.1 mark. This is reassuring, since as mentioned above the average double coding region in HIV2 and Hepatitis B is 171 and 532 respectively. We test the confidence levels of our predictions, trying to create as 'realistic' simulated data as possible. In the light of our real data analysis, we take the Hepatitis B genome NC00397 and split it into 7 different regions, a new one starting whenever there is a change in gene structure. We evolve the sequence according to our indel model with varying transition and transversion rate of a = 0.2 -0.8 and b = a/2 respectively, and fixed selection strength of 0.5 for each of the different regions. Depending on the evolutionary distance and closely related to our results in Figure 1 , we achieve an accuracy of approximately 70 -94% with both the statistical alignment method as well as the fixed alignment method using GenAl, versus 20 -72% for the fixed alignment method using ClustalW. In contrast using the true alignment gives us an accuracy of 78 -96%. Here our estimate is counted as correct if the true value lies within the error bars around the estimated value. This is naturally highly dependent on the width of our error bars, which in some cases are indeed large, simply due to lack of data. However, the error bars for the parameter estimates of both the fixed and the summed alignment are close to identical, and thus the measure is valid if only for the sake of direct comparison. One of the reasons for the comparatively low performance on ClustalW alignments might be that those alignments often do not conserve the reading frame. As we wish to make as fair a comparison as possible, we therefore additionally manually adjust alignments to be more 'reasonable' by adjusting gap placement to conserve the reading frame. This does indeed result in considerable inprovement, thus demonstrating the volatility of results when dependent on one particular alignment. However, even when improving the fixed ClustalW alignment, the resulting accuracy after manual adjustment still falls short of that achieved by the statistical alignment method, reaching only 40 -70%. Finally, we compare our results on the last setup using simulated descendants of the Hepatitis B genome in a pairwise versus a multiple sequence scenario. When adding up to four sequences, we observe the error bars getting notably tighter and simultaneously our estimation error decreasing by about 0.01 per added sequence. This implies, as desired, a more precise estimation of selection factors for multiple sequences. We run our method on the Hepatitis B strand NC003977 and 'descendants' Woodchuck Hepatitis B strand J02442 and Ground Squirrel Hepatitis K02715, with sequences and gene structure downloaded from GenBank. As seed parameters we have all values set to 0.5 and wait between iterations for a difference in our loglikelihood of <1. Our method takes ~40 seconds to reach convergence and results are shown in Figure 3 . To see how a region acts when viewed as a whole, we also calculate the average selection acting on double coding regions, by weighting the expected counts for each mutation by the appropriate selection coefficient -in the case of a single non-synonymous change in gene A or B by the factor f A and f B respectively, and in the case of two nonsynonymous changes by the joint factor f AB . Table 1 shows the values obtained for the different regions, both single and double coding. We can see that when viewed like this, the double coding regions are on average under 0.41 selection, and thus not greatly different to the single coding ones at an average of 0.39. Due to more than 1500 sites in the Hepatitis B genome being multiple coding, we may reasonably test whether the simpler multiplicative model is an equally good fit to the full one used above. Setting f AB = f A ·f B we may perform a likelihood ratio test between the full and the restricted model, where selection acting on two different genes simultaneously gets multiplied up. With -2log Λ = 18 for 3 added parameters, the full model fits the date significantly better than the restricted multiplicative one (P = 0.0004). We apply our method to the HIV2 genomes J04542 with reasonably diverged 'descendants' U27200, M15390, DQ00835 and M30502, by splitting the genome into different regions whenever there is a change in gene structure. Setting all our initial parameters to 0.5, as above, we Simulation Results: Varying Evolutionary Distance Figure 1 Simulation Results: Varying Evolutionary Distance. Simulation results between two sequences for a double coding region of length 600 of varying evolutionary distance. The figure plots the average estimation error of the statistical method and the fixed alignment method using both ClustalW and GenAl, versus the evolutionary distance between the two sequences. The estimation error is measured as the fraction of the average absolute deviation to the true parameter value and the true value itself. The evolutionary distance is measured as a + 2b, where a and b are transition and transversion rates respectively. obtain a selection annotation for the different regions. The results of our parameter estimates are given in Figure 4 . As we can see, there is a marked difference between the estimated selection strengths underlying the different regions, with selection ranging from 0.21 -1.50. Our results seem to suggest that genes encoded by double-coding regions often show contrasting modes of evolution, where one gene is highly conserved, whereas the other is less so. Naturally all these estimates are made on relatively small regions, and thus have relatively large error bars, but tendencies towards a distinction between fast and slow evolving overlaps are nonetheless demonstrated. On the other hand, the selection on the overlap between the fifth and sixth gene in line -the vpr and the tat gene -is close to equal in both reading frames, thus indicating that the otherwise observed high and low selection values are not mere inevitable artefacts of our model. We return to this in the discussion. One of the most remarkable observations is that within each reading frame, selection on single coding regions appears to be more constrained than in double coding ones. As before, we calculate the selection acting on each region as a whole, as shown in Table 2 is in line with the results shown by de Oliveira et al. [2] and more recently by McCauley et al. [13] , but somewhat contrary to general belief [19, 22] . Clearly within the HIV2 genome there is much less data than with Hepatitis B, so it is harder to assign a true significance to these figures. However, our results do appear to suggest less stringent selection on overlapping regions than on single coding ones, thus maybe indicating the overlapping regions to be a relatively young feature in the virus. We have introduced a novel method for estimation of selection strengths that explicitly incorporates uncertainty in estimated alignments. By integrating a statistical alignment procedure into our parameter estimation, we do not rely on a fixed alignment input. Instead, we calculate the expected number of observations, and are thus weighting our parameter estimates by the probability of each possible alignment. We naturally can not compete with knowing the true alignment, something which sufficient and extremely time consuming manual work can get close to. We do however offer a fast, automatic and efficient alternative to the use of a fixed alignment, which provides a quick and easy way for producing selection factors for different regions in a viral genome. We outperform alignments given by ClustalW consistently. We even beat GenAl for sequences of evolutionary distance below 0.8 and only do slightly worse for ones further apart. It is however additionally worth noting that the sequences we have and generally will be dealing with, will generally The average selection acting on each of the seven regions of the Hepatitis B genome, measured by weighing each expected mutation by its appropriate selection coefficient. Selection on double coding regions appears to tend to be more lenient have an evolutionary distance of 0.4 -0.9. We are therefore encouraged to see that our method is competetive compared to the slightly more refined GenAl and hope that this is amplified once also extended to include protein alignment. More importantly our method is statistical, which means it can be more readily incorporated into a maximum likelihood estimation framework, whereas GenAl works on a count-basis. We test our method in a number of different simulation studies against the use of a fixed alignment, which we obtain using ClustalW. We show that on average our statistical approach has up to 30% higher absolute sensitivity, and that both evolutionary distance and the length of a double coding region have a lesser effect on our results than when using a fixed alignment. Our study focuses on trying to understand the selection mechanism underlying overlapping reading frames. On the Hepatitis B genome, which boasts over 1500 multiple coding sites, we investigate several questions such as the selection a mutation is under, when it causes a non-synonymous mutation in two genes simultaneously. That is to say, if gene A and gene B are under selection f A and f B respectively, will a mutation affecting both necessarily be [22] . Another feature which is particular to our method, is that we may seperate selection acting on the different reading frames in an overlapping region. We find especially in HIV2 a certain division of selection occurring, similar to that observed in Potato Leafroll Virus [4] and in Microviridae [16] . Essentially, it appears as though in an overlapping region one gene can take over the fastly evolving function, whilst the other behaves more conserved. Since this is not something we observed in our simulation studies, it seems to be no artefact of our model. One possible evolutionary scenario that could explain this observation is the following: when an overlapping region is 'created'for example by the elongation of one of the genes involved, then it is likely to initially be under nonnegative selection. Since the organism survived both with and without the overlap, it might be expected to be able to evolve without detrimental effects. A thus possible behaviour would be for the newly coding region to be encouraged to evolve quickly, whilst the other gene remains under negative selection as before. The estimated selection strengths may subsequently help deduce which overlaps are the 'newer' regions -for example our study suggests that the pol gene extended itself both onto the gag and the vif gene. The effect would essentially be similar to that noted on selection occurring on duplicated genes, where the duplicated gene reaches fixation in the population due to initially being under positive selection [24] . Up till now, other methods dealing with related issues have made use of the concept of K a /K s ratio, which however creates problems when applied to overlapping reading frames [4, 7, [14] [15] [16] . For this reason McCauley et al. [13] decided on a different HMM based approach and estimated selection as acting on a single nucleotide basis, but at the cost of not being able to pry apart selection acting on different reading frames. Most importantly however, all of the above methods use a fixed alignment and are thus prone to a great variability in their estimated parameters, dependent on the alignment. Our method manages to circumvent this problem by using a statistical approach, and thus we account for the uncertainty inherent in the alignment by considering all, rather than picking a single "best" alignment. The improvement we observe by doing this, makes us suggest that our approach of marginalizing over alignments may benefit other sequence-based inferential methods, such as for example methods for identifying conserved binding motifs. One drawback to our method is the fact that for each descendent sequence we model transition and transversion rates as constant along the genome. This is a gross simplification, and something that should be dealt with in future work. As mentioned above, we would also like to superimpose protein alignment on our existent statistical alignment method, in accordance with the idea behind [6] . Another point is our fixing a partition prior to analysis. It would be even more interesting to be able to incorporate a hidden Markov model approach, in which breakpoints between regions would be chosen organically from the data. We could then truly start questioning which parts of the genome behave in different ways, as opposed to being restricted to the 'trial and error' approach that is the essence of our method now. We describe the type of problem we are confronted with according to a specific example, shown in Figure 5 . Due to the 3-periodicity of the genetic code, there are three global reading frames in which a sequence may code in the forward direction, henceforth referred to as GRF1, GRF2 and GRF3. In viruses these reading frames tend to encode simultaneously for up to three different overlapping genes on each strand, resulting in multiple coding regions. We will be looking at single stranded RNA viruses, which predominantly code in the forward reading direction only. Amendments to our model would have to be made to include reverse reading frame encoded genes. We are given two sequences S 1 and S 2 , descended from a common ancestor, together with the gene structure G of S 1 -in the case of Figure 5 this is a genome with two genes which overlap. Say these genes code in GRFs 1 and 2 respectively. Let us first assume we already have an alignment between our two sequences, and we wish to understand the way selection works on different regions of the genome. An initial question to ask would be, whether single and double coding regions behave in the same way. We thus, as shown, partition the genome into five segments, making a split wherever a gene starts or stops. These five segments we then assign to be of one of three region-types: non-, single-and double coding. When considering the effect a mutation of the indicated nucleotide C in the overlapping region of S 1 might have, we must consider its coding role in both reading frames. In GRF1 it is in the third position of the codon AGC and in GRF2 in the second position of the codon GCT. In the genetic code the codon AGx codes for serine or argenine, depending on whether x is a purine or a pyrimidine, respectively. On the other hand GxT codes for four different amino acids, depending on the nature of x. Therefore a transition in the nucleotide C will have no effect on the amino acid encoded by GRF1, whereas a transversion will. In GRF2 on the other hand, both will result in a non-synonymous substitution. Additionally, the selection strengths acting on either gene might be different, due to one of them evolving faster than the other. Since we wish to analyse selection happening over a reasonable evolutionary distance, our aim is to be able to draw conclusions without relying on a prior alignment. Instead of estimating evolutionary parameters using observed substitution counts froma fixed alignment, we will therefore use an alignment model to generate expected substitution counts and from these use a maximum likeli-hood method to estimate all evolutionary parameters. In this manner we may sum over the uncertainty of the alignment -an uncertainty that will be high for distantly related viruses. Since our alignment model includes a substitution model, we iteratively switch between both it and our ML-procedure. Figure 6 depicts the basic outline of our program. To be able to calculate the probability of a certain alignment between S 1 and S 2 , we need to devise a model for the evolution of a sequence. We will be working with a simple 3-state HMM indel model, using a more complex nucleotide substitution model, given by an emission matrix E, for the emission probabilities. We wish to investigate region-specific selectional behaviour along the genome of S 1 . We may thus apply a partition P to our sequence S 1 , given by a sequence of partition points {p 0 , p 1 , ..., p |P| }, where clearly p 0 = 0 and p |P| = |S 1 |. Because we are interested in certain global features, we may wish to group particular partition segments together into regions of a Sequence Annotation Figure 5 Sequence Annotation. An example of our input data and annotation. We see here the 'ancestral' sequence S 1 , whose genes structure G is given by coding regions in two reading frames GRF1 and GRF2. We apply a partition P to the sequence, where a breakpoint occurs whenever there is a change in gene structure. We annotate this partition with R = 3 different types of region for non-, single and double-coding respectively. Finally we have the descendent sequence S 2 . particular type. Say we have R regions, then each partition segment [p k , p k+1 ] with (0 ≤ k < |P|) gets assigned to a certain 'region-type' r, with (r ≤ R), where regions of the same type are assumed to evolve in a similar way. As stated above, since we are interested in investigating the evolutionary behaviour of viruses in particular, we wish to work with a substitution model, which specifically accounts for the presence of multiple coding regions. For our evolutionary model E we use a model very similar to the one in de Groot et al. [1] . For the convenience of those not familiar with our earlier work, we include a description of it in the following. Most amino acids are encoded by several different codons, meaning that a mutation may often result in no change in amino acid. Regarding a certain nucleotide in a codon, depending on its context, we may generally divide it into the following degeneracy classes, where a substitution would result in • four times the same amino acid. Figure 6 Model Setup. A graphic representation of our method. As input, we give the 'ancestral' sequence S 1 , its gene structure G, our desired partition P and our region annotation R of the partition segments. We also input the 'descendent' sequence S 2 , as well as our seed parameters for , a, and b. From this we may generate both our seed emission matrix E and the typeannotation-array t = [t 1 , t 2 , t 3 ] belonging to each locus along the sequence S 1 . These then get input into our alignment procedure, which subsequently over the sum of all possible alignments, calculates the expected counts C of a certain substitution of a certain type in a certain region. This information gets transferred to our maximum-likelihood (ML) method, which generates our new parameter values, maximizing the expected observations C. The resulting emission matrix E gets fed back into our alignment procedure, and the loop continues until a change in parameters is below some given threshold. • two different amino acids, depending on whether a transition or transversion has occurred. • four different amino acids, regardless of the type of substitution. We shorthand these as being of type 4, 2 and 1 respectively. A few sites are not classifiable into one of the above three classes -ATx codes for three isoleucines and one methionine and CGG and GGG are synonymous although one results from the other by a transversion. Treating, for example, ATG as a type 1 site and ATA, ATC and ATT as type 4 sites, means however, that the approximations made by us are most likely to be minor. We model the evolution of our sequences according to the Hein & Støvlbaek [5] model. When looking at a nucleotide in the ancestral sequence, for each reading frame we assign a certain state-dependent 'degeneracy-type' t to it, depending on its context. This will, in a coding region in a particular reading frame, be either of degeneracy 1, 2 or 4 and for non-coding will always be designated as 0. Since we are considering overlapping reading frames, we thus obtain for each nucleotide in the ancestral sequence a certain state-dependent 'degeneracy-type-array' t = [t1, t2, t3] -an array consisting of the degeneracy annotation of a nucleotide for each of the three reading frames. In our example in Figure 5 we can see an overlap between two genes, say genes A and B. This results in an annotation of [2, 1, 0] for our nucleotide C in the overlap, meaning that we have a degeneracy annotation of 2 and 1 with respect to gene A and B respectively. Using this degeneracy annotation we incorporate the concept of selection factors into our framework: transitions and transversions occur according to the Kimura [8] model, and non-synonymous substitutions get accepted by a selection factor specified in the following. Consider a nucleotide x in a region of type r in S 1 with degeneracy-type array t. Then our factors will be given by where with F determined as explained above. We thus are able to construct an emission matrix E, where E(r, t1, t2, t3, x, y) is the probability of in region r, nucleotide x of type [t 1 , t 2 , t 3 ] mutating into nucleotide y. We would like to note that even though our alignment model does assume independence of sites, we model a local dependency in our evolutionary model by conditioning our emission probabilities on the nucleotide context in the ancestral sequence. An undoubtable improvement would be to model the dependency as continuous throughout the evolutionary process [17] . However, as noted by the authors themselves, the elaborate MCMC method developed in order to do this makes it a computationally intractable option. To compute the probability of an alignment we use a simple indel model with Match, Insert and Delete states. We have as alignment parameters a gap-opening, a gap-extension and a transition probability from any state to the end state. All other state transition probabilities may be derived from these. The Insert and Delete states emit a nucleotide from a uniform distribution, aligned to a gap. In the Match state nucleotide pairs are emitted according to our above model. We wish to eliminate the bias in parameter estimation created by the use of a fixed alignment. For this, we work with a probabilistic alignment, which instead of producing an 'optimal' alignment, handles the set of all possible alignments and their likelihood. It computes posterior probabilities for each state at every nucleotide position. We thus are considering all possible sequence alignments and weighing them appropriately (see [23] ), according to our indel model. This method has been previously used and described in further detail in [9, 10] . Note, that when referring to the insertion and deletion states, the related posteriors are added together so that we obtain the posterior probability of a certain nucleotide not being aligned, as opposed to belonging to a particular gap. During the alignment procedure, our alignment parameters are estimated in a few iterations of the Baum-Welch algorithm [3] . The implementation of the algorithm, including banding to cut computational demands, was generated automatically by the HMM compiler program HMMoC [11] . As shown in Figure 6 , we initially have as an input all the sequence and genome structure data, as well as a set of seed parameters. We subsequently use our alignment model to generate the posterior probabilities of every nucleotide position being in each state. This is done by using the forward and backward algorithms applied to our alignment indel model, as is standard HMM procedure. From these posteriors we may calculate, for each degeneracy in each region, the expected number of times an identity, transition and transversion is used. For a site of degeneracy t = [t 1 , t 2 , t 3 ] in region r, let this be , and respectively. Since , and were the probabilities for a site of degeneracy t in region r of an identity, transition or transversion occurring (see equations 1, 2, 3), we may rewrite the emission term of the log likelihood log L as follows: For this function of the 3R + 2 emission parameters , a and b we now find the maximum likelihood estimates using the Newton-Raphson iteration method. We may do this by taking the explicit derivatives of the likelihood function, possible because of the simple substitution model used. If we had opted for a more complicated model, we would need optimization methods that did not rely on derivatives and would subsequently be slower, though the estimation would still be possible. Once the change in log-likelihood between two iterations has fallen below some given threshold, we take the likelihood to have converged. We then generate a new emission matrix E to be fed back into our alignment procedure in order to generate new posterior probabilities. Once the likelihood function has converged below some set threshold, we output the final set of estimated selection parameters. We may also, if desired, construct an alignment, either using the Viterbi path or posterior decoding. We would like to be able to apply our method to multiple sequences, thus extrapolating more information where possible. We could of course devise a multiple alignment indel model, and develop a new likelihood function from which to maximize over all tree branches simultaneously. This however would be of computational much higher demand, runtime increasing exponentially with the addition of each new sequence. Instead, we therefore opt to work with a multiple pairwise alignment under the assumption of a star shaped tree, with the reference sequence as the root in the star topology. This implies viewing the evolution of the pairwise sequences as independent, which is an approximation which we wish to address in future work. This merely requires per additional sequence an extra transition and transversion parameter, since selection is acting on the gene in the ancestor and we assume this to be constant over all branches. The modification to our program is thus trivial, with only a linear increase in runtime. As an input we have, for lack of better terminology, the ancestral sequence A and its N descendants D 1 , ..., D N , together with the seed parameters for the selection factors (f 1 , f 2 , f 3 ) R on each region R as well as n transition and transversion parameters (a, b) n respectively. We then build a set of N pairwise alignments between the ancestor A and its N descendants. Each one of these obtains a likelihood function log L as given in equation 6. Now we create a new likelihood function log L* which is the sum of the N log likelihoods. If is the number of expected identities of type t in region r between the ancestral sequence A and its n th descendant, then our assumption of independence implies that x ts t r , x id t n r , , is the full likelihood of observing all N sequences under our model. Note here, that the probabilities P are dependent on the sequence-dependent transition and transversion rates (a, b) n and the selection factors (f 1 , f 2 , f 3 ) R which in turn are not dependent on n, since we are assuming selection to occur on the gene in the ancestral sequence. Maximizing this new log likelihood function, we proceed as above and estimate a new set of selection factors and a set of sequence specific transition and transversion rates, from which we may generate a new set of pairwise statistical alignments. log log log log , , ,

Seasonality of Influenza A(H3N2) Virus: A Hong Kong Perspective (1997–2006)

BACKGROUND: The underlying basis for the seasonality of influenza A viruses is still uncertain. Phylogenetic studies investigated this phenomenon but have lacked sequences from more subtropical and tropical regions, particularly from Southeast Asia. METHODOLOGY/PRINCIPAL FINDINGS: 281 complete hemagglutinin (HA) and neuraminidase (NA) sequences were obtained from influenza A(H3N2) viruses, collected over 10 years (1997–2006) from Hong Kong. These dated sequences were analyzed with influenza A(H3N2) vaccine strain sequences (Syd/5/97, Mos/10/99, Fuj/411/02, Cal/7/04) and 315 other publicly available dated sequences from elsewhere, worldwide. In addition, the NA sequence alignment was inspected for the presence of any naturally occurring, known, neuraminidase inhibitor (NAI) resistance-associated amino acid mutations (R292K and E119V). Before 2001, the Hong Kong HA and NA sequences clustered more closely with the older vaccine sequences (Syd/5/97, Mos/10/99) than did sequences from elsewhere. After 2001, this trend reversed with significant clusters containing HA and NA sequences from different locations, isolated at different times, suggesting that viral migration may account for much of the influenza A(H3N2) seasonality during this 10-year period. However, at least one example from Hong Kong was found suggesting that in some years, influenza A(H3N2) viruses may persist in the same location, perhaps continuing to circulate, sub-clinically, at low levels between seasons, to re-emerge in the influenza season the following year, relatively unchanged. None of these Hong Kong influenza A(H3N2) NA sequences contained any of the known NAI-resistance associated mutations. CONCLUSIONS/SIGNIFICANCE: The seasonality of influenza A(H3N2) may be largely due to global migration, with similar viruses appearing in different countries at different times. However, occasionally, some viruses may remain within a single location and continue to circulate within that population, to re-emerge during the next influenza season, with relatively little genetic change. Naturally occurring NAI resistance mutations were absent or, at least, very rare in this population.

Despite many hypotheses and studies, the underlying basis for the annual recurrence of seasonal influenza still remains a mystery [1] . Hammond et al. [2] postulated a rapid, global dispersion of 'airborne aerosolized influenza virus' via the atmosphere, to account for the persistence and spread of the disease. Recent reviews have discussed the various approaches to resolving this question, and identified various factors that may be involved, including: properties of the virus itself (mutation rates and immune escape), properties of the host (seasonal variation in host health and behavior, e.g. crowding and air travel, production and dissemination of bioaerosols through sneezing and coughing), and properties of the environment (temperature, humidity and weather variations, e.g. El Nino) [3] [4] [5] [6] [7] . Some of these factors have been incorporated into mathematical models to attempt to understand the driving forces behind influenza seasonality [6, [8] [9] [10] [11] [12] [13] . Sequence-based analyses have become very popular recently and have shed some interesting insights into possible underlying mechanisms of influenza seasonality. Many of these also urge for (or at least hint at) the need for more sequences from tropical regions to be made publicly available to increase the accuracy of such analyses [14] [15] [16] [17] [18] [19] [20] [21] . Other studies have analyzed genetic data together with the even more scarcely available antigenic data, in attempts to understand and even predict the most likely emerging strains [22] [23] [24] [25] . Even the application of mass spectrometry has been applied to influenza surveillance [26] . Hong Kong is a subtropical region of almost 7 million people, 95% of whom are ethnic Chinese, with a mean temperature of 24uC and mean relative humidity of 79% [27] . It lies geographically in the Northern hemisphere, and its influenza season occurs during February-April, sometimes with a second peak during June-August, each year. In contrast, other Northern hemisphere countries usually have a more extended influenza season from November to March/April, whereas the influenza season of Southern hemisphere countries usually occur from May to September [7, 28] . Hence, Hong Kong may be unique in that its biphasic influenza seasonality seems to straddle those of the Northern and Southern hemisphere countries, making the molecular epidemiology of its circulating influenza viruses of great interest. In addition, Hong Kong and Southern China have been referred to as the 'epicenter' for new influenza A viruses with pandemic potential for over 25 years now [29] . For all of these reasons, any investigation of the underlying basis for influenza seasonality may benefit greatly from a study of influenza viruses isolated from Hong Kong. In this study, an analysis is presented of 281 Hong Kong influenza A(H3N2) hemagglutinin (HA) and neuraminidase (NA) full-length, dated sequences collected over 10 years (1997-2006) to assist the ongoing efforts to elucidate the underlying basis for the seasonality of influenza A(H3N2). The HA and NA ML phylogenetic trees (with and without the additional, down-loaded contemporary sequences from publicly available archives) produced in this study are too large to include as separate figures in this paper and have been published as online Supporting Information in a scrollable PDF format for further inspection on the PLoS ONE journal website (http://www. plosone.org/home.action). For each of these trees, certain clusters of interest have been highlighted using annotated red boxes or ellipses, and will be specifically referred to, in the following text for further description and discussion. All of these 281 Hong Kong influenza A(H3N2) HA and NA sequences have been deposited on GenBank (Accession nos.: EU856814-EU857094 for HA, and EU857095-EU857375 for NA sequences). Figures S1 and S2 show the relationship between the 281 HA and NA sequences for the Hong Kong influenza A(H3N2) samples and 4 World Health Organization (WHO) influenza A(H3N2) vaccine strain HA sequences (Syd/5/97, Mos/10/99, Fuj/411/02 and Cal/7/04). These Hong Kong HA and NA sequences were inspected to determine if there were any sequences from consecutive influenza seasons occurring on the same branch, indicating that viruses with the same or very similar HA and NA gene sequences were occurring in adjacent influenza seasons. This would suggest that that particular virus carrying this gene may have remained 'latent' in that population, to re-emerge in the same population the following season. One example of such possible viral persistence between influenza seasons was found, with HA and NA sequences from the same viruses (5251Jan02 and 5267Jan03, as indicated in Figures S1 and S2 for the HA and NA phylogenetic trees, respectively), showing a similar clustering pattern for both these genes, separated by at least one year. Interestingly, the HA sequence from sample 5250Jan02 clusters closely with those from samples 5251Jan02 and 5267Jan03 (Figures S1 and S3), but its NA sequence lies some distance away on a separate branch (Figures S2 and S4 ). This may suggest a possible reassortment event, either with its HA or NA gene segment. Further full genome sequencing and analysis may resolve this issue. The The relationship between the WHO vaccine HA and NA sequences and those from Hong Kong and elsewhere can be seen even more clearly in Figures S3 and S4 when the 315 JCVI sequences are added to each tree. Although these contemporary JCVI sequences are mainly drawn from just three additional locations, they still represent the Northern hemisphere (New York, USA) and the Southern hemisphere (Western Australia and various locations in New Zealand). Again, for reference, the January Hong Kong HA and NA sequences from each year are again highlighted in red boxes. In addition, in Figures S3 and S4 , red ellipses have been added to show where similar HA and NA sequences from other, non-Hong Kong locations have clustered with Hong Kong sequences on the same branch. The dates of such sequences may be the same (within the limit of the one month temporal resolution used in this study), relatively similar, or very different. These highlighted clusters serve to demonstrate the mobility and ubiquity of this influenza A(H3N2) virus, worldwide, during this 10-year period, i.e. genetically similar viruses can appear in different parts of the world at similar and also different times. These examples are not meant to be exhaustive and other such examples may be found in these trees. These number and position of the clusters indicated by the red ellipses differ between Figure S3 (HA sequences) and Figure S4 (NA sequences) probably because there are different selection pressures acting on these two genes as they have quite different functions (i.e. the HA protein is used by the virus to bind to the host cell for entry, whereas the NA protein is an enzyme that enables new progeny viruses to leave the host cell). Also, in Figures 3 and 4 , for both the HA and NA gene sequences, respectively, there is a large region of transition between Mos/10/99-like and Fuj/411/02-like viruses, containing sequences collected during 1999-2005. Inspection of the protein-coding alignment for NA sequences showed no evidence of any of the known N2 neuraminidase inhibitor (NAI) resistance-associated mutations, R292K and E119V, in any of these Hong Kong influenza A(H3N2) isolates collected between 1997-2006. This comparative analysis of dated HA and NA sequences from influenza A(H3N2) viruses from 4 geographical regions of the world (New York, Western Australia, New Zealand and Hong Kong) attempts to elucidate more clearly the behavior of influenza A(H3N2). This study contributes an additional 10 years of influenza A(H3N2) HA and NA sequences, from Hong Kong, to the publicly available sequence database (GenBank), which should aid other researchers investigating an underlying basis for influenza A(H3N2) seasonality. The approach of the analysis in this study has been to compare accurately dated HA and NA sequences using established phylogenetic techniques to examine which sets of sequences cluster together, and by examining the dates and locations from which they were collected, to infer the movements of the virus within those dated periods. A similar analysis was recently performed using dated whole genome influenza A(H3N2) sequences from New York, New Zealand and Australia, downloaded from publicly available databases, in an attempt to test two competing hypotheses: whether seasonal influenza A(H3N2) viruses continuously 'migrate' around the world, particularly between Northern and Southern hemispheres; or whether the virus remains 'latent' in one location and reactivates each year to produce the familiar pattern of influenza seasonality [19] . As these authors stated, ideally, whole genomes should be used for more accurate phylogenetic analyses of influenza virus as has been reported previously [14, 15] , with at least one good reason for this being the potentially misleading conclusions caused by influenza viral HA and NA gene reassortments [19] . However, many laboratories worldwide do not have the resources to perform whole genome sequencing, which is expensive -particularly those in subtropical and tropical regions from where such influenza sequence data is significantly lacking. In addition, with any phylogenetic study such as this, there is always a limitation on the number of sequences that are available (i.e. the number of respiratory samples containing influenza viruses that have been collected and sequenced in any one influenza season), and the number that can be comfortably analyzed within a given time-frame (i.e. the limitations on computing power, which again may be more of a problem in tropical and subtropical countries that are more resource-limited). In this particular study, there is also a problem of sample bias as these sequences were obtained from only hospitalized children (rather than from those who remained in the community), and may therefore reflect the influenza virus population isolated from the more clinically severely ill patients (or those with more concerned, anxious parents). However, there was a rationale for deliberately selecting children's samples for this study. The reason for this is that, especially in Hong Kong, unlike adults, children of school age (1-10 years old) are more likely to stay close to home and not travel far from home, which would minimize any importation of influenza viruses from overseas. This would therefore reduce this confounding factor when assessing the migration vs latency hypotheses as an explanation for the underlying mechanism of influenza seasonality in Hong Kong. Accepting all of these shortcomings, some of which are inevitable for such phylogenetic studies (since not all samples can be collected and sequenced from all infected individuals from all over the world for any particular virus), there are still some useful conclusions that can be gained from this study: i) from Figures S1 and S2, there is at least one example of a virus that reappears, relatively unchanged between consecutive influenza seasons, and which can be seen as some evidence to support the 'latency' hypothesis [19] . Since the same viruses show this same pattern of clustering in both their HA and NA genes, this reduces the likelihood that this was due to a reassortment event in one of these genes, i.e. it is less likely for the same viruses to have reassorted both the HA and NA genes during the same years -though of course this possibility cannot be ruled out. Whole genome sequencing would be useful to confirm this if resources are available in the future. ii) from Figures S3 and S4 , it is difficult to say whether viruses from Hong Kong preceded (or gave rise to) those from elsewhere, since viruses from outside Hong Kong can be found to both pre-date and post-date those isolated from Hong Kong. However, there are several examples of similar viruses being isolated in different parts of the world at about the same time (as shown in some of the red ellipses in Figures S3 and S4 Figures S3 and S4) , though this may not necessarily mean that the Fuj/411/02-like viruses originated from there. Hence, these results suggest that the seasonality of influenza A(H3N2) may, in fact, result from a combination of these 2 models postulated by Nelson et al [19] , i.e. mainly migration, but with occasional examples of latency. Again, due to the unavoidable, incomplete sampling and sequencing of influenza A(H3N2) viruses worldwide, this support for these hypotheses ('migration' and 'latency'), as presented here, is admittedly, only patchy at best. However, the fact that we can find at least one example supporting the latency hypothesis, in these 281 HA and NA sequences spanning 10 years (1997) (1998) (1999) (2000) (2001) (2002) (2003) (2004) (2005) (2006) , suggests that they may both play a part in the underlying mechanism governing influenza A(H3N2) seasonality. The large variations in both the HA and NA sequences during the transitional period between Mos/10/99 and Fuj/411/02 ( Figure S3 and S4) may have reduced the protective efficacy of any pre-existing influenza antibodies (from prior infection or immunization by Mos/ 10/99-like viruses) before the inclusion of the Fuj/411/02 strain in the World Health Organization (WHO) recommendations for the seasonal influenza vaccine in 2004. Such a reduction in protective efficacy from the contemporary WHO influenza vaccine has indeed been suggested in several reports [14, 30, 31] . Similarly, although, the NA antigen is not usually specified in seasonal influenza vaccine, and its protective effect is unknown, its sequence variation in NA during this transitional period, may have also contributed to the reduced protection provided by the seasonal influenza A(H3N2) vaccine component during this period, when the Fuj/411/02-like viruses were emerging. In summary, this study has provided additional data from Hong Kong to support a mainly migratory mechanism to explain the underlying seasonality of influenza A(H3N2) viruses. However, there may be small localities of so-called 'latency' where viruses remain circulating at low levels within that local population, to reemerge during the influenza season of the following year, with relatively little genetic change. This may affect only a minority of these populations, with the majority being infected by newly imported influenza viruses from elsewhere. These concepts are summarized in Figure 1 . Two recent papers [37, 38] have suggested that the existing evidence tends to support a migration rather than a latency mechanism to explain the annual seasonality of influenza A. However, they did not have access to large numbers of influenza sequences from Southeast Asia. Thus, in view of the new data presented in this study, it is hoped that such hypotheses may be revised, to include a contribution from the latency mechanism. In some populations (perhaps those more localized in Southeast Asia), this latency mechanism may contribute more significantly to the underlying basis for the seasonality of influenza A -a possibility not entirely ruled out by one of these studies [38] . More data is required to explore this hypothesis in those populations. Admittedly, the fact that in this 10-year (1997-2006 ) data set of almost 300 influenza A(H3N2) HA and NA sequences from Hong Kong, we have only found one particular example supporting the latency mechanism, does suggest that this is a relative rarity, and that the migration mechanism is probably responsible for the majority of influenza seasonality patterns seen worldwide. Finally, it is interesting, and to a certain extent, reassuring, to note that during this 10-year period (1997) (1998) (1999) (2000) (2001) (2002) (2003) (2004) (2005) (2006) , no influenza A(H3N2) viruses isolated from this cohort of NAI-naïve children from Hong Kong, showed any evidence of naturally occurring, known NAI resistance-associated amino acid mutations (R292K and E119V). There are many parts of the puzzle remaining, such as exactly how do some of these influenza viruses migrate so widely, yet still remain relatively similar over several years. How is the low level circulation of 'latent' influenza viruses accomplished between seasons? Is this a property of the population's host immune responses, the virus, the environment or some combination of each of these factors? In a collaborative effort to answer these questions and perhaps to improve the accuracy of influenza strain forecasting and vaccine composition [38] , it is also hoped that this study will encourage more researchers in Southeast Asia to make their influenza sequences publicly available for analysis -especially whole genome sequences where resources permit, and particularly sequences from more tropical countries where influenza is prevalent all year round, with less well-defined seasonal peaks. Approximately 30 influenza A(H3N2) isolates for each year of 1997-2006 obtained from the nasopharyngeal aspirates (NPAs) of children aged 1-10 years, admitted to the Prince of Wales Hospital (PWH) in Hong Kong, were retrieved from long-term archives (stored either at 270uC or in liquid nitrogen) for hemagglutinin (HA) and neuraminidase (NA) gene sequencing and analysis. These children presented with acute respiratory illness and did not receive anti-influenza therapy before or during their illness. As influenza A(H1N1) was the predominant circulating virus during 2006, very few H3N2 isolates were obtained for that year. Table 1 shows the age and sex distribution of these children. Verbal consent for the initial diagnostic testing of these samples for respiratory viruses, including influenza, was obtained from the parents of these children. Such verbal (rather than written) consent is routine for such standard diagnostic tests in Hong Kong. The local Joint Chinese University of Hong Kong-New Territories East Cluster (Joint CUHK-NTEC) Research Ethics Committee institutional review board awarded ethics approval for this retrospective sequencing and molecular epidemiological study (study reference number: CRE-2005.390) without the need to obtain further, explicit, written consent. This is also in agreement with the American College of Epidemiology guidelines [33] for such retrospective epidemiological/surveillance studies. The samples retrieved from the deep-freeze archive were first generation isolates, i.e. the original clinical samples (NPAs) had been inoculated, for routine diagnostic testing, into Madin-Darby Canine Kidney (MDCK) cells. These MDCK-cultured viral isolates were originally confirmed to be influenza A(H3N2) before being frozen and archived. For this study, these frozen isolates were retrieved from deep-freeze, thawed and then used directly for sequencing. If any of these newly-thawed, archived samples failed to amplify at this stage, as determined by ethidium bromide staining and gel electrophoresis, it was inoculated into MDCK cells and re-cultured. After this additional step, most isolates were successfully sequenced. If this step still failed, then an alternative isolate was retrieved from deep-freeze for sequencing. Total RNA was extracted using the PureLink TM Viral RNA/ DNA Kit (Invitrogen, Carlsbad, USA) according to the manufacturer's instructions, and resuspended in 50 mL of RNase-free water. Reverse transcription-polymerase chain reaction (RT-PCR) was carried out with SuperScript III One-Step RT-PCR System with Platinum Taq DNA Polymerase kit (Invitrogen, Carlsbad, USA) according to the manufacturer's protocols. In brief, a 25-mL reaction mix containing 0.5 mM of each forward and reverse primers and 10 mL of extracted RNA template were used for the RT-PCR. Sets of primers was designed to amplify the complete influenza A(H3N2) HA and NA genes ( Table 2) . The RT reaction (55uC for 30 min) was followed by 94uC for 2 min and 40 cycles of PCR (94uC for 30 sec, 50uC for 30 sec, and 68uC for 1 min 45 sec, for each cycle) and a final extension at 68uC for 10 min. The PCR products were purified by MicroSpin Sephacryl S-400 HR columns (Amersham Biosciences, UK). Sequencing reactions were performed using BigDyeH Terminator v3.1 Cycle Sequencing Kits (Applied Biosystems, Foster City, USA) with 2.5 mL of template cDNA. For sequencing the HA and NA genes, the primers used are shown in Table 3 . Sequencing reactions were performed on an Applied Biosystems 3130 ABI sequencer (ABI, Foster City, USA) and in both directions to cross-check the results. Alignments of nucleotides sequences were carried out using SeqScape v2.5 (Applied Biosystems, Foster City, USA). These Hong Kong influenza A(H3N2) HA and NA gene sequences were aligned and edited in BIOEDIT v.7.0.9.0 [32] . After alignment and manual editing, to enable all sequences (i.e. both the Hong Kong and JCVI reference sequences) to have the same final length for the construction for each of the phylogenetic trees shown in Figures S1 to S4 , the sequence lengths were: This was due to different HA and NA sequence lengths being available in the respective JCVI and WHO sequence data bases. Phylogenetic tree construction was performed with PAUP* version 4.0b10 [34] by using a maximum likelihood (ML) approach, under an optimum model of evolution as determined by MODELTEST v3.7 [35] . Due to the large dataset and to reduce the time required for computation, optimal trees were searched for by using a nearest neighbor interchange (NNI) heuristic search strategy. Bootstrapping was performed within PAUP* and displayed using exported PDF files created using Figtree v1.0 (previously available from: http://tree.bio.ed.ac.uk/ software/figtree/), which were subsequently annotated using Adobe Acrobat Professional 6.0. The phylogenetic trees for these influenza A(H3N2) HA and NA sequences for the period 1997-2006 were rooted against the reference HA and NA sequences obtained from the influenza vaccine strain A/Syd/5/97(H3N2) downloaded from the Los Alamos National Laboratory (LANL) database (http://www.flu. lanl.gov/vaccine/) [36] . This strain was used as it was representative of the influenza A(H3N2) viruses circulating in 1997, i.e. at the start of this Hong Kong influenza A(H3N2) archive. Other available HA and NA sequences for other vaccine strains (Mos/ 10/99, Fuj/411/02, Cal/7/04) were also downloaded and included in all the phylogenetic trees. For comparison with other contemporary influenza A(H3N2) sequences worldwide between 1997-2006, all publicly available (at the time of this analysis) dated HA and NA sequences from children of similar ages (0-16 years old -the upper range was extended to include more sequences), spanning this period were downloaded from the then TIGR (The Institute for Genomics Presence/absence of established N2 neuraminidase inhibitor (NAI) resistance-associated mutations, R292K and E119V Once the NA nucleic acid sequences were obtained, they were aligned, in-frame for protein coding, and converted to amino acids using the built-in function in BIOEDIT. The presence or absence of the established NAI resistance-associated mutations, R292K and E119V, was then determined by inspection of the resulting amino acid alignment, with reference to the influenza A/Syd/5/ 97 (H3N2) NA sequence, which was isolated before the licensing and widespread use of neuraminidase inhibitors that began after 1999/2000. Table S1 The 315 downloaded TIGR(JCVI) influenza A(H3N2) HA and NA sequences used to construct the multi-country HA and NA phylogenetic trees shown in Figures S3 and S4 . MS Excel file containing the GenBank Accession numbers of the 315 downloaded TIGR(JCVI) influenza A(H3N2) HA and NA sequences which were used to construct the multi-country HA and NA phylogenetic trees shown in Figures S3 and S4 Author Contributions

H5N1 and 1918 Pandemic Influenza Virus Infection Results in Early and Excessive Infiltration of Macrophages and Neutrophils in the Lungs of Mice

Fatal human respiratory disease associated with the 1918 pandemic influenza virus and potentially pandemic H5N1 viruses is characterized by severe lung pathology, including pulmonary edema and extensive inflammatory infiltrate. Here, we quantified the cellular immune response to infection in the mouse lung by flow cytometry and demonstrate that mice infected with highly pathogenic (HP) H1N1 and H5N1 influenza viruses exhibit significantly high numbers of macrophages and neutrophils in the lungs compared to mice infected with low pathogenic (LP) viruses. Mice infected with the 1918 pandemic virus and a recent H5N1 human isolate show considerable similarities in overall lung cellularity, lung immune cell sub-population composition and cellular immune temporal dynamics. Interestingly, while these similarities were observed, the HP H5N1 virus consistently elicited significantly higher levels of pro-inflammatory cytokines in whole lungs and primary human macrophages, revealing a potentially critical difference in the pathogenesis of H5N1 infections. These results together show that infection with HP influenza viruses such as H5N1 and the 1918 pandemic virus leads to a rapid cell recruitment of macrophages and neutrophils into the lungs, suggesting that these cells play a role in acute lung inflammation associated with HP influenza virus infection. In addition, primary macrophages and dendritic cells were also susceptible to 1918 and H5N1 influenza virus infection in vitro and in infected mouse lung tissue.

The influenza pandemic from 1918 to 1919 was the most devastating infectious disease pandemic ever documented in such a short period of time, killing nearly 50 million people worldwide [1] . Unlike the epidemiological profiles of most influenza infections, young adults aged 18-35 yrs old had the highest mortality rate, so much so that the average life expectancy during those years was lowered by 10 years [2] . In 1918, severe destruction of lung tissue observed by pathologists at autopsy was unlike that typically seen in cases of pneumonia [3] and histopathological analysis of lung tissue showed severe tissue consolidation with unique destruction of the lung architecture [3, 4] . Human infections with highly pathogenic avian influenza (HPAI) strains of subtype H5N1 since the first outbreak in 1997 have also been particularly severe for children and young adults [5] [6] [7] . Assessing pulmonary infiltrates in response to influenza H5N1 virus infection has been difficult due to the lack of autopsy material. The basis for the high morbidity and mortality associated with the 1918 virus and recent H5N1 viruses remains inconclusive based on viral genetic analysis alone and accounts of patient lung pathology provide only qualitative information about the host factors contributing to disease [4, 8, 9] . Great concern about a pandemic caused by a novel avian H5 subtype virus warrants comparative studies to better understand the cellular pathology caused by a pandemic virus and potentially pandemic viruses. Identification and quantification of the inflammatory cell types associated with highly pathogenic respiratory infections represent prospective targets for modulation of host innate immune responses. Recent studies using animal models to investigate the mechanism(s) of severe influenza virulence have implicated the innate immune system in complicating lung tissue recovery [10] [11] [12] [13] . Mouse models of highly pathogenic (HP) H5N1 [14] [15] [16] [17] [18] and 1918 [19, 20] influenza virus infection confirm histological observations of severe lung pathology in human patients, however, the types of immune cells present during the peak of lung pathology have not been fully elucidated. Excessive immune cell infiltration during an acute lung injury may impair tissue restoration directly by interfering with gas exchange, or indirectly through the release of soluble immune mediators. In the present study, we determined key immune cellular components in the murine lung following infection with matched H5N1 and H1N1 virus pairs that represent high and low virulence infections of each influenza subtype as previously determined in the mouse model [18, 21] . The two H5N1 viruses used in this study (A/Thailand/16/2004 and A/Thailand/SP/83/2004) were isolated in 2004 from fatal human cases in Thailand but have a differential pathogenic outcome in mice, specifically a low and high mouse lethal does 50 (LD 50 = 1.7 and 5.6 log 10 PFU respectively) [18] . For relevant comparison, we also used a contemporary (non lethal) seasonal H1N1 human isolate from 1991 (A/TX/36/91) and the reconstructed 1918 pandemic virus [21] . A detailed flow cytometry evaluation of lung cells demonstrated that macrophages and neutrophils are the prominent cell types associated with and potentially mediating the severe lung pathology following infection with the highly virulent H5N1 and 1918 viruses. Moreover, inoculation of macrophages and dendritic cells with the HP viruses in vitro or ex vivo reveals that some innate immune cells can themselves serve as targets of viral infection. Highly pathogenic H1N1 and H5N1 viruses exhibit early and sustained replication in murine lung tissue following intranasal infection Female BALB/c mice were infected intranasally with either highly pathogenic (HP) or low-pathogenic (LP) influenza viruses ( Table 1 , Methods) based on known LD 50 's and phenotypes of disease in mouse [21] and ferret [18, 22] models. As shown in Figure 1 Increased cellularity in lungs of mice infected with highly pathogenic H1N1 and H5N1 influenza viruses Whole lungs collected (without perfusion, therefore including the bronchoalveolar lavage contents (BAL)) from both 1918 and Thai/16 virus-infected mice showed an increase in overall tissue cellularity as early as 3 days p.i. (Figure 2A ). By day 3 p.i., lungs infected with the HP influenza viruses (containing between 4.3-5.1610 7 cells) had nearly twice as many total lung cells than were measured in the HP infection groups just 24 hours previously at day 2 p.i. (2.1-2.4610 7 cells). Significant differences (*p,0.05) were observed between HP and LP infection groups in total lung cell number at every time after day 2 p.i. Total lung cell numbers doubled in HP infection groups between days 5 and 7 p.i. and by day 7 p.i., when viral titers were high for all four viruses, total cell numbers in the lungs of mice infected with either HP H1N1 or H5N1 viruses were 6-fold higher than those in PBS-inoculated mice, and at least 3-fold higher than those found in LP virusinfected lungs (Figures 1 and 2A) . On day 7 p.i., there were as many as 1.3610 8 cells in HP-infected lungs compared with 4.0-8.0610 7 cells in LP-infected and 1.8610 7 cells in PBS-inoculated lungs. To quantify the immune cell sub-populations responding to viral infection, we next determined the total cell numbers of specific inflammatory cell populations in the infected lungs using flow cytometry ( Figure 2B ). Compared with PBS-inoculated Patients who succumbed to influenza during the 1918 pandemic had severe lung pathology marked by extensive inflammatory infiltrate, indicating a robust immune response in the lung. Similar findings have been reported from H5N1-infected patients, raising the question as to why people expire in the presence of a strong immune response. We addressed this question by characterizing the immune cell populations in the mouse lung following infection with the 1918 pandemic virus and two H5N1 viruses isolated from fatal cases. Our data shows excessive immune cell infiltration in the lungs contributing to severe consolidation and tissue architecture destruction in mice infected with highly pathogenic (HP) influenza viruses, supporting the histopathological observations of lung tissue from 1918 and H5N1 fatalities. We found that certain cells of the innate immune system, specifically macrophages and neutrophils, increase significantly into the mouse lung shortly following HP virus infection. Interestingly, lung macrophages and dendritic cells were shown to be susceptible to 1918 and H5N1 virus infection in vitro or ex vivo, suggesting a possible mechanism of immunopathogenesis. Identification of the precise inflammatory cells associated with lung inflammation will be important for the development of treatments that could potentially enhance or modulate host innate immune responses. [18, 21, 22] . Generated by reverse genetics [21] . doi:10.1371/journal.ppat.1000115.t001 animals, mice infected with each of the viruses exhibited an increase in the numbers of macrophages (CD11b + , CD11c 2 , Ly6G/c 2 , CD4 2 , CD8 2 ) beginning 3 days p.i. and continuing through day 9 p.i. Strikingly, there were significantly (*p,0.05) more macrophages in HP virus-infected lungs than in LP virusinfected lungs from days 2 through 9 p.i. (Figure 2B , a). As early as 2 days p.i, there was 1-2 million more macrophages in the lungs of HP-infected lungs compared to LP infected lungs. 1918 and Thai/ 16 virus-infected lungs had twice as many and nearly 4 times as many macrophages compared to LP viruses at days 3 and 5, respectively. Numbers of lung macrophages peaked in 1918 and Thai/16 infected mice (1.5 and 1.2610 7 cells, respectively) at day 7 p.i. before waning as demonstrated at day 9 p.i. An increase of at least twice as many neutrophils (CD11b + , CD11c 2 , Ly6G/c + , CD4 2 , CD8-) was observed as early as 1 day p.i. in all infection groups, compared with neutrophils numbers in PBS-inoculated mice. On day 2 p.i. there was on average (n = 3 mice) over four hundred thousand more neutrophils in the lungs of 1918 virus inoculated mice than TX/91 inoculated mice. Significant differences in neutrophil populations (*p,0.05) between the HP and the LP virus groups emerged on day 3 p.i. and were sustained at each subsequent time point measured ( Figure 2B , b). On day 3 p.i., more than twice as many neutrophils were found in HP compared to LP infected lungs. Between days 3 and 7 p.i., lungs infected with HP viruses displayed a three fold increase in neutrophil numbers. At the peak of neutrophil infiltration on day 7 p.i, there was 4-8 million more neutrophils in the lungs of HP-infected animals compared to LP virus-infected mice. Although numbers of dendritic cells (CD11b 2 , CD11c + , Ly6G/c 2 , CD4 2 , CD8 2 ) ( Figure 2B , c) and CD4 + T (CD11b 2 , CD11c + , Ly6G/c 2 , CD8 2 ) and CD8 + T (CD11b 2 , CD11c + , Ly6G/c 2 , CD4 2 ) cells ( Figure 2B , d and e) in all groups of infected mice increased slightly compared with numbers in PBSinoculated animals during the course of infection, no significant differences were found between HP and LP infection groups, suggesting that these cell populations are not major contributors to the histopathological consolidation observed in lungs during HP H5N1 virus infections in mice [16] . Macrophages and neutrophils account for the highest percentage of the total lung leukocytes following infection with highly pathogenic influenza viruses We next determined specific immune cell sub-populations given by percent of the total lung leukocyte population in order to reveal differential population dynamics amongst the immune cell populations measured in these studies (Table 2) . Strikingly, macrophage populations by day 3 pi, represented 24% and 24.4% (nearly one-quarter) of the total gated lung leukocytes of mice infected with 1918 and Thai/16 (HP) viruses, respectively, which was significantly higher than the frequency of macrophages detected in TX/91 and SP/83 virus-infected mice from days 2 though 9 p.i. (^p,0.005). Neutrophil populations were elevated in all infection groups compared to mock levels beginning 1 day p.i ( Figure 2 , panel B) however percentages of neutrophils in HP virus infected mice were significantly higher (* p,0.05) than those in LP virus-infected mice beginning day 2 p.i. and levels remained elevated at each subsequent time point measured ( Table 2 ). In contrast, dendritic cell populations as percent of the total leukocyte population declined over the course of infection with HP viruses while percentages increased in LP infections, peaking at days 3 and 5 p.i. (Table 2 ). Significant differences in percent dendritic cells between HP and LP infection groups were observed day 3 p.i. and at all other subsequent time points (* p,0.05). Percentages of CD4 + T cells in the lungs decreased in all infection groups but no significant differences were observed in CD4 + or CD8 + T cell population dynamics between HP and LP infection groups. Together, these results indicate that macrophages and neutrophils are responsible for the majority increase in total lung cell numbers following infection with the 1918 and HP H5N1 influenza viruses. To better understand potential influences on immune cell population dynamics in the lung following HP influenza virus infection as demonstrated in Figure 2B and Table 2 , we analyzed 17 chemokines and cytokines in the lungs of mice on days 1 and 4 p.i and report data here on 6 of the analytes that revealed significant differences among the viruses tested ( Figure 3 ). These earlier time points were chosen in an effort to understand the temporal relationship with lung immune cell infiltration in 1918 and Thai/16 infections ( Figure 2 ) and day 4 p.i. is a time when significant differences were observed in lung virus titers between HP and LP infection groups ( Figure 1 ). As shown in Figure 3 , on day 1 p.i. TX/91 infected lungs exhibited higher titers of IL-1a, IFN-c, and KC compared to 1918 virus-infected lungs though levels of MIP-1a, MCP-1, and IL-6 were higher in 1918 virus infected lungs. Cytokine levels were similar between the H5N1 viruses at day 1 p.i.. In contrast, lung tissue cytokine and chemokine levels at day 4 p.i. were higher among 1918 and Thai/ 16 virus-infected mice compared to those infected with the subtype-matched LP (TX/91 and SP/83) counterpart viruses for all cytokines and chemokines measured. Protein levels of the potent monocyte chemoattractant MCP-1 were 10-fold higher in 1918 and H5N1 virus-infected lungs than TX/91 virus-infected lungs, whereas levels of the MIP-1a chemokine were notably elevated in Thai/16 virus-infected lungs. KC (the mouse equivalent of human IL-8) [20, 23] levels were 10-fold higher in Thai/16 and 3-fold higher in 1918 virus infected lungs compared to TX/91 virus-infected lungs with SP/83 levels similar to 1918 infected lungs. The HP viruses were also potent inducers of IL-1a and IFNc. IL-6, (a generally pro-inflammatory cytokine) was also Due to the increased presence of macrophages in the lungs following HP influenza virus infection, replication of paired H1N1 and H5N1 viruses (Table 1 ) was assayed over time in primary human PBMC-derived macrophages and mouse lung macrophages to address whether these cells are specific targets of viral infection and can productively replicate HP viruses. Mouse lung macrophages were harvested from naive BALB/c mice and infected in vitro (MOI = 0.1, Figure 4A ) as described. While the HP 1918 and Thai/16 viruses exhibited a slight increase in log titer very early after inoculation (13-24 hrs p.i.), overall these viruses did not replicate efficiently compared to the growth kinetics of these viruses observed in lung epithelial cells [22, 24] . At titers and these differences were statistically significant (^p,0.05) at 72 hrs p.i.. The lack of prolific replication of these four viruses in primary mouse lung macrophages was confirmed further when a higher MOI (1.0) was utilized (data not shown). Human macrophages also supported replication of all four viruses ( Figure 4B , MOI = 0.1). At 48 hrs p.i. 1918 and Thai/16 infected macrophages exhibited 180 times higher virus titers than TX/91 and SP/83 infected cultures. Interestingly, 1918 virus-infected macrophages exhibited a higher baseline titer soon following infection that was found to be statistically significant when compared to the other infection groups ( { p,0.05). In summary, while human macrophages are a target of viral replication and support replication well, mouse lung macrophages support low levels of 1918 and Thai/16 virus production early following infection. Additional experimentation with primary human macrophages revealed that levels of pro-inflammatory cytokines were higher for H5N1-infected cells than either 1918 or TX/91 virus infected cells (48 hrs p.i., MOI = 0.1, Figure 5 ). Significant differences in cytokine levels were observed between H5N1 and H1N1 virus Figure 4B ). Thai/16 infected cultures elicited at least a 2 fold greater cytokine response than the SP/83 virus infected cultures in every cytokine measured except IL-8 and MCP-1 chemokines where Thai/16 levels were only slightly higher than SP/83 levels. Because of their important role in antigen presentation to T cells, we also assessed the ability of primary dendritic cells to productively replicate pandemic H1N1 and human H5N1 viruses ( Figure 6A To determine if innate immune cells are being productively infected in vivo, macrophages and dendritic cells were purified from lungs of infected mice and cultured for infectious virus. Lungs from infected (3 days p.i.) mice were harvested and ex vivo cultures containing either lung macrophages or dendritic cells were sampled for infectious virus over a 65 hour time period. While the seasonal influenza isolate, TX/91 virus was not produced from either macrophages (CD11b+) or dendritic (CD11c+) cells, the 1918 pandemic virus as well as the two human H5N1 isolates were released into the culture supernatant (Figure 7) , indicating that these cells are being productively infected in the mouse lung. In CD11b+ (macrophages) cell cultures ( Figure 7A Figure 7B ). These data further demonstrate that mouse lung macrophages and dendritic cells are susceptible to highly pathogenic influenza virus infection in the lung tissue. Lung consolidation has been described as a pathological feature of severe influenza virus infection caused by the 1918 pandemic virus and H5N1 viruses in humans [9, 25, 26] as well as in animal models [11, 12, 18] . Using a detailed flow cytometry evaluation, the current study set out to characterize differences in the cellular innate immune response in the mouse lung following highly pathogenic (HP) or low pathogenicity (LP) influenza virus infections. Lungs from mice infected with the HP 1918 H1N1 virus and a recent H5N1 human isolate (A/Thailand/16/04 (Thai/16)) exhibited a significant increase in cellularity in comparison to the LP seasonal H1N1 isolate, A/Texas/36/91 (TX/91). Significant differences in titer between HP and LP virusinfected mice were observed as early as day 1 post-inoculation (p.i.), likely forecasting the dramatic increase in immune cell infiltration in the lungs of these virus-infected mice. Interestingly at day 7 p.i. when peak lung cellularity was observed in HP virus infection groups, differences in virus titers between paired subtype viruses (1918 compared to TX/91 and Thai/16 compared to SP/ 83) were minimal and were limited to a maximum difference of 1 log (Figure 1 ), indicating a failure by the immune system to clear the viral infection. Lung cellularity was further investigated by characterizing the comprising immune cell sub-populations. We observed a significant increase in macrophages and neutrophils early following infection with the 1918 and Thai/16 viruses, and their sustained presence in the lung tissue mark a distinction between HP and LP influenza virus infection. These data show that virus replication in the lungs of HP influenza infections are sustained at high levels the first week of infection regardless of the high numbers of immune cells present in the tissue. Although the current study did not determine whether these cells are playing an antiviral role against the HP viruses, it has been previously shown that neutrophils and macrophages assist in the clearance of influenza virus early during infection; these cells appear to be capable of only partially reducing the virus load in the lung despite their presence at high numbers [20] . We also observed a decrease in the percentage of lung-associated dendritic cells and T cell lymphocytes during HP influenza virus infection. A decrease in the number of circulating lymphocyte populations has also been previously observed in the peripheral blood of humans and mice infected with H5N1 viruses [5, 9, 10, 15, 20, 27] . The precise mechanism of leukocyte depletion during H5N1 infections is not well understood, but evidence of apoptosis in the spleen and lungs of HP H5N1-infected mice detected in situ suggests a mechanism for cell loss [16] . Influenza virus growth in the respiratory epithelium and the subsequent release of chemotactic proteins from those cells may encourage the increased presence of macrophages and neutrophils [28, 29] . Macrophages and neutrophils can secrete chemokines and cytokines that can act in an autocrine fashion which in turn can promote the increased migration of those cells and other leukocytes into the lung tissue [30] . Elevated levels of certain chemokines and cytokines have been associated with high viral load and severe disease in H5N1 virus infected patients [26, 31] . These studies show that infection with HP 1918 and Thai/16 H5N1 viruses result in elevated amounts of pro-inflammatory chemokines and cytokines in the lungs of mice day 4 post-infection compared with TX/91 and SP/83 infected mice, a time point that correlates with rising but significantly different lung virus titers between infection groups. Elevated levels of the chemokines MCP-1 [32] and MIP1-a were observed among H5N1 and 1918 virus infected mouse lungs. Although, MIP-1a does not appear to be critical for virus replication and spread in the mouse model [10] , this chemokine exhibits a variety of pro-inflammatory activities including macrophage and neutrophil recruitment and has been associated with fatal outcomes in human H5N1 virus infections [31] . Lungs infected with the 1918 pandemic and Thai/16 H5N1 viruses also exhibited significantly higher levels of IFN-c on day 4 p.i compared to their subtype-paired LP virus counterparts. IFN-c is known to mediate the increased production of nitric oxide [33] which can subsequently result in the recruitment of more neutrophils and macrophages. Higher levels of IL-6 were measured in 1918 and Thai/16 virus infected lungs, supporting observations obtained with the 1997 H5N1 viruses [10] and has been correlated with systemic illness symptoms and fever in experimental human TX/91 infections [34] . By directly measuring cytokine protein levels, these data provide confirming evidence of a heightened lung cytokine response to 1918 and H5N1 infection in mice [13] . Interestingly, while we reveal marked similarities between the HP 1918 and Thai/16 viruses in overall lung cellularity, virus growth and patterns of immune cell sub-population dynamics over time, Thai/16 virus infection consistently resulted in higher levels of chemokines and cytokines both in mouse lungs and human macrophages. Although it has been shown recently by two independent research groups that the lack of key cytokines (through the use of single cytokine gene knockout mice) had no effect on the overall disease outcome or virus replication among H5N1 virus-inoculated mice [10, 35] , the present results along with data from others continue to indicate that pro-inflammatory cytokines correlate with disease outcome [36] . It is thought that these immune mediators do not act singularly in vivo and it will be critical to reveal these concerted interactions (both locally in the lung and systemically) to further our understanding of the pathogenesis of H5N1 infection in animal models and in human patients. Macrophages and dendritic cells play a fundamental role in the lung at all stages of influenza virus infection [20, 37, 38] . We have provided evidence regarding the higher replication efficiency of HP influenza viruses in primary human macrophages and dendritic cells, a property that has also been demonstrated previously in other primary cells [24, 39] . Mouse macrophages were also susceptible to virus infection in vitro, however they did not support productive replication to the level observed in primary human monocyte-derived macrophages ( Figure 4B ) or lung epithelial cells [21, 24] . However, the cytopathic effects (estimated by visual examination of monolayers) among HP virus-infected mouse and human macrophages was observed in a shorter period of time compared to LP virus infected cells. Interestingly, the 1918 virus exhibited higher baseline titers in human and mouse macrophages as early as 2 hrs p.i. compared to the TX/91 H1N1 virus or the H5N1 viruses, indicating a curious property of this pandemic virus. While the importance of the binding properties of the HA molecule has been demonstrated elsewhere extensively [22, 40, 41] , further resolution of this interesting finding and its immunological importance deserves further experimentation. The role of other cell surface molecules on innate immune cells such as the mannose receptor in HP influenza infection should be investigated [42, 43] . The higher viral replication of HP viruses in dendritic cells also correlated with the severe pulmonary disease observed in mice. We also demonstrated definitively that these cells are targets of infection ex vivo by highly pathogenic influenza viruses like the 1918 pandemic virus and recent H5N1 isolates (Figure 7) . Thus, it appears that macrophages and dendritic cells may contribute to the pathogenesis of HP virus infection due to their susceptibility to influenza virus infection. An inability to mount an adaptive immune response due to direct infection of important innate immune cells such as macrophages and dendritic cells may be a critical difference in host outcome during influenza virus infection [44] . This coupled with the phenomenon of T cell depletion in infected mice [17] and human patients [9, 31] may allow for uncontrolled viral replication. While reducing viral load through anti-viral intervention remains the best treatment option for H5N1 patients, therapies that moderate immunopathology may help to reduce the high case fatality rate currently associated with virus infection [45] . (Table 1) . Low pathogenicity in this manuscript refers to the non-lethal phenotype of the seasonal H1N1 TX/91 virus and the low virulence SP/83 H5N1 isolate [18, 21] . The Thai/16 and SP/83 viruses differ from each other in 13 amino acids in 7 proteins and this sequence comparison has been published previously [18] . The A/Texas/36/91 and H5N1 viruses were grown in 10 day old embryonated hen's eggs and the 1918 viruses grown in MDCK cells. All virus stocks were titered by plaque assay on MDCK cells prior to mouse infections. Human peripheral blood monocytes (PBMC's) were obtained by Histopaque (Sigma-Aldrich, St. Louis, MO) density gradient centrifugation of whole blood donated by healthy donors aged 20-40 yrs old without history of influenza vaccination in the past year. Whole blood was obtained through an approved protocol by both the Emory University Institutional Review Board (IRB) and CDC IRB (Emory University Hospital Blood Bank is an FDA-accredited Blood Bank). Human monocytes were obtained by negative selection column enrichment (Miltenyi Biotech, Auburn, CA) yielding approximately 90% CD14+ purity as determined by FACS analysis. For development of macrophages, monocytes were cultured at 37uC in 6 well plates in Macrophage SFM media (Gibco, Grand Island, NY) with 20% heat inactivated autologous serum for 7 days in the presence of GM-CSF (1000 U) before infection [46] . Human macrophages cultured in this manner typically displayed classical morphology with the phenotype: CD11b low , CD11c low , HLA-DR low , CD14 low , CD40 low , CD80 low , CD83 low , CD86-( Figure 4C and D). For the development of dendritic cells (DC's), monocytes were grown in RPMI (Gibco) with 20% heat inactivated autologous serum for 10 days in the presence of IL-4 (1000U) as previously described [46] . Dendritic cells developed in this manner typically displayed a classical morphology with the presence of dendritic processes with the phenotype: CD11b low , CD11c high , HLA-DR high , CD14 low , CD40 high , CD80 high , CD83 high , CD86 high ( Figure 6B) . To obtain primary mouse lung macrophages and dendritic cells, lungs from naïve mice were removed and tissue disrupted as described above through the use of collagenase digestion and cell suspensions prepared. Macrophages (CD11b+) and dendritic cells (CD11c+) were extracted from contaminating cells by selection on magnetic columns (Miltenyi Biotech, Auburn, CA). Cells were washed twice with media containing 20% FCS (Macrophage SFM for macrophages (Gibco) or RPMI (Gibco) for dendritic cells) and cultured for 24 hours before in vitro infection. Primary mouse lung macrophages typically displayed the phenotype: CD11b high , CD11c-, MHCII high , CD40 high , CD80 high , CD83 high ( Figure 4A and B). Primary mouse lung dendritic cells typically displayed typical morphology with dendritic extensions and the phenotype: CD11b-, CD11c high , MHCII high , CD40 high , CD80 high , CD83 high ( Figure 6A ). Primary human and mouse cells were washed 36 with serum free growth media and infected for 1 hour with viruses (Table 1) at a multiplicity of infection (MOI) of 0.1. Following infection, cells were washed 36 with serum free growth media and 1 ml SFM media, containing 1 mg/ml of TPCK-treated trypsin (Sigma-Aldrich), was placed into the wells. Virus growth was measured over time in triplicate wells for each experiment and titered in duplicate by standard plaque assay on MDCK cells. All macrophage and dendritic cell data reflects at least three independent experiments (Figures 4 and 6) . Cytokine levels produced from infected human macrophages (MOI = 0.1) were quantitated 48 hrs p.i. by BioPlex assay (Figure 5 ). Escherichia coli lipopolysaccharide (LPS, 100 ng, Sigma-Aldrich) and Poly I/C (100 ng, Sigma-Aldrich) were used as positive control stimulants. All animal research was conducted under the guidance of CDC's Institutional Animal Care and Use Committee and in an Association for Assessment and Accreditation of Laboratory Animal Care International-accredited facility. 8-10 week old female BALB/c mice (Harlan, Indianapolis, IN) were anesthetized with Avertin [20] (Sigma-Aldrich) and infected intranasally (i.n.) with 50 ml of 10 2 PFU of influenza viruses prepared in phosphate buffered saline (PBS). Avertin was chosen as the anesthetic because it provides consistent mouse infections with the viruses used in these studies. Using the sublethal (10 2 PFU) inoculum, 1918 and Thai/16 virus infected mice survive a prolonged disease course allowing for the measurement of the influx of inflammatory cells into the lung tissue during a full course (,7-9 days) of influenza virus infection. At indicated times post-infection (n = 3 mice per virus group) mice were euthanatized and exsanguinated. Lungs were removed from individual mice without PBS perfusion and included total lung cell counts included cells located in the bronchoalveolar airways. Perfusion was not possible in many cases of HP influenza virus infection due to the presence of microvascular hemorrhage. We obtained similar results when performing these lung cell quantitation assays with or without lung perfusion in LP virus infected mice and therefore did not introduce this variable in our high containment laboratory. Whole lung cell suspensions were prepared in Dulbecco's minimal essential media (DMEM) with 20% fetal calf serum following collagenase-DNase treatment and manual disruption [47] . Red blood cells were removed by lysis buffer treatment (Sigma-Aldrich). Total viable lung cell number was determined for each mouse by trypan blue exclusion on a hemocytometer. For the ex vivo experiment, macrophages and dendritic cells were isolated from the lungs of infected BALB/c mice. Two or three mice were infected i.n. with 10 2 PFU of each of the four viruses described in this study (Table 1) . Three days post-inoculation, lungs were removed without perfusion and cell suspensions were prepared as described above. Lungs were pooled from mice in each virus infection group. Macrophages and dendritic cells were isolated by positive selection on CD11b+ or CD11c+ MACS columns. Columns containing bound immune cells were washed extensively (5x) and CD11b+ or CD11c+ cells were eluted off the magnetic columns and cultured in 6-well plates in 5 ml of RPMI containing 5% BSA. Supernatants were collected at the indicated times and virus content was determined in a standard plaque assay on MDCK cells. Lung cell suspensions were incubated with anti-Fc block (antimouse CD16/CD32) to reduce non-specific antibody binding for 10 min. prior to staining for 1 hr with fluorophore-conjugated antibodies (BD Biosciences, San Diego, CA) specific for immune cell populations according to standard protocols [48] and included: CD11b-PE (pan-macrophage), CD11c-APC (pan-dendritic cell), Ly6G/C-FITC (neutrophil), CD4-PE and CD8-APC T cell markers (Table 2, Figure 2 ). Cells were washed twice with PBS and fixed overnight at 4uC with 2% paraformaldehyde. Samples were safety tested for infectious virus and removed from the BSL3+ laboratory. Flow cytometry was performed on a FACSAria flow cytometer (BD Biosciences). To further characterize primary mouse and human macrophage and dendritic cells we utilized the following fluorescently conjugated antibodies for flow cytometric analysis: MHC II (I-A/I-E)-PE, HLA-DR-PE, CD14-FITC, CD40-FITC, CD80-FITC, CD83-APC, and CD86-APC (BD Biosciences) (Figures 4 and 6 ). At various times post-infection (n = 3 mice per virus group) lungs were removed and stored at 270uC until virus and cytokine levels could be quantified. Lungs were homogenized individually in 1 ml PBS. Virus was titered from clarified lung homogenates by standard plaque assay on MDCK cells in duplicate and titers are reported as plaque forming units per ml PBS (PFU/ml, Figure 1 ). Cytokine protein levels were measured (day 4 post-infection ( p.i.)) by the Bioplex Protein Array system [49] (Bio-Rad, Hercules, CA) using beads specific for mouse G-CSF, IL-1a , IL-1b, IL-3, IL-6, IL-9, IL-12 (p40), IL-12 (p70), IL-13, Eotaxin, TNFa, RANTES, KC, MIP1-a, MIP-1b, MCP-1, and IFN-c. Cytokine protein levels were measured according to the manufacturers instructions by fluorescently conjugated monoclonal antibodies in duplicate against a standard curve (Figures 3 and 5 ). Statistical significance of differences between experimental groups was determined through the use of the unpaired, nonparametric Student's t test. Values of p,0.05 were considered significant.

Patterns of Positive Selection in Six Mammalian Genomes

Genome-wide scans for positively selected genes (PSGs) in mammals have provided insight into the dynamics of genome evolution, the genetic basis of differences between species, and the functions of individual genes. However, previous scans have been limited in power and accuracy owing to small numbers of available genomes. Here we present the most comprehensive examination of mammalian PSGs to date, using the six high-coverage genome assemblies now available for eutherian mammals. The increased phylogenetic depth of this dataset results in substantially improved statistical power, and permits several new lineage- and clade-specific tests to be applied. Of ∼16,500 human genes with high-confidence orthologs in at least two other species, 400 genes showed significant evidence of positive selection (FDR<0.05), according to a standard likelihood ratio test. An additional 144 genes showed evidence of positive selection on particular lineages or clades. As in previous studies, the identified PSGs were enriched for roles in defense/immunity, chemosensory perception, and reproduction, but enrichments were also evident for more specific functions, such as complement-mediated immunity and taste perception. Several pathways were strongly enriched for PSGs, suggesting possible co-evolution of interacting genes. A novel Bayesian analysis of the possible “selection histories” of each gene indicated that most PSGs have switched multiple times between positive selection and nonselection, suggesting that positive selection is often episodic. A detailed analysis of Affymetrix exon array data indicated that PSGs are expressed at significantly lower levels, and in a more tissue-specific manner, than non-PSGs. Genes that are specifically expressed in the spleen, testes, liver, and breast are significantly enriched for PSGs, but no evidence was found for an enrichment for PSGs among brain-specific genes. This study provides additional evidence for widespread positive selection in mammalian evolution and new genome-wide insights into the functional implications of positive selection.

Positive darwinian selection is an important source of evolutionary innovation and a major force behind the divergence of species. The Neutralist-Selectionist debate of the past 30 years has gradually given way to a general consensus that both neutral drift and positive selection play major roles in evolutionary change. Interest has therefore shifted to questions of which genes positive selection has affected, how strong was the effect, when did it occur, and what were its functional consequences. Heightening interest in these questions is a growing appreciation that methods for detecting positive selection can also be valuable tools for gaining insight into gene function [1] . Consequently, a wide variety of methods for detecting positively selected genes (PSGs) have been introduced, including comparative or phylogenetic methods, which make use of patterns of substitutions between species, and population genetic methods, which primarily rely on patterns of intraspecies polymorphism [2, 3] . Using these techniques, strong evidence of positive selection has been found for various genes in various organisms, including many genes involved in sensory perception, immunity, host-pathogen interactions, and reproduction (reviewed in [1, 3] ). Phylogenetic and population genetic methods for detecting positive selection serve as complementary tools for functional and evolutionary analysis. These methods operate at different time scales, with phylogenetic methods being best suited for detecting selection that operates over relatively long periods in evolutionary history, and population genetic methods being best suited for detecting more recent selection. Population genetic methods can potentially detect selection operating at individual sites, through the effects of linkage with flanking alleles, while phylogenetic methods generally require multiple sites to have been affected in a sequence of interest. At the same time, decay of linkage disequilibrium at longer evolutionary time scales can allow phylogenetic methods to more accurately pinpoint the specific locations of functionally important substitutions. In some cases, phylogenetic methods also allow such substitutions to be mapped to particular branches of a phylogenetic tree, thereby providing useful insights about the evolutionary histories of the sequences in question. With the availability of multiple complete genome sequences, it has become possible to apply phylogenetic methods for the detection of positive selection at a genome-wide scale. Within mammals, several genome-wide scans for positive selection on protein-coding genes have been conducted, using both phylogenetic [4, 5, 6, 7, 8, 9] and population genetic [10, 11, 12, 13, 14, 15] methods (reviewed in [16] ). These analyses have identified many new genes showing strong evidence of positive selection and have revealed striking differences in the prevalence of positive selection on different lineages and among different classes of genes. For example, it has been reported that PSGs are enriched for roles in sensory perception, immunity and defense, tumor suppression, apoptosis, and spermatogenesis [4, 5] ; that PSGs are associated with known Mendelian disorders [4] ; that PSGs often coincide with segmental duplications [8] ; and that more genes have undergone positive selection in chimpanzee evolution than in human evolution [9] . Genome-wide scans for PSGs have also helped to stimulate interest in detecting positive selection on noncoding sequences and on gene expression [17, 18, 19, 20] . Nevertheless, much remains to be learned about positive selection in mammalian genomes, even within protein-coding regions. The most comprehensive scans for PSGs so far [4, 5, 7, 8, 9] have been based on at most three genome sequences-typically the highly similar human, chimpanzee, and/or rhesus macaque genomes (.97% average identity in orthologous coding regions [8] ). As a result, the power for detection of PSGs has been relatively weak [5, 8] . In addition, in several of these studies, at least one genome was of draft quality, which reduced the number of genes that could be examined and required additional care in avoiding false positive predictions. Here we present a phylogenetic analysis of positive selection in the six eutherian mammalian genomes for which high-coverage, high-quality sequence assemblies are now available: the human [21] , chimpanzee [6] , macaque [8] , mouse [22] , rat [23] , and dog [24] genomes. The phylogenetic depth of this data set helps considerably in addressing the problem of weak power. Indeed, these genomes have a near-optimal degree of divergence for PSG detection, being distant enough to produce a strong phylogenetic signal, yet close enough that gene structures are well conserved, alignment is fairly straightforward, and synonymous substitutions are not saturated (e.g., [25] ). In addition, our data set for the first time allows positive selection of mammalian genes to be examined genome-wide on a nontrivial phylogeny, so that insight can be gained into the particular ''selection histories'' of individual genes-that is, the branches of the phylogeny on which they experienced positive selection. In our analysis, we employ models of codon substitution that account for variation of selective pressure over branches on the tree and across sites in a sequence, which can capture signatures of molecular adaptation that affect small numbers of sites [26, 27] . Using a series of likelihood ratio tests (LRTs) based on these models, we identify more than four hundred genes that show strong signatures of positive selection during mammalian evolution. Our detailed analysis of the functional roles, selection histories, and expression patterns of these genes follows. Using the latest human, chimpanzee, macaque, mouse, rat, and dog genome assemblies, we identified 17,489 human genes with high-confidence orthologs in at least two of the remaining five species. These ortholog sets (human genes and non-human orthologs) were identified by an automatic pipeline that made use of syntenic whole-genome alignments, sequence quality scores, and other data (see Methods). Briefly, the pipeline began with 21,115 human genes drawn from the RefSeq [28] , UCSC Known Genes [29] , and VEGA [30] gene sets. These genes were mapped to the other genomes via syntenic pairwise alignments, then passed through a series of rigorous filters to ensure correct mapping, high sequence quality, and only minimal changes between species in gene structure. This approach exploits the fact that gene structures are generally well-conserved between mammalian species [22] and avoids any dependency on the non-human gene annotations, which-with the exception of mouse-are significantly less accurate and complete than those for human. Because low-quality sequence can produce a spurious signal for positive selection (e.g., [8] ), all bases with low quality scores (Phred quality ,20) were masked out for subsequent analyses. Masking (or truncation at the 59 or 39 end) was also used to exclude regions of genes in which minor differences in gene structure were apparent. Genes that showed signs of substantial disruptions to their exon-intron structures or open reading frames in one or more species (perhaps indicating pseudogenization) were masked out completely in those species. All masked bases were treated as missing data in the subsequent analysis of positive selection. This masking approach allowed the number of genes to be maximized while ensuring that the analyzed alignments were of high quality (Table 1) . For this study, we chose to avoid recently duplicated gene families and to focus on 1:1 orthologs. This simplified the analysis, allowed for parameter sharing across genes (see Methods), and eliminated an important source of error by avoiding the need for a separate tree reconstruction for each gene family. (All ortholog sets were assumed to obey the species tree shown in Figure 1 ; because only an unrooted tree is needed, the topology is well accepted.) It was therefore necessary to discard any genes that showed evidence of recent duplication. This was accomplished in a pairwise fashion, by examining each human gene and orthologous non-human gene, and determining-based on BLAST matches to other genes and gene predictions in the same genome-whether either gene had a paralog that was more similar to it than the two orthologs Populations evolve as mutations arise in individual organisms and, through hereditary transmission, gradually become ''fixed'' (shared by all individuals) in the population. Many mutations have essentially no effect on organismal fitness and can become fixed only by the stochastic process of neutral drift. However, some mutations produce a selective advantage that boosts their chances of reaching fixation. Genes in which new mutations tend to be beneficial, rather than neutral or deleterious, tend to evolve rapidly and are said to be under positive selection. Genes involved in immunity and defense are a well-known example; rapid evolution in these genes presumably occurs because new mutations help organisms to prevail in evolutionary ''arms races'' with pathogens. Many mammalian genes show evidence of positive selection, but open questions remain about the overall impact of positive selection in mammals. For example, which key differences between species can be attributed to positive selection? How have patterns of selection changed across the mammalian phylogeny? What are the effects of population size and gene expression patterns on positive selection? Here we attempt to shed light on these and other questions in a comprehensive study of ,16,500 genes in six mammalian genomes. were to each other (see Methods). Requiring that each human gene had a high-confidence 1:1 ortholog in at least two other species reduced the total number of ortholog sets to 16,529. These sets contain a human gene and either five (42% of cases), four (28%), three (15%) or two (15%) non-human orthologs. We performed a series of nine different LRTs to identify genes under positive selection on particular branches or clades of interest in the six-species phylogeny. In particular, we tested for selection on any branch of the tree ( Figure 1A ); on the branch leading to, and on any branch within, the primate clade ( Figure 1B ,C); on the branch leading to, and on any branch within, the rodent clade ( Figure 1D ,E); and on each of the four individual branches within the primate clade ( Figure 1F -I). These LRTs were all based on widely used site or branch-site models of codon evolution [31, 26, 27] (see Methods). The test for all branches was applied to all 16,529 ortholog sets. For the branch-and clade-specific tests, ortholog sets were discarded if they did not contain adequate ingroup or out-group data for the test in question, which somewhat reduced the number of tests (Text S1, Table S1 ). The PSGs identified by each test ranged in number from only seven (the hominid branch) to 400 (the test for all branches; FDR,0.05 in all cases). As in previous studies, the numbers of genes identified by the tests for individual primate branches were small, primarily due to weak power caused by low levels of interspecies divergence. The inclusion of additional non-primate mammals does not appear to have improved the power of these tests substantially, but it does allow a distinction to be made between selection on the branches to the hominids and to macaque. The tests for selection on the branch to the primates and in the primate clade also yielded fairly small numbers of PSGs, but the tests for selection in, or on the branch to, the rodents identified somewhat (nearly three-fold) larger numbers. In general, even with the larger data set, our power to detect selection on individual lineages and clades is still fairly weak, and differences in numbers of identified PSGs almost certainly reflect differences in power more than differences in the prevalence of selection. Nevertheless, these LRTs together produced a fairly large set of high-confidence PSGs, permitting a more detailed and thorough functional analysis than has previously been possible in mammals (see below). The identified PSGs are significantly enriched for a large number of functional categories, according to the Gene Ontology (GO) [32] and Protein Analysis Through Evolutionary Relationships (PANTHER) databases (Tables 2, S2, and S3). If these overrepresented categories are clustered by the PSGs that are assigned to them, major groups corresponding to sensory perception, immunity, and defense emerge (Figure 2 ), in agreement with previous genome-wide scans [4, 5] . However, the increased power of our analysis allows biological processes and functions associated with positive selection to be identified at much finer resolution than in previous analyses, as discussed below. The increased power also seems to diminish the dependency of functional enrichments on the database or statistical methodology selected for the analysis. In particular, better agreement was observed between functional categories over-represented among the identified PSGs, as determined by Fisher's exact test (FET), and categories whose genes displayed significant shift toward smaller LRT P-values (whether or not they met the significance threshold for PSGs), as determined by the Mann-Whitney U (MWU) test (see Methods). Better agreement was also observed between analyses based on the GO and PANTHER databases (see Tables S2 and S3 ). The observed enrichments do not appear to be an artifact of differences between categories in gene length or alignment depth per gene (Text S1). In the discussion below, we focus on GO categories and nominal P -values based on the MWU test, as applied to P-values from the LRT for selection on any branch of the tree (except when otherwise indicated); full results are shown in Table 2 and Text S1. The PSGs are enriched for a wide variety of functions related to immunity and defense. Several over-represented categories describe activation in response to external or environmental stresses, such as from bacteria (P = 4.2610 28 ), viruses (P = 3.0610 28 ), wounding (P = 3.2610 28 ), and acute inflammation (P = 4.7610 211 ). In some cases, different categories reflect the same or very similar sets of genes (e.g., ''response to wounding'' and ''acute inflammatory response,'' or ''response to virus'' and ''response to bacterium''), while in others they reflect quite distinct gene sets (''response to wounding'' and ''response to virus'') ( Figure 2 ). Genes involved in both innate (P = 1.9610 29 ) and adaptive (P = 1.5610 25 ) immunity are over-represented, with many PSGs contributing to both classes. The conventional division of adaptive immunity into humoral (P = 1.6610 27 ) and cellular (P = 3.5610 27 ) responses is reflected in the enriched GO categories. Various mechanisms of immune response are represented, including previously identified categories for natural killer cell (P = 1.6610 28 ), B-cell (P = 4.8610 27 ), and T-cell (P = 1.2610 28 ) mediated immunity [5, 8] , and new categories such as cytokine/chemokine-mediated (7.6610 28 ) and complement-mediated immunity (P = 6.0610 26 ; see Table S3 ). Some of the enriched categories point to particular pathways with large numbers of PSGs. A striking example is the complement immunity system, a biochemical cascade responsible for the elimination of pathogens. This system consists of several small proteins found in the blood that cooperate to kill target cells by disrupting their plasma membranes. Of 28 genes associated with this pathway in KEGG [33] , nine are identified as PSGs (FDR,0.05), and five others have nominal P,0.05 ( Figure S1 ). Most of these PSGs are inhibitors (DAF, CFH, CFI) and receptors (C5AR1, CR2), but some are part of the membrane attack complex (C7, C9, C8A), which punctures cell membranes to initiate cell lysis. Many of these PSGs are known to interact with one another, suggesting possible co-evolution. Two of three biochemical pathways known to activate the complement system are also enriched for PSGs (the classical complement pathway [P = 6.1610 27 ] and the alternative complement pathway [P = 1.5610 26 ]), as is the coagulation cascade that interacts with the complement system (''blood clotting,'' MWU P = 2.2610 27 ; Table S3 ). Other pathways that contain multiple interacting PSGs include those for apoptosis, taste transduction, antigen processing and presentation, and cytokine-and chemokine-mediated signaling (e.g., Figures. S4, S5) . Several gene families of the immunoglobulin superfamily (''immunoglobulin mediated immune response,'' P = 1.1610 27 ) show particularly strong enrichments for PSGs. For example, five of the six SIGLEC genes included in our analysis are under positive selection (see [34] ). A detailed examination of one immunoglobulin gene for which structural information was available-a cellsurface receptor for hepatitis A and other viruses called HAVCR1 (LRT P = 6.9610 29 )-revealed several sites under positive selection in its N-terminal V-like immunoglobulin (IgV) domain. Three of these sites correspond to regions of the protein believed to play critical roles in binding to viruses or in regulating the immune function of the gene (Figure 3 ). In addition to its role in viral defense, HAVCR1 is a key player in the hygiene hypothesis explaining the increase in allergies and asthma [35] . It also interacts with IgA (CD79A; P = 5.4610 29 ), whose deficiency is associated with increased susceptibility to autoimmune and allergic diseases [36] . The hierarchical clustering of GO categories ( Figure 2 ) reveals an unexpected similarity between the sets of PSGs involved in fertilization and cytolysis, and some similarity of both sets with immune-related PSGs. This association of immunity, fertilization, and cytolysis is driven by a group of genes that participate in sperm-egg interaction, but also have immune-related functions and destroy pathogens by cytolysis. Interestingly, PSGs with roles in both reproduction and immunity are often also related to cancer, and it has been hypothesized that most cancer genes under positive selection have been subject to antagonistic co-evolution, with lineage-specific variations in dynamics and strength [5, 37] . Several PSGs identified here are associated with both FAS/p53 apoptosis and cancer (Da Fonseca et al., in prep.), such as the protein p53, which also regulates maternal reproduction [38] ; the cell adhesion gene ADAM2 (P = 2.9610 26 ), which is integral to fertilization [39] ; and the related genes ADAM15 (P = 5.4610 24 ) and ADAM29 (P = 3.4610 24 ), which are strong candidates for cancer evolution driven by sexual conflict. In addition, the testes development-related gene CCDC54 (P = 3.3610 24 ) is currently a target of cancer immunotherapy research [40] . A smaller and somewhat less diverse group of enriched categories is associated with sensory perception. Among the most inclusive categories of this type are ''sensory perception of chemical stimulus'' (24 PSGs; P = 4.3610 239 ) and ''G-protein coupled receptor protein signaling pathway'' (39 PSGs; P = 1.4610 27 ). Previously, enrichments for such categories have been attributed primarily to olfactory receptors [4, 5] . Indeed, 15 PSGs are labeled as having ''olfactory receptor activity'' (P = 6.9610 236 ). However, eight PSGs are involved in ''sensory perception of taste,'' including five taste receptors (P = 1.4610 210 ). Interestingly, several of these are bitter taste receptors. The sense of bitter taste is critical in allowing organisms to avoid toxic and harmful substances, and extensive gene expansion of bitter taste receptors is known to have occurred during mammalian evolution [41] , possibly driven by (or helping to drive) positive selection. Bitter taste receptors under positive selection include TAS2R1, TAS2R5, and a recently expanded cluster of genes at chr12p13 (TAS2R13, TAS2R14, TAS2R42, and TAS2R49). Another PSG, TAS1R2, is a receptor of sweet and umami taste, and the PSG RTP3 is a transmembrane protein that is involved in the transport of taste receptors and apparently influences their expression. The PSGs in the ''neurological processes'' category (P = 7.5610 27 ) are dominated by olfactory and taste receptors, but they also include other types of genes. For example, TMC2 (P = 1.1610 24 ) is expressed in the inner ear and is important for balance and hearing [42] . The acid-sensing ion channel gene ACCN4 (P = 1.0610 26 ) has been implicated in synaptic transmission, pain perception, and mechanoperception [43] . SLC6A5 (P = 3.0610 24 ) is associated with hyperekplexia, a neurological disorder characterized by an excessive startle response [44] . The neuromedin receptor NMUR (P = 6.1610 24 ) is involved in the mammalian circadian oscillator system [45, 46] . Finally, the neurotensin receptor NTSR1 (P = 8.1610 24 ) mediates hypotension, hyperglycemia, hypothermia, antinociception, and regulation of intestinal motility and secretion [47] . Similarly, the PSGs associated with diet include but are not limited to taste and olfactory receptors. For example, MGAM (P = 2.4610 28 ) is essential for the small intestinal digestion of starch, giving it a critical role in human metabolism, as starches of plant origin make up two-thirds of most human diets [48] (see also [49] ). MAN2B1 (P = 1610 26 ) is involved in the cleavage of the alpha form of mannose, a sugar monomer. Defects in this gene cause lysosomal alpha-mannosidosis, a lysosomal storage disease characterized by the accumulation of unbranched oligosaccharide chains [50] . TCN1 (P = 2.9610 231 ) is a major constituent of secondary granules in neutrophils and facilitates the transport of vitamin B12 into cells, which is important for the normal functioning of the brain and nervous system, and for the formation of blood [51] . In addition, several PSGs participate in ''steroid hormone metabolism'' (P = 8.3610 24 ) including genes that metabolize xenobiotics and drugs (e.g., SULT1C3, UGT2B7, and CYP2C8). Positive selection in these and other genes is likely to Figure 2 . Hierarchical clustering of 27 over-represented GO categories identified by the Mann-Whitney U test (''biological process'' group only), based on the genes assigned to each category. This dendrogram is derived from a dissimilarity matrix defined such that any two GO categories, X and Y, have dissimilarity 0 when all genes assigned to X are also assigned to Y (or vice-versa), and dissimilarity 1 when the sets of genes assigned to X and Y do not overlap. Specifically, X and Y have dissimilarity have been influenced by changes in food preferences during mammalian evolution. Few functional enrichments were evident for the PSGs identified by the branch-and clade-specific LRTs, primarily because these sets were quite small in size. However, the more powerful LRTs, such as those for the primate and rodent clades ( Figure 1C ,E), did produce significantly lower P-values for genes of certain functional categories than for others. Interestingly, these categories were dramatically different for the primate-and rodentclade LRTs, with nearly all of the primate categories relating to sensory perception, and nearly all of the rodent categories relating to immunity and defense (Table 4) . Indeed, the PSGs identified by the primate-clade test include several taste and olfactory receptors, as well as receptors for the sensation of pain (e.g., MRGPRE, NPFF2) and color vision (e.g., OPN1SW), and receptors involved in immunity (e.g., CCR1). The PSGs identified by the rodent-clade test include few such genes, but they include many genes involved in responses to wounding, inflammation, and stress, as well as genes involved in complement activation and innate immunity. Thus, we find little evidence that genes directly involved in brain development and function have (as a group) been driven by positive selection in primates, but many genes that provide sensory showing the interaction between two receptors that have been implicated in the regulation of HAVCR1's immune function. It is thought that clustering of receptors within the same cell surface might facilitate phosphorylation of the cytoplasmic tail, and that interaction between receptors from different cells might be a mechanism for B-T cell adhesion [91] . Predicted residue 39 falls within the region of these receptors, very near residue 37, which directly interacts with the opposite receptor (according to the available mouse structure). In addition, predicted residues 54 and 56 are adjacent to the virus-binding surface (shown in pink), as defined by a polymorphism in macaque [91] . Interestingly, the residue that falls between them (55) appears to be critical for virus-binding at the homologous loop in the CEA coronavirus receptor [91] . Residue 75 in the IgV domain also shows evidence of positive selection (PP.0.90, shown in orange) but its function is unknown. doi:10.1371/journal.pgen.1000144.g003 information to the brain do appear to have experienced positive selection. These changes in sensory perception could conceivably have been brought on by, or could have contributed to, increased brain size and complexity in primates. To gain further insight into the patterns of positive selection that have shaped present-day mammalian genes, we devised a model that allows for probabilistic inferences about the selection histories of individual genes. A selection history is defined as an assignment to each branch of the phylogeny of one of two evolutionary modes: positive selection (each site evolves with v 0 ,1, v 0 = 1, or v 2 .1) or absence of positive selection (each site evolves with v 0 ,1 or v 0 = 1). The model allows a posterior distribution over selection histories to be inferred for each gene, and it allows for estimates of the number of genes under positive selection on individual branches and clades that consider uncertainty about selection histories. Unlike the branch-and clade-specific LRTs-which are simple one-sided hypothesis tests and are necessarily conservative about rejection of the null hypothesis-this model considers all candidate histories symmetrically, and allows for ''soft'' (probabilistic), rather than absolute, choices of history at each gene. Briefly, the model is defined in terms of a simple switching process along the branches of the phylogeny. It has separate parameters for the rates of gain and loss of positive selection at several switch points on the tree, with two switch points per internal branch and one per external branch (see Figure 4A and Methods). The joint posterior distribution of these parameters and of all selection histories is inferred from the data by a Gibbs sampling algorithm (see Methods and Text S1). The inference procedure is computationally intensive, so it was applied only to the 544 genes identified by one or more LRTs as showing significant evidence of positive selection. Because in these cases the null model of no positive selection had already been rejected by a conservative test, the history without selection on any branch was excluded, leaving 2 9 21 = 511 possible histories for the nine-branch (unrooted) phylogeny. To reduce computational cost, the inference of selection histories was conditioned on the maximum likelihood estimates of the parameters of the codon models (see Methods). The inferred rates of gain and loss are quite variable ( Figure 4A and Figure S2 ), with posterior means ranging from about 0.01 to 0.53. These rates are sharply reduced for the external branches of the tree, probably in large part because of diminished power to detect changes in selective mode on these branches. The number of genes inferred to be under selection also varies by branch, but not as dramatically, with expected values ranging between 207.9 and 393.9 and many 95% credible intervals overlapping ( Figure 4B ). Despite differences at individual branches, gains and losses appear to be roughly in equilibrium overall, with 61% of genes estimated to have been under selection at the root, and between 38% and 62% (averaging 50%) under selection at the leaves. The slight tendency to lose selection over time could reflect an ascertainment bias for genes that experienced selection early in mammalian evolution, which will tend to display signatures of selection on multiple long branches of the tree and therefore will be more easily detectable by the LRTs. The branches with the most genes under selection (such as those leading to the rodent and primate ancestors, and to dog and macaque) are generally long (see Figure 1A ), suggesting power may influence these estimates. Nevertheless, the unusually high rate of gain on the branch to the rodents, and the comparatively low rate of loss on that branch (both having fairly low posterior variance; Figure S2 ), suggest not just differences in power but a real tendency for a net gain of selection on this branch, perhaps due to larger population sizes in the rodents. Whether because of power or a genuine increase in selection, the rodent branch appears to play a major role in the identification of PSGs. An expected 72% of the 544 candidate PSGs are under selection on this branch. The posterior distributions over histories suggest that few genes have experienced positive selection specific to individual branches or clades ( Figure 4B) . Instead, most genes appear to have switched between evolutionary modes multiple times. The estimated number of mode switches per gene (averaging across genes but considering the joint posterior distribution for all selection histories) is 1.6 (95% CI: 1.5-1.7), with 0.6 gains (0.5-0.7) and 1.0 losses (0.9-1.1). An expected 91% of PSGs have experienced at least one mode switch, and an expected 53% have experienced two or more switches. 54% of PSGs have 95% CIs excluding zero switches (i.e., with high confidence, these genes have switched modes at least once), and 10% have 95% CIs also excluding one switch (with high confidence, they have switched modes at least twice). Thus, this analysis suggests that positive selection tends to be gained and lost relatively frequently in mammalian genes. Episodic positive selection has been observed and analyzed in detail at individual loci (e.g., [52, 53] ) but to our knowledge genome-wide evidence of this phenomenon in mammalian phylogenies has not previously been reported. Interestingly, our observations are qualitatively compatible with Gillespie's theoretical model of an episodic molecular clock [54] , although our model differs from his in detail. By pooling information across genes and allowing for uncertainty in selection histories, this method estimates much larger numbers of genes under positive selection on each branch of the tree than do the more conservative LRTs (Figure 1 ). For example, the expected number of genes under selection on the branch to the primates is 360.5 (95% CI 338-382), compared with 21 genes identified by the corresponding LRT, and the expected number under selection on the branch to the rodents is 393.9 (357-426), compared with 56 identified by the corresponding LRT. In this analysis, the estimated numbers of genes that have experienced positive selection on the various primate and rodent lineages are not dramatically different, suggesting that the sharp differences from the LRTs in large part reflect inequalities in power. They also suggest that the numbers of genes under selection in recent human and chimpanzee evolution are not as different as they appear from LRTs, which will identify only the most extreme cases [9] . Indeed, the 95% CIs for the human and chimpanzee estimates heavily overlap. In addition to being useful in a bulk statistical analysis of all PSGs, the Bayesian framework can be used to identify the single most likely selection history for each gene. In some cases, these histories are consistent with known functional differences between species, and help to shed light on the evolutionary basis of these differences. For example, the sweet receptor TAS1R2 has been shown in knock-out experiments to be responsible for differences between species in preferences for sweet tastes [55] . (Humans can taste several natural and artificial sweeteners that mice cannot, such as monellin, thaumatin, aspartame, and neohesperidin dihydrochalcone.) This gene is predicted to have experienced selection on the primate clade and on the branches leading to the primate and rodent clades (posterior probability [PP] = 0.20), suggesting that positive selection on TAS1R2 in both primates and rodents could have contributed to differences in sweet taste preferences. Another example is the integral membrane glycoprotein GYPC, which plays an important role in regulating the mechanical stability of red blood cells. In humans, GYPC has been associated with malaria susceptibility, and predicted to have undergone recent positive selection [56] . However, we find evidence that GYPC has experienced positive selection on all branches of the primate clade (PP = 0.66), suggesting longer-term selective pressure that have also affected nonhuman primates. A third example is CGA, which encodes the alpha subunit of the four human glycoprotein hormones (chorionic gonadotropin, luteiniz-ing hormone, follicle stimulating hormone, and thyroid stimulating hormone). This gene shows strong evidence of positive selection specific to the primate clade (PP = 0.82), consistent with the proposal that relatively recent adaptations in pregnancy and development have played a critical role in the evolution of the human endocrine system [57] . Interestingly, the closely related genes CGB1 and CGB2 (which encode two of the six beta subunits of chorionic gonadotropin) are thought to have originated by gene duplication in the common ancestor of humans and great apes [58] , and these events could have contributed to positive selection on CGA. Finally, the complement components C7 and C8B, which encode proteases in the membrane attack complex, are predicted with high probability to be under selection in rodents only (C7: PP = 0.98 for selection in mouse; C8B: PP = 0.93 for selection in mouse and rat). Differences in complement proteases are thought to explain certain differences in the immune responses of humans and rodents [59] . We examined the human mRNA expression levels of PSGs non-PSGs using public data from the Affymetrix Human Exon 1.0 ST Array, which contains probes for nearly all of our genes and permits accurate estimation of expression levels [60] . Our most striking finding was that PSGs show reduced expression levels in all of the 11 available tissues (breast, cerebellum, heart, kidney, liver, muscle, pancreas, prostate, spleen, testes, and thyroid; see Methods). In particular, a significantly smaller fraction of PSGs than of non-PSGs produce a hybridization signal above the background level for the array (P,4610 24 in all tissues for PSGs defined by the all-branch test, one-sided FET). Moreover, among genes expressed above background, expression levels are significantly lower for PSGs than for non-PSGs (P,7610 25 in all tissues, one-sided MWU test; Figures 5A-C) . PSGs also show significantly greater tissue bias than non-PSGs, as measured by the statistic t [61] ( Figure 5D ) or by an alternative statistic here denoted c [17] (Methods). The differences in expression level and tissue bias between the two sets of genes do not appear to be explained by differences in false negative or false positive rates in the detection of positive selection, and the differences in expression level do not appear to be a consequence of the differences in tissue bias (Text S1). In addition, the observed differences remain if the genes that belong to strongly enriched GO categories (Table 2) are excluded, indicating they cannot be attributed to particular classes of PSGs known to have tissue-specific expression patterns, such as those involved in immunity or spermatogenesis. That expression levels are reduced in all tissues further suggests the existence of a general relationship between expression patterns and the likelihood of positive selection. Consistent with previous observations (e.g., [62] ), we found a significant negative correlation of v with expression level in all 11 tissues (Spearman's rank correlation coefficient r ranged from 20.25 to 20.43). In addition, we observed a positive correlation of v with tissue bias, as measured by t (r = 0.24) [63, 64] . (Similar correlations were observed when the log likelihood ratio in the test for positive selection on any branch-which increases with increasing evidence for selection-was used in place of v.) Unlike in previous studies, however, we were able to examine these correlations separately for positively and non-positively selected genes, using the set of PSGs identified by the all-branches LRT. Interestingly, the correlations of v with expression level t are much stronger within the non-PSGs than within the PSGs, indicating that the observed correlations are primarily driven by negative rather than positive selection (see also [65] ). Thus, while genes expressed at low levels and/or in a tissue-specific manner show an increased tendency to have experienced positive selection, the strength of positive selection does not appear to be strongly correlated with their expression patterns (see Discussion). Of the 15,823 genes that were tested for positive selection and had detectable expression in at least one tissue, 1,509 showed a strong preference for one tissue and were designated as tissue specific (c t .0.25 for some tissue t and c t .0.25 for all t9 ? t; see Methods). Based on this designation, spleen-and testes-specific genes were strongly enriched for PSGs: 22 of 174 (12.6%) spleenspecific genes were PSGs, compared with only 2.2% of other genes (P = 8.7610 211 , one-sided FET); and 45 of 715 (6.3%) testesspecific genes were PSGs, compared with 2.1% of other genes (P = 8.2610 210 ). There were also significant, but weaker, enrichments for PSGs among liver-specific (P = 9.1610 23 ) and breastspecific (P = 1.0610 22 ) genes. Not surprisingly, the spleen-specific PSGs generally appear to be immune-related, and many of the testes-specific PSGs are involved in spermatogenesis or sperm adhesion (they include ADAM2 and SPAM1; Table 3 ). The liver and breast specific genes are more heterogeneous. In contrast, only 2 of 254 (0.7%) cerebellum-specific genes were PSGs, compared with 2.3% of other genes (P = 0.066, one-sided FET). Only a few tissue-specific genes were identified by the clade tests, so it was not possible to compare the relationships between tissue-specific expression and positive selection in primates versus rodents. However, there were significant enrichments for primate PSGs among spleen-specific genes, and for rodent PSGs among testesspecific genes. Despite our large data set, we found no indication of a correlation between expression in the primate brain and recent positive selection in protein-coding regions [66] (see [67, 68] ). Indeed, we found some evidence to the contrary: PSGs identified by the primate-clade test show more sharply reduced expression levels (compared with non-PSGs) in the cerebellum than in any other tissue; cerebellum-specific genes are depleted, not enriched, for PSGs; and none of the primate PSGs show tissue-specific expression in the cerebellum. These findings, of course, do not rule out positive selection in individual genes of great importance in brain development, nor do they rule out positive selection on gene expression. While positive selection was our primary focus, our data set also provides an opportunity to compare the average rates of protein evolution in various mammalian lineages. We estimated a separate nonsynonymous-synonymous rate ratio v for each branch of the six-species phylogeny, pooling data from all ortholog sets ( Figure 1A) . Consistent with previous findings [6, 8] , we observe that protein-coding genes, on average, have experienced moderately strong purifying selection (v « 1) on all branches of the phylogeny, but that estimates of v vary considerably within the mammals. These estimates are largest for the hominids (v<0.25), smallest for the non-primate mammals (0.12,v#0.14), and intermediate for non-hominid primates (0.17,v,0.21). It is thought that increased estimates of v in hominids primarily result from weakened purifying selection, owing to reduced effective population sizes [69, 5] . The intermediate values for non-hominid primates may also be influenced by population size. To examine the relationship between v and population size further, we made use of a theoretical relationship between v and the scaled selection coefficient c (see [70, 71] ), which holds if nonsynonymous substitutions have equal (and small) selection coefficients, if synonymous substitutions are neutral, and if population sizes are sufficiently large (Methods). This relationship allows ratios of population sizes to be estimated from ratios of v estimates, under the assumption of constant selection coefficients across species. Here we further assumed that the ancestral population sizes of humans and the chimpanzee subspecies Pan troglodytes versus (to which the sequenced animal belonged) were roughly the same (N h = N c ) [5] , and estimated the ratio of v m in macaque to v h in human/chimpanzee from our 10,980 humanchimpanzee-macaque ortholog trios. Our estimate of v m / v h = 0.732 implies an estimate for the ratio of the macaque to human ancestral population sizes of N m / N h = 1.41 [bootstrapping 95% CI (1.15, 1.64)]. In comparison, the ancestral macaque population size has been estimated at ,73,000 [72] and ancestral human and chimpanzee population sizes at 40,000-70,000 [73, 74] , which would imply a ratio of 1.04-1.82, in reasonable agreement with our estimate. We used the same theoretical relationship to devise a LRT indicating whether or not each gene deviated significantly from the assumed model with N m / N h = 1.41 (Methods). For the vast majority (96%) of the 10,980 genes examined, no significant deviation was observed, indicating that the differences in selection pressure in macaque and the hominids are generally well-explained by differences in population size. To compare the power of our LRTs with the power of previous tests based on hominid or primate genomes, we simulated data sets under a range of parameter values and measured the fraction of cases in which positive selection was predicted ( Figure 6) . These experiments show that power increases substantially when the set of species under consideration is expanded from the two hominid species to the three primates then to all six mammals. With hominid species only, power is poor even when selection is quite strong (e.g., ,20% with a constant v = 2 and ,40% with v = 4), suggesting that a genome-wide scan will tend to identify only the most extreme cases of positive selection. If a rigorous correction for multiple testing is applied, a test based on hominids only has essentially no power, even for fairly long genes under strong selection ( Figure S3 ; see also [5] ). The situation is considerably improved by the addition of the macaque genome, but power remains poor when controlling for multiple testing unless genes are long and selection is strong. When all six mammals are considered, however, power increases substantially. With the full data set, power is reasonably good ($70%) even when genes are short and selection is moderate in strength; it remains good when multiple comparisons are considered ( Figure S3 ). The absolute estimates of power from these experiments depend on the simplifying assumptions used in the simulations (including the unrealistic assumption of constant v among lineages and among sites), and they must be interpreted cautiously. However, estimates of relative power-which will be less sensitive to these simplifying assumptions-indicate a substantial improvement is achieved by the addition of the three non-primate mammals. Since it first became possible to compare the sequences of complete mammalian genomes about five years ago, a number of genome-wide scans for positively selected genes (PSGs) have been conducted using phylogenetic methods [4, 5, 6, 7, 8, 9] . These studies have provided a valuable initial assessement of the genome-wide landscape of positive selection in mammals, but they have left many important questions unanswered. The analysis presented here, by incorporating non-primate mammalian genomes into a genome-wide scan for positive selection, represents a significant step forward. The larger, more divergent group of species improves power significantly, and the use of a nontrivial phylogeny provides insight into the particular patterns of positive selection that have helped to shape present-day genes. To our knowledge this is the largest and most detailed genome-wide analysis of positive selection to date, not only in mammals but in any group of organisms (although extensive analyses, similar in some respects, have been performed recently in Drosophila [75, 76] ). One finding of particular interest was that several whole pathways are especially rich in PSGs. Examples include the classical and alternative pathways for complement-mediated immunity and the FAS/p53 apoptotic pathway ( Figures S1, S4 and S5). These findings suggest that positive selection may frequently act directly on whole protein complexes or pathways (see [77, 78] ). Alternatively, adaptive changes in one protein may sometimes have a cascade effect, leading to changes in other genes that bring a system back into equilibrium. Whether or not all changes affecting a pathway are driven by positive selection, one might expect to see similarities in the selection histories of gene with closely related functions. Indeed, we have found that genes with similar selection histories on average have substantially greater similarity in their GO categories than do genes with more divergent histories ( Figure S6 ). The observations that multiple interacting genes often show evidence of positive selection and that positive selection is frequently episodic may well be connected. For example, in some cases a transient external force could induce a burst of changes in multiple genes that participate in the same pathway, either separately or by triggering a cascade of interdependent events. Further unraveling the (co-)evolutionary histories of interacting PSGs promises to be a fertile area for future Table 3 . Summary of individual PSGs discussed in this article. Cytokine/chemokine C-C motif: CCL1 (P = 5.2610 24 ), CCL20 (P = 7.6610 24 ); C-X-C and C-X3-C motifs: CXCL5 (P = 8.1610 24 work. Care will be required to distinguish between true coevolution and correlations that can be explained by dependencies on expression levels or other covariates of evolutionary rate [79] . Our finding that PSGs are expressed at lower levels and in a more tissue-specific manner than non-PSGs is consistent with a well-known negative correlation v with expression level, and a positive correlation of v with tissue bias (t or c). Various explanations have been proposed for the observed decrease in v among genes expressed at high levels and/or expressed broadly across tissues, including selection for translational efficiency, selection against misfolding, or increased selection due to pleiotropy [62, 68, 65] . In any case, these genes do appear to experience a reduction in their evolutionary ''flexibility'' compared with genes expressed at low levels and/or nonuniformly across tissues. Our observation of decreased rates of positive selection among these genes-and increased rates among lowexpression/high-tissue-bias genes-is consistent with this characterization. Interestingly, however, we observe that correlations of v with expression level and t hold strongly within non-PSGs, but are much less pronounced within PSGs. Thus, expression levels and patterns are strongly correlated with both the strength of negative selection and the likelihood of positive selection, but they are only weakly correlated with the strength of positive selection. It appears that genes may be more likely to come under positive selection if they are in a state of evolutionary flexibility brought on by reduced or tissue-specific expression, but once positive selection has taken hold their subsequent evolutionary course is not strongly dependent on their expression patterns. As additional mammalian genomes become available, the statistical power to detect positive selection will improve. However, most forthcoming genomes are being sequenced at low coverage, and will inevitably exhibit increased levels of error in base calls, genome assemblies, ortholog identification (due to short contigs), and alignment-all of which can lead to spurious signals for positive selection. (The same errors tend to produce false negatives, rather than false positives, in the identification of conserved elements.) Thus, careful data quality controls will be needed to take advantage of these data. In addition, when considering the impact of additional sequences on statistical power, it is useful to distinguish between positive selection that acts continuously (or in recurrent episodes) over a long evolutionary period, and positive selection that acts transiently or in a lineagespecific manner. Deep phylogenetic sequencing should generally improve detection power for continuous or recurrent positive selection, but power for transient selection depends strongly on the sequenced species and the lineages of interest. For example, the genome sequences of a dozen non-primate mammals will likely have little effect on the power to detect human-specific selection, while the gorilla and neanderthal genomes could help considerably. There are fundamental limitations in the detection of weak, transient, or highly localized positive selection that will not be overcome by any amount of genome sequencing. Nevertheless, the availability of several new primate genomes, including those of the orangutan, marmoset, and gorilla, may significantly improve power for PSGs in primates. Our ability to connect positive selection with function remains rudimentary, but gradual progress is being made. As additional sequence data becomes available, it will become possible to associate selection with individual residues of proteins with greater accuracy. At the same time, more data is becoming available on the specific functional roles of individual amino acids, for example, from structural or mutagenesis studies. As a result, it will increasingly become possible to find direct links between selection and function. Often these links will initially be tentative, as in our site-specific analysis of the HAVCR1 gene. Nevertheless, they provide a valuable starting point for experimental follow-up. At the same time, more can be done to incorporate non-sequence data-such as structural and expression data-into computational methods for detecting positive selection. Thus, improvements in both computational and experimental methods will be needed to establish deeper and more informative connections between evolutionary dynamics and molecular function. The latest human (hg18), chimpanzee (panTro2), rhesus macaque (rheMac2), mouse (mm8), rat (rn4), and dog (canFam2) genome assemblies were obtained from the University of California, Santa Cruz (UCSC) Genome Browser. Humanreferenced whole-genome alignments were constructed from syntenic pairwise alignments with human (the ''syntenic nets'') using the UCSC/MULTIZ alignment pipeline [80, 81] . Low quality bases (Phred score ,20) from the chimpanzee, macaque, rat, and dog genomes were converted to 'N's in these alignments. A starting gene set was composed from of the human RefSeq [28] , UCSC Known Genes [29] , and VEGA [30] annotations (downloaded from UCSC Feb. 19, 2007) . Transcripts that lacked annotated coding regions (CDSs), that had CDSs of ,100 bp, or that had CDSs whose lengths were not multiples of three were discarded, leaving 88,879 nonredundant transcripts. These transcripts were grouped by same-stranded CDS overlap into 21,115 genes (transcript clusters). All transcripts were mapped from human to each of the other five mammalian species via the syntenic alignments, then subjected to a series of filters designed to minimize the impact of annotation errors, sequence quality, and changes in gene structure on subsequent analyses. Briefly, each human transcript was required (1) to map to the non-human genome via a single chain of sequence alignments including $80% of its CDS; (2) after mapping to a non-human species, to have #10% of its CDS in sequencing gaps or low quality sequence; (3) to have no frame-shift indels, unless they were compensated for within 15 bases; (4) to have no in-frame stop codons and to have all splice sites conserved. To allow for genes that are mostly conserved but whose start or stop codons have shifted, incomplete transcripts-with ,10% of bases removed from the 59 and 39 ends of the CDS-were also considered. The final collection of ortholog sets was obtained by selecting, for each gene, the (complete or incomplete) transcript that successfully mapped to the largest number of non-human species. In the case of a tie, the transcript with the greatest total CDS length was selected. This procedure resulted in 17,489 genes with $2 non-human orthologs, averaging ,5 species per gene (including human; see Table 1 ). To establish 1:1 orthology, each human gene and putative nonhuman ortholog was examined for evidence of an inparalog (a Figure S3 .) When v#1, these fractions are estimates of the false positive rate. Each data point is based on 1000 data sets simulated with evolver [84] under the assumption of a constant v among lineages and among sites (model M0). All other parameters (including the transition-transversion ratio k, the codon frequencies, and the branch lengths) were fixed at values estimated from the real data. Results are shown for short (200-codon) and long (500-codon) genes and three sets of species: hominids (human and chimpanzee), primates (human, chimpanzee, and macaque), and all six mammals. Details on the computation of P-values are given in Text S1. Note the logarithmic scale on the x-axis. doi:10.1371/journal.pgen.1000144.g006 paralog arising from a recent duplication [82] ) with respect to the other species. Specifically, if either gene had a BLASTN match within the same species (with $80% CDS alignment) that was more similar than the two orthologs were to each other, then that gene was considered recently duplicated and was excluded from the analyses of positive selection. The removal of a duplicated gene did not require an ortholog set to be discarded entirely, provided a human gene and $2 nonhuman orthologs still remained. A collection of genes and gene predictions from the UCSC Genome Browser were used in the identification of inparalogs. When comparing rodent vs. non-rodent and rodent vs. rodent distances, a simple correction for unequal rates of evolution was applied. Further details are given in Text S1. The LRT for selection on any branch of the phylogeny is essentially Nielsen and Yang's [31] test of site models 2a versus 1a, and the lineage-and clade-specific LRTs are essentially instances of Yang and Nielsen's [26] test 2 (see also [83, 27] ). However, to reduce the number of parameters estimated per gene, the complete set of 17,489 genes was divided into eight equally sized classes by G+C content in third codon positions. The branch lengths and the transition-transversion rate ratio k were estimated for each class under the null model, and these estimates were subsequently held fixed, in a G+C dependent way, for the LRTs. Instead of a complete set of branch lengths, a single scale parameter m was estimated per gene. Thus, only the parameters m, v 0 ,1 and p 0 for the null model and the additional parameters v 2 .1 and p 1 for the alternative model, were estimated per gene (see [31, 26] ). This parameterization speeds up calculations substantially compared to estimating k and a set of branch length per gene, while its sensitivity, specificity and power to detect positive selection are comparable (Text S1). We developed our own software for likelihood computation and parameter estimation to support this parameterization. For the LRT for selection on any branch, P-values were computed empirically, based on simulation experiments. 10,000 alignments were simulated under the 'nearly neutral model' (allowing for a fraction p 0 of sites to evolve with v 0 ,1 and a fraction 12p 0 to evolve with v 1 = 1) for each G+C class using evolver [84] . Alignment lengths and values of m, v 0 and p 0 were drawn from the empirical distribution defined by the real alignments (using estimates obtained under the null model), and the remaining parameters were fixed at global estimates for each G+C class. Log likelihood ratios (LLRs) were then computed exactly as for the real data. The nominal P -value for a LLR of r was defined as the fraction of all simulated alignments with LLR$r, unless the number of such alignments was ,10, in which case we assumed 2r*x 2 df~1 (an adequate approximation for small P-values, according to the simulation experiments). The method of Benjamini and Hochberg [85] was used to estimate the appropriate P-value threshold for a false discovery rate of ,0.05. For the lineage-and clade-specific LRTs, P-values were computed assuming the null distribution was a 50:50 mixture of a x 2 df~1 distribution and a point mass at zero (see [27] and discussion in Text S1). Let X = (X 1 ,…, X N ) be the alignment data, with X i denoting the alignment for the i th gene (1#i#N; here N = 544), and let Z = (Z 1 ,…, Z N )be the set of selection histories, with Z i denoting the selection history for the i th gene (1#Z i #M; here M = 511). Recall that a selection history is defined as a pattern of presence or absence of positive selection on the branches of the unrooted phylogeny. Let Z ib M {0,1} indicate the selective mode (with 1 representing positive selection) for branch b M {1,…,B} (here B = 9) under history Z i . The parameters of the switching model, denoted h, are defined below. The model assumes independence of genes and independence of histories, and conditional independence of X and h given Z. Thus, the complete data likelihood is given by: The probability of a history, P(Z i |h), is a function of the set of switches in selective mode required to explain the history parsimoniously. For each history to be explained parsimoniously, switches must be allowed to occur early (near the ancestor) or late (near the descendant) on each internal branch, as well as (early) on each external branch ( Figure 4A ; see Text S1 for a justification of the model). Thus, there are twelve possible switch points, with three of them adjoining each of the four internal nodes of the tree. It is convenient to denote these points P nb : n[N ,b[B n f g where N is the set of internal nodes and B n represents the branches adjoining node n. Let V nb M {0,1} and W nb M {0,1} indicate the selective states before and after point P nb , respectively. For a given history Z i , these variables are uniquely determined by parsimony according to a simple algorithm (see Text S1). The four possible values of (V nb , W nb ) correspond to four possible scenarios at P nbgain of selection (0,1), loss of selection (1,0), absence of gain (0,0), or absence of loss (1,1). The probabilities of these scenarios (i.e., the conditional probability of each W nb given V nb ) are defined by a parameter for gains (h nbG ) and a parameter for losses (h nbL ) at each point. In addition, the prior probability of selection at the root of the tree is given by a parameter h 0 . (For this analysis, the most recent common ancestor of the primates and rodents is treated as the root of the tree; see Text S1.) The set of parameters can thus be described as h~h nbe : The prior probability of a history Z i is simply a product of the prior and the relevant switching probabilities: where U 0 represents the selective state at the root. The switching model effectively defines a prior distribution over histories, which tends to favor simpler histories over more complex ones (typically h nbe ,0.5). The prior probability for each element of h is defined by a (conjugate) Beta distribution with parameters a and b (here, a = 1, b = 9). Because these elements are independent in the prior, The term P (X i | Z i ) in equation 1 is simply the likelihood at gene i of a branch-site codon model that assumes selection history Z i . A full Bayesian approach would integrate over the parameters of these codon models, but this would be computationally prohibitive. Instead, we make the Empirical Bayes simplification of conditioning the analysis on maximum likelihood estimates of the parameters of the codon models. The maximized log likelihoods L ij for all genes i and histories j are precomputed using existing software (in parallel, on a large computer cluster) and stored in an N6M matrix, which is then used in the inference of selection histories. The variables Z and h are unobserved, and the goal is to infer their joint posterior distribution, This inference was accomplished by a Gibbs sampling algorithm that alternates between sampling each Z i conditional on X i and a previously sampled h, and sampling each element of h conditional on a previously sampled Z. It is straightforward to derive the required conditional distributions and to sample from them (Text S1). The Gibbs sampler converges rapidly and mixes well. Notice that, because the history without selection on any branch is excluded, all of the histories are described by codon models with the same number of parameters. Therefore, no penalty for parameter number is needed when comparing histories. After an appropriate burn-in period, each iteration of the Gibbs sampler produces a sample (Z (t) , h (t) ) from P(Z, h|X). Estimated posterior expected values of interest were obtained by averaging these samples or functions of these samples, and Bayesian 95% confidence intervals were obtained by taking the 0.025 and 0.975 quantiles of the sampled values. For example, the posterior expected number of genes under selection on branch k (see Figure 4 ) was estimated as 1 , where T is the number of samples and the function f k (Z) counts the number of genes under selection on branch k in a set of histories Z. Each gene was assigned categories from the GO [32] and PANTHER [86] databases (downloaded on June 26, 2007) , based on the Uniprot identifiers of associated transcripts. At least one GO category was identified for 14,137 (86%) genes, and at least one PANTHER category for 13,753 (83%) genes. To account for the hierarchical nature of these databases, each gene was also considered to belong to all parent categories of the ones to which it was directly assigned. For each category C and set of PSGs S, a 262 contingency table was constructed for the numbers of genes assigned or not assigned to C, and within and outsideS, then a (one-sided) P-value for independence of rows and columns was computed by Fisher's exact test. In addition, the distributions of LRT P -values among the genes assigned to C and not assigned to C were compared by a (one-sided) Mann-Whitney U (MWU) test. (Notice that S is not considered in this case.) Nominal P -values computed by the FET and MWU tests were corrected for multiple comparisons using the method of Holm [87] . The analysis of gene expression was based on the publicly available ''Tissues+Mixtures'' sample data set for the Affymetrix GeneChip Human Exon 1.0 ST Array (http://www.affymetrix. com/support/technical/sample_data/exon_array_data.affx). The RMA-based probeset summaries [88] and DABG (detected above background) -values were used. Each probeset was assigned genomic coordinates using the ''Affy All Exon'' track in the UCSC browser (hg17), then was associated with any human gene from our set having an exon on the same strand that completely contained the probeset. Nearly every gene (98%) had at least one probeset. To calculate a P-value for each gene6tissue, the DABG Pvalues of all associated probesets (pooling the three replicates per probeset6tissue) were combined using Fisher's method [89] . A gene was considered to be significantly expressed above background if it had (nominal) P,0.001. Similarly, an estimated expression intensity for each gene6tissue was calculated by first taking the median over the three replicates of each RMA-based probeset summary, then taking the median of these values over all probesets associated with the gene. The analysis of expression intensities was restricted to genes significantly expressed above background so that genes expressed at or near the background level did not drive the results. To measure tissue bias, we used: (1) the statistic t [61] , which represents the average difference in normalized expression intensity from that of the tissue of maximal expression, and (2) a statistic, here denoted c, defined as c = max t c t , where c t is the squared cosine of the angle between the expression vector and the coordinate axis associated with t (see [17] ). In defining genes as tissue specific for tissue t we required c t .0.25 and c t9 ,0.125 for all t9 ? t. Further details are given in Text S1. Maximum likelihood estimates of v for each branch were obtained using the codeml program in the PAML software package [84] , with F364 codon frequencies, estimation of k (fix_kappa = 0) and a single v across sites per branch (model = 1, NSsites = 0). The tree topology shown in Figure 1 was assumed. The alignments for all genes were concatenated for this analysis. Assuming all non-synonymous mutations at a given gene have the same selection coefficient and all synonymous mutations are neutral, population genetic theory says that v should be given by [70, 90] : where c = 2Ns. Therefore, c can be estimated as f 21 (v), where v = f(c) denotes the function above. (Values of c can be obtained numerically; see Text S1.) Ratios of population sizes can therefore be estimated from ratios of v estimates: N1 The LRT to test whether differences in population size can explain the differences in v in human and macaque was constructed as follows. The null model assumes v h = v c and v m = 0.732v h (see Results). The alternative model also assumes v h = v c but leaves v m as a free parameter to be estimated from the data. Because the models are nested, a x 2 df~1 distribution is used for significance testing. This test was applied separately to each gene. A website is available at http://compgen.bscb.cornell.edu/ projects/mammal-psg/ with definitions of the candidate genes (accession numbers, genomic coordinates, and descriptions), multiple alignments of orthologous gene sets, GO and PANTHER category assignments, detailed results of the LRTs and the Bayesian analysis, and other resources. In addition, the candidate genes and predicted PSGs are displayed as a track in the UCSC Genome Browser (http://genome.ucsc.edu; assembly hg18).

Mathematical Analysis of Copy Number Variation in a DNA Sample Using Digital PCR on a Nanofluidic Device

Copy Number Variations (CNVs) of regions of the human genome have been associated with multiple diseases. We present an algorithm which is mathematically sound and computationally efficient to accurately analyze CNV in a DNA sample utilizing a nanofluidic device, known as the digital array. This numerical algorithm is utilized to compute copy number variation and the associated statistical confidence interval and is based on results from probability theory and statistics. We also provide formulas which can be used as close approximations.

Digital PCR conventionally utilizes sequential limiting dilutions of target DNA, followed by amplification using the polymerase chain reaction (PCR) [1, 2] . As a result, it is possible to quantitate single DNA target molecules. We utilize the digital array, which is a novel nanofluidic biochip [2, 3] where digital PCR reactions can be performed ( Figure 1 ) by partitioning DNA molecules, instead of diluting them. This chip utilizes integrated channels and valves that partition mixtures of sample and reagents into 765 nanolitre volume reaction chambers. DNA molecules in each mixture are randomly partitioned into the 765 chambers of each panel (the total volume of the PCR mix in each panel: 6 nl6765 = 4.59 ml). The chip is then thermocycled and imaged on Fluidigm's BioMark real-time PCR system and the positive chambers that originally contained 1 or more molecules can be counted by the digital array analysis software ( Figure 2 ). Copy number variations (CNVs) are the gains or losses of genomic regions which range from 500 bases on upwards in size. Whole genome studies have revealed the presence of large numbers of CNV regions in human and a broad range of genetic diversity among the general population [4, 5, 6] . CNVs have been the focus of many recent studies because of their roles in human genetic disorders [7, 8, 9] . Current whole-genome scanning technologies use array-based platforms (array-CGH and high-density SNP microarrays) to study CNVs. They are high throughput but lack resolution and sensitivity. Real-time PCR is a sequence-specific technique which is easy to perform, but is limited in its discriminating power beyond a 2-fold difference [11, 12] . CNV determination on the digital array is based upon its ability to partition DNA sequences. Given the number of molecules per panel and the dilution factor, the concentration of the target sequence in a DNA sample can be accurately calculated. In a multiplex PCR reaction with 2 or more assays, multiple genes can be quantitated simultaneously and independently, effectively eliminating any pipetting errors if separate reactions have to be set up for different genes. When a single copy reference gene (RNase P in this study, [10] ) is used in the reaction, the ratio of the target gene to the reference gene would reflect the copy number per haploid genome of the target gene. Primary contribution of this paper In this paper we will show that the digital array provides a robust and easy-to-use platform to study CNVs. We have derived a mathematical framework to calculate the true concentration of molecules from the observed positive reactions in a panel. We also show how to perform statistical analysis to find the 95% confidence intervals of the true concentrations and the ratio of two concentrations in a CNV experiment using the digital array with multiplex PCR. The copy number variation problem can be stated as follows. Given two counts h 1 and h 2 of positive chambers for two genes in a digital array panel, how can one estimate a ratio of true concentrations r = l 1 /l 2 of the two genes and a confidence interval [r Low , r High ] on the estimation? Our approach is built on well-known tools and techniques from statistics. It decomposes the problem into two parts. 2. Given estimated true concentrations l 1 and l 2 of the reference gene and the target gene, respectively, in the DNA sample and their respective confidence intervals, how can one estimate the ratio r = l 1 /l 2 and a confidence interval [r Low , r High ] on this estimation? It turns out that the first question can be answered by applying sampling and estimation theories from statistics and probability, and the second question can be answered by a numerical algorithm based on generalization of a mathematical theorem. For related work on answering the first question, using Bayesian approach, see unpublished preprint by Warren et al. titled ''The Digital Array Response Curve'' dated March 2007 at http:// thebigone.stanford.edu/papers.htm. Warren et al. assumed a uniform probability distribution of number of molecules, with maximum number assumed to be 4000, and using Bayesian and combinatorial methods, presented a solution. The confidence interval obtained using Bayesian probabilistic framework, is often referred to as credible interval or Bayesian confidence interval which requires one to incorporate problem-specific contextual information from the prior distribution. This paper differs from this prior work by Warren et al. in two different ways. First, we consider the parameter l to be a fixed constant, unlike having a probability distribution as in Bayesian approach. Second, in addition to providing the answer to the first question, we are interested in estimating the confidence interval of the ratio of two concentrations which is new work. For difference between credible interval and confidence interval, see [13] . Both approaches give good results depending upon the question one is trying to answer. We will prove mathematical correctness of our results in this paper and present simulation results to help the reader build useful insight. Finally, we present actual CNV experiments on the digital array with known ratios and show the results using the techniques developed in this paper. DNA quantitation in the digital array is based on the partitioning of a PCR reaction into an array of several hundreds or even few thousands of chambers or wells. One panel of the digital array consists of 765 chambers and one can use up to 12 panels at a time. If the concentration of the target molecules is low in the DNA sample, most of the chambers capture either one or no molecules and the number of positive chambers at the end point of the PCR yields close approximation to the true concentration of the target. However, if the number of molecules is large, then there is greater probability of several molecules being in the same chamber, and therefore the number of positive chambers would be significantly lower compared with the number of molecules in the chambers. We are interested in estimating the true concentration of the molecules in the DNA sample from which we extracted 6 nl6765 = 4.59 ml of sample for each panel. Consider the universe of infinite number of the digital array chambers filled with an infinite amount of the DNA sample where the true concentration of the target molecules is l per chamber (per 6 nl). The true concentration is an unknown population parameter of this infinite DNA sample. If a chamber gets no molecule then it constitutes failure in the sense of Bernoulli experiment. If it gets one or more molecules, that is, if it gets a ''hit'' and is therefore positive, then it constitutes success. Let the probability of success be p. Note that p is an unknown population parameter. We will use the standard hat notation to denote sample estimators of population parameters. For example, p andl l will denote the estimators of p and l, respectively. One can model K, the number of molecules in each chamber as a Poisson process, and this gives the relationship between p and l as follows Alternatively, consider M molecules randomly distributed in C chambers. The probability of any given molecule being in any given chamber is 1 C . So the probability p of a given chamber having at least one molecule is As number of chambers becomes arbitrarily large, the above approaches e 2l . Therefore, l~{ln 1{p ð Þ which establishes the relationship between l and p. Confidence Intervals for estimation of p and l A chamber getting a hit or no hit is a binomial process, same as toss of a coin, with success probability p. Let the number of positive chambers in the panel be H. Considerp p~H C as an estimator of p. It is well known that p is an unbiased estimator of p and has expectation p and standard deviation ffiffiffiffiffiffiffiffiffiffiffi and its sampling distribution f(p) is approximately normal for large C. See Figure 3 for illustration of the above ideas. See [13, 14, 15] for extensive literature on obtaining confidence interval for the estimation of binomial probability. It is referred to as binomial sign test when the test statistic can be approximated with the chi-square distribution, specifically through the use of the chi-square goodness-of-fit. An alternative and equivalent approximation is obtained by using the normal distribution and then the test is referred to as the population proportion test, see [15] . If C is large enough, then the confidence limits are approximately given byp For 95% confidence interval, z c = 1.96. For the digital array, C is an integral multiple of 765 and is comfortably large enough for the above approximation. Define the estimator of l aŝ l l~{ln 1{p p ð Þ Since the probabilities in any given differential area of a probability density function are preserved under change of variables, the 95% confidence interval [l l Low ,l l High ] is directly given as followŝ p p Low,High~p p+1:96 See Figure 4 for illustration. Let a random variable X have probability density function f X (x). If h(x) is either increasing or decreasing in x, then U = h(X) has density function given by which follows from the fact that probabilities in any given differential area have to be invariant under change of variables, see [16] . Furthermore, which can be expanded using Taylor series expansion of h(x) around the mean j = E(X), as follows Since in our case, we have the followinĝ l l~{ln 1{p p ð Þ therefore, in above, we have x = p, u~l l and h(x) = 2ln(12x). Sincê l l is a monotonically increasing function of p, one can get the sampling distribution ofl l from the sampling distribution of p as Note that due to nonlinear relationship betweenl l and p, one can not make assumptions about g. In general, g is not normal and Now we derive an approximation for El l from the Taylor series expansion shown above. Higher order central moments of Gaussian function f(p) with mean p are For proof see [17] . Since f(p) has very small s due to very large number of chambers, the higher order terms for all n.0 in the Taylor expansion are small, and therefore the only contributing term is when n = 0, which It is informative and useful to run a simulation experiment on the computer to see how the real-world matches with the theory developed above. For this purpose, one can use a random number generator and a computer program to simulate the universe of the digital array chambers. If a panel has C chambers, consider a universe of C6K many chambers where K is a large number chosen for simulation. Choose some value of l as the true concentration of molecules in one chamber. Therefore, in total, there will be l6C6K molecules. Assign each of these molecules randomly to one of the chambers. Extract K panels out of this universe and for each of the panels, computep p~H C as an estimator of p and plot its histogram over all the K panels. The mean should be p = 12e 2l and standard deviation should be ffiffiffiffiffiffiffiffiffiffiffi . For each of these panels, estimate l and compute the 95% confidence interval. In 95% of the K panels, the true value of l should lie within the confidence interval. For our simulation experiments we chose M = 400, that is, l~4 00 765 . We chose K = 70000. In Figure 5 we show the histogram of H which is really same as distribution of P scaled by a factor of 765. In Table 1 we show how the predicted values match with the actual simulation values. In the same way, the sampling distribution of number of molecules matched with what is predicted by theory. Though the results of the simulation follow from elementary probability, we conducted these simulations in order to build more advanced simulations for ratios of concentrations later. They also illustrate the meaning of the confidence interval. In previous section, we established a method for estimating the true concentration of the target molecules in the DNA sample from the count of positive chambers as well as the 95% confidence interval for this estimation. We also showed how the sampling distribution gl l is related to the sampling distribution f(p). In CNV, the goal is to determine ratio of true concentrations of two genes, one being reference gene and the other being test gene, and associated confidence interval, which we now accomplish in next subsections. Table 1 . Comparison of the metrics of histogram, shown in Figure 5 , of number of positive chambers obtained in simulation with those predicted by the theory. Let the sampling distributions of the test gene and the reference gene be g 1l l 1 and g 2l l 2 , respectively. If these distributions were normal, then one can make use of Fieller's Theorem [18, 19] . However, as mentioned in previous section, one can not make this assumption in general. It is useful to go through the geometric interpretation of Fieller's theorem so that one can solve the problem for arbitrary sampling distributions. See Figure 6 for geometric interpretation of Fieller's Theorem [20, 21] . Assume g 1l l 1 and g 2l l 2 are normal. Forl l 1 andl l 2 , the ratior r~l l 1 .l l 2 can be shown as the slope of the line in the twodimensional plane which passes through the origin and the 2-D point (l l 2 ,l l 1 ). Luxburg et al. show in [20, 21] how a confidence ellipse in the two-dimensional plane can be constructed. Consider the two lines which pass through the origin and are tangents to this ellipse. The intersection of these lines with the vertical line at l l 2~1 gives the desired confidence interval. In this paper we have presented data in a controlled experimental system, where a synthetic DNA construct was spiked into human cell line DNA at different concentrations. In this case, the synthetic construct, which was to the RPP30 gene, was used as the target, and the RNase P gene which was endogenous to the human cell line, was used as the reference gene. The two genes were identified using two separate PCR reactions, using separate PCR primers and probes. Since there is no reason to assume that the amplification and detection of the target and reference genes are linked,l l 1 andl l 2 are independent variables. It is easy to see from the proof of Fieller's theorem and its geometric interpretation that one can compute sampling distribution q of the ratio estimatorr r~l l 1 .l l 2 as follows: This can be interpreted as cutting out thin wedges in the joint distribution of g 1l l 1 and g 2l l 2 and accumulating the probabilities inside the wedge to compute the function q in the corresponding thin interval of the ratio. This is the basis of our numerical algorithm which implements integration in order to compute q(r): 1. Build histograms of sampling distributions g 1l l 1 and g 2l l 2 . The tails of the histograms where probabilities become very small are approximated by zero. 2. Build a histogram of sampling distribution q(r) ofr r~l l 1 .l l 2 by considering each bin [r 1 , r 2 ] and by adding all the joint probabilities of different values of concentrations which give a ratio rM[r 1 , r 2 ]. 3. Compute the mean and the 95% confidence interval from the ratio histogram. See Figure 7 for illustration of the above algorithm. One can still use direct formulas, as an approximation, to compute confidence interval as follows. The means of g 1l l 1 and g 2l l 2 are l 1 and l 2 respectively. Let the standard deviations be s x and s y respectively. For given estimationsl l 1 andl l 2 , assuming that distributions are normal, it follows from the analysis in [20, 21] that the boundary of the confidence ellipse for a given confidence level z c would be defined by It is easy to generalize this to the case, under a reasonably close approximation, when we have asymmetric distributions which are assumed to be normal in each of the four quadrants of the coordinate system centered at (l l 2 ,l l 1 ). Then the confidence region is made of union of four quadrant-wise elliptic regions. Let the asymmetric confidence intervals for specified z c and the two concentrations be [l l 1 {H B ,l l 1 zH T ] and [l l 2 {W L , l l 2 zW R ]. If W R = W L = z c s x and H T = H B = z c s y , it is symmetric case [20, 21] . Using simple algebraic manipulations, it can be shown, as in symmetric case, that the slopes of lines that will be tangents to this union of four quadrant-wise ellipses will bê r r Low~l The above equations can be used as an approximation though numerical algorithm will give more accurate results as the algorithm does not make any assumptions and works with arbitrary sampling distributions. One detail has to be mentioned. Special care has to be taken if the confidence region gets too close tol l 1 axis whenl l 2 is small. If it touchesl l 1 axis, then r High = '. If eitherl l 1 orl l 2 is too small, one can build respective histogram with smaller bin size to get more accurate results. See Table 2 for summary of equations derived in order to solve the copy number variation problem. Though the numerical approach based on histograms is recommended as it does not make assumptions, these direct formulas can be used as close approximation. We conducted simulation studies, using a random number generator and a computer program as in previous section, by choosing a ratio of 2 of concentrations of two genes, which are independent of each other, and building a distribution of estimated ratios over 50 thousand panels. In 94.9% of the panels, the true chosen ratio did lie in the computed confidence intervals thereby showing the correctness of our mathematical analysis. The copy number variation results for known ratios of 1, 1.5, 2, 2.5, 3 and 3.5 are shown in Figure 8 . Materials and methods for this experiment are discussed in next section. As the number of panels P increases, then the number of chambers C = 765 Pincreases and therefore the estimation of the ratio becomes more accurate as well as the confidence interval shrinks. When only 1 panel is used, there is significant overlap between 95% confidence intervals of certain ratios e.g. between ratio 2 and 2.5. There is no overlap when 3 or more panels are used. In all cases the known ratio lies within the computed 95% confidence interval. Note that using mathematical analysis one can find optimal numbers of positive chambers for each ratio which give smallest confidence intervals and which will therefore improve the results. In summary, Fluidigm's digital array is capable of accurately quantitating DNA samples and is a valuable platform for studying copy number variation. It is a robust technology that is sequencespecific, easy-to-use, and extremely flexible. We have presented mathematical and algorithmic solutions to analyze CNV on a digital array. The solution is an elegant application of statistical sampling and estimation theories to such an important real-world Table 2 . Given number of chambers C and counts H 1 and H 2 of the positive chambers in a digital array for the target gene and the reference gene, respectively, list of formulas needed to analyze copy number variation. problem. We have shown how one can compute the true concentration of a target sequence in a DNA sample and the associated confidence interval on this estimation, and how one can compute the ratio of true concentrations of multiple sequences and the associated confidence interval on the estimation of this ratio. A 10-ml reaction mix is normally prepared for each panel. It contains 16 TaqMan Universal master mix (Applied Biosystems, Foster City, CA), 16 RNase P-VIC TaqMan assay, 16 TaqMan assay for the target gene (900 nM primers and 200 nM FAMlabeled probe), 16 sample loading reagent (Fluidigm, South San Francisco, CA) and DNA with about 1,100-1,300 copies of the RNase P gene. 4.59 ml of the 10-ml reaction mix was uniformly partitioned into the 765 reaction chambers of each panel and the digital array was thermocycled on the BioMark system. Thermocycling conditions included a 95uC, 10 minute hot start followed by 40 cycles of two-step PCR: 15 seconds at 95uC for denaturing and 1 minute at 60uC for annealing and extension. Molecules of the two genes were independently amplified. FAM and VIC signals of all chambers were recorded at the end of each PCR cycle. After the reaction was completed, Digital PCR Analysis software (Fluidigm, South San Francisco, CA) was used to process the data and count the numbers of both FAM-positive chambers (target gene) and VIC-positive chambers (RNase P) in each panel. A spike-in experiment was performed using a synthetic construct to explore the digital array's feasibility as a robust platform for the CNV study. A 65-base oligonucleotide was ordered from Integrated DNA Technologies (Coralville, IA) that is identical to a fragment of the human RPP30 gene. The sequences of the primers and FAM-BHQ probe used to amplify this construct are from Emery et al [22] . The primers and probe were ordered from Biosearch Technologies (Novato, CA). Both RPP30 synthetic construct and human genomic DNA NA10860 (Coriell Cell Repositories Camden, NJ) were quantitated using the RPP30 assay on a digital array. Different amounts of RPP30 synthetic construct was then added into the genomic DNA so that mixtures with ratios of RPP30 to RNase P of 1:1 (no spikein), 1:1.5, 1:2, 1:2.5, 1:3, and 1:3.5 were made simulating DNA samples containing 2 to 7 copies of the RPP30 gene per diploid cell. These DNA mixtures were analyzed on the digital arrays as described above. Five panels were used for each mixture and 400-500 RNase P molecules were present in each panel. The ratios of RPP30/RNase P of all samples were calculated using the techniques developed in this paper. For each ratio, we did pooled analysis by adding the numbers of positive chambers in the first P = 1,2,3,4,5 panels. The results are summarized in the previous section and in Figure 8 .

The geosimulation of West Nile virus propagation: a multi-agent and climate sensitive tool for risk management in public health

BACKGROUND: Since 1999, the expansion of the West Nile virus (WNV) epizooty has led public health authorities to build and operate surveillance systems in North America. These systems are very useful to collect data, but cannot be used to forecast the probable spread of the virus in coming years. Such forecasts, if proven reliable, would permit preventive measures to be put into place at the appropriate level of expected risk and at the appropriate time. It is within this context that the Multi-Agent GeoSimulation approach has been selected to develop a system that simulates the interactions of populations of mosquitoes and birds over space and time in relation to the spread and transmission of WNV. This simulation takes place in a virtual mapping environment representing a large administrative territory (e.g. province, state) and carried out under various climate scenarios in order to simulate the effects of vector control measures such as larviciding at scales of 1/20 000 or smaller. RESULTS: After setting some hypotheses, a conceptual model and system architecture were developed to describe the population dynamics and interactions of mosquitoes (genus Culex) and American crows, which were chosen as the main actors in the simulation. Based on a mathematical compartment model used to simulate the population dynamics, an operational prototype was developed for the Southern part of Quebec (Canada). The system allows users to modify the parameters of the model, to select various climate and larviciding scenarios, to visualize on a digital map the progression (on a weekly or daily basis) of the infection in and around the crows' roosts and to generate graphs showing the evolution of the populations. The basic units for visualisation are municipalities. CONCLUSION: In all likelihood this system might be used to support short term decision-making related to WNV vector control measures, including the use of larvicides, according to climatic scenarios. Once fully calibrated in several real-life contexts, this promising approach opens the door to the study and management of other zoonotic diseases such as Lyme disease.

The WNV is a flavivirus which was isolated for the first time in 1937. Its name comes from the district of West Nile in Uganda. It was detected in human, birds and mosquitoes in Egypt at the beginning of the fifties, and has then been found in various European countries [1] . It is however only with the important 1996 human epidemic in Bucharest, Romania, that WNV became a concern for public health. Moreover, there is no specific treatment of the disease and no vaccine is yet available for humans. The WNV was detected on the American continent in 1999 and more specifically in New York [2] . In Canada, WNV reached southern Ontario in 2001, while the first human cases were detected in August 2002 [3] . WNV made its appearance in Quebec in July 2002. The virus was then propagated, like everywhere else, by the intermediary of mosquitoes and birds. The expansion of this epizooty forced the Government of Quebec to adopt an intervention plan which included in 2003 the implementation of a multi-faceted surveillance system [4] . This system brought together field data on human, avian and entomological infection and deaths. While these monitoring activities were undertaken to better understand the epidemiology of WNV and the level of risk it can represent for the human population, they do not allow for forecasts of the probable propagation of the virus on the territory. Such a forecast, if it proved to be reliable, would allow public health authorities to initiate preventative actions at the right time and places and at the appropriate level of expected risk. Currently, one main control activity is larvicide spraying in urban and rural settings in order to reduce the population of mosquitoes infected with WNV. However, it remains difficult to determine the at-risk zones on a scientific basis and the efficacy of such measures has been challenged [5] , not to mention their high cost and environmental impacts. The identification of vulnerable zones and risk levels in due time remains a significant challenge for public health management due to the complexity of the phenomena related to the virus transmission. Multi-agent geosimulation is an artificial intelligence modeling approach which might be used to develop public health management tools in order to anticipate the progression of the disease and to assess various intervention scenarios. This approach makes it possible to simulate the behaviours of thousands of agents in georeferenced virtual spaces. The MAGS System (Multi-Agent GeoSimulation) recently developed by Dr. Moulin's Groupe de Recherche en Informatique Cognitive at Laval University, can be used to create such simulations in virtual environments generated with georeferenced data obtained from geographic information systems (GIS). These agents are characterized by cognitive capacities such as perception of the environment and its content, autonomous navigation and reasoning [6] . Although one of the first applications of MAGS was related to the simulation of crowd behaviours in urban environments, MAGS is a generic platform allowing the simulation of several types of behaviours in various geo-referenced virtual environments. For example, it has already been used to simulate the behaviour of consumers visiting shopping centers and firemen intervention plans to contain the propagation of forest fires [7] . The main objective of the WNV-MAGS Project reported in this paper, was to develop a system to simulate the behaviours and interactions of populations of indicator birds and of mosquitoes involved in the propagation and transmission of the WNV, taking into account the characteristics of the geographic environment. This simulation takes place in a virtual cartographic world representing a large territory (southern part of the province of Quebec, Canada). The simulation also takes into account various climatic scenarios and regimens of larvicide treatments. In Section 2, we present an overview of the phenomena which are linked to the spread of WNV. Then, we present the conceptual model which was developed after setting some carefully chosen hypotheses. Next, we present the geosimulation of the populations of interest, using agents' roosts to represent the dynamics of the bird populations and an intelligent density map to represent the populations of mosquitoes. Some short-term climate scenarios and the calibration of the system are also presented in this section. In Section 3, we present a conclusion and some new work currently underway. In Section 4 we briefly present the design method used to develop the system, including the conceptual architecture and an overview of the mathematical model formalizing the evolution of relevant populations. We also comment upon the quality and availability of data used to feed the system. Finally, we briefly present the implementation context of the system. Overview Figure 1 presents a synthetic view of the phenomena which are involved in the spread of WNV, as adapted for the Quebec context [8] . Indeed, there are mainly two populations involved in the transmission of the WNV: the population of mosquitoes (Culex sp.) and the population of birds. In this paper, we mainly consider the Corvidae family and more specifically crows which have been chosen by public health authorities as indicator birds for the WNV. Mosquitoes spawn eggs in sumps and other shelters. The larvae hatch from eggs and evolve into nymphs that emerge to become adult mosquitoes. This cycle mainly depends on temperature and humidity [9] . Besides, human intervention can reduce the population of mosquitoes through larvicide treatments (e.g. Methoprene) in order to kill larvae. The transmission of the WNV occurs mainly mosquitoes biting birds. An infected mosquito can infect a bird, which can in turn infect healthy mosquitoes that will subsequently bite the infected bird before its death [10] . Regarding the populations of Corvidae, their spatial and temporal characteristics depend on geographic areas and on the periodicity of displacements and grouping. During early spring bird couples spread over the whole territory and remain for few months around their nesting areas. By the end of July, which happens to be the very beginning of human infections in Quebec [3] , Corvidae change their social behaviour and regroup in roosts at night. During the day, the birds fly to surrounding areas in search of food, but they go back to the roosts at night [11] . At the end of fall, many of them migrate to warmer areas south of the province [12] . Furthermore, the transmission of WNV to the populations of Corvidae can occur either by mechanical infection (an infection after a direct contact between birds) or through the biting of a healthy Corvidae by an infected mosquito ( Figure 1 ). Since we wanted to simulate the progression of the WNV infection involving a large number of individuals of two main species interacting in a particular geographic region, we selected a geosimulation approach which allows for the study of the spatial and temporal characteristics of the populations' interactions in a virtual geographical environment [6] . However, given the enormous complexity involved in representing such phenomena and the lack of detailed data, we had to raise a number of reasonable simplifying hypotheses with regard to the species of mosquitoes and Corvidae of interest, to the factors influencing the evolution of these populations, the geographical region selected for the analysis, the period of simulation and the space-time scale. These hypotheses led us to identify a set of key parameters to carry out the simulations, based on the epidemiologic and surveillance experience with WNV in North America and more precisely in the province of Quebec [4] . For example, considering the availability of surveillance data, we selected the American crow as the main indicator bird species; and the Culex pipiens and Culex restuans as the main mosquito's species susceptible to bite crows (and possibly humans). Another example is the period of simulation for the WNV propagation: July 1st to October 1st was selected as the critical time window during which human cases have appeared in Quebec so far (Table 1) . The objective of the conceptual model is to introduce a synthetic view of the phenomena of interest while taking into account the above mentioned simplifying hypotheses ( Figure 2 ). Let us briefly comment upon this model which represents the evolution and interactions of Culex sp. (pipiens/restuans) and crows. Moreover, we simplified the biological cycle of Culex to only consider the change from a larval state to an adult one. From a public health management's point of view, these two states are the most important ones since the virus is spread by adult females and treatments against the progression of the WNV are carried out using larvicides. This simplification has been validated by domain experts (see below). In addition, considering the spatial dynamics of crow populations, we selected the period of the year when Corvidae regroup in roosts. In our model, a roost is considered as the spatial extension of an aggregate of crows (a sub-population of crows which gathers in this roost for the period of the year of interest). During the day, crows fly a variable distance from the roost in search of food, and return at night. Hence, the spatial phenomenon of gathering and dispersion of this sub-population of crows can be represented in a synthetic way in the form of an expansion and a contraction of the area occupied by this sub-population. The surface over which the birds spread during the day ("roost expansion") depends on the roost size. Consequently, we can take into account the variable density of crows in this dynamically changing area. Another fact to consider is that the Culex mosquito has a mostly nocturnal activity [13] . Therefore, the crows located in roosts at night will be good targets for them. Moreover, preset variables have been used in order to compute parameters such as the infection probabilities and the mortality proportions. This conceptual model has been validated by domain experts (from the GDG Company, Université de Sherbrooke and Université du Québec à Trois Rivières -UQTR) and was used to orient the development of the geosimulation tool. According to our conceptual model, the progression of the WNV infection involves a large number of individuals of two main species and their interactions depend on the probabilities of finding sub-populations of these species within the same geographic areas at specific times. We already mentioned the interest of using a multi-agent geosimulation approach in such a context. However, we had to adapt it to take into account the large geographic area of interest and the very large size of the involved populations, especially for Culex. To create the virtual geographic environment representing the studied region ( Figure 3 ), we first collected geo-referenced data and generated the various spatial data layers needed by the MAGS platform. Then, we modelled the two populations involved in the transmission of the virus as well as their locations in this virtual environment. Indeed, the population of Culex represents an extremely large number of individuals and cannot be represented using individual agents. Instead, we decided to model the mosquito population as an 'intelligent density map' which is characterized by population data being attached to reference areas (municipalities) in the virtual space. The idea is to associate to each reference area a list of variables corresponding to the numbers of the different categories of mosquitoes (larvae, healthy and infected adults) located in this place. These numbers evolve during the simulation as a consequence of various parameter changes (temperature, degree of humidity resulting from Representative species of Corvidae populations: Corvus brachyrynchos (American crow). Other birds are also considered. Representative species of mosquito's populations: Culex pipiens and Culex restuans. Main factors influencing the behaviours Culex Climatic conditions (primarily the temperature and precipitations). crows Zones and periodicity of displacements and grouping. The Ecumene zone of the following administrative areas: Québec, Chaudière-Appalaches, Mauricie-Centre du Québec, Montérégie, Estrie, Montréal-centre, Laval, Laurentides, Lanaudière et Outaouais. We decided to simulate the WNV propagation from July 1 until October 1. An interval of a daily step with a weekly assessment. The micro-space scale specifies the size of a pixel (4 km 2 ). The macro-space scale give an idea on the spread of the WNV propagation in order to bring a help to the decision-making concerning the larvicides treatments (interpolated by aggregation of pixels between 1 and 50 km 2 ). rainfall, etc.) as well as encounters with crows. For the population of crows, we used agents to model groups of crows associated with specific areas where roosts have been observed in the field. The interactions of the two populations have also been modeled thanks to the geosimulation which enables the system to automatically determine the places and times where groups of crows (pertaining to roosts) will cross areas in which the Culex sub-populations are located. The populations of Culex do not move much and they are present practically everywhere in the selected territory [3, 14] . Because of the extremely large number of individual mosquitoes, we represent sub-populations of mosquitoes as characteristics of the virtual geographic environment and we use what we call an "intelligent density map" that represents the distribution of Culex subpopulations over the different reference areas depending on the geographic characteristics and the locations of favourable habitats for mosquitoes. This intelligent density map is a kind of cellular automaton associated with rules that enable the system to simulate the evolution of the different categories of mosquitoes (larvae, adults, healthy, infected, etc.) in each reference area under the conditions that influence the mosquitoes' life cycle (temperature and precipitations). The system gets these conditions either from actual meteorological data (from specific databases: see Section 4.3) or from the parameters set in scenarios that the user wants to explore. This map contains the polygons representing all the municipalities of the Quebec province (the reference areas). On the user's screen, the color of each municipality changes according to the relative densities of the Culex populations (ratios of healthy, and infected adults) computed by the simulator (Figure 4 ). A roost synthesizes the behaviour of a group of crows. It is modeled by an agent having some initial characteristics such as the number of individuals, the position of the roost on the map, and the maximum radius of its expansion area. These characteristics are computed using various field data as presented in Section 4.3. Moreover, this agent inherits from all the functionalities of MAGS agents. For example, it uses some behaviour rules in order to model how crows scatter around the roost. In addition, an operating range parameter is computed for each roost in order to estimate the maximum distances covered by the crows when they search for food during the day. Each roost agent is implemented as a particle system [15] which simulates the way crows spread around a roost during the day. Hence, such particle systems behave as agents, as described above. Each particle represents one or several crows, depending on the number of individuals attached to the roost. In Figure 5 we can see a snapshot of a simulation in which roosts are displayed as "clouds of blue particles". Each particle has different characteristics (velocity, direction of movement) that enable it to travel at a distance from the roost location during a number of simulation steps representing a day. Hence, the set of particles associated with a given roost covers a circular area with a maximal radius set by the operating range parameter. We calibrated the parameters of the particle system by computing the density of crows in the area covered by the expansion of the roost and comparing it to observed field data and other estimated data. The interaction between the two main populations is a very important functionality of our system, since it is the way of representing the evolution of the infection. Indeed, while traveling in the geographic space, one or several crows represented by a particle can cross areas in which Culex mosquitoes are located. Consequently, there is a probability that some of these crows will be bitten. Technically, in order to determine the probabilities of encounters between mosquitoes and crows, the corresponding particle takes into account the characteristics of the Culex population associated with each reference area of the 'intelligent density map' over which it travels. Therefore, the system can estimate the number of infected individuals, based on the likelihood that a number of individual crows be bitten by mosquitoes and be infected with WNV (using the equations of the mathematical model described in Section 4.2). Moreover, the user can visualize the extent of the spread of WNV on the map in different ways. The system can either change the color of the particles representing the infected crows or the color of the polygon representing a municipality containing a high density of infected Culex. Our initial simulations involved the two main species of American crows and Culex pipiens/restuans that we selected and that enabled us to apply Wonham's mathematical model [16] (see Figure 12a in section 4.2). We quickly found out some limits with this model, since it does not take into account the influence of temperature on the evolution of mosquito's populations. When we applied this model, it led, after a number of iterations, to the complete "extinction of crows". This, obviously, does not conform to reality, although a dramatic decrease of Corvidae popu- The geographic area of interest which contains all the municipalities belonging to the ecumene (Southern Quebec, Canada) Figure 3 The geographic area of interest which contains all the municipalities belonging to the ecumene (Southern Quebec, Canada). lations have been observed in recent years due to the spread of WNV [17] . Hence, we proposed an extension of this model which enables us to model several species of birds and to take into account the impact of the temperature in terms of cumulated degree-days which influence some parameters of the model (see Figure 12b in section 4.2). We cannot discuss here the details of such a model and its implementation (for more details, see [18] ). In the current experiments, we modeled the interactions between crows and mosquitoes as described in the previous section. Since surveillance systems provide data about crows as indicator birds, we used this species to set the simulation parameters and to calibrate the system. However, we added other bird species in the simulation to increase "the biting opportunities" for mosquitoes, so that the "crow popula-tion" does not become extinct by the end of the simulation period. Indeed, this is a plausible hypothesis: mosquitoes bite other birds as well as crows. We thus introduced in the WNV-MAGS system another 'global' population of birds, that we called "generic birds" (Common Raven, Blue Jay, American Robin, House Sparrow, European starling and Mourning Dove) which are resident in the municipalities and known to carry WNV [17] . This population of generic birds appears in the mathematical model with similar equations as those used for the crows (each bird population is represented by a different index j in the equations: see Figure 12b in section 4.2). However, the parameters for each bird family may be different. Due to lack of data, we currently set some average parameters to the equations of the "generic birds". Getting more accurate parameters will require further Using an intelligent density map to represent the population of Culex Figure 4 Using an intelligent density map to represent the population of Culex. research from bird specialists. In the simulation, one distinction that we established between crows and "generic birds" is that we assumed that birds do not move outside the municipality (as crows may do while flying away from the roosts). Hence, generic birds stay in contact with the same mosquito population during the simulation. Indeed, this is a simplification. Since our system is parameterised, we will be able to introduce parameters for other bird species as soon more precise data will be available with respect to the ecology and epidemiology of other birds affected by WNV. Using various short-term climate scenarios In our system, multi-agent geosimulation is at the heart of a decision support tool. Hence, our approach is somewhat different from more traditional simulations used for prediction purposes [19] . The WNV-MAGS System simulates the WNV epidemics and enables a user to specify scenarios in order to explore various situations including climate change and different intervention strategies. The user may choose one among four different scenarios which influence the dynamics of the Culex population ( Figure 6 ). The first scenario is the default scenario which can be set in order to use average conditions of temperature and precipitations (using in this case the Canadian Climate Normals [20] ). In order to estimate the number of mosquitoes located in each municipality, we computed the number of sumps that are along the municipality's roads (see section 4.3). Sumps offer ideal locations for the maturation of larvae and the emergence of adult mosquitoes. They are also the main targets of larvicide spraying. But abundant rains may flush sumps, killing a large proportion of larvae. In a second type of scenario, the user can choose a date during which abundant rains may flush Using roosts to represent the populations of crows Figure 5 Using roosts to represent the populations of crows. sumps in some municipalities (Figure 7 ). In the same way, the third scenario is used to simulate the use of larvicides in a certain area (municipality). The last scenario is a combination of the second and third scenarios. Hence, it is possible to choose a date for the flushing of sumps and another date for the application of larvicides. Most larvae are supposed to die after the flushing of a sump, although the dynamics of the larval populations starts all over again since there are always Culex adults in the vicinity of the sump that will spawn new eggs. Moreover, the WNV-MAGS System offers a variety of functionalities to the user in order to modify the parameters of the mathematical model, to visualize the progress of the infection in and around the crows' roosts, to extract data from the simulation and to generate graphs showing the evolution of the involved populations. The qualitative results of the model which represent the distribution of the populations were satisfactory. Indeed, the resulting curves reflect the biological behaviours of the studied species according to the opinion of the consulted domain experts (from GDG and UQTR). However, the quantitative data needed to be calibrated in order to be used in real-life situations. In fact, we calibrated the model by comparing simulation results and field observations (ISPHM-WNV data [4] ). We evaluated the ratio between the real populations of mosquitoes and the samples of mosquitoes captured in traps (absolute densities) as well between crows and the collected dead crows. Regarding the populations of Culex, we used Reisen's work [21, 22] to estimate the mosquito density ratio. A Using management scenarios Figure 6 Using management scenarios. The dynamics of the larval populations before (a) and after (b) the flushing of sumps in Laval Municipality on August 15 th (hypo-thetical scenario defined by user) Figure 7 The dynamics of the larval populations before (a) and after (b) the flushing of sumps in Laval Municipality on August 15 th (hypothetical scenario defined by user). captured mosquito was considered to represent a population of 300 Culex over one km 2 . Since we did not have data for all regions, we only calibrated simulation results for some key municipalities where human infections had occurred. It appeared thereafter that there was a significant difference between the data generated by the model and those obtained from the field. Hence, we tuned up the initial settings of the simulation (e.g. the initial percentage of infected Culex or infected crows, distance between sumps, emerged Culex per sump, percentage of sumps containing larvae, etc) as well as some parameters of the mathematical model (e.g. mosquitoes biting rate of crows per capita, WNV transmission probabilities from Culex to crows of from crows to Culex, etc). These changes have helped us to quantitatively calibrate the model for the processed municipalities. Figure 8a presents the evolution of the total number of mosquitoes for the municipality of Laval between July 1 st and October 1 st . The smooth blue curve represents the data generated by the simulation while the rugged red curve represents averages of real data over four years (2003 to 2006) . We had to consider these averages because we do not have sufficient data from the field (trap measurements are sparse and not carried out regularly in Quebec municipalities). Moreover, these data averages enabled us to adjust our initial data in the simulation (mainly the initial number of mosquitoes) as it can be observed in Figure 8a . The simulation curve and the real data curve fit nicely between July 1 st and August 15 th . The two big drops that are observed in the real data curve are difficult to explain at this point since we consider the average measures over four years. This may be the result of systematic larvicide applications in July and August (3 applications in some municipalities during a WNV season), but we have no sufficient data to confirm this conjecture. In figure 8b , we also observe a similarity between the curves representing the infected mosquitoes. Again, the rugged red curve represents averages of real data over 2003-2006. All drops in the curve result from lack of sufficient field data. Regarding the populations of crows, we used the results presented by David and colleagues [23] in order to determine whether the numbers of dead birds sighted and tested for WNV are representative of the true bird mortality. We also used the index trend obtained from the ÉPOQ database [24] and from the North American Breeding Bird Survey [25] to adjust the population of crows as well as the population of generic birds. Moreover, changes in the population of crows have been calibrated using field data collection of dead birds and their analyses in the laboratory, as it was done for the population of Culex. Model calibration using the average total mosquitoes cap-tured in traps (a) and those among them which are infected with the WNV (b) during the considered simulation period represents a comparison of the simulated data (smooth blue curve) and real data (red rugged curve) for the collected dead crows. The general shapes of the curves are similar. This is encouraging since data available for dead birds are even sparser than for mosquitoes. Then, we looked at real data for Laval Municipality for 2003, the year for which we have the most complete data set. We created the temperature scenario for 2003 and launched the simulation. Figure 10a presents the difference between the simulated data (blue smooth curve) and the real data. In order to explain this difference, we checked with the SOPFIM Company if larvicides had been sprayed in Laval Municipality in 2003. It was indeed the case, with interventions on June 18, July 17, and August 13. We created a new scenario using these three dates for larvicide spraying and we got the curve displayed in Figure 10b . The curve of simulated data now approximates the real data fairly well (the rugged curve of real data being again explained by missing data). This is an encouraging result showing that the parameters adjusted for calibration provide reasonable results. In order to improve precision and validate our models and the simulation parameters, we will carry out the simulation on a different data set. We are currently collecting data (for mosquitoes and crows) for the city of Ottawa. We expect to get a more complete data set, since measurements have been more frequent and regular in the Ottawa region (Canada) over the past 6 years. This work is in progress. In this paper, we presented a system that simulates the interactions of the populations of mosquitoes and birds which are involved in the propagation and transmission of the WNV. Moreover, we used a multi-agent geosimulation approach which takes into account the influence of the geographic characteristics of the various regions, thanks to the use of GIS data. For example, we determined the geo-referenced co-ordinates of crows' roosts in order to locate them on the map and we were able to develop rules which control the expansion/contraction of roosts over space and time. We also pre-processed climate data in a GIS in order to feed it to the simulation. We also used the geographic characteristics and the location of favourable habitats for mosquitoes in order to represent the populations of Culex using an intelligent density map. Consequently, we were able to implement the interactions between the mosquitoes' and birds' populations which can cause an outbreak of the virus and epidemic propagation of the disease. Even if other works [26, 27] also used GIS data to simulate the spread of the WNV, they did not offer a decision support system as we do. In contrast, our system enables users to simulate the propagation of WNV under various short-term climate scenarios and allows for local parameterization. This approach may be useful for practical decision-making. For instance, it has been shown [28] that the number of degree-days below -5°C in the winter and the number of degree-days greater than 25°C in the summer may contribute to a highly epidemic emergence of the virus during the summer under specific climatic conditions. Consequently, our system may be used to predict such an epidemic if we simulate the propagation of the WNV using a scenario in which seasonal forecasts of climatic data are favourable for the emergence of the virus [29] . By assessing the simulation results and comparing the outcomes of different intervention scenarios, the users of the WNV-MAGS System can make more informed decisions about the actions to be taken such as the application of larvicides or the stepping up of personal protection measures. An important limit of this kind of approach is the lack of field data. As we have already shown in this paper, a good calibration and validation of the models depends on the availability of a large variety of data sets (related to mosquitoes and to different species of birds). There is also a difficulty in estimating the parameters needed in the mathematical model, which would require in some cases that additional field studies be carried out by entomologists and ornithologists. In addition, the potential effects of changing resistance and immunity in wild birds remain unknown and need to be studied by domain experts. Obviously, they have not been included in our models yet. If we were able to collect sufficient data about the WNV spread in different regions during the past years, we could develop scenarios and simulations whose results could be compared to recorded field data. Consequently, we would be able to further validate the system and adjust the various parameters that are used for the simulations, taking into account the specificities of the considered regions and species. Nonetheless, the system can already be used to compare different scenarios involving variations of the climatic data in relation to the potential spread of WNV in particular regions. As we have shown, the system can also be used to estimate the influence of human intervention based on larvicide application. However, since it still remains difficult to get accurate weather forecasts over long periods (several months), public health authorities will have to take into account this inherent limitation of meteorological science when developing intervention plans using such a tool. Our MAGS approach and tool can be used not only to simulate the propagation of the WNV, but they can also be adapted to various other vector-borne diseases. We are currently working on the simulation of Lyme disease in Quebec. Moreover, the tool and approach can be extended to take into account the specificities of other similar diseases (e.g. SARS) in other geographic areas. In order to develop the WNW-MAGS System, we applied an 'Agile' [30] analysis and design method which favours the collaboration with domain specialists and users, as well as quick adaptations of the software under development. We also applied classical knowledge engineering techniques [31] in order to acquire domain knowledge from the specialized literature and from domain experts (entomologists and ornithologists) after many work sessions. We then went through an exploration phase of the field by gathering the maximum useful information in order to understand all the phenomena which are related to the spread of WNV. We present in this section the conceptual architecture which is used as a basis of the simula-tion system. We also present the mathematical model which was chosen and adapted in order to compute the dynamics of the populations involved in the transmission of the WNV. We also present a subset of the relevant data which is used to feed the system. Finally, we present the implantation of the system. Based on the requirement specifications and using the conceptual model of the phenomena (see Figure 2 in section 2.2), we designed a conceptual architecture of the WNV-MAGS system which includes all the needed system components and their relationships [32] . While constructing this architecture, we identified all the processes to be developed (represented as rectangles numbered as Pi in Figure 11 ) and all the data stores (represented as ovals numbered as Ai in Figure 11 ) that gather data and feed the system processes. Indeed, most of the necessary data are obtained from external databases (EPOQ [24] , Weather data, etc.) and GIS. They are represented as 'cylinders' at the bottom of Figure 11 . The architecture is divided into four parts. The first part (processes P1 to P4) deals with data preparation, including the extraction of data from all the required databases. The second part (processes P5 and P6) computes the evolution of populations using the mathematical model presented in Section 4.2. The third part (processes P7 and P8) deals with the interactions between the sub-populations of crows and the sub-populations of Culex. It reflects the interactions between the agents' roosts and the intelligent density map. The last part (processes P9 to P12) is responsible for the management of scenarios, as well as for the display, analyses and calibration of the results. We used different databases (as presented in Section 4.3) in order to initialize the populations of Culex and crows at time t 0 (beginning of the simulation). We used these populations to compute the dynamics of crows, the dynamics of Culex and the interactions between the two populations while increasing the time by one step, at time t. Then, we get the new populations at time t+1 (taking into account the state changes of the sub-populations of mosquitoes and crows reflecting the infection spread and deaths) and the system triggers again the same processes to simulate the joint evolution of the two populations. The simulation results can then be displayed on the map. Domain specialists can also calibrate the simulation using the WNV surveillance data (that we have pre-processed). The user can manipulate the results and create various scenarios. Then, the simulation results can be assessed and compared. We need to compute the evolution of the populations of Culex and the populations of crows in order to simulate their interactions using the geosimulation system. To this Conceptual architecture of the system (based on EPAS method [32] ) Figure 11 Conceptual architecture of the system (based onEPASmethod [32] ). end, we selected the model proposed by Wonham and colleagues [16] to compute the dynamics of the two populations. This model is based on 8 differential equations (Figure 12a ) which can compute over time the evolution of the different categories of individuals (called 'compartments'): susceptible,infected, recovered and dead birds, the larvae of mosquitoes and the susceptible, exposed and infected adult mosquitoes. We proposed some modifications in order to correct some discrepancies that we found in the model. We also included climate effects in the model using the work of Madder and colleagues [33] as a starting point. This was a difficult task because the model was no longer in equilibrium and this required several modifications to the differential equations [18] . Figure 12b presents an overview of the new equations of the proposed model. We notice that the birds equations (dS BJ /dt, dI BJ /dt, dR BJ /dt, dX BJ /dt) have an index j which represents a different bird species that we want to include in the simulation. Climate effects are computed using another set of equations that is not presented in this paper. These equations modify certain parameters in the differential equations of Figure 12b . For example v xL in the dL M /dt equation represents the rate of progress of larvae toward the state of nymphae, this rate depends on temperature conditions defined in a different set of equations. The adjusted model gives satisfactory results in terms of quality (e.g. distribution of the mosquitoes' generations). Indeed, the pace of the established curves reflects the biological behaviours of the studied species if we refer to the specialized literature. However, the quantitative results (e.g. the number of larvae, eggs, emerged Culex, dead crows, etc.) had not been conclusive with the first results of the simulation. We corrected this problem with the calibration of the system as presented in Section 2.4. We already mentioned that our approach takes advantage of GIS data in order to properly locate the agents' roost in space. Indeed, we used the Geomedia GIS software in order to handle the geo-referenced data of the DMTI Spatial (CanMap Streetfiles) and the digital maps of INSPQ. Using this data, we created the bitmap from which the MAGS platform generates the simulation environment. This bitmap contains polygons representing a list of 945 municipalities (out of a total of 1476 for the whole province of Quebec) being part of the ecumene (inhabited part of the studied region) of the geographical area of interest. Moreover, this bitmap is also used by the intelligent density map presented earlier (see Figure 3 in section 2.3). In addition, we had to pre-process all the data needed to feed the system (see processes P1 to P4 of the conceptual architecture in Figure 11 , section 4.1). We estimated the initial populations of Culex and crows at the beginning of the simulation. For the population of crows, we used the SAS statistical software and the MapInfo GIS to compute a specific density of birds per region (number of individuals by square kilometer). This done by estimating an average of the sighting mentions provided by professional or amateur ornithologists (from 1997 to 2005 and located inside the ecumene) using the ÉPOQ database [24] (Table 2) . We also processed the data relative to crows' roosts (obtained through an email survey involving expert birders or extracted from the ÉPOQ database [24] ). This data included the co-ordinates (latitude/longitude) as well as an approximate number of individuals for each mentioned roost. We computed an average of the number of individuals in the case of several records of the same roost (same latitude and same longitude). This data is used to characterise the roost agents (Table 3) . Moreover, we studied the literature and discussed with domain experts in order to collect relevant information about crows' behaviours. We used this information in order to specify some behavioural rules for the roost agent such as the way that particles spread around the roost to simulate the birds' daily search for food. Considering the Culex population, we computed the number of individuals of the initial population, estimat- ing the number of adults that emerge from the larvae laid down in sumps (which we supposed to be the main reservoirs of mosquitoes in urban and sub-urban areas). To this end, we developed a Visual Basic application in order to query the geo-referenced databases in Geomedia and to compute the total length of roads for each municipality of interest. We then computed the number of sumps in each municipality by using the total length of roads (Table 4 ). We assumed that there is an average of one sump for each 30 linear meters of road. This average distance conforms to the standards used by the Ministère des transports du Québec (MTQ) when they install sumps. However, the user of the WNV-MAGS system can modify this value, since it can vary in relation to the considered regions such as urban and rural areas. We also assumed that there are only 20% of sumps containing larvae. This default value can also be modified by the user as well as the average number of Culex adults which emerge from a sump at the beginning of a simulation. These numbers were obtained by consulting data provided by the SOPFIM Company [34] in relation to the monitoring of larval and adult Culex in certain Quebec regions during the summer of 2005. Furthermore, we used a DLL which enables us to integrate the climate data into the system. This DLL represents one of the functionalities of the BIOSIM software [35] . We used it to interpolate values for temperature and precipitations at certain precise locations on the territory, taking into account the data of the four neighbouring weather stations and the elevation data. This computation can be done using either real data or the Canadian Climate Normals [20] which are produced over several years. After the data preparation, we implemented the system using the MAGS platform which is developed in C++. This simulator includes several modules performing various tasks. It contains a controller to manage the threads of application (Processes Module), a user Interface and a module managing the simulation environment (Environment Module). In addition to these modules, MAGS contains a module in charge of data display, modules to specify agents' populations and agents' behaviours, and a module simulating particle systems. In order to simulate the propagation of the WNV, we added a module which simulates an epizooty (Epizootic Manager) which has been developed in a generic way in order to be easily extended to simulate epizooties different from WNV (Figure 13 ). The authors declare that they have no competing interests. Various components of the technical system architecture (the epizootic module was added to the pre-existing MAGS components) Figure 13 Various components of the technical system architecture (the epizootic module was added to the preexisting MAGS components).

Alternative medicines for AIDS in resource-poor settings: Insights from exploratory anthropological studies in Asia and Africa

The emergence of alternative medicines for AIDS in Asia and Africa was discussed at a satellite symposium and the parallel session on alternative and traditional treatments of the AIDSImpact meeting, held in Marseille, in July 2007. These medicines are heterogeneous, both in their presentation and in their geographic and cultural origin. The sessions focused on the role of these medications in selected resource poor settings in Africa and Asia now that access to anti-retroviral therapy is increasing. The aims of the sessions were to (1) identify the actors involved in the diffusion of these alternative medicines for HIV/AIDS, (2) explore uses and forms, and the way these medicines are given legitimacy, (3) reflect on underlying processes of globalisation and cultural differentiation, and (4) define priority questions for future research in this area. This article presents the insights generated at the meeting, illustrated with some findings from the case studies (Uganda, Senegal, Benin, Burkina Faso, China and Indonesia) that were presented. These case studies reveal the wide range of actors who are involved in the marketing and supply of alternative medicines. Regulatory mechanisms are weak. The efficacy claims of alternative medicines often reinforce a biomedical paradigm for HIV/AIDS, and fit with a healthy living ideology promoted by AIDS care programs and support groups. The AIDSImpact session concluded that more interdisciplinary research is needed on the experience of people living with HIV/AIDS with these alternative medicines, and on the ways in which these products interact (or not) with anti-retroviral therapy at pharmacological as well as psychosocial levels.

A large number of new treatments offered to people living with HIV/AIDS (PLWA) have appeared over the last fifteen years in the therapeutic domain of AIDS. These med-icines are particularly heterogeneous, both in their presentation and in their geographic and cultural origin. They constitute a group of products with a therapeutic aim that occupies a space between the customary traditional, popular and biomedical sectors of health care [1] . These products often mix reference to biomedicine and science with notions of traditional health culture and nature in a syncretic way. They consist mainly of herbs and nutritional substances and are packaged as 'modern' pharmaceuticals: capsules, tablets, and solutions. The names of these alternative treatments reflect their reference to biomedicine: Immunocomplex, Viralgic, Virjint, etc. Their accompanying leaflets provide detailed information on substance, as well as dosage, indications, and biomedical efficacy claims. Their diffusion follows contemporary paths in the global economy and makes use of new information technologies. In this paper, we will use the term "alternative" to consider a generic category including medicines that recently appeared for AIDS which have not been authorised by drug regulatory authorities, nor recommended by WHO. Other terms, such as neo-traditional or neo-phytotherapeutic, may be discussed for the characterization of some of these treatments, related to their local meanings or their social status. The emergence of alternative medicines for AIDS in Asia and Africa was discussed at a satellite symposium and the parallel session on alternative and traditional treatments of the AIDSImpact meeting, held in Marseille, in July 2007. We were especially interested in the role of these medications since the introduction and rapid scale-up of highly active anti-retroviral therapy (HAART) in resource poor settings. Twenty anthropologists and health researchers attended the satellite session and presented exploratory findings from Asia and Africa (Uganda, Senegal, Benin, Burkina Faso, China and Indonesia). The aims of the satellite, the results of which were presented at the parallel session [2] , were to (1) identify the actors involved in the diffusion of these alternative medicines for HIV/AIDS, (2) explore uses and forms of these medicines, and the way they are given legitimacy, (3) reflect on underlying processes of globalisation and cultural differentiation, and (4) define priority questions for future research in this area. We present here the insights generated at the meeting, illustrated with some findings from the studies that were discussed. There has been an increased professionalisation and commercialisation of traditional medicine in response to the development of biomedicine. This trend is not specific to AIDS and not necessarily a recent development. Social scientists first noted this trend in the late 1980s: Charles Leslie [3] for example has shown how, in India, in response to an increased authority of biomedicine and the globalisation of health markets, Unani and Ayurvedic medicine production changed; and Afdhal and Welsch [4] described the rise of 'modern' jamu in Indonesia. Jamu is the traditional term for Indonesian indigenous medicines usually prepared from herbal medicines such as leaves, bark, roots and flowers. Nowadays a multimillion dollar industry is involved in the production of Ayurvedic and Unani medicines in India, and of jamu in Indonesia. A rapidly expanding assortment of powders, creams, pills, capsules and cosmetics has been manufactured both in small cottage industries as in large factories with increasingly sophisticated technologies [3, 4] . The modernization of the manufacturing of these drugs has been accompanied with more modern biomedical modes of presenting their efficacy [5] . Under globalization, similar trends occurred in other regions and these products diffused more rapidly. At the seminars in Marseille, we discussed the ways in which such alternative remedies operate in the therapeutic domain of AIDS care. In the first decade of the AIDS epidemic there was no effective treatment for HIV/AIDS and patients were faced with nearly certain premature death. At that time, there were regular hypes offering hope for life. But with the introduction of ART, alternative treatments are now marketed for many additional purposes too: to prevent AIDS, to kill viruses, to delay the need for ART, to restore and enhance health while on ART, to treat opportunistic infections, and to alleviate adverse side effects of other treatments. Biomedical practitioners generally discourage the use of alternative medicines, fearing interactions with ART and also through the concern that patients may stop using ART. At the AIDSImpact sessions Egrot and colleagues [6] presented findings on the supply of what they label "neo-traditional medicines" to refer to the boundary-crossing nature of these treatments in West Africa. The "designers" of the inventoried products are extremely heterogeneous. In some cases these people are nationals of African countries who present themselves as healers. Some say they have undertaken "research" on the basis of therapeutic products that were already known locally. Others refer to a dream revelation (classic in the universe of healers in Africa) of a plant composition that is "efficacious" against AIDS, while yet others speak of a divine revelation. Physicians, scientists and academics are solicitated, brought into involvement or spontaneously engage themselves in the exploitation of neo-traditional products. The case studies in West Africa show that other treatments, such as Immunicomplex or Aloe Vera, originate in Europe and the USA. Alternative medicines from Europe and the USA occupy the same shelves in ordinary pharmacies as those originating from Africa and China, often along with a few 'immune-boosting' food products (honey, olive oil). Specialized "bio", "natural health" and "health food" shops make these products available to the more affluent. The distributors and marketing men of these products also target health workers and clinics directly. The West Africa case studies noted that health workers also have started to prescribe alternative products such as Immuboost (NHi2T) or Viralgic (Pharma Concept) (see Figure 1) . A case study from Uganda showed how health workers operating an anti-retroviral treatment program adopted a locally available traditional ointment as an alternative medication for skins problems of people living with HIV and AIDS. The skin problems result from adverse effects of ART or symptoms of opportunistic infections. The health workers obtained the recipe from local traditional healers (patients had told them that the cream works well), and the patients help collect the ingredients. They 'repackage' this traditional remedy into what is now called 'mobile cream' (to make clear it is produced by the so called 'Mobile' ART program). One of the nurses reports: "The mobile cream, which we ourselves prepare either at our chief nurse's home or here at the office depending on how busy we are at the office, is very efficacious for many kinds of skin related conditions. We are quick to prescribe it to the patients because we know it works and it is popular among patients too because it works for them [7] ." Content analysis of drug information leaflets, advertisements, product catalogues, and brochures distributed by medical representatives in the West Africa case studies [6, 8] casts light on the range of effects that are attributed to these drugs. Most commonly cited (biomedical) properties are immune-stimulation and antioxidant. Some manufacturers suggest that the products have antiviral properties as well. The antiviral dimension refers either to the opportunistic infections such as herpes (mentioned for example in the product information for Immuboost) or eventually to the immunodeficiency syndrome itself. Indeed, some products boldly claim anti-HIV activity as well, and are marketed as natural ART (see Figure 2 ). However, such efficacy claims are not static. The producer of Virusinest (Nesto-Pharma) recently withdrew the antiviral claim, stating in its information leaflet: "the analyses carried out among patients do not allow the anti-HIV assertion to be upheld". There may be also inconsistencies between various information sources. The brochure for Viralgic (Pharma Concept) says that this is a product which renders the virus undetectable, but the website of the manufacturer presents the drug as immunostimulant (result of trials published on the web site), and present the product as treatment for opportunistic infections: "antiherpes...for healthy persons". A case study on Indonesia [9] dealt with the demand for alternative medicines among PLWA. As Afdhal and Welsch noted two decades ago, Indonesia has a thriving market for jamu. Jamu are sold for a wide range of indications: common colds, influenza, headaches, aches and pains, high blood pressure, beauty, improvements in sexual performance, and recently to treat and prevent HIVrelated health problems. AIDS prevalence is below 1% (i.e. this is a low prevalence area), but the disease is stigmatised, because of its association with intravenous drug use and prostitution. Hardon and her colleagues conducted interviews with women and men who live with HIV and use anti-retroviral therapy, mainly intravenous drug users and their partners. All of them had better health since taking these modern drugs. Nonetheless, all of them see the need to take jamu as well. They do so in Tobacoak's, West Africa part out of their intention to live positively (i.e. eating and sleeping well, and keeping a positive outlook on life), as promoted by many of the support groups in which PLWA participate. The respondents do not make distinctions between modern medicines and jamu in these health maintenance and restoring practices. Rather they distinguish the drugs by their effects. They use popular jamu to treat side effects of HAART, such as itchiness. These jamus are not specifically promoted for HIV and AIDS in Indonesia, perhaps partly because the disease is so stigmatised. However one neo-traditional preparation stood out in the narratives of our respondents as a product which can treat HIV/AIDS: virgin coconut oil. Ceri, for example, started using coconut oil shortly after she found out she was HIVpositive. She says: Actually, the effect is not only for your immune system. So, I feel better, don't feel tired, and have more energy. I think what influences most is self-suggestion. It's self-suggestion that matters... Mia (a 28 year old woman from Jakarta) was given virgin coconut oil by a friend from Yogjakarta: I got 70 boxes. A box contains 60 capsules. It took it every day until I felt sick, but there was no effect. My CD4 level did not increase. Three months, three months made me look like a coconut you just needed to squeeze (laughing). I became very oily. The good effect when you take VCO is that your skin is silk smooth, your face is fairer and if you take a shower, you don't need any lotion, because your skin is naturally oily. That is the positive effect. Your hair is also stronger. But Buli, a 29-year-old ex-drug user from Karawang, one of the most active members of the support group in Karawang says: In Indonesia, the drug sellers were not very willing to discuss the effects of VCO. They would acknowledge that indeed these drugs are used by PLWA, or they would deny knowing anything about the drugs. But their pharmacies are full of advertisements for the products and they have prominent positions on their shelves (see Figure 3 ). Content analysis of the package information for VCO in Indonesia revealed that they are marketed as real 'curealls', i.e. to kill viruses and bacteria and/or strengthen the immune system, efficacy claims that we also found in West Africa. For example the package leaflet for Vicofit (manufactured by Sumber Dinamis in Bogor) states that the drug has "a high content of lauric acid which has antivirus, anti-bacterials and anti-protozoa properties." And that it is "believed to help improve the health condition of those with cholesterol, diabetes, coronary heart disease, hepatitis C, HIV positive, cancer, prostate, uric acid, osteoporosis, influenza and weight problems". The package for Virjint (produced by PT Vermindo International) states that the medicine is safe for daily use and without side effects. It lists two dosages: one for prevention (2 × 2 capsules per day) and another for treatment (2 × 3 capsules per day). The leaflet stipulates that the indications are: -"to increase energy and body stamina -to increase body resistance (Meningkatkan daya tahan tubuh) against bacterial, viral and fungal pathogens -to reduce weight -anti-oxidant, anticancer, and anti-HIV -to overcome uric acid, hypertension, stroke, heart disease, atherosclerosis, osteoporosis, influenza, hepatitis, chickenpox, herpes, TB, diabetes, epilepsy, eczema, liver, haemorrhoids, kidney, peradangan (burning sensation), infection, degenerative disease." The packages cite clinical research conducted elsewhere (Philippines, USA) to give legitimacy to the products. For example the leaflet of Holistic Virgin Coconut Oil states: "Based on research conducted in the Philippines, Holistic Virgin Coconut Oil is very effective to fight against SARS and AIDS". One of the key characteristics of alternative medicines in Asia and Africa is that they move from one cultural and geographic space to another, apparently without being constrained by trade-barriers, or regulatory mechanisms. Some governments promote the production and diffusion of neo-traditional medicines. They do so for economic reasons: alternative medicines are big business, but they also do so for ideological reasons: neo-traditional medicines reflect an attractive hybrid of modernity and national heritage, providing a sense of national identity in the globalized health economy [10] . The governments of India, China, Indonesia, and some African countries support research programs to further advance these neo-traditional products, and facilitate market diffusion. While registered pharmaceuticals are regulated heavily upon market entry (proof of efficacy is assessed by national drug regulatory authorities), this is not the case for alternative medicines. ART programs, which are sponsored by the same governments, usually discourage the use of alternative medicines, fearing the toxicity of the drugs, or that these medicines will interact with anti-retroviral medication and lead to discontinuation of ART therapy [11] . Governmental agencies may have contradictory attitudes towards the use of alternative medicines for AIDS, discouraging it within ART programs and supporting it within divisions of traditional medicine. An exception is the Chinese government, which officially supports a complementary medicine program for AIDS care and research [12] . Mass-produced alternative medicines meet an increasing demand for health products, a trend which has been labelled the "commodification of health" [13, 14] : from the slums of Djakarta to rural settings in Burkina Faso, people believe more and more that they need pharmaceutical 'things' to protect their health and to treat illness symptoms. People living with HIV and AIDS are particularly uncertain about their health and their future: ART may be accessible and improve health now, but they wonder if this will be the case in the future. This uncertainty makes them an attractive market for the 'best of both worlds', alternative medicines, which come with assertions of 'natural' safety and 'biomedical' efficacy [15] . However the case studies presented in Marseille suggest that people especially want to use alternative medicines to delay onset of ART, treat opportunistic infections, restore health and alleviate adverse effects once on ART. Immuneboosters are popular, though our case studies suggest that PLWA are often ambivalent about alternative medicines that claim anti-HIV efficacy. The case studies make clear that the market of alternative medicines for HIV/AIDS is dynamic. It adapts to progress in biomedicine, which has produced potent anti-retrovi-ral medications. In some cases, the efficacy claims for alternative medicines reinforce a biomedical paradigm for HIV/AIDS, and fit with a healthy living ideology promoted by AIDS care programs and support groups. More interdisciplinary research is needed on the experience of people living with HIV/AIDS with these alternative medicines, the ways in which the products and their representations move from one cultural setting to another, and on the ways in which these products interact (or not) with anti-retroviral therapy at pharmacological as well as psychosocial levels. More research is also needed to assess the economic impact of these therapies, since people seem to be spending much on these 'other' medicines while ART is provided for free. A blanket denial of the relevance of these products for the quality of life of PLWA does not make sense for patients, who need precise information that make clear which products are likely to have negative interactions with ART, and which ones could be beneficial. Unfortunately research on the interactions between alternative medicine and antiretroviral drugs is sparse [11] . To be able to inform patients better, more clinical research is needed on the benefits and risks of those alternative medicines that are perceived to be beneficial by people living with HIV and AIDS.

Evolutionary and Transmission Dynamics of Reassortant H5N1 Influenza Virus in Indonesia

H5N1 highly pathogenic avian influenza (HPAI) viruses have seriously affected the Asian poultry industry since their recurrence in 2003. The viruses pose a threat of emergence of a global pandemic influenza through point mutation or reassortment leading to a strain that can effectively transmit among humans. In this study, we present phylogenetic evidences for the interlineage reassortment among H5N1 HPAI viruses isolated from humans, cats, and birds in Indonesia, and identify the potential genetic parents of the reassorted genome segments. Parsimony analyses of viral phylogeography suggest that the reassortant viruses may have originated from greater Jakarta and surroundings, and subsequently spread to other regions in the West Java province. In addition, Bayesian methods were used to elucidate the genetic diversity dynamics of the reassortant strain and one of its genetic parents, which revealed a more rapid initial growth of genetic diversity in the reassortant viruses relative to their genetic parent. These results demonstrate that interlineage exchange of genetic information may play a pivotal role in determining viral genetic diversity in a focal population. Moreover, our study also revealed significantly stronger diversifying selection on the M1 and PB2 genes in the lineages preceding and subsequent to the emergence of the reassortant viruses, respectively. We discuss how the corresponding mutations might drive the adaptation and onward transmission of the newly formed reassortant viruses.

The H5N1 highly pathogenic avian influenza (HPAI) virus was originally isolated from a farmed goose in Guangdong province of China in 1996 [1] , and soon spread to live-poultry markets in Hong Kong [2] , resulting in 18 cases of human infection in 1997, 6 of which were fatal [3, 4] . The first wave of H5N1 infection ceased after the depopulation of all poultry in Hong Kong, although the H5N1 virus was later found to circulate continuously in Southern China without causing apparent disease symptoms among infected poultry [5] . H5N1 outbreaks recurred in 2003, persistently affecting poultry farms in many Southeast Asia countries, such as China, Thailand, Vietnam, Indonesia and Cambodia. The viruses also spread outside Asia, including to some European countries. More importantly, occasional zoonotic transmissions to humans occurred in most of the affected Asian countries and the virus continued to pose a serious threat to global public health [6] . H5N1 outbreaks in Indonesia were initially detected in poultry farms in December 2003 [7] . It was suggested that the H5N1 virus was first introduced to Java and subsequently spread to other parts of the country [8] . The virus rapidly became endemic in Indonesia [9, 10] , and continued to cause sporadic zoonotic transmissions to humans beginning in July 2005 [9] . Three clusters of H5N1 transmission among family members were identified in 2005, raising concerns of possible human-to-human transmission of the virus [11, 12] . As of April 8, 2008 , Indonesia had 132 confirmed human cases with 107 deaths [13] , the largest number of deaths among all affected countries. Previous studies have shown that several H5N1 genotypes have emerged in Asia through reassortment between H5N1 viruses and other subtypes [14, 15] . One of these genotypes, Z, predominated the H5N1 outbreaks throughout 2003-2007, causing most H5N1 outbreaks in Asian countries, including Indonesia [16] . Moreover, a variety of antigenically distinct sublineages of Z genotype virus have been established [16] . Unlike Vietnam and Thailand, Indonesia was invaded by only a single sublineage of genotype Z virus. Previous phylogenetic analyses suggested that Hunan province of China may be the source of the initial H5N1 outbreak in Indonesia [17] , and classified the Indonesian H5N1 HPAI viruses into three groups [10] ; however, further statistical analysis is necessary to characterize and compare different aspects of their evolutionary histories. In this study, we examined molecular phylogeny of the most recent Indonesian H5N1 viruses isolated from avian and mammal hosts. A group of putative reassortant viruses was discovered and their genetic parents were identified. In addition, we investigated the evolutionary behaviors (including spatial migration, growth of genetic diversity, and evolutionary drift and selection) of the reassortant viruses and compare with those of the parental strain, thereby providing insights into the nature and impact of this emerging reassortant strain. Phylogenetic trees of Indonesian H5N1 viruses were reconstructed from 12 separate gene datasets (Table S6) , using a maximum likelihood (ML) approach with bootstrapping analyses to assess clade robustness (Figures 1, S1-S3; computer files of dendrogram are available as Dataset S1). In all the phylogenies, viruses sampled from avian species during earlier years of outbreaks (predominantly 2003-2004) tended to cluster near the root as expected, but with a poorly resolved branching structure that is likely due to relatively low sequence divergence. In contrast, viruses sampled from recent infections (2005) (2006) (2007) from avian and mammalian hosts formed three well-supported lineages with bootstrap support (or posterior probabilities) over 90 (or 0.9) under neighbor-joining (NJ), ML and Bayesian Markov Chain Monte Carlo (BMCMC) methods. We denote these lineages as groups 1, 2, and 3 in the hemagglutinin (HA) and neuraminidase (NA) phylogenies ( Figures 1A and S2C ). This structure was preserved in the phylogenies of other genes for which sufficient sequence data were available (viruses from group 3 were missing sequence data for the NP, NS, NS1, NS2, and PB2 genes). The group 3 lineage in the MP, M1, M2, and PB1 phylogenies was only represented by the A/Indonesia/6/05 strain. It is important to recognize that our phylogenetic groupings (groups 1, 2, and 3) of Indonesian H5N1 viruses ( Figure 1 and Table S6 ) are slightly different to those by Smith and coworkers [10] who did not require the same level of clustering support for each group, leading to the inclusion of earlier viruses (predominantly 2003-2004) . We chose to be conservative, and did not include poorly supported branches (e.g., earlier viruses) in our viral group definition. Therefore, we did not define a group corresponding to the group B of Smith et al., because most of group B taxa are earlier viruses. Groups 1 and 2 in this study correspond to groups C and A defined by Smith et al. respectively, plus some more recent viruses. Smith did not report group 3, because the sequences were unavailable at that time. We found previously unrecognized phylogenetic discordance between gene trees involving human and cat isolates (n = 25, denoted in red in Figures 1, S1-S3)-the main focus of our study-suggesting that they are reassortant viruses descending from group 2 and 3 lineages. In addition, the placement of two avian viruses isolates from Java (Ck/IDN/Semerang1631-62/07 and Ck/IDN/Magelang1631-57/07, shown in blue in Figures 1A and S2C ) differed between HA and NA phylogenies, suggesting another reassortment event. To further investigate the putative reassortant human and cat viruses, a selected dataset (n = 24) of manually concatenated full genomes (Figure 2A ; see Methods) of Indonesian H5N1 HPAI viruses were analyzed using more sophisticated analysis methods, including similarity plots, bootscan analyses and GARD analyses (genetic algorithm for recombination detection). In the similarity and bootscan plots ( Figure 2B and 2C), the putative reassortants (represented by a consensus sequence) showed a high degree of sequence similarity and phylogenetic clustering with the group 3 strain A/Indonesia/6/05 in the MP and PB1 segments, but not in other genomic regions, where they were more similar to the consensus sequence of group 2 viruses ( Figure 2B and 2C). Moreover, GARD detected two well-supported breakpoints near the boundaries of MP and PB1 segments in the concatenated genomes ( Figure 2D ), suggesting that the phylogenetic incongruence was significant between the three regions. In summary, all three analyses agreed that the newly reassortant strains had arisen from acquiring PB1 and MP genome segments from the group 3 lineage and the remaining segments from the group 2 lineage. Based on the HA phylogeny ( Figure 1A ), we further classified the reassortant viruses into three subgroups (R1, R2, and R3) with bootstrap support of 80% or better, as shown in the phylogenies containing only reassortant viruses ( Figure S4 ). Similar groupings were observed in the NA phylogeny ( Figures S2C and S4B ), although here subgroup R3 clustered with subgroup R1, and two reassortant viruses isolated in 2007 (IDN/CDC1046/07, IDN/ CDC1047/07) moved to a different subgroup. These inferred clustering patterns can be explained by multiple reassortment events, or by a single reassortment followed by divergence due to mutation and selection in different populations. We note that some group 2 viruses also cluster inside the reassortant subgroups ( Figures 1A and S2C ) and may indicate more reassortment events; however, most of them formed polytomies close to the most recent common ancestor (MRCA) of the reassortant subgroups and had poor bootstrap support for their exact placement. As the divergence between the reassortant subgroups and other intercalating group 2 viruses are low, the three subgroups may actually be linked uninterruptedly, implying a single origin. Therefore, the times and number of reassortment events that generated the putative mosaic reassortant viruses remains elusive. We examine both the single and multiple origin hypotheses in subsequent analyses, excluding the intercalating group 2 viruses from the reassortant group. To estimate the times of the reassortment events that generated the putative reassortant viruses, the times of the MRCA (tMRCA) of the three reassortant subgroups were estimated using BMCMC methods [18, 19] . Bayes Factors (BF) [20] were used to select Author Summary H5N1 highly pathogenic avian influenza (HPAI) virus emerged in China in 1996, and has spread beyond Asia since 2003. Following the first outbreak reported in Indonesian poultry farms in December 2003, the virus spilled over to 27 Indonesian provinces by June 2006, and became endemic in the country. In the following years, repeated sporadic human infections in Indonesia had been attributed to H5N1 HPAI viruses. Nonetheless, the viral evolution and transmission have not been fully understood. Here, we report phylogenetic evidence of a group of interlineage reassortant viruses isolated from human and cats in Java. Our comparative study of the reassortant viruses and one group of genetic parents found that although their rates of evolution were similar and both of their phylogenies were not geographically structured within mainland Java, the growths of genetic diversity were different. We also detected significant positive selection on the viral matrix and polymerase genes preceding and subsequent to the emergence of the reassortant viruses, which might correspond to viral adaptation. Based on our findings, we discuss the possibility of host switching in facilitating the emergence of the reassortant strain, and call for more extensive viral surveillances in the non-avian population in Indonesia. among strict and relaxed clock models of evolution [21] . The uncorrelated exponentially-distributed clock model (UCED) significantly outperformed the other models (lnBF.3) for most datasets, except for the NA gene of the reassortant viruses, for which the strict clock model was not rejected (lnBF,1; Table S4 ). The results of tMRCA estimation are summarized in Figure 3C and 3D. In addition, sequence isolation dates were plotted against their genetic distance (units of substitutions/site) to their MRCA, to graphically show the accumulation of mutations through time ( Figure 3A and 3B). The tMRCA of all reassortant viruses (All- were isolated from central and east Jakarta (Table 1) . Parsimony reconstruction (see Methods) of binary ancestral geographical states (either Greater Jakarta or West Java) upon the HA and NA ML phylogenies suggested that the MRCA of all reassortants (and the MRCAs of each reassortant subgroup) likely originated from Greater Jakarta and surroundings (Table S2 ; result robust to random resolution of polytomies; see Methods). The mean numbers of observed geographical state changes (GSC) of the reassortant and of the group 2 parental strains were estimated independently and compared with the null distribution of GSC values under the null hypothesis of completely unrestricted migration (i.e. panmixis; Figure S5 ) [22] . For the reassortant strain, the observed GSC value was not significantly lower than the GSC value expected under panmixis (Slatkin-Maddison test: p.0.2). Therefore the observed geographic structure is not significantly different to that expected by chance alone. For group 2 viruses, the observed GSC value for all geographical state pairs is significantly (p,0.0002) lower than the null value. However, the observed value of GSC within Java (i.e., migrations between Greater Jakarta and the rest of Java) and between the three islands (i.e. migrations between mainland Java, Sumatera and Sulawesi Selatan including Papua) are insignificantly (p.0.2) and significantly (p,0.0002) lower than the corresponding null values respectively, suggesting that the phylogeny of group 2 viruses is not geographically structured within Java, but is subdivided by island-to-island migrations. However, we could not address whether the viral migrations inside Sumatera and Sulawesi Selatan including Papua are panmictic or structured due to limited operative localities in our dataset to distinguish between different regions inside these islands. We also found that the migration of group 2 viruses from Greater Jakarta and surroundings to Sumatera and Sulawesi Selatan including Papua was more frequent than expected under the null hypothesis, and there is relatively little viral migration from the rest of Java to Sumatera and Sulawesi Selatan including Papua (Table S3 ). This observation suggests Greater Jakarta played a more salient role in dispersing group 2 viruses to other Indonesian islands than other parts of Java did. There is another family member (brother) who was confirmed with H5N1 infection (index #45, fatal); however, virus sequences are not available. f Numbers in parentheses are the unique references to the localities shown in Figure 4 . Numbers 1-6 were assigned to Greater Jakarta and surroundings; numbers 7-12 were assigned to West Java. n/a = Information not available. doi:10.1371/journal.ppat.1000130.t001 We used the Bayesian skyline plot (BSP) [23] to estimate the change of relative genetic diversity of the reassortant viruses and of the group 2 parental strain over time, as shown in Figure 3E -3H. For both the HA and NA datasets, the group 2 viruses consistently show a slow growth in relative genetic diversity over time which appears to follow a constant size or exponential growth model, whereas the reassortant viruses initially exhibited an abrupt rise in relative genetic diversity followed by stabilization, which visually resembles a logistic growth curve with two phases [24, 25] ( Figure 3E and 3F) . When the BSPs are superimposed upon the demographic results obtained under parametric growth models (i.e., constant, exponential and logistic growth; Figure S8 ), then a similar observation can also be made. However, BF tests (Table S4) indicate there is insufficient statistical power to discriminate between the three parametric growth models (lnBF,2.99), suggesting a lack of strong demographic signal in these data. When the parametric demographic models were fitted to the data, the median estimates of growth rates for the reassortant datasets are generally higher than those estimated for the datasets of group 2 viruses (Table S1 ). However, the confidence intervals of some growth rate estimates are fairly large and overlapped among the reassortant and group 2 viral datasets. Diversifying selection in the PB2 and M1 genes Using the Random Effects Likelihood (REL) method [26] we found sites under positive selection in the PB2 gene (codons 76, 534, 627, 677 and 740) and the PA gene (codon 409) of the reassortant viruses. The Fixed Effects Likelihood (FEL) method [26] was more conservative and only identified PB2 codon 534 as being positively selected. For the group 2 viruses, HA codon 129 (starting from HA1) and M1 gene codon 205 were the only selected sites identified by the FEL and REL methods, respectively. Using a lineage-specific selection model (see Methods), we identified elevated rates of diversifying selection, measured by the ratio of non-synonymous to synonymous substitutions (dN/dS), on the M1 gene in the lineage leading to the MRCA of the group 3 viruses and preceding the emergence of the reassortant viruses (highlighted in Figure 1B ). The dN/dS values for the M1 gene in this lineage (which we call the preemergence lineage) was estimated to be 1.514 (95% CI: 0.447-3.814; see Table S5 ), significantly higher (LRT p,0.002) than the mean estimates for other lineages (dN/dS = 0.077) in the Indonesian clade and for lineages in other H5N1 HPAI clades (e.g., Fujian, Qinghai, Thailand and Vietnam clades which have dN/dS ranging from 0.05 to 0.09). This lineage-specific elevation of dN/dS was not significant (LRT p.0.1) for other genes (i.e. HA, NA, M2, PB1; see Table S5 ). Four amino acid changes in M1 occurred along the pre-emergence lineage, including threonine to alanine at reside 37, arginine to lysine at reside 95, threonine to alanine at reside 137, glutamine to histidine at reside 249. Three (residue 37, 95, and 137) of them are located close to the electrostatic positive surface of the N-terminal domain of the M1 protein molecule ( Figure S7 ), and one (residue 249) is located in the remaining C-terminal fragment. This study classified H5N1 HPAI viruses in Indonesia into three distinct viral lineages (groups 1, 2, and 3) and discovered a group of naturally occurring reassortant viruses that represent a newly emergent H5N1 HPAI strain in Java in 2006. Several phylogenetic methods concurred that two (MP and PB1) of the reassortant viruses' genome segments descended from the group 3 ancestral viruses, and the remaining six (PB2, PA, HA, NP, NA, NS) segments descended from the group 2 ancestral viruses. Although the majority of reassortant viruses (24/25) are human isolates, few of the associated human infections are epidemiologically linked (Table 1) , suggesting multiple sporadic zoonotic transmissions from birds. The phylogeographic results indicate that the parental viruses of the reassortants have been co-circulating in Java since 2005. Despite the identification of parental lineages, the exact number of reassortment events remains difficult to assess. Although the three fairly consistent phylogenetic subgroups (subgroups R1, R2, and R3 in Figure S4 ) formed by the reassortant viruses suggest three independent reassortments, the underlying uncertainty in our estimated phylogenies means that we cannot rule out the possibility of a single origin. The hypothesis of three reassortments implies that the viruses have acquired exactly the same genome segments from the same group of parental viruses, which seems unlikely to occur by chance (probability = 0.0089, assuming panmixis and that exactly two genomic segments are swapped out). This probability might be increased if reassortments confer a selective advantage. We did detect a significantly stronger selection pressure on the M1 protein in the pre-emergence lineage of group 3 parental strain that led to the reassortant viruses (Table S5) . Previous reports suggested a few amino acid changes in M1 of influenza A and B viruses can confer a growth advantage in mouse lungs [27] [28] [29] . Although the M1 mutations identified in this pre-emergence lineage have not been functional characterized elsewhere in the authors' knowledge, one (residue 137, TRA) of them is close to a previously characterized mutation (residue 139, TRA) which controls the virulence in mouse model [27, 29] . Three of the inferred residue changes are located close to the electropositive surface of N-terminal domain of M1 protein ( Figure S7 ) that acts to bind viral RNA [30, 31] . The M1 matrix protein mediates encapsulation of viral RNAnucleoproteins into membrane envelope during packaging [31] , and has close contact with other viral proteins inside the viral particle. It seems possible that some of these changes may be involved in the adaptation of reassortant viruses, through promotion of structural interactions among viral proteins. According to our analyses, the common ancestor of the reassortant viruses is dated to July 2005 (HPD: April-October), approximately 5 months prior to the first case of human infection caused by the reassortant virus (index case #15 defined by WHO; see Table 1 ). Our analysis of virus phylogeography suggests the ancestors of these reassortant viruses first arose in Greater Jakarta and surroundings, which agrees with the observation that the first two cases of human infection by the reassortant viruses occurred in Central and East Jakarta (index cases #15 and #16). The molecular dating and phylogenetic analyses suggest that nascent reassortant viruses might take several months to spread and expand their diversity in the local bird population, eventually leading to the exposure of human population. The subsequent spread of the reassortant strain seems to become more rapid and extensive, as human cases were reported outside Greater Jakarta one month later, and the reassortant virus spread to as far as the south and east of West Java in the following six months (Table 1 and Figure 4) . Commercial poultry transportation, as well as carriage by migratory birds, may facilitate the viral migration, but their tangible contributions need further studies. Our results suggest that the circulations of reassortant viruses and their genetic parent (group 2) were not restricted by geography within Java. The viral migration back to Greater Jakarta could be driven by the inter-province transfer of infected poultry, in particular the importation of live poultry or fresh poultry products to the densely human populated Jakarta from the remote provinces engaged in poultry-farming. Future studies on economic and social geography (e.g., addressing the modes of inter-provincial poultry transport) in Indonesia might help to further elucidate the effect on the viral dispersal by human, agricultural and industrial activities. In this study, we opted for a lower geographical resolution (i.e., four widely ranged geographical states instead of distinct geographical coordinate for each viral isolate) in our phylogeographic analyses because of the varying precision of the geographical data we have. Therefore, more complex hypotheses of viral origin and migration trajectory cannot be investigated here, but can be explored when more high-quality geographical data of Indonesian H5N1 viral samples is available. The BSP analyses ( Figure 3E-3H) indicate that the reassortant viruses follow a logistic-like growth curve, which is typical for virus invasion and maintenance, especially in a structured population [24, 25] . In contrast, the group 2 viruses followed a more continuous and relatively slow growth in diversity. There was insufficient data in our samples to definitively discriminate between alternative population growth models and provide narrow confidence intervals for parameter estimates, but our results are suggestive and future sequencing will add to the needed statistical power. What factors have contributed to the apparent difference in the growth of genetic diversity? Rates of molecular evolution between the two groups were similar ( Figure S6 ) and therefore are not likely to be the cause. Since our analyses could not resolve the temporal dynamics of population subdivision by geography, we cannot directly investigate how viral genetic variation is affected by the population structure. We expect future development of analysis methods will help to shed more light on the interaction between viral migration and genetic diversity. Analysis of clinical records (Table 1) found that the mean duration from onset to death in those fatal human cases caused by Indonesian reassortant H5N1 viruses is 9.1 days (standard deviation [SD] = 3.9; n = 23) and those caused by other Indonesian H5N1 viruses is 7.7 days (SD = 2.7; n = 10), and their means are not significantly different (student t-test, p.0.25, two-tails). Therefore, based on the clinical records, the reassortant viruses did not kill human faster than other Indonesian H5N1 viruses did. However, we would recommend more experimental studies addressing the virulence, pathogenicity and immunogenicity of the reassortant viruses and the parent strains to verify this claim in the future. Mechanisms of viral transmissions are sometimes correlated with genetic diversity dynamics. For example, hepatitis C viruses transmitted by drug injection or blood products have a faster rate of spread than endemic strains circulating in Asia and Africa [32] . It has also been suggested that mosquito susceptibility may affect the growth of dengue viruses [33] . Therefore, it is possible that a change of host species could generate the difference in the viral dynamics we observe. In our study, the majority of the reassortant viruses (24/25) were isolated from humans, whereas only a minority of the group 2 viruses were isolated from humans (10/57 and 10/41 in the HA and NA datasets, respectively). It has been previously shown that the receptor binding specificity of hemagglutinin [34] and mutations in the viral polymerase (e.g., lysine at residue 627 of PB2) [35] [36] [37] can determine viral transmissibility and replication in different host species. None of the aforementioned HA mutations which confer recognition to human-type host cell receptors [34, 38] were found in the Indonesian reassortant viruses; however, our detection of positively selected sites in the PB2 gene of the reassortant viruses could potentially reflect adaptation to mammalian hosts, and requires further investigation. In particular, amino acid changes on two positively selected sites (threonine to methionine at reside 76, glutamic acid to glycine and alanine at reside 677) were found on the internal branches of the reassortant lineage, corresponding to molecular changes during sustainable transmissions. However, some of these positively selected changes may also result from the compensatory evolution as the mix of genome segments from different strains might alter their epistatic physiochemistry [39] . Although most of the human isolates in our datasets were epidemiologically unlinked, such linkage is theoretically possible if many asymptotic or mildly manifested human infections are not reported. Recently, some evidence of subclinical or asymptotic H5N1 infection in humans has been put forward [40, 41] ; however, the ability of the viruses to transmit from these infected individuals to other susceptible individuals remains unknown. The possible role of other animal host species in the transmissions of reassortant viruses in Indonesia should not be neglected. In particular, one of the reassortant viruses was isolated from a dead cat in Jakarta, where H5N1 outbreaks in poultry and sporadic human infections have been reported [42] . Moreover, unusual high mortality of cats in the vicinity of H5N1 HPAI outbreaks has been reported [43] . An unofficial report also detected H5N1 HPAI sero-positivity in around 20% of 500 blood samples taken from stray cats near poultry markets in Java and Sumatera [44] . In addition to small cats in Germany [45] , Iran [46] , and Indonesia [42] , dogs and zoo tigers were also found infected with H5N1 HPAI viruses in Thailand [47, 48] . Furthermore, previous experimental studies have demonstrated that cats can be infected with H5N1 HPAI virus [49, 50] , and that cat-to-cat transmission is possible [49, 51] . Could cats, or other non-avian species, have played a role in spreading the reassortant viruses in Java? Similarly, could cats act as amplifying hosts facilitating viral expansion and cross-species transmission, as civets did in the SARS outbreaks [52] ? Future experimental studies on these reassortant viruses, that assess viral transmissibility between species, together with epidemiological studies, such as viral monitoring within Indonesian animal populations using serological tests and PCR detection, would give more clues to these questions. H5N1 HPAI viruses have been endemic and evolved into different genetic lineages that have spread across Indonesia. Areas where more than one lineage of virus is co-circulating, such as Jakarta, are most likely to generate novel viruses by inter-lineage reassortment. These reassortant viruses have distinctive evolutionary and transmission dynamics, as shown in this study. We suggest that more intensive and timely field surveillance and analysis of influenza viruses, including H5N1 HPAI and human H3N2, H1N1, and H1N2 epidemic strains, should be employed, so that bio-security can be undertaken promptly and appropriate strains can be selected for vaccine production whenever a novel reassortant strain emerges. The reassortant viruses reported in this study should be also added to the watch list for the future epidemiological surveillance. Sequence data collection and alignment H5N1 influenza viruses isolated from avian and mammalian hosts in Indonesia during 2003-2007 were studied. Their genomic sequences (n = 807) were extracted from the Influenza Virus Resource [53] and the Influenza Sequence Database [54] in September 2007, and aligned using MUSCLE version 3.6 [55] . Columns with gaps were removed from the alignments, and sequences from the same virus strain (duplicated submission in the two databases) were filtered such that one copy was retained. Eight genome segment alignment datasets (PB2, PB1, PA, HA, NP, NA, MP, and NS), as well as four coding sequences (M1, M2, NS1, and NS2), were generated. Full details of our datasets can be found in Table S6 and S7. Phylogenetic trees of 12 alignment datasets were reconstructed using the ML approach implemented in PhyML 3.412 [56] . The robustness of the ML tree topology was assessed by comparing the ML topology with the topologies sampled in the BMCMC analysis performed in MrBayes version 3.1.2 [57] , and with bootstrapping analyses of 1,000 pseudo-replicate datasets. ML and NJ trees were estimated from the bootstrap datasets using PhyML [56] and PAUP* version 4beta10 [58] , respectively. A general-time-reversal (GTR) substitution model with gamma distributed rate heterogeneity of 4 rate categories (C 4 ) and a proportion of invariable sites were used in all tree reconstruction methods. Phylogenies were rooted with the H5N1 HPAI strain A/Ck/HK/YU324/2003, which is genetically close to the newly reported Hunan strains [17] , and shares comparable genetic proximity to Indonesian clade. Homologous recombination within each gene segment among Indonesian H5N1 isolates was extensively searched using Recombination Detection Program version 2 (RDP2) [59] , and the datasets are found to be free of homologous recombination. Putative reassortant viruses were preliminarily identified by their topological incongruity across the phylogenies of different gene segments. This was further investigated using a smaller set of Indonesian H5N1 virus isolates with full genome sequences, which included sequences of early viruses (n = 2), group 1-3 lineages (n = 12) and putative reassortant viruses (n = 10). The eight gene segment alignments were manually concatenated in the order of their length to generate a single alignment of complete genome sequences, and was further analyzed using 1) similarity plots and 2) bootscan analyses [60] implemented in SIMPLOT version 3.5.1 [61] , and 3) GARD [62] available via the Datamonkey website [63] . The hypothesis of reassortment was supported if the recombinant breakpoints were detected near the junctions where the genome segments were manually concatenated. The geographic locations of virus isolation were either obtained from the sequence databases, or obtained through personal communication with Catherine Smith (from Disease Control and Prevention, Atlanta, USA), or inferred from their strain names (Tables 1 and S6 ). The locations of isolates were indicated on the map of main island of Java in Indonesia (Figure 4 ). Due to the limit of our geographical data, the localities of the isolates shown in the map (Figure 4 ) should be regarded as arbitrary within the province which is the highest precision level shared by all viral samples. Each of the reassortant viruses was assigned with a state of either Greater Jakarta (surroundings included) or West Java depending on its place of origin ( Table 1 ). The migratory history of the reassortant viruses (n = 25) between these geographical states were inferred based on the refined ML phylogeny of HA and NA ( Figure S4 ) independently using two parsimony optimization methods, called ACCTRAN (accelerated transformation) and DELTRAN (delayed transformation) implemented in PAUP* software. The geographical states of all ancestral nodes in the tree were estimated to achieve minimum state changes in overall, and therefore the number of state changes and state of the MRCA of the reassortant was obtained. Polytomies were randomly resolved 1,000 times, and state changes were estimated separately for each resolution. The mean number of state changes was then calculated. To test against the null hypothesis of completely unrestricted migration between geographical states (panmixis), the mean number of observed state changes was compared with the frequency distribution of the mean number of expected state changes under the null hypotheses. The null distribution and critical values were generated by randomly shuffling the states of isolates 5,000 times (the Slatkin-Maddison test [22, 24, 64] ). The migratory history of group 2 viruses was also studied using the HA gene in a similar manner, while each group 2 virus was assigned to either of four widely ranged geographical states: Greater Jakarta and surroundings, the rest of Java, Sumatra, and Sulawesi Selatan, including Papua. This assignment scheme is comparable to that of reassortant viruses, as West Java is part of Java. Parameters of codon-partitioned substitution rates, demographic functions, tMRCA and tree topologies were co-estimated from HA and NA gene datasets of reassortant and group 2 viruses separately in a BMCMC framework [18] using BEAST version 1.4.6 [65] . Substitution model HKY+C 4 with invariable site portion was used. Isolation dates were used to calibrate the molecular clock. Three clock models including strict clock, UCEN and UCLN relaxed clocks [21] were attempted independently, and the best-fit clock model was selected by comparing the BF calculated from their posterior distributions [20] . The Bayesian skyline plot [23] was used to estimate population dynamics, in terms of relative genetic diversity. Less complex parametric demographic models (constant size, exponential growth and logistic growth) were applied independently, and the best-fit models selected by BF tests were used to quantitatively estimate the growth rate and other demographic parameters. The BMCMC analyses contained 2610 8 states, with sampling every 1,000 states, and the first 10% of each chain was discarded as burn-in. Convergences and effective sample sizes of the estimates were checked using Tracer v1.4 [66] . Positively selected sites were detected using random effect likelihood (REL) and fixed effect likelihood (FEL) methods [26] via the Datamonkey website [63] . Bayes factors larger than 50 and pvalues smaller than 0.1 were used as thresholds for strong evidence of selection in REL and FEL, respectively. To test lineage-specific positive selection, the two-ratio branch model was used, which pre-specifies a single rate of synonymous substitution (dS) for the whole phylogeny and two rates of non-synonymous substitution (dN 1 and dN 2 ). The dN 1 was specified for the pre-emergence lineage (indicated as the ancestral branch connecting the group 3 MRCA; see Figure 1B ) for the group 3 viruses (including the reassortant viruses for M1, M2, and PB1 genes). The dN 2 was specified for other lineages across the phylogenies. The ML estimates of these rate parameters were performed in HYPHY version 0.99 [67] . The resulting likelihood score of the two-ratio model was then compared with that of the one-ratio model, which assumes the same dN and dS across the phylogeny, using the likelihood ratio test (LRT, with degree of freedom = 1). The substitution model MG94XGTR+C 4 was used. The ancestral nucleotide sequences of all internal nodes were reconstructed using joint ML method [68] implemented in HYPHY. Amino acid changes along the pre-emergence lineage were determined, and were then mapped onto the three-dimensional structure of the N-terminal domain of M1 matrix protein molecule [30] available (PDB-ID: 1EA3) in RCSB Protein Data Bank. Figure S6 Substitution rates of HA and NA genes from reassortant and group 2 parental strains. 95% higher probability densities (HPDs) are indicated by the error bars. 1st, 2nd, 3rd, and C denote the rate for the 1st codon position, 2nd codon position, 3rd codon position, and whole sequence (non-partitioned), respectively. Substitution rate units for codon partitioned and non-partitioned sequences are substitution/codon/year and substitution/site/year, respectively. Table S5 Estimations of dN/dS using 1-ratio and 2-ratio lineage-specific selection models. These estimations were performed in HYPHY software. Gene datasets other than PB1, HA, NA, M1, and M2 were not analyzed because group 3 is represented by the single virus IDN/6/05. Found at: doi:10.1371/journal.ppat.1000130.s014 (0.03 MB DOC) Information and phylogenetic groupings of sequences used in this study. 1, 2, 3, and X denotes groups 1, 2, 3, and unclassified (early viruses; see main text for explanation). Empty entries indicate the unavailability (e.g., no sequence found, too short, too many ambiguous codes, and too many gaps) of the sequence.

Endothelial Cells Support Persistent Gammaherpesvirus 68 Infection

A variety of human diseases are associated with gammaherpesviruses, including neoplasms of lymphocytes (e.g. Burkitt's lymphoma) and endothelial cells (e.g. Kaposi's sarcoma). Gammaherpesvirus infections usually result in either a productive lytic infection, characterized by expression of all viral genes and rapid cell lysis, or latent infection, characterized by limited viral gene expression and no cell lysis. Here, we report characterization of endothelial cell infection with murine gammaherpesvirus 68 (γHV68), a virus phylogenetically related and biologically similar to the human gammaherpesviruses. Endothelial cells supported γHV68 replication in vitro, but were unique in that a significant proportion of the cells escaped lysis, proliferated, and remained viable in culture for an extended time. Upon infection, endothelial cells became non-adherent and altered in size, complexity, and cell-surface protein expression. These cells were uniformly infected and expressed the lytic transcription program based on detection of abundant viral gene transcripts, GFP fluorescence from the viral genome, and viral surface protein expression. Additionally, endothelial cells continued to produce new infectious virions as late as 30 days post-infection. The outcome of this long-term infection was promoted by the γHV68 v-cyclin, because in the absence of the v-cyclin, viability was significantly reduced following infection. Importantly, infected primary endothelial cells also demonstrated increased viability relative to infected primary fibroblasts, and this increased viability was dependent on the v-cyclin. Finally, we provide evidence for infection of endothelial cells in vivo in immune-deficient mice. The extended viability and virus production of infected endothelial cells indicated that endothelial cells provided a source of prolonged virus production and identify a cell-type specific adaptation of gammaherpesvirus replication. While infected endothelial cells would likely be cleared in a healthy individual, persistently infected endothelial cells could provide a source of continued virus replication in immune-compromised individuals, a context in which gammaherpesvirus-associated pathology frequently occurs.

Endothelial cells create a physical barrier on the luminal surface of blood and lymphatic vessels. This barrier must be traversed by blood-borne pathogens and immune cells trafficking between tissues and the bloodstream. Many herpesviruses require systemic spread for persistence within a host, and therefore must cross such an endothelial cell barrier. To date, herpesviruses have been implicated as potential initiators of arterial injury, endothelial dysfunction, and local inflammation, possibly contributing to the pathogenesis of atherosclerosis [1] [2] [3] [4] . Human cytomegalovirus (HCMV), a betaherpesvirus, infects endothelial cells in vivo. Studies have shown that infected endothelial cells play a role in HCMV dissemination and pathogenesis [5] . Endothelial cells exhibit regional specialization in gene expression and morphology depending on the local physiologic demands of their respective organs and tissues [6] . In light of this diversity, it is not surprising that endothelial cells from different tissues differ in their susceptibility to HCMV infection [7] . Kaposi's sarcoma-associated herpesvirus (KSHV), a human gammaherpesvirus, also infects endothelial cells in vivo. More importantly, KSHV is the causative agent of the endothelial cell neoplasm, Kaposi's sarcoma (KS). The murine gammaherpesvirus, cHV68, has been detected in aortic endothelium after infection of apoE deficient mice, as well as on the luminal surface of explanted aortas infected in vitro [8] . While a relationship between gammaherpesviruses and endothelial cells has been noted, the role of endothelial cells in chronic gammaherpesvirus infection and pathogenesis is ill-defined. Gammaherpesviruses are a lymphotropic family of viruses associated with a broad spectrum of malignancies and lymphoproliferative diseases. These oncogenic viruses persist for the life of the host by establishing and maintaining a latent infection. Because gammaherpesviruses are extremely host-specific, studying the human viruses Epstein-Barr virus (EBV) and KSHV in nonhuman systems does not mimic natural infection. cHV68 is a natural pathogen of wild murid rodents and therefore provides a valuable small animal model of gammaherpesvirus infection [9] [10] [11] [12] [13] . This virus has important biological and genetic similarities to the human gammaherpesviruses, and results in a variety of pathologies in defined mouse models. Primary cHV68 infection is characterized by virus replication in lung epithelial cells, and the establishment of latency in B cells, dendritic cells, and macro-phages [12] [13] [14] [15] [16] . Persistent cHV68 infection also occurs in lung epithelial cells [17] . The role of endothelial cells in cHV68 infection has not been explored to date. At the cellular level, gammaherpesvirus infection encompasses two broadly defined outcomes: productive, lytic replication and non-productive, latent infection. During lytic replication, viral DNA is amplified and host cell machinery is utilized for the production of viral progeny. Viral genes are actively transcribed and translated, contributing to new virus production. Ultimately, the cell is lysed and infectious virus released. In light of the fact that this process occurs quickly (24 to 48 hours in vitro) our understanding of virus-host interactions during primary infection is quite limited. During latency, viral DNA is not amplified, but instead is maintained as a nuclear episome in latently infected cells, and little viral gene transcription occurs. Latently infected cells remain intact and do not produce new infectious virus. Furthermore, it is thought that these long-lived latent cells may serve as the major mechanism by which gammaherpesviruses promote life-long infection of their hosts. Here we characterize cHV68 infection of endothelial cells. To date, in vitro infection of most cells with cHV68 supports significant viral replication and results in complete cell lysis (data not shown, van Dyk). We have infected both primary endothelial cells and endothelial cell lines and demonstrated that they produce comparable amounts of virus as fibroblasts. However, whereas fibroblasts were mostly lysed by 36-48 hours post-infection, a significant percentage of infected endothelial cells remained intact. Analysis of endothelial cell lines revealed that these intact cells 1) were actively infected and undergoing the lytic transcriptional program, 2) continued to proliferate for a prolonged time after infection, 3) were altered in morphology and cell-surface protein expression, compared to uninfected endothelial cells, and 4) continued to release infectious virions as far as 30 days postinfection. In the absence of the cHV68 viral cyclin, endothelial cell viability was significantly decreased after infection, indicating that an active viral process was responsible for promoting the outcome of infection in endothelial cells. Of major significance, we also provided evidence of endothelial cell infection in primary cells and in vivo in immune-deficient mice. These data demonstrated that endothelial cells supported persistent, productive gammaherpesvirus infection, an outcome not previously reported for cHV68 or other gammaherpesviruses. Together, these data indicated the potential of the endothelial cell as a critical cell type in cHV68 pathogenesis, and open a new avenue of research into herpesvirus manipulation of endothelial cell biology. Endothelial cells are heterogeneous in nature, express unique surface antigens, and differ in their susceptibility and response to infection by various pathogens [6] . In light of this diversity, we investigated endothelial cell lines from various anatomic locations, and from BALB/c and C57BL/6 mice, for their ability to support cHV68 replication. We analyzed the outcome of cHV68 infection in endothelial cell lines from lung (CD3), brain capillary (MB114), and lymph node (SVEC) by single-step and multi-step growth assays. Single step growth proceeded with kinetics similar to previously published infection of the fibroblast cell line, NIH 3T12 cells [18] in each of the three endothelial cell lines analyzed (Fig. 1A , top panel). Though equivalent numbers of cells were plated per well, the differences amongst the endothelial cell lines in titer at time zero were not surprising given that cell lines grow at varying rates and to varying densities. A comparison of viral titers at 36 hours revealed that the three endothelial cells lines produced titers comparable to those achieved in 3T3 fibroblast cells. Additionally, multi-step growth in MB114 and CD3 endothelial cells was similar to growth in 3T3 fibroblasts (Fig. 1A, bottom panel) . These data demonstrated that cHV68 is capable of replication in endothelial cells infected in vitro. Using a GFP-labeled virus, we next sought to determine if the percentage of infected endothelial cells was comparable to percentage of infected fibroblasts. Following infection of MB114 endothelial cells and 3T12 fibroblasts, cells were analyzed for expression of GFP from the cHV68 genome by flow cytometry. At 24 and 48 hours post-infection the majority of endothelial cells and fibroblasts expressed GFP (Fig. 1B) . These data indicated that, like fibroblasts, endothelial cells were uniformly infected and supported the cHV68 lytic transcription program. A population of endothelial cells remained intact following cHV68 infection and uniformly expressed lytic antigens Uninfected endothelial cells and fibroblasts grew as adherent monolayers ( Fig. 2A, top panel) . However, in infected endothelial cell cultures, we observed a population of intact, phase-bright, and non-adherent cells as early as 96 hours post-infection, whereas only cellular debris remained in the infected fibroblast cultures (data not shown). At six days post-infection, we collected non-adherent cells and cellular remains by centrifugation. After washing, we resuspended this material in complete media for culture and analysis. The cells harvested from the infected endothelial cultures appeared as individual, intact cells suspended in culture, whereas the material harvested from the infected fibroblast cultures appeared as clumps of cellular debris ( Fig. 2A, bottom panel) . Prior to centrifugation, we measured cell viability by trypan blue exclusion. Very few infected fibroblasts remained as viable, nonadherent cells at six days post-infection (Fig. 2B ). However, a significant proportion of endothelial cells remained as viable, nonadherent cells at six days post-infection. When MB114 endothelial cells were treated with a non-toxic dose of phosphonoacetic acid Various human diseases are associated with gammaherpesvirus infections, including neoplasms of lymphocytes (e.g. Burkitt's lymphoma) and endothelial cells (e.g. Kaposi's sarcoma). Gammaherpesvirus infection of cells usually results in either productive infection that is characterized by new virus production and rapid destruction of the host cell, or latent infection that is characterized by long-term carriage of viral DNA in intact cells that do not produce virus. Here, we characterize endothelial cell infection using a small animal model of gammaherpesvirus infection and disease. While infection of endothelial cells resulted in virus production as most cells do, infection of this cell type was unique in that cells remained intact and continued to proliferate. These intact, infected endothelial cells were significantly altered in appearance and gene expression compared to uninfected cells, changes predicted to impact endothelial cell growth and function. Endothelial cells were unique in their ability to support this type of persistent infection. These data demonstrate an additional mechanism, beyond latency, by which gammaherpesviruses may achieve long-term propagation, particularly in conditions of immune suppression. Our results suggest persistently infected cells as a therapeutic target for prevention/ treatment of chronic disease, and may provide a mouse model for future testing. (PAA), an inhibitor of viral DNA replication and late gene synthesis, most cells remained adherent, and there were significantly fewer detached, viable cells (Fig. 2B) . PAA alone had no effect on the viability, adherence, or phenotype of uninfected MB114 cells in culture (data not shown). These data revealed that a population of endothelial cells became non-adherent and remained viable at six days post-infection, whereas fibroblasts were destroyed by cHV68 infection. Additionally, this outcome of infection in endothelial cells was influenced by late viral gene expression, as most cells remained adherent in the presence of PAA. Next we determined the status of virus replication in the viable and non-adherent cells. Cells collected at six days post-infection were analyzed by flow cytometry for expression of the cHV68 glycoprotein, gp150, which is expressed during the lytic transcription program on the surface of infected cells [19] . Cell viability was also determined using propidium iodide (PI), a cell impermeant dye that is excluded from intact cell membranes. In these experiments we also included the S11E cell line, a mouse B cell lymphoma line which harbors latent cHV68, and therefore is negative for gp150 surface expression [20] . The majority of MB114 endothelial cells and 3T12 fibroblasts were positive for surface gp150 expression (Fig. 2C ). While 3T12 fibroblasts were mostly PI positive, MB114 endothelial cells excluded PI. These data indicated that endothelial cells which remained viable at six days post-infection were undergoing the lytic transcription program and were not latent or uninfected. At 12 days post-infection we determined the viability of cells collected and cultured at six days post-infection. PI staining of S11 cells, after six days in culture, revealed that 51.7% (62.5 SEM) of cells were PI negative and thus viable. PI staining of post-infection fibroblasts, treated in parallel with infected endothelial cells, revealed that only 14.5% (61.5 SEM) of the cells harvested remained intact. In contrast, PI staining of post-infection endothelial cells revealed that 92.5% (60.3 SEM) of the cells harvested were intact and viable (Fig. 2D) . Analysis of the viral replication program in MB114 endothelial cells at 12 days postinfection revealed that the majority of cells continued to express gp150 (Fig. 3C) . Therefore, infection of endothelial cells with cHV68 resulted in a population of cells that escaped lysis and remained intact as far as 12 days post-infection, while undergoing the lytic transcription program. The observed viability of endothelial cells after cHV68 infection could be the result of escape from infection or latent infection of these cells. To determine the infection status of the intact endothelial cells, we performed RT-PCR analysis of viral gene transcripts. The cHV68 M2 gene transcript is synthesized during both lytic and latent infection, as are the polymerase III (pol III) transcripts encoded at the left end of the viral genome [21, 22] . Latently infected S11 cells contained both pol III-1 and M2 transcripts (Fig. 3A) . We detected these transcripts in 3T3 fibroblasts at 24 and 36 hours post-infection, as well as from the 3T3 cellular debris harvested at six days post-infection, while no cellular b-actin transcript was detected from infected 3T3 cells at 36 hours post-infection. No RNA could be detected from infected 3T3 debris at 12 days post-infection, whereas we recovered RNA from infected MB114 cells as far as 12 days post-infection. Pol III-1 and M2 transcripts were detected in infected MB114 cells at 24 and 36 hours post-infection and from the intact, cultured cells at six and 12 days post-infection. We detected b-actin RNA in infected MB114 cells at 24 and 36 hours post-infection, but not at six and 12 days post-infection. However, we detected another cellular mRNA transcript, cyclophilin A, at six days post-infection in 3T3 fibroblasts and MB114 cells (Fig. S2) , and comparable detection of the cellular 18S rRNA transcript was observed in all conditions. Gradual loss of the cellular b-actin transcript in viable cells was not surprising given that b-actin has been documented to vary significantly in the setting of virus infection [23] , and selective degradation of certain mRNAs frequently occurs during virus infection [24] [25] [26] [27] [28] . Analysis of viral gene transcripts indicated that those endothelial cells which remained intact after cHV68 infection contained viral gene transcripts as far as 12 days postinfection. To determine whether the intact, infected endothelial cells were undergoing active viral replication, we analyzed infected cells for the expression of early and late lytic replication-associated M3 and gB. As predicted, 3T3 fibroblasts, which support lytic replication, expressed both M3 and gB (Fig. 3B ). In contrast, S11 B cells, which contain latent cHV68, did not express either M3 or gB. Intact, infected MB114 endothelial cells not only expressed M3 and gB early (36 hours), but also at six and 12 days post-infection. A faster migrating gB-specific band was also detected at six days post-infection, and may indicate differential glycosylation or degradation. Additionally, infected MB114 endothelial cells were positive for surface protein expression of gp150 at both 12 and 24 days post-infection (Fig. 3C ). Our analysis of lytic protein expression indicated that endothelial cells which remained intact after cHV68 infection expressed early and late lytic proteins, indicating that not only did cHV68 infected endothelial cells still contain virus, these cells also expressed viral proteins indicative of active viral replication, and not latency. To further test the status of viral replication in endothelial cells, we infected MB114 endothelial cells with GFP-cHV68 and collected the intact, non-adherent cells at six days post-infection. GFP expression is driven by the CMV immediate early promoter and has been demonstrated to be expressed only during lytic infection [29] . At the time of harvest, these cells uniformly expressed GFP (Fig. 3D histogram) and continued to do so at 20 days post-infection, (Fig. 3D micrographs) . Therefore, the nonadherent endothelial cells which remained intact after cHV68 infection did not escape infection and were not latently infected. After establishing that the intact endothelial cells harvested at six days post-infection were indeed infected, we determined whether these cells produced mature virions. Transmission electron microscopy (TEM) of these cells revealed virions at various stages of maturation throughout the nucleus and cytoplasm of 100% of the 20 cells imaged at 12 days post- The culture conditions we established for maintaining the nonadherent cells included centrifugation and resuspension in fresh media every six days (Fig. S1 ). To test the contribution of nonadherent cells to virus production, aliquots of the supernatants were collected for measurement of cell-free virus. A culture of latently infected S11 cells did not yield any detectable cell-free virus after six days in culture by plaque assay (Fig. 4B ). In contrast, after six days in culture (12 days post-infection), cell-free supernatant from infected MB114 cells yielded significant viral titer, and this titer was significantly higher than that of supernatant taken from infected 3T3 cultures. Therefore, endothelial cells continued to release new infectious virions for at least 12 days post-infection. Infected endothelial cell supernatant contained 100-fold more virus than that of a lysed, infected, fibroblast culture. After the initial observation that cHV68 infection of several endothelial cell lines from different anatomic locations resulted in a population of cells that escaped lysis, we focused our analysis on this outcome in MB114 endothelial cells. Beginning at the time of collection (day six post-infection), we counted intact, trypan blue excluding cells every three days. Infected MB114 cells remained viable for at least 30 days post-infection (Fig. 5A) . During this course of infection, the number of intact cells in the culture increased from 7.08610 5 (60.16 SEM) to 8.51610 6 (60.28 SEM) during the first six days of culture, a ten-fold increase. A corresponding increase in the percentage of membrane viable cells also occurred during this time. Subsequent to day 12 post-infection, the total number of intact cells began to decrease, as did the overall viability of the culture (Fig. 5A ). The observed increase in both intact cell number and percent viability during the first six days of culture, concurrent with positive gp150 staining ( Fig. 2C and Fig. 3C ), indicated that endothelial cells were intact and proliferating in culture while maintaining the lytic transcription program. To further examine the proliferation observed during the first six days of culture, we stained MB114 cells with carboxyfluoroscein (CFSE) prior to infection. CFSE is a fluorescent molecule used to measure cell proliferation, in that each time a cell divides, the two daughter cells contain half the CFSE of the parent cell. S11 cells demonstrated a drop in CFSE signal intensity after six days in culture, consistent with cell division. At the time of harvest (day six post-infection), MB114 cells also demonstrated a drop in CFSE signal intensity consistent with multiple cell divisions (Fig. 5B ). After six more days in culture (12 days post-infection) CFSE signal intensity dropped further, consistent with continued cellular proliferation. Thus, endothelial cells which remained intact after cHV68 underwent multiple rounds of proliferation. Note that by 30 days post-infection, however, the viability of the culture was quite low in comparison to earlier time points (Fig. 5A ). To further test the fate of the surviving endothelial cells, we examined viability of infected cells by staining with annexin V and PI (Fig. 5C ). MB114 cells were 92.4% (60.6 SEM) viable (annexin V negative, PI negative) at two days post-infection and 72.5% (614.5 SEM) viable at four days post-infection. At six days postinfection all MB114 cells were non-adherent and 31.0% (610.74 SEM) of these non-adherent cells were viable. In contrast to MB114 cells, dying cells dominated infected 3T12 cultures at six days post-infection These data indicate that many MB114 cells died early during infection, and in agreement with our proliferation data ( Fig. 5A and 5B ) the population of viable cells increased to 86.4% (60.6 SEM) of the culture by 12 days postinfection. However, by 30 days post-infection dying cells finally dominated the infected culture. These data support that the subset of cells surviving and proliferating at six days post-infection were unique their extended survival, but did not remain viable indefinitely. Of significance, during the course of this infection most of the cells expressed the lytic glycoprotein gp150 on their surface at both early and late time points (Fig. 2C and Fig. 3C) . Moreover, the percentage of cells expressing gp150 did not change markedly with time ( Fig. 2C and Fig. 3C ). These data indicate that the vast majority of viable, infected endothelial cells were undergoing active viral replication, and that these cells were neither latently infected nor had escaped infection. Intact endothelial cells were altered in size, shape, and surface marker expression after cHV68 infection To investigate the morphologic differences between uninfected endothelial cells and the endothelial cells harvested at six days post-infection, we quantified cell size and internal granularity by flow cytometric determination of forward versus side scatter. Uninfected MB114 cells exhibited a broad range of light scatter, indicative of a cell population diverse in size and internal granularity. In contrast, infected cells were very uniform in forward and side scatter (Fig. 6A) . PAA treatment immediately following MB114 infection resulted in significantly fewer cells becoming non-adherent (Fig. 2B) . Like the uninfected MB114 cells, MB114 cells treated with PAA following infection exhibited a broad range of forward and side scatter (Fig. 6A) . Thus, cHV68 infection of endothelial cells resulted in a uniform population of cells with altered morphology, and this outcome was dependent on productive viral infection and late viral gene expression. To further investigate the differences between uninfected endothelial cells and the intact cells harvested six days after infection, we compared the expression of cell-surface proteins on these two cell populations by flow cytometry. Because the cells harvested at six days post-infection persisted in culture as nonadherent cells, we chose to investigate surface expression of the adhesion markers intercellular adhesion molecule-1 (ICAM-1, CD54) and vascular cell adhesion molecule-1 (VCAM-1, CD106). Additionally, we analyzed cells for surface expression of Thy1, a cell surface protein which is upregulated on the surface of activated endothelial cells and functions in cell-cell interactions [30] [31] [32] . Uninfected MB114 cells expressed ICAM-1, VCAM-1, and Thy1 (Fig. 6B) . In contrast, infected endothelial cells expressed neither ICAM-1 nor VCAM-1, but did express Thy1. Based on these data, cHV68 infected endothelial cells downregulate cell surface expression of two adhesion molecules, while maintaining surface expression of an activation marker. In the absence of the viral cyclin, endothelial cell viability after cHV68 infection is reduced We demonstrated that cHV68 infection of endothelial cells resulted in a population of intact cells which persisted in culture as non-adherent cells and continued to produce new virus. Our analysis of endothelial cell infection in the presence of PAA indicated that this outcome is influenced by viral DNA replication and/or late gene synthesis ( Fig. 2B and Fig. 4A ). To begin dissecting the mechanism of persistent cHV68 in endothelial cells infected in vitro, we tested the role of the cHV68 viral cyclin (v-cyclin) in this system. We measured expression of the cHV68 vcyclin in infected MB114 endothelial cells and infected fibroblasts by RT-PCR and western analysis. Notably, the cHV68 v-cyclin gene transcript and protein were detectable in infected endothelial cells and fibroblasts ( Fig. 7A and Fig. S3A) . To determine the role of the v-cyclin in endothelial cell infection, we infected MB114 cells with either wildtype cHV68 or a v-cyclin deficient cHV68 (v-cyclin.STOP cHV68) [18] . Cells were harvested at six days post-infection and cultured as described previously. As per previously published reports in fibroblasts, we determined that the v-cyclin was dispensable for cHV68 growth in endothelial cell lines by multi-step growth assays (data not shown Because viability of MB114 cells after infection was impaired in the absence of the v-cyclin, we next determined the effect of this viral gene on persistent viral replication in viable endothelial cells. Cell free supernatants from infected MB114 cells were titered by plaque assay (Fig. 7C) . Infections with both wildtype and vcyclin.STOP virus resulted in persistent viral production up to 30 days post-infection. Total viral titers from v-cyclin.STOP infections were significantly less than wildtype infections at 12 and 24 days, however, an equivalent amount of virus was produced per cell as in wildtype cHV68 infections. Cell surface protein expression of ICAM-1, VCAM-1, and Thy1 was the same following infection with either wildtype or v-cyclin.STOP virus ( Fig. 6B and Fig. S3B ), thus the v-cyclin was not required for the surface phenotype of infected endothelial cells and the surface phenotype of infected endothelial cells did not predict survival. Therefore, upon infection, the primary function of the v-cyclin was to promote the viability of endothelial cells. Primary endothelial cells were infected in vivo, and ex vivo demonstrated prolonged viability while supporting cHV68 growth Given that immortalized endothelial cell lines are inherently different from endothelial cells in vivo, we next determined the outcome of cHV68 infection in primary endothelial cells. Primary endothelial cells were isolated from C57/BL6 mouse lungs and characterized as per previously published methods (Fig. S4) . First, we determined if primary lung endothelial cells could support growth of wildtype cHV68 by multi-step growth assay (Fig. 8A) . Because of the apparent role for the v-cyclin in infection of endothelial cell lines, we also analyzed growth of v-cyclin.STOP virus in primary endothelial cells. Infection proceeded with kinetics comparable to previously published infection of NIH 3T12 fibroblasts [18] . Growth curves did not differ between wildtype and v-cyclin.STOP virus. These data revealed that primary endothelial cells supported growth of cHV68, irrespective of the vcyclin. Second, we analyzed the percent of cells that remained viable after cHV68 infection at a low MOI. Immediately following removal of the viral inoculum (t = 0), mouse embryonic fibroblasts (MEFs) and primary lung endothelial cells did not differ significantly in viability (p.0.05) (Fig. 8B) . However, at 48 hours post-infection with wildtype cHV68, primary lung endothelial cells were significantly more viable than MEFs (p,0.001), and remained significantly more viable at 96 (p,0.05) and 120 hours post-infection (p,0.001). Notably, endothelial cells infected with the v-cyclin deficient cHV68 had reduced viability at 48, 96, and 120 hours post-infection. Given that cells were infected at a very low MOI, it was not surprising that a proportion of MEFs (28.5%611.1 SEM) were viable at 96 hours post-infection. However, unlike endothelial cells, MEFs were mostly lysed by 120 hours post-infection with wildtype virus. These data demonstrate that primary endothelial cells surpass MEFs in viability following cHV68 infection, and that this outcome is promoted by the v-cyclin. Lastly, we examined lung endothelial cells following acute cHV68 infection in vivo. Our in vitro studies revealed that viable, infected endothelial cells were not latently infected, but uniformly supported a lytic viral program (Fig. 2C, Fig. 3C and 3D ). Because persistent viral infection and lytic viral antigen expression is likely to be cleared by an intact immune response, we analyzed endothelial cell infection in lung tissues of immune-deficient CD8-alpha null mice. At six days post-intranasal infection, lung cells enriched for those bearing the endothelial cell marker CD31 were analyzed alongside the remaining CD31 depleted lung cells. Given the possibility that a small percentage of non-endothelial cells can transiently express CD31 (i.e. macrophages and neutrophils), we depleted the CD31 positive cell population of these potential contaminating cells and analyzed resultant cell populations by flow cytometry (Fig. S5B ). PCR analysis with single copy sensitivity for cHV68 gene50 (Rta) was performed to determine the frequency of viral genome positive cells in the CD31-enriched and remaining CD31-depleted lung cell populations (Fig. 8C) . Data from this analysis revealed that approximately 1/102 CD31-enriched cells were viral genome positive, whereas 1/525 remaining CD31depleted lung cells contained viral DNA. Thus, these data support that a surprisingly large proportion of lung endothelial cells were viral genome positive following in vivo infection. To further demonstrate that these viral genome positive cells are actively infected, and to exclude the possibility of endocytic uptake of virus or abortive infection, we performed RT-PCR analysis on endothelial enriched and depleted lung cells (Fig. 8D) . The transcript for the endothelial cell marker CD31 was robustly detected in the endothelial enriched lung cells from each of the four mice analyzed, whereas a very low level of this transcript was detected from the CD31 depleted lung cells of only one out of the four mice analyzed. In combination with flow cytometric analysis (Fig. S5B) , these data support that our enrichment strategy was effective in isolating CD31+ cells from total lung cells. Additionally, the viral pol III-1 transcript was detected from both endothelial cell enriched and depleted lung cells of cHV68 infected mice, and was absent from lungs of mock infected mice. Detection of viral gene transcripts from lung cells indicates virus infection of these cells. Notably, within each infected mouse, pol III-1 detection was comparable between endothelial enriched and CD31-depleted lung cell populations, suggesting that detection of this viral transcript in the endothelial cell population was not due a few contaminating infected non-endothelial cells. Taken together, these data demonstrate cHV68 infection of endothelial cells in vivo. Herpesviruses have been implicated as potential initiators of a variety of endothelial pathologies [1] [2] [3] [4] . Given the intimate interactions observed between herpesviruses and endothelial cells, and the systemic spread of cHV68 during infection, we characterized the outcome of endothelial cell infection with cHV68, a small animal model for the human gammaherpesviruses. While cHV68 replicated comparably in endothelial cells and fibroblasts up to 36 hours post-infection, infected cultures of endothelial cells had a high percentage of viable, non-adherent cells that remained following infection. Of significance, these viable, non-adherent cells had not simply escaped infection, but instead were actively infected, as determined by the presence of multiple markers of the lytic replication program. While the absolute number of viable, infected endothelial cells varied among different cell lines and primary cells, the prolonged viability of endothelial cells in comparison to fibroblasts was remarkably consistent in endothelial cell lines from distinct anatomic locations, as well as in primary endothelial cells. Furthermore, optimal survival of both endothelial cell lines and primary endothelial cells was dependent on the presence of the cHV68 v-cyclin, indicating that this outcome was a process actively promoted by virus infection. This conserved outcome of endothelial cell infection is particularly striking given that endothelial cells display phenotypic heterogeneity in structure and function depending on anatomic location [33] [34] [35] . Based on the persistent infection observed in diverse endothelial cell lines, and heightened viability of primary endothelial cells following infection, we propose that cHV68 may have evolved machinery to specifically promote persistent infection in endothelial cells. Endothelial cells serve as a natural site of infection and possible viral reservoir of HCMV [36] [37] [38] [39] [40] , suggesting a role for HCMVinfected endothelial cells in viral spread and persistence. Additionally, recent reports implicate circulating endothelial progenitor cells as potential reservoirs of KSHV and possible precursors of KS spindle cells [41, 42] . However, the specific mechanisms by which infected endothelial cells contribute to the pathogenesis of these human viruses remains unclear. Murine cHV68 pathogenesis involves dissemination from the lung to lymph nodes, spleen, and peritoneum [12, 15] . In light of this systemic spread, cHV68 likely encounters an endothelial cell barrier. The human gammaherpesvirus KSHV causes a serious endothelial cell malignancy, KS, which predominantly occurs in immunecompromised individuals (e.g. AIDS patients). KS tumors are comprised of distinctive spindle cells of endothelial origin and a variable inflammatory infiltrate [43] [44] [45] [46] . KSHV is detected primarily in the endothelial component of the lesion, and though most of these cells harbor latent KSHV, a subset of them enter the lytic cycle [47, 48] . While there is precedent for a mixed infection type within KS tumors (i.e. both lytic and latent infection), the history of the lytically infected cells remains elusive, though recent reports point to circulating endothelial progenitor cells [41, 42] . In vitro, endothelial cell infection with KSHV is predominantly latent [43, [49] [50] [51] . However, when these cells are transferred into mice, they show evidence of lytic gene expression and virus production [52] . Because viral DNA replication and/or late gene synthesis were important for endothelial cell outcome of cHV68 infection (Fig. 2B and Fig. 6A ), our findings indicate that an active viral process occurred within the cells to yield these changes. Consistent with this hypothesis, we identified that the cHV68 v-cyclin is required for optimal endothelial cell viability after cHV68 infection in vitro. The cHV68 v-cyclin promotes cell cycle progression in primary lymphocytes and can function as an oncogene in transgenic mice [18] . While the v-cyclin is critical for reactivation from latency, to date the v-cyclin has been dispensable in all assays of lytic replication in vitro [53, 54] . Given the lytic nature of endothelial cell infection, the contribution of the v-cyclin to optimal endothelial cell viability after cHV68 infection indicates a role for the v-cyclin outside of its requirement in reactivation from latency. Additionally, the apparent role of the v-cyclin in this prolonged infection of both endothelial cell lines and primary lung endothelial cells may explain the slight decrease in lung titers that resulted in mice infected with low doses of v-cyclin.STOP cHV68 and other vcyclin mutant viruses compared to wildtype cHV68 [55] . Although the precise mechanism by which the v-cyclin promotes endothelial cell viability is unknown at this time, it is possible that the v-cyclin provides growth cues that allow for anchorageindependent growth of endothelial cells after cHV68 infection. While our initial analysis focused on the role of the v-cyclin, it is very likely that additional viral genes facilitate persistent endothelial cell infection. Lead candidates include the anti-apoptotic viral bcl-2 gene M11 and the viral GPCR (ORF 74), whose homologs in KSHV have been implicated in endothelial cell survival and transformation [56] [57] [58] [59] . Additional candidates for optimal endothelial cell infection are the ribonucleotide reductase homologs, ORF 60 and 61 of cHV68. Although the role of these genes in cHV68 infection of endothelial cells is currently untested, the MCMV ribonucleotide reductase homolog is required for in vivo replication and pathogenesis of this endothelial cell-tropic virus [60] . One of the most noticeable alterations of the persistently infected endothelial cells described here is the extent of change in their cellular morphology and properties. Though we did not test the oncogenic potential of these cells, endothelial cells achieved anchorage-independent growth, a property frequently associated with oncogenic transformation. These cells also underwent significant changes in protein expression on the cell surface, with down-regulation of the cellular adhesion proteins ICAM-1 (CD54) and VCAM-1 (CD106). While there was decreased expression of ICAM-1 and VCAM-1, changes in cell surface expression were not global, since Thy1 expression remained positive on these cells (Fig. 6B) . It is worth noting that infection of endothelial cells with KSHV results in down-regulation of MHC class I, PE-CAM (CD31), and ICAM-1 (CD54), but not LFA-3 (CD58) or Fas (CD95), and the viral genes K3 and K5 have been demonstrated to regulate this outcome [61] [62] [63] [64] [65] [66] . The contribution of the mK3 gene of cHV68 to persistent endothelial cell infection remains untested. Down-regulation of certain surface markers but not others during viral infection indicates a specific phenomenon, rather than global down-regulation, and provides further evidence that an active process was responsible for the observed endothelial cell outcome of cHV68 infection. While cHV68 may have evolved mechanisms that promote persistent infection of endothelial cells (e.g. the v-cyclin), it is also possible that endothelial cells have a cellular program that limits the cellular lysis typically observed during productive infection. This putative cellular adaptation may be particularly important in limiting destruction of blood vessel integrity during viral infection. A recent report of host cell response to cHV68 in three different cell types (fibroblasts, endothelial precursor cells and macrophages) identified 148 genes whose expression was altered in endothelial precursor cells, but not macrophages or fibroblasts [67] . Taken together with the unique endothelial cell outcome of cHV68 infection reported here, these data make a compelling argument for cell type specific responses to productive cHV68 infection. We have provided evidence for cHV68 infection of endothelial cells in vivo at early times post-infection. Interestingly, infected cells (including endothelial cells) were not abundant in acutely infected lung tissue, as little to no GFP was detected by flow cytometric analysis of unfractionated lung cells after infection with a GFPmarked virus (data not shown), but required sensitive PCR methods for detection. Our in vitro analysis demonstrated a change in surface phenotype of infected endothelial cells which could correspond to substrate detachment and release into circulation. This is an intriguing idea in light of published reports of circulating endothelial cells and progenitor endothelial cells in virus infection. Our in vitro studies implicated endothelial cells as a persistent source of virus production, however, the extent to which endothelial cells contribute to cHV68 persistence in vivo, and to what degree it might be influenced by host immune status, are important issues that remains to be addressed. While it is unlikely that an intact immune response would permit such long-term expression of viral antigens in vivo, we hypothesize that persistently infected endothelial cells could provide a significant source of continued virus replication in immune-compromised individuals (i.e. AIDS patients), a context in which gammaherpesvirusassociated pathology frequently occurs. Given the unusual properties of these viable and infected endothelial cells (e.g. anchorage-independent growth and altered cell surface protein expression), it will be important to critically address the potential role of endothelial cells as a reservoir for infection in vivo in both immune-competent and immune-deficient individuals. In conclusion, our data provide evidence for prolonged gammaherpesvirus infection in endothelial cells. This outcome appears to be the result of a specific interaction between cHV68 and endothelial cells, as it is promoted by a viral gene (the cHV68 v-cyclin) and, to date, is unique to endothelial cells. These data further refine the concept of gammaherpesvirus infection and demonstrate that a gammaherpesvirus can undergo robust productive replication in the context of prolonged host cell viability. Identification of intermediate outcomes of gammaherpesvirus infection, such as the one we have described here, has major implications for our understanding of the nature of gammaherpesvirus infection as it relates to specific cell types. Such a course of infection provides an additional mechanism, beyond latency, by which gammaherpesviruses can achieve longterm propagation. Mouse fibroblast cell lines 3T3 (ATCC CRL-1658) and 3T12 (ATCC CCL-164) and mouse endothelial cell lines MB114 [68] , SVEC4-10 (ATCC CRL-2181), and CD3 [69] were cultured in Dulbecco's Modified Eagle Media (DMEM) supplemented with 5% FBS (Hyclone, Logan, UT), 2 mM L-glutamine, 10 U/mL penicillin, and 10 mg/mL streptomycin sulfate. S11 [20] and S11E tumor cells [22] were cultured in RPMI 1640 medium (Gibco) supplemented with 10% FBS, 50 mM b-mercaptoethanol, 1 mM sodium pyruvate, 2 mM L-glutamine, 10 U/mL penicillin, and 10 mg/mL streptomycin sulfate (complete RPMI). Mouse embryonic fibroblasts were isolated from C57/BL6 mice as previously described [70] and cultured in DMEM supplemented with 10% FBS, 2 mM Lglutamine, 10 U/mL penicillin, 10 mg/mL streptomycin sulfay7te, and 250 ng/mL fungizone. Isolation, characterization, and culture of primary endothelial cells from C57/BL6 mice was done according to previously published methods [71] and is described in Protocol S1, Text S1 and Figure S3A . cHV68 WUMS (ATCC VR-1465) and all recombinant viruses were grown and titered as previously described [72] . DK3TET 2 cHV68 (cHV68-GFP) containing a GFP cassette under the control of an immediate early CMV promoter was generated and characterized by Dr. Phillip Stevenson [73] . cHV68 containing a stop codon within ORF 72 (v-cyclin.STOP. cHV68) was previously described [18] . All infections were carried out at a multiplicity of infection (MOI) of 5 plaque forming units (PFU) per cell. Inoculum was removed after one hour of infection at 37uC, cell monolayers rinsed three times with sterile phosphate buffered saline (PBS), and complete media added. Intact and non-adherent cells were collected at six days post-infection, at which time cells and media were collected, counted to determine post-infection viability by trypan blue exclusion, and then centrifuged for 10 minutes at 2086g. Cell pellets were washed twice in sterile PBS and then resuspended in complete RPMI at a concentration of 5610 5 viable cells/mL for continued culture. Cells were counted every three days of culture, and adjusted to a concentration of 5610 5 viable cells/mL. Every six days of culture, cells were counted and centrifuged for 10 minutes at 2086g to remove cell-free cHV68 and cellular debris. Pellets were washed twice in sterile PBS and resuspended in RPMI at a concentration of 5610 5 viable cells/ mL. Culture scheme is depicted in supplement (Fig. S1 ). To measure virus replication, infected samples were analyzed by plaque assay at various times post-infection. To measure cell-free virus titer from infected cultures, supernatants were collected every six days of culture and analyzed by plaque assay. Samples were thawed, serially diluted, and plated onto NIH 3T12 cells in 12 well plates in triplicate. Infection was performed at 37uC for one hour. Cells were overlaid with a 1:1 mix of DMEM and carboxymethylcellulose plus fungizone (final concentration 250 ng/mL). Plates were incubated for seven days at 37uC. On day seven, carboxymethylcellulose was removed, and plates were stained with 0.35% methylene blue in 70% methanol and rotated for 15-20 minutes, before rinsing with water and counting on a light box. All titers were determined in parallel with a laboratory standard. For propidium iodine (PI) viability studies, cells were washed twice in PBS (five minutes, 10006g), resuspended in a 0.5 mg/mL PI solution, and incubated for 15 minutes. Following incubation, cells were centrifuged (five minutes, 10006g) and washed in a solution of PBS, 2% fetal calf serum, and 0.1% NaN 3 (buffer A). Cells were fixed in 1% paraformaldehyde, and analyzed by flow cytometry. For two parameter viability analysis, cells were washed in 1X annexin V binding buffer (BD Bioscience, San Jose, CA), resuspended in 0.5 mg/mL PI and 5 mL annexin V-FITC (BD Bioscience), incubated for 15 minutes, washed in binding buffer, fixed in 1% paraformaldehyde, and analyzed by flow cytometry. For carboxyfluorescein (CFSE) proliferation studies, MB114 cells and Sll cells were washed twice in PBS (five minutes, 10006g) and resuspended in a solution of PBS and 2% fetal calf serum (buffer B) at a concentration of 1610 6 cells per mL. An equal volume of 4 mM CFSE in buffer B was added to the cell suspension (2 mM final concentration) and pipet mixed. After three minutes the labeling reaction was quenched with an equal volume of fetal calf serum for 30 seconds, and buffer B was added for a total volume of 50 mL. Cells were centrifuged (five minutes, 10006g), resuspended in buffer B, and aliquots collected for day 0 analysis. Remaining, labeled cells were resuspended in complete media and cultured. Labeled MB114 cells were infected at an MOI = 5 PFU/ cell. Infected cells were harvested as described at day six postinfection and analyzed by flow cytometry at six and 12 days post-infection. Stained and unstained S11 cells were analyzed at the same time as a positive control. For surface marker staining, cells were washed twice in buffer A (five minutes, 10006g) and resuspended in primary antibody (1:200 in buffer A, 25% 24G2 [74] ). The following primary antibodies were used: CD106-biotin (Rat, IgG2a k, clone 429 (MVCAM.A)), CD54-FITC (Armenian hamster, IgG1 k, clone 3E2), Thy1.2-APC (Rat, IgG2a k, clone 53-2.1) (BD Bioscience), monoclonal mouse anti-cHV68 gp150 (mouse IgG2a, a kind gift from Dr. Phillip Stevenson) [75] . Cells were then incubated for 45 minutes at room temperature. Following incubation, cells were centrifuged twice in buffer A (five minutes, 10006g), resuspended in either buffer A or secondary staining reagent in buffer A (1:500), and incubated for 20-30 minutes at room temperature. Secondary staining reagents were either streptavidin-APC (BD Bioscience) or anti-mouse IgG2a-FITC (Rat IgG1 k, clone R19-15) (BD Bioscience). Cells were centrifuged twice in buffer A (five minutes, 10006g) and analyzed by flow cytometry. For forward versus side scatter analysis, MB114 cells harvested at six days post-infection were analyzed. Uninfected MB114 cells were detached from flasks with 0.5 mM EDTA and analyzed in parallel with infected cells. In certain experiments, to determine the effect of viral DNA replication and/or late gene synthesis on forward and side scatter properties, MB114 cells were treated with 200 mg/mL phosphonoacetic acid (PAA) after one hour of infection. At six days post-infection, non-adherent cells were harvested and counted by trypan blue exclusion to determine the percent of cells infected that were viable and non-adherent at time of harvest. Cells which remained adherent to the flask were then detached with 0.5 mM EDTA, combined with the harvested nonadherent cells, and analyzed by flow cytometry for forward and side scatter properties. Effective block of late gene synthesis was confirmed by plaque assay titer of treated cells. Viral titer was reduced 96.2% after 24 hours and 99.9% after six days as compared to untreated cells. 10 micron control beads were run in parallel with each experiment (Beckman Coulter, Fullerton, CA). Cells were collected and boiled in Laemmli buffer for 10 minutes. Total protein concentration of each cell extract was determined by Lowry assay (DC protein assay kit, Bio-Rad, Hercules, CA). 10-20 mg of total protein per extract was separated by sodium dodecyl sulfate-polyacrylamide gel electrophoresis. Proteins were electrotransferred (ThermoFisher Scientific, Portsmouth, NH) to PVDF membranes (Millipore, Billerica, MA) and blocked in PBS with 0.05% Tween-20 and 5% nonfat milk for one hour at room temperature. Western blots for cHV68 protein expression were incubated with 10 mg/mL monoclonal mouse anti-cHV68 gB [76] (kind gift of Dr. Phillip Stevenson), monoclonal mouse anti-cHV68 M3 at 1:50, polyclonal rabbit anti-cHV68 v-cyclin at 1:2000 (kind gifts of Dr. Herbert Virgin), and monoclonal antibody to mouse b-actin at 1:1000 (Sigma Chemical, St. Louis, MO). Blots were washed for 45 minutes in PBS with 0.05% Tween-20, then incubated with donkey antimouse or donkey anti-rabbit antibodies at 1:6000 for one hour. Blots were then washed in PBS containing 0.1% Tween-20 for 45 minutes. Proteins were visualized using an ECL Plus Western blotting detection kit (Amersham Pharmacia Biotech). Total RNA was extracted from S11 cells, infected NIH 3T3 cells, infected MB114 cells, and murine lung cell fractions using mirVANA TM miRNA Isolation kit (Ambion, Austin, TX), per manufacturer's instructions. Amplifications were conducted using the following primer sets: cHV68 polIII-1 forward 59 CAA CAG GTC ACC GAT CC 39, cHV68 polIII -1 reverse 59 GGA AGT ACG GCC ATT TC 39, cHV68 M2 forward 59 TAA GGA CCT CGT AGA GAT TGG C 39, cHV68 M2 reverse 59 ACG TTA AAG TCC CCA TGG AAG C 39, cHV68 v-cyclin forward 59 ATT AGC ACT GGG CGT TTC ATG 39, cHV68 v-cyclin reverse 59 GAC CTC CGT CAG GAT AAC AAC Cells were pelleted (five minutes, 2086g), resuspended in 1 mL PBS, and pelleted again (five minutes, 2086g). Supernatant was removed and 2.5% glutaraldahyde solution (adjusted to pH 7.4 using HCl and to 400 mOsm using CaCl 2 ) added. Samples were processed and imaged by Dr. Gary Mierau, The Children's Hospital Department of Pathology/Laboratory Services, Aurora, CO. Briefly, samples were post-fixed in 2% cacodylate buffered osmium tetroxide (pH 7.4), dehydrated in a graded series of alcohols, and embedded in epoxy resin. Sections, approximately 80 nm in thickness, were stained with uranyl acetate and lead citrate prior to examination at 60 kV with a Zeiss EM-10 transmission electron microscope (Carl Zeiss Inc, Thornwood, NY). Lung tissues were removed from CD8-alpha knock-out mice six days post-intranasal inoculation of 1610 6 PFU cHV68. Tissues were enzymatically digested and endothelial cells purified using the endothelial cell marker CD31 (PECAM-1) (Fig. S4A) . Briefly, cells were stained with the following antibodies: anti-CD31-biotin (PECAM-1, clone MEC 13.3, rat IgG2a k) (BD Bioscience), anti-CD45-PE (B220, clone RA3-6B2, Rat IgG2a k) (BD Bioscience), anti-F4/80-PE (clone BM8, Rat IgG2a k) (eBioscience, San Diego, CA), anti-CD4-PE (L3T4, clone RM4-5, Rat IgG2a k) (eBioscience), anti-CD8a-PE (Ly-2, clone 53-6.7, Rat IgG2a k) (BD Bioscience), and anti-Ly-6G and Ly-6c-PE (Gr-1, clone RB6-8C5, Rat IgG2b k) (BD Bioscience). Cells were magnetically labeled with Anti-Biotin MultiSort beads, separated with Octo-MACS separation unit, and released from the MultiSort beads as per manufacturer's instructions (Miltenyi Biotec Inc., Auburn, CA). Released cells were incubated with Anti-PE MultiSort beads and depleted of contaminating PE positive cells by magnetic separation. Flow cytometric analysis was performed to confirm flow through fraction from PE column as enriched in CD31 positive cells and depleted of PE positive cells (Fig. S5B) . Limiting dilution nested PCR detection of cHV68 genome-positive cells The frequency of lung cells from CD8-alpha knock-out mice containing cHV68 genome was determined using a previously described nested PCR assay (LD-PCR) with single-copy sensitivity to detect gene 50 of cHV68 [77] . Briefly, freshly isolated cells were counted, resupsended in isotonic solution, and then diluted in 10 4 uninfected NIH 3T12 cells prior to serial dilution plating. Plated cells were lysed overnight in proteinase K, followed by two rounds of nested PCR. Reactions were performed on 12 replicates per dilution per sample and products resolved on a 2% agarose gel and identified by ethidium bromide staining. PCR sensitivity was quantitated using 10, 1, or 0.5 copies of a gene 50 containing plasmid (pBamH I N) diluted in 10 4 uninfected NIH 3T12 cells. Data were analyzed using GraphPad Prism software (GraphPad Software, San Diego, CA). Data were analyzed using the paired Student's t test to determine statistical significance. Limiting dilution data were subjected to nonlinear regression analysis, and the frequency of genome-positive cells was calculated using the Poisson distribution to assume that the cell number at which 63.2% of events were detected corresponded to the occurrence of a single event. Text S1 Endothelial cell specific markers included CD31 and CD54. CD80 and CD86 were included as non-endothelial cell specific markers, though a previous report has identified low level CD80 expression on primary murine lung endothelial cells [71] . Fluorescence was determined relative to unstained cells (grey). Cell morphology and surface expression was similar to previously characterized primary endothelial cells [71] . Found at: doi:10.1371/journal.ppat.1000152.s006 (0.89 MB TIF) Figure S5 Isolation and characterization of murine lung endothelial cells infected in vivo. (A) Primary murine lung endothelial cells were isolated for analysis of infection in vivo as described in Materials and Methods. Lung cells were stained for the endothelial cell marker CD31, as well as with a cocktail of PE labeleled antibodies specific to the following cell types (which includes potential contaminating infected cells): macrophages, granulocytes, CD8 + T lymphocytes, CD4 + T lymphocytes, and B lymphocytes. CD31 positive cells were enriched from total lung cells and then finally depleted of cells stained with the PE cocktail. (B) Flow cytometric analysis of lung cell separation following in vivo infection was performed on total lung cells and PE depleted/ CD31+ enriched cells. Left and middle panels show representative data (1610 6 cells per stain) from an infected mouse (Fig. 8D ). Gates were set on unstained cells, and data are representative of the four mice analyzed in Fig. 8D . Splenocytes (right panel) were included as a staining control. Found at: doi:10.1371/journal.ppat.1000152.s007 (0.20 MB TIF)

Analysis of synonymous codon usage and evolution of begomoviruses

Begomoviruses are single-stranded DNA viruses and cause severe diseases in major crop plants worldwide. Based on current genome sequence analyses, we found that synonymous codon usage variations in the protein-coding genes of begomoviruses are mainly influenced by mutation bias. Base composition analysis suggested that the codon usage bias of AV1 and BV1 genes is significant and their expressions are high. Fourteen codons were determined as translational optimal ones according to the comparison of codon usage patterns between highly and lowly expressed genes. Interestingly the codon usages between begomoviruses from the Old and the New Worlds are apparently different, which supports the idea that the bipartite begomoviruses of the New World might originate from bipartite ones of the Old World, whereas the latter evolve from the Old World monopartite begomoviruses.

Synonymous codon usage bias has been investigated in many organisms, as the genetic code is degenerate. The synonymous codons are also non-randomly used in viruses infecting living organisms. Several factors such as mutational bias (Jenkins and Holmes, 2003; Gu et al., 2004; Zhou et al., 2005) , translational selection (Sau et al., 2005a; 2005b; 2005c) , gene function (Wang et al., 2002; Gu et al., 2004; Zhou et al., 2005) , gene length (Sau et al., 2005a) , and CpG island (Shackelton et al., 2006) were found to influence codon usage in animal viruses and phages, and mutational bias was found as the major determinant factor. Adams and Antoniw (2004) also suggested that mutational bias rather than translational selection was the major determinant of codon usage variation amongst plant viruses. Geminiviruses (family Geminiviridae) are single-stranded DNA (ssDNA) viruses that cause severe disease in major crop plants worldwide. Most geminiviruses belong to the genus Begomovirus, which are transmitted exclusively by the whitefly Bemisia tabaci (Harrison and Robinson, 1999) . Many begomoviruses have bipartite genomes known as DNA A and DNA B. DNA A contains the AV1 (coat protein) and AV2 ORFs (open reading frames) in the virus strand and, on the complementary strand, four ORFs: AC1 (replication initiation protein), AC2 (transcriptional activator protein), AC3 (replication enhancer protein) and AC4. The virus and complementary strands of DNA B contain two ORFs: BV1 (nuclear shuttle protein) and BC1 (movement protein). Some begomoviruses have a monopartite genome and lack DNA B. Phylogenetic analysis shows that begomoviruses can be generally divided into two groups, the Old World begomoviruses (eastern hemisphere, Asia, Africa, Europe, the Mediterranean areas) and the New World begomoviruses (western hemisphere, the Americas). All the New World begomoviruses are evolved to bipartite with lack of AV2 ORF in DNA A, whereas both bipartite and monopartite begomoviruses in the Old World encode AV2 ORF (Harrison and Robinson, 1999) . Because of their destructive effect on cash crops (Moffat, 1999; Moriones and Navas-Castillo, 2000; Mansoor et al., 2006) , numerous studies on begomoviruses have been conducted to understand their symptoms, host range, distribution, genome structure, gene function, and so on Zhou et al., 2003) . In this paper, we report the analysis of codon usage bias in begomoviruses and also perform an evolutionary analysis based on their codon usage pattern. The complete genomic sequences of 147 begomovirus species were downloaded from the GenBank database, from which a total of 932 variants of the 8 known protein-coding genes were extracted. To minimize sampling errors, 915 variants were selected for further analysis by the following sifting standard: (1) the selected genes should be complete coding DNA sequences (CDS) with correct initial and terminal codons; (2) only those CDS including at least 80 codons were selected in the dataset; (3) those CDS with uncertain annotation or annotated as hypothetical protein-coding genes were excluded from this study. GC content is the frequency of G+C in a coding gene. GC1, GC2 and GC3 contents are the frequencies of G+C at the first, second and third positions of codons, respectively. A3, T3, G3 and C3 contents are the frequencies of A, T, G and C at the synonymous third position of codons, respectively. Effective number of codons (N c ), ranging from 20 to 61, is generally used to measure the bias of synonymous codons. When N c value approaches 20, only one codon is used with extreme bias for one amino acid and, if the value is up to 61, the anonymous codons are used equally with no bias (Wright, 1990) . Relative synonymous codon usage (RSCU) is defined as the ratio of the observed frequency of codons to the expected frequency given that all the synonymous codons for the same amino acids are used equally. RSCU values have no relation to the amino acids usage and the abundance ratio of synonymous codons, which can directly reflect the bias of synonymous codon usage (Sharp and Li, 1986) . The codon adaptation index (CAI) was used to estimate the extent of bias towards codons that were known to be preferred in highly expressed genes (Sharp and Li, 1987) . It is now proved that CAI values mostly approach the theoretical values to reflect the expression level of a gene. Thus it has been widely utilized to measure the gene expression level (Naya et al., 2001; Gupta et al., 2004) . A CAI value ranges from 0 to 1.0, and a higher value indicates a stronger codon usage bias and a higher expression level. CodonW version 1.4.2 (John Peden, available at http://sourceforge.net/projects/codonw/), an integrated program, was utilized for calculating GC, GC3 contents and N c values and then carrying out correspondence analysis (CA), while GC, GC1, GC2 and GC3 contents were calculated by practical extraction and report language (PERL) scripts which were written by us. A3, T3, G3 and C3 contents, as well as RSCU and CAI values, were also calculated by using PERL scripts. CA is the most commonly used multivariate statistical analysis at present (Greenacre, 1984) . This method can successfully present variation trends among genes, and then distribute them along the continuous axis by using RSCU value as variable data. In CA, all genes were plotted in a 59-dimensional hyperspace, according to the usage of the 59 sense codons. Major variation trends can be determined using these RSCU values and genes ordered according to their positions along the major axis, which can also be used to distinguish the major factors influencing the codon usage of a gene. The set of reference sequences used to calculate CAI values in this study were the genes coding for coat proteins. According to the calculated CAI values, 5% of the total genes with extremely high and low CAI values were regarded as the high and low dataset, respectively. Then we calculated the average RSCU values of the two gene samples and subtracted them subsequently in each dataset group (ΔRSCU). If the ΔRSCU values are larger than 0.08, then this codon will be defined as the optimal codon (Duret and Mouchiroud, 1999) . In order to examine the base composition variation among different genes, the base composition of different protein-coding genes was calculated. Table 1 shows that with the exception of the AC3 gene that has a lower average GC content (0.391), no obvious difference in GC content was found among other tested genes. However, differences in GC content at the different synonymous positions of codons were apparent. For example, GC content at the synonymous first position of codons for the AV2 gene is 0.563, while that for the AC4 gene is only 0.412. GC content at the synonymous third position of codons for the AC4 gene is 0.534, while that for the BV1 gene is only 0.399. It was also observed that the average percentage of GC content was generally higher at the first than at the second codon position (Table 1) , except for the AC4 gene whose GC2 content was larger than GC1 content. This result suggested that the AC4 gene might have a special codon usage pattern. In addition we found that AV1 and BV1 genes had a tendency to usage bias at the synonymous third codon position. The AV1 gene does not tend to use A(T)-ending or G(C)-ending codons but tends to use T-ending codons relative to A-ending codons. However, BV1 gene apparently uses T-ending codons and seldom used C-ending codons. Therefore, AV1 and BV1 genes should show a stronger codon usage bias among the begomovirus genes. An N c plot (a plot of N c vs GC3 content) was widely used to investigate the determinants of the codon usage variations among genes in different organisms. It was suggested that if GC3 content was the only determinant of the codon usage variation among the genes, then the N c value would fall on the continuous curve between N c value and GC3 content (Wright, 1990 ). In general, if genes are distributed in N c plots approaching the expected continuous curve with no selection, then codon usage bias of genes is mainly influenced by compositional constraints. Otherwise the codon usage bias of genes is more affected by other factors such as translational selection, etc. The N c plot for genes of the 147 begomoviruses showed that although a small number of genes were located on the expected curve, most points lay far below the expected curve ( Fig.1) , suggesting that apart from mutation bias, other factors might also play a role in shaping the codon usage bias of begomoviruses (Guo et al., 2007) . Table 1 The GC content at the different codon positions and the A, T, G and C contents at the third position of begomovirus genes Base composition analysis suggested that the codon usage of the AC4 gene, which is always embedded in the AC1 gene, might differentiate from other genes. In addition, strong codon bias was observed in AV1 gene. Because of the special characteristics of the abovementioned genes, we selected them for examining the influence factors in shaping the codon usage (Fig.2a) . The N c plots of AV1, AC1 and AC4 genes suggested that apart from compositional constraints, other factors might play important roles in shaping their codon usage, although high translational selection seems to lay stress on the AC4 gene because of the wide range of N c values for the same GC3 content. To examine the reason for N c variation under the same GC3 content, the relationship between the GC1, GC2 and GC3 contents for AV1, AC1 and AC4 genes was further examined (Fig.2b) . It was observed that the GC1 content was always higher than GC2 content for the AV1 and AC1 genes, whereas the GC1 content of the AC4 gene was generally lower than the GC2 content. This result was coincident with the base composition analysis (Table 1) . Thus the variations of the synonymous first and second codon positions for the AC4 gene might be caused by the translational selection utilizing G or C at the synonymous second position. As the AC4 gene is always embedded in the AC1 gene, the codon usage pattern of the AC4 gene might be influenced by the AC1 gene. The sequences of AC1 and AC4 genes were aligned. It was found that the synonymous third codon position of the AC1 gene corresponded to the second codon position of the AC4 gene for all the begomoviruses. Thus the base compositional environment of the AC1 gene might also influence the codon usage pattern of the AC4 gene. It could be postulated that the second codon position of the AC4 gene that tended to use G or C should be influenced by compositional constraints other than translational selection. According to the variation analysis of base composition for the genes of begomoviruses, the AV1 gene was selected as a reference dataset to calculate the CAI values of all genes. The correlation analysis between CAI value and the positions of the genes along the first two major axes (generated by correspondence analysis), as well as other indices, was then calculated. It was observed that the CAI values were negatively correlated with the GC3 and GC contents (r=−0.234 and r=−0.535 respectively, P<0.01), while significantly positively correlated with axis 1 (r=0.587, P<0.01). Moreover the CAI value was also significantly negatively correlated with the N c value, indicating a tendency to a higher CAI value or a lower N c value and a higher expression level for begomovirus genes. Thus it is feasible to use the AV1 gene as a reference dataset for our estimation of the CAI value of begomovirus genes. Based on the calculated CAI values, 5% of the total genes with extremely high and low CAI values were regarded as the high and the low datasets, respectively. Then we compared the codon usage of the high dataset to the low dataset. Table 2 shows that 14 codons that code 13 amino acids were apparently used at a high level, and can be determined as translational optimal codons. Out of the 14 optimal codons, 5 were ended with G, 1 with C and 8 with T. CA was performed on the RSCU value based on the concatenated genes for each begomovirus genome. Fig.3 shows the positions of all tested begomoviruses along the first two major axes. The first major axis accounted for 15.2% of the total variations, while the second major axis accounted for 9.9% of the total variations. In order to detect the codon usage variation of different genomes, the begomoviruses were divided into three groups including the Old World begomoviruses with monopartite genomes (OM), the Old World begomoviruses with bipartite genomes (OB), and the New World begomoviruses with bipartite genomes (NB). The distribution of the three groups along the first two major axes showed that the Old World monopartite begomoviruses and the New World bipartite ones were located in two independent fields, indicating that the two groups of begomoviruses exhibit a different codon usage pattern. Because the species with a close genetic relationship always present a similar codon usage pattern (Sharp et al., 1988) , the genetic relationship of the two groups of begomoviruses should be far removed from each other. As to the Old World begomoviruses with bipartite genomes, we found that the majority of them exhibited a similar codon usage pattern with the Old World monopartite begomoviruses, and a few of them showed a similar codon usage to the New World bipartite begomoviruses. An explication of this result might be that a number of Old World bipartite begomoviruses evolved to adopt the codon usage pattern of some Old World monopartite begomoviruses. Moreover, the New World bipartite begomoviruses were closely related to a small number of the Old World bipartite ones and far removed from the Old World monopartite ones. The results of the N c -plot and base composition analysis indicated that the codon usage pattern of begomoviruses was influenced by mutation bias as well as other factors such as translational selection. Comparative analysis of AC1 and AC4 genes showed that the compositional environment of the former genes might play a role in dictating the codon usage of the latter gene. Thus, although it seems that strong translational selection might have an influence on shaping the codon usage of AC4 genes, the compositional constraints derived from AC1 genes might be the major determinant in determining codon usage. Consequently it can be speculated that mutation bias might play a major role in shaping the codon usage pattern of begomoviruses. As to the gene itself, the selection pressures from the external environment always act as effective factors in promoting the gene to adapt to the change in the external environment. On the other hand, direct changes to a gene will interfere with or be harmful to the gene. Therefore, the base mutation at the synonymous third codon position may not affect the protein expression for AC1 gene because of the degeneration of genetic codons. But for the embedded AC4 gene, the corresponding base mutation occurs at the second codon position, which probably results in the loss of function of the translated protein. Thus we suggested that AC4 gene might degenerate step by step during the long period of evolution, which might be an important reason for explaining the loss of function for the AC4 gene in the bipartite begomoviruses with DNA B. The variation analysis of the base composition for begomovirus genes showed that AV1 and BV1 genes exhibit a stronger codon usage bias and a higher gene expression level. Thus CAI values of different gene samples were calculated using the AV1 gene as a reference set. The results of correlation analysis indicated the reliability of choosing AV1 gene as a high expression gene sample for begomoviruses. Then 14 codons were determined as the major optimal codons for begomoviruses. That will be very important during the design of degenerate primers, the introduction of point mutation, the modification of the virus genes and the investigation of the evolution mechanism of species at the molecule level. It was speculated that monopartite begomoviruses emerged approximately 130 million years ago (Rybicki, 1994; Mansoor et al., 2006) , suggesting that the begomoviruses should have evolved from the original monopartite viruses. Rybicki (1994) suggested that the significant expansion of the New World begomoviruses might have occurred after the transmission of the Old World begomoviruses to those of the New World by whiteflies. Based on codon usage pattern analysis, it could be inferred that there was no direct relationship between the Old World monopartite begomoviruses and the New World bipartite ones. Interestingly, a small number of the Old World bipartite begomoviruses exhibited a similar codon usage pattern to the New World bipartite ones, which suggested that the ancestor of begomoviruses could have evolved from monopartite to bipartite ones before they were transferred to the New World areas. Subsequently it was not the Old World monopartite begomoviruses but the Old World bipartite ones that transmitted to the New World in a certain way and finally evolved into the New World bipartite begomoviruses. In other words, the New World bipartite begomoviruses probably resulted directly from the Old World bipartite ones, while the latter evolved from the Old World monopartite begomoviruses. It still remains unclear whether the New World begomoviruses directly evolved from the Old World bipartite begomoviruses because of the sharp environmental change after their transfer to the New World, or whether the Old World bipartite begomoviruses had evolved into the New World bipartite ones before their transmission to the New World. Rybicki (1994) suggested that the absence of the AV2 gene in all the New World begomoviruses could be attributed to its earlier loss after the New World begomoviruses arriving in the New World. These results suggest that the current New World begomoviruses evolved from ancient viruses after transferring to the New World. Ha et al.(2006) speculated that the new identified begomovirus termed Corchorus yellow vein virus (CoYVV) might belong to a New World begomovirus group that previously existed in the Old World, which suggested that the common ancestor of the New World begomovirus might originate from the Old World begomovirus. However, both the New World and Old World begomoviruses had began to evolve and coexisted in this area for a long time before the separation of the continents. In other words, the present New World begomoviruses might have evolved in the Old World, and then moved to the New World by some unknown means (Ha et al., 2006) .

Intrapulmonary administration of recombinant activated factor VII in diffuse alveolar haemorrhage: a report of two case stories

BACKGROUND: Diffuse alveolar haemorrhage (DAH) is a serious pulmonary complication characterised by a high mortality rate and the absence of specific treatment. The intrapulmonary administration of activated recombinant factor VII (rFVIIa) in DAH was recently published in six patients by Heslet et al with an efficient hemostatic effect. We describe two cases of DAH treated with intrapulmonary rFVIIa. METHODS: Two cases of DAH were admitted to the ICU after presenting abrupt desaturation, tachypnea, cough and haemoptysis, requiring orotracheal intubation and mechanical ventilation. The diagnosis was achieved by the bloody return during the bronchoalveolar lavage, during the procedure rFVIIa (50 μg/Kg in 50 ml of isotonic saline) was administered via the bronchoscope. RESULTS: Immediate cessation of bleeding was observed. Prior to intrapulmonary administration of rFVIIa, the FiO(2 )was 1, which was reduced to 0.4 24 hours later. Following the procedure, the haemostatic effect made blood transfusion superfluous. No thrombotic complications associated with administration of the drug were observed. After the intervention both cases progressed fast and was discharged from the ICU with no further episodes of bleeding. CONCLUSION: 1. Local intrabronchial deposition of DAH with rFVIIa has been shown to be effective in controlling life-threatening DAH. 2. In the case described above, no thrombotic complications were observed following the intrapulmonary administration of rFVIIa.

Conclusion: 1. Local intrabronchial deposition of DAH with rFVIIa has been shown to be effective in controlling life-threatening DAH. 2. In the case described above, no thrombotic complications were observed following the intrapulmonary administration of rFVIIa. Diffuse alveolar haemorrhage (DAH) is a serious pulmonary complication, characterised by the presence of haemoptysis, dyspnea, hypoxemia and anaemia, with a high mortality rate of over 50% of patients requiring mechanical ventilation [1] . Diffuse opacities found on X rays in patients with DAH, however, are unspecific [2] . DAH is a complication of systemic diseases, and frequently manifests as an initial sign of these [3] . Bronchoalveolar lavage (BAL) is the most useful procedure for confirming initial clinical suspicion, BAL with a bloody return is the only way to confirm the diagnosis and, at times, fiberoptic bronchoscopy provides the treatment. In a new development, its use in intrapulmonary administra-tion of recombinant activated factor VII (rFVIIa) has been reported in six patients [4] . The drug (rFVIIa) promotes local formation of thrombin when it combines with tissue factor exposed at the level of the endothelium. It was initially indicated to treat coagulopathies in patients suffering from haemophilia, although its use has extended to other haematological conditions. Beyond these indications, it is administrated on a more compassionate level. The requirements for its effective use in controlling serious haemorrhages are fibrinogen and platelet readings of over a 50 mg/dl and with 50.000/μl respectively, in addition to a pH > 7.20. We present two cases of massive haemoptysis in which local administration of rFVIIa (Novoseven ® from Novonordisk) via BAL was used as an emergency measure. Two cases of life-threatening DAH were admitted to the ICU after presenting abrupt desaturation, tachypnea, cough and haemoptysis, requiring orotracheal intubation and mechanical ventilation. Both cases were diagnosed with fiberoptic bronchoscopy and treated with local rFVIIa. A 39 year old woman, with a personal history of acute promyelocytic leukaemia treated with chemotherapy and renal failure with recent arteriovenous fistula intervention. The patient was admitted to the ICU with hypoxemic respiratory failure with cough and haemoptysis, requiring orotracheal intubation and mechanical ventilation. The thorax X-ray showed a "patchy" infiltrate affecting bases and middle fields. Values of haemoglobin with 11,4 decreased to 7,9 g/dl concomitant with reduced platelet count from 100 × 10[3]/μL to 40 × 10 [3] /μL during the first 24 hours. Suddenly, the patient deteriorated with an abrupt desaturation and frank haemoptysis through the orotracheal tube. In order to achieve platelet readings of over 50,000/μL a transfusion of 8 units of platelets was necessary before the drug could be administered. An emergency fibrobronchoscopy confirmed DAH. Systemic administration of rFVIIa was considered, but the potential thrombogenic effect of the drug and the risk of obstruction of the recent arteriovenous fistula prompted us to decide to administer 50 μg/Kg of rFVIIa in 50 ml of isotonic saline via the bronchoscopy channel, 25 ml in each main bronchus; following this, immediate cessation of bleeding was observed. Prior to this, the families had been informed and their written consent obtained. The inspired fraction of oxygen (FiO 2 ) which, before intrapulmonary administration of rFVIIa had been 1, was reduced to 0.6 during the first three hours subsequent to the administration of rFVIIa, and to 0.4 over the following 24 hours (figure 1). Haemoglobin readings remained unchanged without the need for blood transfusion. Although there were no new episodes of active DAH, weaning from ventilator was retarded due to muscular weakness. The patient was extubated at day 16 in the ICU and discharged to stationary ward without recurrent bleeding. Changes in the FiO 2 following rFVIIa administration in two cases treated with intrapulmonary rFVIIa Figure 1 Changes in the FiO 2 following rFVIIa administration in two cases treated with intrapulmonary rFVIIa. A 46 year old man, with a personal history of smoking, ex parenteral drug abuse, hepatitis B and C infection, hepatic cirrhosis evolving for years and HIV infection diagnosed in 1988; currently receiving antiretroviral treatment. He was admitted to the ICU with acute inferoposterior myocardial infarction. Treatment was commenced with low molecular weight heparin and double anti-aggregation therapy with aspirin and clopidogrel. After 24 hours, he suddenly developed haemoptysis, acute hypoxemic respiratory failure and bilateral crackles, requiring orotracheal intubation and mechanical ventilation with FiO 2 of 1 to maintain a SpO 2 of 85-90%. The BAL return was increasingly bloody from both lungs. Due to the potential thrombogenic effect of systemic rFVIIa administration in acute myocardial infarction, and after informing and obtaining consent from the family, we decided on intrapulmonary administration of the drug at a dosage of 50 μg/Kg in 50 ml of isotonic saline via the bronchoscopy channel, observing the immediate cessation of bleeding. The FiO 2 was reduced to 0.5 over the first three hours and to 0.35 after 24 hours (figure 1). After 12 days on mechanical ventilation the patient was extubated and transferred to a stationary ward. Systemic administration of rFVIIa to patients with lifethreatening conditions due to active haemorrhage has increased in clinical practise, based more on presumed expectation than on scientific evidence supported by controlled and randomised studies [5] . Edward et al carried out a survey in the American College of Chest Physicians (ACCP) in 1998 on the treatment of acute haemoptysis: 85% of specialists answered that intubation and connection to mechanical ventilation must be performed at an early stage and 64% considered it mandatory to carry out a fibrobronchoscopy during the first 24 hours [6] . Consequently, treatment undertaken in both clinical cases presented -mechanical ventilation support and fibrobronchoscopy -meets these recommendations and was performed as an emergency measure to treat a massive haemoptysis episode. In middle-sized hospital, where selective embolization of the bronchial artery techniques and/or thoracic surgery are not available, systemic rFVIIa has been used in massive haemoptysis cases, with good results [7] . DAH in haematological patients requiring mechanical ventilation has a high mortality rate of over 70% in series described [8, 9] . This is due to the absence of specific treatment for DAH of haematological origin. The results of a multicentre, randomised study on the efficacy and safety of three different dosages of systemic rFVIIa compared with a placebo in treating haemorrhagic complications in 100 bone-marrow transplant patients (seven with DAH) were inconclusive, 8% of thromboembolic events were observed in the group treated with rFVIIa [10] . Pulmonary haemorrhage associated with myocardial infarction thrombolysis is an unusual complication; in 1996, Chang YC et al published a retrospective study, finding an incidence of 0.4% [11] . The same is true of platelet antiaggregation treatment, where isolated cases of DAH have been described [12] . In spite of its low incidence, DAH secondary to acute coronary syndrome treatment is a complication which maybe undetected, due to the common radiological findings in both DAH and ALI/ARDS and acute pulmonary oedema [13] . Bearing in mind the thrombogenic risk of a systemic administration of local rFVIIa, and with Heslet et al's publication as a reference, in which 6 consecutive critically ill patients with acute DAH are treated with local rFVIIa, intrapulmonary administration of the drug was chosen in the cases here presented with a view to avoiding the risk of thrombosis of the arteriovenous fistula in the first case, and of reinfarction in the second [4, 10] . DAH is a life-threatening disease characterised by the lack of specific treatment and a high mortality of patients requiring mechanical ventilation. Bronchoscopy BAL with increasingly bloody return is the only diagnostic procedure for the diagnosis of DAH, and at times, provides the treatment. Local administration of rFVIIa via the fibrobronchoscope channel was used as an emergency measure in two cases of massive haemoptysis with an excellent hemostatic effect and without adverse effects. ALI: Acute lung injury; ARDS: Acute respiratory distress syndrome; BAL: Bronchoalveolar lavage; DAH: Diffuse alveolar haemorrhage; FiO 2 : Inspired fraction of oxygen; rFVIIa: Recombinant activated factor VII; HIV: Human inmunodeficiency virus; SpO 2 : Pulse oxygen saturation.

Activation of the Unfolded Protein Response Is Required for Defenses against Bacterial Pore-Forming Toxin In Vivo

Pore-forming toxins (PFTs) constitute the single largest class of proteinaceous bacterial virulence factors and are made by many of the most important bacterial pathogens. Host responses to these toxins are complex and poorly understood. We find that the endoplasmic reticulum unfolded protein response (UPR) is activated upon exposure to PFTs both in Caenorhabditis elegans and in mammalian cells. Activation of the UPR is protective in vivo against PFTs since animals that lack either the ire-1-xbp-1 or the atf-6 arms of the UPR are more sensitive to PFT than wild-type animals. The UPR acts directly in the cells targeted by the PFT. Loss of the UPR leads to a normal response against unrelated toxins or a pathogenic bacterium, indicating its PFT-protective role is specific. The p38 mitogen-activated protein (MAPK) kinase pathway has been previously shown to be important for cellular defenses against PFTs. We find here that the UPR is one of the key downstream targets of the p38 MAPK pathway in response to PFT since loss of a functional p38 MAPK pathway leads to a failure of PFT to properly activate the ire-1-xbp-1 arm of the UPR. The UPR-mediated activation and response to PFTs is distinct from the canonical UPR-mediated response to unfolded proteins both in terms of its activation and functional sensitivities. These data demonstrate that the UPR, a fundamental intracellular pathway, can operate in intrinsic cellular defenses against bacterial attack.

Pore-forming toxins (PFTs) are the single most prevalent protein virulence factor made by disease-causing bacteria and are important for the virulence of many important human pathogens including Staphylococcus aureus, Streptococcus pyogenes, Clostridium perfringens, and Aeromonas hydrophilia [1, 2] . Crystal (Cry) toxins produced by the invertebrate pathogen Bacillus thuringiensis (Bt) are a large family of PFTs that target the intestinal cells of insects and nematodes [3, 4, 5] . The fact that some Cry proteins target nematodes, in particular C. elegans, has been exploited to provide the only in vivo genetic model for studying PFTs. This system led to the discovery of the first signal transduction pathway that protects cells against PFTs, the p38 mitogen-activated protein kinase (MAPK) pathway, which has been confirmed in mammalian cells [6, 7] . There is growing evidence that the response of cells to PFTs is, however, complex and there is a great deal yet to learn [8] . The unfolded protein response (UPR) of the endoplasmic reticulum (ER) is a fundamental stress response used by eukaryotic cells to match protein synthesis demand to its capability to fold proteins within the ER to maintain cellular homeostasis [9] . In C. elegans and other animals there are three transducers that signal from the ER to activate this response. These three distinct arms of the UPR are mediated by IREI, ATF6, and PERK in mammals [10] , which correspond to the genes ire-1, atf-6, and pek-1 in C. elegans [11, 12, 13] . All three pathways are regulated by the ER chaperone BiP in response to an increase in unfolded proteins [9] . Here we demonstrate that the ER stress response, in particular the ire-1 arm, is activated upon exposure of C. elegans and mammalian cells to PFTs. We demonstrate for the first time that the ire-1 -xbp-1 arm of the UPR (and to a lesser extent the atf-6 arm) is functionally important for defense against a pathogenic attack since loss of this pathway leads to animals hypersensitive to PFT, but not to other toxic insults. Furthermore, we demonstrate that activation of the ire-1-xbp-1 pathway by PFT requires p38 MAPK and its associated MAPK kinase and that the in vivo response of the UPR to a PFT can be separated from its response to unfolded proteins. These results indicate that activation of the UPR plays an important role in cellular defenses against pathogens. In a genetic screen for genes involved in the cellular response of C. elegans to the PFT Cry5B, we found a mutant predicted to be defective in protein N-glycosylation in the ER (L.J.B. and R.V.A., manuscript in preparation). Since defects in protein glycosylation induce the UPR, this result suggested that perhaps the UPR might play a role in protection against PFTs. To test this hypothesis, we first investigated whether or not the UPR was activated by a PFT. The xbp-1 gene is spliced upon activation of the IRE-1 branch of the UPR, and its splicing is one marker for IRE-1 (and UPR) activation [13] . In C. elegans, the xbp-1 intron spliced by IRE-1 is 23 nucleotides and the induction of this splicing event can be detected by RT-PCR [14] . To analyze xbp-1 mRNA transcript splicing, animals were fed Escherichia coli expressing Cry5B and compared to worms fed control E. coli ( Figure 1A ). While there is abundant unspliced xbp-1 mRNA transcript in both samples, there is an increase in the spliced xbp-1 transcript from worms ingesting Cry5B, indicating activation of the IRE-1 pathway. Quantitative analyses indicate that the xbp-1 spliced transcript increases 2.3, 3.0, and 3.0 fold at the 7, 8, and 9 h time points respectively. To independently test this result, we analyzed the in vivo expression of an ire-1 regulated gene, hsp-4, a BiP homolog. In vivo analysis of the hsp-4 promoter coupled to green fluorescent protein (GFP) demonstrated expression of this gene requires ire-1 and xbp-1 [13] . A C. elegans strain containing hsp-4::GFP was fed either control E. coli or Cry5B expressing E. coli for 8 hours at 20uC. As shown, a strong and specific increase in GFP expression in the intestine can be seen in the presence of the PFT ( Figure 1B , middle panel), consistent with activation of the ire-1-xbp-1 pathway by Cry5B. Heat shock of this strain in the absence of Cry5B confirms GFP could be induced in other cell types in addition to the intestine ( Figure 1B, right panel) , as was demonstrated with the N-glycosylation inhibitor tunicamycin [13] . The fact that Cry5B only induced expression in intestinal cells suggests the PFT is only targeting these cells (see below). To address whether the ire-1-xbp-1 pathway is also activated in mammalian cells in response to a PFT, activation of the pathway was ascertained in HeLa cells exposed to the Aeromonas hydrophila PFT, aerolysin. As detected by the presence of the spliced protein isoform of XBP-1, treatment of mammalian cells with a PFT also results in robust activation of the ire-1-xbp-1 pathway ( Figure 1C ). The ER stress response is required for defense of C. elegans against Cry5B To determine whether the ER stress response played a role in the defense of C. elegans against the PFT, C. elegans mutants in the ER stress response pathway were qualitatively compared to wildtype N2 animals in their susceptibilities to Cry5B. The mutants that were tested included those encoding the three ER stress transducer genes, atf-6(ok551), pek-1(ok275), and ire-1(v33), as well as xbp-1(zc12); these mutations are predicted or known to be loss of function mutations in their respective genes [11, 12, 13] . In the absence of Cry5B, the wild type and mutant worms are healthy adults with similar appearance, except ire-1(v33), which is clearly smaller than the other strains ( Figure 2A ). In the presence of lowmoderate levels of the PFT Cry5B, wild-type worms are slightly intoxicated compared to those found on control no-toxin plates, as evidenced by their smaller sizes and paler appearances (Figure 2A ). To the same extent seen with wild-type worms, atf-6(ok551) and pek-1(ok275) mutant animals are also slightly intoxicated on lowmoderate levels of the PFT Cry5B, indicating lack of either of these genes does not result in overt hypersensitivity or hyperresistance to Cry5B (Figure 2A) . However under the same conditions, the ire-1(v33) and xbp-1(zc12) mutant worms are more severely intoxicated than wild-type worms as they are relatively smaller and considerably paler compared to their corresponding no toxin controls. The hypersensitivity to Cry5B resulting from lack of ire-1 and xbp-1 was also seen using RNA interference (RNAi; data not shown), confirming the phenotype is caused by loss of function in these genes. We call this hypersensitivity phenotype ''Hpo'' for hypersensitive to pore-forming toxin. The sensitivity to Cry5B of animals mutant for the three ER stress response pathways was quantitatively assessed using a dosedependent lethality assay ( Figure 2B ). From these data, an LC 5s s (lethal concentrations at which 50% of the animals die) were obtained (Table 1) . Our quantitative results confirm that ire-1(v33) and xbp-1(zc12) mutant animals are statistically more sensitive to PFT than wild type animals (Table 1 ) and thus are Hpo relative to wild type (caution is called for in interpreting the ire-1(v33) data since many of these animals also have significant overt defects, e.g., developmental delays which prevents them from being as well synchronized at the start of the assay compared to the other strains [11] ). Our results indicate that atf-6(ok551) mutant animals are also Hpo, albeit to a lesser extent (2.8 vs. 5.8 fold increase in sensitivity for atf-6 vs. xbp-1). Although atf-6(ok551) hypersensitivity was not discerned with the plate assay, it is likely that the quantitative lethality assay is a more sensitive test for Cry5B hypersensitivity than the qualitative plate assay. In contrast to xbp-1 and atf-6 mutant animals, the sensitivity of pek-1(ok275) mutant animals is not statistically different from that of wild-type animals ( Table 1) . To independently confirm these results, we used a developmental assay to assess the relative sensitivity of the four ER stress response mutants to Cry5B. This experiment was performed by placing newly hatched L1 stage worms on plates containing different percentages of Cry5B expressing E. coli and then counting the worms that developed to either the L4 stage or adulthood ( Figure 2C ). In the absence of Cry5B, nearly all worms developed to the L4 stage or adulthood for all strains with the exception of ire-1(v33). This result confirms developmental defects previously seen with this mutant [11] , and it was therefore excluded from subsequent analyses. Wild type N2 and pek-1(ok275) were both similarly inhibited in their development by increasing percentages of Cry5B. Compared to N2 and pek-1(ok275) animals, though, both atf-6(ok551) and xbp-1(zc12) were Hpo, i.e., each is more developmentally inhibited by Cry5B than wild-type animals ( Figure 2C ). Because the ire-1-xbp-1 pathway has a more discernible effect on protection against Cry5B than atf-6, further experiments were focused on this arm of the ER stress response. Pore-forming toxins (PFTs) are bacterial toxins that form holes at the plasma membrane of cells and play an important role in the pathogenesis of many important human pathogens. Although PFTs comprise an important and the single largest class of bacterial protein virulence factors, how cells respond to these toxins has been understudied. We describe here the surprising discovery that a fundamental pathway of eukaryotic cell biology, the endoplasmic reticulum unfolded protein response (UPR), is activated by pore-forming toxins in Caenorhabditis elegans and mammalian cells. We find that this activation is functionally important since loss of either of two of the three arms of UPR leads to hypersensitivity of the nematode to attack by PFTs. The response of the UPR to PFTs can be separated from its response to unfolded proteins both at the level of activation and functional relevance. The response of the UPR to PFTs is dependent on a central pathway of cellular immunity, the p38 MAPK pathway. Our data show that the response of cells to bacterial attack can reveal unanticipated uses and connections between fundamental cell biological pathways. Taken together, the above results suggest that the ire-1-xbp-1 pathway functions to protect the host against the PFT Cry5B. However, an alternative explanation for our results is that animals mutant in this pathway (e.g., xbp-1 mutant animals) are sickly and have compromised health and therefore would respond poorly to any toxic insult. To address this alternative hypothesis, we tested whether xbp-1(zc12) animals are hypersensitive to two toxic chemical compounds, the heavy metal CuSO 4 (a toxic insult that kills with kinetics similar to Cry5B) and the oxidative stress agent H 2 O 2 (a toxic insult that kills rapidly). The mutant xbp-1(zc12) has the same sensitivity as wild type to killing by either CuSO4 or H 2 O 2 ( Figure 2D and 2E; Table 1 ). These data argue against the supposition that this mutant is hypersensitive to the PFT merely because it is generally unhealthy. Rather, the protective response is somewhat specific against the PFT. These conclusions are strengthened by the finding that C. elegans lacking the UPR respond normally to attack by the pathogenic bacteria Pseudomonas aeruginosa, which does not make a PFT ( Figure 2F and Table 1 ). The xbp-1 pathway functions in the intestine to protect against Cry5B PFT Mosaic and expression analyses have shown that the targeting of intestinal cells by the PFT Cry5B is both necessary and sufficient to intoxicate worms [15, 16] . If the ire-1-xbp-1 pathway is functioning directly to protect against the effects of the PFT, then we would predict that the ire-1-xbp-1 pathway should function in the target cells of the toxin, the intestinal epithelial cells. Alternatively, the pathway might be functioning indirectly to protect against the effects of the PFT (e.g., it might hypothetically function in neurons that then sends protective signals to the intestine). Consistent with the first hypothesis, that the pathway is functioning directly in the target cells to protect against the PFT, we previously noted that a marker for downstream activation of the pathway, hsp-4, is turned on exclusively in intestinal cells ( Figure 1B , middle panel), although the pathway is capable of being activated throughout the worm by a more general stress, such as heat shock ( Figure 1B, right panel) . elegans fed E. coli expressing Cry5B compared to control E. coli not expressing Cry5B. The time the worms were allowed to feed on the E. coli before total RNA was prepared for RT-PCR is indicated at the top, and the positions of the nucleotide size markers are indicated at the left. (B) Compared to worms fed control non-Cry5B expressing E. coli, in vivo activation of hsp-4::GFP occurs specifically in the intestines of worms fed Cry5B expressing E. coli at 20uC for 8 hours. As a comparison for GFP induction, separate worms on control bacteria were heat shocked at 30uC for 8 hours to induce the ER stress response by causing unfolded proteins. The heat shock worms have a strong increase in GFP throughout the body including the head, intestine and hypodermis. Thus, although the entire worm is capable of activating the ire-1-xbp-1 pathway as judged by hsp-4 induction, activation in Cry5B-fed animals is occurring only in those cells targeted by the PFT. Images taken by light microscopy are compared to images with fluorescence microscopy. Scale bar is 0.2 mm. The experiment was performed three times, and representative worms are shown. (C) Aerolysin induces activation of IRE1 in mammalians cells. Exposure of HeLa cells to proaerolysin (2 ng/mL) leads to increased production of spliced XBP1 protein as shown on this immunoblot (upper) and quantitated relative to no toxin control (lower). DTT (10 mg/mL for 2 h) was used as a positive control. Positions of molecular weight markers (kDa) are indicated on right side of the figure. A nonspecific antibody-reacting band was used as a loading control and normalization of the XBP1 signal in each lane. doi:10.1371/journal.ppat.1000176.g001 Figure 2 . Loss of specific UPR pathways cause hypersensitivity to PFT but not other toxins or a pathogenic bacteria. (A) Comparison of ER stress response mutants to wild-type N2 on 25% Cry5B-expressing E. coli plates indicate ire-1(v33) and xbp-1(zc12) are hypersensitive to Cry5B intoxication. Two representative worms are shown for each strain 48 hours after feeding either on E. coli without Cry5B or on E. coli of which 25% expressed Cry5B. Scale bar is 0.2 mm. (B) A lethal concentration assay was performed using purified Cry5B toxin to quantitatively compare sensitivities of wild-type N2 and the ER stress mutants. Lethality was determined after 8 days. This semi-log graph represents three independent experiments, and each data point is the mean and standard deviations of the experiments. (C) A Cry5B developmental inhibition assay was performed beginning with synchronized worms at the first larval stage. Worms were grown on plates containing different percentages of Cry5B-expressing E. coli (% Cry5B as indicated under the figure), and the percent of worms reaching the L4 stage or adulthood 72 hours later is indicated. ire-1(v33) was included only on the plates with 0% Cry5B. Data are presented as mean and standard deviation. (D) A lethal concentration assay comparing sensitivity to CuSO 4 revealed xbp-1(zc12) is not hypersensitive compared to wild-type N2. Lethality was determined after 8 days of CuSO 4 exposure, the same time frame as the Cry5B lethality assay. Data, plotted semi-log, are the mean and standard deviation of three independent experiments. (E) A lethal concentration assay comparing sensitivity to H 2 O 2 revealed xbp-1(zc12) is not hypersensitive compared to wild-type N2. Lethality was determined after 4 hours of H 2 O 2 exposure. Data, plotted semi-log, are the mean and standard deviation of three independent experiments. (F) A lifespan assay was used to compare the ER stress mutants to slow killing by P. aeruginosa PA14. This graph represents combined data from three experiments. doi:10.1371/journal.ppat.1000176.g002 To directly demonstrate the role of xbp-1 in protecting intestinal cells against Cry5B, the intestinal specific app-1 promoter [17] was used to drive expression of xbp-1 in xbp-1(zc12) mutant animals to determine if expression in the intestine is sufficient to rescue the Hpo phenotype. As a negative control, GFP was similarly expressed under control of the app-1 promoter in xbp-1(zc12) mutant animals. In control animals, expression of the GFP solely in intestinal cells was confirmed (data not shown). As expected, the majority of wild-type N2 animals showed only a low-modest degree of intoxication upon exposure to 25% Cry5B-expressing E. coli ( Figure 3A , B; they were smaller and somewhat paler than the wild-type worms on control plates but were still quite active). Also as predicted, both xbp-1(zc12) mutant animals and xbp-1(zc12) mutant animals transformed with app-1::GFP were Hpo and intoxicated to similar extents ( Figure 3A , B; most animals were very pale, small, and inactive). In contrast, xbp-1(zc12) worms expressing xbp-1 under the app-1 promoter were significantly healthier than either untransformed or app-1::GFP transformed xbp-1(zc12) animals fed with Cry5B ( Figure 3A, B) . However, these app-1::xbp-1-transformed xbp-1(z12) worms were not as healthy as wild-type N2 under the same conditions. This partial rescue could indicate the expression of the artificial xbp-1 transgenes did not fully recapitulate wild-type xbp-1 expression levels and/or that there is some role for the ire-1 -xbp-1 pathway in other cell types. Nonetheless, our results support a significant protective function for xbp-1 within the cells targeted by Cry5B. Induction of ire-1-xbp-1 pathway's role in response to PFT but not unfolded proteins is regulated by the p38 MAPK pathway ER stress responses have been studied extensively for their role in protecting cells against unfolded proteins [10, 18] . One way to assess the role of the ER stress pathways in protecting against unfolded proteins is with the drug tunicamycin (a natural compound that leads to the accumulation of unfolded proteins in the ER due to its inhibitory effect on N-linked protein glycosylation [19] ). Previous data in C. elegans have indicated different sensitivities of the three ER stress response pathways for tunicamycin [11, 12] . Using a different toxicity assay, we have confirmed these observations: atf-6(ok551) mutant animals have a similar sensitivity to tunicamycin as wild-type animals whereas both xbp-1(zc12) and pek-1(ok275) mutant animals are more readily killed by tunicamycin (Figure 4) . These results are in contrast to the response of these different ER stress pathways to Cry5B, to which atf-6 mutant animals are more sensitive than pek-1 mutant animals. These data suggest that there are differences in how ER stress pathways are activated in response to unfolded proteins and to the PFT Cry5B. It is known that PFTs trigger the activation of p38 MAPK, which promotes cell survival and cellular defenses and which seems to play a central role in cellular responses to PFTs [6, 7, 20] . We therefore investigated whether PFT-mediated activation of the UPR and the p38 MAPK pathway might be connected. We first investigated whether the ire-1-xbp-1 pathway plays a role in the PFT-induced activation of p38 by comparing the activation of the p38 MAPK in wild-type and xbp1(zc12) animals. We find that addition of Cry5B to wild-type C. elegans results in an increase in phosphorylated p38, indicating the p38 pathway is activated by a PFT in C. elegans just as it is in mammalian cells [20] ( Figure 5A ). We find that p38 activation occurs normally in xbp-1(zc12) mutant animals ( Figure 5A ), indicating that the UPR is not required for activation of p38 MAPK pathway in response to PFT. We extended this result using ttm-2, a downstream transcriptional target of the p38 MAPK pathway in response to Cry5B and a gene required for normal defense against Cry5B PFT [6] . Upregulation of ttm-2 mRNA was dependent on the p38 MAPK pathway but not dependent on xbp-1 ( Figure 5F ). We next analyzed the reverse relationship between the ire-1-xbp-1 and the p38 MAPK pathways, namely whether the p38 MAPK pathway is required for PFT-induced activation of the ire-1-xbp-1 pathway. We find that activation of the ire-1-xbp-1 pathway in response to PFT is dependent on the p38 MAPK pathway, namely on sek-1, the MAPK kinase (MAPKK) gene upstream of p38, and on pmk-1, the p38 MAPK downstream of sek-1 ( Figure 5 ). We find that increased splicing (activation) of xbp-1 in response to Cry5B does not occur in sek-1(km4) MAPKK mutant animals ( Figure 5B) . Quantitatively, at the 3 h time point the spliced form of xbp-1 is induced 1.9 fold in animals with an intact p38 MAPK pathway and depressed 1.8 fold in sek-1(km4) MAPKK mutant animals relative to untreated controls. However, sek-1 is not absolutely required for splicing of xbp-1 since, in response to tunicamycin, splicing of xbp-1 is normal in sek-1(km4) mutant animals ( Figure 5C ). In agreement with these results, we find that in vivo activation of the downstream target of the ire-1-xbp-1 pathway, hsp-4::GFP, by Cry5B within intestinal cells does not occur in pmk-1(km25) p38 MAPK mutant animals ( Figure 5D ), whereas activation of hsp-4::GFP by tunicamycin does occur normally in pmk-1(km25) mutant animals ( Figure 5E ). To independently confirm and extend these results, we analyzed a different downstream target of the ire-1-xbp-1 pathway. Using proteomics, we identified a protein, Y41C4A.11 (a homolog of the beta-prime subunit of the coatomer complex), that increased 4.6 fold in C. elegans animals exposed to Cry5B and whose increase was completely dependent on xbp-1 (see Materials and Methods and Protocol S1). The gene encoding this protein was previously demonstrated to be transcriptionally regulated by tunicamycin in an ire-1 and xbp-1 dependent manner [12] . Using real time PCR, we find that both hsp-4 mRNA and Y41C4A.11 mRNA are induced by either Cry5B or tunicamycin ( Figure 5F ). Consistent with activation of the ire-1-xbp-1 pathway by p38 MAPK in response to PFT but not unfolded proteins, the full induction of both mRNAs by Cry5B, but not tunicamycin, is dependent on sek-1 MAPKK. Interestingly, whereas induction of both mRNAs by Cry5B is lacking in xbp-1(zc12) mutant animals (confirming that activation of hsp-4 and Y41C4A.11 by PFT is via the UPR), both mRNAs are still somewhat induced by Cry5B in a sek-1(km4) mutant, albeit at lower levels than in wild-type animals. These data suggest that some of the UPR-mediated transcriptional response is p38 pathway independent. Based on these data, we predicted that animals mutant in the p38 pathway should be more sensitive to PFT than animals mutant in the UPR pathway. This hypothesis is based on the fact that the p38 pathway is upstream of the UPR, is required for full activation of the UPR in response to PFT, and is involved in UPRindependent PFT defense pathways (e.g., ttm-2). Comparison of sek-1(km4) and xbp-1(zc12) mutant animals on Cry5B indicates sek-1(km4) animals are more severely intoxicated than xbp-1(zc12) animals at the same dose of Cry5B ( Figure 5G ). This conclusion was quantitatively confirmed by performing LC 50 experiments on N2 and sek-1(km4) animals (Table 1) . Whereas the LC 50 of xbp-1(zc12) animals on Cry5B is 5.8 fold lower than N2, the LC 50 of sek-1(km4) animals on Cry5B is 170 fold lower than N2. Here we demonstrate that ER stress response pathways play a central but heretofore unknown role in innate defenses in vivo. Specifically, we find that bacterial pore-forming toxins (PFTs) activate the ire-1-xbp-1 branch of the ER Unfolded Protein Response (UPR) in C. elegans and mammalian cells and that the ire-1-xbp-1 and atf-6, but not the pek-1, branches of the UPR are important for C. elegans cellular defenses against a PFT since elimination of either of these two branches leads to hypersensitivity to the PFT Cry5B. The ER stress response has been previously associated with pathogenic attack, mostly in the opposite direction shown here, e.g., aiding viral replication and pathogenesis ( [21] and references therein). In a few cases, the ER stress response has been linked with innate immunity since induction of ER stress can activate CREB-H, which in turn promotes the acute inflammatory response [22] . It has also been suggested that IRE-1 could influence immunity via its association with TRAF-2, which in turn can regulate NF-kB [23] . Data from studies in plants suggest that in response to pathogens, signals can be produced that lead to an ''anticipatory'' UPR to handle the massive synthesis of new secretory proteins required [24] . Here we definitively demonstrate a functional role of the UPR in defense against a pathogen in vivo. Loss of xbp-1 leads to animals nearly 6 fold more susceptible to PFT whereas loss of atf-6 leads to animals nearly 3 fold more susceptible. Our data suggest that cells have adapted the UPR pathway for a specific response to PFTs in order to promote cellular defense against this common form of pathogenic attack. First, we found that loss of the xbp-1 arm of the UPR does not lead to hypersensitivity to a heavy metal or hydrogen peroxide nor does loss of either xbp-1 or atf-6 lead to decreased protection against a bacterial pathogen that lacks a PFT. Second, the ire-1-xbp-1 and atf-6 arms of the UPR are involved in the defense but the pek-1 arm is not. Third, the activation and function of the UPR in PFT defenses can be separated from the role of the UPR in dealing with unfolded proteins (here tested using the drug tunicamycin) in two ways: 1) the relative importance of the various arms of the UPR for defense against PFT is different than their importance for protection against unfolded proteins and 2) the activation of the ire-1-xbp-1 pathway by PFT, but not unfolded proteins, requires p38 MAPK (see below). A link between the p38 and UPR pathways has been shown in previous studies, although not with the level of functional relevance demonstrated here. Various arms of the UPR have been shown as both upstream or downstream of the p38 pathway, depending on the circumstances [25, 26, 27, 28, 29, 30] . The p38 pathway itself is implicated extensively in innate immune protection of many organisms against pathogens [31] and against PFTs in worms and mammals [6, 7] . Our data presented here for the first time functionally link the UPR to this major innate immune signal transduction pathway. Our findings on the activation and role of the UPR and p38 pathways in defense against PFT are summarized in Figure 6 . Why would induction of the ER stress response play a protective role against PFTs? It is possible that PFTs somehow lead to the accumulation of unfolded proteins in a cell. For example, PFTs are known to perturb calcium homeostasis and changes in calcium homeostasis are known to affect protein folding [32, 33] . In this model, cells would respond to the toxin via p38 MAPK and turn on the UPR to anticipate and ameliorate the detrimental effects of unfolded proteins. Arguing against this model, however, is our data showing that sensitivity of the three arms of the UPR to Cry5B is different than their sensitivity to a global unfolder of ER proteins, tunicamycin. A second model is that activation of the ER stress response by Cry5B in a p38 MAPK dependent manner may prepare the cell to handle an altered biosynthetic load in the ER to defend against a toxin. For example, transcriptional array analysis indicate that over 1000 genes are differentially regulated in C. elegans by Cry5B ingestion [6] , which could in turn lead to significant changes in the protein load of the ER. A third model is 4(bn2) and glp-4(bn2);sek-1(km4) after 3 hours of exposure to either control (DMSO) or tunicamycin (2 mg/mL). This is a representative experiment of three independent experiments. (D) In vivo induction of hsp-4::GFP by Cry5B requires pmk-1 (p38 MAPK). The strains hsp-4::GFP and hsp-4::GFP;pmk-1(km25) were fed either control E. coli or E. coli expressing Cry5B for 8 hours and the expression of GFP was then analyzed. Cry5B induces GFP within the intestinal cells of the strain hsp-4::GFP but not in the strain containing the pmk-1(km25) mutant. The experiment was performed three times and representative worms are shown. Scale bar is 0.2 mm. (E) In vivo induction of hsp-4::GFP by tunicamycin does not require pmk-1 (p38 MAPK). The strains hsp-4::GFP and hsp-4::GFP;pmk-1(km25) were exposed to either control (DMSO) or tunicamycin (2 mg/mL) for 8 hours and the expression of GFP was then analyzed. Tunicamycin induces GFP throughout both the strains hsp-4::GFP and hsp-4::GFP;pmk-1(km25), including within the intestinal cells. The experiment was performed three times and representative worms are shown. Scale bar is 0.2 mm. (F) Downstream targets of the UPR require the p38 MAPK pathway for induction by PFT but not unfolded proteins. The fold change in the levels of hsp-4 and Y41C4A.11 mRNA transcripts by Cry5B and tunicamycin were determined for glp-4(bn2), glp-4(bn2);xbp-1(zc12) and glp-4(bn2);sek-1(km4) using real-time PCR. In addition, the fold change in ttm-2 transcripts was determined in response to Cry5B. Data are mean and standard deviation of three independent experiments. (G) Animals lacking sek-1 MAPKK are more sensitive to Cry5B than animals lacking xbp-1. Wild-type N2, sek-1(km4), and xbp-1(zc12) animals were placed on plates spread with E. coli transformed with empty vector (0%) or spread with empty vector E. coli diluted 9:1 (10%) or 3:1 (25%) with Cry5B-expressing E. coli (% thus gives toxin dose on a plate relative to undiluted Cry5B-expressing E. coli). The assay was initiated with L4 stage worms and photographs were taken 48 hours later. In the absence of Cry5B, the worms developed into dark, gravid, active, healthy adults. On 10% Cry5B-expressing E. coli, xbp-1(zc12) were slightly smaller than N2 but healthier than sek-1(km4), which were as small, pale, inactive, and severely intoxicated. On 25% Cry5B-expressing E. coli, xbp-1(zc12) was more intoxicated than N2 but not as intoxicated as sek-1(km4) animals. Scale bar is 0.2 mm. doi:10.1371/journal.ppat.1000176.g005 based on the fact that activation of the ire-1-xbp-1 pathway leads to increased phospholipid biogenesis [34] . It is possible that the defensive role of the ire-1-xbp-1 pathway is to produce phospholipids that play a protective role against PFTs. Consistent with this, it has been shown that inhibiting the activation of SREBPs, the central regulators of membrane biogenesis, leads to hypersensitivity of mammalian cells to the PFT aerolysin [35] . In summary, we have identified specifically the ire-1-xbp-1 and atf-6 ER stress transducer pathways as components of cellular defenses against a PFT. While p38 MAPK was previously demonstrated to function in this regard [6] , we have discovered a major and unexpected downstream target of this pathway for PFT defenses, namely the UPR. These results demonstrate the fundamental requirement for specific cell responses to bacterial PFTs and support the notion of intrinsic cellular defenses (or INCED, formerly, cellular non-immune defenses), a budding concept in immunity that emphasizes the intrinsic ability of epithelial cells to defend against bacterial toxins and the importance of these defenses as a supplement to the innate immune and adaptive immune systems [36] . Additionally, the differential importance of the three ER stress transducer pathways in response to Cry5B versus tunicamycin, the differential activation of ire-1-xbp-1 by p38 MAPK in response to Cry5B versus tunicamycin, and the divergent pathways regulated by p38 MAPK in protective responses reveal how studying pathogenesis can uncover a wonderful complexity and new connections among intracellular pathways. C. elegans strains were maintained at 20uC on NG plates using Escherichia coli strain OP50 as the food source [37] . Strains used in this study were wild-type Bristol strain N2 [37] , atf-6(ok551), glp-4(bn2), ire-1(v33), pek-1(ok275), pmk-1(km25), sek-1(km4), SJ4005 (zcIs4 [hsp-4::GFP]) and xbp-1(zc12). atf-6(ok551) and pek-1(ok275) were each outcrossed a total of 6 times. SJ4005 was outcrossed an additional 4 times as it had been outcrossed twice upon receipt from the Caenorhabditis Genetics Center. xbp-1(zc12) was created by outcrossing strain SJ17 (xbp-1(zc12); zcIs4 [hsp-4::GFP]) four times and removing the integrated hsp-4::GFP during the outcrosses. Images were acquired with an Olympus BX60 microscope with the 106 objective linked to a 0.56 camera mount and a DVC camera. Worms were placed on 2% agarose pads containing 0.1% sodium azide for photography. All assays were performed at 20uC unless indicated elsewhere. Qualitative toxicity assays based on visual comparison of worm intoxication were performed on plates with E. coli-expressed Cry5B as described [6, 38] . Beginning with the 4 th larval (L4) stage worms, worms were fed for 48 hours either on control plates with E. coli JM103 that did not express Cry5B (empty vector) or plates prepared with E. coli JM103 expressing Cry5B diluted 1:3 with empty vector transformed JM103. This amount of Cry5B (25%) mildly intoxicates wild-type C. elegans, which allows for identification of strains that are hypersensitive to Cry5B as these strains will be more severely intoxicated than wild type. Quantitative lethal concentration assays were performed as described [38] except the worms were scored after 8 days for Cry5B, CuSO 4 , and tunicamycin. Lethal concentration assays with H 2 O 2 did not include E. coli or 5-fluoro-29-deoxy-uridine, and worms were scored after 4 hours. Concentrations of each toxin were set-up in triplicate for each assay, and each assay was performed independently three times. Purified Cry5B was prepared as described [39] and dissolved in 20 mM HEPES (pH 8.0) prior to use. Approximately 1500 worms were scored for each strain in the calculation of the LC 50 values for each toxin. For tunicamycin assays, the set up was identical to the lethality assay with Cry5B. For the developmental inhibition assay, Cry5B plates were prepared as described [6, 38] . Approximately 100 L1 stage worms (from bleached embryos hatched off overnight) were placed on each plate (60 mm) and the number of worms at the L4 or adult stage 3 days later was determined. This assay was performed independently three times. The P. aeruginosa lifespan assay was performed on slow-killing plates as described [40] , with the following modifications: PA14 was cultured overnight in tryptic soy broth instead of King's broth and then spread on slow-killing plates complemented with 75 uM mM 5-fluoro-29-deoxy-uridine. The experiment was performed three times with approximately 100-150 worms total per strain, at 20uC. To determine if there was rescue of the hypersensitivity phenotype in the intestinal-specific promoter studies, 25% E. coli-expressing Cry5B plates were used to compare Cry5B sensitivities of wild-type N2, xbp-1(zc12), and xbp-1(zc12) that were transformed with constructs to express either green fluorescent protein (GFP) or xbp-1 mRNA within intestinal cells using the app-1 promoter (plasmids are described in Protocol S1). Transgenic L4 stage worms were placed on the 25% E. coli expressing Cry5B plates and their health status was assessed 72 hours later. Specifically, the relative health of each worm was determined qualitatively by comparing body size, darkness of the intestine as an indicator of feeding, and activity, including whether the worm demonstrated spontaneous movement. For scoring of the transgenic worms, comparisons were made using both N2 as a reference for healthy worms, as they demonstrated dark intestines and continuous spontaneous movement, and xbp-1 (zc12) as a reference for intoxicated worms that had pale intestines and demonstrated rare or no spontaneous movement. The glp-4(bn2) strain was used for these experiments (including the double mutants glp-4(bn2);xbp-1(zc12) an glp-4(bn2); sek-1(km4)) since it has a greatly reduced number of germ cells when grown at Figure 6 . Schematic illustrating relationship between p38 MAPK, ire-1-xbp-1, and PFT defense pathways. PFTs at the cell surface of epithelial cells activate p38 MAPK that activates IRE-1 that induces splicing of xbp-1, which then turns on defense against PFTs. Residual activation of xbp-1 targets in the absence of the p38 MAPK pathway suggests there might be p38-independent activation of the ire-1-xbp-1 pathway in response to PFT as well (not shown). Independent of IRE-1 activation, p38 MAPK can also activate TTM-2 and other PFT defenses. Tunicamycin, which causes the accumulation of unfolded proteins in the ER, activates IRE-1 via a mechanism independent of the PFT and p38 MAPK. doi:10.1371/journal.ppat.1000176.g006 20uC. This helps remove the background of macromolecules not isolated from the intestine. The response to Cry5B is not altered in this strain compared to wild type [6] . Primers used for these experiments are described in Protocol S1. Approximately 15,000 L4 stage worms were used per 100 mm dish for each treatment group. Worms were exposed to Cry5B for the indicated period of time on either E. coli JM103 containing empty vector or E. coli JM103 expressing Cry5B as described [6, 38] . After exposure to each treatment, worms were rinsed from plates with water, centrifuged at 500 g for 45 seconds, and washed two additional times with water. RNA was prepared from worms using TRIZOL (Invitrogen) and further purified with RNeasy columns (Qiagen). cDNA was prepared by reverse transcription using oligo-dT. Standard PCR was used to detect xbp-1 splicing, and products were analyzed on 2% agarose gel. Unspliced xbp-1 transcript is 220 nucleotides and spliced transcript is 197 nucleotides. To quantitate the amount of xbp-1 splicing, loading was normalized by quantitating cDNA levels using real time PCR and eft-2 primers [6] . Equal amounts of cDNA were used for the xbp-1 splicing PCR experiments and 10 microliters of each reaction were loaded onto a 2% agaose gel and stained with ethidium bromide. NIH ImageJ was then used to quantitate the intensities of xbp-1 spliced forms in Cry5B treated samples relative to untreated samples at the same time point. Real time PCR was performed on an ABI 7000 Instrument using SYBR Green detection (Applied Biosystems). eft-2 was used as the real time PCR normalization control [6] . Experiments with Cry5B used either a control plate (E. coli not expressing Cry5B) or a Cry5B plate on which 100% of the E. coli expressed Cry5B. Tunicamycin experiments used E. coli OP50 as a food source and either DMSO as the control or tunicamycin at 2 mg/mL incorporated into the plates. Three independent experiments for the splicing and real time PCR were performed for each treatment. HeLa cells were cultured in MEM media supplemented with 10% fetal calf serum, 1% penicillin-streptomycin, 1% glutamine and 1% non-essential amino acids, in a humidified incubator with 5% CO 2 at 37uC. Aerolysin was purified as described [41] . Cells were continuously treated with 2 ng/mL (0.02 nM) of proaerolysin. At different time points, cells were washed with PBS and lysed at 4uC in 0.25 M sucrose supplemented with proteases inhibitor (Roche, Germany) using a needle. The whole cell extracts were subjected to SDS-PAGE and Western blotting. XBP1 (R-14) antibody was from Santa Cruz Inc. Band intensities were quantified, after background removal, using ImageJ software (NIH). The loading in each lane was normalized relative to the intensity of a nonspecific antibody-reacting band on the blot. Approximately 750 L1 stage worms were grown in a single well of a 48 well plate containing 150 mL S media [42] and E. coli OP50. When worms had reached the L4 to young adult stage, glucose was added to 100 mM and either 20 mM HEPES (pH 8.0) or Cry5B dissolved in 20 mM HEPES (pH 8.0) to give a final concentration of 100 mg/mL was added. After one hour, worms were removed, centrifuged, and 175 mL of media was removed. Twenty five mL of 26 sodium dodecyl sulfate loading buffer was added, and worms were boiled for 5 minutes. Ten microliters of lysate were used for immunoblotting. Monoclonal antibody to phospho P38 MAPK (Cell Signaling Technology cat. no. 9215) was used at 1:300 and monoclonal antibody to a-tubulin (Sigma-Aldrich cat. no. T6199) was used at 1:4000. L4 stage glp-4(bn2) and glp-4(bn2);xbp-1(zc12) worms were used for this experiment. Approximately 80,000 worms of each strain were used for both control and Cry5B treatments. Control plates consisted of 100 mm plates spread with E. coli that did not express Cry5B, while Cry5B treatments consisted of plates in which 100% of the E. coli expressed Cry5B. Approximately 20,000 worms were used per plate. Worms were fed on the bacteria for 6 hours at 20uC. For details of mass spectrometry, please see Protocol S1. All experiments were performed a minimum of three times. LC 50 values were determined by PROBIT analysis [43] . The lethal concentration assays are represented graphically using nonlinear regression performed with the software GraphPad Prism. Statistical analysis between two values was compared with a paired t-test. Statistical analysis among three or more values was compared with matched one way ANOVA using the Tukey post test. Lifespan data was analyzed with Kaplan-Meier survival curves. Statistical significance was set at p,0.05. Protocol S1 Found at: doi:10.1371/journal.ppat.1000176.s001 (0.03 MB DOC)

Composition and Function of Haemolymphatic Tissues in the European Common Shrew

BACKGROUND: Studies of wild animals responding to their native parasites are essential if we are to understand how the immune system functions in the natural environment. While immune defence may bring increased survival, this may come at a resource cost to other physiological traits, including reproduction. Here, we tested the hypothesis that wild common shrews (Sorex araneus), which produce large numbers of offspring during the one breeding season of their short life span, forgo investment in immunity and immune system maintenance, as increased longevity is unlikely to bring further opportunities for mating. In particular, we predicted that adult shrews, with shorter expected lifespans, would not respond as effectively as young animals to infection. METHODOLOGY/PRINCIPAL FINDINGS: We examined haemolymphatic tissues from wild-caught common shrews using light and transmission electron microscopy, applied in conjunction with immunohistology. We compared composition and function of these tissues in shrews of different ages, and the extent and type of inflammatory reactions observed in response to natural parasitic infections. All ages seemed able to mount systemic, specific immune responses, but adult shrews showed some signs of lymphatic tissue exhaustion: lymphatic follicles in adults (n = 21) were both smaller than those in sub-adults (n = 18; Wald = 11.1, p<0.05) and exhibited greater levels of depletion (Wald = 13.3, p<0.05). CONCLUSIONS/SIGNIFICANCE: Contrary to our expectations, shrews respond effectively to their natural parasites, and show little indication of immunosenescence as adults. The pancreas of Aselli, a unique lymphoid organ, may aid in providing efficient immune responses through the storage of large numbers of plasma cells. This may allow older animals to react effectively to previously encountered parasites, but infection by novel agents, and eventual depletion of plasma cell reserves, could both still be factors in the near-synchronous mortality of adult shrews observed shortly after breeding.

The immune system is the primary mechanism through which animals defend against parasites and pathogenic organisms. Immunity is believed to be a major factor in regulating host survival, as under natural conditions, even a mild parasitic infection may weaken an animal sufficiently to increase the chances of mortality through starvation or predation [1] . However, maintenance and up-regulation of the immune system requires energetic and nutritional resources [1] , resulting in a trade-off between investment in immunity and other physiological processes, including growth and reproduction [2] [3] [4] . An appreciation of the extent to which hosts invest in immunity is therefore critical to understanding the strategies through which animals maximise fitness, and how trade-offs are mediated between current offspring production, longevity and future reproductive success [3, 4] . While there have been a number of studies of ecological immunology in birds and insects [4] [5] [6] , there has been little effort to understand immune function of small mammals in this context outside of the laboratory. Here, we examined the unique haemolymphatic system of the European common shrew (Sorex araneus) to investigate the capacity of this short-lived mammal, restricted by a fast metabolism and extremely limited fat reserves, to defend against its unusually diverse parasite fauna, both as a young animal and an adult. Common shrews have attracted considerable attention from both ecologists and parasitologists, and have a life history strategy characterized by a high investment in reproduction and a short life span [7] [8] [9] . Their life cycle takes 14 to 16 months to complete, with the first young born in mid-May [10] [11] [12] , and only one breeding season in the spring of the second year of life [9, 12, 13] . Both sexes can mate with multiple partners, and females are extremely promiscuous [14, 15] . Females can produce up to three litters of around seven offspring [10, 11, 16] , with energy intake during lactation increasing to around three times the non-reproductive level [8] . Both males and females die shortly after breeding, such that there is little overlap between generations [12, 13, [17] [18] [19] . S. araneus has an unusually diverse parasite fauna, which includes ectoparasites [20] [21] [22] , Bartonella and trypanosome infections [23] [24] [25] [26] [27] , Anaplasma phagocytophilum [27, 28] , and Pneumocystis carinii [29] , as well as over 20 helminth species [30] . The extent to which shrews are able to mount immune responses to these parasites is unknown: commons shrews have a fast metabolic rate even for their small body size [31] but store very little energy as fat [32] . Lack of resources may therefore limit the capacity of S. araneus to mount immunological responses, particularly as reproductive adults, and parasitism has been suggested as one of the causes of mortality of adults after breeding [12, 13, 18] . Common shrews possess a large lymphatic organ, known as the Pancreas of Aselli, which may function in defence against parasites, though its exact role is unknown, and remains the subject of discussion [33] [34] [35] . To date, there have been no studies of spleen or bone marrow function in S. araneus. We hypothesised that common shrews, which are not expected to survive beyond the first breeding season, would gain little benefit from investing their limited resources in immunity and immune system maintenance, at the expense of reproduction. Instead, we predicted that wild shrews would demonstrate only limited responses to parasites, and that their immune system would show signs of deterioration with age. The aim of the study was therefore to evaluate the capacity of sub-adult and adult shrews to mount immunological responses. We examined and compared the structure, composition and function of relevant haemolymphatic tissues including the pancreas of Aselli, in wild-caught common shrews of different ages pre and post maturation, and the extent and type of inflammatory reactions produced in response to naturally occurring parasitic infections. Light and electron microscopy were applied in conjunction with immunohistological characterisation of leukocyte populations. Contrary to our predictions, our results indicated that shrews are capable of mounting immune/inflammatory responses throughout their entire life span. While some degree of lymphatic exhaustion was obvious in adult animals (perhaps as a result of age-related changes, or reduced investment in immunity as a consequence of breeding effort), there was also evidence of some degree of compensation, in the form of storage of plasma cells particularly in the pancreas of Aselli, possibly as a defence against previously encountered parasites. Forty-three common shrews (19 male, 24 female) were live-caught in Cheshire, England between September 2001 and June 2003. The work was performed with approval of and under a licence from English nature (licence number 20030767) held by PS. Shrews were classified into three age categories: 18 sub-adults, showing no sign of sexual development (11 female, 7 male), 3 animals undergoing sexual maturation (2 female, 1 male) and 22 sexually mature animals (11 female and 11 male) caught during or after the breeding season. Adult females all exhibited signs of mating, pregnancy and/or lactation. All animals appeared healthy when captured, and were killed humanely by overdose of inhalation anaesthetic (Fluothane, Schering-Plough Animal Health, UK). Animals were inspected for ectoparasites (data not presented) before full necropsy was performed and body mass (minus gastrointestinal tract) recorded. Bladder and oesophagus were removed and dissected in Hanks saline (406 magnification), with nematode and digenean parasites in both tissues counted and identified using keys [30] . Stomachs and guts were removed, weighed, stored in 10% formalin and later dissected in Hanks saline (406 magnification). Recovered helminths were identified as nematodes, cestodes or digeneans [30] and counted. Tissue samples from all major organs from all animals were fixed in 4% buffered paraformaldehyde for 24-48 h prior to routine embedding in paraffin wax. Sections (5 mm) were stained with haematoxylin-eosin for histological evaluation. Sections from haemolymphatic tissues (spleen, pancreas of Aselli, bone marrow (sternum) and in selected cases mesenteric or mediastinal lymph nodes and thymus) were prepared for immunohistological examinations and the TUNEL method. Samples from the pancreas of Aselli of one sub-adult common shrew were fixed in 2.5% glutaraldehyde and 4% paraformaldehyde in cacodylate buffer (pH 7.4) and subsequently embedded in epoxy resin. Semi-thin (1 mm) and thin sections were prepared and the latter examined using transmission electron microscopy. Labelling of leukocytes, proliferating and apoptotic cells by immunohistology and the TUNEL method Leukocytes, proliferating and apoptotic cells in haemolymphatic tissue samples from 21 common shrews were identified using immunohistology and the TUNEL (terminal deoxynucleotidyl transferase-mediated dUTP-biotin nick end labelling of DNA fragmentation sites) method respectively. For immunohistology, monoclonal and polyclonal antibodies (cross) reacting in other species were used in conjunction with peroxidase anti-peroxidase and avidin biotin peroxidase complex methods as previously described [36] [37] [38] [39] . Antibodies, their sources and detection methods are listed in Table 1 . Cellular turnover in haemolymphatic tissues was assessed by counting proliferating and apoptotic cells. Proliferating cells were identified by their expression of the proliferating cell nuclear antigen (PCNA; [38] [39] [40] ), while apoptotic cells were demonstrated in situ by the TUNEL method ( [39, 41] ) using a commercially available kit according to the manufacturer's instructions (ApopTag TM In Situ Apoptosis Detection Kit; Chemicon, California, USA). Consecutive tissue sections, incubated with normal rabbit or rat serum or a non-reacting mouse monoclonal antibody, were used as negative controls for polyclonal and monoclonal antibodies respectively. For TUNEL, terminal deoxynucleotidyl transferase (TdT) was replaced by distilled water on negative control slides. All antibodies used cross-reacted with shrew leukocytes. The B cell markers CD45R and CD79a were expressed by B cells, CD45R being strongly expressed in follicular germinal centres, but relatively faintly in well-differentiated B cells within follicular mantle zones, whereas CD79a was mainly expressed by mature B cells (weak expression in germinal centres (dark zone), strong expression in the periphery of secondary follicles, positive reaction of all cells in primary follicles). Plasma cells were negative or exhibited faint staining for both B cell markers. CD3 acted as a pan T cell marker, being expressed by entire T cell zones. Both the myeloid/histiocyte antigen and lysozyme were markers for mature monocytes/macrophages and their precursors, myelomonocytic cells, as they were expressed by monocytes and macrophages and a high percentage of cells in the bone marrow. However, they seemed not to be a marker for all neutrophils, as in both the splenic red pulp and sinuses of the pancreas of Aselli, a proportion of cells with the morphology of neutrophils were negative for both antigens. PCNA-positive, proliferating cells were present in follicle germinal centres, T cell zones, bone marrow (including megakaryocytes) and splenic red pulp, the sites in the haemolymphatic tissue expected to contain proliferating cells. The TUNEL method identified cells with the morphology of apoptotic cells [39, 41] as well as apoptotic bodies (both free and within tingible body macrophages), located predominantly in follicular germinal centres. Assessment of lymphatic follicle and bone marrow activity, and the severity of inflammatory infiltration in the liver Spleen and lymph node exhibited a composition very similar to laboratory mice, which allowed direct comparison for evaluation. Lymphatic follicles in the spleen and lymph nodes of all animals were classified as primary and/or secondary follicles and as small, medium or large. The presence and degree of follicle depletion (none, mild, moderate or severe) was assessed on the basis of the cellularity of the germinal centres. Large, undepleted secondary follicles were interpreted as evidence of high activity, whereas small, depleted primary follicles indicated lowest activity. Similarly, bone marrow activity was classified as low, moderate or high based on the ratio of haematopoietic cells to adipose tissue in a cross section of the marrow in the sternum. The degree of inflammatory infiltration in the liver was assessed semi-quantitatively as mild, moderate or severe, based on the number of cells and cell layers in the portal areas or between hepatic cords. Statistical analyses were restricted to sub-adult and adult animals as only 3 pubescent shrews were caught. Ordinal logistic regression examined whether bone marrow activity, lymphatic follicle activity (primary/secondary follicles, presence and degree of follicular depletion) in the spleen and pancreas of Aselli and severity of inflammatory infiltration in the liver (mild, moderate or severe) varied with age class or sex. Sex and age category were entered simultaneously as independent variables into models for each dependent variable listed above, and the significance of both terms assessed using Wald tests. Mann-Whitney U tests were used to test for differences in helminth abundances between sub-adult and adult shrews. The spleen in adult shrews exhibits lesser activity in the white pulp but higher cellularity in the red pulp compared to the spleen in sub-adults Red and white pulp of shrews of all age groups were examined for their composition and cellular turnover. The white pulp (lymphatic follicles and T cell zones) was generally confined to the organ's centre, where follicles were arranged singly or in groups ( Fig. 1 ). Germinal centres exhibited numerous proliferating, PCNA-positive cells (up to 50%) and few (#5%) apoptotic cells. T cell zones were arranged around medium-sized arteries, forming periarterial lymphatic sheaths similar in size and cell density in all animals, exhibiting up to 10% proliferating, PCNA-positive cells and generally few apoptotic cells. The generally cell-rich red pulp contained neutrophils, erythrocytes and smaller numbers of lymphocytes (each ,5% up to 30% T cells and mature B cells) and macrophages as well as numerous evenly distributed megakaryocytes (Fig. 1B) ; the red pulp exhibited some degree of cellular turnover, with approximately 10% proliferating, PCNApositive cells and scattered apoptotic cells. Follicles and follicle groups were often delineated by a variably distinct rim of macrophages and neutrophils (Fig. 1C) , together with variable numbers of mature, strongly CD79a-positive B cells, T cells and erythrocytes. In general, 30-50% of cells in the red pulp were positive for myeloid/histiocyte antigen (Fig. 1D ) and lysozyme and had the morphology of monocytes/macrophages or neutrophils. Age groups differed in both the composition and functional state of the spleen. Within the red pulp, the amount of neutrophils and Table 1 . Antibodies and detection methods used to identify leukocytes and proliferating cells in the common shrew with references to suppliers and use in other species. [36, 53] . f [37] . g [53] . h [38] [39] [40] . megakaryocytes often appeared greater in adult animals than in sub-adults. The white pulp of sub-adults was exclusively comprised of secondary follicles, which often appeared interconnected and formed large groups (Fig. 1A) . These follicles were for the most part large and without signs of depletion and included numerous apoptotic cells and tingible body macrophages, as well as several mitotic cells. In contrast, the majority of adults exhibited fewer follicles which were a mixture of primary and secondary follicles (Figs. 1B, 2) and generally smaller than those of sub-adults (n = 17, Wald = 17.17, p,0.05, Figs. 1A, 2). Follicles in adults also differed from those in sub-adults, in that they were only partially surrounded by distinct perifollicular rims. Where present, germinal centres in adults contained both mitotic as well as apoptotic cells. Follicle centres in four adults exhibited collagen deposition, and while there was a tendency for greater follicle depletion in older animals (Fig. 2) , the difference between adults and sub-adults was not statistically significant (Wald = 2.10, not significant (ns)). The white pulp of the three pubescent animals exhibited primary and/ or secondary follicles, the latter with features similar to those found in sub-adult shrews. Males (n = 18) and females (n = 21) did not differ with respect to either size (Wald = 0.67, ns) or depletion (Wald = 0.13, ns) of follicles in the spleen. Lymph node composition is similar in all age groups, with a relatively high proportion of plasma cells, particularly in adults Mesenteric or mediastinal lymph nodes were examined from selected sub-adult and adult animals. All features normally associated with mammalian lymph nodes were represented: a cortex containing primary and secondary follicles, paracortex, lymphatic cords, medulla and both marginal and medullary sinuses. Compared to lymph nodes in other mammalian species [39, 42] , the medulla often appeared to contain a high number of plasma cells, particularly in adults. No other differences were observed between the age groups. The pancreas of Aselli represents a specialised abdominal lymph node that appears to function as a plasma cell store, in particular in adult shrews In the past, there has been some controversy as to the composition and function of the pancreas of Aselli (lymph nodelike or equivalent to the avian bursa of Fabricius [35] ). The aim of this study was therefore to clarify this matter with up-to-date methodology. In general, the composition of the pancreas of Aselli was very similar to that of a lymph node. Beneath the capsule were marginal sinuses of variable width, containing disseminated lymphocytes (mostly T and B cells in equal proportions), macrophages and neutrophils, the latter either disseminated or as small accumulations. In four adult shrews, the marginal sinuses exhibited focal to extensive fibrosis. Beneath the sinuses lay a cortex containing exclusively secondary follicles ( Fig. 3A-C) , with the exception of one adult male where both primary and secondary follicles were present (Fig. 4) . Germinal centres generally exhibited a high cellular turnover, with variable but high numbers of PCNA-positive, proliferating cells (often more than 50% of cells; Fig. 3E ) and often numerous apoptotic cells (Fig. 3F ). T cell zones formed a paracortex located immediately beneath the follicles (Fig. 3D ) and were generally similar in size and cell density in animals of all age groups. The organ's centre (medulla) contained loosely arranged sinuses with only low numbers of macrophages. The remainder of the medulla was made up almost entirely of plasma cells (Fig. 3G, H) . In sub-adult animals the cortex generally appeared tightly packed with large follicles that exhibited no, mild or moderate depletion (Figs. 3A, 4) . In adult shrews, the cortex often contained only a small number of follicles (9/20 animals; 45%), frequently with large areas of cortex devoid of follicles (5/20; 25%; Fig. 3B ). Follicles occasionally seemed to extend outwards into marginal sinuses and in two animals exhibited central collagen deposition. Follicles in adults (n = 21) were both smaller than those in sub-adults (n = 18; Wald = 11.06, p,0.05, Fig. 4 ) and exhibited greater levels of depletion (Wald = 13.28, p,0.05, Fig. 4) . No difference was found between males (n = 17) and females (n = 22) with respect to follicle size (Wald = 0.30, NS) or depletion (Wald = 0.42, NS). Cortical areas devoid of follicles were also devoid of T cell zones. As a consequence, overall numbers of T cells in the pancreas of Aselli often seemed lower in adult than sub-adult animals. Where the cortex was devoid of follicles, plasma cells extended from the medulla to the marginal sinuses or beyond to the capsule, such that the medulla often occupied most of the organ in adult animals (Fig. 3B) . Accordingly, in adult animals, the whole organ frequently appeared as an accumulation of plasma cells, surrounded by a fragmentary cortex and paracortex. Extramedullary haematopoiesis was observed in 9/21 (43%) adult animals, as represented by scattered megakaryocytes within the outer medulla. This was not seen in sub-adult animals. Thymic tissue was recovered from seven animals across all age groups. This generally consisted of a variable number of lymphocyte layers which were arranged around blood vessels and encased by a thin capsule of fibrous connective tissue. Taking the histological features of a mouse thymus in account, the thymus appeared to exhibit a variable, but generally low degree of involution in all examined shrews, regardless of age. The bone marrow generally exhibits moderate to high activity, regardless of age The bone marrow generally exhibited moderate to high activity. All haematopoietic cell types known from other mammalian species were present. Approximately 10% of cells were identified as T cells, 10% as B cells (CD79a-positive, CD45R-negative; interpreted as circulating mature B cells) and 30% to 50% were myeloid/histiocyte antigen-and/or lysozyme-positive cells. At least 30% to 40% of all cells were PCNA-positive, representing relatively high levels of proliferation. No difference in bone marrow activity was found between sub-adult (n = 15) and adult (n = 16) shrews (Wald = 1.91, ns) or between males (n = 12) and females (n = 19; Wald = 0.74, ns). All animals harboured helminths ( Table 2 ). The abundance of infection was greater in adults than sub-adults (Table 3) . Nematodes (Liniscus incrassatus; Roots, 1992) recovered from the lumen and between epithelial cells of the urinary bladder were associated with mild to moderate acute to chronic lymphocytedominated cystitis and/or mild degeneration and sloughing of epithelial cells. Infestation of the gall bladder by the digenean Dicrocoelium soricis resulted in only a mild lymphocytic submucosal infiltration in one of two animals infected. Neither helminths within the gastrointestinal tract nor Porrocaecum sp. larvae within interscapular adipose tissue were associated with inflammatory reactions or other histological changes. A high proportion of shrews (14/18 sub-adults; 78%, 2/3 pubescent animals; 67%, 12/22 adults; 55%) exhibited protozoan parasites within vessel walls that could not be further identified ( Table 2 ). These occurred predominantly in kidneys (21/43 shrews; 49%) and myocardium (12/43 shrews; 28%), but also occasionally in the liver, the splenic red pulp, the medulla of the pancreas of Aselli and the wall of the urinary bladder. In general, these cysts did not induce any alterations apart from an occasional slight thickening of the affected vessel wall, or a mild granulomatous inflammatory infiltration. Protozoan cysts with features of Sarcocystis sp. were found within skeletal muscle myocytes of one adult female, without any associated reaction. The liver of all animals exhibited a variable degree of mixed cellular (neutrophils, lymphocytes and macrophages) portal inflammatory infiltration, where T cells and B cells were present in equal amounts. No difference in the severity of the inflammatory infiltration was apparent between sub-adults (n = 18) and adults (n = 22; Wald = 0.03, ns), or between males (n = 22) and females (n = 18; Wald,0.01, ns). Infiltrates often occurred together with follicle-like accumulations of lymphocytes, which occasionally exhibited germinal centres. Additional findings in the liver included a granuloma with central necrosis in one sub-adult, focal necrosuppurative hepatitis in two adults and variably intense (multi)focal hepatic necrosis with haemorrhage and pyogranulomatous inflammation in another three adults. In one pubescent shrew helminth parasites were observed within and outside multifocal suppurative hepatitis and haemorrhage. Five of the 43 animals exhibited granulomas within the pancreas of Aselli, with a central area of necrosis and/or mineralization (four shrews) or an embedded nematode (one animal). Focal pyogranulomatous (two adults) or suppurative inflammation (two adults), the latter in one case surrounding a nematode, were also observed. The marginal sinuses of one subadult shrew contained extensive focal accumulations of neutrophils, occasionally surrounding areas of necrosis. None of the animals exhibited other major gross or histological changes. In particular there were no findings suggestive of a systemic non-infectious disease and/or a neoplastic disease. While studies of laboratory animals allow aspects of immunity to be studied in controlled, repeatable environments, they may not reflect how wild animals, constrained by limited resources and at threat from a variety of infectious agents, respond to parasites and disease. We predicted that common shrews, which are short lived, store little energy as fat, and invest heavily in reproduction, would show limited responses to parasites, and their haemolymphatic tissues would deteriorate with age as expected survival decreased. Our study is the first to examine haemolymphatic tissue structure, compoosition and function in shrews using modern techniques, and one of only a small number to explore immune responses of wild animals in their natural environment. With regards to both morphology and composition, the spleen, lymph nodes, thymus and bone marrow in S. araneus were found to be very similar to their equivalents in other mammalian species [39, 42] . In common with a number of species, including the musk shrew Suncus murinus, both bone marrow and spleen were identified as sites of haematopoiesis in S. araneus [43] . The results of our study also confirm that the pancreas of Aselli, which is specific to shrews, can be considered as a large, specialised lymph node [33] . The presence of a cortex with both follicles and paracortical T cell zones renders previous controversial assumptions regarding the organ's function incorrect: the pancreas of Aselli is neither a specific site of exclusive B cell production nor a functional analogue of the bursa of Fabricius in birds [35] . However, it differs from normal lymph nodes in that the centre (medulla) contains a very high proportion of plasma cells. In adulthood, the number of plasma cells and the relative size of the medulla seem to increase, until almost the entire organ is composed of plasma cells. Such a feature has not been described under physiological circumstances in any other species, and suggests that the pancreas of Aselli in S. araneus functions as a storage site for plasma cells, particularly in older animals. Lymph nodes in S. araneus were also found to contain a higher number of plasma cells than normally observed in other species [39, 42] , which may emphasise a general tendency towards progressive plasma cell storage in common shrews. In comparing the lymphatic tissues of sub-adult and adult shrews, young animals were generally found to have an activated immune system, as represented by a predominance of large, active, secondary follicles in the spleen and pancreas of Aselli. This suggests sub-adults were responding effectively to a diverse array of infectious agents, including the helminth and protozoan parasites detected in a number of tissues. Post-reproductive animals, however, exhibited characteristics indicative of immune system exhaustion: follicles were generally smaller and were often depleted, with a smaller proportion of secondary follicles, particularly in the spleen. This could indicate decreased follicular activity in adult animals, with impaired germinal centre reactions resulting in reduced B cell production [44] . Impairment of germinal centre reactions is a known feature of immunosenescence in vertebrates and has been studied extensively: in humans it has been shown to be a product of defective T cell-dependent B cell activation [44, 45] . Reduced lymphocyte production as a consequence of follicular and T cell impairment could explain why significantly lower numbers of white blood cells, and specifically lymphocytes, have been reported in old common shrews [46] . Differences between sub-adults and adults were also evident in the so-called ''marginal'' or ''intermediate zone'' of the spleen, the variably distinct rim of macrophages, lymphocytes, neutrophils and erythrocytes surrounding follicles and follicle groups seen in S. araneus and previously described in other species, including the musk shrew [43, 47] . This zone is considered to be the site of most intensive blood filtration in the spleen [43, 47] , and its loss of integrity/intensity in older animals may indicate a reduction in filtration capacity. This might however be counterbalanced by an increase in the capacity of the peripheral phagocytic response, as represented by an increase in neutrophil numbers within both the spleen (as observed here) and the peripheral blood [46] . Both in bone marrow and splenic red pulp, the degree of haematopoiesis was similar in animals from all age groups. This concurs with similar findings in the musk shrew [43] , where both the splenic red pulp and bone marrow have been identified as physiological sites of erythropoiesis, leukocytopoiesis and platelet production over the animal's lifespan. In this aspect, shrews are similar to some reptiles, whereas in other mammals the haematopoietic capacity of the spleen seems to cease after birth [43] . We also found evidence of haematopoiesis in the pancreas of Aselli in adult S. araneus, as demonstrated by the presence of megakaryocytes in the medulla of some individuals. Interestingly, we found no evidence of major thymic involution in S. araneus, even in older animals. The rate of thymic involution is known to vary between species and breeds and with intraspecific factors such as sex and diet [48, 49] . Perhaps in shrews thymic involution is delayed to maintain production of T cells into adulthood. It has been suggested that short-lived species should limit their investment in immunity to immediate, innate responses, as the energetic costs associated with mounting specific immune reactions are unlikely to be outweighed by the benefits of increased long-term survival [1] . The dependence on innate responses may be greater for species with limited energetic reserves (such as S. araneus), as even a mild immune challenge is likely to result in starvation if allowed to persist for more than a short time [1] . Here, however, the presence of numerous active secondary follicles in the spleen and pancreas of Aselli, the development of small lymphatic follicles in portal areas in the liver and the generally high number of plasma cells in the pancreas of Aselli all indicate that common shrews remain consistently able to mount systemic, specific immune responses. We also observed macro- phage-dominated (granulomatous) inflammatory reactions with lymphocyte involvement in both sub-adult and adult shrews, which included reactions to helminths in tissues. The increasing number of plasma cells in the medulla of the pancreas of Aselli and in lymph nodes with advancing age might even suggest a 'refocusing' of the immune system, from reacting to novel antigens in follicles as a young animal, to combating previously experienced parasites or pathogens with appropriate antibody responses as an adult. Plasma cells are long-lived and can survive for weeks after immunisation, particularly when not too tightly packed [50] ; perhaps young common shrews invest in long term immunity by producing and storing plasma cells in the pancreas of Aselli, which can then be used to mount efficient responses against previously encountered parasites in adulthood, when reproduction places greater demands on internal resources [8] . While this strategy may allow older animals to react effectively to previously encountered parasites, infection by novel agents or eventual depletion of plasma cell reserves, could still be factors in the near-synchronous mortality of adult shrews observed shortly after the breeding season [12, 13, 18]

Prevention of Cytotoxic T Cell Escape Using a Heteroclitic Subdominant Viral T Cell Determinant

High affinity antigen-specific T cells play a critical role during protective immune responses. Epitope enhancement can elicit more potent T cell responses and can subsequently lead to a stronger memory pool; however, the molecular basis of such enhancement is unclear. We used the consensus peptide-binding motif for the Major Histocompatibility Complex molecule H-2K(b) to design a heteroclitic version of the mouse hepatitis virus-specific subdominant S598 determinant. We demonstrate that a single amino acid substitution at a secondary anchor residue (Q to Y at position 3) increased the stability of the engineered determinant in complex with H-2K(b). The structural basis for this enhanced stability was associated with local alterations in the pMHC conformation as a result of the Q to Y substitution. Recombinant viruses encoding this engineered determinant primed CTL responses that also reacted to the wildtype epitope with significantly higher functional avidity, and protected against selection of virus mutated at a second CTL determinant and consequent disease progression in persistently infected mice. Collectively, our findings provide a basis for the enhanced immunogenicity of an engineered determinant that will serve as a template for guiding the development of heteroclitic T cell determinants with applications in prevention of CTL escape in chronic viral infections as well as in tumor immunity.

Despite the antigenic complexity of microbes, primary pathogen-specific cytotoxic CD8+ T lymphocyte (CTL) responses are commonly directed to just one or a few determinants. Furthermore, even when multiple epitopes are targeted, distinct patterns of epitope hierarchy often emerge. Such immunodominant epitopes commonly elicit high-magnitude CTL responses characterized by potent cytolytic function, whereas subdominant determinants generate responses that are relatively lower in magnitude and often less efficacious. In general, potent anti-viral CTL strongly correlate with control of infection and less clinical disease. Viral progeny selected on the basis of CTL surveillance can evolve to evade T cell responses. This selective pressure results in mutations in immunodominant CTL determinants that abrogate recognition. CTL escape virus is commonly observed in humans and nonhuman primates infected with HIV-1, hepatitis C virus (HCV) or simian immunodeficiency virus (SIV) and its selection often correlates with disease progression [1] [2] [3] [4] . Escape mutations may diminish binding to the restricting MHC class I molecule, interfere with T cell receptor (TcR) recognition or interfere with antigen processing [5] [6] [7] . Escape mutations are usually detected in epitopes targeted by CTL that exhibit high functional avidity because the corresponding potent CTL response exerts high selective pressure on the virus [2] . In some HIVinfected patients, CTL escape occurs without an associated enhancement of virus replication, suggesting that the mutations compromised virus fitness or, alternatively, the variant determinant elicited a de novo CTL response [8, 9] . Virus fitness is sometimes restored, with concomitant increased virus replication, when a second, compensatory mutation is selected [10, 11] . Collectively these results suggest that given the importance of virus diversification (CTL escape) in disease progression, suppression of selection or outgrowth of CTL escape variants should improve outcomes in persistently infected animals and humans. Modulating the immunogenicity of subdominant CTL determinants could potentially lead to the development of more efficacious vaccines that are more broadly protective and prevent or minimize the appearance of variant viruses that have mutated in dominant epitopes targeted by high-avidity CTL responses. Enhancement strategies, which result in augmented responses to the native, subdominant epitope, have been described for both MHC class I and class II-restricted determinants, whereby the most common approaches involve generating a series of conserved and non-conserved mutations at MHC anchor residues, followed by an empiric determination of whether each individual substitution augments T cell effector function [12] . Furthermore, evaluating the effect of epitope enhancement in vivo has generally been achieved via heterologous infection systems [13] [14] [15] . Thus, while these results demonstrate proof of principle, direct evidence for enhanced protection against autologous microbial infection in vivo is lacking. The design of heteroclitic determinants in which non-MHC anchor residues are targeted for substitution are also usually determined empirically. Studies of heteroclitic tumor epitopes have demonstrated the clinical utility of such determinants [16, 17] . Notably, however, there are no well defined examples of viral epitopes which demonstrate enhanced immunity and the molecular basis for the enhanced immunogenicity is not well understood. Potential interventions to directly manipulate host-pathogen interactions and thereby diminish CTL escape variant selection are often difficult to evaluate since most examples of CTL escape occur in infected humans or non-human primates. By contrast, mice persistently infected with mouse hepatitis virus (MHV) strain JHM (JHMV) serve as a useful system for investigating anti-viral CTL responses and CTL escape [18] [19] [20] . Two JHMV-derived CTL epitopes are recognized in C57BL/6 (B6) mice. The immunodominant H-2D b -restricted CTL epitope (S510, CSLWNGPHL, spanning residues 510-518 of the Spike (S) glycoprotein) elicits a highmagnitude, high-avidity CTL response that drives virus diversification during persistent infection [18] [19] [20] [21] [22] . A second subdominant CTL epitope, S598 (H-2K b -restricted, RCQIFANI, spanning residues 598-605 of the S glycoprotein) also elicits an appreciable CTL response [21] ; however, S598-specific CTL exhibit ,100-fold lower functional avidity and do not protect from CTL escape in S510. The presence of a readily mutable dominant and a subdominant epitope with high and low functional avidity, respectively, in JHMV-infected mice is useful for investigating both epitope enhancement and approaches to diminishing CTL escape. Consistent with this notion, we have previously shown that the introduction of a second dominant CD8 T cell epitope into the JHMV genome (GP33 from lymphocytic choriomeningitis virus (LCMV)) protected mice from the development of CTL escape in S510 and enhanced virus clearance [22] . Here we determined whether the CD8 T cell response to S598 could be enhanced so that it now elicited a more potent T cell response that protected mice against the development of CTL escape at S510 and subsequent clinical disease. We modified epitope S598 (S598 Q600Y ) such that it elicited a high-avidity CTL response and using the crystal structures of the H-2K b /S598 and H-2K b /S598 Q600Y complexes, determined the basis of this enhanced immune response. We then introduced this more immunogenic S598 epitope into a recombinant version of JHMV and showed that these high-avidity S598-specific CTLs protected against escape variants in the immunodominant S510 epitope. Immunization with the modified peptide resulted in an improved response to the native S598 epitope, demonstrating a true heteroclitic effect and suggesting that this strategy may have clinical applications for reducing viral titer and preventing CTL escape during chronic infections. As we have previously observed with other MHC/Cyscontaining peptide complexes [19, 23] , complexes did not readily form unless the cysteine of the S598 peptide (RCQIFANI) was modified with L-a-aminobutyric acid (Aba, an isostereomer of cysteine). The Aba-modified peptides maintained immunogenicity since a higher frequency of splenic CD8+ T cells from JHMVimmune mice reacted to Aba-modified S598 peptide, relative to unmodified S598 peptide ( Figure S1A ). Next, we used the consensus H-2K b binding motif [24] to engineer a novel, high-avidity S598 CTL epitope. Importantly, Gln-3 diverges from the consensus H-2K b -restricted ligand binding motif, in which a tyrosine is often present at position 3 [24, 25] . Therefore, we substituted a glutamine residue for tyrosine (Q600Y, CAA to TAT, RCYIFANI) at position 3 of the determinant with the aim of creating a peptide that bound more tightly to H-2K b . Stability of the H-2K b /S598 and H-2K b /S598 Q600Y complexes was assessed by circular dichroism (CD). As shown in Figure 1 , H-2K b /S598 Q600Y was considerably more thermostable than the native complex (Tm 54uC vs 64uC). To probe the biological properties of the Q600Y substitution, we used reverse genetics to engineer a recombinant version of JHMV expressing the S598 Q600Y epitope (Figure 2A) . Recombinant viruses encoding this substitution replicated as efficiently as wild type JHMV (rJ) in vitro during one-step growth kinetics analyses and in vivo virus competition assays ( Figure 2B , C). The immunogenicity of S598 Q600Y was assessed by intracellular expression of IFN-c by central nervous system (CNS)-derived lymphocytes. Since Cys-containing peptides are often diminished in Enhancing the immune responses to pathogens is a chief goal of vaccine development. Here, we describe the development of an engineered CD8+ T cell epitope that elicits an immune response to the native epitope that is more potent than the one that occurs during the natural infection. We showed that this ''improved'' (heteroclitic) epitope protects against clinical disease and against cytotoxic T cell escape that frequently occurs in the immunodominant epitope expressed by the virus. We also performed structural analyses and showed that enhanced immunogenicity was associated with changes in the conformations of both the peptide and the region of the MHC class I molecule that is in close association with the peptide. These studies provide a model for designing T cell epitopes with enhanced immunogenicity that will be useful in vaccine development, with particular emphasis on diseases, such as HIV and hepatitis C, in which epitope mutation and escape is common. Seven days p.i., total RNA was harvested from the brains of mice and relative representation of virus template was determined via RT-PCR and direct sequencing of PCR products. The relative proportion of animals in which only rJ, only rJ.S Q600Y , or a mixture of the two viruses is shown. (D) High-ability to elicit a CTL response, we included a reducing agent (TCEP, Tris [2-carboxyethyl] phosphine) in the cultures. TCEP enhanced the stimulatory capacity of the S598 peptides ( Figure S1B ), indicating that a proportion of the unmodified S598 peptide stock had undergone oxidation. Thus, we stimulated CNS-and spleen-derived CTL ex vivo in the presence of 500 mM TCEP, a concentration consistent with other work with Cys containing peptides [23, 26] . The S598 Q600Y epitope elicited a CTL response in the CNS of rJ.S Q600Y infected mice, with nearly 30% of all CD8 T cells recognizing the determinant ( Figure 2D ). In addition, CTL primed by S598 Q600Y cross-reacted with the native S598 determinant. The converse was not true, however, as cells primed by the native epitope failed to produce IFN-c when stimulated with S598 Q600Y peptide ( Figure 2D ). Of note, the CTL response to the dominant D b -restricted S510 epitope in rJ.S Q600Y -infected mice was diminished relative to responses in mice infected with wildtype rJ virus ( Figure 2D , E). Next, we assessed the relative functional avidity of CTL populations primed by native S598 and S598 Q600Y determinants, as a surrogate measure of the potency of the anti-virus CTL response in vivo. CNS-derived mononuclear cells were harvested from mice infected with rJ or rJ.S Q600Y and examined for IFN-c expression after stimulation ex vivo in the presence of 10-fold dilutions of the appropriate peptide. Cells primed to the native S598 epitope (cells harvested from rJ-infected mice) required approximately 100-fold more peptide than did S598 Q600Y -primed cells to elicit a half maximal response (100 nM vs. 1 nM, Figure 3A ,B). Because we observed that a subpopulation of S598 Q600Y -primed cells cross-reacted with the native determinant, we next determined whether this subpopulation also exhibited high functional avidity. For this purpose, we isolated cells from the CNS of rJ-and rJ.S Q600Y -infected adult B6 mice and stimulated them with 10-fold dilutions of native S598 peptide. S598 Q600Y -primed cross-reacting cells exhibited 7-fold higher functional avidity, relative to those primed by the native determinant ( Figure 3C , D) and are therefore distinct from CTL primed by the native S598 determinant. Consistent with the presence of a distinct population of cells, there were modest differences in Vb chain utilization when total populations from rJ and rJ.S Q600Y -infected mice and when total and cross-reacting cell populations from rJ.S Q600Y -infected mice ( Figure S2 ) were compared. In all instances, Vb5.1/5.2 expression was relatively over-represented, but some Vb elements were preferentially utilized by specific responding populations (Vb11, Vb13 and Vb14). Also consistent with this observation, alanine scanning mutagenesis of the two determinants revealed that the CTL response to each cognate peptide following infection was also subtly different reflecting the altered repertoire. The CTL response to each cognate peptide was very sensitive to mutation at every position except 2 and 9 although mutations in the S598 Q600Y determinant were tolerated slightly better than changes in the S598 epitope ( Figure S2C -E). If CTL recognizing S598 Q600Y exhibit high functional avidity in vivo, they should protect from CTL escape in S510 and might select for S598 Q600Y CTL escape variants. Diversification at S510 is observed in pups infected with neurovirulent JHMV at 10 days of age and nursed by JHMV-immune dams [27] . These mice are largely protected from developing acute lethal encephalitis, but a variable percentage (30-90%) later develop a demyelinating encephalomyelitis. Infectious virus isolated from these mice with late onset clinical disease is mutated in S510, resulting in enhanced virus replication [20] , with demyelination occurring during the process of virus clearance [28] . Thus, we next infected maternal antibody-protected suckling mice with rJ or rJ.S Q600Y viruses and monitored persistently infected mice for the development of clinical signs for 60 days post infection (p.i.). The presence of the highly immunogenic S598 Q600Y epitope did not protect mice from acute encephalitis ( Figure 3E ), perhaps because the presence of the improved S598 epitope was accompanied by a diminished response to S510 ( Figure 2D ). However, we found that among survivors (defined as survival past day 14 p.i.) there was a significant reduction in the incidence of clinical disease as well as in the development of CTL escape in S510 (Table 1) . Additionally, S598 Q600Y did not undergo CTL escape in mice persistently infected with rJ.S Q600Y , even though single nucleotide changes in the region of the S glycoprotein gene encoding the S598 determinant could potentially result in fifty-one CTL escape mutations. Thus, as expected, the ''improved'' S598 Q600Y epitope was protective in vivo in infected mice, likely because S598 Q600Yspecific CTL are present in higher numbers and exhibit higher functional avidity than the native S598-specific response. Since this lack of mutation at S598 Q600Y might reflect enhanced suppression of virus replication mediated by co-dominant CTL responses directed against S510 and S598 Q600Y , we developed a recombinant virus encoding S598 Q600Y in the context of a common S510 CTL escape mutation, S510 W513R (rJ.S W513R+Q600Y , Figure 2A ). The CTL response is predicted to be largely directed at S598 Q600Y in mice infected with this virus. The W513R mutation (position 4 substitution in S510 epitope, CSLRNGPHL) occurs in 13% of all CTL escape variants [19, 20, 22, 29, 30] , and completely abrogates native S510 CTL recognition [31] . We verified that virusspecific CTL responses were focused on S598 Q600Y in adult B6 mice infected with rJ.S W513R+Q600Y ( Figure 3F ). To examine the phenotype of the S598 Q600Y /S510 W513R double mutant, we infected antibody-protected suckling B6 mice with this virus and appropriate controls and monitored mice for survival ( Figure 3G ). As expected, 93.3% of mice infected with rJ survived the acute infection (day 0-14 p.i.). All mice infected with rJ.S W513R developed fatal encephalitis but, in marked contrast, 66.6% of mice infected with rJ.S W513R+Q600Y survived. We also found that survival correlated with virus clearance ( Figure 3H ). Relative to rJ-infected mice, replication was suppressed in mice infected with virus encoding the S598 Q600Y epitope and greatly elevated in mice infected with rJ.S W513R . In mice infected with rJ.S W513R+Q600Y , virus titers were intermediate between rJ and rJ.S W513R . Thus, the presence of the heteroclitic S598 Q600Y determinant contributed to suppression of virus replication and to increased survival, even when the highmagnitude, high-avidity CTL response to S510 was largely abrogated. Surprisingly, S598 Q600Y still did not undergo sequence diversification in mice that survived the rJ.S W513R+Q600Y infection ( Table 2) . This result was unexpected, as the majority of CTL in the rJ.S W513R+Q600Y -infected CNS specifically target the magnitude, unidirectional cross-reactivity. Representative dot plots demonstrating the frequency of epitope-specific CD8 T cells in a mouse infected with rJ (top panels) or rJ.S Q600Y (bottom panels). Numbers represent the frequency of epitope-specific CD8 T cells among total CD8 T cells recovered from the brains of mice 7 days p.i. (E) Summaries of the frequency (left panel) and absolute number (right panel) of epitope-specific CD8 T cells recovered from the brains of rJ and rJ.S Q600Y -infected mice 7 days p.i. Data shown in D represent mean6SEM for 4 independent experiments. doi:10.1371/journal.ppat.1000186.g002 S598 Q600Y determinant and exhibit high functional avidity ( Figure 3A-C) . One possible explanation for this result is that S598 is not as plastic as S510, even though both determinants are derived from a region of the spike gene that is hypervariable and even deleted in some strains of MHV [32] . To determine whether cells primed to S598 Q600Y that crossreact with the native S598 determinant were more protective in vivo than S598-primed cells, we vaccinated mice with bone marrow-derived dendritic cells (BMDC) alone, or BMDC pulsed with peptides corresponding to S598 or S598 Q600Y (Figure 4) . Seven days later, mice were challenged via intranasal inoculation of 4610 4 PFU of wild type, non-recombinant JHMV. Similar to results observed following rJ.S Q600Y infection (Figure 3) , CTL primed via DC-S598 Q600Y vaccination exhibited higher functional avidity when reacted against the native S598 determinant when compared to those arising after DC-S598 vaccination ( Figure 4A ). In other experiments, we examined the survival of mice vaccinated with each determinant, but we observed no significant differences between groups (data not shown), probably because mortality is largely CD4 T cell-mediated in adult mice with acute encephalitis [33, 34] . In terms of virus titers, vaccination with either the native or enhanced S598 determinants resulted in ,70-80% reduction in virus burden compared to mice that received un-pulsed BMDC ( Figure 4B ). When we examined the One-fourth of the pups in an individual litter were infected with each virus. At the indicated day p.i., brains were aseptically harvested, homogenized in sterile PBS and clarified by centrifugation. Supernatants were collected and infectious virus was titered on Hela-MHVR (Hela cells transfected with CEACAM1, the JHMV receptor [59] ) as previously described [27] . Symbols on graph represent individual mice assayed from multiple independent litters. The limit of detection (LOD) for the assay is 80 PFU/brain. frequency and numbers of virus-specific CTL in the CNS of these same mice, we detected markedly fewer S598-specific CTL in the CNS of mice that received BMDC pulsed with S598 Q600Y peptides ( Figure 4C) . Calculation of the product of virus titers and CTL numbers within individual mice, as an approximate measure of CTL potency, indicated that the S598-specific CTL in S598 Q600Ycoated BMDC vaccinees were ,6-7-fold more efficacious on a per cell basis ( Figure 4D ). Thus, S598-specific CTL induced by the S598 Q600Y determinant show similar enhancement in function compared to S598-primed cells, whether measured in vitro ( Figure 3C ,D) or in vivo ( Figure 4D ). While these studies clearly demonstrated that S598 Q600Y is heteroclitic, they did not provide a mechanism for the immune enhancement. To address this, we determined the crystal structures of the H-2K b /S598 (PDBid 2ZSV, Protein Data Bank Japan (http://www.pdbj.org/)) and H-2K b /S598 Q600Y (PDBid 2ZSW) complexes to 1.8 Å and 2.8 Å resolution respectively. The structure of H-2K b /S598 consists of two heterodimers in the asymmetric unit (r.m.s.d. of 0.18 Å for Ca atoms), with the S598 peptide clearly bound in the antigen binding cleft of the heavy chains (HC, Figure S3A ). The two peptide copies display a virtually identical configuration with root mean square deviation (rmsd) values of only 0.09 Å for all peptide atoms (0.05 Å for Ca atoms). The mode of S598 and S598 Q600Y binding within the Agbinding cleft is unambiguous, with the exception of Arg-1, whose side chain is partially disordered (Figure 5A,D; Figure S3C ). The S598 peptide adopts an extended conformation, with the side chains of Arg-1, Ile-4 and Asn-7 extending prominently out of the cleft ( Figure 5C ). Ala-6 is also largely solvent exposed with its side chain pointing towards the a2 helix, while the side chains of Cys (Aba)-2, Gln-3, Phe-5 and Ile-8 are buried within the cleft. While the cysteine analogue's side chain is not involved in any hydrophilic interactions, there are a number of suitably positioned hydrogen bonding partners (Glu-24, Tyr-45 and Asn-70), with which the original thiol side chain could potentially interact ( Figure S3B ; Table S2 ). In addition to main chain interactions across the length of the peptide, S598 is anchored to the MHC via the side chains of Gln-3, Phe-5 and Ile-8. Gln-3 forms hydrogen bonds with the Ser-99 and Gln-114 of the HC (Figure 5B ), while Phe-5 and Ile-8 are buried within hydrophobic pockets. In addition, the side chains of Phe-5 and Gln-3 pack against each other and between the aromatic rings of HC residues Tyr-159 and Phe-74, constraining the peptide's backbone conformation at those positions. The structure of H-2K b /S598 Q600Y consists of four heterodimers in the asymmetric unit (rmsd of 0.09 Å for heavy chain (HC) Ca's), the four copies of the peptide adopting virtually identical conformations (rmsd of 0.12 Å for all peptide atoms and 0.05 Å for Ca atoms). S598 Q600Y displays the same conformation as the S598 determinant (the average rmsd for peptide Ca atoms between the two structures is 0.24 Å ) and forms equivalent interactions with the MHC. The only prominent structural difference is observed at the mutated position (Q3RY3) ( Figure 5B,E, 6A,B) . In contrast to Gln-3, the Tyr-3 side chain is oriented towards the a-2 helix rather Figure 5E ). The side chain of Tyr-3 also forms a number of close contacts with the residues in its immediate environment, specifically HC residues Gln-114, Arg-155, Leu-156 and Tyr-159, as well as intramolecular interactions with Phe-5 (Table S2 ). In addition to Tyr-3, deviations of potential significance (.0.5 Å ) in the peptide structures that are attributable to the Q660Y mutation are observed at Ile-4 and Phe-5. Overall the S598 determinant displays greater complementarity for the antigen binding cleft of H-2K b in its N-terminal region, with few stabilizing interactions observed between the HC and positions 6 and 7. (Figure S3B, D; Table S2 ). Nevertheless, a pocket is observed between the a-2 helix and the peptide near position 3 in the index structure that is filled by the steric bulk of Tyr-3 in the Q600Y structure ( Figure 5C,F) . This increase in surface complementarity and the greater number of observed interactions resulting from the Q3RY3 mutation would account for the enhanced thermostability (,10uC) measured for the H-2K b /S598 Q600Y -Aba complex by circular dichroism (Figure 1 ). This increase in complementarity of the MHC for S598 Q600Y and the greater stability of the resulting complex were predicted from comparisons of the WT determinant complex with existing structures of H-2K b bound with peptides similar to S598 and possessing the consensus tyrosine anchor residue at position 3, as well as another aromatic residue at position 5. With respect to the HC, the a-1 and a-2 domains of the two structures superimpose well with an rmsd of 0.37 Å for Ca atoms (residues 1-176). Nevertheless, significant deviations (.0.5 Å ) are observed between the two structures at a number of positions in the region of the antigen binding cleft. In particular, changes in the conformation of Ser-99, Gln-114, and Gly-151-Glu-154 are associated with the bound peptides. Changes in the side chain conformations of Ser-99 and Gln-114 in the Q600Y complex structure are consistent with the loss of hydrogen bonding interactions with position 3. In the wild type peptide structure, Glu-152 forms a salt bridge with Arg-155, the guanadinium group of which also forms a hydrogen bond with the main chain carbonyl group of the peptide's Ile-4. While these interactions are maintained in the Q600Y complex, the side chain of Glu-152 displays a conformational shift consistent with the formation of a hydrogen bond with the peptide's Tyr-3 ( Figure 6 ). Interestingly the region of the a-2 helix around Glu-152 (Gly-151 -Glu-154) also displays significant main chain and side chain deviations between the two structures (rmsd of 0.77 Å for Ca atoms; Figure 6 ). Thus, consistent with the functional analyses, the structure of the heteroclitic variant of the S598 epitope displays relatively small changes in conformation yet the combination of these subtle changes in the TcR accessible residues and the structural landscape of the MHCp in addition to the enhanced stability of the complex lead to more efficacious CTL responses. We describe the identification of a heteroclitic determinant that enhances recognition by virus-specific CD8 T cells, and use the crystal structure of the new determinant (S598 Q600Y ) to provide a basis why it elicits an enhanced CTL response. Comparison of S598 to the consensus binding motif suggested a suboptimal interaction with the H-2K b molecule at the secondary anchor position (Gln-3) and, consequently, an approach to enhance the immunogenicity of the determinant. Replacement of the Gln-3 with Tyr-3 (Q600Y) did, indeed, result in an determinant with increased thermostability without diminishing the CD8 T cell response. Most strikingly, the Q600Y change resulted in subtle changes in the conformation of the a-2 helix locally in the vicinity of Glu-152. These subtle changes are likely critical for the enhanced TcR recognition that we detected. S598 Q600Y elicited a response with higher functional avidity to both the cognate and native determinants than S598, and this was not reflected in differential Vb usage. The T cell response to S598 in rJ-infected mice is very diverse [35] . As assessed by Vb usage, the response to S598 and S598 Q600Y in rJ.S598 Q600Y was similarly diverse with only modest differences noted when cells from mice primed by S598 and S598 Q600Y were compared. While we cannot exclude the possibility that cross-reacting S598-specific cells primed by S598 Q600Y are biased for Vb chains not analyzed in this study, it is more likely that the fine specificity of the complementarity-determining region 3 (CDR3) determines their greater affinity for H-2K b /S598. Although an increase in stability of the MHC class I/peptide complex is not generally expected to enhance TcR affinity for the complex, similar results have been observed in mice immunized with analogues to a common tumor antigen [36] . One unexpected result was that S598 exhibited both low MHC class I and low TcR avidity. Previous studies showed that this determinant exhibited low functional avidity [21] , but it was not known whether this reflected low binding to the MHC class I or to the TcR. Assuming that low affinity for MHC class I results in a low effective concentration of H-2K b /S598 complex on the cell surface, the responding T cells should be high avidity, based, primarily, on in vitro studies [37] . While the relationship between level of surface peptide and avidity of the responding T cells generated in vivo is not as clearcut [38] , there is no obvious explanation for how a peptide with low MHC class I binding also elicits a low avidity T cell response. This selection of only a subset of CD8 T cells capable of responding to S598 may partly explain why S598-primed cells do not recognize the Q600Y determinant. The biological significance of the heteroclitic Q600Y determinant was shown by its ability to protect JHMV-infected mice from CTL escape at S510 and to diminish clinical disease. This was important to demonstrate because other studies, using tumor models, have shown that immunogenicity and tumor recognition are not necessarily concordant [39] . Mutations resulting in CTL escape occur most commonly in determinants that are exposed to high selective pressure [40] [41] [42] and outgrowth of CTL escape variants is efficiently suppressed by effective and rapid virus clearance [43] [44] [45] [46] , as occurs in mice infected with rJ.S Q600Y . Thus, even though CTL escape is not detected in normal mice infected with LCMV or influenza, escape does occur when mice transgenic for a single LCMV-specific TcR are infected with high amounts of virus [44, 47] . Under these conditions, the immune response is highly focused on a single CD8 T cell determinant and virus replication continues for extended periods of time, facilitating mutation at the targeted determinant. In mice infected with wild type JHMV, the CTL response is functionally focused on S510 [21] ; the Q600Y substitution effectively prevents CTL escape at either S510 or S598 Q600Y by the induction of a second high avidity CTL response. Mutations in S598 Q600Y were not detected even when the CTL response was directed primarily at this determinant (e.g. mice infected with rJ.S W513R + Q600Y ). Consistent with this inability to readily tolerate mutations, we were unable to generate recombinant virus mutated in position Ile-4 (I601D,E,K,R,T) and recombinant virus mutated at Phe-5 (F602A) was highly attenuated (data not shown). The combination of induction of high avidity CTL and inability to tolerate mutation without adversely affecting virus fitness make S598 Q600Y an ideal target for the anti-JHMV CTL response. Further, the ala scanning results suggest that S598 Q600Y -specific CTLs may more readily tolerate changes in the H-2K b /peptide complex, and this plasticity would also minimize the likelihood of CTL escape. In contrast, we have previously shown that introduction of the LCMV-specific GP33 determinant, which also elicits CTL with high functional avidity, into JHMV greatly diminished clinical disease but did not prevent CTL escape [22] . The GP33 determinant was introduced at a site in the genome that tolerated mutation and deletion and intact determinant was no longer present in virus by day 20 p.i. Collectively these results suggest that a response to a second determinant that elicits CTL exhibiting high functional avidity at early times p.i. results in enhanced suppression of virus replication, but its presence throughout the infection is required to protect against CTL escape. In conclusion, we have demonstrated that crystal structures are useful in gaining an understanding of the basis of heteroclitic epitopes and can also prove valuable in guiding the rational design of ''better'' CTL epitopes. In our mouse system, immunization with the heteroclitic determinant resulted in the generation of unique populations of CTL that respond with high functional avidity to an otherwise modestly immunogenic viral epitope. The generation of unique populations of CTL that respond with high functional avidity to weakly immunogenic epitopes will be useful for the treatment and prevention of human infectious diseases. Our proposed structure-guided approach has direct application to HIV, HCV and other chronic infections in which virus persistence and CTL escape occurs. By modulating T cell immunity through prophylactic or therapeutic peptide-based vaccination, virus titers may be reduced and CTL escape and other consequences of viral persistence circumvented. Specific pathogen-free B6 and BALB/c mice were obtained from National Cancer Institute (Bethesda, MD). To obtain infected mice in which CTL escape at S510 was detected, suckling B6 mice were infected intranasally with 2-4610 4 PFU of recombinant JHMV at 10 days of age and nursed by dams that were immunized with JHMV, as described previously [27] . For experiments comparing multiple JHMV variants, each litter served as an internal control: equal numbers of pups were infected with rJ and one to three recombinant variant viruses, depending on litter size. All animal studies were approved by the University of Iowa Animal Care and Use Committee. Mononuclear cells were harvested from the brains of acutely ill mice 7 days p.i. and analyzed for expression of IFN-c by an intracellular cytokine staining protocol as previously described [28] . Unless otherwise noted, peptides corresponding to epitopes were used at a final concentration of 1 mM and cells were stimulated in the presence of 500 mM TCEP (Sigma, St. Louis, MO). Cells were analyzed using a FACScan flow cytometer (BD Biosciences, Mountain View, CA). Data sets were analyzed using FlowJo software (Tree Star, Inc, Ashland, OR). All antibodies and reagents were purchased from BD Pharmingen (San Diego, CA). Recombinant wild-type and S510 and S598 variants of JHMV were generated as previously described [48, 49] . Briefly, overlapping extension polymerase chain reaction (PCR) was used to generate the S598 Q600Y and S510 W513R variants. Primers that overlapped the glutamine at residue 600 of the spike glycoprotein were (59 to 39) Q600Y fwd, ATGATCGCTGCTATATTTTTGCTAACATAT-TG; Q600Y rev, AATATGTTAGCAAAAATATAGCAGCGAT-CAT. Primers that overlapped the tryptophan at residue 513 were (59 to 39) W513R fwd, GTGAGTGTTCTCTTCGGAATGGGC-CCCATTTGCGCTCGGC; W513R rev, AGCGCAAATGG-GGCCCATTCCGAAGAGAACACTCAC. The outer primers for each targeted change were fwd, TGTTGATTGCGCCAG-CAGCTACATTAG; and rev, ACCTACGGATTGAACGCTAT-CATTGAC. Underlined nucleotides correspond to the nucleotides encoding the Gln to Tyr and Trp to Arg substitutions within S598 and S510, respectively. Recombinant viruses encoding the variant epitope(s) were selected, propagated and titered as previously described [49] . At least two independent isolates of each recombinant virus were analyzed. Virus was inoculated onto confluent 17Cl-1 monolayers in a 12well plate at a multiplicity of infection (MOI) of 1.0. Groups of cells were harvested at the indicated time points and total virus (cellassociated and cell-free) was titered as previously described [27] . Equal PFU (2-4610 4 ) of rJ and rJ.S Q600Y were combined and inoculated intranasally into 5-week-old B6 and BALB/c mice. Total RNA was harvested from the brains of mice 7 days p.i. and the relative representation of WT (wild type) versus variant template was determined by RT-PCR and sequencing. This assay can specifically detect a given species of template when that species comprises at least 20% of a heterogeneous pool [50] . Primers used were : forward, AACCCCTCGTCTTGGAATAGGAGG-TATGG; and reverse, CCTACGGATTGAACGCTATCATT-GACTAAC. PCR products were sequenced directly by the University of Iowa DNA Core. For alanine scanning, cells were stimulated ex vivo with the indicated concentration of native or variant peptide and stained for CD8 and IFN-c as described above. Data were normalized to the frequency of cells that reacted to the unmodified S598 or S598 Q600Y peptides. For functional avidity determination, mononuclear cells were harvested from the brains of rJ or rJ.S598 Q600Y -infected mice 7 days p.i. and stimulated ex vivo in the presence of EL-4 cells pulsed with tenfold dilutions of peptide corresponding to native S598 or S598 Q600Y epitopes. After 5.5 hours, cells were stained for intracellular IFN-c as described above. For each epitope-specific population, data were normalized to the frequency of antigenspecific CTL detected using the highest titration of peptide (1 mM). CD spectra were measured on a Jasco 810 spectropolarimeter using a thermostatically controlled cuvette at temperatures between 30 and 90uC. Far-UV spectra were collected and analyzed as described [23] . Cells were harvested from the CNS of mice 7 days p.i. and stimulated ex vivo with S598 or S598 Q600Y peptides. Cell aliquots were subsequently stained for CD8 (PE-Cy7-anti-CD8a) and Vb (FITC-anti-Vb2, 3, 4, 5.1/5.2, 6, 7, 8, 9, 10 b , 11, 12, 13 or 14, BD-Pharmingen) followed by intracellular staining for IFN-c (PEanti-IFN-c). Data were collected using a Becton Dickinson LSR II instrument at the University of Iowa Flow Cytometry Facility. Data are expressed as the proportion of antigen-specific CD8 T cells that express each Vb chain. Total RNA was purified with TRIzol (Invitrogen, Carlsbad, CA) from the CNS of mice. The 1055 base pair region of the spike glycoprotein encompassing both S510 and S598 was amplified by RT-PCR and sequenced directly as previously described [19] . Bone marrow-derived DC were prepared, pulsed with peptides and injected into mice as previously described [51] . Briefly, 5610 5 LPS-matured DC were left uncoated, coated with S598 or S598 Q600Y peptides and injected via tail vein into groups of 4week-old mice. Seven days following DC-vaccination, mice were infected intranasally with 4610 4 PFU of JHMV. Seven days following virus infection, brains were harvested from mice and the frequencies of epitope-specific CD8 T cells were determined by ex vivo stimulation and intracellular cytokine staining as described above. In independent studies, DC-S598 and DC-S598 Q600Y priming was verified by harvesting spleens from several mice seven days following DC-vaccination. Crystal structure of H-2K b /S598 and H-2K b /S598 Q600Y complexes Crystals of H-2K b S598-Aba were grown at 21uC in 0.1 M cacodylate pH 6.6, 16% PEG (polyethylene glycol) 8,000, 0.2 M Ca(OAc) 2 , using a protein concentration of 9 mg/ml. Crystals were cryoprotected by stepwise equilibration against mother liquor containing 5, 10 and 15% glycerol and flash frozen by placing in a nitrogen stream. A 1.8 Å resolution dataset was collected on an inhouse X-ray source. Crystals of H-2K b S598 Q600Y -Aba were grown at 21uC in 0.1 M cacodylate pH 6.5, 13% PEG 8,000, 0.2 M Ca(OAc) 2 , using a protein concentration of 6 mg/ml. Crystals were cryoprotected by gradual equilibration against mother liquor containing 20% PEG 8,000 and 5% glycerol before flash freezing. A 2.8 Å resolution dataset was collected on an in-house source. The WT data were integrated in MOSFLM [52] and scaled/merged using SCALA [53] . The Q600Y variant data were processed using HKL2000. Both structures were solved by molecular replacement in PHASER [54] , against previously solved H-2K b complexes (PDBid's: 1G7Q and 1RJY, respectively). The resultant models underwent iterative cycles of refinement in PHENIX [55] and, REFMAC5 [56] (restrained and TLS refinement) followed by model building/ correction in COOT [57] . The solvent structures were built using ARP/wARP [58] and COOT. A summary of the processing and refinement statistics is presented in Table S1 . Statistical significance was determined by nonpaired, two-sided Student's t test or chi-squared test, where indicated. (A) Total mononuclear cells were harvested from mice infected with rJ or rJ.S Q600Y and stimulated ex vivo with S598 and S598 Q600Y peptides, respectively. Following stimulation, aliquots of cells were surface stained for CD8 and the indicated Vb chain followed by intracellular cytokine staining for IFN-c. (B) CNSderived cells from rJ.S Q600Y -infected mice were stimulated ex vivo with peptides corresponding to native S598 or S598 Q600Y epitopes. Following stimulation, cells were analyzed as described for (A). Data represent the fraction of IFN-c+CD8+ T cells expressing each Vb chain and are derived from cells pooled from 2-3 individual mice. (C) Alanine scanning of the native S598 determinant. CNS-derived mononuclear cells were recovered from rJ-infected mice 7 days p.i. and stimulated ex vivo in the presence of 500 mM TCEP and 1 mM of the indicated peptide then stained for CD8 and intracellular IFN-c. Data are normalized to the frequency of epitope-specific cells detected when stimulated with the native S598 determinant. (D) Alanine scanning of the S598 Q600Y determinant. Cells were harvested and tested as described for B except in this case the cells originated from the rJ.S Q600Y -infected CNS and were stimulated with 10 nM S598 Q600Y peptide. (E) Alanine scanning of the S598 determinant recognized by S598 Q600Y -primed, cross-reactive CTL. As in C, but cells were stimulated with 150 nM S598 peptide. For B-D, concentrations of peptide equivalent to 106 that required for half maximal stimulation were used; data are mean6SEM from four independent experiments. Note that the differential responses to RCAIFANI in panels C and D reflect the differing amounts of peptide used in the two assays. Found at: doi:10.1371/journal.ppat.1000186.s002 (0.76 MB TIF) Figure S3 Refined structures of WT and Q600Y S598-Aba bound to H-2K b . (A) View of the H-2K b antigen binding cleft from above. The HC is shown as a cartoon representation and coloured slate. The peptide is in stick format with carbon atoms coloured yellow. The unbiased F o -F c map density for the peptide contoured at 2.5 s is shown as a magenta mesh. (B) The same view as in A displaying key interactions (dashed lines) between H-2K b and S598-Aba. Selected residues of the HC are drawn in stick format (slate carbon atoms) and ordered water molecules are shown as red spheres. Peptide residues are labelled in italics. C and D, Equivalencies to A and B, respectively, for the H-2K b / S598 Q600Y -Aba structure. In these panels the HC is drawn in green and the peptide in cyan.

HIV-Specific T-Cells Accumulate in the Liver in HCV/HIV Co-Infection

BACKGROUND AND AIMS: Hepatitis C Virus (HCV)-related liver disease progresses more rapidly in individuals co-infected with Human Immunodeficiency Virus-1 (HIV), although the underlying immunologic mechanisms are unknown. We examined whether HIV-specific T-cells are identified in the liver of HCV/HIV co-infected individuals and promote liver inflammation through bystander immune responses. METHODS: Ex-vivo intra-hepatic lymphocytes from HCV mono-infected and HCV/HIV co-infected individuals were assessed for immune responses to HIV and HCV antigens by polychromatic flow cytometry. RESULTS: HCV/HIV liver biopsies had similar frequencies of lymphocytes but lower percentages of CD4(+) T-cells compared to HCV biopsies. In co-infection, intra-hepatic HIV-specific CD8(+) and CD4(+) T-cells producing IFN-γ and TNF-α were detected and were comparable in frequency to those that were HCV-specific. In co-infected individuals, viral-specific CD8(+) T-cells produced more of the fibrogenic cytokine, TNF-α. In both mono- and co-infected individuals, intra-hepatic HCV-specific T-cells were poorly functional compared to HIV-specific T-cells. In co-infection, HAART was not associated with a reconstitution of intra-hepatic CD4(+) T-cells and was associated with reduction in both HIV and HCV-specific intra-hepatic cytokine responses. CONCLUSION: The accumulation of functional HIV-specific T-cells in the liver during HCV/HIV co-infection may represent a bystander role for HIV in inducing faster progression of liver disease.

Approximately 25% of Human Immunodeficiency Virus-1 (HIV) infected individuals are also infected with Hepatitis C Virus (HCV) [1] . HIV adversely affects each stage of the natural history of HCV infection. Fewer individuals recover spontaneously from HCV infection when also infected with HIV [2] . Among those with persistent HCV infection, HIV co-infection is associated with higher HCV viremia and more rapid progression to cirrhosis and hepatocellular carcinoma [3] . A recent meta-analysis showed that HIV co-infection increased the risk of histological hepatic cirrhosis by two-fold and clinically decompensated liver disease by six-fold [4] . In addition, HCV co-infection is associated with increased incidence of HAART (highly active antiretroviral therapy) related liver injury [5] . The mechanisms for hepatic damage in HCV/HIV co-infection are poorly defined. Although intra-hepatic T-cell immune responses are necessary for HCV clearance, they have also been shown to play a central role in mediating hepatocellular injury by direct cytotoxicity or indirectly by releasing cytokines. In this regard, IFN-c has been shown to be anti-fibrogenic, whereas, TNF-a activates hepatic stellate cells, which induce fibrosis, and likely contributes to progression to cirrhosis [6, 7] . Potent and broad CD4 + and CD8 + T-cell immunity are important for virologic control in both HCV and HIV viral infections. Ex-vivo HCV-specific CD8 + T-cell responses in peripheral blood mono-nuclear cells (PBMCs) from mono-infected individuals are generally weak [8] . Although, peripheral HCVspecific CD4 + and CD8 + T-cell responses are somewhat weaker in HCV/HIV co-infected individuals [9] , similar frequencies of intra-hepatic HCV-specific responses appear to be obtained in HCV versus HCV/HIV co-infection [10, 11] . However, ex-vivo HIV-specific CD8 + T-cell responses in PBMCs from HIV monoinfected individuals are about one log higher than ex-vivo HCVspecific responses in HCV mono-infection. In addition, impairment in cellular immune responses to HCV compared to HIV has been shown in HCV/HIV co-infection [12] . HIV-specific CD8 + T-cells are easily detectable in blood of untreated HIV infected individuals [13] . Such high frequencies of HIV-specific T-cells circulating in peripheral blood led us to question whether these cells could also migrate to the liver in HCV/HIV co-infection and through bystander responses add to the inflammation induced by HCV-specific T-cells. HCV mono-infected and HCV/HIV co-infected individuals who required liver biopsies for work up of liver disease were recruited for the study (see Results and Table 1 ). All study participants provided informed, written consent and the study protocol was approved by the research ethics board at the University of Toronto and St. Michael's Hospital. Both blood and liver biopsy samples were received from each participant. Liver biopsy samples were washed in RPMI-1640 to remove contaminating blood lymphocytes, manually homogenized with a plastic plunger, and treated with DNase (0.002%, Sigma) and collagenase IV (0.02%, Sigma) for 30 minutes, stirring at 37uC. The digested cell suspension was filtered through a 70 mm strainer, washed and re-suspended in R-10 medium (10% fetal calf serum). In order to identify candidate epitope-specific responses to be detected in ex-vivo liver samples, we first mapped antigen-specific T-cell responses in blood against the entire HIV-1 clade-B and HCV-1a proteome using the matrix approach by IFN-c ELI-SPOT assay as described previously [14] . Mapped peptides were then pooled to evaluate hepatic responses. In order to address the possibility that differing epitopes were only targeted in the liver, we also used four peptide pools that previously were shown to target a majority of responses. These pools spanned HIV-Gag and HCV-NS3, HCV-NS4 and HCV-Core protein (2 mg/peptide/ml, from National Institute of Health Reagent Program). Of the HCV pools, the pool that gave the strongest ELISPOT response in PBMCs was used for hepatic cell stimulation (see below). All the extracted cells from each liver biopsy were split in three wells and stimulated on the same day as PBMCs. 1610 6 PBMCs and liver-isolated cells were stimulated with either DMSO, HIV or HCV peptide pools as described previously [14] . HIV pools consisted of peptides that were screened by the matrix approach in that individual plus the HIV-Gag pool. Likewise, HCV pools consisted of mapped peptides plus an HCV pool that gave the strongest response in PBMCs. CD107a antibody (PE-Cy5, BD Pharmingen) was added at the time of stimulation. The following antibodies were used for staining: CD8-PE Texas-Red (Beckman Coulter), CD4-Pacific Blue (e-Bioscience), CD3-APCCY7, IFNc-FITC, TNFa-PECY7, IL2-APC, MIP-1b-PE (BD Bioscience), PD-1 FITC (Biolegend) and dead cell stain Aqua (Invitrogen, Molecular Probes). Cells were analyzed on a multi-color FACSAria flow cytometer (BD Biosciences). For Blood samples between 500,000 to 1,000,000 total events and for liver biopsy samples between 50,000 to 200,000 total events were collected. Data analysis was performed using FlowJo version 8.6 (Treestar Inc., San Carlos, CA). Polychromatic FlowJo data were analyzed with PESTLE software, and pie-chart graphs were generated using SPICE software (obtained from M. Roederer, National Institutes of Health, Bethesda MD). Multi-parameter analysis of HIV-Gag: 77-85 (SLYNTVATL: SL9) specific CD8 + T-cells was conducted in both blood and liver of co-infected individuals initially identified with a positive ELISPOT response to the 15-mer HIV peptide including SL9 epitope, using the corresponding tetramer (iTAg MHC Class-I tetramer, Beckman Coulter). Tetramer staining was performed prior to peptide stimulation, at room temperature for 20 minutes. Tetramer stained cells were then washed and stimulated with 10 mg/ml of SL9 peptide followed by ICS staining as mentioned above. Additional HIV, HCV and CMV-specific pentamer staining (Pro5 MHC class I Pentamers -Proimmune) was conducted followed by PD-1 staining. Data were analyzed by performing two-tailed non-parametric Mann-Whitney test using GraphPad Prism version 4.00. P-values#0.05 were considered significant. Three groups of individuals were studied as depicted in Table 1 ; HCV mono-infected (n = 6), HCV/HIV co-infected who were not receiving HAART (n = 8) and HCV/HIV co-infected who were receiving HAART for greater than one year at the time of evaluation (n = 12). All individuals never received prior treatment for HCV and underwent liver biopsies for staging and evaluation for pegylated-interferon/ribavirin treatment. HCV/HIV co-infected individuals had higher HCV viral loads. On average, CD4 T-cell counts of HIV infected individuals were .400/ml in both groups. Of note, the mean hepatic fibrosis scores were higher in the HAART treated and mono-infected groups in this cohort, indicating that individuals in these groups had more advanced disease at the time of biopsy in this study. Although similar frequencies of intra-hepatic lymphocytes were obtained in dual versus mono-infection, HAART-treated individuals showed a trend towards greater percentages of lymphocytes in their biopsies ( Figure 1a ). The percentage of intra-hepatic CD4 + T-cells was significantly reduced in dual infection [31.6%613.8 for HCV vs 6.5%62.9, for HCV/HIV therapy naïve, p,0.01], and was not associated with any improvement in HAART-treated individuals, as previously shown in the gut [15] (Figure 1b) . However, compared to HCV mono-infected individuals the percentage of intra-hepatic CD8 + T-cells was higher in both co-infected groups [33.8%65.5% for HCV vs 67.3%615.5% for HCV/HIV therapy naïve vs 59.5615.1% for HCV/HIV on HAART, p,0.01] (Figure 1b ). To determine the presence of intra-hepatic viral specific T-cell responses, liver isolated cells were stimulated with HCV and HIV peptide pools. Summary data of viral specific responses are depicted in Figure 2 . In response to stimulation with HIV peptide pool, untreated co-infected individuals showed significantly higher frequencies of intra-hepatic CD4 + T-cells producing IFN-c, compared to HCV mono-infected [0.1660.05% vs 0.0260.01%, p,0.05], and HAART-treated co-infected individuals [0.1660.05% vs 0.0360.05%, p,0.05] (Figure 2a ). Untreated co-infected individuals showed a trend towards lower frequencies of intra-hepatic IFN-c producing CD4 + T-cells in response to HCV peptides. Surprisingly, HAART-treated co-infected individuals had significantly reduced HCV-specific IFN-c producing CD4 + T-cells when compared to untreated co-infected individuals [0.0260.01% vs 0.4660.11%, respectively, p,0.01] (figure 2a). Therapy naïve co-infected subjects had greater IFN-c producing CD8 + T-cells in response to HIV peptides compared to HCV mono-infected individuals [1.3960.37% vs 0.0260.0%, p,0.05], and HAART was associated with a significant reduction in the frequencies of these cells [1.3960.37% vs 0.3060.26%, p,0.05] (figure 2b). Although there was a trend for enhanced intra-hepatic CD8 + T-cells producing IFN-c in response to HCV peptides in therapy-naive co-infection compared to HCV mono-infection, this was not found to be statistically significant. HAART on the other hand, was associated with a significant reduction in HCV-specific, intra-hepatic CD8 + T-cells producing IFN-c [1.360.37% vs 0.0360.01%, p,0.05] (figure 2b). Similarly, co-infected individuals had significantly greater intrahepatic TNF-a expressing CD4 + T-cells after HIV peptide stimulation compared to HCV mono-infected [0.260.05 vs 0.0260.01, p,0.01], although HAART had no significant effect on their frequencies (Figure 2c ). HCV mono-infected individuals showed significantly higher frequencies of HCV-specific TNF-a producing CD4 + T-cells compared to HAART-treated co-infected individuals [0.9160.25% for HCV vs 0.2360.20 for HCV/HIV on HAART, p,0.01] (figure 2c), but did not show significant differences with the untreated co-infected group. The therapy-naïve co-infected group showed significantly higher frequencies of intra-hepatic TNF-a producing CD8 + Tcells in response to both HIV-1 and HCV antigens. Both types of To study the functional profile of virus-specific T-cells in HCV/ HIV co-infection, simultaneous expression of 5 distinct CD8 + T-cell markers were analyzed in 3 individuals within each cohort using a previously developed multicolor flow cytometry method [16] . Expression levels of degranulation marker CD107a and cytokines IFN-c, TNF-a and IL-2, as well as the chemokine MIP-1b were simultaneously measured in response to HCV or HIV peptides in both blood and liver of each individual. Figure 3 depicts a representative multi-parameter analysis of CD8 + T-cell responses in liver and blood of a therapy-naïve, co-infected subject in response to HIV and HCV peptide pools. These data indicate that both HCV and HIV specific CD8 + T-cells expressing one or more functions are detectable in the liver and blood. Compared to blood, the frequency of HIV-specific T-cells producing CD107a and IL-2 was shown to be significantly higher in the liver of therapy-naïve, co-infected individuals. Consistent with previously reported data [10] , HCV-specific responses were compartmentalized to the liver and stronger than peripheral HCV-specific responses (Figure 3c ). Recognition of functional CTL, specific for the HLA-A0201-restricted HIV-SL9 epitope in HCV/HIV co-infected liver To determine if T-cells specific for an HIV immuno-dominant epitope are present in HCV/HIV co-infected liver, we quantified CD8 + T-cells specific for HLA-A*0201-restricted SLYNTVATL (SL9) epitope in the liver of individuals with positive SL9 responses in their blood. We identified 3 co-infected HLA-A*0201 individuals, among them one showed a response to the SL9 epitope of HIV-Gag antigen. Figure 4 shows the multi-parameter analysis of tetramer positive CD8 + T-cells in blood and liver of this therapy-naïve, co-infected individual. The tetramer cytokine response pattern was shown to be different in the liver compared to blood of the same individual, with diminished intra-hepatic tetramer-specific IFN-c responses and an increase in both CD107a and TNF-a responses, with the majority of SL9 tetramer positive cells expressing these two markers. We also included CMV as a non-hepatotropic control virus in our liver analysis. Using a pool of HLA-A*0201-and HLA-A*2402-restricted, CMV-specific pentamers, we did not detect any CMV-specific CD8 + T-cells in HCV/HIV co-infected liver, although we could readily detect them in blood (Figure 4c ). Using the panel of markers TNF-a, IFN-c, MIP-1b, IL-2, and CD107a, we characterized the ability of CD8 + T-cells to simultaneously exert these 'functions' in response to both HCV and HIV peptides. Figure 5 depicts a representative functional profile of virus specific CD8 + T-cells in blood and liver during HCV/HIV co-infection (Fig. 5a) and HCV mono-infection (Fig. 5b) . Analysis of co-infected subjects demonstrates a very limited functional hierarchy of HCV-specific T cells in the blood, with majority of T-cells producing one function. HIV-specific Tcells from blood had a more expanded functional hierarchy. In accordance with previous reports on HIV mono-infected individuals [16] , cells expressing all 5 functions were absent in the blood of co-infected subjects, mainly due to lack of IL-2 production. In co-infection, intra-hepatic CD8 + T-cells responding to HCV peptides were within the single-responding and 2+ populations. Intra-hepatic CD8 + T-cell responses to HIV peptides produced a larger spectrum of responses. CD107a responding cells were represented in nearly all of the different HIV-specific populations in the liver of co-infected individuals. As expected, HCV-specific responses in the blood of HCV mono-infected subjects were mainly single functional. The profile of HCV-specific responses in the liver of HCV mono-infected individuals showed the appearance of a very small population of 4+ responding cells in the liver. We and others have considered a cutoff point of more than 2 simultaneously expressed markers to demonstrate poly-functional characteristic of responding T-cells [16, 17] . Figure 5c shows a comparison between average frequency of intra-hepatic viralspecific responses within the pool of CD8 + T-cell populations expressing 2 markers or less, and those within the pool of populations expressing more than 2 markers simultaneously. For both HCV and HIV-specific CD8 + T-cells the majority of responses had two or less functions. However, intra-hepatic HIV-specific responses demonstrated more poly-functionality, compared to HCV-specific responses either within co-infected or mono-infected individuals [0.0560.01 vs 0.00760.00, p,0.05; HIV-specific responses vs HCV-specific responses in HCV/HIV co-infected group]; [0.0560.01 vs 0.0160.00, p,0.05; HIVspecific responses in HCV/HIV co-infected group vs HCVspecific responses in HCV mono-infected group]. In summary, although viral-specific T-cells, simultaneously expressing all 5 measured markers were rarely found in the liver, intra-hepatic HIV-specific T-cells showed greater functional capacity when compared to those being HCV-specific. Based on the recently highlighted role of PD-1 contributing to the dysfunction of T-cells in chronic viral infections, we also determined whether HIV and HCV-specific intra-hepatic T-cells differ in the degree of PD-1 expression. In an HCV/HIV coinfected liver, we found that 100% of intra-hepatic HCV-specific CD8 + T-cells were PD-1 positive, compared to 48.8% of those cells that were HIV-specific (Figure 5d ). This is the first study to demonstrate the presence of HIVspecific T-cells within the liver of HCV/HIV co-infected individuals. The finding of HIV-specific T-cells within liver of co-infected individuals may not altogether be surprising, given the high frequencies of HIV-specific CD8 + T-cells found in the peripheral blood in untreated HIV infection. Nevertheless, it is surprising to find functional T-cells of such viral specificities to be accumulating in liver. In contrast, we could not detect CMVspecific T-cells in co-infected liver despite their abundance in blood indicating that different viruses target T-cells to the liver. Recent studies have demonstrated that systemic viral infections may recruit viral-specific T-cells to the liver. The significance of non-hepatotropic viral-specific T-cells that are found in liver is unclear. It has been postulated that the liver can non-selectively trap activated T-cells during any infection, and thus act as a 'sink' or 'graveyard' [18] , however it is unclear whether these cells are rendered anergic while traveling in the liver or contribute to inflammation and damage as a result of bystander activation. Of note, is that hepatitis has been observed in measles [19] , SARS [20] and in 20% of individuals with acute HIV infection [21] . Polakos et. al. [22] found that some individuals infected with influenza-A develop transaminitis and showed in a murine influenza model that influenza-specific CD8 + T-cells migrate to Figure 3 . Polychromatic FACS analysis of viral-specific T cells in HCV/HIV co-infection. Shown are representative FACS data of the HIV and HCV specific multi-parameter CD8 + T-cell responses from (a) liver and (b) blood of subject OM 405, a therapy-naïve HCV/HIV co-infected individual, after in vitro stimulation using pool of HIV and HCV peptides. Initial gating on forward scatter area (FSC-A) versus height (FSC-H) was used to remove doublets. The events were further gated on forward scatter (FSC) versus the dead cell marker to remove dead cells. Lymphocytes were gated on the remaining live cells on a FSC versus SSC plot. Gates on CD3 + /CD8 + cells were then generated. All responses are gated on a CD3 + /CD8 + population and presented against TNF-a on the x-axis. Figure (c) shows a comparison of the frequency of HIV and HCV-specific CD8 + T-cells in the liver and blood of therapy-naïve, co-infected individuals. All intra-hepatic HCV-specific responses are significantly stronger than peripheral HCV-specific responses. doi:10.1371/journal.pone.0003454.g003 Non-hepatotropic viruses such as HIV, CMV and EBV, in general do not induce chronic hepatitis, thus, it is possible that the coexistence of hepatotropic viruses may alter the hepatic environment to allow recruitment of activated T-cells non-specifically. This could be due to an up-regulation of integrins such as ICAM-1 and VCAM-1 in hepatic sinusoids as previously shown during HCV infection [23] that could enhance T-cell recruitment. In this regard, Spangenberg et. al. [24] demonstrated the presence of influenza-specific T-cells in about 50% of liver biopsies from HCV mono-infected individuals. There are several lines of evidence demonstrating that the liver efficiently clears many foreign pathogens, including RNA viruses. It is shown that liver is a major organ for clearing Simian Immunodeficiency Virus in rhesus monkeys [25] . There is also evidence for the detection of HIV RNA in the liver of HIV infected individuals [26] . These findings support the identification of HIV-specific T-cells in the liver. In HCV/HIV co-infection, it is possible that intra-hepatic HCV-specific CD4 + T-cells become infected with HIV and recruit HIV-specific immune responses to this site. Evidence for these potential mechanisms will need further analysis on liver biopsies of co-infected individuals. Our analysis of liver biopsies from HCV/HIV co-infected individuals not only demonstrate that HIV-specific T-cells producing IFN-c and TNF-a are detected in the liver, but also exhibit comparable frequencies of responses to those that are HCV-specific. This observation may explain the added contribution of HIV-specific immune responses to the ongoing intrahepatic damage induced by HCV-specific T-cell responses that are inefficient in clearing the virus. Therapy naïve co-infected individuals demonstrated a higher frequency of intra-hepatic CD8 + T-cells that produce TNF-a in response to both HCV and HIV antigen stimulation compared to HCV mono-infected individuals. In addition, we identified CD8 + T-cells specific for an immunodominant HIV epitope in co-infected liver, demonstrating high frequency of TNF-a expression. Intra-hepatic TNF-a has been previously associated with liver fibrosis, and the accumulation of cells expressing this marker may explain in part the faster rate of liver disease progression found in HCV/HIV coinfection. Further comparisons of TNF-a responses between immunodominant HCV and other HIV epitopes in a larger cohort of individuals are warranted. Contrary to our expectation, viral-specific, intra-hepatic levels of IFN-c were also higher in the therapy-naïve co-infected group, which would be against the expected protective role of IFN-c. However, we interpret this observation as a potential effect of the fibrogenic TNF-a to mask IFN-c protection. On the other hand, viral-specific T-cells are composed of several major populations with unique functional patterns. Therefore, measurement of only one or two T-cell functions may not provide a comprehensive picture of the quality of T-cell responses. Recent lines of evidence demonstrate the importance of the qualitative rather than quantitative characteristics of CD8 + T-cell responses to efficient viral control [13, 27] . The significance of Tcell populations simultaneously representing 5 different functions has been discussed as a hierarchical functional model in viral infections such as CMV and EBV which are effectively controlled by respective CD8 + T-cells [16] . HCV-specific CD8 + T-cells were not poly-functional which is consistent with the notion that although HCV-specific T-cells are found in hepatic tissue, their loss of poly-functionality may be associated with inefficient control of HCV replication. HIV-specific T-cells in the liver of co-infected individuals however, simultaneously could express 4 and 5 of the measured markers. Recently, T-cell exhaustion has been related to signaling pathways through PD-1 [28, 29] . Our analysis of PD-1 levels of antigen-specific CD8 + T-cells from co-infected liver demonstrates higher expression of PD-1 on HCV-specific T-cells, compared to those specific for HIV, supporting the notion that the former are less functional. The observed poly-functionality of intra-hepatic HIV-specific T-cells, should have little effect on HCV replication but would further enhance the cytokine milieu induced from bystander activation, and contribute to liver damage during co-infection with HCV. In this regard, we found that the degranulation marker CD107a dominates the HIV-specific CD8 + T-cell responses in the liver, with the majority of the responding cells expressing CD107a, a surrogate marker for the cytotoxic function of CD8 + T-cells. Activated HIV-specific CD8 + T-cells with the potential to degranulate could induce bystander damage. In addition, the release of chemokines such as MIP-1b by the same cells could also attract further lymphocytes without HCV specificity to the liver. Bystander function of these non-specific T-cells could expand the tissue damage triggered by HCV infection and ultimately activate fibrogenesis. We found that the frequency of CD4 + T-cells within livers of coinfected individuals was reduced compared to HCV monoinfection. Surprisingly, HAART did not appear to reconstitute the CD4 + T-cell population within liver. Despite this defect of CD4 + T-cell help, comparable frequencies of HCV-specific-CD8 + T-cells were found in co-infected livers. HAART-treated biopsies showed further reduced frequencies of HCV-specific responses. These data support previous findings that show HAART induces CD4 + T-cell recovery but not any restoration of HCV-specific Tcell responses peripherally [30] . Further investigation is needed to clarify the role of CD4 + T-cell help in affecting the frequencies of HCV-specific CD8 + T-cells in HCV/HIV-1 co-infection. HAART was also associated with a reduction in frequencies of HIV-specific T-cell responses within liver, indicating that removing the HIV antigenic load may also reduce the opportunity for such cells to accumulate within hepatic tissue. Here, we propose a novel mechanism for enhanced HCVrelated liver disease progression in HIV co-infection; that of bystander activation and induced inflammation from HIV-specific T-cells accumulated in the liver. Our data however are limited in the cross-sectional nature of our cohort, the low number of analyzed liver biopsies and the narrow range of CD4 + T-cell counts among the studied individuals. We should also acknowledge that ex-vivo functional T-cell capacity may not exactly reflect the situation in-vivo. Further studies, particularly, those which are prospective are warranted in order to understand the role that HIV-specific T-cells play in contributing to fibrosis and in particular how HAART modulates these responses. frequency of CD8 + T-cells specific for HCV, HIV and CMV in HCV/HIV co-infected liver (OM 385). Liver isolated mononuclear cells were stained with pools of HLA-A*0201 and HLA-A*2402-restricted pentamers (Pro5 MHC class I Pentamers, Proimmune), followed by staining for cell surface markers CD3 and CD8. The following pentamers were used for each group: HCV pentamers: NS3-CINGVCWTV and NS3-KLVALGINAV; HIV pentamers: Pol-ILKEPVHGV and Gag p24-TLNAWVKVV; CMV Pentamers: pp65-NLVPMVATV and pp65-QYDPVAALF. No CD8 + T-cells specific for CMV were detected in this co-infected liver sample. Similar findings were found in another individual (data not shown). doi:10.1371/journal.pone.0003454.g004 background is shown to become extremely low when examining combinations of functions, nearly reaching 0 events for multiple functions simultaneously. This permits a very low threshold for detection of positive responses from multiple combinations. Consequently, for multi-parameter analysis, the results were thresholded based on a minimum criterion of positivity, as calculated by SPICE software and presented as the 90 th percentile of negative values for each analysis. Each pie chart generated by SPICE software, represents the hierarchy of responses to either HCV or HIV antigen stimulation. For simplicity, responses are grouped by number of functions and matched to the colored bars, with black bars representing the percentage of responding cells to HIV peptides and gray bars representing the percentage of responses to HCV peptide stimulation. In all pie charts, color red represents the 5+ responding population and the colors blue, green, turquoise, and yellow representing the 4+, 3+, 2+, and 1+ populations respectively. Color-coded arcs represent the dominant marker within each pie slice, with color blue representing TNF-a, red for CD107-a, green for IFN-c and black for MIP-1b. Although IL-2 is included in the presentation and demonstrated by bar graphs, the software would not allow for arc colors for more than 4 responses. As a result there is no arc representative for IL-2. Figure (c) represents the average frequency of intra-hepatic viral specific responses within the pool of CD8 + T-cell populations simultaneously expressing 2 functions or less, compared with those within the pool of populations expressing more than 2 functions simultaneously; as analyzed in 3 subjects within each cohort of HCV mono and HCV/HIV co-infected individuals. The cutoff point of simultaneous expression of more than 2 measured markers is considered to show CD8 + T-cell poly-functionality.

Multiorgan failure due to hemophagocytic syndrome: A case report

INTRODUCTION: Hemophagocytic syndrome (HFS) is a potentially lethal disorder due to an uncontrolled immune response to a triggering agent. Our objective is to raise the importance of HFS early diagnosis by presenting a representative case. CASE PRESENTATION: A sixteen-year-old girl with Still disease diagnosis developed a progressive multiorgan failure including acute respiratory distress (ARDS), anemia and thrombopenia, elevated liver enzymes, renal failure, coagulopathy with hypofibrinogenemia, and acute phase reactants elevation despite broad-spectrum antibiotics. A bone marrow puncture-biopsy was performed, and hemophagocytosis was found. Prolonged fever, splenomegaly, bicytopenia, hypofibrinogenemia, hyperferritinemia and hypertriglyceridemia confirmed HFS diagnosis. She received intensive care support therapy including mechanical ventilation and specific therapy according to HLH 2004 protocol, with a very good response. CONCLUSION: Our case shows complexity of HFS diagnosis, due to septic shock-like manifestations. Early diagnosis is essential to start appropriate treatment achieving a better outcome.

Hemophagocytic syndrome (HFS) is a rare disorder characterized by prolonged fever, cytopenias, hepatosplenomegaly, hypertriglyceridemia, disseminated intravascular coagulation (DIC)-like coagulopathy and bone marrow, spleen, liver or lymphatic nodes histiocytosis [1] [2] [3] . A sudden presentation, like a septic shock is possible making its early recognition a challenging diagnosis [4] . It is well known that HFS could be a severe complication in some infections (mainly virals) or in some underlying diseases, such as chronic juvenile arthritis (CJA) [3, 5, 6] . Moreover, it is one of the differential diagnosis in fever of unknown origin [7] . Taking into account that a better outcome has been related to an early treatment [3, 8] , presentation of a difficult diagnosis case in a young lady could be helpful and interesting. A sixteen-year-old girl presented with rash, elevated fever and joints swelling. She was admitted to the hospital to make further examinations as the symptoms persisted for 15 days. A diagnosis of probable Still disease was made. In the following days, her clinical state worsened, with persistence of elevated fever. Serologic tests were negative for Bartonella, Parvovirus B19, Rickettsia, viral hepatitis (A, B and C), HIV, Toxoplasma, Salmonella and Yersinia. Blood cultures were also negative. No abnormalities were found in abdominal and cardiac ultrasonography and in cranial computed tomography (CT). A bone gammagraphy, showed enhanced captation in right tibial malleolus and in proximal interphalangeal joints of fingers 2 and 4 of the right hand, and finger 4 of the left hand. An abdominal CT showed a biliary bladder wall enlargement and a laparotomy was performed, but no signs of cholecystitis were found. A broad spectrum antibiotherapy was started with ciprofloxacin and imipenem. The patient was transferred to pediatric intensive care unit (PICU) as a consequence of progressive multiorgan failure including acute respiratory distress syndrome, liver failure, anemia, thrombopenia and increasing acute phase reactants (APR). On admission to PICU, the patient was on mechanical ventilation. Examination revealed generalized oedema, hypoventilation in both lung bases, abdominal distension and hepatoesplenomegaly. There was also active bleeding around puncture points and through nasogastric tube. Blood analysis on admission is shown in table 1. A high positive end-expiratory pressure (PEEP) was set due to hypoxemia (up to 14 cmH 2 O), with a PO 2 /FiO 2 of 141. Thorax radiography showed bilateral diffuse infiltrates, and slight cardiomegaly. Inotropic support was needed (dopamine at 15 μg/kg/minute) and a perfusion with furosemide (0.5 mg/kg/hour) was started. Laboratory analysis showed abnormal values for haemoglobin (7.7 g/dL), platelets (29,000/mm 3 ) and coagulation times including hypofibrinogenemia (85 mg/dL). She received red blood cell concentrates, platelets and fresh frozen plasma. The same antibiotherapy was maintained and acute liver failure support treatment was started. A bone marrow puncture-biopsy was performed. Activated macrophages with hemophagocytosis were found (figure 1), which together with the clinical and analytical data (bicytopenia, and coagulopathy with hypofibrinogenemia, and afterwards a ferritin of 190594 μg/L and triglycerides of 677 mg/dL) confirmed the HFS diagnosis. Specific treatment for this syndrome was started, following Hemophagocytic Lymphohistiocytosis (HLH) 2004 guidelines. She had a very good response, and at forth day from admission she was extubated to non invasive ventilation using a full-face mask. A progressive analytical normalization was observed and she was discharged after 12 days in PICU, without any sequelae. She is currently being followed as an outpatient. She had two reactivations of her rheumatoid disorder, with a good response to corticoids. HFS is an activation of mononuclear phagocyte system cells, with hemophagocytosis in bone marrow and the rest of reticuloendothelial system. This syndrome can be either primary/familial (familial hemophagocytic lymphohistiocytosis -FHL) or reactive/secondary. FHL has a recessive autosomal inheritance and it develops in chil- Bone marrow smear (optic microscope) Figure 1 Bone marrow smear (optic microscope). Activated macrophages and hemophagocytosis from bone marrow puncture-biopsy. dren younger than 2 years, even though in rare cases it can feature later on [9] . It is rapidly lethal and it is sometimes related to some immunological diseases (X-linked lympho-proliferative, Chediak Higashi and Griscelli syndromes). Secondary HFS has a better outcome than primary HFS. It is triggered mainly by viral infections (especially Ebstein-Barr virus) [10] , and also by bacterial, parasitic and fungal infections. It can also develop during malignancies and rheumatoid disorders (kwown in this case as macrophagic activation syndrome), as in our patient [6] . The activation of mononuclear system cells occurs due to a hypersecretion of proinflammatory cytokines (IFNγ, TNFα, IL6, IL10, M-CSF), as a consequence of a triggering agent, which is often a viral infection [11] . The underlying problem is a T and Natural Killers cells dysfunction, which leads to an uncontrolled immunological response [12] due to inability to eliminate the triggering agent. All viral, bacterial, parasitic and fungal cultures performed in our case were negative. Impaired perforine function due to gene mutations seems to play an important role in HFS pathogenesis, as reported in literature [12] . They are implicated in cytotoxicity by forming a death-inducing pore in target cell [13] . HFS diagnosis is made basing on clinical and histological criteria. Five out of 8 criteria must be fulfilled. Absence of hemophagocytosis does not exclude the diagnosis [2] . Multiorgan failure is the most severe presentation of HFS. In pediatrics, multiorgan failure is usually caused by sepsis. In the present case, the initial diagnosis was septic shock. Therefore, HFS has to be included between the causes of multiorgan failure in pediatrics to permit an early diagnosis and treatment. Central nervous system (CNS) is often involved, which has been linked with a poor prognosis [14] . Even though our patient developed a very severe form of HFS, there seemed to be no CNS involvement, and this agrees with the good outcome. Treatment is nowadays applied according to HLH 2004 protocol, which is designed for primary HFS and also used in severe secondary HFS cases. Aggressive immunochemotherapy is given (dexamethasone, cyclosporine A, etoposide and in patients with CNS symptoms or abnormal CSF, also intrathecal methotrexate and corticoids) [2] . After initial treatment, bone marrow transplantation is indicated in primary disease and in severe and persistent secondary HFS [1] . Our patient had a very good outcome, without any sequelae. Early recognition of this syndrome to apply specific therapy as well as multiorganic failure treatment in PICU, are management key factors. HFS is probably underdiagnosed, as multiorgan failure is usually explained by other more common causes like septic shock. [4] Abbreviations APR: acute phase reactants; CJA: chronic juvenile arthritisl CNS:

Studying copy number variations using a nanofluidic platform

Copy number variations (CNVs) in the human genome are conventionally detected using high-throughput scanning technologies, such as comparative genomic hybridization and high-density single nucleotide polymorphism (SNP) microarrays, or relatively low-throughput techniques, such as quantitative polymerase chain reaction (PCR). All these approaches are limited in resolution and can at best distinguish a twofold (or 50%) difference in copy number. We have developed a new technology to study copy numbers using a platform known as the digital array, a nanofluidic biochip capable of accurately quantitating genes of interest in DNA samples. We have evaluated the digital array's performance using a model system, to show that this technology is exquisitely sensitive, capable of differentiating as little as a 15% difference in gene copy number (or between 6 and 7 copies of a target gene). We have also analyzed commercial DNA samples for their CYP2D6 copy numbers and confirmed that our results were consistent with those obtained independently using conventional techniques. In a screening experiment with breast cancer and normal DNA samples, the ERBB2 gene was found to be amplified in about 35% of breast cancer samples. The use of the digital array enables accurate measurement of gene copy numbers and is of significant value in CNV studies.

Variation in the human genome occurs on multiple levels, from single nucleotide polymorphisms (SNPs) to duplications or deletions of contiguous blocks of DNA sequences (1) (2) (3) (4) (5) . Copy number variation (CNV) is an important polymorphism of DNA segments across a wide range of sizes and one of the primary sources of variation in the human genome (6) . Recently, CNV has been studied extensively because of its close association with large numbers of human disorders (7, 8) . An understanding of this variation is important not only to understand the full spectrum of human genetic variation but also to assess the significance of such variation in disease-association studies. The first human CNV map was constructed from a study of 270 normal individuals with a total of 1447 CNV regions in the whole genome (9) ; more than 15 000 CNVs have been found in the human genome (http://projects. tcag.ca/variation). A recent paper demonstrated the presence of 525 novel insertion sequences across the genomes of eight unrelated individuals, which were not present in the human reference genome, and showed that many of these have different copy numbers (10) . However, the current CNV analysis is mainly dependent upon microarray-based SNP and comparative genomic hybridization (CGH) platforms, or DNA sequencing, and is therefore subject to low sensitivity and low resolution. These techniques are high throughput but lack the flexibility of analyzing individual genes or sequences of interest. Other existing technologies, such as quantitative polymerase chain reaction (PCR), are limited because of their inability to reliably distinguish less than a twofold difference in copy number of a particular gene in DNA samples (11) (12) (13) . In this study we demonstrate the use of a unique integrated nanofluidic system, the digital array, in the study of CNVs. The digital array (14, 15) is able to accurately quantitate DNA samples based on the fact that single DNA molecules are randomly distributed in more than 9000 reaction chambers and then PCR amplified. The concentration of any sequence in a DNA sample (copies/ml) can be calculated using the numbers of positive chambers that contain at least one copy of that sequence. In order to ensure that the apparent difference in gene copy numbers in different samples are real, and not distorted by differences in sample amounts, we use the expression 'relative copy number'. The relative copy number of a gene is the number of copies of that gene per haploid genome. It can be easily expressed as the ratio of the copy number of a target gene to the copy number of a single copy reference gene (two copies per cell) in a DNA sample, which is always 1 per haploid genome. By using two assays for the two genes (the gene of interest and the reference gene) with two fluorescent dyes on the same digital array, we are able to simultaneously quantitate both genes in the same DNA sample. The ratio of the numbers of molecules of these two genes is the relative copy number of the gene of interest in a DNA sample. A single copy gene should have a relative copy number of 1. A relative copy number greater than 1 indicates the presence of duplication of the target gene while a number smaller than 1 implies deletion of this gene. Our data show that the digital array is able to distinguish less than twofold differences in gene copy number and differentiate between 1, 2, 3, 4, 5, 6 and 7 copies of a gene with great accuracy. It provides a reliable and robust platform to study copy number variations and has great advantages over conventional techniques. The sequence of the RPP30 synthetic construct and the sequences of the primers and probe used to amplify this construct are shown in Supplementary The TaqMan assay for the RNase P gene (VIC) was ordered from Applied Biosystems (Foster City, CA). The feasibility of digital PCR has previously been demonstrated by performing PCR on a single DNA sample obtained by a serial dilution process (16, 17) . Target molecules in a DNA sample could be quantitated by counting the number of positive reactions. We utilize the principle of partitioning instead of dilution in order to identify and quantitate individual DNA molecules. The Fluidigm digital array is a novel nanofluidic biochip where digital PCR reactions can be performed (14, 15) . Utilizing nanoscale valves and pumps, the digital array delivers up to 12 mixtures of sample and PCR reagents into 12 individual panels. Each panel contains 765 independent 6-nl chambers. This nanofluidic platform utilizes soft lithography and silicone rubber to create nanoscale valves and pumps that can be used in serial or parallel applications. The digital array is composed of a PDMS (silicone rubber) Integrated Fluidic Circuit, an Integrated Heat Spreader to ensure rapid heat transfer and temperature uniformity within the array and an SBS-formatted carrier with inputs and pressure accumulator to act as an interface between the user and the PDMS chip. There are 12 carrier inputs corresponding to 12 separate sample inputs to the chip. Individual samples of a minimum volume of 8 ml each are delivered into 765 6-nl preprogrammed partitioning chambers in the chip by pressure-driven 'blind filling' in the PDMS. Control lines are primed with control fluid and are pressurized to actuate valves between the reaction chambers. The valves partition individual chambers that are kept closed during the PCR experiment. One of the important applications of the digital array is absolute quantitation (14, 15) . The DNA molecules in each mixture are randomly partitioned into the 765 chambers of each panel. The chip is then thermocycled on Fluidigm's BioMark system and the positive chambers that originally contained one or more molecules will generate fluorescent signals and can be counted by the Digital PCR Analysis software. Since the volumes and dilution factors of the DNA samples are known prior to loading into the digital array, the DNA concentrations can be accurately calculated. The precision of this test is only dependent upon the sampling randomness and, like any biological experiments, will improve with multiple tests (panels). Digital array has been routinely used by us to quantitate DNA samples of unknown concentration and, especially, cDNA samples whose concentrations of the sequences of interest are hard to determine otherwise. When duplication occurs, multiple copies of a gene might be closely linked on the same chromosome and therefore might not be separated from each other, even on the digital array. As a result, multiple copies might behave as a single molecule and the total number of copies of the gene would be underestimated. When two copies are separated by a large genomic distance, some of them might be separated when DNA molecules are fragmented during purification. However, in most cases this would not be sufficient (see Table 2 , sample NA11994 genomic DNA data). Specific target amplification (STA) is a good solution to this problem. STA is a simple PCR reaction with primers for both the reference gene and the gene of interest. It is typically performed for a limited number of thermal cycles (five in this study). The copy numbers of both genes are proportionally increased. Using this process, multiple copies of the gene of interest will be amplified separately and later randomly partitioned into chambers in the digital array. Since the newly generated molecules of both genes reflect the original ratio and they are not linked any more, a digital chip analysis can quantitate the molecules of the two genes and measure their ratio, and therefore the copy number of the gene of interest, very accurately ( Figure 3 ). It is very important that the amplification efficiencies of the two pairs of primers be approximately equal in order not to introduce any bias in the ratio of the two gene copy numbers in the limited number of STA thermal cycles, although this is likely to have an insignificant effect on our results since we utilized only five cycles of preamplification. The amplification efficiency of any pair of primers can be easily measured using real-time PCR (18) . STA was performed on a GeneAmp PCR 9700 system (Applied Biosystems, Foster City, CA) in a 5 ml reaction containing 1  TaqMan PreAmp master mix (Applied Biosystems, Foster City, CA), 225 nM of primers for both RNase P and the target gene and 10-50 ng DNA. Thermocycling conditions were 958C, 10 min hot start and five cycles of 958C for 15 s and 608C for 2 min. The products were diluted prior to the copy number analysis on the digital array based on their initial concentrations so that there would be about 500-600 RNase P molecules per panel. Copy number analysis using the digital array on the BioMark system Each panel of a digital array contains a total of 4.59 ml (6 nl  765 chambers) PCR reaction mix. However, 10 ml reaction mixes were normally prepared for each panel, containing 1  TaqMan gene expression master mix (Applied Biosystems, Foster City, CA), 1  RNase P-VIC TaqMan assay, 1  TaqMan assay (900 nM primers and 200 nM probe) for the target gene, 1  sample loading reagent (Fluidigm, South San Francisco, CA) and DNA with about 1100-1300 copies of the RNase P gene. The reaction mix was uniformly partitioned into the 765 reaction chambers of each panel and the digital array was thermocycled on the BioMark system (http:// www.fluidigm.com/products/biomark-main.html). Thermocycling conditions included a 958C, 10 min hot start followed by 40 cycles of two-step PCR: 15 s at 958C for denaturing and 1 min at 608C for annealing and extension. Molecules of the two genes were independently amplified. FAM and VIC signals of all chambers were recorded at the end of each PCR cycle. After the reaction was completed, Digital PCR Analysis software (Fluidigm, South San Francisco, CA) was used to process the data and count the numbers of both FAM-positive chambers (target gene) and VIC-positive chambers (RNase P) in each panel. There are 765 chambers in each of the 12 panels in a digital array. When single DNA molecules are randomly partitioned into these chambers, it is possible that multiple molecules could partition into the same chamber. As a result there could be more molecules in each panel than positive chambers. The true number of molecules per chamber can be estimated using a simple Poisson distribution equation as described by Sindelka et al. (15) . We have developed a more robust computational algorithm to analyze CNV data obtained from the digital array. This algorithm has been integrated into the Digital PCR Analysis software and is detailed in (19) . A proof-of-principle spike-in experiment was performed using a synthetic construct to explore the digital array's feasibility as a robust platform for the CNV study. A 65-base oligonucleotide that is identical to a fragment of the human RPP30 was ordered from Integrated DNA Technologies (Coralville, IA, USA). RNase P, a single copy gene, is used as reference in this study (20, 21) . Both RPP30 synthetic construct and human genomic DNA NA10860 from the Coriell Cell Repositories (Camden, NJ, USA) were quantitated using the RPP30 assay on a digital array. Different amounts of RPP30 synthetic construct were then spiked into the genomic DNA so that mixtures with ratios of RPP30 and RNase P of 1 : 1 (no spike-in), 1 : 1.5, 1 : 2, 1 : 2.5, 1 : 3 and 1 : 3.5 were made, simulating DNA samples containing two to seven copies of the RPP30 gene. These spike-in mixtures were analyzed on the digital arrays. Five panels were used for each mixture and 400-500 RNase P molecules were present in each panel. The ratios of RPP30/RNase P of all samples were calculated and are plotted against the expected ratios in Figure 1 . A good linear relationship can be observed. Also shown in Figure 2 is an example of a typical digital array experiment. CNVs of the CYP2D6 gene CYP2D6 belongs to the cytochrome P450 system responsible for the metabolism of many commonly prescribed medications (22, 23) . The CYP2D6 gene is highly polymorphic and this can significantly influence the metabolic activity of the enzyme it codes for (debrisoquine 4-hydroxylase) and the therapeutic efficacy of the drugs. Therefore, the pharmacogenetic polymorphism information of this gene would be of great clinical importance in therapeutic decision-making (24) (25) (26) (27) . More than 100 alleles of the CYP2D6 gene have been identified (http://www.cypalleles. ki.se/cyp2d6.htm). Allele-associated variations in the activity of the CYP2D6 enzyme have been observed and individuals carrying these alleles are classified into poor, intermediate, extensive and ultrarapid metabolizers (28, 29) . Genotyping patients would be able to identify those who are at risk of severe toxic responses (poor metabolizer) or in need of more than standard level of drugs (ultra rapid metabolizer). It has been shown that some poor metabolizers and ultra rapid metabolizers are caused by the deletion or duplication of the entire CYP2D6 gene (30, 31) . These large structural changes can be detected using conventional technologies such as Southern blot and longrange PCR. However, it is believed that real-time PCR is Figure 1 . Quantitation of the RPP30 copy number in spike-in samples that contain two to seven copies of the RPP30 molecules per two haploid genomes. The x-axis shows the expected ratio of the numbers of RPP30 molecules to RNase P molecules. The y-axis shows the observed ratios. Each value is calculated using five panels of the same sample mix and the error bars represent standard errors. A good linear correlation can be seen with a coefficient of determination (R 2 ) of 0.996. currently the only promising technique that is able to provide information about the exact copy number of the CYP2D6 gene in a routine clinical setting (32) (33) (34) . We used the digital array to measure the CYP2D6 copy numbers of three DNA samples from ParagonDX (Morrisville, NC). The CYP2D6 genotypes of these DNA samples had been characterized (Table 1 ). The samples were STA-treated (see Figure 3 and Materials and methods section) and the products were analyzed using five panels each on the digital arrays. The relative copy numbers of these three samples are 0, 0.49 and 0.98, respectively, highly consistent with their assumed CYP2D6 diploid copy numbers (0, 1 and 2) based upon their genotypes. We also studied five cell line DNA samples from Coriell Cell Repositories (Camden, NJ). First, we measured their relative copy numbers using genomic DNA. The results showed that two of them have a single copy and two have two copies of the CYP2D6 gene per cell ( Table 2) . One sample had a relative copy number of about 1.17, equal to a diploid copy number of 2.34. We then STA-treated these five samples and ran the products on digital arrays. The relative copy numbers of the 1-and 2-copy samples remained the same and the fifth sample showed a relative copy number of about 1.5 or a diploid copy number of 3. Apparently this sample had a duplication of the CYP2D6 gene on one of the two chromosomes (35) . It has been previously demonstrated (31, 36) that when CYP2D6 duplication occurs, the two copies are separated by 12.1 kb. Therefore, the diploid copy number of 2.34 obtained when genomic DNA was used is likely the result of DNA breakage in this 12.1 kb genomic region in some DNA molecules that separated the two CYP2D6 copies. To confirm this, we ran a long range PCR [see (31) CYP2D6 duplication was observed only in the sample with a relative copy number of 1.5 ( Figure 4) . ERBB2 (also known as HER2) is a receptor tyrosine kinase gene overexpressed in up to 30% of invasive breast cancer, resulting in a loss of normal cellular growth control. Most of these cases (97%) are caused by the amplification of this gene and the number of extra copies is closely related to the protein expression level (37) (38) (39) (40) . ERBB2 amplification is well correlated with an aggressive phenotype characterized by reduced response to chemotherapy, high recurrence rate and short survival time and serves as a significant prognostic predictor for breast cancer patients (37, 41) . Trastuzumab (Herceptin), an FDA-approved monoclonal antibody against the ERBB2 protein, has been shown to dramatically increase response rate and extend survival in breast cancer patients with ERBB2 amplification. Given Trastuzumab's proven efficacy and substantial benefit in multiple clinical trials, detection of ERBB2 amplification has become critical (42) (43) (44) (45) . There are different methodologies of determining the ERBB2 status in breast cancer. Immunohistochemistry (IHC) and fluorescence in situ hybridization (FISH) are two FDA-approved technologies for the detection of ERBB2 amplification. The former detects overexpression of the ERBB2 receptor on the cell membrane while the latter detects the copy number of the gene itself relative to the chromosome 17 centromere. IHC is less expensive and easy to perform but is prone to a high rate of inaccuracies due to variations in tissue preparation, protein stability, antibody sensitivity and scoring subjectivity. On the other hand, FISH is accurate with good clinical correlation but it is expensive, time consuming, and labor intensive and requires very experienced personnel. Therefore, suggestions have been made to use a combination of IHC and FISH, where IHC is used as a screening procedure followed by a FISH confirmation if necessary (46, 47) . We used digital arrays to analyze the ERBB2 copy numbers of 40 breast cancer and 8 normal breast tissue DNA samples from BioChain (Hayward, CA). All DNA samples were from Asian individuals except one normal sample that was from a Caucasian. Of the 40 breast cancer samples, 3 are adenocarcinoma, 1 is fibroadenoma, 2 are invasive lobular carcinoma, 1 is infiltrative ductal carcinoma and 33 are invasive ductal carcinoma. The samples were STA-treated and, for screening purpose, the products were analyzed using only two panels for each sample on digital arrays. The results are shown in Figure 5 . Fourteen breast cancer samples (35%) had a diploid ERBB2 copy number of more than five while all control samples were below five copies [an absolute number of ERBB2 copies greater than 4.0 per cell is considered amplification in FISH analysis (47) . Here we use five as the threshold]. The copy numbers shown are not all integers due to (i) heterogeneity of the cancer cells and (ii) sampling variations as only two panels were used for each sample. A real-time PCR reaction was also performed on these 48 samples. Twenty-four replicates were used for each sample. Although the average copy numbers were close to the digital array data, large fluctuations (SDs of up to 0.5) were observed in the 24 reactions of each sample. Studies on other genes (for example, CYP2D6) showed that real-time PCR does not always produce accurate results (data not shown). Genomewide analyses have shown the existence of large numbers of CNVs in the entire human genome with large interindividual diversity (48) (49) (50) (51) (52) (53) . Many of these CNVs colocalize with genes involved in a variety of diseases or disease susceptibility and are believed to play some role in pathogenesis (54) (55) (56) (57) . The first Mendelian disorder associated with the amplification of a 750 kb DNA fragment was reported recently (54) . It appears to only be a The results of both genomic DNA and STA products are shown. The ratios of the CYP2D6 gene to the RNase P gene should be close to multiples of 0.5. The genomic ratio of 1.17 for sample NA11994 (corresponding to a diploid copy number of 2.34) reflects the partial separation of the duplication alleles in the genomic DNA. A ratio of 1.51 (diploid copy number of 3) was obtained when the sample was subjected to STA prior to the digital PCR analysis. question of time before more genetic conditions related to CNV are identified. Two standard genomewide scanning methods for CNV detection are array-based CGH and high-density SNP genotyping arrays and both were employed in the construction of the first human CNV map (58) . These microarray techniques are able to generate whole-genome CNV data and are important in CNV discovery. Their resolution is also improving with the development of new probes. However, since they are both based on hybridization, the detection of copy number changes largely depends on signal-to-noise ratio, which is sensitive to reagent and manufacturing variability. Therefore, false positive and false negative results are sometimes inevitable (59) . Additionally, the lack of standard reference genomes in the studies using these technologies further complicates the interpretation of the results (60). On many occasions, gene-or locus-specific (other than the whole genome) copy number information is required. This is especially true in the cases of CYP2D6 and ERBB2 described above in which therapeutic decision needs to be made based upon the copy numbers of these genes. In addition to other conventional methods (Southern blot, long-range PCR and FISH), the possibility of using quantitative PCR in the CNV study of these two genes has been previously explored (61) (62) (63) (64) (65) . Quantitative PCR is simple and easy to perform. However, since the copy number of the target gene is derived from the Ct difference between the target gene and a reference gene, the results are very sensitive to the efficiency of the amplification reaction. Even if one compensates for the amplification efficiency, it is considered difficult to obtain a discrimination power of better than twofold (66) . The digital array has the ability to absolutely quantitate any type of DNA sample. In a multiplex PCR reaction with two assays, the quantitation of two or more genes/sequences in a single sample becomes possible, effectively eliminating pipetting variations inherently occurring in any quantitation experiment. The accuracy of the results is only subject to the random distribution of the molecules and, like any biological experiments, can improve with the use of multiple replicates for each sample. STA can efficiently separate the linked copies of a gene on the same chromosome when duplication occurs while other methods, such as restriction digestion are also valid (data not shown). We performed three experiments to test the feasibility of the digital array in the CNV study. First we measured the copy numbers of the RPP30 gene of a series of mixtures made of a human genomic DNA and a synthetic RPP30 construct. We observed a very good correlation between the results and the expected outcome. We then studied the CYP2D6 copy numbers of some DNA samples that were either genotyped elsewhere or characterized by us using conventional techniques. The results were also consistent. Lastly, we screened 40 breast cancer samples for the amplification of the ERBB2 gene. Although the clinical data (other than pathological classification) of these samples were lacking, about 35% of the samples had an increased number of this gene above 5, very close to the ERBB2 amplification frequency reported in the literature (67) . In conclusion, this study shows that the digital array provides a new and robust technology to study geneand sequence-specific CNV and is able to detect gene copy numbers with great accuracy. Digital arrays provide a much greater discrimination power than quantitative PCR. CNV studies on the digital array are easy to perform, fast and the data obtained is easy to interpret. Furthermore, the platform is very flexible and can be tailored to any gene/sequence. It can also serve as an independent measure to verify results from the whole-genome scans using array technologies. The digital array is an excellent CNV platform for both basic research and clinical investigation. Supplementary Data are available at NAR Online. Funding for Open Access charge: Fluidigm Corporation. Conflict of interest statement. The authors declare competing financial interests. All are employees of Fluidigm Corporation. (19) .

An analysis of hospital preparedness capacity for public health emergency in four regions of China: Beijing, Shandong, Guangxi, and Hainan

BACKGROUND: Hospital preparedness is critical for the early detection and management of public health emergency (PHE). Understanding the current status of PHE preparedness is the first step in planning to enhance hospitals' capacities for emergency response. The objective of this study is to understand the current status of hospital PHE preparedness in China. METHODS: Four hundred hospitals in four city and provinces of China were surveyed using a standardized questionnaire. Data related to hospital demographic data; PHE preparation; response to PHE in community; stockpiles of drugs and materials; detection and identification of PHE; procedures for medical treatment; laboratory diagnosis and management; staff training; and risk communication were collected and analyzed. RESULTS: Valid responses were received from 318 (79.5%) of the 400 hospitals surveyed. Of the valid responses, 264 (85.2%) hospitals had emergency plans; 93.3% had command centres and personnel for PHE; 22.9% included community organisations during the training for PHE; 97.4% could transport needed medical staff to a PHE; 53.1% had evaluated stockpiles of drugs; 61.5% had evaluated their supply systems; 55.5% had developed surveillance systems; and 74.6% could monitor the abnormity(See in appendix). Physicians in 80.2% of the analyzed hospitals reported up-to-date knowledge of their institution's PHE protocol. Of the 318 respondents, 97.4% followed strict laboratory regulations, however, only about 33.5% had protocols for suspected samples. Furthermore, only 59.0% could isolate and identify salmonella and staphylococcus and less than 5% could isolate and identify human H5N1 avian flu and SARS. Staff training or drill programs were reported in 94.5% of the institutions; 50.3% periodically assessed the efficacy of staff training; 45% had experts to provide psychological counselling; 12.1% had provided training for their medical staff to assess PHE-related stress. All of the above capacities related to the demographic characteristics of hospitals and will be discussed in-depth in this paper. CONCLUSION: Our survey suggested that, at the time of the survey, hospital preparedness for PHE in China was at an early stage of development. Comprehensive measures should be taken to enhance hospital capacity in the prevention and management of PHE.

Public health emergency (PHE) is an event or events that cause or may cause harm to the health of a community or nation [1] . To prevent and/or minimize the harm caused by PHE, early detection and management are necessary. As hospitals are the main location for PHE surveillance and treatment, their preparedness is critical for PHE's early detection and management [2] . Evaluating the current status of PHE preparedness within the hospital system is the first step in improving a nation's preparedness for a PHE. Yet, there is no national data on China's hospital PHE preparedness capacity aside from two studies that addressed the issues at local level [3, 4] . To understand the current status of hospital PHE preparedness in China, a sample survey of hospitals in four representative city/ provinces were conducted between November 2004 and March 2005. The survey used a cross-sectional study design to survey hospitals in different regions of China. Respondents were all secondary and tertiary hospitals(the detail of hospital classification see in appendix) in the city of Beijing and provinces of Shandong, Guangxi, and Hainan. The selection of hospitals in these four regions is intended to represent a variety of regional economic status. Broadly speaking, Beijing and Shandong are economically well developed, Hainan moderately developed, and Guangxi developing [5] . According to the Hospital Classification Method issued by the National Bureau of Statistics of China, the surveyed hospitals included general hospitals, hospitals of traditional Chinese medicine (TCM), hospitals of integrated traditional Chinese medicine and western medicine (TCM-WM), specialized hospitals, community health center, and medical emergency center (the definition of community health center and medical emergency center see in appendix) [6] . Four hundred secondary and tertiary hospitals were surveyed. The study was approved by the Institutional Review Board (IRB) of the School of Basic Medicine, Peking Union Medical College in Beijing, China. Based on a literature and government document review, a detailed methodological approach for research framework and questionnaire development was followed to inform the development of this study [3] . An indicator system framework was created and questionnaire designed based on the framework. The questionnaire consists of 17 sections and 192 items. The questionnaire and the survey protocol (including field work manual and quality control procedures) were tested by a pilot study. For the purpose of this study, we analyzed the data focused on the following nine areas of interest: (1) hospi-tal's demographic data (including region, SARS crisis experience, teaching function, hospital type, and number of medical staff in related departments); (2) hospital PHE preparation (emergency plans, response initiating time, accessibility, and revision and implementation of emergency plan); (3) response to a community PHE (cooperation with local organizations, relationship with the community PHE network, medical treatment, and rescue work in the community); (4) stockpiles of drugs and materials (stockpiles of drugs and other resources and personal protective equipment); (5)PHE detection and identification (syndrome surveillance); (6) procedures for medical treatment (protocol for diagnosis, treatment, and transfer of PHE victims); (7) laboratory diagnosis and management (laboratory regulation and management system, sample disposal and evaluation system, collection and disposal of suspected samples, and diagnosis of pathogen/etiology); (8) staff training (organization of PHE training, current training of medical staff, curriculum development and training effectiveness assessment); and (9) risk communication (organization for communication of risk psychological counseling to victim and medical staff, and communication with public). Excluding aspect 1, items 2-9 (covering 88 survey questions) represent 8 types of PHE preparedness capacities. Each answered item was scored 1 for "yes" and 0 for "no" or "unknown". Item scores were calculated by adding together "yes" answers. Items scores were used as a proxy for measuring PHE preparedness in an institution. A total item score was measured by calculating the score across all 8 items. The higher the total item score, the better the hospital PHE preparedness capacity. Further analyses were conducted to understand the correlation between preparedness capacity and demographic information. The distribution of the related preparedness capacities across 10 categories of PHE [1] and 15 types of etiology was also assessed. A computerized questionnaire stored in a CD was sent to the targeted hospitals accompanied by an official letter from each of the four city and provincial health departments stating the importance of the survey and requiring that each hospital designates a department director to be responsible for coordinating the completion of the questionnaire. Each returned questionnaire was carefully reviewed for its completeness and consistency. For those questionnaires with incomplete and/or inconsistent responses, one or two follow-up telephone calls were made to ensure completeness and consistency. The data from returned questionnaires were then transferred into a database for analysis. A database was set up using Microsoft Excel 2003. Data was checked, cleaned, and analyzed using SPSS software version 11.5. Ninety-five percent confidence interval of means (95% CI) was used to describe PHE preparedness capacities. Categorical variables were analyzed with frequency and percentage. Comparisons of mean score of each of eight PHE preparedness capacities among different types of hospitals were performed with P < 0.05 as statistical significance using parameter test (Independent-Samples T Test (two-tailed) or One-way Analysis of Variance) and/or non-parameter test (Mann-Whitney Test or Kruskal-Wallis Test) based on data distribution characteristics and homogeneity. Four hundred hospitals responded, with a response rate of 100%. However, seventy-seven questionnaires were excluded from analysis due to one of the following reasons: (1) if less than 50% of items in the questionnaire were not answered, or (2) hospital did not meet secondary and/or tertiary hospital standard according to the hos-pital classification system. Therefore, the valid response rate was 79.5%. Of analyzed hospitals (318), 29.9% were in Beijing, 24.5% in Shandong, 40.6% in Guangxi and 5.0% in Hainan. In terms of hospital type, 72.4% were teaching hospitals. The mean number of physicians and nurses per hospital was 174.5, and the mean number of total medical staff per hospital was 206.1. The mean number of physicians and nurses in emergency department and infectious-disease department were 24.3 and 12.0, respectively. Table 1 shows the demographic characteristics of the analyzed hospitals. Of 318 hospitals, 264(85.2%) had an emergency plan. Among the 264 hospitals that had an emergency plan, 92.6% reported that the institution possessed a protocol to initiate the emergency plan, 75.5% had a classification system for different PHE events, 55.3% had evaluated and revised their emergency plan at least once, and 79.6% reported that their emergency plan was accessible to all 2 and table 3 . Of all analyzed respondents, 64.2% were designated as the local emergency hospital for PHE patient admissions and 53.0% of them were the designated hospitals to provide medical rescue services during a national disaster. Of all analyzed respondents, 97.4% could promptly transport needed medical staff to the PHE field, 84.5% reported that they were prepared to respond to the needs of vulnerable people (including women, children, pregnant women and the disabled) during a PHE, however, only 49.8% had evaluated their ability to increase beds and equipment for PHE. When performing a PHE preparedness drill, 22.9% of respondents reported that they would invite relevant community organizations to participate. With regard to capacity comparison, the statistics test showed: the total item score of hospitals in Beijing(95% CI:5.9,6.9) was lower than that of hospitals in Shandong (95% CI:7.0,7.9) and Guangxi(95% CI:6.7,7.4); the score of teaching hospitals(95% CI:7.0,7.5) was higher than that of non-teaching hospitals(95% CI:5.7,6.6); and the score of tertiary grade A (95% CI:6.8,8.0) and B (95% CI:6.7,8.4) hospitals was higher than that of secondary grade B ones(95% CI:5.4,6.9), respectively. Among all types of hospitals, community health center scored highest on this aspect. Our results revealed that 53.1% of respondents had evaluated their stockpiles of drugs, and 61.5% had established a relationship with suppliers to provide emergency drug- supplies, however, only 43.2% had signed written contracts with suppliers. Of all analyzed respondents, 47.8% had drug-distribution plans, and 21.5% knew where the national or local pharmacy distribution centers were located. In regards to other medical materials, 80.1% had stockpiles of materials for responding to PHE. As for the stockpiles of drugs for infectious diseases, about 93.2%, 91.9% and 43.5% of responding hospitals had drug stockpiles for treating infectious diarrhea, influenza and botulismo toxin, respectively. When hospitals were compared on this item, statistical analysis showed that institutions in Beijing (95% CI:5.8,6.9) had a higher score than that of Shandong (95% CI:6.6,7.9). Tertiary hospitals generally had a higher score than secondary ones. Among all the respondents, 55.5% reported that they had developed syndromic surveillance systems for certain diseases and 84.4% required that physicians on duty should report any abnormity to the hospital's presidents (the definition of abnormity see in appendix). Abnormity in admission diagnosis, routine microbiological tests, emergency room patients, and death with unknown causes were systematically monitored by 74.6% of institutions and 47.4% of hospitals shared their surveillance information with the local health authority. There were statistically significant differences between tertiary grade hospitals (Grade A 95% CI: 5.6,6.6; Grade B 95% CI: 5.6,7.3) and secondary grade B hospitals (95% CI:4.1,5.9) for this capacity, with tertiary hospitals scoring higher on their ability to detect and identify a PHE. Physicians in 80.2% of the responding institutions reported being familiarized with the latest treatment protocol for a PHE, 92.8% could transfer PHE victims to corresponding medical agencies for appropriate treatment, and 98.0% could provide training on the protocol system. However, only 69.0% had specific procedures for patient transfer in a PHE. As for infectious disease treatment protocol, 80.1% had protocols for SARS, but only 37.3% for brucellosis. With regard to the capacity comparison Among all the respondents, 94.5% reported that they had a training program for the following medical staff: infection managers (56.3%); emergency department physicians and nurses (92.2%); and infectious disease ward physicians and nurses (71.8%). Staff training was supervised by a designated person in 82.3% of institutions and 65.8% had training curriculums, 66.5% of which was updated regularly. Effectiveness of PHE training was periodically assessed in 50.3% of respondents. For this capacity, statistical significance indicated that respondents in Shandong (95% CI:5.9,7.0) scored higher than participating institutions in Guangxi (95% CI:5.1,6.0). Serious PHE concerns were raised in China during the 2003 SARS crisis when it became apparent that hospitals possessed poor emergency preparedness [7] . Even the upcoming 2008 Olympics Game in Beijing and the 5.12 Earthquake Disaster in China have dramatically evoked the awareness of PHE preparedness capacity for hospital. Based on the experience of the SARS pandemic, all hospitals should possess fundamental PHE programs, including preparedness of drugs, equipment, staff, emergency education and staff training [3, 8, 9] , coordination with relevant community bodies [10], medical treatment [11] , early detection and warning [12] , laboratory diagnosis [13] [14] [15] and psychological intervention [8] . Since the SARS crisis, the central Chinese government has become more active in the construction of public health system, especially in regards to the medical emergency response system [16] . One major effort involved a 11.4 billion RMB investment in local governments to initiate the construction of regional PHE medical treatment systems [17] . In order to offer some insight into the development of hospital PHE preparedness capacity, this study examined the current status of hospital preparedness in Beijing, Shandong, Guangxi, and Hainan. Emergency preparedness refers to the processes involved in ensuring an institution: (1) has complied with the preventive measures; (2) is in a state of readiness to contain the effects of a forecasted disastrous event in order to minimize loss of life, injury, and damage to property; (3) can provide rescue, relief, rehabilitation, and other services in the aftermath of the disaster; and (4) holds the capability and resources to continue to sustain its essential functions during a PHE [18] . An emergency preparedness systems primarily composed of emergency plans and organizational structures and lays the foundation for dealing with PHE [19] . Emergency plans establish the protocol for operation under a PHE [16] . For a hospital to mobilize all PHE resources in a short period of time, contingency plans must be issued in advance [9] . In addition, periodic review and updating of emergency plans enhance an institution's emergency response capacity [3] . Our study showed that most hospitals had emergency plans and that these plans focused on infectious diseases control with less attention to preparedness for biological, nuclear radiation and other terrorism attacks. Most of the hospitals had PHE command departments and emergency response teams, however, only 55.3% of hospitals with emergency plans reported they had evaluated and revised their PHE systems. Overall, tertiary hospitals performed better in PHE preparation than secondary hospitals. Meanwhile, no statistical significance was found between hospitals that had admitted SARS patients and those that had not, suggesting that after the SARS crisis, all hospitals raised awareness of emergency plans and implementation. No hospital or medical system can manage a public health emergency without community networks and public involvement. Therefore, hospitals need to communicate and cooperate with other local health agencies, functioning as a networked public health provider. Problems like lack of communication and coordination between hospital departments and inter-agency networks hinder the availability of resources in a community and limit timely forecasting, public communication and effective regulation of a PHE [10]. Our survey revealed that if a PHE occurred, most of hospitals reported that they could take responsibility for PHE rescue service, transport the medical staff in a timely manner, and provide priority health services to vulnerable populations. Yet, less than one third of respondents attended regulation and revision workshops for emergency plans for infectious epidemic control held by local agencies. This lack of cross-institutional interaction indicated that the ability of hospitals to coordinate with community agencies in preparation for, or in the event of a PHE was generally poor. The survey showed that among all the types of respondents community health center were best able to respond to PHE and the respondents with multiple functions performed better suggesting that communication and coordination between hospitals and community agencies should be strengthened. Characteristics of a PHE include suddenness and unpredictability [9] . For most hospitals, medicine storage may be in great demand when faced with a sudden increase in patients. Therefore, hospitals must have programs to ensure appropriate levels of emergency supplies including drugs, medical equipment, electricity, water and oxygen, disinfectant, etc. Our survey suggested that most of the hospitals could establish an emergency-drug-supply system for most of the infectious diseases we addressed in the questionnaire except anthrax, brucellosis, botulism toxin poisoning and tetramine poisoning. For most of surveyed hospitals possessed emergency resource reserves, but less than half of them had corresponding drug distribution programs. In addition, hospital capacity was affected by economic level and classification of the hospital, suggesting that the importance of local economic development strengthens hospital ability to provide PHE. Early detection and identification of a PHE are amongst the most important objectives for prompt and effective public health response to a PHE [12] as well as an essential precondition for selecting appropriate prevention and treatment measures. This study showed that most of the hospitals could regularly train medical staff on how to report and identify suspicious PHE and that the institutions possessed surveillance systems to monitor various aspects of abnormity. Approximately half of the respondents could share surveillance information with the local health authorities. There were statistically significant differences among various classification of the respondents, which demonstrated that after the SARS crisis, hospitals at all levels attached high importance to PHE monitoring and early warning system, however, the capacity was affected by the comprehensive strength of hospital. PHE happens suddenly and its incidence rate is relatively low, which leaves most medical staff inexperienced and unprepared [11] . Therefore, it is important that hospitals develop emergency plans for PHE treatment programs. In this survey, more than half of respondents showed that their physicians were aware of current PHE protocols. Most hospitals had transfer and treating procedures for infectious diseases, including SARS, influenza, and infectious diarrhea, but less held these procedures for biochemical incidents, leakage of nuclear, and terrorist attacks. Because they are easily used as biological terrorist attacks materials [20] , therefore, the prevention and control of these emergencies become very important. Our statistical analyses showed that tertiary-grade, teaching and TCM-WM hospitals performed better on medical treatment procedures preparedness, which might reflect the fact that different types of hospitals have different functions and mission in the community, however, for this capacity, the statistical significance among different regions showed the important role that economic factor plays. Hospital laboratories not only have the task of clinical diagnosis, but take some responsibility in the surveillance of public health [13, 14] . Therefore, laboratory informa-tion plays an important role in detection of the PHE [13, 15] . Detecting PHE related pathogen/etiology can not only confirm clinical diagnosis, but also identify newly emerging infectious diseases [15, 21] . The presence of SARS in China in 2003, and the slow response to its emergence, revealed that China's public health laboratory systems were weak [13] . This survey indicated that many of the hospitals did not report adequate laboratory diagnostic capacities. Although hospital laboratory regulations seemed relatively good, only one-third of hospital laboratories had programs for dealing with suspicious samples collecting, disposal and delivery. [23] . When PHE occurs, hospital medical staff are usually the first responders and information providers, therefore, education and training are key measures to enhance PHE response [24] . Our survey suggested that after SARS crisis, most hospitals re-evaluated the importance of medical staff training for PHE. The majority of respondents offered training programs to their related medical staff. However, the effectiveness of these training programs needs to be periodically evaluated. PHE can cause psychological as well as physical problems for the public and medical staff attending to victims [9, 25] . In a public health crisis or emergency, effective risk communication can help people cope, make decisions, and return their lives to normal. Crisis communication, as an important part of a PHE response [8] , is key to ensuring complete, transparent and prompt information exchange, and to help hospitals make timely responses and reduce the serious consequences [26] . The results of this survey revealed that medical staff in 12.1% of the hospitals underwent training for evaluation of PHE-related stress and only one-third of respondents had specific programs and spokespersons for communicating critical messages and information to the media, public, governments and stakeholders. These results indicated that most of the surveyed hospitals do not understand the importance of psychological care in a PHE emergency, do not have the resources to deal with it, or presume that it is not their place to do so. Indeed, this capacity evaluation revealed that when a PHE occurred, most hospitals' response plans focused on physiological medical treatment, but health education, psychological counseling, and crisis communication plans were rare. However, for this capacity, the statistical significance among different regions and levels showed the important role that economic factor and comprehensive level play. The study has several limitations. First of all, the surveyed hospitals were restricted to four city and provinces, even some types of hospitals were rare (the number of the surveyed community health center and emergency center was just one, respectively), therefore, the results may not fully represent the PHE capacity of all hospitals in China. Secondly, because of self-report method there may be a respondent reporting bias. The inclusion of official documents from respective Health Bureaus, for example, may have encouraged respondents to complete survey but have also been interpreted as an official assessment of capacity leading some hospital representatives to overestimate PHE capacity. Thirdly, only quantitative data were collected to measure certain capacities of PHE preparedness. Most questions required a "yes" "no" or "unknown" answer which restricts the collated data to these three categories. Finally, this data set is not complete as some hospitals did not respond and others had to be excluded on the basis of incomplete answers or for ineligibility for hospital classification. To a certain extent, this loss of respondents caused a loss of information. After several years of construction and development, the capability of hospitals in China to deal with PHE, in particular infectious diseases control, has improved greatly [3, 4] . Nevertheless, this research suggests that China has more progress to make before PHE preparedness is satisfactory. To enhance hospital preparation for dealing with PHE, governments at all levels should increase investment in the construction of infrastructure to create and sustain appropriate PHE capacity. On the other hand, hospitals at all levels should enhance their management, including updating and revising of emergency plans; strengthening communication and cooperation with other local agencies; enhancing the capacity of abnormity monitoring and laboratory diagnostic capability for infectious diseases; improving the treatment program for various PHE scenarios; and strengthening psychological intervention and risk communication capabilities. Finally PHE preparedness in relation to terrorism caused by nuclear radiation and biochemical substance was low in this study and should be further assessed for areas of need and improvement.

Molecular evidence for the evolution of ichnoviruses from ascoviruses by symbiogenesis

BACKGROUND: Female endoparasitic ichneumonid wasps inject virus-like particles into their caterpillar hosts to suppress immunity. These particles are classified as ichnovirus virions and resemble ascovirus virions, which are also transmitted by parasitic wasps and attack caterpillars. Ascoviruses replicate DNA and produce virions. Polydnavirus DNA consists of wasp DNA replicated by the wasp from its genome, which also directs particle synthesis. Structural similarities between ascovirus and ichnovirus particles and the biology of their transmission suggest that ichnoviruses evolved from ascoviruses, although molecular evidence for this hypothesis is lacking. RESULTS: Here we show that a family of unique pox-D5 NTPase proteins in the Glypta fumiferanae ichnovirus are related to three Diadromus pulchellus ascovirus proteins encoded by ORFs 90, 91 and 93. A new alignment technique also shows that two proteins from a related ichnovirus are orthologs of other ascovirus virion proteins. CONCLUSION: Our results provide molecular evidence supporting the origin of ichnoviruses from ascoviruses by lateral transfer of ascoviral genes into ichneumonid wasp genomes, perhaps the first example of symbiogenesis between large DNA viruses and eukaryotic organisms. We also discuss the limits of this evidence through complementary studies, which revealed that passive lateral transfer of viral genes among polydnaviral, bacterial, and wasp genomes may have occurred repeatedly through an intimate coupling of both recombination and replication of viral genomes during evolution. The impact of passive lateral transfers on evolutionary relationships between polydnaviruses and viruses with large double-stranded genomes is considered in the context of the theory of symbiogenesis.

Approximately two-thirds of these wasps are endoparasites, meaning that the larval stages develop within the body cavity of their hosts, typically other insects. Among the most successful of these endoparasitic wasps are those that use lepidopteran larvae as hosts. Owing to the economic importance of these insects and the utility of their wasp parasites as biological control agents, the ability of these parasites to develop within lepidopteran hosts without triggering an intense immune response has been the subject of numerous studies over the past forty years. Early studies of the Mediterranean flour moth, Ephestia kuhniella, parasitized by the ichnemonid, Venturia canescens, showed that eggs of this species are coated with particles that resemble virions [2] [3] [4] and contain surface proteins that mimic host proteins, thus keeping the eggs and larvae from being recognized as foreign material by their host. These particles lack DNA, and thus are not considered virions [5] . With respect to both species number and mechanisms that lead to successful parasitism, endoparasitic wasps are known to inject secretions at oviposition, but only a few lineages use viruses or virus-like particles (VLPs) to evade or to suppress host defences. In the family Ichneumonidae, for example, four types of host defence suppression mediated by the injection of fluids or suspensions are known that lead to successful parasitism. 1) Fluid injected with eggs bypasses host defences without the aid of viruses or VLPs [6] . 2) Wasps inject a virus that replicates in both the wasp and lepidopteran host. One example is the wasp Diadromus pulchellus, which injects an ascovirus, DpAV4 [7] into host pupae to circumvent host defence response. 3) The wasp injects VLPs capable of molecular mimicry and/or direct defence suppression. 4) The wasp injects polydnavirus particles that contain genes coding for proteins that interfere with host defence responses. The last mechanism is by far the best-studied type of direct immune suppression by ichneumonid wasps, and occurs in many species belonging to genera Campoletis, Hyposoter and Tranosema (Ichneumonidae, Campopleginae), and Glypta (Ichneumonidae, Banchinae) [8] . In these cases, female wasps inject eggs along with ichnovirus particles into their hosts. Similarly, in certain lineages of endoparasitic braconid wasps, other types of immunosuppressive particles containing DNA occur in the fluid injected along with eggs [ [9] ; for a review, [10] ]. Once in the host, ichneumonid and brachonid particles enter host nuclei and their DNA is transcribed, producing proteins that selectively suppress various steps in the host defence response. As a result of this unusual biology, these particles were described as symbiotic viruses belonging to new viral family, Polydnaviridae [10] [11] [12] Since the 1970's, it was assumed that the DNA in the polydnavirus particles, as with all other viruses, encoded typical enzymes and proteins for viral replication and virion assembly and structure. However, several recent genomic studies have shown that only a small number of the genes vectored into lepidopteran hosts, less than 2%, have homologs in other viruses. Most viral DNA is noncoding, except that which codes for wasp proteins involved in suppression of immune pathways, such as phenoloxidase activation and the toll pathways [8, 13, 14] . Even before these genomic studies, it was suggested that these particles were more similar to organelles than viruses [15] . The similarities between particle structure and virions of known types of complex DNA insect viruses are striking, and suggest these immunosuppressive particles originated by symbiogenesis between viruses and endoparasitic wasps, the same evolutionary process by which mitochondria and plastids originated from symbiotic bacteria [16] . For example, most braconid wasps produce enveloped bacilliform particles classified as bracoviruses, and these resemble baculovirus and nudivirus virions [10, 15] . Similarly, ichneumonid wasps produce enveloped spindle-shaped particles classified as ichnoviruses that resemble virions of ascoviruses, viruses lethal to lepidopterans, which, interestingly, are vectored by endoparasitic wasps [15] . It must also be noted that ichnoviruses resemble other true virus particles that are structurally very similar to virions of ascoviruses, but which remain unclassified because the lack of information about their genomes [17] [18] [19] [20] [21] . However, ascoviruses and ichnoviruses display very different genome properties; similar genomic differences occur between bracoviruses and baculoviruses or nudiviruses, suggesting that convergent evolution led to the origin the different polydnavirus types from at least two different types of viruses. In ascoviruses, the genome consists of a single circular DNA molecule ranging from 119-to 180-kpb in size [7] . Phylogenetic analyses of several viral genes have revealed that ascoviruses are closely related to iridoviruses [22] , and likely evolved from them. In contrast, the genome of ichnoviruses is composed of multiple circular DNA molecules (25 to 105) representing a total size of 250 to 300kbp, all of which are replicated from the wasp chromosomes. The ichnovirus proviral genome is specifically excised and amplified in several segments in the female calyx cells, the only wasp tissue in which ichnovirus virogenesis occurs. After assembly, these particles are secreted into the female genital tract. Once injected into the host, the ichnovirus genome does not replicate, and does not lead to the production of a new virus generation. The third characteristic of ichnoviruses is that most of the genes borne by the particles are not related to viral genes. Among the 7 annotated ichnovirus gene families, there are four (rep, PRRP, N, and TrV) for which no homology with known eukaryotic (or prokaryotic) proteins has been detected and for which no function has been proposed. Among the remaining three (cys, ank and inx), cys-motif proteins have no clear homologs among eukaryotic (or prokaryotic) proteins, although the "cysteine knot" that they form is a folding domain found in many proteins, but not one that is necessarily related to eukaryotic host immune systems [10, 14] . However, some protein domains and their putative functions suggest that they might be related to regulatory components of eukaryotic host defence systems that are not sufficiently elucidated. Although the resemblance of the polydnavirus virions to those of conventional insect viruses suggests that the former evolved from the latter, to date no molecular evidence supports this hypothesis. In the case of ascoviruses and ichnoviruses, well-conserved genes found among the three ascoviruses sequenced so far (SfAV1a [23] , TnAV2c [24] , and HvAV3e [25] ) are not found in ichnovirus genomes. As noted above, the principal reason for this is that the genomes of the latter viruses appear to contain mainly wasp genes, not viral genes. This highlights the need for new and alternative types of sequence data obtained from pertinent biological systems. In this regard, DpAV4 has features that could provide important insights. Indeed, it is the only ascovirus known to replicate in both its wasp and caterpillar hosts. It is transmitted vertically from wasp to caterpillars to suppress the defence response of the latter host, thereby enabling parasite development [26, 27] . Moreover, in males and females of D. pulchellus, the DpAV4 genome resides in the nuclei of all hosts cells, providing a possible example of what may have been an intermediate stage in the symbiogenesis that led to the evolutionary origin of ichnoviruses. We recently sequenced the DpAV4 genome, and a combination of our analysis of this genome and recent data from new types of ichnoviruses, as well as new software programs that elucidate protein relationships based on structural analysis, have enabled us to detect phylogenetic relationships between proteins encoded by open reading frames of DpAV4 and the Glypta fumiferanae (GfIV) and Campolitis sonorensis (CsIV) ichnoviruses. In support of the symbiogenesis hypothesis for the origin of ichnoviruses, data and analyses suggest two independent symbiogenic events, in agreement with what was previously proposed [28] . The first led to the ichnoviruses in Banchinae lineage. This hypothesis is based on the occurrence of a gene cluster present in GfIV and DpAV4. The second symbiogenic event led to ichnoviruses in the Campopleginae wasp lineage. This hypothesis is based on relationships of the major capsid proteins among CsIV, ascoviruses and iridoviruses. Extending our investigations to proteins encoded by open reading frames of certain ascoviruses and bracoviruses, hosts and bacteria, in the light of recent analyses about the involvement of the replication machinery of virus groups related to ascoviruses in lateral gene transfer [29] , we discuss the robustness and the limits of the molecular evidence supporting an ascovirus origin for ichnovirus lineages. The DpAV4 genome sequenced by Genoscope (France) is 119,334-bp in length. Its organization, gene content and evolutionary characteristics will be detailed in a separate publication (manuscript in preparation; Additional file 1). However, BLAST results obtained with several ORFs in the DpAV4 genome provide evidence that certain ichnovirus ORFs have their closest relatives in an ascovirus genome. Specifically, we identified a 13-kbp region that contains a cluster of three genes ( Fig. 1 , ORF90, 91 and 93; Additional files 1 and 2) that have close homologs in a GfIV gene family composed of seven members [28] . All contain a domain similar to a conserved domain found in the pox-D5 family of NTPases. To date, this pox-D5 domain has been identified as a NTP binding domain of about 250 amino acid residues found only in viral proteins encoded by poxvirus, iridovirus, ascovirus and mimivirus genomes. These genes seem to be specific to GfIV, as they are absent in the three sequenced genomes of other ichnoviruses, namely CsIV, Tranosema rostrales ichnovirus (TrIV), and Hyposoter fugitivus ichnovirus (HfIV). More specifically, in DpAV4, ORF90 encodes a protein of 925 amino acid residues that is 40% similar from position 140 to 925 to a protein of 972 amino acid residues encoded by the ORF1 contained in the segment C20 in the GfIV genome (Fig. 2) . These two proteins can therefore be considered putative orthologs. The 480 C-terminal residues of this DpAV4 protein are also 42% similar to the Cterminal domain of the protein homologs encoded by the ORF1 of the D1 and D4 GfIV segments, 36% similar to the N-terminal and the C-terminal domains of the protein encoded by the ORFs 184R and 128L of the iridovirus CIV and LCDV, and 30% similar with those encoded by ORFs 119, 99 and 78 in the ascovirus genomes of HvAV3e, SfAV1a and TnAV2c, respectively. Overall, this indicates that this DpAV4 protein is more closely related to that of GfIV than to those found in other ascovirus and iridovirus genomes currently available in databases. ORF091 encodes a protein of 161 amino acid residues similar only with the C-terminal domain of three proteins encoded by the ORFs 1, 1 and 3, contained, respectively, in GfIV segments D1, D4 and D3. In contrast, ORF93 is closer to iridovirus and ascovirus genes than to GfIV genes. This protein of 849 amino acid residues is 43% similar over all its length to CIV ORF184R orthologs in all iridoviral and ascoviral genomes and is only 36% similar over 350 amino acid residues to the C-terminal domain of the GfIV protein homologs encoded by the ORF1, 2, 1, 1, 1 and 1 in, respectively, the C20, C21, D1, D2, D3 and D4 segments of this virus. Analysis of the genes surrounding the DpAV4 ORF-90-91-93 cluster confirms that this virus has an ascovirus origin since this region contains ORFs that are close homologs of genes in iridovirus and ascovirus genomes. Upstream from the ORF-90-91-93 cluster, an ORF encoding the DNA-dependent RNA polymerase 1 subunit C is present, which is an ortholog of the iridoviral CIV ORF176R and the ascoviral SfAV1a ORF008. Downstream from this cluster, there are two genes, absent in known ascoviral genomes, but similar to the iridoviral CIV ORF115L and CIV ORF132L. These two genes encode, respectively, a chromosomal replication initiation protein and zinc finger protein. In between them, a gene encoding a small protein is present that is similar to that encoded by the ORF069L of the iridovirus CIV, and which corresponds to the ALI-like protein also found in entomopoxviruses [30] . Since the three DpAV4 genes have relatives in all ascovirus and iridovirus genomes sequenced so far, their presence in the DpAV4 genome cannot result from a lateral transfer that occurred from an ichnovirus genome related GfIV to DpAV4. Thus, as these DpAV4 genes are the closest relatives of the pox-D5 gene family present in GfIV identified so far, they could be considered a landmark of the symbiogenic ascovirus origin of the ichnovirus lineage to which this polydnavirus belongs. An alternative explanation is that the presence of DpAV4-like genes in the genome of GfIV resulted from a lateral transfer from viral genomes closely related to those of GfIV and DpAV4. Indeed, this might have happened when a Glypta wasp was infected by an ancestral virus related to DpAV4. Nevertheless, the symbiogenic origin of GfIV from ascoviruses is also supported by morphological features of its virions [28] , which, aside from similarities in shape, also show reticulations on their surface in negatively stained preparations, a characteristic of the virions of all ascovirus species examined to date [7] . Because ascovirus virions and ichnovirus particles display structural similarities, we developed an approach to search for homologs of virion structural proteins in ichnoviruses. These approaches were initiated in 2000 and recently finalized, but some of the conclusions have been published [14] . To date, only two virion proteins from the Campoletis sonorensis ichnovirus (CsIV) have been characterized [31, 32] . The first is the P44 (Acc N° AAD01199), a structural protein that appears to be located as a layer between the out envelope and nucleocapsid, and the second, P12, a capsid protein (Acc N° AF004367). Presently, there are more than one hundred ascoviral or iridoviral MCP sequences in databases. BLAST searches using these sequences failed to detect any similarities between CsIV virion proteins and ascoviral or iridoviral MCPs, or any other proteins [33] . To evaluate the possibility that homology between ichnovirus and ascovirus virion proteins may simply not be detectable by conventional Blastp searches, we used a different method, WAPAM (weighted automata pattern matching; [34] ). The models were designed on the basis of a previous study [22] demonstrating that MCP encoded by ascovirus, iridovirus, phycodnavirus and asfarvirus genomes are related, and all contain 7 conserved domains separated by hinges of very variable size. We investigated these conserved domains further using hydrophobic cluster analysis (HCA, [35] ). This Map of the 13-kbp region of the DpAV4 genome (EMBL Acc. N° CU469068 and CU467486) that contains the gene cluster with direct homologs in the genome of the Glypta fumiferanae ichnovirus Amino acid sequence comparison resulting from a BLAST search done with the DpAV4 ORF90 as a query, and the best hit corresponding to the protein encoded by the ORF1 of the ichnovirus segment GfV-C20 (Subject; Genbank Acc. N° YP_001029423) Figure 2 Amino acid sequence comparison resulting from a BLAST search done with the DpAV4 ORF90 as a query, and the best hit corresponding to the protein encoded by the ORF1 of the ichnovirus segment GfV-C20 (Subject; Genbank Acc. N° YP_001029423). analysis revealed that most conservation occurred at the level of hydrophobic residues, as expected for structural proteins (Additional file 3a and 3b). The size variability of the hinges between conserved domains and the conservation of hydrophobic residues might explain why BLAST searches using iridoviral and ascoviral MCP sequences have limited ability to detect MCP orthologs in phycodnavirus and asfarvirus genomes. We designed two syntactic models (see Materials and Methods), which together were able to specifically align all MCP sequences of the four virus families. Importantly, WAPAM aligned the CsIV ichnovirus P44 structural protein with both models. Complementary structural and HCA confirmed the presence of the seven conserved domains in this CsIV structural protein ( Fig. 3a and Additional file 3c). In addition to the above analysis, ten syntactic models were developed using proteins conserved in the three sequenced ascovirus species (SfAV1a, TnAV2c, and HvAV3a) and twelve iridoviruses [36] . None of these 1 and 4, typed in black) , DpAV4 (lanes 2 and 5, typed in blue) and SfAV1a (lanes 3 and 6, typed in purple) . Conserved positions among the amino acid sequence of CsIV and those of DpAV4 and SfAV1a are highlighted in grey. Secondary structures in the three SfAV1a ORF061 orthologs were calculated with the Network Protein Sequence Analysis at http://npsa-pbil.ibcp.fr/ and the statistical relevance of the secondary structures were evaluated with Psipred at http://bioinf.cs.ucl.ac.uk/psipred/. C, E and H in lanes 4 to 6 respectively indicated for each amino acid that it is involved in a coiled, b sheet or a helix structure. Using default parameters of Psipred, upper case letters indicate that the predicted secondary structure is statically significant in Psipred results. Significant secondary structures are highlighted in yellow. In (a), the comparisons were limited to three of the seven conserved domains (Additional file 3a, b and 3c), the 2, 5 and 7. Indeed, classical in silico methods appeared to be inappropriate to predict statistically significant secondary structures in conserved structural protein rich in b strand such as iridovirus and ascovirus MCP. In contrast, a complete and coherent domain comparison was obtained by HCA profiles (fig. S3b, c) . , developed from small proteins encoded by the DpAV4 ORF041, SfAV1a ORF061, HvAV3a ORF74, and TnAV2c ORF118 in the ascovirus genomes, and iridovirus CIV ORF347L and mimivirus MIV ORF096R genomes, respectively. Importantly, these proteins have orthologs in vertebrate iridoviruses, phycodnaviruses, and asfarvirus. In SfAV1a, the peptide encoded by ORF061 is one of the virion components. In ascoviruses, iridoviruses, phycodnaviruses, and the asfarvirus, they have been annotated as thioredoxines, proteins that play a role in initiating viral infection [37] [38] [39] . Database mining with our model revealed four hits with CsIV sequences (Acc N°. M80623, S47226, AF236017, AF362508) each a homolog ORF of SfAV1a ORF061. In fact, these sequences correspond to several variants of a single region contained in the B segment of the CsIV genome. To date, these have not been annotated in the final CsIV genome, probably because they overlap a recombination site. HCA analyses confirmed that the hydrophobic cores were conserved ( Fig. 3b and Additional file 3d and 3e). The chromosomal locations of genes encoding these two CsIV proteins, i.e., P44 and P12, were also consistent with the symbiogenesis hypothesis. In fact, the ORF encoding P44 is not found in proviral DNA. It is notable that no ORFs encoding orthologs of P44 or other structural proteins such as MCPs are found in any of the other three ichnovirus genomes sequenced -TrIV, GfIV, HfIV [8, 14] . Therefore, this indicates that the orthologs of ichnovirus MCPs and other virion structural proteins are also probably located in the genomes of these wasps, i.e., not in proviral DNA. In contrast to this, we found that the gene encoding the CsIV ortholog of SfAV1a ORF061 is located within the proviral DNA. Whether ortholog proteins are similarly involved in the TrIV, GfIV and HfIV biology, their genes are not found in proviral DNA, since no matches were detected in their viral genomes. The phylogenetic analysis performed previously on P44 and the SfAV1a ORF061 orthologs [15] indicated that they have an ancestor close to that of the ascoviruses and iridoviruses. As in the case of genes encoding pox-D5 family of NTPases in all ascoviruses, iridoviruses, and GfIV, genes encoding virion proteins cannot result from a horizontal transfer from a Campoplegine or Banchine ichnovirus genome to all ascovirus, iridovirus, phycodnaviruses and asfarvirus genomes. As the ascovirus genes encoding the two virion proteins investigated here are the closest relatives of virion proteins in CsIV, they can be considered a landmark reflecting the symbiogenic origin of the two ichnovirus lineages from ascoviruses closely related to DpAV4. In fact, the difficulty encountered in elucidating their sequence relationships can be explained by a combination of the marked transition from ascovirus to ichnovirus, and the significant selection constraints that resulted as the latter virus type evolved from the former. Analysis of available ascovirus, iridovirus and ichnovirus genomes provides some of the first molecular support for the hypothesis that ichnoviruses evolved from ascoviruses by symbiogenesis. However, examining genes shared only by ascovirus, iridovirus and ichnovirus genomes likely limits the sources of genes that contributed to the evolution and complexity of these viruses, especially of the role of lateral gene transfer. Relevant to this is the recent finding that an important part of the mimivirus and phycodnavirus genomes had a bacterial origin [28] . Obviously, this did not lead to the conclusion that these viruses had a bacterial origin. The cytoplasmic environment in which these viruses replicate is rich in bacterial DNA because their amobae and unicellular algae hosts feed on bacteria that they digest in their cytoplasm. Thus, it has been proposed [28] that lateral transfers of bacterial DNA within these viral genomes were driven by intimate coupling of recombination and viral genome replication. Indeed, replication of these viruses is similar to that of bacteriophage T4. This mode of replication has been called recombination-primed replication. It permits integration of DNA molecules with sequence homology as short as 12-bp [28, 40] . The replication machinery used by ascoviruses, iridoviruses, mimiviruses, phycodnaviruses, and other nucleocytoplasmic large DNA viruses (NCLDV) [41, 42] is common to all of them, despite differences in the specifics of replication in each virus family. It can therefore be expected that recombination-primed replication occurred repeatedly during evolution of both these viruses and the genome of their eukaryotic hosts. In an eukaryotic cellular environment in which bacteria, chromosomes, NCLDV viruses and non-NCLDVs (such as baculoviruses) intimately cohabit temporarily or permanently, recombination-primed replication is able to allow reciprocal passive lateral transfers between viral genomes, host chromosomes, and bacterial DNA. Under these conditions, lateral transfers are considered passive since they just result from the intimate environment and not from an active mechanism dedicated to genetic exchanges. In ascoviruses and iridoviruses, the occurrence of such lateral transfers is supported by BLASTp searches that detected the presence of ORFs whose closest relatives have their origin within eukaryotic genomes (e.g., for DpAV4, in Additional data 1, ORFs 029, 049, 077, 080, 083, 118), bacterial genomes (e.g., for DpAV4, in Additional data 1, ORFs 056, 057, 059, 112, 115 and119) or viruses belonging to other NCLDV and non-NCLDV families (e.g., for DpAV4, in Additional data 1, ORFs 007, 037, 062, 068). The transmission of ascoviruses is unusual in that they are poorly infectious per os and appear to be transmitted among lepidopteran hosts by parasite wasp vectors at oviposition [7, 43] . The genome of the ascoviruses can be replicated in presence of polydnavirus DNA either within the reproductive tissues of female wasps or within the body of the parasitized hosts infected by both polydnavirus and ascovirus. Consequently, integrated sequences of ascovirus origin can be expected within wasp and polydnavirus genomes. Reciprocally, sequences of polydnavirus origin may have been integrated in ascovirus genomes, whatever the wasp origin, ichneumonid or braconid. One gene family related to a bacterial family of N-acetyl-L-glutamate 5-phosphotransferase (Acc. N° of the closest bacterial relatives YP_001354925, CAM32558, ZP_00944224, ZP_02006449), identified only within the SfAV1a, HvAV3e and TnAV2c genomes, supports this conclusion. It has been found in the genome of a bracovirus, Cotesia congregata BracoVirus (CcBV [13] ; Fig. 4 ). Since this gene is absent in the genome of Microplitis demolitor BV, a related bracovirus [8] , it is difficult to infer the direction of the lateral transfer between the common ancestors of the three ascoviruses and of the wasp C. congregata. However, they unambiguously indicate that there was at least one lateral transfer for this gene between the common ancestor of ascoviruses and the parasitic wasp. Since iridoviruses, like ascoviruses and other virus species [44, 45] , are, in some cases, vectored by parasitic wasps, databases were mined using all the available ichnovirus virus proteins as queries. We found no significant relationships between CsIV, HfIV and TrIV genomes and genomes of their putative closest relatives NCLDV and non-NCLDV relatives. This indicates that passive lateral gene transfers from virus to eukaryotes that are successfully spread and maintained in ichnovirus genomes remain rare events. One case of such lateral transfer was described in the CcBV genome. In this genome, aside from the presence of cardinal endogenous eukaryotic retrotranposon and Polintons that transposed in the chromosomal DNA of the proviral form of CcBV [46] [47] [48] , two genes encoding AcMNPV P94-related proteins, which have their closest relatives among granuloviruses (XcGV), were found. This suggests that CcBV contained at least two cases of lateral transfers between non-NCLDV and a bracovirus. Our results provide another source of evidence that passive lateral gene transfers have occurred regularly during evolution from bacteria to viruses and eukaryotes, and between viruses and eukaryotes [49] [50] [51] [52] . Even if the pox-D5 NTPase genes in the GfIV genome, and the MCP and SfAV061-like genes in the CsIV genome, indicate that they have an ascovirus origin, they provide only limited evidence supporting an ascovirus origin of ichnoviruses. Indeed, their sequence conservation and biological characteristics suggest that there were repeated lateral transfers during evolution between ascoviruses and wasp genomes, including the proviral ichnovirus loci. This raises an important issue about the role of lateral transfers during co-evolution of the NCLDVs and non-NCLDVs, ichnovirus, wasp and parasitized host. Indeed, genetic materials of various origins have been exchanged and maintained during co-evolution. This therefore suggests that ichnoviruses might be chimeric entities partly resulting from sev- Symbiogenesis was first proposed as an evolutionary mechanism when it became widely recognized that mitochondria and plastids originated from free-living prokaryotes [7] . The genomes of the endosymbiotic cyanobacteria and proteobacteria, respectively, at the origin of chloroplasts and mirochondria have evolved by reduction of several orders of magnitude to the approximate size of plasmids. Concurrently, nuclear genomes have been the recipients of plastid genomes. This relocation of the genes encoding most proteins of the endosymbiotic bacteria to the host nucleus is the ultimate step of this evolutionary process, so-called endosymbiogenesis [7, 53] . Recent studies of plants have revealed a constant deluge of DNA from organelles to the nucleus since the origin of organelles [54] . This allows the host cell to have the genetic control on its organelles, in a relationship that is closer to enslavement or domestication than to a symbiosis or a mutualism in which the organelles would recover benefits from their contribution to the eukaryotic cell well-being. To date, this deluge of DNA is considered to correspond to passive lateral transfers that result from the interactions between the life cycle of the organelle and nuclear replication. Numerous cases of symbiogenesis between endocellular bacteria and a wide variety of eukaryotic hosts have been characterized. However, recent work has demonstrated that this evolutionary process was not restricted to bacteria. It also occurred between endocellular eukaryotes such as unicellular algae and fungal endophyte in plants [55, 56] . Endosymbiogenesis was also proposed as the evolutionary mechanism that allowed some invertebrate viruses with a large double-stranded DNA genome related to the nudiviruses and the ascoviruses [22] , to have led, respectively, to the origin of bracoviruses and ichnoviruses, which are currently recognized as forming two genera within the family Polydnaviridae. Although presently there is no definitive evidence ruling out the hypothesis that the resemblance between ichnovirus and ascovirus virions is only an evolutionary convergence, the genomic differences between ascovirus and ichnoviruses are in good agreement with the symbiogenetic hypothesis. Indeed, they match an evolutionary scenario of endosymbiogenesis during which, from a single integration event of symbiotic virus genome, viral genes were lost and/or translocated from the provirus to other chromosomal regions (Fig. 5 ). In parallel, host genes of interest for the wasp parasitoid were integrated and diversified by selection and gene duplication in the proviral DNA. In this scenario, the more ancient symbiogenesis, the rarer the traces of genes from viral origin in the ichnovirus genome would be. This constitutes a constraint that dramatically limits the possibility to investigate the evolutionary links between ascovirus and ichnovirus. Results of our analyses demonstrate that the situation is also complicated by the fact that lateral gene transfers unrelated to the origin of ichnoviruses cause important misleading background noise. Moreover, the scenario in Figure 5 is close to a previously proposed version [57] , but is not consistent with results presented here, nor with recently accumulated knowledge on DNA transfer from organelles into the nucleus. Since endocellular environments favour lateral transfers between virus and wasp nucleus, it can be proposed that genes of virus origin that are involved in the ichnovirus biology were passively integrated in one or several loci, step by step over time, alone or through transfers of gene clusters, or even the entire viral genome. Since parasitoid wasps are able to vector different viruses [44, 45] , this second scenario opens the exciting possibility that virus genes involved in the ichnovirus biology might correspond to a gene patchwork resulting from transfers from viruses belonging to different NCLDV and non-NCLVD families. Because of the background noise due to lateral gene transfers found in these systems, elucidating the origins of ichnoviruses will be very time-consuming, requiring new accurate experimental approaches to generate more robust evidence. Sequencing wasp genomes to identify proteins of viral origin that are components of virions and involved in the assembly of these may well contribute to our understanding of how ichnoviruses and bracoviruses evolved from other insect DNA viruses. Searches for similarities were mainly developed using facilities of BLAST programs at two websites http:// www.ncbi.nlm.nih.gov/blast/Blast.cgi and http:genoweb.univ-rennes1.fr/Serveur-GPO/out ils.php3?id_rubrique=47. For DpAV4 genes having their origin within eukaryotic, bacterial or virus genomes belonging to NCLDV and non-NCLDV families, the closest gene was located using the distance trees supplied with each BLAST search at the NCBI website. Construction of syntactic models: Conserved amino acid blocks and positions described previously [15, 22] and with new data sets were verified or determined using MEME at http://meme.sdsc.edu/meme/meme.html. In the first step, we used motifs resulting from MEME to make MAST minings in databases at http:// meme.sdsc.edu/meme/mast.html. Since MEME motifs depend significantly on the data set use to calculate them, this approach did not enable an exhaustive detection of homologs among ascoviruses, iridoviruses, phycodnaviruses, mimiviruses and asfarviruses, and the detection sensitivity was ultimately very similar to that obtained with BLAST. To reach our detection objectives, we therefore constructed syntactic models that only included the most conserved positions and their variable spacing using WAPAM at the website. http://genoweb.univ-rennes1.fr/ Serveur-GPO/ outils_acces.php3?id_syndic=185&lang=en. Defining these models was obtained empirically until they allowed an exhaustive detection in refseq-protein and Genbank databases of the homologs among ascoviruses, iridoviruses, phycodnaviruses, mimiviruses and asfarviruses. The procedures were done until we were only able to detect exact match with the syntactic model. Whatever obtained with WAPAM, they required a confirmation with other approaches. Here, we used Psipred result comparison for regions with scores over 7 and HCA analyses for regions having scores lower than 7 with Psipred. This simplified the statistical treatment of the result obtained with WAPAM, since all exact matches have significance or a score of 100%. Syntactic Hypothetical mechanism for the integration and evolution of ascovirus genomes in endoparasitic wasps Figure 5 Hypothetical mechanism for the integration and evolution of ascovirus genomes in endoparasitic wasps. Schematic representation of the three-step process of symbiogenesis, and DNA rearrangements that putatively occurred in the germ line of the wasp ancestors in the Banchinae and Campopleginae lineages, from the integration of an ascoviral genome to the proviral ichnoviral genome. Sequences that originate from the ascovirus are in blue, those of the wasp host and its chromosomes are in pink. Genes of ascoviral origin are surrounded by a thin black or white line, depending on their final chromosomal location. Two solutions can account for the final chromosomal organisation of the proviral ichnovirus genome, monolocus or multilocus, since this question is not fully understood in either wasp lineage. More complex alternatives to this three-step process might also be proposed and would involve, for example, the complete de novo creation of a mono or multi locus proviral genome from the recruitment by recombination or transposition of ascoviral and host genes located elsewhere in the wasp chromosomes. This model for the chromosomal organization of proviral DNA in polydnaviruses is consistent with data recently published [58] .

Nasal Delivery of an Adenovirus-Based Vaccine Bypasses Pre-Existing Immunity to the Vaccine Carrier and Improves the Immune Response in Mice

Pre-existing immunity to human adenovirus serotype 5 (Ad5) is common in the general population. Bypassing pre-existing immunity could maximize Ad5 vaccine efficacy. Vaccination by the intramuscular (I.M.), nasal (I.N.) or oral (P.O.) route with Ad5 expressing Ebola Zaire glycoprotein (Ad5-ZGP) fully protected naïve mice against lethal challenge with Ebola. In the presence of pre-existing immunity, only mice vaccinated I.N. survived. The frequency of IFN-γ+ CD8+ T cells was reduced by 80% and by 15% in animals vaccinated by the I.M. and P.O. routes respectively. Neutralizing antibodies could not be detected in serum from either treatment group. Pre-existing immunity did not compromise the frequency of IFN-γ+ CD8+ T cells (3.9±1% naïve vs. 3.6±1% pre-existing immunity, PEI) nor anti-Ebola neutralizing antibody (NAB, 40±10 reciprocal dilution, both groups). The number of INF-γ+ CD8+ cells detected in bronchioalveolar lavage fluid (BAL) after I.N. immunization was not compromised by pre-existing immunity to Ad5 (146±14, naïve vs. 120±16 SFC/million MNCs, PEI). However, pre-existing immunity reduced NAB levels in BAL by ∼25% in this group. To improve the immune response after oral vaccination, the Ad5-based vaccine was PEGylated. Mice given the modified vaccine did not survive challenge and had reduced levels of IFN-γ+ CD8+ T cells 10 days after administration (0.3±0.3% PEG vs. 1.7±0.5% unmodified). PEGylation did increase NAB levels 2-fold. These results provide some insight about the degree of T and B cell mediated immunity necessary for protection against Ebola virus and suggest that modification of the virus capsid can influence the type of immune response elicited by an Ad5-based vaccine.

The ability of human adenoviruses to induce strong innate and adaptive immune responses makes them powerful adjuvants that facilitate the immune response against an encoded antigen. Recombinant adenoviruses have been shown to elicit significant immune responses to bacterial (anthrax, plague), viral (Hepatitis C, Rabies, SARS) and tumour-associated antigens [1] [2] [3] . While these results are encouraging, immunity eventually develops against virus capsid proteins. This severely reduces the immunogenicity of adenovirus-based vaccines in mice, [4] [5] [6] [7] [8] , primates [9] and humans [10] . This problem is also significant since a large portion of the Western world has marked levels of anti-adenovirus serotype 5 (Ad5) antibodies and is also prominent in regions of sub-Saharan Africa and Southeast Asia, where many of these vaccines are needed [11, 12] . Thus, assessment of the impact of pre-existing immunity on immune protection and alternative vaccination strategies may be needed for successful use of many adenovirusbased vaccines. Several strategies have been developed to address the prevalence of pre-existing immunity to Ad5 in the general population. Increasing the vector dose or adopting a prime-boost regimen in order to overcome pre-existing immunity to the virus is a common approach [2] . There is mixed enthusiasm for this plan, however, due to the documented toxicity associated with high doses of adenovirus and the length of time required for primeboost regimens when only a prime could be sufficient [13, 14] . 'Seroswitching', using recombinant adenoviruses constructed from chimpanzees or rare human serotypes with limited exposure rates such as adenoviruses 35 and 11 can elicit potent immune responses that are minimally affected by pre-existing immunity [7, 8, [15] [16] [17] [18] [19] [20] [21] . Hexon-chimeric adenoviruses can also avoid neutralization [22, 23] . Both approaches offer promise in the context of addressing pre-existing immunity, but require further investigation in response to concerns regarding safety and feasibility of largescale production. Covalent attachment of polyethylene glycol or incorporation of the virus into polymer matricies can also effectively protect Ad5 from neutralization [24] [25] [26] [27] [28] [29] [30] [31] . Delivery of Ad5-based vaccines by mucosal routes can also circumvent the effect of pre-existing immunity and induce a significant immune response against an encoded antigen [32] . Recombinant adenoviruses are one of the few well-studied vectors currently under development for vaccination against Ebola virus infection. The first protocol for an Ebola vaccine employed a prime-boost regimen consisting of naked DNA expressing either Ebola glycoprotein (GP) or nucleoprotein (NP) and recombinant Ad5 expressing Ebola GP to successfully protect non-human primates against a lethal challenge of Ebola [33] . This has since led to several Phase I clinical trials [34, 35] in which each component of the vaccine is administered by intramuscular injection. To date, there have been only two reports describing mucosal administration of an Ebola vaccine [36, 37] . Nasal administration of recombinant human parainfluenza virus type 3 (HPIV3) vectors expressing Ebola GP and/or NP to Guinea pigs and rhesus monkeys conferred complete protection against a lethal challenge with Ebola. We have previously found that a single dose of a recombinant adenovirus expressing Ebola Zaire GP given by either the oral or the nasal route is capable of affording protection against lethal challenge in naïve mice and that mucosal immunization can stimulate a broad, prolonged T cell-mediated immune response in both the systemic and mucosal compartments [37] . The primary objective of this study was to test the hypothesis that administration of an Ad5-based Ebola vaccine by either the nasal or oral route can circumvent pre-existing immunity and confer full protection upon challenge. Systemic and mucosal T and B cell responses to Ebola GP were assessed in both naïve mice and those with pre-existing immunity. The influence of PEGylation of the vaccine carrier on the immune response after oral immunization is also described. The E1/E3-deleted adenovirus vector expressing the Ebola Zaire glycoprotein was created by cloning the open reading frame (ORF) sequence of the glycoprotein in the plasmid pShuttle (Adeno-X Expression system I, BD Clonetech, Palo Alto, CA) for subsequent insertion in the E1 region of the human adenovirus serotype 5 genome. The human cytomegalovirus (CMV) promoter included in the Adeno-X expression system was used to drive the expression of the Ebola Zaire glycoprotein in the final recombinant adenovirus serotype 5 construct. Authenticity of the final product was confirmed by sequencing of the recombinant virus rescued by transfecting the linearized DNA into 293 cells. Virus was sequentially amplified to large-scale infections (5610 8 cells) and purified on an affinity column (Adeno-X virus purification mega kit, BD Clonetech, Mountain View, CA) according to the manufacturer's instructions. Genome structures of vectors were analyzed by restriction digestion of isolated viral DNA and compared with those of the original molecular clones. Particle number and infectivity of vectors were determined by standard optical density reading and immunodetection of the hexon protein, respectively, following infection of 293 cells with limiting dilutions of each vector preparation according to the recommendations by the manufacturer (Adeno-X rapid titer kit, Clontech, Mountain View, CA). Purified virus was administered in sterile phosphate buffered saline (pH 7.4) and had particle to plaque forming unit (pfu) ratios of 100:1 or less. First generation recombinant adenovirus expressing Ebola Zaire glycoprotein was prepared and purified as described above. The protein content of the virus preparation was determined using BioRad DC Protein Assay reagents (BioRad, Hercules, CA) and bovine serum albumin as a standard in a microplate format. According to established protocols, 10 mg of monomethoxypoly(ethylene) glycol, activated by tresyl chloride (Sigma Aldrich, St. Louis, MO), was added for each microgram of protein present [38] . The coupling reaction was performed at 25uC with gentle agitation. The reaction was stopped by the addition of L-lysine, in a 10-fold excess with respect to the amount of PEG added. Unreacted PEG, excess L-lysine, and reaction byproducts were removed by buffer exchange over a second Econo-Pac 10DG disposable chromatography column equilibrated with 100 mM potassium phosphate-buffered saline (pH 7.4). PEGylated preparations were administered in sterile potassium phosphate buffered saline (pH 7.4) and had particle to plaque forming unit (pfu) ratios of 100:1 or less as determined by the Adeno-X rapid titer kit. Characterization of these preparations revealed significant changes in biophysical properties of the virus once the reaction was complete such as the PEG-Dextran partition coefficient and peak elution times during capillary electrophoresis (as described previously [30, 38] ). Approximately 13,000 PEG molecules were associated with each virus particle in the studies outlined here as determined by a PEG-biotin assay [30] . B10.BR mice were immunized with 1610 10 particles of recombinant virus per mouse either by intramuscular injection (50 ml) in the right hindlimb, or by oral gavage (100 ml) using oral feeding needles (18G, 2.25 mm dia., Popper & Sons, Inc, New Hyde Park, NY). For nasal immunization, mice were anesthetized with isoflurane. Once anesthesia was achieved, 1610 10 particles of virus slowly delivered as a bolus into the nostrils using a standard micropipette (Gilson, Middleton, WI) as previously described [39] . Pre-existing immunity to adenovirus serotype 5 was established by injecting 5610 10 particles of adenovirus expressing betagalactosidase (AdlacZ) by intramuscular injection in the right hindlimb 30 days prior to vaccination with Ad5-ZGP. This protocol has been documented to activate T and B cells against virus capsid proteins and elicit humoral immunity [8, 11] . At the time of vaccination, mice had an average anti-adenovirus circulating NAB titer of 1:320, which falls within lower range of average values reported in humans after natural infection [11] . Mice were challenged by intraperitoneal injection of 2006LD 50 of mouse-adapted Ebola virus, Zaire strain (MA-ZEBOV) in 200 ml sterile saline [40] . After challenge, the animals were weighed daily for 13 days and monitored for clinical signs of Ebola infection using an approved scoring sheet. All procedures and the scoring method were approved by the Institutional Animal Care Committee at the National Microbiology Laboratory (NML) of the Public Health Agency of Canada (PHAC) according to the guidelines of the Canadian Council on Animal Care. All infectious work was performed in the 'Biosafety Level 4' (BSL4) facility at NML, PHAC. A) Anti-ebola neutralizing antibody (serum). Sera collected from immunized mice were inactivated at 56uC for 45 minutes. Serial dilutions of each sample (1:10, 1:20, 1:40, etc, in 50 ml of DMEM) were mixed with equal volumes of recombinant Ebola Zaire expressing the enhanced green fluorescent protein (EGFP) reporter gene (ZEBOV-EGFP, 100 transducing units/well, according to EGFP expression) and incubated at 37uC for 90 minutes [41] . The mixture was then added to subconfluent VeroE6 cells in 96-well flat-bottomed plates and incubated for 5-10 minutes at room temperature. Control wells were infected with equal amounts of either ZEBO-EGFP with media without serum or that containing non-immune serum. 100 ml of DMEM supplemented with 20% FBS was then added to each well and plates were incubated at 37uC in 5% CO 2 for 48 hr. Cells were subsequently fixed with 10% buffered formalin for 24 h and examined under a fluorescent microscope. Sample dilutions which showed .50% reduction in the number of green cells compared to controls scored positive for neutralizing antibody. B) Anti-ebola neutralizing antibody (mucosal). For the evaluation of the specific levels of IgG and IgA antibodies, bronchoalveolar lavage (BAL) fluid was collected in situ with a 20gauge catheter inserted into the proximal trachea, flushing the lower airways three times with 1 milliliter of L15 media (Sigma). BAL from each animal was incubated at 56uC for 45 minutes. Two-fold serial dilutions were added to 96 well plates pre-coated overnight with 30 ng of ZEBOV-like particles per well and incubated at 37uC for 1 hr. Goat anti-mouse secondary antibody conjugated to horseradish peroxidase (HRP) was then added and the plate was incubated for one additional hour at 37uC. The ABTS Peroxidase Substrate System (KPL) was used for detection and data collected using an ASYS UVM 340 ELISA plate reader (Isogen Life Science) at OD405. All infectious in vitro work was performed in the BSL4 laboratory at the NML, PHAC. C) Anti-adenovirus serotype 5 neutralizing antibody (serum). Pre-existing immunity against recombinant adenovirus 5 was assessed by determining the amount of neutralizing antibody present in serum according to established methods [30] . In brief, serum was incubated at 56uC for 30 minutes and then diluted in DMEM in twofold increments starting from a 1:20 dilution. Each dilution (100 ml) was mixed with an aliquot of a standard stock of adenovirus type 5 expressing E. coli beta-galactosidase (10 6 pfu), incubated for 1 hour at 37uC, and applied to HeLa cells in 96-well plates (2610 4 cells/well). One hundred microliters of DMEM supplemented with 20% FBS was then added to each well. Cells were incubated at 37uC for 24 hours. Neutralizing antibody titers were calculated as the highest dilution at which 50% of the cells stained blue by visual inspection. For the evaluation of INF-c positive CD8+ T cells 10 days postimmunization, splenocytes were harvested and cultured (1610 6 / sample) for 5 hours at 37uC in 96-well round bottom microtiter plates in DMEM supplemented with 10% FBS, 2-beta-mercaptoethanol (10 26 M) and GolgiStop (1 ml/ml, BD PharMingen, San Diego, CA). The TELRTFSI peptide which carries the Ebola Zaire GP immunodominant MHC class I epitope for mice of the H-2 k haplotype (B10.BR) was used for stimulation at a concentration of 1 mg/ml [42] . Control cells were treated with either an unrelated peptide or no peptide. After washing, cells were stained with 100 ml of a FITC-anti mouse CD8a antibody (1:100 dilution, PharMingen) at 4uC for 30 minutes. Cells were washed again, permeabilized in 16Cytofix/Cytoperm (PharMingen) for 20 minutes at 4uC, washed with 16Perm/Wash (PharMingen) and stained with 100 ml of a PEanti mouse IFN-c antibody (1:100 dilution, PharMingen) in the same buffer at 4uC for 30 minutes. Quantitation of INF-c positive CD8+ T cells isolated from splenocytes or mononuclear cells from bronchioalveolar lavage (BAL), mesenteric lymph nodes (MLN) and Peyer's patches (PP) 45 days after vaccination was performed using an ELISPOT assay (ELISPOT Mouse Set, BD PharMingen, San Diego, CA) according to the manufacturer's instructions. Briefly, a 96-well ELISPOT plate was coated with 5 mg/ml anti-mouse IFN-c capture antibody. Cells pooled from 4 B10.BR mice per experimental group were added to microwells along with the TELRTFSI peptide (2 mg/well). Control cells were incubated either without peptide or with the non-specific stimulator, SEB (200 ng/ml). After incubation with a biotinylated anti-mouse IFNc detection antibody and Streptavidin-horseradish peroxidase antibody, wells were counted using an ELISPOT reader (AID EliSpot reader system, Cell Technology, Colombia, MD). Data were analyzed for statistical significance by performing unpaired T tests (two-tailed p value) or one-way analysis of variance (ANOVA) when appropriate. The differences in the mean or raw values among treatment groups were considered significant when p,0.05. In an effort to correlate markers of immunity with protection against Ebola infection after mucosal immunization, T and B cell specific immune responses against Ebola glycoprotein were analyzed in mice in the presence or absence of pre-existing immunity (PEI) to adenovirus 10 days after vaccination with a first generation adenovirus serotype 5 vector expressing the Zaire Ebola glycoprotein (Ad5-ZGP). Intracellular staining and flow cytometry (FACS) revealed that Ebola glycoprotein peptide-specific activation of CD8+ T cells, as measured by production of IFN-c, occurred in naïve animals immunized by the nasal route at a frequency of 3.961% (I.N., Figure 1A ). Pre-existing immunity was induced by intramuscular administration of recombinant Ad5 expressing a nonrelevant antigen, beta-galactosidase (AdlacZ), 30 days prior to vaccination with Ad5-ZGP. At the time of vaccination, mice had an average anti-adenovirus circulating NAB titer of 1:320. Pre-existing immunity did not significantly alter activation of CD8+ T cells when the vaccine was given intranasally (I.N.+PEI, 3.661%, p = 0.07). Oral immunization lowered the response against Ebola glycoprotein (P.O., 260.5%). Samples obtained from animals with pre-existing immunity and vaccinated in the same manner were barely above the positive threshold (three times the average background level obtained from animals treated with the irrelevant virus, AdlacZ). Samples obtained from naïve animals immunized by the intramuscular route (I.M.) contained the largest population of IFN-c positive CD8+ T cells (1062%). Animals treated with the AdlacZ vector alone (AdlacZ-control) served as negative controls and produced 0.5% CD8+, IFN-c+ T cells in response to the Ebola glycoproteinspecific peptide. The B cell response against Ebola glycoprotein achieved by administration of the vaccine was determined by incubating a recombinant Ebola virus (Zaire strain) expressing green fluorescent protein (ZEBOV-EGFP) with serum collected 25 days after vaccination [8, 37] . Neutralizing antibody (NAB) levels equivalent to 40610 reciprocal dilution were detected in both naïve animals and those with pre-existing immunity after intranasal immunization ( Figure 1B) . Samples from naïve animals immunized by the oral route contained NAB levels equivalent to 2065 reciprocal dilution whereas anti-Ebola NAB could not be detected in samples obtained from animals immunized by the same route with preexisting immunity (P.O.+PEI). NAB levels of 80610 were detected in samples obtained from naïve mice immunized by the intramuscular route. Anti-Ebola NAB could not be detected in samples from mice given the AdlacZ vector alone (AdlacZ Control). Since the mucosa is often the primary sight of exposure, longterm, localized immune responses against Ebola virus are desirable for providing optimal protection against infection. In this context, T cell mediated immune responses were assessed by ELISPOT from various tissue compartments specific to the route of immunization 45 days after vaccination. The number of activated IFN-c secreting mononuclear cells harvested from splenocytes was significantly reduced by 38 and 59% in mice vaccinated nasally or orally when compared to those immunized by intramuscular injection (p#0.05, Figure 2A ). In contrast, samples obtained from the spleen of animals with pre-existing immunity that were immunized by the I.N. route contained the highest number of activated IFN-c secreting cells (8156190 spot-forming cells (SFC)/ million mononuclear cells (MNCs), Figure 2C ) although this was approximately 30% lower than that seen in naive animals (1,3406146 SFC/million MNCs, Figure 2A ). Pre-existing immunity also significantly reduced the number of these cells produced by animals immunized by the I.M. (210668 SFC/million MNCs) and P.O. routes (71612 SFC/million MNCs) with respect to those seen in naïve animals (2,145612 (I.M.) and 8856168 (P.O.)), p#0.05, Figure 2A) . A very limited number of activated IFN-c secreting mononuclear cells were detected in splenocytes of naïve animals immunized with the AdlacZ vector (50620 SFC/million MNCs, Figure 2A ) and those with pre-existing immunity to the same virus (1261, AdlacZ, Figure 2C ). Significant levels of INF-c positive cells were detected in bronchioalveolar lavage fluid (BAL) after intranasal immunization (146614 SFC/million MNCs, p#0.05, Figure 2B ). This was not significantly compromised by pre-existing immunity to adenovirus (120616 SFC/million MNCs, p#0.06, Figure 2D) . A limited number of INF-c secreting cells were detected in BAL from naïve animals immunized by the I.M. or P.O. routes (462 and 1165 SFC/million MNCs respectively, Figure 2B ) and were similar to that seen in samples obtained from animals given the AdlacZ vector (1261, negative control). Pre-existing immunity to adenovirus reduced these cell populations further to values that were not statistically significant (161 (I.M.), 562 (P.O.) and 261 (AdlacZ) SFC/million MNCs (p = 0.08, Figure 2D) ). INF-c positive cells were found in mesenteric lymph nodes (MLN) and Peyer's Patches (PP) (146611 and 2965 SFC/million MNCs, respectively) only after oral vaccination with Ad-ZGP ( Figure 2B ). Pre-existing immunity, however, reduced production of these cells to 18612 SFC/million MNCs (MLN) and 1662 SFC/million MNCs (PP) ( Figure 2D ). The NAB response was also monitored from bronchioalveolar lavage fluid 45 days post-vaccination. NAB to ZEBOV-EGFP was undetectable in samples obtained from control (AdlacZ) or I.M. immunized mice, whereas levels of 40610 and 1065 reciprocal dilution were detected in the BAL of nasally and orally vaccinated animals, respectively ( Figure 3 ). Those with pre-existing immunity and immunized by the nasal route had NAB levels of 30610. NAB was not detected in mice with pre-existing immunity to Ad5 and immunized by either the intramuscular or oral route. Further characterization of Ebola glycoprotein-specific immunoglobulin isotypes obtained from bronchioalveolar lavage fluid revealed that pre-existing immunity correlated with a marked decrease in the production of both IgG and IgA antibodies in mice immunized by intramuscular injection (Figure 4A and 4B ). IgG and IgA levels were the lowest in that were vaccinated orally regardless of whether they had pre-existing immunity ( Figure 4A and 4B). Immunization by the nasal route induced a strong IgG response that was reduced by an average of 25% in the presence of pre-existing immunity. The strongest IgA response was detected in samples obtained from animals given the vaccine by the nasal route. IgA levels, however, were also reduced by 25% in mice with pre-existing immunity to Ad5. The most direct means of evaluating vaccine efficacy in mice is to assess protection by monitoring weight loss and death rates after a lethal challenge of Ebola [43] . Therefore, mice were immunized with 1610 10 pre-existing immunity. At the time of vaccination, animals had an average anti-adenovirus neutralizing antibody titer of 1:320 reciprocal dilution. Naïve mice vaccinated by intramuscular injection with saline (vehicle) served as controls for complete lethality following challenge with mouse-adapted Ebola virus. Twenty-eight days after vaccination, animals were challenged with 200 LD 50 of mouse-adapted (MA)-ZEBOV. Only mice immunized by the nasal route survived the lethal challenge ( Figure 5A ). This was evident as early as 5 days after challenge when controls and mice immunized by the other routes began to lose weight ( Figure 5B ). All eventually expired 7 days post-challenge ( Figure 5A , some data not shown for clarity). In contrast, naïve mice given the vaccine survived challenge regardless of the route of administration. The studies outlined above strongly suggest that only intranasal immunization can successfully afford protection against Ebola virus in the absence and presence of pre-existing immunity. It was also clear that additional strategies were needed to improve vaccination by either the oral or intramuscular route in the presence of preexisting immunity. Covalent attachment of activated monomethoxypolyethylene glycol to the protein capsid of the Ad5-ZGP vector was evaluated as a potential way to improve the immune response after oral vaccination. Based upon previous observations, we hypothesized that this modification would protect the virus from neutralization by immune sera and improve survival in gastrointestinal tract [26, 27, 30, 44] . The frequency of Ebola glycoprotein peptide-specific activation of IFN-c positive CD8+ T cells was not significant in naïve animals given the PEGylated virus (0.360.3%) with respect to those given the unmodified virus (2.060.5%, p = 0.09, Figure 6A ). Similar results were also seen in the presence of pre-existing immunity in either treatment group (0.460.2% PEGylated vaccine, 1.760.5%, unmodified vaccine). In contrast, NAB levels equivalent to 30610 reciprocal dilution were detected in naïve animals given the PEGylated vaccine ( Figure 6B ). Samples obtained from animals given the unmodified virus had levels of 2065 reciprocal dilution. Animals with preexisting immunity against adenovirus and treated with the PEGylated preparation were also able to produce detectable levels of NAB (1065 reciprocal dilution) whereas NAB was not found in sera from mice given the unmodified vector orally. Further characterization of Ebola glycoprotein-specific Ig isotypes revealed that the PEGylated vector could possibly stimulate production of IgG in the presence or absence of pre-existing immunity with respect to levels found in animals given the unmodified vector ( Figure 6C ). Modification of the virus induced a slight increase of IgA in the presence of pre-existing immunity with respect to that seen in naïve animals given unmodified virus (Vaccine, Figure 6D ) and those with pre-existing immunity (Vaccine+PEI). Despite this, animals with pre-existing immunity to adenovirus and immunized with the PEGylated vaccine orally did not survive after challenge with 200 LD 50 of (MA)-ZEBOV. With mortality rates as high as 95%, Ebola infection occurs largely by direct contact with blood, tissues or skin of patients, and through mucosal exposure [45, 46] . To date, few efforts have been made to focus on the mucosa as the primary sight of exposure to Ebola virus and priming it for participation in the host defense by vaccination [47] . This is surprising given that the pulmonary, nasal and oral immune systems which comprise the mucosal-associated lymphoid tissues (MALT) are responsible for the production of approximately 80% of all immunocytes [48, 49] and mucosal immunity is often the first line of defense against pathogens coming in contact with susceptible hosts. Many studies in rodents indicate that systemic immunization produces strong anti-viral systemic responses while mucosal vaccination can stimulate both the mucosal and systemic immune systems and can confer long-term immunological memory against a given pathogen despite the fact that the magnitude of these responses are often reported to be somewhat reduced [37, 48, 50] . We have found this to be the case in the studies outlined here since the systemic cellular response and neutralizing antibody levels against Ebola Zaire GP were consistently lower following nasal and oral vaccination. Mucosal vaccination did, however, generate significant cellular and antibody responses in the periphery (BAL, MLN or Peyer's patches) and a single intranasal immunization with Ad5-ZGP conferred 100% protection even in the presence of pre-existing immunity. Administration of recombinant adenovirus-based vaccines to the mucosa has also conferred sufficient protection against challenge with a variety of pathogens in the presence of preexisting immunity to the vaccine carrier in mice and other preclinical models of disease [5, 19, 32, 51, 52] . This served as a basis for the present study. A single intranasal dose of a recombinant Ad5 vaccine expressing the Zaire Ebola glycoprotein conferred 100% protection in both naïve mice and those with pre-existing immunity despite the fact that the strength of the immune response generated by this route of administration was quantitatively lower than that seen in animals vaccinated by intramuscular injection. It is also important to note that pre-existing immunity induced by intramuscular injection did not severely compromise the T cell-mediated response at either the systemic or mucosal levels in these animals. The level of anti-Ebola GP antibodies in the circulation and in the lung was also not significantly compromised by pre-existing immunity to adenovirus. While these results suggest that intranasal vaccination with an Ad-based vaccine is indeed a promising strategy to overcome pre-existing immunity, one might wonder how accurately our results translate to natural exposure to the wild type virus via the respiratory tract. This is somewhat difficult to establish in the mouse model since the ability of the wild-type adenovirus to replicate is limited [53] . In addition, the amount of neutralizing antibody present in the nasal cavity of those in the general population with pre-existing immunity to adenovirus 5 has not been assessed to the degree that serum neutralizing antibodies have, making it difficult to set a relative parameter for one to target in pre-clinical animal models. Lack of this data may be due to the invasive nature of the technique for acquiring samples and/or the fact that antibody levels in the mucosa of individuals with established pre-existing immunity to adenovirus type 5 are quite low and transient in contrast to systemic levels of anti-adenovirus neutralizing antibodies which are quite robust and persist over time (unreported observations). Thus, for these studies, we decided to induce pre-existing immunity against adenovirus by intramuscular injection at a dose that has been shown to induce production of systemic neutralizing antibodies at 1:320, a level mostly below what was reported in humans with documented preexisting immunity [8, 11, 12] . Additional studies designed to assess the amount of neutralizing antibody to AdHu5 in the lung over time after intranasal administration of varying doses of AdHu5 are currently underway in an effort to further define stringent conditions under which pre-existing immunity can be established for experimental testing. Data obtained from animals vaccinated by the oral route provided several insights about the immunological requirements for protection against Ebola in a mouse model. Immunization by this route induced quantitatively lower T and B cell mediated immune responses against Ebola glycoprotein with respect to that achieved by either intramuscular or intranasal immunization. Despite this, every naïve animal immunized with a single oral dose of Ad-ZGP survived challenge, giving better precision on the minimal threshold of immunity required to achieve protection. As seen with intranasal immunization, the T cell-mediated response was not compromised by pre-existing immunity 10 days after oral vaccination. The number of IFN-c secreting T cells in both mucosal and systemic compartments of these animals at day 45, however, was significantly reduced by pre-existing immunity. It is possible that the peak T cell response at day 10 does not reflect the extent by which pre-existing immunity decreases the Ad5-ZGP-induced T cell response. Alternatively, it is also possible that the ELISPOT assay could detect subtle variations that flow cytometery could not monitor accurately due to differences in the sensitivity of each assay. Neutralizing antibody was not detected in the serum of animals with pre-existing immunity given a single oral dose of the vaccine. More importantly, none of these animals survived challenge with mouse adapted Ebola. We attempted to improve the immunogenicity of the adenovirus-based vaccine by protecting it from the harsh environment of the gastrointestinal tract and from neutralization by antiadenovirus antibodies by PEGylation. Interestingly, the T-cell mediated immune response was significantly reduced and antibody levels increased in naïve animals given a single oral dose of the PEGylated vaccine with respect to that seen in animals given the same dose of unmodified Ad5-ZGP. It has been shown that modification of virus capsids by PEGylation can significantly dampen the T-cell mediated immune response against the virus and stimulate the antibody response against a secreted antigen [26, 27, 30, 54] . Taken together, these data support the notion that the exposition of the virus capsid proteins facilitates the immune response against the encoded antigen. Optimization of PEGylation chemistries and/or densities on adenovirus-based vaccine that promote and strengthen protective immune responses following oral immunization is currently underway. Delivery of recombinant adenoviral vaccines to either the nasal or intestinal mucosa is an attractive vaccination strategy for many reasons. Vaccines administered in this manner will offer improved safety with respect to disease transmission and needle-stick injuries among health care workers, significant issues of concern in developing countries where the demand for many vaccines is high [55] . Mucosal administration of vaccines reduces the pain associated with vaccination, eliminates the need for specialized training programs for large vaccination campaigns and makes selfadministration of the vaccine possible. This route of administration may also significantly reduce systemic toxicity associated with recombinant adenovirus despite the fact that it has been shown that nasal immunization with recombinant adenovirus-based vaccines can facilitate translocation of the virus to the central nervous system [52] . Even though testing of other virus-based vaccines such as influenza have reported similar findings and are currently used in the clinic [56] , studies designed to fully assess the toxicological profile of adenovirus vaccine after nasal administration are also currently underway. We have shown that nasal immunization with an Ad5-based vaccine can induce a long-term protective immune response against Ebola virus in a mouse model which is not impeded by preexisting immunity to adenovirus serotype 5. While these results are extremely encouraging, further characterization of the immune response against both the encoded antigen and the adenovirus vector in larger, clinically relevant animal models is vital for both understanding the biology of Ad vaccines and for the development of an effective Ebola vaccine suitable to populations with different requirements [57, 58] . The issue of pre-existing immunity must also be adequately addressed in order to develop efficient recombinant adenovirus-based vaccines. While the majority of the literature suggests that pre-existing immunity significantly hamper the effective use of AdHu5 vaccine carriers, other investigators have reported that pre-existing immunity did not interfere with the potency of recombinant Ad5-based vaccines in both pre-clinical models of disease and in humans [59, 60] . Thus, additional studies identifying clinically relevant conditions under which to test Ad-based vaccine candidates are necessary to assess the full impact of pre-existing immunity on vaccine potency, including in different compartments. Better define the role of preexisting immunity on vaccine-induced immunity will further the understanding of how individuals previously exposed to adenovirus will respond to these immunization regimens.

Screening Pneumonia Patients for Mimivirus

Acanthamoeba polyphaga mimivirus (APM), a virus of free-living amebae, has reportedly caused human respiratory disease. Using 2 newly developed real-time PCR assays, we screened 496 respiratory specimens from 9 pneumonia-patient populations for APM. This virus was not detected in any specimen, which suggests it is not a common respiratory pathogen.

I nvestigation of a suspected Legionnaire's pneumonia outbreak in 1992 led to the isolation of a new microorganism from a water cooling tower in Bradford, England. This pathogen was thought to be a bacterium because it resembled small gram-positive cocci; however, in 2003 it was correctly identifi ed as a virus (1) . Acanthamoeba polyphaga mimivirus (APM), named for its ameba host and bacteria-mimicking characteristics, is a double-stranded DNA virus with the largest viral genome described to date (1.2 Mb) (2) . Mimiviridae is the newest member of the nucleocytoplasmic large DNA virus (NCLDV) group, which also contains Poxviridae, Iridoviridae, Asfarviridae, and Phycodnaviridae (1) . APM encodes specifi c translation proteins that are more commonly associated with cellular organisms than with viruses (2) . Other ameba-associated microorganisms from environmental sources, such as Legionella pneumophila, are known to cause outbreaks of acute pneumonia in immunosuppressed and elderly persons, although person-to-person transmission is uncommon. Whether APM is similarly responsible for individual cases or outbreaks of respiratory disease has yet to be conclusively determined. Previous studies have reported serologic evidence of APM infection in 7.1% to 9.7% of patients with community-or nosocomially acquired pneumonia (3, 4) . APM DNA was also amplifi ed by a nested PCR assay from a bronchoalveolar lavage specimen of a 60-year-old patient receiving intensive care for hospital-acquired pneumonia (3) . In this study, we used newly developed real-time PCR assays to screen pneumonia patients from a variety of epidemiologic settings for APM infections. Real-time PCR assays for APM were developed from multiple primers and probes designed for conserved re-gions of class I NCLDV genes L396 and R596, class III NCLDV gene L65, as well as the R656 gene, from the published APM genome sequence (GenBank accession no. NC_006450) by using Primer Express 3.0 software (Applied Biosystems, Foster City, CA, USA). All probes were labeled at the 5′ end with 6-carboxy-fl uorescein and quenched at the 3′ end with Black Hole Quencher-1 (Biosearch Technologies, Novato, CA, USA). Different primer and probe combinations were evaluated, and the 2 PCR assays that gave the best performance were selected for further studies (Table 1 ). Assays were performed by using the iQSupermix Kit (Bio-Rad, Hercules, CA, USA) in 25-μL reaction volumes. Amplifi cation was performed on an iCycler iQReal-Time Detection System (Bio-Rad) by using the following cycling conditions: 95ºC for 3 min for 1 cycle; 95ºC for 15 s and 55ºC for 1 min for 45 cycles each. Total nucleic acid was extracted from all specimens by using either the NucliSens Automated Extractor (bioMérieux, Boxtel, the Netherlands) or the automated BioRobot MDx (QIAGEN, Valencia, CA, USA) according to the manufacturers' instructions. Each clinical specimen was also tested for the human ribonuclease P gene to measure nucleic acid integrity as previously described (5) . For PCR-positive controls, recombinant plasmids containing APM DNA (kindly provided by Didier Raoult, Unite des Rickettsies, Universite de la Mediterranee, Marseille, France) were constructed. Primer pairs bracketing the L396 and R596 genes were used to amplify 1,560bp and 879-bp full gene regions, respectively, using 300 nmol/L of forward primers 396 F (5′-TTA ATC ATC TTC CAA AAA ATT TAA TTC-3′) and 596 F (5′-ATG TCG TTA TCA AAA CAA GTA GTT CC-3′), and 300 nmol/L of reverse primers 396 R (5′-ATG GCG AAC AAT ATT AAA ACT AAA A-3′) and 596 R (5′-CTA ATT TTC AAT ATA GTG CGT AGA TTC TA-3′). These PCR products were purifi ed by using the QIAquick Gel Extraction Kit (QIAGEN) and then cloned into a pCR-II TOPO vector by using a TOPO TA Cloning Kit (Invitrogen, Carlsbad, CA, USA). Recombinant plasmids were then isolated by using the QIAprep Spin Miniprep Kit (QIAGEN) and quantifi ed by UV spectroscopy. Standard curves were prepared from serial 10-fold dilutions of the quantifi ed plasmid in nuclease-free water containing 100 μg/mL of herring sperm DNA (Promega, Madison, WI, USA). The L396 and R596 real-time PCR assays could detect as few as 10 copies of plasmid DNA per reaction with amplifi cation effi ciencies of 99.6% [slope -3.33 and r 2 = 0.99] (Figure, left panels) and 99.2% [slope -3.34 and r 2 = 1.00] (Figure, right panels) , respectively. No amplifi cation was obtained by either assay with pooled total nucleic acid extracts from respiratory samples from healthy humans or from other common DNA respiratory viruses, including adenovirus, human bocavirus, or herpesviruses. The real-time PCR assays were used to test respiratory specimens from 496 pneumonia cases representing 9 distinct patient populations, which consisted of hospitalized pneumonia patients from population-based pneumonia surveillance studies in Thailand and the United States, transplant recipients with pneumonia, and isolated pneumonia outbreaks in either retirement homes for the elderly or familial clusters ( Table 2) . Of the 496 specimens tested, no positive results were obtained for APM DNA by either assay. We developed a rapid method of screening samples for APM DNA by using 2 sensitive and specifi c realtime PCR assays designed to target conserved NCLDV class I genes. With only 1 APM sequence published (NC_006450) (2), little is known of APM strain variation; therefore, use of assays that target different genes increases the likelihood that genetic variants of APM will not be missed. A suicide-nested PCR method for APM detection has been reported (3); however, the quicker turnaround time and lower risk for amplicon contamination makes the real-time PCR method more attractive for screening large numbers of samples. ACC TGA TCC ACA TCC CAT AAC TAA A Reverse Helicase GGC CTC ATC AAC AAA TGG TTT A seroprevalence study of APM among Canadian patients with community-acquired pneumonia identifi ed APM antibodies in 9.7% of 376 patients compared with 2.3% of 511 healthy controls (3). However, seropositivity may refl ect exposure to APM antigen rather than active infection, and the potential for nonspecifi c cross-reactions with the serologic assays used may have infl ated the true prevalence of APM infection (6) . In a separate report, a laboratoryacquired APM infection was linked to acute pneumonia by seroconversion in a technician in Marseille, France, thus providing evidence that this virus can occasionally cause clinical disease (7) . However, using sensitive real-time PCR assays, we failed to detect APM DNA in 496 respiratory specimens from 9 epidemiologically varied pneumonia patient populations. If we assume an APM prevalence of 0.2% (1 case in the study sample), the estimated probability of obtaining our results by chance, based on binomial analysis, would be 0.37. Most of the specimens we tested were from the upper respiratory tract, whereas the only reported APM PCR-positive sample was from a lower respiratory bronchoalveolar lavage specimen (3) . Moreover, the patient populations sampled may not represent those at highest risk for APM infection. Nevertheless, our study supports the fi ndings of an Austrian study that failed to detect APM in 214 nasopharyngeal specimens from hospitalized children with respiratory symptoms (8) . Our study did not detect APM in a large collection of specimens from patients with pneumonia, which indicates that this virus is not a common cause of severe acute respiratory disease. Because APM is an ameba-associated pathogen like Legionella, exposures to APM are most likely to occur from environmental sources. Further studies of more epidemiologically appropriate populations may be necessary to adequately access the importance of APM as a potential human respiratory pathogen. The real-time PCR assays described here will help facilitate these studies. Mr Dare is a microbiologist in the Division of Viral and Rickettsial Diseases at the Centers for Disease Control and Prevention. His research focuses on developing molecular diagnostic assays for respiratory viruses.

Resource Allocation during an Influenza Pandemic

Resource Allocation during an Influenza Pandemic

To the Editor: Planning for pandemic infl uenza is accepted as an essential healthcare service and has included creation of national and international antiviral drug stockpiles and novel approaches to emergency vaccine development (1) . The effectiveness of these strategies in a pandemic may be substantial but is unknown. More certain is that effective management of severe and complicated infl uenza will reduce deaths and that demand will exceed available treatment resources (2) . Appropriate allocation of treatment resources is therefore essential, perhaps more important than any specifi c treatment such as administering antiviral medication to symptomatic patients. Re- (4) . Even more important for most severely ill patients, however, will be deciding whether to admit them to the hospital at all. The UK pandemic-planning criteria currently recommend a scoring system for hospital admission based on an assessment of poor outcome rather than on capacity to benefi t (2). Indeed, age >85 years and severe underlying cognitive impairment, which would rule out admission to critical care in Canada, would strongly favor admission to hospital care in the United Kingdom, the opposite of the situation for a younger cognitively intact person with similar disease severity. If tools are to be developed to support triage at all stages of the patient pathway in a pandemic, societies must consider the ethical issues raised (4,5), debate them, and take a position on the values that should underpin decision making in a pandemic. Even when clear societal goals are established, much work remains to ensure that the healthcare community is equipped to steer healthcare resources to deliver these effectively (6) . Community-acquired pneumonia has been used as a surrogate for infl uenza to test predictive scoring systems for assessing severity and assisting triage decisions (7) . Seasonal infl uenza epidemics would provide the most realistic setting available, in particular, if protocols were in place to test criteria when a relatively severe infl uenza season occurs. In addition to identifying criteria for setting priorities within infl uenza management, such testing will need to consider the balance of resources between infl uenza treatment and treatment of other usual noninfl uenza conditions that will require emergency care during the pandemic. Decisions that must be made during a pandemic are complex, varying from when to stop major elective surgery so critical care capacity can be opened up, to how to triage those who have experienced major trauma and those with infl uenza. These decisions could differ from those same decisions made outside a pandemic, and an adequate evidence base is needed if they are to be of good quality. The third component of our preparation for optimally deploying standard care in a pandemic is being able to change our approach quickly as new knowledge emerges. In the so-called Spanish infl uenza pandemic of 1918-19, the unfamiliar clinical course meant that infl uenza was not even considered when the fi rst cases appeared (8) , and expectations had to be revised concerning who was most vulnerable and at what stage in their clinical course they were most at risk. Therefore, healthcare professionals must develop and test the public health infrastructure to capture patient factors associated with outcome and treatment response during a pandemic and feed this information back into clinical practice rapidly and reliably, as occurred during the epidemic of severe acute respiratory syndrome (9) . International collaboration will be important for sharing this work (10) and developing useful tools early in a pandemic. Having recognized the risk for pandemic infl uenza, we must now complement the research into novel infl uenza treatments by addressing our knowledge gap on how best to use our resources to deliver optimal clinical care in the management of infl uenza guided by effective clinical surveillance. To the Editor: Tick-borne relapsing fever in western North America is a zoonosis caused by spirochetes in the genus Borrelia that are transmitted by argasid ticks of the genus Ornithodoros (1) . Human disease occurs in many focal areas and is associated with infections of Borrelia hermsii, B. turicatae, and possibly B. parkeri (2, 3) . Although the ecologic parameters that maintain B. hermsii and B. turicatae differ, human infections usually occur in rustic cabins (B. hermsii) and caves (B. turicatae) inhabited by ticks and their terrestrial vertebrate hosts (1) . Recently, Gill et al. (4) provided evidence that the argasid bat tick, Carios kelleyi, feeds upon humans. Subsequently, Loftis et al. (5) used PCR analysis and DNA sequencing to detect in C. kelleyi an unidentifi ed Borrelia species that was closely related to B. turicatae and B. parkeri. We report the partial molecular char-acterization of another novel tickborne relapsing fever spirochete in C. kelleyi, which expands our knowledge for this group of pathogenic spirochetes and their potential vertebrate hosts and tick vectors. C. kelleyi were collected August 18, 2005, from a house in Jones County, Iowa, built in 1857. Bats had been excluded from the attic since 1992. Nine months before ticks were collected, bats were prevented from roosting under the eaves. DNA was extracted from 31 nymphal C. kelleyi, as described previously (6) . For each tick, regions of the glpQ, fl aB, and 16S rRNA genes were amplifi ed and sequenced as described (3, 7, 8) . Sequences were assembled by using the SeqMan program in the Lasergene software package (DNASTAR, Madison, WI, USA). Fourteen (45.1%) of 31 ticks were positive by PCR for >1 of the genes tested. Partial DNA sequences were determined from tick no. 16, for which amplicons for all 3 genes were obtained. The partial fl aB sequence had 4 bases different from the 300-base sequence (98.66% identity) reported previously (GenBank accession no. AY763104) for another Borrelia sp. found in C. kelleyi (5) . We constructed a 1,992-bp concatenated sequence that contained 1,273 bp of the 16S rRNA, 351 bp of fl aB, and 368 bp of glpQ. This concatenated sequence was aligned with homologous, trimmed DNA sequences of the same length obtained from representative full-length sequences determined previously for B. hermsii, B. turicatae, and B. parkeri (3, 9) (Figure) . This C. kelleyi spirochete was more closely related to B. turicatae and B. parkeri than to B. hermsii but was clearly distinct from all 3 species (DNA sequence identities of 98.89%, 98.75%, and 95.98% to B. turicatae, B. parkeri, and B. hermsii, respectively). A glpQ amplicon from another nymphal tick (no. 3) was sequenced (GenBank accession no. EF688578) and was unique in the database; it was also considerably different from the glpQ sequence determined from tick 16, with 325 of 368 bases matching (88.3% identity). The Borrelia glpQ sequence from tick 3 had 85.1%-89.1% identity compared with glpQ sequences from B. hermsii, B. turicatae, and B. parkeri. This fi nding suggests the presence of at least 2 relapsing fever group spirochetes in C. kelleyi that await further characterization. We found a novel Borrelia in bat ticks that is closely related to, but distinct from, the other known species of tick-borne relapsing fever spirochetes in North America. The human health implications of the new relapsing fever group spirochete are not yet known. The willingness of C. kelleyi to feed on humans and the fact that infection with bacteria closely related to true relapsing fever spirochetes occurs in

Factors influencing psychological distress during a disease epidemic: Data from Australia's first outbreak of equine influenza

BACKGROUND: In 2007 Australia experienced its first outbreak of highly infectious equine influenza. Government disease control measures were put in place to control, contain, and eradicate the disease; these measures included movement restrictions and quarantining of properties. This study was conducted to assess the psycho-social impacts of this disease, and this paper reports the prevalence of, and factors influencing, psychological distress during this outbreak. METHODS: Data were collected using an online survey, with a link directed to the affected population via a number of industry groups. Psychological distress, as determined by the Kessler 10 Psychological Distress Scale, was the main outcome measure. RESULTS: In total, 2760 people participated in this study. Extremely high levels of non-specific psychological distress were reported by respondents in this study, with 34% reporting high psychological distress (K10 > 22), compared to levels of around 12% in the Australian general population. Analysis, using backward stepwise binary logistic regression analysis, revealed that those living in high risk infection (red) zones (OR = 2.00; 95% CI: 1.57–2.55; p < 0.001) and disease buffer (amber) zones (OR = 1.83; 95% CI: 1.36–2.46; p < 0.001) were at much greater risk of high psychological distress than those living in uninfected (white zones). Although prevalence of high psychological distress was greater in infected EI zones and States, elevated levels of psychological distress were experienced in horse-owners nationally. Statistical analysis indicated that certain groups were more vulnerable to high psychological distress; specifically younger people, and those with lower levels of formal educational qualifications. Respondents whose principal source of income was from horse-related industry were more than twice as likely to have high psychological distress than those whose primary source of income was not linked to horse-related industry (OR = 2.23; 95% CI: 1.82–2.73; p < 0.001). CONCLUSION: Although, methodologically, this study had good internal validity, it has limited generalisability because it was not possible to identify, bound, or sample the target population accurately. However, this study is the first to collect psychological distress data from an affected population during such a disease outbreak and has potential to inform those involved in assessing the potential psychological impacts of human infectious diseases, such as pandemic influenza.

Equine influenza (EI) is an acute, highly contagious viral disease which can cause rapidly spreading outbreaks of respiratory disease in horses and other equine species. It does not infect humans, but the virus can be physically carried on skin, hair, clothing, shoes, vehicles and equipment and through these means can be transferred to other horses. In addition, the windborne virus can be spread for distances up to eight kilometres [1] . Australia's first outbreak of EI was confirmed on August 24 th 2007. It spread quickly, but was successfully contained within areas of South East Queensland (Qld) and New South Wales (NSW). Although EI was not detected in other States and Territories, stringent disease control procedures were put in place across all States; which included an initial stand-still of all horse movements and subsequent controls, movement restrictions, and biosecurity requirements for many months. Colour-coded EI control zones were established within four weeks of the outbreak based on the level of disease/disease risk in Local Government Areas in NSW and Qld; these were adjusted as the disease spread, and each zone was subject to specific controls and restrictions. Controls were reviewed, revised and expanded as the disease spread, subsequent disease containment and control progressed, and policies were revised. These zones are summarized in Table 1 . Further details of the outbreak, restrictions and zoning are available via the NSW Department of Primary Industries (NSW DPI) and Qld. Department of Primary Industries and Fisheries (DPI&F) websites [2, 3] . Throughout the outbreak movement restrictions and biosecurity requirements remained in place, and no (or very limited) horse movement was ever allowed from higher risk zones to lower risk zones. The disease outbreak peaked in late September/early October 2007, and then declined as successful containment and eradication strategies were progressed. The last new infections of EI were reported in NSW and Qld in December 2007. In total approximately 6,000 properties and 47,000 horses were infected in NSW and at least 3,000 properties were infected in Qld. Current data from disease surveillance and monitoring indicates that no active infection is present in Australia and the expectation is that Australia will be declared EI-free by the end of June 2008; if successful, Australia will be the only EI infected country in the world to have eradicated the disease. The effects of EI and the disease containment strategy, like the horse industry itself, were varied and wide-ranging; impacting differentially on horse owners and those involved with the horse industry nationally. In terms of support to those affected, a range of government financial support and assistance was available to many of those affected within a short time of outbreak onset and financial and economic impact surveys were undertaken to provide feedback information to government [4, 5] . The current study was conducted to gain additional complementary data to assess the impacts of EI on the social and emotional health and well-being of those affected. This paper reports data collected on non-specific psychological distress; however the full study covered many other aspects, such as adherence to biosecurity requirements, effects of social isolation due to quarantine and the consequences of restricted horse movement and related activities, and sources of support and coping during the EI outbreak. Although EI is endemic in Europe and North America, and has occurred as an epidemic in many other countries, e.g. Japan, South Africa, Hong Kong, there does not appear to be any published studies of the human response or impacts to EI or the containment strategies used to control this disease. The best reported and documented research with respect to the impacts of infectious animal disease on people is the outbreaks of foot and mouth disease (FMD) in Europe in 2001, specifically in the UK and The Netherlands. Like EI FMD is highly contagious, however, FMD is considerably more serious as it spreads to cloven-hoofed animals including cattle, sheep, pigs, and goats. During these FMD outbreaks an estimated 4 million livestock were slaughtered on 9,000 farms in the UK (including many healthy animals as part of 'contiguous' or preventative culling on farms neighbouring infected farms) and 270,000 were culled in The Netherlands. The impacts on people were both economic, through financial/business/ tourism-related losses, and psychological, through the exposure to loss of livestock, culling, and massive funeral pyres; the latter affecting not just farmers and their families, but also the wider population through media images on the television and in newspapers [6] [7] [8] . In the UK higher 'caseness' as indicated by the GHQ(G) was found in farmers from 'badly infected' areas, although higher psychological morbidity generally, was reported in farmers from both badly infected and unaffected areas [9] . In a study of Dutch dairy farmers [10] around half of those whose animals were culled suffered from severe post-traumatic distress, (identified as a clinical level of distress (> 25) using the 15-item Impact of Events Scale), with this reducing to one in five for those where severe restrictions were imposed (but where no culling took place). Higher levels of symptoms were reported for older respondents and those with lower levels of education. In this same study differences in stress, psychological marginalization, and depression were reported for different disease control areas, i.e. culled-area, buffer-area, FMD-free area [11] . Within Australia, the psycho-social impacts of Ovine Johne's disease have been reported [12, 8] in which grief, depression, and anxiety were profound in affected farming families, and the perceptions of the management control process were the cause of much of the distress. Government policies on quarantining and de-stocking farms were suspended due to mounting reports of severe emotional and social distress in farmers, rural families, and government employees implementing those policies. Further discussion of stress in emergency responders managing agricultural emergencies is considered in an Australian context in a recent paper by Jenner [13] . The role of the animal-human bond on disaster preparedness and response is key feature in human response to animal disease, and has been review by Hall et al. [8] . These authors report several aspects of relevance to the current study, including the increasing role of horses as companion animals as opposed to livestock or economic investments, and hence an increasing emotional attachment to horses; the complex and dynamic emotional relationship between farmers and their livestock; the emotional and practical implications of the animal-human relationship in disaster management, e.g. compliance with disaster management behaviours; and the impacts on veterinarians as first responders in disasters. These authors conclude that recognizing the mental health aspects of the animalhuman bond is an important factor in public health approaches to disaster and can be critical in promoting the resilience of individuals and communities. Therefore, it follows that in an animal-centred disease outbreak, such as EI affecting horses, the potential disruption of the animal-human bond, and the impact of policies restricting animal-human activities could have significant implications for the mental health and resilience of those affected. The main outcome measure in this study is non-specific psychological distress, as measured by the Kessler 10 (K10) [14] . The K10 was selected because it is a well-established and validated measure that is used widely in population research in Australia, it has been used in population health surveys in NSW [15] , Victoria [16] , South Australia [17] , and Western Australia [18] , as well as in National surveys conducted by the Australian Bureau of Statistics [19] , and therefore State and National prevalence data are available as benchmarks for the current study. Scores from the K10 can be related to levels of intervention, with 'very high' psychological distress scores (> 30) equating to 'caseness' for a mental disorder, and high scores are strongly associated with current diagnosis of anxiety and depression using the Composite International Diagnostics Interview (CIDI) [20] . The K10 is also able to discriminate between DSM-IV cases and non-cases, and is felt to be an appropriate screening instrument for identifying likely cases of anxiety and depression in the population providing a strong marker for a possible mental health disorder [21] [22] [23] . In the most recent (2007) data from the NSW Adult Population Health Survey the combined proportion of the population reporting 'high' or above psychological distress (22-50) is 12.1% [24] . In addition, recent data collected in rural communities suggests that these figures may be slightly higher in rural-dwellers with 'very high' psychological distress of 5% reported in one study [25] and 13.4-13.8% for combined 'high'/'very high' psychological distress in another [26] . These findings are of relevance in the current study as it would be expected that horse-ownership would be linked to rural and peri-urban residency. The questionnaire was designed for online completion to expedite data gathering whilst the EI outbreak was occurring. Questionnaire content was reviewed by subject matter experts, including a small group of public health professionals in NSW Health, some of whom had been involved in aiding the NSW DPI in disease control management, a NSW DPI Local District Control Centre Controller who was responsible for leading control activities, and representatives of the Australian Horse Industry Council (AHIC). Ethics approval for the study was obtained through the University of Western Sydney ethics committee. Horse owners, and those involved in the horse industry were invited to take part in the study via an e-mail alerting service administered by AHIC; using the national Horse Emergency Contact Database (HECD). The HECD had been established before the EI outbreak and was used as a network to contact and inform horse-owners during emergencies, such as bushfires, and disease outbreaks, and had been used previously by AHIC for collecting financial impacts information relating to EI earlier in the outbreak. This alerting service was used regularly during the EI crisis to update registrants with government support agency communications and general industry news and support information. Approximately 8,000 addressees were registered on the HECD; most were individuals, but also included were industry associations, pony clubs, and horse groups that would forward information to their own memberships nationally. Horse owners in NSW were encouraged by the NSW DPI to register on the HECD to receive up to date information. The initial invitation to participate was sent to those registered on the HECD on 14 November 2007 (Week 12 of the outbreak). The survey remained open until 7 January 2008 (Week 21 of the outbreak) and date of completion was recorded with each respondent's data. The full survey comprised 166 questions, covering a wide range of subject areas; those reported here include demographic information, i.e. gender, age category, number of children, highest level of educational qualification, and State/Territory of residence. In addition, respondents were asked about the nature of their current main involvement with horses (i.e. their industry sector), for example breeding, equestrian, recreational; whether their primary source of income was linked to a horse-related industry, and their current colour-coded EI control zone. The main outcome measure reported in this paper is nonspecific psychological distress as measured by the K10. This measure comprises 10 questions that ask respondents how often they have experienced certain symptoms during the preceding four weeks and responses are scored on a scale of 1 to 5 depending on how frequently each symptom is experienced, where 1 = 'none of the time', and 5 = 'all of the time'. Thus, a minimum score is 10, indicating no psychological distress, and a maximum score is 50, indicating the most severe level of psychological distress. Scores on the K10 are subsequently categorized into four levels: low (scores of 10-15); moderate (scores of [16] [17] [18] [19] [20] [21] ; 'high' (scores of [22] [23] [24] [25] [26] [27] [28] [29] and 'very high' (scores of 30-50) [27] . Statistical analyses were undertaken using STATA, version 9.2 (2004; Stata Corporation, College Station, TX, USA). Exploratory data analysis was conducted using frequency distributions for categorical variables. In the logistic model, a binary coding of psychological distress was used in which high psychological distress was a combination of 'high' + 'very high' levels of psychological distress = 1 (i.e. K10 scores of 22 or greater) and low psychological distress was a combination of 'low' + 'moderate' levels of psychological distress = 0 (i.e. K10 scores of 21 or less). Simple binary logistic regression and backward stepwise multiple logistic analyses were performed to identify factors influencing high psychological distress. All variables were entered into the model initially, with the least significant variables removed one at a time until only significant variables associated with values of p ≤ 0.05 remained. All statistical tests were two-tailed. Details of the study sample are presented in Table 2 . In total, 2,760 respondents completed the online survey, and of these 15% were male and 84% were female. More than a half of the sample (58.9%) had no children. A total of 40.2% of the respondents had a tertiary level educational qualification. Just under half the sample (47%) was from NSW and respondents from Qld. and Victoria (Vic) comprised a further 40% of the sample (20% from each State). Thirty percent of respondents were in uninfected white zones in States other than NSW and Qld, and 22% were from the restricted high EI risk red zones in NSW and Qld. Around three quarters of the sample (73%) were from three industry sectors; recreational, equestrian, and breeding/stud sectors (30%, 27%, and 16%, respectively). The majority of respondents (76%) reported that their main source of income was not linked to a horse-related industry. The prevalence of the four levels of psychological distress for the whole sample during the equine influenza outbreak; were 39% of respondents reporting 'low', 27% reporting 'moderate' 20% reporting 'high' and 14% reporting 'very high' levels of psychological distress. Table 3 shows the proportion of respondents reporting each level of psychological distress for the main socio-demographic survey variables. The greatest prevalence of 'very high' psychological distress was reported for those respondents in the 16-24 age group (21.2%), and the lowest prevalence was reported by those in the 55-64 age group and those under 16 (8.6% and 7.4%, respectively). With regard to the remaining socio-demographic variables the highest prevalence of 'very high' psychological distress were recorded for those respondents who were female, those with one child, and those with no formal educational qualifications. The prevalence of 'very high' psychological distress was greater for respondents from Qld. (19.4%) with prevalence figures being slightly lower for respondents from NSW (14.5%) and lower again for respondents from Vic. (10.8%). The highest prevalence of 'very high' psychological distress was found for respondents in the red zones (18.2%) and lowest for those in the white zones (9.3%). Those whose incomes were linked to horse-related industry had a higher prevalence of 'very high' psychological distress as compared to those whose main income was not linked to a horse-related industry (20.7% and 11.8%, respectively). The four levels of psychological distress were combined in pairs ('low'/'moderate', and 'high'/'very high') to form a binary variable for subsequent statistical modelling. Figure 1 shows the prevalence of this binary high/low psychological distress variable by EI disease zones. Respondents in the red and amber zones reported higher prevalence of high psychological distress (41% and 39%, respectively) than those in the purple, green, and white zones (36%, 34%, and 26% respectively). Univariate analysis Table 4 shows the unadjusted and adjusted odds ratios (ORs) for the associations between high psychological distress (≥ 22) and socio-demographic variables. Total count = 2760 unless otherwise given in brackets Respondents whose main source of income was from horse-related industry (unadjusted: OR = 2.19, 95% CI: 1.80-2.67; p < 0.001) were at a greater risk of high psychological distress than those whose main income was not linked to horse-related industry. chological distress as compared to those whose income was not linked to horse-related industry. The most salient finding was the extremely high prevalence of high psychological distress in horse owners and those involved in the horse industry during a serious horse disease epidemic; with just over one third (34%) reporting levels of psychological distress that might require some form of external intervention, and 40% of these (14% of the sample) reaching levels that may be considered indicative of 'caseness' for a DSM-IV disorder. The prevalence of 'very high' psychological distress in this sample was approaching five times the level reported in recent population health data for NSW [24] . Although this prevalence is very high, and there are some methodological reasons why this may be distorted (see study limitations section) it is certainly true that many of those impacted by EI, or the threat of EI, were subject to a wide range of acute stressors over a prolonged period, in a country where EI and such rigorous disease containment and control measures were previously unknown. Analysis of psychological distress prevalence within the sample indicated that EI control zone was associated with psychological distress. Those in the areas where EI was present had higher risk of high psychological distress, furthermore, risks were higher in areas where EI was more active or threatening and the tightest levels of disease control were in place (i.e. red and amber/buffer zones). This finding suggests high levels of anticipatory anxiety. Interestingly, the risks of higher psychological distress in the purple zone (the region in NSW with the highest infection rate and earliest infections) were lower than in the red and amber areas. It is probable that during the timing of the study EI was more of a 'known' threat to those in the purple zone and there would have been some habituation to this risk; with many properties already infected or recovering, and restrictions eased due to the decision to let EI 'run its course' in this area at that time. As disease control (and zoning) was controlled at a State level there is geographical overlap and co-linearity of the Australian State/Territory and EI control zone variables in the analysis; in the backward stepwise multiple logistic analyses excluding one made the other a significant factor. With regard to analysis by State, it is interesting to note the high levels of psychological distress reported in Victoria. Although Victoria remained EI-free throughout the crisis, those in Victoria were 1.57 times more likely to experience high psychological distress that those in the other uninfected States. There are probably a number of reasons for this effect: Victoria has a very extensive horse industry and is geographical closer to the infected States and diseaseaffected areas of NSW and Qld; there is also a high level of business interaction and physical movement of horses between Victoria and NSW and Qld; hence the level of proximal threat and the degree of disruption caused by disease control measures was probably experienced more widely in Victoria and may explain some of this effect. It should also be noted that although the remaining States were similarly uninfected, the overall prevalence of high psychological distress in horse owners from these States was still far higher than in the general population; those uninfected were not unaffected. One of the other primary factors associated with high psychological distress was age. Those in the 16-24 year age category reported the highest levels of high psychological distress and analysis indicated that although prevalence and comparative risks of high psychological distress reduced from age 24 onwards, these reduced risks only became reliably statistically significant from age 45 onwards, and high psychological distress was certainly still a risk to those in the 35-44 year age category. This is interesting because in the general population psychological distress is generally found to peak around middle age (40s-50s). The study findings would suggest that younger people were particularly vulnerable and were coping less well with the consequences of EI. The reasons for this finding are not known, however, research literature suggests that younger people form stronger emotional attachments to animals [28] , and they are also less likely to be resilient or practised, generally, when it comes to coping with adversity. From the general perspective of mental and physical health of younger people, it is interesting to consider the longer term consequences and potential burden of disease if these effects are enduring. It is also interesting to note here the association of psychological distress with having children. Data in this study indicated that those with one child had a 1.2 times higher risk of high psychological distress than those with no children; and having three or more children appeared somewhat protective against high psychological distress. National statistics would support the suggestion that those with one child are generally younger adults and/or are 'young families' with a single younger child. In this study, 17.8% of respondents with one child reported 'very high' levels of psychological distress (K10 score = 30-50). Given these family circumstances such a finding may be a cause for concern. The final main factor associated with high psychological distress was having an income linked to horse-related industry. Unsurprisingly, those with financial dependence on an industry facing such a crisis are likely to be significantly predisposed to high psychological distress. Nothing has been mentioned in this paper on the industry sector from which respondents had their main involvement with horses. These data were reported as part of the sample description to illustrate the wide range of industries affected by EI and the complexity of the affected population, and to provide information to aid interpretation of the findings. The nature of the potential psychological impacts of EI on those in different sectors is extremely diverse; from purely economic impacts, to loss of leisure pursuits and disruption of social networks, to loss of futures and missed opportunities in time, and many other possible impacts. Time, money and support will help most recover but it is possible that some people's mental and physical health will be permanently affected by EI and some will take many years to recover professionally if they choose to stay in these professions. Given the level of psychological distress noted in the current study, it is interesting to consider the distress that might result from other epizootics, such as foot and mouth disease or avian influenza, and how this, in turn, might compare to the levels of distress resulting from human epidemics, such as SARS and H5N1/pandemic influenza. As mentioned earlier, foot and mouth disease in Europe resulted in high distress and PTSD in farmers. In relation to avian influenza, most research has focussed on risk perception and compliance with protective behaviours. A large European Union project on risk perception to avian influenza in Europe and Asia found moderate levels of risk perception generally, with higher levels of risk perception noted in Europe, and in females in most countries [29] . Considering distress and risk perception in relation to human epidemics; it is likely that psychological distress would be far greater, since these present a threat to human health and possibly death. Certainly data collected during and after SARS in Hong Kong found high levels of fear and PTSD in health care workers and hospital workers [30] , and high levels of emotional disturbance in the general population [31] . Research in Canada found enduring psychological distress, up to two years following SARS, among health care workers in a hospital that treated SARS patients [32] . This study had a number of limitations that should be considered when interpreting the data. Firstly, the target population; those affected by EI, is a complex, disparate, and unknown population and therefore it is difficult to comment accurately on the representativeness of the sample. All horse owners in Australia are not registered on a centralized database, or otherwise controlled, and as a result, it is not possible to know how extensive the database used to access horse owners (the HECD) was. However, at a national cross-industry sector level it is believed that this was the most extensive and efficient online route to access the target population, and the use of the database as a central communication facility during the EI crisis meant that this was likely to have been a focus for those affected during the epidemic. Due to demographic bias in the sample, in particular, a greater proportion of women, and those with higher levels of education it is possible that there will be response bias in the data. The high proportion of women in the sample may be due to greater interest and participation in studies of this nature, but may also be indicative of higher levels of females in the target population, in particular in the main industry sectors represented in the data, i.e. recreational and equestrian. There are no official statistics on gender breakdown across horse industry sub-populations in Australia, but data indicate that the equestrian sector in the United States may comprise 80% women [33] , so the gender bias may reflect a gender bias in the main industry sub-populations in our data. Research data often report higher levels of psychological distress in women in the general population, and therefore, the gender demographic bias in our study might have led to an elevation in the levels of psychological distress reported in this study. However, the absence of a significant gender effect in this study, and the close matching of relative levels of psychological distress in men and women with data from the Australian general population, suggests that EI, as an adversity, was exerting similar impacts on males and females. It is not possible to explain why there was an absence of a gender difference in the data. One possible explanation is that the timing of the study; around the height of the EI epidemic, and the high levels of psycho-logical distress generally, reflected peak, acute levels in which gender differences were minimised and insignificant. As with gender bias, it is hard to define the impacts of education level in the data. Unlike the (female) gender bias in the data, higher levels of education offer a protective effect (as identified in the univariate analysis). Therefore, this source of bias may have led to an under-reporting of high psychological distress. Again, it is not possible to define or quantify the extent of this. Finally, the use of an online survey imposes potential limitations. It is probable that the study findings under-represent the responses of those in certain demographics, e.g. those who are less educated (as noted), those less affluent, and older respondents. Not all horse owners would have access to the internet, and online survey methodology is relatively uncontrolled, e.g. the sample was self-selected and therefore may be more prone to response bias than a sample that was randomly selected or otherwise controlled. Also, those experiencing higher distress may have been more motivated to respond. The extent of this response bias on the data cannot be accurately estimated, however, in anticipation of potential response bias, actions were taken to ensure that the study was presented to potential respondents in a way that would minimize such effects; e.g. the study was presented as independent of any industry group or government organization and it was clearly identified as a university research study. It was hoped that such presentation of the study would reduce political or self-interest motivation for completing the study. Despite some methodological limitations, this study was able to determine the psychological impact of Australia's first outbreak of equine influenza on a substantial sample of horse owners and those involved in horse-related industry. Study findings indicated that this affected population had highly elevated levels of psychological distress and that, although prevalence of high psychological distress was greater in infected EI control zones and States, elevated levels of psychological distress were experienced in horse-owners nationally, and not just in areas where equine influenza was present. Statistical analysis indicated that certain groups were more vulnerable to high psychological distress; specifically younger people, those with no formal educational qualifications, and those whose main income was linked to a horse-related industry. Findings from this study generate further questions: What were the determinants of elevated psychological distress? Was it the risk of the disease itself, e.g. fear of the disease, or concern for horses? Was it the social and emotional impacts of disease control measures and restrictions, e.g. social isolation, quarantine, loss of freedom or control, stigma of being 'infected'? Was it loss of income or sporting aspirations? More importantly, how enduring is this elevated psychological distress, and what are the longer term mental or physical health consequences for those affected? The latter is of critical importance given the increased prevalence of high psychological distress reported in young people in this study. Some of these questions can be addressed using additional data collected in the wider study; however, the issue of enduring psychological distress will require further assessment. Publish with Bio Med Central and every scientist can read your work free of charge

Influenza A Virus Inhibits Type I IFN Signaling via NF-κB-Dependent Induction of SOCS-3 Expression

The type I interferon (IFN) system is a first line of defense against viral infections. Viruses have developed various mechanisms to counteract this response. So far, the interferon antagonistic activity of influenza A viruses was mainly observed on the level of IFNβ gene induction via action of the viral non-structural protein 1 (NS1). Here we present data indicating that influenza A viruses not only suppress IFNβ gene induction but also inhibit type I IFN signaling through a mechanism involving induction of the suppressor of cytokine signaling-3 (SOCS-3) protein. Our study was based on the observation that in cells that were infected with influenza A virus and subsequently stimulated with IFNα/β, phosphorylation of the signal transducer and activator of transcription protein 1 (STAT1) was strongly reduced. This impaired STAT1 activation was not due to the action of viral proteins but rather appeared to be induced by accumulation of viral 5′ triphosphate RNA in the cell. SOCS proteins are potent endogenous inhibitors of Janus kinase (JAK)/STAT signaling. Closer examination revealed that SOCS-3 but not SOCS-1 mRNA levels increase in an RNA- and nuclear factor kappa B (NF-κB)-dependent but type I IFN-independent manner early in the viral replication cycle. This direct viral induction of SOCS-3 mRNA and protein expression appears to be relevant for suppression of the antiviral response since in SOCS-3 deficient cells a sustained phosphorylation of STAT1 correlated with elevated expression of type I IFN-dependent genes. As a consequence, progeny virus titers were reduced in SOCS-3 deficient cells or in cells were SOCS-3 expression was knocked-down by siRNA. These data provide the first evidence that influenza A viruses suppress type I IFN signaling on the level of JAK/STAT activation. The inhibitory effect is at least in part due to the induction of SOCS-3 gene expression, which results in an impaired antiviral response.

Influenza A viruses are negative-stranded RNA viruses that belong to the family of orthomyxoviruses. The segmented genome of influenza A virus encodes for up to 11 viral proteins. As many other viruses, influenza viruses have evolved strategies to counteract cellular antiviral responses, especially to circumvent the type I IFN system as a first line of defense against the pathogenic invader. Among the influenza viral proteins, the NS1 has been identified as the main type I IFN antagonistic factor. So far two major mechanisms have been described by which NS1 suppresses the initial expression of IFNb. On the one hand NS1 inhibits vRNAmediated induction of the transcription factors interferon regulatory factor-3 (IRF-3), activating protein-1 (AP-1) and NF-kB that target the IFNb promoter. This most likely occurs via binding to the RNA-sensor retinoic acid inducible gene (RIG-I) and inhibition of RIG-I-mediated signaling in response to viral RNA [1, 2] . On the other hand NS1 inhibits maturation [3, 4] and nuclear export of host mRNAs [5] . Other functions of the multifunctional protein include block of activation of the dsRNAactivated protein kinase PKR by direct interaction [6] or activation of the phosphatidylinositol-3 kinase PI3K/Akt pathway to prevent premature apoptosis induction [7, 8] . While the NS1-mediated antagonistic activities of influenza viruses mainly affect the induction of genes such as IFNb, so far no viral suppression of IFN signaling has been described. IFN are among the first molecules synthesized in response to viral infections [9] . The IFN family includes three classes. Type I comprises the well known IFNa and IFNb. The only member of type II IFN is IFNc. Type III IFN comprises IFNl1, -l2, and -l3. All classes of IFN bind to different receptors and are structurally not related [10, 11] . Type I IFN belong to the key cytokines produced by influenza A virus-infected epithelial cells [12, 13] . The antiviral activity of type I IFN is mediated by a set of IFN-induced genes (ISGs). Binding of IFNa/b to its receptor is the initial step in this signaling process, followed by activation of the JAK family and subsequent activation of STAT proteins [14] . Ligand binding leads to dimerisation of the type I IFN receptor subunits IFNAR1 and IFNAR2 and causes their conformational change. The JAK kinase Tyk2, which is constitutively bound to IFNAR1, phosphorylates the receptor at tyrosine residues and creates a docking site for STAT2. Subsequently, Tyk2 phosphor-ylates STAT2 at Y690. At the same time the receptor-bound JAK1 phosphorylates STAT1 at Y701 [15, 16] . The phosphorylated transcription factors dimerise and bind to IRF-9 [17] . The newly formed heterotrimer, called IFNstimulated gene factor 3 (ISGF3), translocates into the nucleus and binds to IFN-stimulated response elements (ISRE), to initiate gene transcription of ISGs. Treatment of cells with type I IFN upregulates expression of an array of genes including SP110, IRF-1 and many others [18] . Among these ISGs the 29, 59oligoadenylate synthetase 1 (OAS1), the Mx proteins and the dsRNA-activated protein kinase (PKR) are described to directly interfere with viral replication [19] . Both, PKR and the OAS1/ RNaseL system are capable of inhibiting cellular and viral translation. IFN-induced JAK/STAT signaling can be inhibited at different levels by several viral and cellular factors through various mechanisms. The large T-antigen of murine polyomavirus (MPyV) binds to JAK1 and inhibits downstream signaling [20] , whereas the VP24 of Ebola virus (EBOV) binds to karyopherina-1 thereby blocking nuclear accumulation of STAT1 [21] . Endogenous cellular key regulators, capable of negatively regulating JAK/STAT-mediated signal transduction, include suppressor of cytokine signaling (SOCS) proteins, protein tyrosine phosphatases (PTP) and protein inhibitor of activated STATs (PIAS). The family of SOCS proteins comprises eight members (cytokine-inducible SH2 domain-containing protein (CIS) and SOCS1-7). All members contain a central SH2 domain, an Nterminus of variable length and sequence and a C-terminal 40 amino-acid module called SOCS box [22] . The SOCS box is necessary for recruitment of the ubiquitin transferase system and for stabilization and/or degradation of SOCS proteins [23] [24] [25] . The N-terminus contains a kinase inhibitory region (KIR), which functions as pseudo substrate for the JAK [26] . SOCS-1 and SOCS-3 differ in their mode of action. For inhibition of the kinase activity of JAKs, SOCS-1 binds directly to the activation loop of JAKs [26] [27] [28] . In contrast, SOCS-3 first binds to the receptor [29, 30] . Induction of SOCS-3 gene transcription by viruses was reported for HSV-1, HCV [31] [32] [33] and for respiratory viruses, such as SARS and RSV [34, 35] . The level of induction of SOCS-3 by HSV-1 seems to determine whether infection turns to acute or persistent progression [31] . For HCV it has been suggested that upregulation of SOCS-3 may contribute to the non-responsiveness of HCV patients to IFN therapy [33, [36] [37] [38] . Elevated SOCS-3 mRNA levels during RSV infection were linked to Th2 cell-mediated immune disease as atopic dermatitis and asthma [39, 40] . In the present study we show that influenza A virus can be added to the list of viruses that induce SOCS-3 expression. The protein functionally interferes with viral replication by providing a virus-supportive IFN-antagonistic activity on the level of type I IFN-signaling that has not been described so far. Phosphorylation of STAT1 and STAT2 by members of the JAK tyrosine kinase family is a prerequisite for activation of these transcription factors to drive type I IFN-induced gene expression. Therefore, we analyzed whether STAT phosphorylation patterns are altered in influenza A virus infected cells that were stimulated with IFN at different time points post infection (p.i.). The human alveolar epithelial cell line A549 was infected with the influenza A virus strain A/Puerto-Rico/8/34 (H1N1) (PR8) ( Figure 1A ). Cells were subsequently stimulated with IFNb at given time-points p.i. and STAT phosphorylation was assessed in Western blots. Both STAT1 and STAT2 were readily phosphorylated upon cytokine stimulation in uninfected cells or in infected cells up to 4 h p.i. ( Figure 1A ). Furthermore, virus infection alone resulted in a significant induction of STAT phosphorylation 4-6 h p.i., presumably caused by virus-induced IFN expression. However, at later time points (6-10 h p.i.), in A549 cells both virus-and IFN-induced STAT1 and STAT2 phosphorylation was markedly reduced ( Figure 1A ). Similar patterns were observed upon stimulation of cells with IFNa or upon infection with other viruses, such as the human influenza virus A/Victora/3/75 (H3N2) (data not shown). In addition, this phenomenon could also be detected in other epithelial cells such as the human embryonic kidney cell line HEK293 ( Figure 2E ) or the human umbilical vein endothelial cells (HUVEC) ( Figure S1B ). Inhibition was not caused by indirect disturbing effects on cellular metabolism or enzyme activities due to ongoing virus replication, since IFNc-induced STAT1 phosphorylation was not affected at all ( Figure 1C ). Finally, involvement of any auto-or paracrine action of virus-induced type I IFN could be ruled out, as the inhibitory effect was also observed in Vero cells lacking functional type I IFN genes ( Figure 1E ). With regard to the molecular basis of impaired IFNa/b-induced STAT phosphorylation in infected cells it was striking that the inhibitory effect correlated with the accumulation of viral proteins, as monitored in PB1 Western blots ( Figures 1A and 1E) . Thus, the question arose whether individual expression of viral proteins may result in the interference with STAT1 phosphorylation. Out of the 11 viral proteins of PR8 we choose the nucleoprotein (NP), the NS1 protein, the matrix protein (M1) (Figure 2A ) and the subunits of the viral polymerase, PA, PB1 and PB2 ( Figure 2C ), for a representative experiment. These proteins are known to bind to vRNA/RNPs or to interfere with the RNA-mediated innate immune response. For efficient transfection of the expression The type I interferon (IFN) system is one of the most powerful innate defenses against viral pathogens. Most RNA viruses are sensitive to the action of type I IFN. Therefore, these pathogens have evolved strategies to evade this response. For example, influenza viruses express a viral protein, the non-structural protein 1 (NS1), that suppresses production of IFNb by lowering cellular sensitivity to viral nucleic acid as a pathogen pattern. Here we present data indicating that influenza A viruses are not only capable of suppressing production of the IFNb gene but also inhibit action of this antiviral cytokine on cells. This occurs by viral induction of a cellular protein, the suppressor of cytokine signaling (SOCS)-3, a potent endogenous inhibitor of IFN signaling. This is a novel mechanism by which influenza viruses inhibit the antiviral response of the host and paves the path to efficient virus replication. This may be especially relevant for influenza viruses that induce high cytokine responses (cytokine burst), such as highly pathogenic avian influenza viruses of the H5N1 subtype. Induction of SOCS-3 expression would allow efficient replication despite high IFN and cytokine levels. constructs we used the highly susceptible cell line HEK293 that also exhibits impaired IFNb-induced STAT1 phosphorylation at later stages of infection ( Figure 2E ). 24 h post transfection cells were stimulated with IFNb and STAT phosphorylation was monitored in Western blots (Figures 2A and C) . Expression of none of the viral proteins resulted in a significant decrease of IFNb-induced STAT1 or STAT2 phosphorylation (Figures 2A and 2C ). Similar results were obtained in the human bronchial epithelial cell line H1299 when expressing M1, NS1 or NP alone or in different combinations (data not shown). Thus, we concluded that viral proteins most likely do not play a prominent role as blockers of IFNa/b-induced JAK/STAT signaling. Decrease of STAT phosphorylation might also be due to the action of virus-induced phosphatases. On the one hand these enzymes may cause direct dephosphorylation of STAT proteins. On the other hand phosphatases could act via an indirect mechanism by dephosphorylation and inactivation of JAKs resulting in an attenuated phosphorylation of STATs. Several protein tyrosine phosphatases (PTPs) are known to mediate dephosphorylation of both, JAKs and STATs [41] . In order to investigate whether influenza A virus activates phosphatases that subsequently target JAKs or STATs, we treated infected or uninfected A549 cells with the well-known tyrosine phosphatase inhibitor sodium vanadate [42, 43] . Uninfected cells or cells infected with PR8 for 10 h were incubated with increasing amounts of this compound 10 min prior to stimulation with IFNb. This time point of infection was chosen since we observed considerable inhibition of IFN-induced STAT1 phosphorylation in the course of infection ( Figures 1A and 1E ). Increasing concentrations of vanadate lead to a gradual shift of the steady state balance of phosphorylation/dephosphorylation. Accordingly, a gradual increase of STAT1 phosphorylation was observed that was similar in both infected and uninfected cells, albeit starting from different basal levels of phospho-STAT1 ( Figures 3A and B ). This is illustrated by an almost identical slope of the regression line in the graphical analysis of the band intensities of the IFNb stimulated samples ( Figure 3B ). If the blockade of IFNb-induced STAT1 phosphorylation would be mediated by specific virus-activated phosphatases, a much steeper slope for vanadate-treated infected cells would be expected. However, the result in Figure 3B indicates that the virus-induced suppression of phosphorylation is not compensated by phosphatase inhibition and consequently no virusactivated phosphatase appears to be involved. In support of these data, phosphatase assays revealed that the overall activity of tyrosine phosphatases in infected cells was not elevated compared to uninfected cells. This is indicated by constant levels of free phosphates released from two different phospho-peptides that represent common tyrosine phosphatase substrates ( Figure 3C ). Thus, involvement of phosphatases in influenza virus-induced alteration of STAT1 phosphorylation can be greatly ruled out. Phosphorylation of STATs in the IFNb signaling cascade may not only be counter-regulated by phosphatases but also by other cellular factors, such as proteins of the suppressors of cytokine signaling (SOCS) family. Action of these proteins is mainly controlled on the level of transcriptional activation. SOCS proteins are described to have high affinity for JAK and STAT proteins and to inhibit the transmission of IFNa and IFNb induced signaling [44, 45] . To examine whether expression of SOCS genes is induced in influenza virus infected cells, A549 cells were infected with PR8 for different time points. Subsequently total RNA was analyzed for the amount of SOCS-1 and SOCS-3 mRNA by means of quantitative real time-PCR (qRT-PCR). The mRNA Table 1 for accession numbers of viral genes) using L2000 according to manufacturer's instructions. Note that the Pol II constructs in use also give rise to expression of second reading frames in the NS, M and PB1 genes (NS2, M2, PB1-F2). 48 h post transfection cells were stimulated with human IFNb (500 U/ml) for 15 minutes. Total protein lysates were subjected to Western blot analysis using anti-phospho-STAT1, anti-phospho-STAT2, anti-STAT1 antibodies. Expression of influenza viral proteins was monitored with antibodies against NP, M1, NS1, PA, PB1 or PB2. (E) HEK293 cells were infected with the human influenza virus PR8 (H1N1) (MOI = 5) for the indicated time points and were subsequently stimulated for 15 min with either human IFNb at a concentration of 100 U/ml. Cell lysates were subjected to Western blots as described. (B, D, F) Quantification of relative pSTAT1 and pSTAT2 band intensities in A, C and E using AIDA software and 2D densitometry (Fuji). doi:10.1371/journal.ppat.1000196.g002 levels of SOCS-1 and SOCS-3 differed notably in the time course ( Figure 4A ). While SOCS-3 mRNA is strongly and transiently elevated in the early phases of infection, SOCS-1 gene transcription is only significantly induced 15 h p.i.. Elevated SOCS-3 mRNA levels were also observed in other host cell types, such as HUVEC starting 3 h p.i. ( Figure S1A ). Although elevation of SOCS-3 mRNA levels in infected cells was rather transient, there appears to be a robust induction on protein level ( Figure 4C ). First detected at 4 h p.i., SOCS-3 protein levels increased and stayed on a high level throughout the observation period. Strikingly, expression kinetics of the SOCS-3 protein perfectly matched the kinetics of virus-induced inhibition of STAT1 phosphorylation ( Figure 4C ), indicating that both processes are functionally linked. Virus mediated SOCS-3 gene induction at early stages of infection ( Figures 4A, 4C and S1A) appeared to occur concomitant with an immediate and strong induction of IFNb ( Figures 4B and S1D) . This prompted us to analyze whether SOCS-3 transcription might be induced due to an auto-or paracrine action of IFNb expressed during infection. A549 cells were stimulated with IFNb for different time points and SOCS-3 gene induction was measured by qRT-PCR ( Figure 4D ). As a control we monitored expression of 29, 59oligoadenylate synthetase (OAS1) and MxA, genes that are typically induced by IFNb. While OAS1 and MxA mRNAs were readily upregulated upon IFNb treatment SOCS-3 mRNA was not significantly elevated ( Figure 4D ). Similar results were obtained from HUVEC stimulated with IFNb ( Figure S1E ). To further confirm these results we knocked down the IFNAR in A549-cells by an siRNA approach. Although the knock down was efficient and leads to more than 60% inhibition of IFNb induced STAT1 phosphorylation ( Figure 4E ), the induction of SOCS-3 expression was not impaired ( Figure 4F ). SOCS-3 levels in the knock down cells were similar compared to wild type cells and even higher than in the vector control ( Figure 4F ). These results are consistent with data gained from previous experiments in Vero cells ( Figure 1E ) and indicate that neither induction of SOCS-3 mRNA nor inhibition of STAT phosphorylation is dependent on virus-induced type I IFN expression. Since accumulation of viral RNA in infected cells is a potent inducer of antiviral gene expression we investigated its ability to induce SOCS-3 gene transcription. As a source for viral RNA, A549 cells were infected with influenza A virus for 10 h and total RNA from these cells was isolated. RNA from uninfected A549 cells served as a negative control. Different amounts of these RNAs were used for stimulation of A549 cells for 3 h (Figures 5A, 5B and 5C). Transfection of RNA from uninfected cells did not result in an increase of SOCS-1 or SOCS-3 gene transcription ( Figure 5A ) or IFNb induction as a control ( Figure 5B ). However, transfection of RNA from virally infected cells led to strongly elevated SOCS-3 mRNA amounts while SOCS-1 mRNA is only induced weakly ( Figure 5A ). This dose dependent induction of SOCS-3 by stimulation with increasing amounts of RNA from infected cells corresponds with a gradual decrease in the ability of this RNA to induce or potentiate STAT1/2 phosphorylation ( Figure 5C ). In contrast to cellular RNA, influenza viral RNA carries a triphosphate group at its 59 terminus that was previously shown to be a major pathogen pattern that triggers cellular signaling [46] . To verify that indeed the viral 59 triphosphate RNA in the pool of RNAs from infected cells is the major trigger for induction of SOCS-3 expression, RNA from infected or uninfected cells was treated with phosphatase to remove the 59 triphosphate termini prior to stimulation of A549 cells ( Figure 5E and 5F). The dephosphorylated viral RNA was only poorly capable to induce SOCS-3 ( Figure 5E ) or IFNb ( Figure 5F ) mRNA expression. In addition, poly(I:C) was transfected to mimic action of double-stranded (ds) RNA ( Figure 5E and 5F, right bars). However, the dsRNA analog showed surprisingly little effects on SOCS-3 and IFNb mRNA induction. Since viral RNA is able to induce IFNb gene transcription ( Figure 5B and 5F) we again wanted to rule out that induction of SOCS-3 by viral 59 triphosphate RNA is mediated by auto-or paracrine action of de novo synthesized IFNb. In order to do so, cells were stimulated with viral RNA after treatment with the protein synthesis inhibitor anisomycin at two different concentrations ( Figure 5G ). SOCS-3 mRNA was still induced to the same extent in the presence of the protein synthesis inhibitor, providing the ultimate proof that de novo protein synthesis is not required for SOCS-3 induction. So far, our data suggest that influenza virus-induced transcriptional upregulation of the SOCS-3 gene is not mediated by the . Equivalent mRNA amounts were normalized to GAPDH mRNA levels and calculated as n-fold of the levels of untreated cells that were arbitrarily set as 1. To detect SOCS-3 protein expression (C) cells were infected for time points indicated or left uninfected. Total cell lysate was subjected to Western blot analysis using anti-SOCS-3 antibody. To allow better comparison of SOCS-3 protein expression and STAT1 phosphorylation phospho-STAT1 and STAT1 Western blots from figure 1A are shown again here. (E) To functionally test effective knock down of the IFNAR, A459 wild type, A549 vector control cells or A549 cells stably expressing IFNAR II-1specific shRNA were stimulated with human IFNb (100 U/ml) for 15 min. Subsequently cells were lysed and levels of phospho-STAT1 were determined by Western blotting using specific antibodies. In addition, the relative pSTAT1 band intensities were quantified. doi:10.1371/journal.ppat.1000196.g004 autoregulatory action of type I IFNs ( Figure 4D and 4F) but is directly induced through accumulation of viral RNA during infection. This raises the question, which RNA-induced signaling pathways are responsible for SOCS-3 expression. The MKK/p38 mitogen activated protein kinase (MAPK) pathway [47] [48] [49] as well as the IkB kinase (IKK)/nuclear factor of kB (NF-kB) cascade [50] [51] [52] are both known to be activated by RNA or influenza virus infection and to be involved in the control of SOCS-3 expression. To assess whether the MKK6/p38-or the IKK/NF-kB-module is required for SOCS-3 gene induction, we generated A549 cell lines expressing dominant negative forms of either MKK6 (MKK6Ala) or IKK2 (IKK2KD) (Figure 6A to 6D). These mutants have been previously shown to efficiently block p38 or NF-kB signaling, respectively [52] [53] [54] . To monitor SOCS-3 gene induction, wild type, vector or mutant expressing cell lines were infected with PR8 ( Figure 6A ) or stimulated with RNA from virally infected or uninfected A549 cells ( Figure 6C ). Induction of IFNb mRNA was monitored as a control ( Figure 6B and 6D) . While MKK6Ala expression did not result in significant reduction of SOCS-3 in either infected ( Figure 6A ) or RNA-stimulated cells ( Figure 6C ), transcription is markedly reduced in IKK2KD expressing cell lines. To obtain independent evidence for NF-kB dependence of SOCS-3 gene transcription, A549 wild type cells were incubated with the NF-kB specific inhibitor BAY 11-7085 prior to stimulation with RNA from virally infected or uninfected A549 cells ( Figure 6E ). Again, IFNb mRNA levels were assessed for control purposes ( Figure 6F ). Both, SOCS-3 and IFNb mRNA levels were strongly reduced in BAY 11-7085 treated cells. This indicates that virus-induced SOCS-3 expression strongly depends on IKK2 and NF-kB activation, while the MKK6/p38 appears not to play a prominent role. To further verify that influenza virus induces SOCS-3 via an RNA sensory pathway and in an NF-kB dependent manner we infected cells with the influenza A virus mutant deficient for NS1 (DNS1) (Figure 6G and 6H) . The NS1 protein is known to block RNA dependent signaling and NFkB activation [55] . Accordingly, infection of cells with the mutant virus resulted in a more pronounced and sustained, albeit delayed induction of SOCS-3 ( Figure 6G ) if compared to infection with the isogenic wild type, that is a very poor inducer of SOCS-3 but still reasonably well induces IFNb. Noteworthy, this isogenic wild type strain differs from the PR8 wild type virus used in the other experiments shown here (see Materials and Methods for details). To analyze whether NF-kB activation is sufficient for SOCS-3 gene induction we stimulated cells with IL-1b ( Figure S2A ) or TNFa ( Figure S2B ) that are both strong activators of the transcription factor. While mRNA levels of IL-6, a strictly NF-kB dependent cytokine, are strongly elevated, SOCS-3 gene transcription is not significantly induced. Under the assumption that these cytokines do not additionally induce counteracting processes one can conclude that NF-kB is required, yet not sufficient for the induction of SOCS-3. Thus viral induction of SOCS-3 may require additional factors that are only active in virus-infected cells. Furthermore, these results rule out a potential role of virus-induced IL-1b or TNFa in the induction of SOCS-3. This is supported by the observation that neither expression of IL- To further assess a functional role of SOCS-3 in virus-induced suppression of STAT1 phosphorylation we analyzed mouse cells with a targeted deletion of the SOCS-3 gene [56] . Wild type and SOCS-3 deficient mouse embryonic fibroblasts (MEF) were infected for different time points with PR8. The time of infection was prolonged in comparison to the infection of A549 cells because the human PR8 replicates less efficiently in mouse than in human cells. Following infection lysates of these cells were assessed for STAT1 phosphorylation ( Figure 7A ). Both cell types showed no phosphorylation of STAT1 in the uninfected state. In contrast, infection of SOCS-3 knock out cells resulted in strongly elevated phosphorylation of STAT1 in a sustained fashion. To rule out that this STAT1 phosphorylation is due to altered secretion of IFNb or Figure 7 . Enhanced STAT1 phosphorylation in infected SOCS-3 deficient MEF correlates with elevated induction of IFNb-stimulated genes. Wild type MEF and SOCS-3 knock out MEF were infected with PR8 (MOI = 5) for the indicated times. Subsequently, cell lysates were analyzed for STAT1 phosphorylation (A). For control of productive virus replication, cell lysates were analyzed for viral protein PB1 expression. In (E, F, G) wild type and knock out cells were lysed at indicated time-points of infection. Subsequently RNA was subjected to reverse transcription. cDNA was analyzed in quantitative real time PCR to assess mRNA amounts of three prototype type I IFN-stimulated genes, SP110 (E), interferon regulatory factor-1 (IRF-1) (F) and OAS1 (G). Equivalent mRNA amounts were normalized to GAPDH mRNA levels and calculated as n-fold of the levels of untreated cells that were arbitrarily set as 1. In (C) wild type MEF and knock out MEF were infected with PR8 (MOI = 5) or left uninfected. Supernatants were taken 6 p.i. and used for stimulation of wild type MEF for 15 minutes. As control wild type MEF were stimulated with 500 U/ml mouse IFNb for 15 minutes. Cells were harvested and analyzed for the amount of STAT1 and phospho-STAT1 in Western blot analysis by specific antibodies. In (B) and (D) the relative band intensities of phospho-STAT1 of the blots in (A) and (C) were quantified as described. doi:10.1371/journal.ppat.1000196.g007 other STAT1-activating cytokines in SOCS-3 deficient cells, we performed conditioned medium experiments ( Figure 7C ). MEF wild type and MEF SOCS-3 deficient cells were infected for 6 h and supernatants were subsequently harvested. Stimulation of MEF wild type cells with these different supernatants for 15 min. revealed no differences in STAT1 phosphorylation, indicating that both infected cell types secrete similar amounts of IFNb and other STAT1 activating cytokines. This is a strong indication that the observed differences in virus-induced STAT phosphorylation are directly due to the presence or absence of SOCS-3 in wild type and knock out MEF, respectively. To answer the question whether enhanced STAT phosphorylation in SOCS-3 deficient cells would also lead to enhanced expression of ISGs, total RNA was isolated at different time points p.i. from infected wild type and knock out cells and monitored for induction of SP110, IRF-1 and OAS1 ( Figure 7E, 7F and 7G ). These genes are described as type I IFN-induced genes [18] . Indeed mRNA levels of all three representative ISGs were elevated in SOCS-3 knock out versus wild type cells at almost every time point during the course of infection. This indicates that enhanced STAT1 phosphorylation and activation in SOCS-3 deficient cells results in elevated expression of ISGs. The remaining question was, whether the elevated IFN-induced gene response in knock out cells might also affect propagation of influenza A viruses. Thus, both wild type and knock out cells were infected with PR8 ( Figure 8A ) or the strain A/Victoria/3/75 (H3N2) ( Figure 8B ). Virus titers were assessed at different time points post infection. Progeny virus titers from SOCS-3 knock out cells were significantly reduced compared to titers from infected wild type cells. To independently confirm these results and to verify that the observed effects are really due to the lack of SOCS-3, we used an siRNA approach to specifically knock down SOCS-3 mRNA in A549 cells. Cells were transfected with 150 nM siRNA for 48 h and SOCS-3 protein levels were compared to control transfected samples ( Figure 8C, right) . Subsequently, cells were infected and progeny virus titers were determined by plaque assay (Figure 8C, left) . Similar to the results gained from infected knock out cells, knock down of SOCS-3 resulted in decreased virus titers. On the contrary, over-expression of SOCS-3 resulted in elevated virus titers ( Figure 8D ) concomitant with an inhibition of IFNb-or virus-induced STAT1 phosphorylation ( Figure 8E) . Taken together the data indicate that in the absence of SOCS-3, infection leads to a stronger activation of STAT1, resulting in enhanced expression of ISGs and reduced virus titers. Vice versa, over-expression of SOCS-3 leads to an inhibition of STAT1 activation and elevated virus titers, probably due to inhibited expression of ISGs. This highlights the important role of virus induced SOCS-3 to limit the type I IFN-induced antiviral response program. The type I interferon (IFN) system is one of the most powerful innate defenses of vertebrate cells, which limits replication and spread of viral pathogens including avian and human influenza viruses. Influenza virus propagation is sensitive to IFN activities and therefore, like other viral pathogens, these viruses do not only induce type I IFN but also antagonize the production and effects of these cytokines at the same time [55] . For influenza A and B viruses, this is accomplished through their non-structural NS1 proteins that are structurally related polypeptides of 26 kDa (A/ NS1) and 32 kDa (B/NS1), which are abundantly expressed in infected cells [55] . NS1 proteins predominantly act on the level of IFN gene induction in infected cells by obstructing RIG-Idependent signaling through interaction with cellular factor(s) and/or sequestration of RNAs generated during virus replication [1, 2, 57] . Some NS1 proteins were also described to inhibit the maturation of cellular pre-mRNAs raising the possibility that this activity additionally reduces production of IFNa/b in infected cells [58, 59] . While NS1 also interferes with the activity of some ISGs, such as the dsRNA dependent kinase PKR [5, 60] , so far no type I IFN antagonistic mechanism was described for influenza viruses that act on the level of IFN signaling rather than gene induction. Here we present data, showing that RNA-induced expression of SOCS-3 in early phases of infection leads to a functional inhibition of IFN-induced STAT activation and gene expression. This is a novel mechanism by which influenza virus suppresses the antiviral response of the host and paves the path to efficient virus replication. While it was reported in the literature that expression of SOCS proteins can be induced upon stimulation with IFN [61] we could not detect any significant gene induction by IFNb in A549 cells. Instead we observed a significant up-regulation of SOCS-3 by viral 59 triphosphate RNA, indicating that gene induction occurs via accumulation of vRNA during infection. This appears to occur through the RNA-mediated activation of the IKK/NF-kB pathway, most likely activated through engagement of the RNA sensor RIG-I. At a first sight, this might appear controversial since NF-kB activation is among the RIG-I-induced signaling responses and NS1 was reported to inhibit this signaling pathway. However, despite the action of NS1 it is well known that NF-kB is still significantly activated upon influenza virus infection and many NF-kB and IFNb dependent genes are still expressed. We hypothesized previously that the incomplete inhibition conferred by NS1 is an indication that the virus exploits the remaining signaling activities for efficient replication [52, 62, 63] . The findings described here, namely NF-kB dependent induction of SOCS-3 and limitation of type I IFN signaling responses, provide yet another example how influenza viruses take advantage of NF-kB activity. While the data show that NF-kB is required for viral SOCS-3 induction, the factor appears not to be sufficient, since prototype inducers of NF-kB, such as IL-1b or TNF-a would not induce SOCS-3. Thus there seems to be the need of additional virus or RNA-induced transcription factors. The most likely candidate would be the constitutively expressed interferon regulatory factor 3 (IRF-3), that is known to be simultaneously activated with NF-kB upon virus infection directly via the RIG-I RNA sensing pathway without the need of type I IFN. Furthermore IRF-3 is a factor suppressed by the NS1 protein. Recently it was reported that IFN-induced gene expression responses are potentiated in cells, which lack the NF-kB factors p50 or p65 [64] . Although these authors described an inhibitory binding of NF-kB transcription factors to some IFN-induced gene, this mechanism might be cell type dependent since we could not observe similar effects in the cell types used here (data not shown). Thus, the underlying molecular mechanisms appear to be not fully clear. It is striking that the effects described for p50 and p65 knock out cells in these studies fully correlate with our observations in SOCS-3 deficient cells. While in the latter case cells lack the IFNb signaling inhibitor SOCS-3, the p50 and p65 knock out cells are deficient for the factors required for SOCS-3 induction. Thus, given the NF-kB dependent induction of SOCS-3 described in the present manuscript, we provide an additional molecular mechanism that may explain the phenomenon described by Wei et al. [64] . First indications for beneficial effects of SOCS-3 gene expression on viral replication came from studies using the HCV core protein as a replacement for the influenza A viral NS1 in the context of infections with a NS1 deficient influenza virus [33] . One of the hallmark responses of HCV core expression is a rapid induction of SOCS-3 expression. Given the role of SOCS-3 described here, it was not surprising that HCV core could partially rescue growth of the NS1 deficient virus [33] . While this manuscript was in preparation it was demonstrated by Pothlichet et al. that influenza A virus-induced SOCS-1 and SOCS-3 upregulation requires a TLR-3-independent, RIG-I/ MAVS-dependent pathway [65] . Moreover, over-expression of SOCS-1 and SOCS-3 in infected cells revealed that both molecules inhibit antiviral responses. These studies are perfectly complemented by our findings. Here we confirm involvement of RIG-I/MAVS by showing that 59 triphosphate RNA, the ligand for RIG-I, is a major inducer of SOCS-3. Furthermore, the finding that dsRNA is only a weak inducer of SOCS-3 is also consistent with the independence from the dsRNA sensor TLR-3. The only discrepancy of this work and the study of Pothlichet et al. is that they show a dependence on the type I IFN receptor. This may be due to the different virus-strains and cell types used. It is well known that the capability of type I IFNs to induce SOCS proteins is strongly cell type specific [31] . While in some cell types SOCS-3 expression appears to be type I IFN dependent (e.g. fetal liver cells) [31] it is clearly independent of IFN in other cell types [66] . Recently it was shown that SOCS-3 is not significantly induced by IFNa in A549 cells [18] , the major cell type used in our study. Evidence that cell type specificities may be the cause of discrepancy is additionally provided by the fact that Pothlichet et al. show identical induction kinetics of SOCS-1 and SOCS-3. In contrast the kinetics of the two proteins differ clearly in the cells we used, with SOCS-3 being induced much earlier than SOCS-1 on mRNA and protein level. Finally, it should be stated that regardless whether SOCS-3 is additionally induced by type I IFNs at a later stage of infection, it is important that it can be induced earlier and in parallel to IFNb directly by vRNA accumulation. This is supported by the finding that IFNb and SOCS-3 induction occurs in parallel kinetics ( Figure 4A and 4B) while IFN-induced genes such as OAS1 and MxA are only up-regulated later in a delayed and more sustained fashion ( Figure 4D ). This makes a qualitative difference since the blocking effect of SOCS-3 on IFNb signaling already kicks-in during the first wave of IFNb action. Taken together we describe here for the first time that at least some influenza A virus strains are able to suppress type I IFN signaling by a mechanism involving NF-kB dependent activation of SOCS-3 expression, which negatively affects STAT phosphorylation. This adds a new aspect to our knowledge of the strategies used by influenza A virus to antagonize type I IFN responses. Human influenza A/Puerto-Rico/8/34 (H1N1) (PR8) (Giessen variant) and A/Victora/3/75 (H3N2) (Victoria) were originally taken from the strain collection of the Institute of Virology, Giessen, Germany. The human NS1 deficient influenza virus mutant DNS1 and its isogenic wild type variant were propagated and used as described earlier [7, 67] . It should be noted that this isogenic wild type strain as described by Garcia-Sastre et al. [68] is different from the PR8 (Giessen variant) used in the other experiments and in many previous studies [52, 67] . The supernatant was aspirated and cells were incubated with specific medium containing 0.2% BSA and antibiotics. To score for production of viral plaques the overlay was stained for 1 h using 1 ml neutral red in PBS per well [69] . To trigger JAK/STAT signaling cells were stimulated using human IFNa/b or c as well as mouse IFNb. For stimulation of A549 cells or HUVECs 100 U/ml human IFNa or human IFNb was used. For stimulation of the green monkey epithelial cell line Vero or HEK 293 cells 500 U/ml human IFNb was applied. IFNc was always used in the concentration of 500 U/ml. Mouse embryonic fibroblasts (MEF) were incubated with 100 U/ml mouse IFNb. The different IFN were diluted in infection medium. For stimulation after infection, viral supernatants were aspirated and diluted cytokine was incubated for 15 minutes at 37uC. To investigate the potential of other cytokines to induced SOCS-3 gene expression A549 cells were stimulated with 100 U/ml IL1b or 20 ng/ml TNFa at 37uC for times indicated. After stimulation cells were lysed and subjected to immune blotting. To block the activity of phosphatases after infection with influenza virus, sodium vanadate was used. Dilutions were prepared using infection medium. Sodium vanadate was added to the virus-containing infection medium at the time points indicated. After 10 minutes of incubation IFNb, diluted in infection medium, was added to the medium containing virus and sodium vanadate. The cells were stimulated with IFNb for 15 minutes. Incubation with sodium vanadate started 25 min before cells were lysed and subjected to Western blotting as described. For conditioned medium experiments wild type and SOCS-3 knock out MEF were infected with PR8 (MOI = 5) for 10 h or left uninfected. Supernatants were used for stimulation of MEF wild type for 15 minutes. Cell lysates were subjected to Western blot analysis. To investigate the induction of SOCS-3 expression by viral RNA, RNA isolated from infected or uninfected cells (control) was used. A549 cells were infected with PR8 (MOI = 5) or left mock infected. 10 h post infection RNA was isolated using the RNeasy mini Kit from Qiagen according to manufacturer's instructions. To dephosphorylate viral 59 triphosphate RNA, calf intestine alkaline phosphatase (CIAP) (Fermentas) was used. Briefly, RNA was isolated using Trizol according to manufacturer's instructions. For dephosphorylation the reaction mix was set up in a 50 ml volume with 50 mg RNA, 25 U CIAP and 80 U RiboLock RNase inhibitor (Fermentas) and was incubated for 3 h at 42uC. Thereafter the RNA was isolated using the RNeasy mini Kit from Qiagen. RNAs used as control were mock-treated replacing CIAP by glycerol. For stimulation, the different RNA species and analogues were transfected using Lipofectamine 2000 (L2000) according to manufacturer's instruction (Invitrogen). In brief, L2000 was incubated with OPTI-MEM for 5 minutes at room temperature; different amounts of RNA were added and incubated for additional 15 minutes. For stimulation of cells with cellular or viral RNA 400 ml RNA-L2000 mix were added to 2 ml serum-free medium. Cells were stimulated for 3 hours and subjected to either Western blot analysis or quantitative real time PCR. For silencing SOCS-3 mRNA, A549 cells were transfected with 150 nM human SOCS-3 siRNA 48 h before infection using Hiperfect (Qiagen) according to manufacturer's instructions. In brief, 150 nM siRNA was added to a mixture of D-MEM without FCS/antibiotics and Hiperfect and incubated for 10 min at room temperature. For transfection 400 ml of this mixture were added to the cells. Subsequently cells were subjected to plaque assay analysis or Western blot analysis. Control siRNA was purchased from Qiagen. The sequences for the human SOCS-3 siRNA in use are: human SOCS-3 siRNA sense 59-CCA AGA ACC UGC GCA UCC AdTdT-39, human SOCS-3 siRNA anti-sense 59 -UGG AUG CGC AGG UUC UUG GdTdT-39 ) (see Table 1 for accession number of the human SOCS-3 gene). To determine whether tyrosine phosphatases become activated upon infection with influenza virus a phosphatase assay using the Tyrosine Phosphatase Assay System (Promega) was performed. A459 cells were infected for 10 h (MOI = 5) or left uninfected. Cells were harvested in assay buffer (100 mM tris-HCl pH 5.2, 100 mM CaCl 2 , 100 mM MgCl 2 , 0.02% b-mercapto ethanol), cracked by a single freeze/thaw step at 280uC and disrupted by ultrasonic pulsing. Lysates were precleared from cell debris and residual free phosphates according to the manufacturer's instruction. Tyrosine phosphatase activity was measured by enzymatic release of free phosphate of two given pseudosubstrates (phosphorylated peptides representing target sequences for the most common tyrosine phosphatases). Quantification was performed in comparison to a given standard according to the manufacturer's instruction. For Western blot analysis cells were lysed with RIPA [25 mM Tris/HCl, pH 8.0, 137 mM NaCl, 10% Glycerol, 0.1% SDS, 0.5% NaDOC, 1% IgePal, 2 mM EDTA, pH 8.0, pyrophosphate 5 mg ml 21 aprotinin; 5 mg ml 21 leupeptin; 1 mM sodium vanadate and 5 mM benzamidine] on ice for a minimum of 30 minutes. Supernatants were cleared by centrifugation in a standard tabletop centrifuge (Eppendorf) at maximum speed. Protein concentration was determined by Bradford assay. The phosphorylated and unphosphorylated forms of STAT1 were detected using anti-STAT1 (Y701) antibody and anti-STAT1 (BD Bioscience). An antibody directed against Y690 of STAT2 was used for detection of the phosphorylated form of STAT2 (Upstate). Antibodies to detect influenza viral proteins were purchased from Serotec (NP, M1), Santa Cruz (PB1, PB2). The anti-PA antibody was kindly provided by J. Ortin (Madrid/ Spain). A monoclonal antibody directed against the viral NS1 was generated at the IMV, Muenster, Germany [70] . A monoclonal anti-Myc-tag antibody to detect Myc-M1 was kindly provided by Viktor Wixler. IMV, Muenster, Germany. All secondary antibodies were purchased from Amersham and diluted 1:2500 in TBS-T. Secondary antibodies were incubated for a minimum of 60 minutes at room temperature. To synthesize cDNA from cells, RNA was isolated using Qiagen RNeasy mini kit according to manufacturer's instruction. In brief, cells were lysed in the presence of b-mercaptoethanol and lysates were loaded to a column, washed and eluted in RNase-free water. For reverse transcription 3 mg total RNA, 0.5 mg oligo dT primer in a total volume of 12 ml were heated for 10 minutes at 70uC. Enzyme mix was prepared (56 Enzyme Buffer (Fermentas), water and 500 mM dNTPs) and pre-warmed at 42uC for 2 minutes before adding 535 U/100 ml RevertAid H 2 M-MuLV (Fermentas). Reverse transcription was performed at 42uC for 1 hour. The enzyme was inactivated at 70uC for 10 minutes. Samples were stored at 220uC or directly used in quantitative real-time PCR. For analysis of gene expression relative quantification of the DNA amount was applied. In order to do that gene expression of the housekeeping gene GAPDH was determined. To ascertain changes in expression of the gene of interest the differences between expression of GAPDH and the gene of interest was calculated using the 2 2DDCT method [71] . For quantitative real time Brilliant QPCR SYBR Green Mastermix (Stratagene) was used according to manufacturer's instructions. The fragment of interest was amplified in 40 cycles. The following primers were used (see Table 1 The pCFG5-EGZ retroviral vector used for transfection [72] as well as the constructs to express dominant negative MKK6 (MKK6Ala) or IKK2 (IKK2KD) have been described earlier [52, 73] . The Phoenix amphotropic retroviral producer cells (a gift from G. Nolan, Stanford, CA) [74] were cultured in Dulbecco's modified Eagle's medium containing 10% fetal bovine serum, 100 units/ml penicillin and 100 mg/ml streptomycin. Generation of MKK6Ala or IKKKD expressing producer cells as well as transduction of A549 cells to stably express these transgenes was performed as previously described [52, 53] . Figure S1 Infection of HUVEC results in inhibition of STAT1 phosphorylation and IFNb independent SOCS-3 gene transcription. HUVEC were infected with PR8 (MOI = 5) (A, B, D) or stimulated with 100 U/ml IFNb (E) for time points indicated. To assess the mRNA levels of SOCS-3 (A, E), IFNb (B) and MxA (E) RNA was reverse transcribed and cDNA was subjected to quantitative real time PCR. Equivalent mRNA amounts were normalized to endogenous GAPDH and calculated as n-fold of untreated cells that were arbitrarily set as 1. To assess the amount of phosphorylated STAT1 (B) A549 cells were infected with PR8 (MOI = 5) for time points indicated. Total cells lysate was subjected to Western Blot analysis using anti-phospho-STAT1, anti-STAT1 antibodies. To assess effective viral replication viral NS1 was detected using an anti-NS1 antibody. (C) Quantification of relative band intensities of (B) using AIDA software and 2D densitometry (Fuji). Figure S2 IL1b and TNFa do not affect induction of SOCS-3 gene transcription. A549 wt cells were stimulated with 100 U/ml IL1b (A), 20 ng/ml TNFa (B) or infected with PR8 (MOI = 5) (C) for time points indicated. Cells were lysed, and RNA was subjected to reverse transcription. cDNA was analyzed in quantitative real time PCR to assess mRNA amounts of SOCS-3 and IL6 (A and B) or IL1b (C). Equivalent mRNA amounts were normalized to GAPDH mRNA levels and calculated as n-fold of the levels of untreated cells that were arbitrarily set as 1. Found at: doi:10.1371/journal.ppat.1000196.s002 (4.87 MB TIF)

Key Role of Splenic Myeloid DCs in the IFN-αβ Response to Adenoviruses In Vivo

The early systemic production of interferon (IFN)-αβ is an essential component of the antiviral host defense mechanisms, but is also thought to contribute to the toxic side effects accompanying gene therapy with adenoviral vectors. Here we investigated the IFN-αβ response to human adenoviruses (Ads) in mice. By comparing the responses of normal, myeloid (m)DC- and plasmacytoid (p)DC-depleted mice and by measuring IFN-αβ mRNA expression in different organs and cells types, we show that in vivo, Ads elicit strong and rapid IFN-αβ production, almost exclusively in splenic mDCs. Using knockout mice, various strains of Ads (wild type, mutant and UV-inactivated) and MAP kinase inhibitors, we demonstrate that the Ad-induced IFN-αβ response does not require Toll-like receptors (TLR), known cytosolic sensors of RNA (RIG-I/MDA-5) and DNA (DAI) recognition and interferon regulatory factor (IRF)-3, but is dependent on viral endosomal escape, signaling via the MAP kinase SAPK/JNK and IRF-7. Furthermore, we show that Ads induce IFN-αβ and IL-6 in vivo by distinct pathways and confirm that IFN-αβ positively regulates the IL-6 response. Finally, by measuring TNF-α responses to LPS in Ad-infected wild type and IFN-αβR(−/−) mice, we show that IFN-αβ is the key mediator of Ad-induced hypersensitivity to LPS. These findings indicate that, like endosomal TLR signaling in pDCs, TLR-independent virus recognition in splenic mDCs can also produce a robust early IFN-αβ response, which is responsible for the bulk of IFN-αβ production induced by adenovirus in vivo. The signaling requirements are different from known TLR-dependent or cytosolic IFN-αβ induction mechanisms and suggest a novel cytosolic viral induction pathway. The hypersensitivity to components of the microbial flora and invading pathogens may in part explain the toxic side effects of adenoviral gene therapy and contribute to the pathogenesis of adenoviral disease.

Adenoviruses (Ads) cause mild disease in humans, but are hazardous pathogens in immuno-compromised individuals [1] . Human Ads are dsDNA viruses grouped into six species. Species A, C, D, E, and F and species B Ads use different infectious entry pathways [2] . Human Ads enter mouse cells and express their early genes; however, the virus genome is not replicated and no viral progeny is made during the infection of mouse cells in vitro or in vivo [3, 4] . Furthermore, early viral gene expression can be abolished by UV-inactivation and well-defined mutants with defects of viral early genes or viral entry are available [3, 5] . Thus, the effects elicited by different components of the virus-host interaction preceding viral replication can be accurately evaluated. Ads transduce many different cell types and can be produced in vitro in sufficiently high amounts for in vivo administration. While these properties make them attractive for gene therapy applications, they can also trigger a severe systemic toxic reaction [6, 7] . Upregulation of inflammatory mediators, including cytokines and chemokines such as IL-1, IL-6, IL-8, IL-12, macrophage inhibitory protein-1/2, tumor necrosis factor-a (TNF) and recently also type I IFN has been observed in experimental and clinical infections with wt as well as with recombinant Ads [6, 7, 8, 9, 10] . Type I IFNs represent one of the host's most important antiviral defense mechanisms. The type I IFN family comprises different IFN-a subtypes, a single IFN-b and other less well characterized proteins [11] . All IFN-a species and IFN-b interact with the same IFN-ab cellular receptor, the activation of which mediates a wide range of direct and indirect innate antiviral or antimicrobial effects and modulates the antiviral adaptive immune response [12, 13, 14] . At present, two main mechanisms of type I IFN induction by viruses resulting from the extracytoplasmic or cytoplasmic virus recognition, respectively, are known [12, 13, 14, 15] . The extracytoplasmic induction is initiated by triggering the surfaceexpressed transmembrane protein toll-like receptor (TLR) 4 with certain non-nucleic viral constituents [16, 17, 18] or upon recognition of viral nucleic acids in the endosomes of specialized cells (dendritic cells and macrophages) via different members of the TLR family. These include TLR3, TLR7/TLR8 and TLR9, sensing dsRNA, ssRNA and CpG DNA, respectively. For IFN-ab induction TLR3 and TLR4 signal through the adaptor molecule TIR domain-containing adaptor protein inducing interferon b (TRIF). This results in the activation of interferon regulatory factor (IRF)-3. TLR7, 8 and 9 signal through the adaptor molecule Myeloid differentiation factor 88 (MyD88). An important result of the MyD88-mediated pathway is the activation of IRF-7 (but not of IRF-3), which together with the transcription factors NF-kB and AP-1 initiates the induction of both the IFN-a and IFN-b genes [13, 15] . This induction pathway is responsible for the strong, early IFN-ab response to several replicating and inactivated viruses in pDCs, which express preferentially TLR7 and TLR9 [19, 20, 21] . The ''classical'' cytosolic pathway is the major IFN-ab producing mechanism in cells other than pDCs [22, 23] . Signal transduction leading to type I IFN gene induction is initiated by the recognition of intracellular virus-associated molecular patterns. dsRNA and 59-triphosphate RNA produced during viral replication are sensed by the RNA helicases RIG-I and MDA-5 [23, 24, 25, 26, 27, 28] . This pathway has been extensively studied, mainly in virus-infected fibroblasts. Triggering of the aforementioned RNA helicases leads to the activation of the transcription factors NF-kB, IRF-3 and IRF-7 that are important for the induction of IFN-ab and proinflammatory cytokines, including IL-6. In the cytosolic pathway, type I IFN gene induction is a sequential event and both IRF-3 and IRF-7 were shown to be important in the early phase when mostly IFN-b is produced. The late phase of the IFN-ab response is regulated by positive feedback via the increased levels of IRF-7 elicited by IFN-b production during the early phase [29, 30, 31] . In addition to fibroblasts, the potential of mDCs and macrophages to produce significant amounts of type I IFNs in response to viral replication has been demonstrated in vitro [32, 33] ; however, in vivo the specific contribution of these cells to systemic levels of IFN-ab is not well documented. Recently, detection of bacterial DNA in cells infected with L. monocytogenes and recognition of transfected B-DNA has been shown to trigger IFN-b production. This type of response strictly requires IRF-3 [15, 34, 35, 36, 37, 38] . Such a sensing system has been suggested to represent a further mechanism of cytosolic DNA virus recognition [14, 15] and the Z-DNA binding protein 1 (Zbp1, also referred as DNA-dependent activator of IFN regulatory factors, DAI) was shown to be a candidate DNA sensor in this pathway [39] . Notably however, in a follow-up study the same group found a critical role for DAI in L-929 cells but not in mouse embryonic fibroblasts (MEFs) [40] . Furthermore, recent experiments with Zbp1/DAI knockout mice did not show the essential role of Zbp1/DAI in the induction of innate and adaptive responses to B-DNA in vivo and in macrophages, dendritic cells and MEFs in vitro [41] . The induction of type I IFN in Ad-infected mice has been recently studied [10, 42] and associated with both the extracytoplasmic and intracytoplasmic pathways. It was claimed that a part of the IFN-ab response is initiated by TLR9 and MyD88 signaling in pDCs and another part by cytosolic DNA recognition in non-pDCs. However, the identification of IFN-ab producing cell types directly in infected mice was not carried out. In the present study we investigated the IFN-ab responses of Ad-infected mice and showed that the bulk of the in vivo induced IFN-ab is produced by splenic mDCs. Furthermore, we found that TLRs, including TLR9 play no major role. The Ad-elicited IFN-ab response required viral endosomal escape, suggesting a cytosolic induction pathway. Surprisingly however, the induction was independent of IRF-3 and dependent on stress-activated protein kinase/c-Jun NH2-terminal kinase (SAPK/JNK) activity, which is in contrast to the known induction mechanism initiated through cytosolic DNA recognition. Instead, the induction required IRF-7, and a positive feedback regulation via the type I IFN receptor. Although this does not exclude a role for cytosolic nucleic acid sensors, our data do not support the involvement of MDA-5, RIG-I and Zbp1/DAI in the induction of the IFN-ab response to Ad. Furthermore, our results reveal distinct mechanisms in the induction of IFN-ab and IL-6 by Ad. Finally, we show that Ad-induced IFN-ab is a key mediator of hypersensitivity to bacterial lipopolysaccharides in infected mice. Enhanced susceptibility to LPS and to other microbial inducers of inflammation may contribute to toxic reactions observed during adenoviral gene therapy and to the clinical symptoms of adenoviral diseases. In order to characterize the induction of type I IFNs by Ad in vivo, we first examined how two types of human Ad, Ad3 (species B) and Ad R700 (species C) [2] known to use distinct infectious entry routes, elicit an IFN-ab response in vivo. The results summarized in Fig. 1A and supplementary Fig. S1A show that all mice infected with either of the two viruses exhibited similar IFN-ab responses. IFN-ab was first detectable in plasma at 4 h, peaking at 8 h and declining to low levels 18 h after infection. We then investigated, whether the expression of viral genes is required and/or regulate the induction of IFN-ab by Ads. To this end, we injected mice with an UV-inactivated Ad3 (Fig. S1B ) or Ad R700, incapable of viral gene expression, or with a recombinant Ad5 (species C) that contains a deletion of the E1 and E3 Ad early regions and expresses GFP (Ad5-GFP) (Fig. S1C ). As shown in Fig. 1A , UV-inactivated Ads also induced a strong IFN-ab response which, however, peaked at 6 h after injection, i.e. 2 h earlier than the response to intact Ads. A similar early-peaking IFN-ab response was obtained in mice injected with UV inactivated Ad R700 (Fig. S2A ) and with the recombinant Ad5-GFP (Fig. S2B ). Adenoviruses (Ads) are important pathogens and promising vectors for gene therapy applications. In the course of adenoviral infections innate immune responses are activated, which can be beneficial for the antiviral host defense but also detrimental if activated in a deregulated manner. Type I IFNs are crucial for the innate immune control of various viral infections in the mammalian host. So far, the early, systemic release of IFN-ab during viral infections has been attributed to specialized immune cells, the plasmacytoid dendritic cells. Here, in a mouse infection model, we show that wild type Ads, as well as adenoviral vectors, elicit rapid IFN-ab production almost exclusively in another cell population, the splenic myeloid dendritic cells. This IFN-ab storm depends on viral escape from endosomes to the cytosol and the requirements of the response are suggestive of a novel viral induction pathway. Furthermore, we show that virus induced IFN-ab is the key mediator of Ad-induced hypersensitivity to the cytokine-inducing and toxic activity of lipopolysaccharide, a common constituent of Gram-negative bacteria. Since these bacteria comprise several commensals and pathogens, enhanced susceptibility to lipopolysaccharide may contribute to toxic reactions observed during adenoviral gene therapy and to the clinical symptoms of adenoviral diseases. Titration of the viral preparations in mice revealed that a positive correlation between the viral dose and the height of the IFN-ab response existed only when relatively small doses of intact Ads were used. Higher doses of the Ads either did not elicit a further increase of the IFN-ab response (Ad3, Fig. 1B ) or led to a decrease of the response (Ad R700, Fig. S2C ). In contrast, injection of the corresponding UV-inactivated Ads always led to a gradual increase of the IFN-ab response and, at higher viral concentrations, even exceeded the response obtained with intact viruses. Thus, the expression of Ad genes is not required for the induction of IFN-ab in mice. The data also indicate that the expression of early adenoviral genes negatively regulates type I IFN production. Notably, the Ad E1A gene has been shown previously to suppress Newcastle Disease Virus and IRF-3 induced IFN-a4 promoter induction in transient expression assays in fibroblasts [43] . Interestingly, however, viral gene expression did not inhibit the production of IL-6 in either Ad3-or Ad R700-infected mice ( Fig. 1C and Fig. S2D ). Further experiments revealed that, in contrast to the UV-inactivated Ad, heat-inactivated Ad did not elicit IFN-ab in mice (data not shown). Since heat inactivation prevents the entry of Ad into cells ( [44] and Fig. S2E ), we conclude that signal transduction leading to IFN-ab production is activated during Ad entry. Previous studies have shown that, depending on the inducing virus, IFN-ab is produced ubiquitously or in a cell type specific manner [13, 14] . Here we stimulated different primary mouse cells, including bone marrow derived mDCs (BMDC), bone marrow derived macrophages (BMM), bone marrow derived pDCs and mouse embryonic fibroblasts (MEFs), with Ad2 (species C) or Ad3 (species B). Six hours later, the IFN-ab content of culture supernatants and the expression of the viral E1A gene in the cells were determined. BMDC and BMM, but not MEFs, produced IFN-ab (Fig. 1D ), although all cell types were successfully infected as shown by RT-PCR ( Fig. 1E and not shown). Very similar results were obtained when, instead of Ad2, Ad3 or the recombinant Ad5-GFP were used for infection of the three cell types (not shown). In addition, pDCs also produced IFN-ab in response to Ad infection in vitro (Fig. S3A) ; however, only at high multiplicities of infection. Finally, like MEFs, L-929 cells infected with Ad2 produced no IFNab either (Fig. S3B ). In agreement with previous studies [36, 37] , MEFs and L-929 cells produced IFN-ab following transfection with purified DNA (Fig. 1D and Fig. S10C ). In addition to mouse DCs and macrophages we also found that human monocyte-derived DCs produced IFN-ab upon infection with the adenoviral vector Ad5-GFP (Fig. S3C) . The present data confirm that Ads can trigger IFN-ab production in various immune cells in vitro [10, 45, 46] . However, it furthermore indicates that production does not proceed ubiquitously in all types of infected cells. In order to identify the organ site of viral uptake and IFN-ab production in vivo, we analyzed the expression of the early viral gene E1A and of type I IFN mRNAs, respectively, in the spleen, liver, lung and kidney of Ad3-infected mice. We found expression of the early viral E1A gene in spleen and liver (Fig. S1B ), but not in lung and kidney (not shown), which agrees with earlier findings on in vivo Ad tropism [47, 48] . Expression of IFN-a and IFN-b mRNA was below the level of detection in the organs of non-infected controls. We also found that between 4 and 18 h after infection IFN-a and IFN-b mRNAs were expressed at high levels in the spleen (Fig. 1F ), but surprisingly not in the liver, despite the presence of Ad in both organs. As expected, IFN-ab was not expressed in the virus-free lung or kidney of infected mice (Fig. 1F) . In order to identify the cell type(s) producing IFN-ab in vivo, we isolated splenocytes from Ad infected animals 8 h after virus treatment and separated them into CD11c + (DC-containing) and CD11c 2 (non-DC) populations. Both CD11c + and CD11c 2 populations contained viral DNA (not shown). This finding is in accordance with the report of Morelli et al describing that both splenic DC and non-DC contain the virus in Ad-infected mice [49] . However, quantitative RT-PCR determination of IFN-a and IFN-b mRNA in both populations revealed the presence of IFNab mRNAs predominantly in the DC-containing CD11c + fraction ( Fig. 2A, B ), but not in the macrophage containing CD11c 2 fraction. In contrast to IFN-a and IFN-b, the mRNA levels of IL-6, another cytokine known to be induced by Ad, were comparably upregulated in both the CD11c + and CD11c 2 population (Fig. 2C ). Since the latter non-DC population comprises more than 95% of all mouse splenocytes [50] , we conclude that most of the IL-6 made in the Ad-infected spleen is of non-DC origin. According to published data, the CD11c + cell population contains different types of DCs, as well as other cells such as lymphocytes, macrophages, granulocytes and NK cells [51, 52, 53] . We therefore further separated the purified CD11c + cells into mDCs (CD11c + CD11b + Gr1 2 ), pDCs (CD11c + CD11b 2 GR1 + B220 + ) and a CD11c + CD11b-F4/80 + subpopulation ( Fig. S4A -D) and measured the expression of IFN-a, IFN-b and b-actin with real-time RT-PCR. After normalization to b-actin expression, we found that on a per cell basis mDCs expressed significantly more IFN-b than pDCs (Fig. 2D ), but both mDCs and pDCs expressed similar amounts of IFN-a (Fig. 2E ). In contrast, CD11c + , CD11b 2 cells carrying the macrophage marker F4/80 + did not express detectable amounts of IFN-a or IFN-b. Since mDCs comprise approximately 60% of all analyzed CD11c + splenocytes and their numbers are approximately 10-times higher than those of pDCs (Fig. S5A , B and [50] ), these results suggested that the vast majority of IFN-ab in Ad-infected mice was produced by splenic mDCs. To verify this assumption, we analyzed the Ad-elicited cytokine responses in mice depleted of CD11c high MHC II + myeloid DCs. To ablate these cells, we injected diphtheria toxin into the CD11c-diphtheria toxin receptor CD11cDTR/GFP transgenic mice [54] 24 h prior to infection with Ad. In agreement with previous reports [55, 56] , DT pre-treatment of DTR/GFP transgenic mice resulted in the ablation of CD11c high MHC II + CD11b + splenic mDCs, whereas the CD11c int Siglec H + CD11b 2 plasmacytoid DCs remained unaffected (Fig. S5A, B) . When DT pre-treated CD11cDTR/GFP transgenic mice were challenged with Ad3 and examined for IFN-ab in plasma 4 and 8 h after infection, only marginal IFN responses were found at both timepoints, in contrast to the strong responses of similarly infected transgenic control mice that had not received DT (Fig. 3A) . The same pretreatment with DT had no effect on the Ad elicited IFN-ab response of wild-type mice (Fig. 3A) . Interestingly, the determination of IL-6 levels in plasma 8 h after infection revealed that the ablation of mDCs in CD11cDTR/GFP transgenic mice affected only moderately the induction of IL-6 ( Fig. 3B) , confirming that non-DCs contribute significantly to the Ad-elicited IL-6 response in vivo. The fact that different cell types are responsible for IFN-ab and IL-6 response to Ad may explain at least in part why viral gene expression did not inhibit the production of IL-6 in Ad3-or Ad R700-infected mice (see Fig. 1C and Fig. S2D ). In order to functionally evaluate the possible participation of pDCs to the systemic production of IFNab to Ad we also checked this response in mice depleted of pDCs. As shown in Fig. S6A and B, the injection of anti PDCA-1 antibody led to the substantial decrease of the number of splenic pDCs. Nevertheless, the production of IFN-ab was not changed in response to Ad infection (Fig. 3C) . The data collectively show that the vast majority of IFN-ab but not of IL-6 in Ad infected-mice is produced by splenic mDCs. Recent studies have shown the involvement of different TLRs including TLR 2, 3, 4, 7 and 9 in the innate recognition of different viruses [16, 17, 18, 57, 58, 59, 60, 61] . Signaling via TLR9 was shown to be responsible for the strong type I IFN response of pDCs to Ad in vitro [10, 45, 46 ]. Here we investigated the possible involvement of TLR9 in the induction of type I IFNs by measuring IFN-ab in Ad infected wt and TLR9 2/2 mice. Fig. 4A shows that TLR9 2/2 mice produced normal levels of IFN-ab. upon infection with Ad3. Comparable results were obtained with the recombinant Ad5-GFP (Fig. S7A) . Moreover, Unc93B mice, deficient in signalling by intracellular TLRs showed normal IFN-ab responses to Ad5-GFP (Fig. S7A ). We further checked the possible involvement of the TLR system using mice deficient for TLR2, TLR4 or for the TLR adaptor proteins TRIF and MyD88. Fig. 4A shows that the Ad3 induced IFN-ab levels in all strains of mice were as high as in the respective wt controls. Similarly, comparable IFN-ab responses were also found in TLR-, MyD88-or TRIFdeficient mice and in the corresponding wt animals after infection with UV-inactivated Ad3, AdR700 and Ad5-GFP (not shown). Furthermore, comparable IFN-ab responses were found in cultures of Ad-infected BMDCs from wt, MyD88-and TRIFdeficient mice (Fig. 4B ). The various TLR deficient mice showed impaired responses to the corresponding ligands in control experiments (Fig. S7B ). Collectively, our data show that the TLR system plays no major role in the systemic IFN-ab responses to Ads in mice. The type of virus, the IFN-ab-producing target cell, and the activation mechanism determines whether positive feedback signaling is involved in the induction of the IFN-ab response or not [13, 20] . Here, we studied the possible involvement of IFN feedback signaling in the IFN-ab and IL-6 response to Ad3 by using wt and IFN-abR-deficient mice. The absence of the IFN-ab receptor resulted in dramatically decreased levels of IFN-a and IFN-b protein in the plasma as well as of IFN-a and IFN-b mRNAs in the spleen (Fig. 5A , B) 4 and 8 h after Ad infection. The difference between the protein or mRNA levels of IFN-a and IFN-b in wt versus mutant animals was approximately 100 and 20fold, respectively. Furthermore, in contrast to wt mice, the characteristic rise in IFN-ab levels between 4 and 8 h after infection was absent in IFN-abR 2/2 mice. Thus, in Ad-infected mice, production of both IFN-a and IFN-b is strictly dependent on positive IFN-ab feedback. The determination of IL-6 protein and mRNA levels in the same plasma and splenic tissue samples revealed that the induction of IL-6 is also positively regulated by IFN-ab signaling (Fig. 5C ), which is in agreement with a previous study [10] . We also tested whether IFN-abR-dependent signaling is involved in the IFN-ab response of Ad-infected BMDCs in vitro. As with the in vivo results, we found that cells from IFN-abR 2/2 mice infected with Ad3 produced significantly less type I IFN than similarly infected cells from wt mice (Fig. S8 ). The loss of IFN-ab signaling also resulted in a strong inhibition of the Ad induced IL-6 production in BMDCs (Fig. 5D ) and BMMs (Fig. 5E ). Furthermore, as shown using Ad-infected BMMs, it resulted also in a strong reduction of inducibility of IL-6 mRNA expression (Fig. 5F ). Because transcriptional changes are often determined by epigenetic factors [62] we checked the levels of hyperacetylated histone H4 (acH4), a permissive factor for transcription at the IL-6 promoter in control and Ad-infected BMMs from wt and IFNabR 2/2 mice. Chromatin immunoprecipitation (ChIP) assays showed a significant enrichment of acH4 at the IL-6 promoter in infected wt, but not IFNabR 2/2 BMMs (Fig. 5G) . In a control experiment, as expected, an enrichment of acH4 was observed at the promoter of the constitutively active Topoisomerase 3b but not of the l5 (not expressing, active only in early B-cells) gene in both cells types. From these results we conclude that IFN-ab exerts a positive regulatory effect on the Ad-induced IL-6 transcription and that its loss is at least partly responsible for the strong reduction of the IL-6 response in Ad-infected IFN-abR knockout mice. Using real-time RT-PCR we then analyzed the spectrum of IFN-ab genes in wt mice as well as the impact of IFNabR deficiency on their induction. Included were IFN-a2, 4, 5, 6, 9, 11, 12, 13, 14 and IFN-b. All of them were induced in the spleen by Ad3 in vivo and in BMDC in vitro. IFN-ab subtypes were not detectable in the spleen of unstimulated mice or BMDCs (not shown). Fig. 6 shows the patterns of IFN-ab genes induced in vivo and in vitro in the presence or in the absence of IFN-ab feedback signaling. In wt mice and cells IFN-a5 and IFN-b were the most strongly expressed genes and IFN-a13 was the least activated IFNa subtype. IFN-a11 was not induced at all. IFNabR deficiency resulted in an inhibition of IFN-ab gene expression, the strength of which was subtype dependent. In some cases the inhibition was weak (IFN-a2, 4, 5 and IFN-b), in others strong or complete (IFN-a12, 13, and 14) , showing that the expression of different subtypes of IFN-ab are differentially affected by IFN-ab feedback signaling. Collectively, these data show that the adenovirus triggered production of IFN-ab and IL-6 in BMDCs in vitro and in mice in vivo is strongly dependent on intact IFN-ab signaling. The critical role of IRF-7 but not IRF-3 in Ad-induced type I IFN production The transcription factors IRF-3 and IRF-7 have distinct and important roles in IFN-ab production induced by viruses or other pathogens and their involvement can be characteristic for the induction mechanisms involved [13] . Specifically, IRF-3 has been shown to be critically involved in cytoplasmic DNA sensing and in the Ad-induced IFN-ab production in BMMs in vitro [63] . We show here that IRF-3 is critical for the induction of IFN-ab by isolated adenoviral DNA, but not by infection with whole virions in BMDCs (Fig. 7A ). In order to analyze the individual contribution of IRF-3 and IRF-7 to the Ad-induced IFN-ab response in vivo, we infected mice deficient for these transcription factors with Ad3. We also compared these responses to those triggered by poly I:C, a known activator of the cytosolic IFN-ab producing pathway in vivo. As shown in Fig. 7B , the lack of IRF-3 did not significantly influence the plasma levels of IFNab 4 or 8 h after infection in response to Ad. In contrast, Ad-infected IRF-7deficient mice did not exhibit detectable amounts of IFN-ab in the plasma. Very similar data were obtained with Ad5GFP (Fig. S9) . Thus, IRF-7, but not IRF-3, is essential for the induction of the IFN-ab response during Ad infection. Compared to wt mice, poly I:C injected IRF-7 deficient mice produced significantly less, but still well detectable amounts of IFN-ab. The nucleic acid sensors MDA-5, RIG-I and DAI/Zbp1 are not critical for Ad-induced type I IFN production Next, we investigated whether the cytoplasmic RNA sensors MDA-5 or RIG-I, or the putative DNA sensor DAI/Zbp1 may play a major role in the induction of the IFN-ab response to Ad. The possible involvement of MDA-5 was tested using BMDCs and BMMs from MDA-5 deficient mice. Normal IFN-ab responses to Ad were obtained in MDA-5-deficient BMDCs cells (Fig. 8A ) and also in BMMs (not shown), whereas the responses to the known MDA-5 ligand Poly I:C were abrogated (Fig. 8A) . The possible involvement of RIG-I was checked using a dominant negative form of RIG-I (RIG-IC) [64] . BMDCs from IRF-3 2/2 mice were transfected with a GFP-expressing plasmid (transfection control) with or without RIG-IC and subsequently stimulated with Ad or control leader RNA. We used IRF-3 2/2 cells to avoid any induction of IFN-ab by the plasmid itself, which is, contrary to that by Ad, strictly IRF-3 dependent. The induction of IFN-b mRNA was analyzed in sorted GFP-positive cells. As shown in Fig. 8B , left, the transfection of RIG-IC prevented induction of IFN-b mRNA by the leader RNA, but not by Ad. The role of Zbp1/DAI was analyzed using siRNA-mediated knockdown of DAI/Zbp1 in BMDCs. For this purpose, cells from IRF-32/2 mice were co-transfected with DAI/Zbp1 targeting siRNAs and a GFP expressing plasmid and subsequently stimulated with Ad. As shown in Fig. 8B , right, the transfection of BMDCs resulted in strongly reduced DAI/Zbp1 expression but not in a reduced IFNb mRNA induction by Ad in GFP-positive cells. Control experiments with L-929 cells showed that transfection of the gene-specific siRNA downregulated the expression of DAI/Zbp1 on both the mRNA and protein levels and efficiently inhibited the IFN-b response to B-DNA ( Fig. S10A-C) . Collectively, our results indicate that known nucleic acid sensors such as MDA-5, RIG-I and Zbp1/DAI are not involved in the Ad-induced type I IFN production. MAPKs have been previously shown to be activated by Ad in vitro, in different non-immune cell types and to be important for the induction of chemokines in response to Ad [65, 66, 67, 68] . Here we investigated whether members of the MAPK family play a role in the Ad-induced IFN-ab and IL-6 response. We infected BMDCs with Ad3 in the presence or absence of MAPK inhibitors and determined the levels of IFN-ab and IL-6 produced. Fig. 9A and Fig. S11 show that the pharmacological blockade of the SAPK/JNK MAPK almost completely inhibited the Ad3-induced production of both IFN-ab and IL-6. In contrast, the inhibition of the p38 MAPK pathway partially inhibited the production of IL-6, but had no effect on the production of IFN-ab. Finally, the blockade of ERK1/2 had no effect on the production of either IL-6 or IFN-ab. Very similar data on the effects of MAPK inhibitors were obtained using Ad2 and mutant Ad5-GFP to stimulate BMDC (data not shown). Next, we analyzed the levels of activated SAPK/JNK MAPK proteins in BMDCs and found their robust phosphorylation 2 h after either Ad3 or Ad5-GFP infection (Fig. 9B) . We also tested the importance of SAPK/JNK signaling on Ad-induced IFN-ab production in vivo. Fig. 9C shows that the blockade of the SAPK/ JNK signaling pathway in mice completely inhibited the production of IFN-ab at 4 and partially at 8 h after Ad infection. Taken together, these data strongly indicate that the Ad-activated SAPK/JNK MAPK pathway plays an important role in the virusinduced production of type I IFNs and IL-6. Endosomal escape triggers the Ad-induced production of IFN-ab and IL-6, but prevents TLR9-dependent innate recognition During the course of our adenovirus preparations, we regularly found ''empty capsids'' which we separated and purified in addition to the mature virions. We tested the IFN-ab stimulating activity of these preparations in vivo and observed that they were not active (Fig. S12A ). Since empty capsids lack viral DNA and exhibit an altered protein composition [69] , the absence of an inducing viral constituent(s) from these capsids could explain their inability to provoke an IFN-ab response. Another possible explanation could be that endosomal escape is required for IFNab induction, since empty capsids cannot escape from the endosome [69] . To test the latter possibility, we infected mice with 3.6610 10 viral particles of wt Ad2 and Ad2Ts1, a viral mutant deficient in endosomal escape [5] . As shown in Fig. 10A , in contrast to the wt virus, Ad2Ts1 did not induce detectable levels of type I IFN at 4 and 8 h after infection. Furthermore, the IL-6 response was also severely reduced in these animals (Fig. S12B) . In a control experiment, mice infected with Ad2Ts1 grown at permissive temperature (32uC) and thus capable of endosomal escape, exhibited normal IFN-ab responses (Fig. 10B) . The inability of Ad2Ts1 to escape from endosomes of mDCs was confirmed by electron microscopy (Fig. 10E-H) and in vivo by the lack of Ad early E1A gene expression in the spleen of mice infected with Ad2Ts1 (Fig. 10I) . Thus, escape from the endosome is critical for the induction of IFN-ab and IL-6 by adenoviruses. It should be noted that a further increase in the Ad2Ts1 dose (2.16610 11 particles) resulted in detectable, albeit very low levels of plasma IFN-ab that were released with different kinetics (Fig. 10C) . In this case, IFN-ab was detectable as early as 2 h after infection and, in contrast to the results we obtained with wt viruses ( Figures 1A, 10A and S1A), the levels of IFN-ab did not increase significantly at the later time-points. This already suggests a mechanism for type I IFN induction by Ad2Ts1 that is fundamentally different from the IFN-ab induction seen with wt Ad2. Since Ads are DNA viruses, they can possibly be detected by TLR9. In fact, the innate immune recognition of Ad in pDCs is TLR9 dependent. We therefore repeated the experiment with a high dose of Ad2Ts1 using TLR9 2/2 mice. As shown in Fig. 10C , TLR9 2/2 mice did not produce IFN-ab in response to the mutant Ad2Ts1, quite in contrast to the results obtained with the wt virus (see Fig. 4A ) Likewise, there was no detectable IFN-ab release in Ad2Ts1-infected mice deficient in MyD88, an essential component of TLR9 signaling (not shown). Furthermore, the corresponding amount of empty particles (DNA-free) of Ad2Ts1 elicited no IFN-ab response in wt mice (Fig. 10C) . These data illustrate the critical role of TLR9 in the induction of IFN-ab by means of high doses of Ad2Ts1. To exclude the possibility that contaminating DNA on the surface of the virions was responsible for the TLR9dependent IFN-ab induction, we treated the mutant virions with bensonase, which destroys all kinds of free nucleic acids. Bensonase-treated Ad2Ts1 still induced an IFN-ab response in We also tested the role of endosomal escape of Ad in the in vitro induction in BMDCs of IFN-ab and IL-6. Fig. 10D shows a dose dependent production of IFN-ab by Ad2Ts1 grown at permissive temperature and Figures S12C and D the production of IFN-ab and IL-6 by wt Ad2, but no production of either of the cytokines by Ad2Ts1 grown at the restrictive temperature. Interestingly, liposomal transfection of whole Ad2 Ts1 virions (but not of empty virus particles) in BMDCs resulted in the significant production of IFN-ab that however, was critically dependent on IRF-3 (Fig. 7A) . Similarly, Ad2 Ts1 did not induce IFN-ab production in human monocyte derived DCs (Fig. S12E) . The requirement of low pH for Ad3 infection was also tested using bafilomycin A1, a drug known to inhibit the acidification of endosomes. Experiments shown in Fig. S13A , B revealed that cells treated with this drug produced significantly reduced levels of IFN-ab and IL-6 respectively, in response to Ad3. This is consistent with the notion that Ad3 and Ad7 infection of cultured cells requires low endosomal pH [70, 71] . Similar results were obtained using treatment with ammonium chloride, another acidification inhibitor known to block Ad escape from endosomes (not shown). Collectively, the data in vitro and in vivo provide evidence that the late phase of the Ad infectious entry, in which the virus escapes from the endosome, triggers an innate response characterized by the production of type I IFNs and IL-6. Induction of type I IFNs was shown to be critical for some innate immune responses to Ads [10] .We therefore investigated whether a characteristic consequence of infection with Ad, the induction of LPS hypersensitivity [72] , might be mediated by type I IFNs. For this purpose, we infected wt and IFNabR 2/2 mice with Ads, challenged them 16 h later with LPS and measured the TNF-a response. Non-infected LPS-treated mice served as a control. Unlike the infected wt mice, the Ad3-infected IFNabR 2/2 mice did not exhibit enhanced responses to LPS (Fig. 11A) . Interestingly, LPS hypersensitivity developed also in mice injected with very small amounts of Ad. Such amounts were capable of the elicitation of IFN-b mRNA in the spleen of infected animals, but incapable of inducing detectable circulating IFN-ab ( Fig. S14A and B) . Similar results were obtained in Ad5-GFP-infected mice (not shown). Furthermore, in IRF-7 2/2 mice that exhibit a severe impairment of the IFN-ab response to Ad a severe impairment of LPS sensitization by Ad5-GFP was also observed. (Fig. 11B) . In contrast, sensitization to LPS developed normally in Ad5-GFPinfected TLR9 2/2 mice (Fig. 11B) , which is in agreement with our finding that TLR9 is not critical for the induction of IFN-ab by Ad. Furthermore, only Ad5-GFP-infected wt, but not IFNabR 2/2 mice exhibited enhanced susceptibility to LPS shock (Fig. 11C) . Notably, also Ad vector-infected or IFN-a treated human monocyte-derived mDCs exhibited enhanced sensitivity to LPS and overproduced TNF-a upon LPS challenge (Fig. S14C) . On the whole, our results indicate that IFN-ab is an essential mediator of the LPS hypersensitivity induced by adenovirus infection. An increased expression of the LPS receptor complex on target cells, may contribute to the enhanced reactivity to LPS [73, 74] . We therefore investigated whether macrophages of Ad infected mice overexpress the receptor components mCD14 and TLR4/ MD-2. We found that splenic macrophages from Ad infected mice overexpress mCD14 in an IFN-ab dependent manner (Fig. 11D ) but the expression of TLR4/MD-2 was only minimally affected (Fig. S14D) . Furthermore, we found increased acetylation levels of histone H4 at the TNF-a promoter in Ad-infected wt, but not in IFNabR-deficient macrophages (Fig. 11E ). This suggests that Adinduced IFN-ab increases LPS-induced TNF-a production, at least in part by epigenetic changes at the TNF-a promoter. In the present study we investigated the IFN-ab response of mice infected with human Ads. We showed that the response is characterized by high levels of IFN-a and IFN-b, which are produced simultaneously and almost exclusively by splenic mDCs. Furthermore, the response is entirely independent of viral replication, TLR-, MyD88-, TRIF-and IRF-3-signaling, but dependent on viral escape from the endosomes, activation of the SAPK/JNK pathway and IRF-7, and subsequent IFN-ab feedback signaling. These data suggest an IFN-ab induction pathway that is different from the extracytoplasmic and cytoplasmic pathways so far described. Finally, we show that IFN-ab is a key mediator of the hypersensitivity to bacterial lipopolysaccharide, which develops, in Ad-infected mice. As shown previously and in this study, a wide spectrum of cells including pDCs, mDCs and macrophages [10, 45, 46, 63] produce IFN-ab in response to Ad in vitro. The present finding that in vivo, in Ad-infected mice, splenic mDCs are the major source of IFN-ab is surprising, since in mice infected with different viruses (MCMV, HSV, VSV, MHV, influenza, vaccinia and Sendai viruses) [27, 57, 58, 59, 75, 76, 77, 78] pDCs activated by the TLRs constitute the major contributors to the systemic levels of type I IFNs. In the present study the expression levels of IFN-ab mRNAs in organs and cells from Ad-infected mice suggested a dominant role for splenic mDCs in the IFN-ab response. Furthermore, the IFN response to Ad was practically absent in mice depleted of CD11c high MHC II + myeloid DCs. Also, there was a striking similarity between the IFN-ab subtypes induced by Ads in the spleen of infected mice and those induced in mDC cultures in vitro. It should be emphasized that in mice infected with VSV, in which splenic pDCs are the main IFN producers, the spectrum of in vivo induced IFN-a subtypes was markedly different [76] . Finally, the finding that the response of Ad-infected mice was completely independent of TLR signaling and strongly dependent on IFN-ab feedback provide further arguments against a major role for pDCs. In a number of viral infection models, IFN-a and -b production by pDCs was mediated by TLR7/9 [13, 19, 23] and was at least partially independent of a positive IFN-ab feedback [76, 77, 79, 80] . In variance to our data, pDCs and various types of non-pDCs [10, 42] were suggested to be responsible for the type I IFN responses to adenoviral vectors in mice. In [10] the loss of TLR9 signaling resulted in a reduction of the IFN-a response in mice infected with the recombinant species C Ad-lacZ [10] . This finding suggested a significant contribution of TLR9, and therefore of pDCs, to the Ad-induced IFN-ab response in vivo. The ratio of pDCs to mDCs (approximately 1:1) in the spleen of animals used in the above study was quite different from that we and others [50] have found (approximately 1:10). This may explain the conflicting results on the role of TLR9 in vivo, between this previous study and ours. In another study [42] the induction of IFN-ab was studied in mice with an artificially enlarged pool of DCs (due to prior pretreatment with sFLT-3L), 24 h after recombinant Ad administration. In our study naïve mice were used and 24 h after Ad infection the levels of IFN-ab were already below the detection limit. Of note also, the bone marrow stromal antigen 2 (BST2) (that is recognized by the PDCA-1 antibody used for isolation of pDCs in [42] ), was shown to be up-regulated on numerous cell types following stimulation that triggers an IFN response [81] . Thus, the use of the PDCA-1 antibody for the isolation of pDCs seems to be more reliable in the case of uninfected mice [81] . Recently, an absolute IRF-3 dependency of the in vitro IFN-ab response of bone marrow derived macrophages to Ad has been reported [63] . In our study IRF-3 deficiency had no significant effect on the levels of IFN-ab induced in Ad-infected mice. In addition, in Ad-infected macrophages, the relatively low level of IFN-a4 mRNA and its negative regulation by the autocrine Figure 11 . Adenovirus triggered IFN-ab production is the mediator of LPS hypersensitivity. Wild-type and IFN-abR 2/2 mice (4-6 mice/ group) were infected i. p. with 1610 10 Ad3 particles or left uninfected. 16 h later the animals received 1 mg LPS i. p. or diluent only. TNF-a was determined in plasma 2 h after challenge (A). One representative experiment of three is shown. n.d.: not detectable. IRF-7 2/2 and TLR9 2/2 mice were infected i. p. with 2610 9 Ad5 GFP particles and injected 16 h later with 1 mg LPS i. p. TNF-a was determined in plasma 2 h later (B). IFN-ab signaling is required for Ad infection augmented LPS lethality. Wt and IFNabR 2/2 mice were infected with 1610 10 Ad5 GFP particles i. p. or left uninfected. 16 h later mice were injected with 3.5 mg LPS/gr. bw. The number of dead/total animals are shown at the indicated time after LPS challenge (C). Ad infection increases mCD14 expression on splenic macrophages in an IFN-ab signaling dependent manner. Wt and IFN-abR 2/2 mice were infected with 1610 10 Ad5 GFP particles i. p. or left uninfected. Expression of mCD14 was measured on the surface of F4/80 + splenic macrophages by FACS 16 h after infection (D). Ad infection increases acetylation of histone H4 in wt macrophages at the TNF promoter. Wt and IFNabR 2/2 BMMs were mock-infected or infected with 5400 Ad5 GFP particles/cell and levels of acetylation of histone H4 were measured at the TNF-a promoter with ChIP assays (E). Representative experiments of two are shown. doi:10.1371/journal.ppat.1000208.g011 feedback was different from our in vivo findings (strong induction and positive regulation by feedback). This is in agreement with the concept that in vivo macrophages make no significant contribution to the IFN-ab response to Ad. A likely explanation is that the induction of IFN-ab in macrophages requires the high numbers of virions used in vitro, and that such multiplicities were never reached in vivo. Interestingly however, in contrast to IFN-ab induction, the induction of IL-6 proceeded in the spleen of infected mice in both DC and non-DC populations. In accordance, the depletion of IFN-ab-producing mDCs in mice prior to infection lowered, but did not entirely prevent the IL-6 response. It is possible that the pathways leading to the induction of IFN-ab and IL-6 by Ad are different, at least in different cell types. Alternatively, in vivo, part of the IL-6 formed is induced indirectly via secondary mechanisms. In this context it is interesting that blockade of the p38 MAPK in mDCs in vitro had no effect on Ad-induced IFN-ab production, but partially inhibited the production of IL-6. Our data showing that IFNabR 2/2 mice exhibit strongly impaired IL-6 responses to Ad is in agreement with a previous report [10] and shows that IFN-ab is a positive regulator of IL-6 production. Moreover, our data indicate that this effect could be explained at least in part by the positive regulatory effect of IFN-ab on Ad-induced IL-6 transcription and is achieved by the alteration of the chromatin structure at the IL-6 promoter. The experiments carried out in this study in MyD88-, TRIF-, Unc93B and various TLR-deficient mice excluded a participation of TLRs in the IFN-ab responses to Ads, including the recombinant Ad5-GFP. The only exception was the TLR9-and MyD88-dependent IFN-ab response elicited by Ad2Ts1, a mutant virus deficient in endosomal escape [5] . However, compared to all other Ads used in this study, Ad2Ts1 induced very low levels of IFN-ab, with faster kinetics and only when used in very high amounts. This is consistent with the finding that the mechanisms of type I IFN induction by Ad2Ts1 and the other adenoviruses are not the same. We suggest that the fast escape of Ads from the endosome circumvents activation of endosomal TLRs. The requirement for a longer endosomal retention time in TLR9dependent IFN-ab production has been recently demonstrated [82] . On the whole, our experiments indicate that Ad endosomal escape is required for the induction of IFN-ab in vivo and suggest a cytosolic pathway. The same requirement was ascertained in the present study for the IL-6 response. Sensors of nucleic acids are powerful initiators of the TLRindependent cytosolic IFN-ab induction [12, 14, 15] . We excluded in this study a major involvement of the cytoplasmic RNA sensors RIG-I and MDA-5 in the induction of IFN-ab by Ads. Likewise, our study does not support a major participation of the cytosolic DNA sensor DAI/Zbp1 in this induction either. However, our study did not formally exclude the existence of all potentially redundant recognition pathways. So far reported, the pathway(s) of IFN-ab induction activated by cytosolic DNA is (are) strictly dependent on IRF-3 and minimally on IRF-7 [34, 35, 36, 37, 39] . IRF-3 was also reported to be essential for the adenoviral DNAdependent induction of IFN-ab in BMMs in vitro [63] . In agreement, in this study we show that transfection of BMDCs with naked adenoviral DNA or with whole virions of the endosomal escape deficient Ad2Ts1 results in a strictly IRF-3 dependent IFN-ab response. Evidently however, the IFN-ab response of Ad-infected BMDCs and mice is independent of IRF-3. These findings do not exclude a requirement for dsDNA recognition in the induction of IFN-ab, but suggest a different induction mechanism and show the importance of cellular compartmentalization during normal Ad entry. A possible important factor involved in cytosolic Ad sensing might be the adenoviral cysteine protease L3/23, whose activation requires the presence of Ad DNA [83] . This assumption is supported by our finding that a small fraction of the protease-lacking Ad2Ts1 still reaches the cytosol (Fig. 10F, H) , but is devoid of any IFN-abinducing activity. This enzyme, apart from its involvement in maturation of viral proteins and endosomal escape, is essential for the stepwise disassembly of Ads in the cytosol and the release of viral DNA at the nuclear pore [5] . The absolute requirement of IRF-7 for type I IFN induction in Ad-infected mice shows that this transcription factor participates not only in the positive IFN-ab feedback, but also in the initial IFN-ab production and indeed plays a master role in the regulation of type I interferon response [29] . A further finding of the present study is the essential role for the MAPK SAPK/JNK in the IFN-ab induction by Ads in vitro and in vivo, although the DNA-mediated cytosolic induction of IFN-ab has been reported to be independent of MAPKs activation [14, 15, 36] . A cytosolic dsDNA-signaling pathway, mediated by the RNA helicase RIG-I and MAVS, and leading to the induction of IFN-b has been recently demonstrated in human hepatoma cells [84] . This induction pathway is absent in murine systems [22, 37, 84, 85] and is therefore unlikely to participate in the IFN-ab response of Ad-infected mice. On the whole, our findings do not support the participation of any known nucleic acid-mediated mechanisms in the elicitation of IFN-ab responses to human Ads. In this context it is interesting that, as shown here, mouse embryonal fibroblasts do not produce IFN-ab upon Ad infection, although they posses efficient cytosolic induction pathways for dsRNA or dsDNA [13, 36, 37] , the latter shown also in this study. Rather, our data support the possibility that the IFN-ab response to Ads occurs via a novel, not yet characterized cytosolic DNA or protein recognition pathway. Further studies are required to identify the viral components and the host receptors involved. We also emphasize that the known extra-and intra-cytosolic induction pathways may contribute to the IFN-ab response in a host in which Ads replicate and free viral dsRNA and dsDNA are generated. As mentioned above, the production of IFN-ab in Ad-infected mice is strongly dependent on IFN-ab feedback signaling. Inhibition of positive IFN-ab feedback is a likely explanation for the negative regulation of the IFN-ab response, observed, in mice infected with intact Ads (expressing early genes) in this study. Likely candidates for negative regulators of the IFN-ab production are the E1A proteins. Previously, they were shown to inhibit Stat1 signal transduction [86] , which plays a role in positive feedback signaling by means of the IFN-ab receptor. Because the adenoviral vectors used in this study correspond to some of those used in gene therapy trials, the present findings in mice may have important implications for Ad gene therapy applications. The adverse effects observed in therapeutic trials, such as systemic inflammatory response and toxicity [6, 7, 9] can be at least partly explained by the enhanced susceptibility of the Adinfected host to microbial components, such as LPS [72] and lipopeptides (unpublished results) from incoming secondary pathogens or from the patient's own flora. Our in vitro finding of a strongly enhanced TNF-a response to LPS in Ad5-GFP-infected human DCs suggests that enhanced susceptibility to LPS may develop also in patients treated with adenoviral vectors. As shown here, this hypersensitivity is mediated by viral-induced IFN-ab, which is in accordance with the role of IFN-ab as a key mediator of sensitization to LPS [87, 88] . The increased mCD14 expression on LPS target cells and epigenetic changes on promoters of relevant genes (both shown here), may at least in part explain the role of IFN-ab in the development of Ad induced LPS hypersensitivity. Type I interferon induction was recently found in recombinant Ad-treated human cells, especially in pDCs [46, 89] and in monocyte-derived DCs in the present study. Moreover, this response was observed also in patients administered with recombinant adenoviral vectors [42, 90] . Since as shown here, IFN-a pre-treatment is capable of increasing susceptibility to LPS of human DCs in vitro, we assume that Ad-induced IFN-ab can induce LPS hypersensitivity in humans in vivo. Undesirable complications mediated by IFN-ab can occur not only during gene therapy, but also in immunocompromised patients where Ads are major pathogens [1] . Our finding that blocking SAPK/ JNK signaling inhibits the IFN-ab response to Ads, is of potential interest for prevention or treatment of the direct and indirect adverse effects of IFN-ab in Ad gene therapy. Wt C57BL/6, C57BL/10 and 129Sv mice, as well as all knockout mice were bred under SPF conditions at the MPI. Breeding pairs of IRF-3 2/2 and IRF-7 2/2 mice were kindly provided by T. Taniguchi [54] were on a C57BL/6 background, TLR2/4-deficient mice [95] on C57BL/10 background and IFNabR-deficient mice [96] on 129Sv and for the lethality experiments on C57BL/6 background. When the strain of the mouse was not indicated C57BL/6 mice were used. Mice of both sexes, 8-12 weeks of age were used for the experiments. All of the experimental procedures were in accordance with institutional, state and federal guidelines on animal welfare. Human Ads of species B (Ad serotype 3) and C (Ad R700, an Ad serotype 5 derivative, Ad serotype 2, Ad2Ts1 and Ad5-GFP an early gene expression defective Ad were grown, purified and stored as previously described [5, 72] . The ratio of infectious/total viral particles was determined on susceptible cells and was typically 1:20-50. If not otherwise stated, Ad2Ts1 used for the experiments was grown at non-permissive temperature 38.5uC (resulting in the absence of incorporation of the adenoviral protease L3/23 into viral capsids). Empty particles were identified by their light density in CsCl density gradients and the absence of viral DNA and were purified simultaneously with mature virions. UV inactivation of Ad was done as described previously [5, 72] using the minimal essential dose preventing viral gene expression and replication in susceptible cells. Heat inactivation was done at 56 uC for 60 min [44] . All virus preparations were LPS-free (less than 1 pg LPS/ 10 11 viral particles) as determined by the Limulus amebocyte lysate test (Pyroquant Diagnostic GMBH, Mörfelden, Germany). MEFs, BMMs and GM-CSF induced BMDCs were generated as described [95, 97] . The purity of BMMs was higher than 98%, of BMDCs approximately 80%. BM derived pDCs were generated in the presence of Flt3L and CD11c + CD11b 2 B220 + CD62L + pDCs were MoFlo sorted (purity higher than 95%) as described [55] . MEFs were grown in a-MEM (Invitrogen). Immature, monocyte derived human DCs were obtained by incubating adherent monocytes with GM-CSF as described [98] Human mDCs were infected with the indicated amounts of Ads in growth medium containing 2% of donor serum. Mouse L-929 cells were grown in DMEM with 10% FCS. MAPK inhibitors UO126 (MEK1/2, Cell Signaling), SB203580 (p38, Sigma) and SP600125 (SAPK/JNK Calbiochem) were used at 15 mM in vitro and SP600125 at 20 mg/kg in vivo. Bafilomycin-A1 (Sigma) was used at 100 nM. The plasmids pmaxGFP (Amaxa) and RIG-IC [64, 99, 100] were purified with the Endo-Free Plasmid kit (Qiagen) and PEG purification. Murine IFN-ab activity was measured using an L-929 cell line (provided by B. Beutler and Z. Jiang, Scripps Research institute, La Jolla) as described [18] Human IFN-ab bioactivity was measured using the HL116 cells from G. Uzé as described [101] . The contribution of IFN-a or IFN-b to the total IFN-ab activity was determined by pre-incubating plasma for 1 h with excess amounts of neutralizing anti-IFN-b antibody (Yamasa Corporation, Japan) or control antibody. Murine TNF-a was measured by a bioassay as described [95] . IL-6 was detected with an ELISA from Pharmingen BD. Human TNF-a was detected with an ELISA from R&D Biosystems. JNK/SAPK MAPKs were detected on immunoblots with antibodies detecting all or phosphorylated isophorms of the proteins (Cell Signaling). Zbp1/DAI and b-actin were detected on immunoblots with antibodies from Santa Cruz Biotechnology. Conventional RT-PCR, real-time RT-PCR, analysis of IFN-áâ subtypes and Chromatin Immunoprecipitation Total RNA was isolated from organs and cells with guanidinium-thiocyanate-phenol-chloroform extraction or with TRI reagent (Sigma). To exclude DNA contamination, RNA samples were treated with RNase free DNase I (Fermentas). cDNA was prepared using Expand reverse transcriptase (Roche) and oligo-dT. Conventional RT-PCRs were performed with primers as follows. [102, 103] , detecting all IFN-a mRNAs. To determine the levels of different IFN-ab subtypes the HT7900 quantitative PCR system (Applied Biosystems) was used. cDNAs were measured in duplicates or triplicates using the following genespecific assays (TaqMan Gene Expression Assays, Applied Biosystems): IFN-alpha2 (Mm00833961_s1), IFN-alpha4 (Mm00833969_s1), IFN-alpha5 (Mm00833976_s1), IFN-alpha6 (Mm02524285_g1), IFN-alpha9 (Mm00833983_s1), IFN-alpha11 (Mm01257312_s1), IFN-alpha12 (Mm00616656_s1), IFN-al-pha13 (Mm00781548_s1), IFN-alpha14 (Mm01703465_s1),IFNbeta (Mm00439546_s1). The gene for mouse hypoxanthine guanine phosphoribosyl transferase-1 (HPRT-1, Mm00446968_m1) was used to calibrate the mRNA levels. Quantitative analysis was performed using the SDS 2.1 software (Applied Biosystems). mRNA levels were calculated by the following formula: relative expression = 2'(2(Ct(Target)-Ct(Endogenous control))*f, with f = 10 000 as an arbitrary factor. Chromatin Immunoprecipitation (ChIP) assays were performed as described [104] . Briefly, cells were cross-linked with 1% formaldehyde and sonicated nuclear extracts were mock-immunoprecipitated or immunoprecipitated with anti-tetraacetylated histone H4 (Upstate). Recovered DNA aliquots from these samples and from input extracts were amplified with real-time PCR using the LightCycler II system (Roche) and Quantitect SYBR Green PCR Kit (Qiagen). Enrichments at specific chromatin loci are shown as the amount of immunoprecipitated DNA in the percent of total input chromatin. The following primers were used: IL -6 promoter: TGG GGA TGT CTG TAG CTC ATT and CAT AGC GGT TTC TGG AAT TGA, TNF promoter: GGG CAG CCC CAG AGG GAA TGA ACT C and TAT GGC AGA GGC TCC GTG GAA AAC TCA CT, Topoisomerase 3b promoter: AGT CCG AGA ACA GCC TGG GT and AGT TGT GCT GCC CAC AGA GG, l5 promoter: TCC CCA TTG CCA GAT AGA GAC ACA and TGG GCC CAA CAG ATT AAC ACA GAG. BMDCs were cold synchronized with saturating amounts of Ad2 and Ad2-ts1 (60 mg/ml, 0.25 ml per 4610 4 cells on a 12 mm glass coverslip) for 1 h, washed and incubated at 37uC for the indicated times. The samples were fixed with 2.5% glutaraldehyde in 0.1 M ice-cold Na-Cacodylate buffer (pH 7.2) containing 0.5 mg/ml ruthenium red for 1 h, washed with 0.1 M Na-Cacodylate buffer (pH 7.2), post-fixed with 2% OsO4 in the same buffer containing 0.5 mg/ml ruthenium red for 1 h at room temperature, and embedded in Epon as described [105] . Virus particles at the plasma membrane, endosomes and the cytosol were determined, and results expressed as means of analyzed cells (n) with standard errors of the mean. List of accession numbers/ID numbers for proteins mentioned in the text

Hemorrhagic shock and encephalopathy syndrome – the markers for an early HSES diagnosis

BACKGROUND: The hemorrhagic shock and encephalopathy syndrome (HSES) is a devastating disease that affects young children. The outcomes of HSES patients are often fatal or manifesting severe neurological sequelae. We reviewed the markers for an early diagnosis of HSES. METHODS: We examined the clinical, biological and radiological findings of 8 patients (4 months to 9 years old) who met the HSES criteria. RESULTS: Although cerebral edema, disseminated intravascular coagulopathy (DIC), and multiple organ failure were seen in all 8 cases during their clinical courses, brain computed tomography (CT) scans showed normal or only slight edema in 5 patients upon admission. All 8 patients had normal platelet counts, and none were in shock. However, they all had severe metabolic acidosis, which persisted even after 3 hours (median base excess (BE), -7.6 mmol/L). And at 6 hours after admission (BE, -5.7 mmol/L) they required mechanical ventilation. Within 12 hours after admission, fluid resuscitation and vasopressor infusion for hypotension was required. Seven of the patients had elevated liver enzymes and creatine kinase (CK) upon admission. Twenty-four hours after admission, all 8 patients needed vasopressor infusion to maintain blood pressure. CONCLUSION: CT scan, platelet count, hemoglobin level and renal function upon admission are not useful for an early diagnosis of HSES. However, the elevated liver enzymes and CK upon admission, hypotension in the early stage after admission with refractory acid-base disturbance to fluid resuscitation and vasopressor infusion are useful markers for an early HSES diagnosis and helpful to indicate starting intensive neurological treatment.

Since the original description of the hemorrhagic shock and encephalopathy syndrome (HSES) by Levin et al. [1] , numerous cases have been reported in the literature. Although the etiology of HSES remains unknown, this syndrome is associated with acute onset of encephalopathy, shock, watery diarrhea, severe disseminated intravascular coagulopathy (DIC), and renal and hepatic dysfunction. As some authors have defined the HSES criteria [2] [3] [4] , patients meeting them will usually have very poor prognoses with a fatal course or severe neurological sequelae. Our experience suggests that early detection of HSES plays an important role in survival and the reduction of neurological sequelae. We described the clinical courses of 8 patients (age range, 4 months to 9 years) who met the HSES criteria of Bacon et al. [3] , were admitted to our Intensive Care Unit between November 2001 and August 2007, and whose patient records were reviewed to detect markers for an early diagnosis of HSES. Patients were excluded if they had an elevated plasma ammonium concentration (>130 μmol/l), historic evidence of Reye's syndrome, inadvertent heating, features of the staphylococcal toxic shock syndrome and/or the hemolytic-uremic syndrome, or if any recognized bacterial pathogens or metabolic disorders were discovered that would explain the illness. Status epilepticus was defined as an epileptic seizure or seizures lasting more than 30 minutes or recurring within 30 minutes without recovery of consciousness. The biological investigations performed for all 8 patients included white blood cell (WBC) counts, C-reactive protein (CRP), platelet counts, hemoglobin, asparate aminotransferase (AST), and alanine aminotransferase (ALT), creatinine, base excess, creatine kinase (CK), and cell counts of cerebrospinal fluid (CSF). Serum lactate level was measured in 6 patients. DIC was defined as a decreased platelet count and an increase in fibrinogen/ fibrin degradation production. Metabolic acidosis was defined as base excess (BE) lower than -3 mmol/L. Lumbar puncture (LP) was performed upon admission for all patients except one whose LP was performed at the previous hospital and could not be performed at our center because of the CT findings of moderate cerebral edema. Specimens of blood, urine, stool, sputum, and CSF of all 8 patients were obtained to determine any bacterial and/ or viral agents as soon as possible after admission. Computer tomography (CT) was performed upon admission, electroencephalogram (EEG) was done within 3 hours of admission, and CT was repeated for each patient the following day. All 8 patients needed mechanical ventilation due to coma and/or seizure. After admission, they required continuous diazepam or barbiturate infusion for seizure or brain edema with the head of the bed elevated 30°. When hypotension was recognized, fluid resuscitation with a crystalloid or colloid and norepinephrine infusion was set up. The fluid resuscitation target was a central venous pressure (CVP) of ≥ 8 mmHg, urine output >1 ml/ kg/hr. Norepinephrine infusion was started when hypotension was refractory with fluid resuscitation. After these cultures were obtained, all the patients took broad-spectrum antibiotics until their bacterial infections were resolved. The changes of base excess and serum lactate levels were expressed as mean ± SD. When a median was used, the range was given. The Osaka City General Hospital ethics committee approved this retrospective analysis of patients' data and informed consent was obtained from the patients' next of kin. All 8 patients were admitted comatose or with febrile convulsions. Five patients had a history of diarrhea and/or vomiting. The patients' ages ranged from 4 months to 9 years old (median, 1.6 years). Six patients had normal neurological development, however, 2 patients had previously been diagnosed as epileptic. Four patients were admitted in winter, between December and February. Seven patients were transferred to our center within 24 hours after the onset of coma or convulsions. Three patients survived (Table 1 ). All 8 patients had normal platelet counts, blood pressure and a normal or slightly elevated CRP level; 7 patients had a normal hemoglobin level and renal function. All 8 patients had metabolic acidosis and abnormal serum lactate levels. Seven patients had slightly or significantly elevated liver enzymes, CK, and abnormal WBC counts. Bacterial cultures of blood and CSF were negative for all the patients, however, viral pathogens were detected by PCR in 4 patients (Tables 2, 3) . Although abnormal cerebral edema was seen in all the patients during their clinical courses, 5 patients appeared normal or only slightly edematous as revealed on their brain CT scans upon admission. On the initial EEG, multi-focal paroxysmal discharges were seen in 4 patients, and low-amplitude patterns were seen in 4 other patients. The CSF cell counts were within a normal range in 7 patients, while the serum level of IL-6 and soluble IL-2 receptors increased with varying ranges in all the patients ( Table 4 ). All 8 patients had hemodynamic failure within 24 hours after being admitted; therefore, fluids were infused to maintain arterial pressure with the range of fluid balance from -6 to 275 ml/kg (median, 61 ml/kg) for 24 hours from admission for hypotension. Norepinephrine was given to all patients ranging from 0.1 to 0.5 μg/kg/min (median, 0.3 μg/kg/min). Twenty-four hours after admission, 6 patients had normal renal function, and 4 patients had normal platelet counts. However, 5 patients exhibited a decrease in hemoglobin ( Table 5 ). All the patients exhibited a severe metabolic acidosis with the BE range from -16.0 to -4.4 mmol/L (median, -10.3 mmol/L) upon admission. The acid-base disturbances were maintained with the BE range from -14.4 to -4.1 mmol/L (median, -7.6 mmol/L) at 3 hours, and from -15.2 to -3.1 mmol/L (median, -4.7 mmol/L) at 12 hours after admission with infusion of fluids and/or norepinephrine. The metabolic acidosis was refractory to intensive treatment with mechanical ventilation, infusion of fluids and/or norepinephrine at 24 hours with the BE range from -8.3 to -3.1 mmol/L (median, -4.9 mmol/L; Figure 1 ). Sodium bicarbonate for metabolic acidosis was not administered because the blood pH was kept in the normal range (7.35-7.45) with effective ventilation. Similarly, there was a tendency to maintain the elevated serum lactate levels, which were measured in 6 patients, with the range from 2.2 to 11.5 mmol/L (median, 4.2 mmol/L) upon admission, from 2.3 to 11.8 mmol/L (median, 6.0 mmol/L) at 12 hours after admission, and from 2.1 to 10.8 mmol/L (median, 6.2 mmol/L) at 24 hours after admission. CT scan revealed abnormalities between 24 and 72 hours after onset of coma or seizures ( Figure 2 ). When the CT revealed bilateral cortical and subcortical areas of low density, those patients had DIC, anemia and multiple organ failure. Nevertheless, the respiratory function was maintained during the clinical course of all patients. We started intracranial pressure (ICP) monitoring (REF 110-4BT, Camino, USA) in Cases 1 and 2. In Case 1 the ICP monitoring was started from when the abnormal CT finding was discovered; however, in Case 2, it was started immediately after admission, i.e., before an abnormal CT In 1983, Levin et al. described a devastating disease occurring in early childhood called the hemorrhagic shock and encephalopathy syndrome [1] . HSES is not a common disease. Sofer et al. reported 20 patients diagnosed with HSES in a population of about 400,000 over an 11-year period [5] . Our center at Osaka City General Hospital is one of two tertiary pediatric centers in Osaka prefecture. The area our center services for primary referrals has a population of about 4 million residents. However, we do not in any way hypothesize that the 8 patients in our 6year study were the only cases of HSES in this large region, as compared with the 20 patients in the 11-year Sofer et al. report of a significantly smaller population [5] , as there were most likely other HSES patients who could not be transferred to our center from other hospitals due to the rapid deterioration of their conditions. Most reported cases of HSES occurred in winter [3, 5] ; as in the present study, further supporting this evidence, 50% of our patients were admitted in winter, between December and February. The outcome of HSES was often fatal or with severe neurological sequelae [2] [3] [4] [5] [6] . In the present study, 5 patients died, and the 3 survivors had neurological sequelae. The HSES criteria have been defined in previous reports [2] [3] [4] . The clinical presentation includes shock, coma and/or seizure, hemorrhage, diarrhea, and oliguria. Laboratory investigations reveal decreased hemoglobin and platelet counts, evidence of DIC, elevated creatinine, AST and ALT, and metabolic acidosis. However, when patients met the HSES criteria, their condition was always critical with multiple organ failure. We determined that these HSES In these cases, the first problem the physician is faced with is the difficulty of the differential diagnosis. Some reports have indicated in the differential diagnosis the diseases of the toxic shock syndrome, the hemolytic-uremic syndrome, and Reye's syndrome, among others [6] [7] [8] . In our patients, those diseases were excluded because of the lack of skin or mucosal manifestations, hemolysis, and/or blood ammonia levels. The clinical course of our patients was not indicative of any of these diseases. Heatstroke has similar clinical and pathological features to HSES; however, there was no history of over wrapping or excessive heating in any of our patients in the present study. The most common first symptoms were seizure or coma following fever in our cases. From our experience, we have found that common febrile convulsion is often the most difficult disease in the differential diagnosis of HSES in the early stage. Concerning febrile convulsions: prolonged seizures lasting 5 to 10 minutes are relatively common, the laboratory data including CSF is near normal, CT scans reveal normal or slightly edematous conditions, and elevation of CK levels and WBC counts are often seen in cases with prolonged seizures [9] [10] [11] [12] . Even though followup CT scans could provide useful information about cerebral edema [13] [14] [15] , our results showed that the initial CT finding was not useful in making the HSES diagnosis. When an abnormality on the CT was discovered, the patient's condition was critical. Harden et al. have reported the importance of EEG features and evolution [16] , as confirmed by the observations in the present study, the initial EEG features appeared abnormal in all patients. However, in the initial EEG features, it is usually difficult to distinguish HSES from other diseases with convulsions including febrile convulsions. Rosman has reported that the initial EEG features in patients with febrile convulsions are abnormal in as many as 88% of the patients [17] . Dunn reported that the outcome of the status epilepticus was not associated with acidosis on admission [9] . Imuekemhe et al. reported that mean serum lactate on admission was significantly higher in patients with prolonged febrile convulsions compared to the corresponding mean value in patients with only brief convulsions [18] . Conversely, Levin et al. noticed metabolic acidosis in HSES patients upon admission [2] . Ince et al. also reported that metabolic acidosis was the common laboratory value in HSES patients upon admission [15] . Little et al. reported a marked metabolic acidosis being refractory to fluid-resuscitate in HSES patients [6] . And Idro et al. reported that the level of base excess of < -8 mmol/l was a prodromal risk factor for death among children with acute seizures [19] . In the present study, deterioration of the patients' conditions, especially hemodynamic failure, was dramatic up to Hours after admission 24 hours after admission. The most effective treatment for metabolic acidosis with high levels of serum lactate is the adequate and timely treatment of fluid resuscitation and vasopressor [20, 21] . All patients needed large amounts of fluids and/or norepinephrine infusion. The median rate of fluid administration needed was 61 ml/kg for 24 hours with the infusion of norepinephrine. However, neither metabolic acidosis nor abnormal serum lactate improved in this study. Sepsis from bacterial infection was excluded by the negative bacterial cultures and the normal CRP levels. However, the hemodynamic course of our patients was very similar to severe septic shock. The etiology of HSES is still unknown. It has been reported that cytokine storm may be associated with the progress of acute encephalopathy including HSES [22] [23] [24] as septic shock, and levels of some cytokines were useful markers for HSES [25, 26] . The serum levels of IL-6 and soluble IL-2 receptors were increased in our patients, however, the degree of the increase varied in each patient. These results suggest that the increase in cytokines may be associated with HSES. As septic shock is characterized by severe vascular leakage, this is the reason that large amounts of fluids and/or norepinephrine infusions were needed for the patients in this study. As previously expressed, the respiratory functions of the HSES patients in the present study were maintained throughout their clinical courses. This is the most essential difference between HSES and sepsis in the state of multiple organ failure. Sepsis-induced acute lung injury (ALI)/acute respiratory distress syndrome (ARDS) has been reported [27] , and the inflammatory response in ALI/ARDS is associated with the release of cytokines [28] . A limitation of the present study was its small population of only 8 children. Therefore, we did not find a definitive reason the cytokine storm with HSES did not have a sig- The changes of CT findings at the level of the basal ganglia of Case 1 Figure 2 The ICP control was difficult when monitoring was started after the abnormal CT finding was discovered in Case 1, therefore, the maximum ICP was increased to 109 mmHg. ICP monitoring started before the presence of abnormal CT findings in Case 2 in which the maximum ICP was increased to 59 mmHg; however, the CPP could be maintained above 50 mmHg. This patient was the only case with mild sequelae among all our cases. Hospital days nificant influence on the pulmonary vascular permeability, as would a prodromal marker of ALI/ARDS. It is our supposition that this difference is a key point that would detect the etiology of HSES. Further studies are warranted to discover what kinds of cytokines would be the most useful markers. The current problem that we face in our center is that it takes considerable time and effort to get the necessary results for these cytokines. We detected viral pathogens in 4 of 8 patients in the present study. Viral infection may be the trigger for the pathogenesis of HSES as reported in the case report by Gooskens et al. [22] . It may be associated with the most common season of HSES -winter. Even though there were abnormalities in coagulation, hemoglobin, and renal function, they were useful to make a diagnosis of HSES, it was evident that these laboratory abnormalities were not appreciated as diagnostic markers until at least 24 hours after admission [14, 29] . Therefore, we did not have sufficient time to make a proper diagnosis of HSES using the established criteria [2] [3] [4] . A diagnosis of HSES ought to be made within 24 hours of admission, otherwise the patients' conditions worsen and the window to provide adequate treatment closes. The second problem is the treatment of patients with HSES. Even though prolonged metabolic acidosis and/or high levels of serum lactate refractory to large amounts of fluids and/or norepinephrine infusion are useful markers for an early diagnosis, they are not useful as markers when only respiration and circulation management is provided. Because the brain appears to be the main target organ of HSES, in our non-surviving patients, severe diffuse brain edema with loss of differentiation between the gray and white matter was found on the CT scan during the clinical course. And similar CT abnormalities have been described in non-surviving patients in other reports [13, 15] . The cause of brain edema following HSES remains unclear, however, Unterberg et al. reported that brain edema of traumatic brain injury was associated with various mediators including cytokines, lactate, free oxygen radicals, etc. [30] . Brain edema in HSES patients seems to occur in such a situation. Because the serum levels of cytokines and lactate had increased, an increase in vascular permeability was suggested by our patients needing large amounts of fluids. Furthermore, large amounts of fluid-resuscitate within 24 hours of admission may lead to brain edema under the state of increasing vascular permeability. We propose that for the most efficacious management of ICP, whenever possible, ICP monitoring ought to be started before detection and observation of any decrease in the platelet count and/or any abnormal CT findings because ICP monitoring could not be performed with DIC. This is the reason that a diagnosis of HSES should be made within the early stage, i.e., within 24 hours of admission. However, to our knowledge, there have not been any reports published proposing an effective treatment for HSES. Controlling brain edema might be the optimal therapy to help HSES patients survive. Even though there are currently only palliative therapies, e.g., mild hypothermia, infusion of fluids and osmotic diuretics, administration of anticonvulsants under mild hyperventilation, and vasoconstrictor infusion to prevent edema and to maintain the cerebral perfusion pressure (CPP). There are no reports about the efficacy of the control of ICP and CPP upon the outcome of the HSES; however, we suggest that the control of ICP and CPP is an essential part of any therapeutic treatment. A patient characterized by coma or seizure following hyperpyrexia might be diagnosed as having common febrile convulsions. However, when such a patient also presents with elevated liver enzymes and CK upon admission, hypotension within 24 hours after admission, with refractory acid-base disturbance and an abnormally high serum lactate level, even with fluid-resuscitate and/or vasopressor infusion, these signs may be useful markers for an early HSES diagnosis and indicators to start intensive neurological treatment. HSES is not a disease that can be diagnosed easily using the current diagnostic criteria, however, HSES can be predicted in the early stage of its clinical course using these new prodromal diagnostic markers. • When the patients met the HSES criteria, their condition was always critical with multiple organ failure. • CT scan, DIC, EEG, or renal function upon admission did not prove useful for an early diagnosis of HSES. • Elevated liver enzymes and CK upon admission, hemodynamic failure in the early stage after admission, and a prolonged metabolic acidosis refractory to intensive treatment were useful markers for an early diagnosis of HSES. • Controlling brain edema might be the most important therapy to help HSES patients survive. • HSES is a disease that should be predicted within the early stage of its clinical course.

Unifying evolutionary and thermodynamic information for RNA folding of multiple alignments

Computational methods for determining the secondary structure of RNA sequences from given alignments are currently either based on thermodynamic folding, compensatory base pair substitutions or both. However, there is currently no approach that combines both sources of information in a single optimization problem. Here, we present a model that formally integrates both the energy-based and evolution-based approaches to predict the folding of multiple aligned RNA sequences. We have implemented an extended version of Pfold that identifies base pairs that have high probabilities of being conserved and of being energetically favorable. The consensus structure is predicted using a maximum expected accuracy scoring scheme to smoothen the effect of incorrectly predicted base pairs. Parameter tuning revealed that the probability of base pairing has a higher impact on the RNA structure prediction than the corresponding probability of being single stranded. Furthermore, we found that structurally conserved RNA motifs are mostly supported by folding energies. Other problems (e.g. RNA-folding kinetics) may also benefit from employing the principles of the model we introduce. Our implementation, PETfold, was tested on a set of 46 well-curated Rfam families and its performance compared favorably to that of Pfold and RNAalifold.

where r(σ) is the root node of τ M (σ). Since we are not using the parse tree τ M (σ) explicitely in the main text, we will write Pr(r(σ), A) as short for Pr τM (σ) (r(σ), A). PETfold uses a Nussinov style algorithm to calculate the consensus structure of an alignment with maximal expected overlap. The Nussinov algorithm uses dynamic programming to find the structure with the highest score. Let F (i, j) denote the maximal score of an RNA structure for the sequence s i . . . s j . Thus, we have where s(x i ) (and s(x j )) is the score for a single-stranded position x i and s(x i , x j ) is the score for paired bases x i and x j . In PETfold the single-stranded score of position x i consists of the evolutionary reliability R sg A,T,M (i) and the thermodynamic probability 1 n u q u f −1 A (i) over all sequences s u (1 ≤ u ≤ n) in the alignment, and the base pair score of the positions x i and x j consists of the evolutionary reliability R A,T,M (i, j) and the thermodynamic probability 1 n u p u f −1 A (i,j) . The optimal structure σ can be reproduced by backtracking from F (1, L) when L is the sequence length. In PETfold, we define ex-over(σ) = F (1, L). We present a statistical method to estimate reliability thresholds for conserved functional regions. Single stranded positions and base pair positions are collected that have a high evolutionary reliability. We write down only the base pair part. Single-stranded positions are treated analogously. For this purpose, do the following 1. Generate shuffled alignment A shuffle by shuffling the alignment columns. Then, we generate again the most likely structure under the shuffled alignment, i.e., we generate Then, we collect all the reliability scores for base pairs that are contained in this structure, and iterate this several times: Finally, we order them in size p 1 > p 2 > · · · > p |B| and select a significance level θ (e.g., θ = 0.01). Then the probability p ⌈θ|B|⌉ is the base pair probability p threshold such that any base pair We applied the previously described stepwise approach on our data set consisting of 46 RNA families. We shuffled for each family 1000 times with a conservative method which mononucleotidely shuffles only columns with the same pattern of gaps and conservation. Then we averaged over the significance values of all families. Using a significance level θ = 0.01, we got a threshold for high reliable base pairs of p threshold However, the parameter tuning has indicated that the performance of reliability thresholds depend on another parameter (the weighting factor for single-stranded positions α) which has high impact in the RNA structure prediction of PETfold, and that slightly different reliability thresholds perform better for the data set. Given two structures in bracket notation, a more stringent secondary structure evaluation can be carried out by considering all pairs of positions, and evaluate the agreement in their structural notation (i.e., dots, opening and closing brackets) in both structures. For each pair of positions (i, j), there are five possible cases. The two positions can be unpaired (4) or paired with each other (1) . Furthermore, only the left (2) (resp. right (5)) position can have an opening (resp. closing) bracket. Finally, both positions can be paired, but with different partnersi (3). Formally, we have the following five categories (K = 5): (1) This can be evaluated by the R K correlation coefficient (K = 5) [1] . This correlation coefficient of two assignments represented by two N × K matrices of data X and Y is defined as . The covariance between X and Y is defined as the expected covariance between the respective k th columns X k and Y k in the matrices: where X k = (1/N ) N n=1 X nk and Y k are the respective means of column k, and X nk are elements of X. Note that Matthews correlation coefficient (M CC) applies to the two categories (K = 2) base paired (i bp j) and not base paired (i ¬bp j) for any pair of bases (N = M (M − 1)/2 where M is length of sequence). Correction for sliding base pairing is not used. When extending the consideration of unpaired bases, we obtain R 5 correlation coefficients of PETfold: 0.72, Pfold: 0.58, RNAalifold: 0.65. This evaluation is more strict as the two-category Matthews correlation coefficient. Nevertheless, both evaluations show almost the same differences between the three methods. SI Table 1

A sensitive array-based assay for identifying multiple TMPRSS2:ERG fusion gene variants

Studies of gene fusions in solid tumors are not as extensive as in hematological malignancies due to several technical and analytical problems associated with tumor heterogeneity. Nevertheless, there is a growing interest in the role of fusion genes in common epithelial tumors after the discovery of recurrent TMPRSS2:ETS fusions in prostate cancer. Among all of the reported fusion partners in the ETS gene family, TMPRSS2:ERG is the most prevalent one. Here, we present a simple and sensitive microarray-based assay that is able to simultaneously determine multiple fusion variants with a single RT–PCR in impure RNA specimens. The assay detected TMPRSS2:ERG fusion transcripts with a detection sensitivity of <10 cells in the presence of more than 3000 times excess normal RNA, and in primary prostate tumors having no >1% of cancer cells. The ability to detect multiple transcript variants in a single assay is critically dependent on both the primer and probe designs. The assay should facilitate clinical and basic studies for fusion gene screening in clinical specimens, as it can be readily adapted to include multiple gene loci.

Chromosome rearrangements are a characteristic feature of cancer. More than 350 gene fusions, as a consequence of chromosome aberrations, have been identified (1) . While gene fusions are common in hematological malignancies, their presence in solid tumors is not as well studied due to several technical and analytic problems related to tumor heterogeneity (1) . Only very limited gene fusion events were discovered in solid tumors, mostly in sarcomas, until the recent discovery of TMPRSS2:ETS fusion genes in prostate cancer (2) . This finding has since changed the general view that gene fusions play only a minor role in the pathogenesis of epithelial tumors. Therefore, there is renewed interest in searching for fusion genes in solid tumors, due to their potential impact on basic research and clinical application as has been demonstrated in chronic myelogenous leukemia (CML) (3, 4) . The recurrent gene fusion event in prostate cancer involves an androgen controlled gene, TMPRSS2, and members (ERG, ETV1 and ETV4) of the ETS transcription factor family (2, 5, 6) . Among these fusion genes, TMPRSS2:ERG is the most prevalent and the only member detected in the majority of reports. This fusion transcript results from $3 Mb interstitial deletion between these two loci at chromosome 21q22. It was found in approximately half (15-78%) of all prostate cancers (2, (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) (16) (17) . As an androgen-related transcription factor controlling cell proliferation, TMPRSS2:ERG has been associated with disease pathogenesis and is a promising biomarker for prostate cancer progression, prognostication and early detection (18) (19) (20) (21) . While the presence of TMPRSS2:ETS fusion genes is highly prostate cancerspecific, its significance as a prognostic biomarker is still controversial partly because many of the clinical studies have been relatively small scale. Therefore, it is important to develop a simple and robust assay for identifying various TMPRSS2:ETS and potential fusion genes in other solid tumors. However, this could be challenging due to high heterogeneity in prostate cancer and other solid tumors, compared to leukemias and lymphomas (22) . Several approaches that have been used previously for hematological malignancies have been applied to detect TMPRSS2:ERG exon fusion variants. These include fluorescent in situ hybridization (FISH) (2, 12, 14, 17, 23) , RT-PCR and sequencing (2, 7, 9, 13) , quantitative PCR (qPCR) (2, 8, 24) and array-based comparative genome hybridization (array CGH) (10) (11) (12) . FISH may be the most commonly used method, but it has relatively low resolution, and therefore, cannot accurately determine different fusion variants. Array CGH has a higher resolution but is costly and often fails when there is normal cell contamination. RT-PCR and qPCR are relatively easy to perform. However, to assess multiple potential fusion variants requires multiple sets of primers and probes, and a corresponding large quantity of RNA templates. Moreover, sequencing RT-PCR products is laborious and difficult to adapt in routine clinical laboratories. Here, we describe an exon array-based detection system, combined with a RT-PCR reaction, that accurately determines multiple TMPRSS2:ERG fusion transcripts in specimens with only a minor population of tumor cells. The method adopts several features of the Virochip (25) protocol to establish a specific, sensitive and semi-quantitative assay that is very useful for analyzing highly heterogeneous solid tumors. The cell lines described in the article were obtained from the American Type Culture Collection (ATCC, Manassas, VA, USA) and cultured as recommended. Frozen unpurified prostate tissues were obtained during routine surgery, and classified pathologically by one of us. The total RNA was extracted with Trizol (Invitrogen, Carlsbad, CA, USA) following the manufacturer's instructions. The primary tumor samples were purified by Qiagen RNeasy mini kit (Qiagen, Valencia, CA, USA). The exon and junction probes are 30-mer oligonucleotides synthesized by Integrated DNA Technologies (Coralville, IA, USA) or Illumina/Invitrogen (San Diego, CA, USA) and printed on poly-L-lysine slides at 50 mM along with Human Cot-1 DNA (Invitrogen), which is enriched for repetitive sequences, and herring sperm DNA (Promega, Madison, WI, USA), which was used as nonspecific control. The printing procedure has been described and essentially follows the manual of the DeRisi arrayer with silicon microcontact printing pins (Parallel Synthesis Technologies, Inc. Santa Clara, CA, USA) (25) (26) (27) . Arrays were postprocessed with succinic anhydride-based method for blocking before hybridization as previously described (27) . The protocols related to array printing and hybridization in this article generally can be found in the following link: http://cat.ucsf.edu/equipment/arrayer/index.html. The RT-PCR reaction was performed with an OneStep RT-PCR kit (Qiagen) essentially following the manufacturer's protocol, except that the final reaction volume was scaled down to 20 ml. The forward (GTT TCC CAG TCA CGA TCC AGG AGG CGG AGG CGG A) and reverse primers (GTT TCC CAG TCA CGA TCG GCG TTG TAG CTG GGG GTG AG) are located at exon 6 of ERG and exon 1 of TMPRSS2 respectively, as described (2, 9) . The 5 0 -ends of both primers have the sequence of primer B (GTT TCC CAG TCA CGA TC) for the subsequent step of PCR labeling with a single primer B as described previously (25) . The procedure is a modification of the previously reported 'Round ABC' protocol (25) . Briefly, the RT-PCR reaction was assembled at 48C in a PCR workstation and transferred to a thermocycler with the block preheated to 508C. The initial reverse transcription was performed at 508C for 30 min and followed by 958C for 15 min to activate HotStarTaq DNA polymerase as well as to inactivate the reverse transcriptases (Round A). The PCR conditions were 35 cycles at 928C for 30 s, 558C for 30 s and 688C for 1.5 min with a final extension step at 688C for 5 min (Round B). One microliter of unpurified product was subsequently used as a template for another 20 cycles of amplification to label the amplicons via a previously described 'Round C' PCR protocol (948C for 30 s, 408C for 30 s, 508C for 30 s and 728C for 1 min) with primer B and a 4:1 mixture of aminoallyl dUTP (Ambion, Austin, TX, USA) and dTTP for probe labeling (25) . The labeled amplicons were purified with DNA Clean-up and Concentrator-5 columns (Zymo Research, Orange, CA, USA), eluted in 9 ml of sodium bicarbonate (pH 9.0) and coupled with 1 ml of DMSO dissolved Cy3 NHS esters (GE Healthcare, Piscataway, NJ, USA) for 30-60 min. The Cy3-labeled amplicons were purified with DNA Clean-up and Concentrator-5 columns and eluted with 10 ml of 10 mM Tris-HCl (pH 8.0). Then, the Cy3-labeled amplicons were diluted in water and combined with 3.6 ml of 20Â SSC, 0.5 ml of HEPES (pH 7.0) and finally 0.5 ml of 10% SDS to reach final volume of 25 ml. The mixed solution was heated for 2 min at 958C, cooled to room temperature and hybridized to the exon mapping arrays at 638C overnight essentially as previously described (25) (26) (27) . The hybridized arrays were washed and scanned with a GenePix 4000B scanner (Molecular Device, Sunnyvale, CA, USA) and analyzed by GenePix Pro 6.0 software. Frozen prostate cancer samples were sectioned onto slides. Cell nuclei were isolated in situ with ice-cold cytoskeleton buffer (CSK: 300 mM sucrose, 100 mM NaCl, 10 mM PIPES, 3 mM MgCL 2 , 1 mM EGTA and 0.5% Triton X-100) (28) . The slides were fixed by dipping in ice-cold methanol for 3 min, followed by ice-cold acetone. After air drying the slides, they were allowed to age for at least 1 week in ethanol. DNA-FISH was carried out according to a method for single copy loci detection (28) . The protocol was adapted with few modifications, using 50-mer oligonucleotides specific to the loci of interest and labeled with a desired hapten. Two probes were used for FISH for breakapart assay. A probe located at the promoter region of TMPRSS2 was labeled with biotin (Bio/GACTCCA GGAGCGCTCCCCAGAATCCCCTTCCTTAACCCA AACTCGAGCC). The other probe at exon 2 of ERG was labeled with 5 0 -6-carboxyfluorescein (56FAM) (56FAM/ GATCTTTGGAGACCCGAGGAAAGCCGTGTTGA CCAAAAGCAAGACAAATG). Detection of probes was achieved by using antibodies conjugated to quantum dots (Qdot) against the hapten label. Conditions were optimized to use a combination of two antibodies (1 : 200) obtained from Invitrogen-Molecular Probes TM (Qdot 655 Ã sheep anti-Bio primary antibody conjugate; Qdot 525 Ã Goat anti-FITC whole IgG primary antibody conjugate). Image acquisition was done with a Zeiss Axioplan 2e microscope (Carl Zeiss, Inc.). All pictures in the corresponding three channels were deconvolved and optical sections merged to produce 2D pictures using Axiovision 4.0 software (Carl Zeiss, Inc.) and Image J (Rasband, W.S., ImageJ, U. S. National Institutes of Health, Bethesda, MD, USA, http://rsb.info.nih.gov/ij/, 1997-2006.) To develop a multiplexing assay that is highly sensitive in clinical samples of high complexity, we adopted our Virochip system (25) . The key protocol, Round ABC, designed for unbiased amplification (29) , is crucial for identifying various fusion variants in this application. Through literature review, we found that most of the TMPRSS2:ERG fusion junctions are between exons 1 or 2 of the TMPRSS2 and exons 2-5 of the ERG (2, 7, 9, 13) . Such constraints perhaps are related to whether a functional ERG protein can be made from the gene fusions (9) . Therefore, we initially used a pair of primers at exon 1 of the TMPRSS2 and exon 6 of the ERG for RT-PCR. As shown in Figure 1A , PCR products were only generated when there was a gene fusion, since the two primers are located at different genes. Subsequently, the PCR products were labeled and hybridized to an exon array for mapping the exons near the fusion junction. Printed on the array are 30-mer oligonucleotide probes derived from exons 1-3 of the TMPRSS2 and exons 1-5 of the ERG (Table 1) . Each selected sequence is represented by two complementary probes (F: forward and R: reverse complement) since sometimes PCR-labeled amplicons may bind to only one strand of the probe, based on empirical observations (25) . We observed that probes with reverse complementary (R) orientation worked consistently with our RT-PCR labeling protocol. A prostate cancer cell line, VCaP, (30) with a TMPRSS2 and ERG fusion (2) was used for initial feasibility testing. The total RNA was subjected to RT-PCR with a pair of primers located at exon 6 of ERG and exon 1 of TMPRSS2 (2, 9) . The unpurified product was labeled and hybridized on the microarray ( Figure 1B ). Only spots corresponding to exon 1 of TMPRSS2 and exons 4-5 of ERG developed strong signals. This result indicates the fusion junction is at the exon 1 of TMPRSS2 and exon 4 of ERG, which is consistent with a previous report (2) . To mimic a typical clinical situation, in which small population of cancer cells are present among normal host cells in a primary tumor, we spiked decreasing amounts of total RNA extracted from VCaP cells into an excess of HeLa RNA, which does not have the fusion transcripts. The detection limit is 32 pg of VCaP RNA in the presence of 100 ng of HeLa RNA (Figure 2 ). This translates into only 1-3 cancer cells in the presence of 3000 times more normal cells. The level of sensitivity is superior to previous methods for detecting fusion transcripts (24) . To test the ability of the exon mapping array to detect and characterize TMPRSS2:ERG fusion transcripts in clinical samples, we isolated total RNA from frozen unpurified primary prostate tissues obtained during surgery. T1F GGGCGGGGAGCGCCGCCTGGAGCGCGGCAG T2F ACATTCCAGATACCTATCATTACTCGATGC T3F GGTCACCACCAGCTATTGGACCTTACTATG T1/2F TGGAGCGCGGCAGGTCATATTGAACATTCC G1F AGGGACATGAGAGAAGAGGAGCGGCGCTCA G2F AGACCCGAGGAAAGCCGTGTTGACCAAAAG G3F GCTGGTAGATGGGCTGGCTTACTGAAGGAC G4F TTATCAGTTGTGAGTGAGGACCAGTCGTTG G5F CTCTCCTGATGAATGCAGTGTGGCCAAAGG T1R CTGCCGCGCTCCAGGCGGCGCTCCCCGCCC T2R GCATCGAGTAATGATAGGTATCTGGAATGT T3R CATAGTAAGGTCCAATAGCTGGTGGTGACC T1/2R GGAATGTTCAATATGACCTGCCGCGCTCCA G1R TGAGCGCCGCTCCTCTTCTCTCATGTCCCT G2R CTTTTGGTCAACACGGCTTTCCTCGGGTCT G3R GTCCTTCAGTAAGCCAGCCCATCTACCAGC G4R CAACGACTGGTCCTCACTCACAACTGATAA G5R CCTTTGGCCACACTGCATTCATCAGGAGAG T, TMPRSS2; G, ERG. F, forward probe; R, reverse complement probe. Many of these tumors had a substantial fraction of normal stromal cells. Total RNA (5-50 ng) from prostate cancers (n = 20) and nonmalignant hyperplastic prostate tissues (n = 10) were subjected to RT-PCR labeling and array hybridization. The results showed that 7/20 cancers but 0/10 nonmalignant samples had TMPRSS2:ERG fusion genes. To confirm the presence of the gene fusions, direct sequencing was performed for the seven samples. The sequencing data validated the exon fusion findings revealed by the array assays. Similar to other reports (7, 12, 13) , some samples clearly showed two or more bands on the agarose gel when the PCR products were subjected to electrophoresis, corresponding to two or more fusion transcripts in the same specimens. The multiple fusion transcripts in a single prostate cancer sample may reflect tumor heterogeneity or alternative splicing events. In order to map multiple fusion junctions in a single assay, we redesigned the exon array to include junction probes between exons 1 and 2 of the TMPRSS2 gene and exons 1-6 of the ERG gene ( Table 2 ). The modified probe set showed very clearly that the patient sample #15 had two fusion transcripts and also revealed the relative ratios of the two fusion transcripts through their respective signal intensities ( Figure 3 ). In this case, the two fusion transcripts are between exon 4 of the ERG fused to either exon 1 (T1G4) or exon 2 (T2G4) of the TMPRSS2. The signal intensity of T2G4 junction probe is weaker than that of the T1G4 junction probe ( Figure 3A) , consistent with the intensities of the probes within the exons. These two fusion transcripts are very likely due to alternative splicing. Figure 4 summarizes the cluster analysis (31) of the seven arrays; the results are shown in Table 3 . Multiple fusion variants were found in 4/7 positive samples. Table 3 also lists the percentages of cancer cells in the tumors, the Gleason tumor grades and the detected variants of TMPRSS2:ERG fusion transcripts. In this small sample set, there is no clear association between tumor Figure 2 . Assay sensitivity. The VCaP total RNA was serially diluted in a solution containing HeLa RNA to mimic the heterogeneous cell population in primary tumors or human body fluids. The total amount of RNA for each reaction is 100 ng. The laser power (PMT 600, 100% output) was adjusted to maximize the sensitivity of detection. Therefore, the intensity of the each expected feature (T1, G4, G5) is at the saturated level. The signal disappeared when the VCaP RNA was diluted from 1:3125 (32 pg) to 1:15625 (6.4 pg). T1G1F CCTGGAGCGCGGCAGCCCCCGAGGGACATG T1G2F CCTGGAGCGCGGCAGGTTATTCCAGGATCT T1G3F CCTGGAGCGCGGCAGCCGTCAGGTTCTGAA T1G4F CCTGGAGCGCGGCAGGAAGCCTTATCAGTT T1G5F CCTGGAGCGCGGCAGATGCCACCCCCAAAC T1G6F CCTGGAGCGCGGCAGATCCTACGCTATGGA T2G1F ATGGCTTTGAACTCACCCCCGAGGGACATG T2G2F ATGGCTTTGAACTCAGTTATTCCAGGATCT T2G3F ATGGCTTTGAACTCACCGTCAGGTTCTGAA T2G4F ATGGCTTTGAACTCAGAAGCCTTATCAGTT T2G5F ATGGCTTTGAACTCAATGCCACCCCCAAAC T2G6F ATGGCTTTGAACTCAATCCTACGCTATGGA T1G1R CATGTCCCTCGGGGGCTGCCGCGCTCCAGG T1G2R AGATCCTGGAATAACCTGCCGCGCTCCAGG T1G3R TTCAGAACCTGACGGCTGCCGCGCTCCAGG T1G4R AACTGATAAGGCTTCCTGCCGCGCTCCAGG T1G5R GTTTGGGGGTGGCATCTGCCGCGCTCCAGG T1G6R TCCATAGCGTAGGATCTGCCGCGCTCCAGG T2G1R CATGTCCCTCGGGGGTGAGTTCAAAGCCAT T2G2R AGATCCTGGAATAACTGAGTTCAAAGCCAT T2G3R TTCAGAACCTGACGGTGAGTTCAAAGCCAT T2G4R AACTGATAAGGCTTCTGAGTTCAAAGCCAT T2G5R GTTTGGGGGTGGCATTGAGTTCAAAGCCAT T2G6R TCCATAGCGTAGGATTGAGTTCAAAGCCAT T, TMPRSS2; G, ERG. F, forward probe; R, reverse complement probe. grade and the presence of fusion transcripts. A relatively larger study also showed that the presence of fusion transcript was associated with tumor stage but not tumor grade (12) . Consistent with the VCaP titration study (Figure 2 ), the clinical assay can detect the fusion transcript when only 1% tumor cells is present in the prostate tissue (sample 10). We used FISH analysis to independently confirm our array approach. Our FISH procedure (28) employed 50-mer probes that were labeled with small haptens for target hybridization in conjunction with individual hapten-specific, quantum dot conjugated antibodies for signal detection. The resolution of this method is $50 kb. We designed two probes for the FISH assays, one at the promoter region of TMPRSS2 (green in Figure 5 ) and the other at exon 2 of ERG (red in Figure 5 ). We observed heterogeneity of the FISH patterns in some primary prostate cancer samples ( Figure 5 ). It is more difficult to find interstitial deletions between TMPRSS2 and ERG in tumor samples containing low percentages of cancer cells by FISH. Therefore, we used samples that contained 480% cancer cells without detectable fusion genes to confirm the results of the arrays. The FISH experiments revealed no genomic deletions at this location for all the selected samples (#5, #7, #9, #19 and #20) that were similarly nondeleted by array hybridization. We have established a simple assay that can concurrently profile variants of TMPRSS2:ERG fusion transcripts by combining a single RT-PCR with an exon array. The modified 'Round ABC' protocol, which was TMPRRS2e1R TMPRRS2e1/2R TMPRRS2e2R TMPRRS2e3R ERGe1R ERGe2R ERGe3R ERGe4R ERGe5R ERGe6R T1G1R T1G2R T1G3R T1G4R T1G5R T1G6R T2G1R T2G2R T2G3R T2G4R T2G5R T2G6R #1 #15 #17 #4 #13 #10 #18 Figure 4 . Cluster analysis of seven prostate cancer samples having fusion transcripts. The signal intensity of each feature is divided by the intensity of a nonspecific control (herring sperm DNA) to normalize the data for cluster analysis. The result is shown in Table 3 . The samples having similar fusion transcript variants were clustered together by the program. 1 30 7 T1-G4; T2-G4 2 2 0 5 3 5 0 5 4 20 6 T1-G4 5 8 0 9 6 1 6 7 9 0 8 8 2 0 4 9 8 0 8 10 1 6 T1-G2 11 2 6 12 70 7 13 20 9 T1-G4 14 1 6 15 70 8 T1-G4; T2-G4 16 20 8 17 80 8 T1-G4; T2-G4 18 50 7 T1-G2; T1-G3; T1-G4 19 80 7 20 80 7 Figure 5 . Heterogeneity of FISH patterns of interstitial deletion between TMPRSS2 and ERG in a primary prostate tumor. An unpaired green dot (TMPRSS2 probe, indicated by arrows) suggests an interstitial deletion. Nonrandom variation of FISH patterns is shown by the fact that most of the green and red signals (two different but nearby probes) are paired in each panel. This variation is expected on a heterogeneous aneuploid cancer cell population, which often makes it difficult to distinguish meaningful events from random background aberrations. originally designed for genomic amplification (29) and has been widely adopted for chromatin immunoprecipitation (ChIP) and whole-genome DNA microarrays (ChIP-chip) (32, 33) and Virochip (25, 34) experiments, is a simple and relatively unbiased amplification procedure to semiquantitatively measure the fusion variants in a complex sample. Previously, the same simple procedure was used to obtain 83% (25 kb/30 kb) of the SARS coronavirus genome with total nucleic acids isolated from a viral culture (25) . The inclusion within the array of probes derived from individual exons and potential fusion junctions simplifies the breakpoint mapping and increases the confidence of data interpretation (Figures 3 and 4) . In contrast to some reports that used multiple primers targeted to every potential fusion junction in hematological malignancies (35) (36) (37) (38) , we used a single set of primers for target amplification (Figure 1 ). The fusion junctions were subsequently decoded by array. This design significantly reduces the problems associated with primer dimers in the multiplex PCR reaction, and creates more room for future assays to include additional fusion genes. Furthermore, most searches for fusion genes have focused on blood cancers, because the cells can be purified before analysis. The application of the previous methodologies is less useful for highly complex solid tumors that are inevitably admixed with normal cells. For example, a previously reported MLLFusionChip could not be applied to samples with cancer cells of 55-10% in 1 mg of total RNA (39) . The current assay should facilitate a thorough compilation of the gene fusion variants in primary prostate specimens, which may be useful for stratifying the aggressiveness of prostate cancer (13) . In this regard, fusion variants of EWS with another member of the ETS family, FLI1, have been shown to be an independent predictor of disease progression in Ewing's sarcoma (40, 41) . It will be of interest to compare in transfected cells the biological activities of the different TMPRSS2:ERG variants from patients with contrasting clinical outcomes (41) . While some studies have suggested that the presence of TMPRSS2:ERG fusions is associated with more aggressive disease or higher Gleason tumor grade, other investigators did not reach the same conclusion (12, 14, 17, 20, 23, 42) . We also did not find such an association in a small series of samples. However, all of these results are defective due to small sample size. The technology described here should make possible a larger scale investigation to find whether there is a correlation between the aggressiveness of the disease and the presence of specific fusion genes. It is crucial to have true cancer-specific biomarkers for early cancer detection as well as for minimal residual disease monitoring, which has been extensively demonstrated in hematologic maligancies (43) . Such biomarkers could help to avoid under-or over-treatment. Thus, there is past interest (24, 44) in applying TMPRSS2:ERG fusion assays for such application since PSA and many other markers in development are not truly prostate cancer-specific (45) . A recent study reported a TMPRSS2:ERG assay with a sensitivity of detecting 1600 VCaP cells (24) . However, this level of sensitivity might not be sufficient for broad clinical application, especially with small biopsy specimens or urine samples. We were able to achieve an assay sensitivity of 532 pg of total RNA derived from VCaP cells, an equivalent to 1-3 cells (Figure 2 ). Because our assay is simple and amenable to automation, it is readily adaptable to clinical studies. While it has been challenging to adapt microarray-based technology to the clinic, some tests (e.g. AmpliChip CYP450 and MammaPrint) have been approved by FDA (46) . The same strategy can be applied to detect other less prevalent fusion transcripts (TMPRSS2:ETV1 and TMPRSS2:ETV4) in prostate cancer (2, 5, 6) . In addition, the exon array approach can also be applied to other fusion genes, such as BCR-ABL in CML and clonal Ig/TCR rearrangements in lymphocytic malignancies. While this methodology development was motivated by the clinical need, it is generally applicable to other research requirements that are analogous to the situation for detecting fusion genes in the single cell level when a large excess of normal cells are present. For example, a developmental biologist may use a similar approach to screen mutants that have a desirable gene fusion when direct gene targeting is not feasible. There are some shortcomings of using RNA transcripts as prostate cancer biomarkers, despite our ability to achieve very sensitive detection of TMPRSS2:ERG fusion variants. First, RNA is unstable and difficult to process in routine clinical assays. Second, commonly used drugs that inhibit androgen growth pathways, including GnRH agonists and testosterone antagonists, may diminish the production of the TMPRSS2:ERG mRNA fusion transcript, thereby producing false-negative results in patients on hormonal therapy with evolving androgenindependent tumors. Indeed, it has been reported that TMPRSS2:ERG mRNA fusion transcripts are not expressed in androgen-independent tumors in spite of the presence of interstitial deletions in between TMPRSS2 and ERG at chromosome 21q22 (10) . While FISH is useful for identifying genomic rearrangements, it has relatively lower resolution and is difficult to use in highly heterogeneous samples with small percentages of tumor cells. We have recently developed a technology, designated Primer Approximation Multiplex PCR (PAMP) for identifying breakpoints in genomic DNA without the need to purify cancer cells from normal tissues (26) . We are optimizing this assay for detecting the breakpoints between TMPRSS2 and ERG loci for primary prostate tumors to overcome any potential problems associated with RNA based biomarkers. In addition, the DNAbased assay will provide information about whether multiple fusion transcripts in a sample are derived from alternative splicing or tumor heterogeneity. The best approach may ultimately be to combine DNA and RNA based assays in a common format. Institutes of Health (CA119335 to UCSD NanoTumor Center of Excellence for Cancer Nanotechnology, CA133634 to Y. T. Liu, NS034934 to

Delivery of steric block morpholino oligomers by (R-X-R)(4) peptides: structure–activity studies

Redirecting the splicing machinery through the hybridization of high affinity, RNase H- incompetent oligonucleotide analogs such as phosphoramidate morpholino oligonucleotides (PMO) might lead to important clinical applications. Chemical conjugation of PMO to arginine-rich cell penetrating peptides (CPP) such as (R-Ahx-R)(4) (with Ahx standing for 6-aminohexanoic acid) leads to sequence-specific splicing correction in the absence of endosomolytic agents in cell culture at variance with most conventional CPPs. Importantly, (R-Ahx-R)(4)–PMO conjugates are effective in mouse models of various viral infections and Duchenne muscular dystrophy. Unfortunately, active doses in some applications might be close to cytotoxic ones thus presenting challenge for systemic administration of the conjugates in those clinical settings. Structure–activity relationship studies have thus been undertaken to unravel CPP structural features important for the efficient nuclear delivery of the conjugated PMO and limiting steps in their internalization pathway. Affinity for heparin (taken as a model heparan sulfate), hydrophobicity, cellular uptake, intracellular distribution and splicing correction have been monitored. Spacing between the charges, hydrophobicity of the linker between the Arg-groups and Arg-stereochemistry influence splicing correction efficiency. A significant correlation between splicing correction efficiency, affinity for heparin and ability to destabilize model synthetic vesicles has been observed but no correlation with cellular uptake has been found. Efforts will have to focus on endosomal escape since it appears to remain the limiting factor for the delivery of these splice-redirecting ON analogs.

Protein transduction domains as penetratin or Tat 48-60 and synthetic cell penetrating peptides (CPP) as oligoarginine have generated a large interest for their seemingly unique mechanism of membrane translocation and for their capacity to transport various biomolecules across biological membranes (1) . Both assumptions have had to be re-visited since cellular uptake does involve endocytosis (2) and transport of biomolecules does not occur as efficiently as anticipated at least at low concentrations. In a series of experiments carried out independently by several groups, CPPs mentioned above turned out inefficient in transporting uncharged splice correcting oligonucleotide (ON) analogs as peptide nucleic acids (PNA) or phosphorodiamidate morpholino oligomers (PMO) for a large part because CPP-conjugated material remained entrapped in endocytic vesicles (3) . Accordingly, peptides or drugs (such as chloroquine) leading to endosome destabilization did significantly increase splicing correction. We have recently described a new (R-Ahx-R) 4 -CPP (in which Arg residues are interspersed with nonnatural 6-aminohexanoic acid amino-acid spacers) which leads to efficient splicing correction at low concentration in the absence of endosomolytic agents. (R-Ahx-R) 4 is less cytotoxic and much more active to deliver splice correcting PMO and PNA in vitro than the parent oligoarginine (R n ) peptide and than the prototypic Tat 48-60 peptide (4) . Importantly, (R-Ahx-R) 4 -PMO conjugates also lead to efficient exon skipping in murine and dog Duchenne muscular dystrophy (DMD) models (5) and inhibit the replication of viruses in several murine models (6) (7) (8) (9) . We had no clear explanation for the improved efficiency of (R-Ahx-R) 4 -PMO and (R-Ahx-R) 4 -PNA conjugates as compared to Tat or (Arg) n steric block ON constructs. Increased cellular uptake could not be the answer since, on the contrary, (R-Ahx-R) 4 -PMO conjugates were taken up less efficiently than Tat-PMO and (Arg) n -PMO conjugates in our model system (3) . Differences could originate from a different mechanism of cellular uptake with (R-Ahx-R) 4 -PMO conjugates taking profit of a more favorable route than the other CPP conjugates. Again available data did not support this hypothesis since all three conjugates were taken up by an energy-dependent pathway involving binding to cell surface proteoglycans. Along the same lines, we recently established that blocking energy-dependent processes through incubation of cells at low temperature or through ATP depletion decreased splicing correction by (R-Ahx-R) 4 -PMO conjugates and by Tat-PMO or (Arg) n -PMO ones to the same extent (10) . Of possible relevance, (R-Ahx-R) 4 -PMO conjugates bind less strongly to heparin (taken as a model heparan sulfate) than Tat-or (Arg) n -PMO conjugates (3) . This could provide an explanation if one assumes that heparan sulfate-bound material has to be released during endocytosis in order to escape to the cell cytoplasm. While sufficient affinity is essential for cell binding and cellular uptake, a too high affinity could become detrimental at later steps, a hypothesis which we aim to investigate here. On the other hand, the inclusion of nonnatural amino acids as aminohexanoic acid in (R-Ahx-R) 4 was expected to increase the metabolic stability of these conjugates and as a consequence to increase their biological efficiency. This assumption has to be tempered since the (R-Ahx-R) repeats are linked by arg-arg peptide bonds which are amenable to proteolysis by trypsin-like enzymes. Indeed recent studies in our group indicated that (R-Ahx-R) 4 -PMO conjugates were rapidly degraded in cells at these arg-arg bonds (11) . We have therefore investigated a (r-Ahx-R) 4 -PMO conjugate (with r standing for D-Arg) in terms of cell uptake and splicing correction activity. The present structure-activity relationship (SAR) studies were initiated for the following additional reasons. Fluorescence microscopy evaluation of the intracellular distribution of FAM-labeled (R-Ahx-R) 4 -PMO conjugates indicated that the majority of the material was entrapped in endocytic vesicles even at concentrations leading to efficient splicing correction. It implies that the biological activity of these conjugates is due to the small (and not detectable by fluorescence microscopy) portion of material escaping from endocytic vesicles. Finally, (R-Ahx-R) 4 -PMO conjugates have shown signs of toxicity when injected to mice at >20 mg/kg dose despite their absence of cytotoxicity in cell culture experiments (12) . This presents a dosing challenge for in vivo systemic applications. Altogether, it is clear that CPP-steric block ON conjugates need to be active at lower doses for systemic administration and clinical applications. It is hoped that a better understanding of the structural determinants required for cell binding, cellular uptake and endosome escape will be helpful for the rational design of more potent and/or less cytotoxic CPPs. The manuscript essentially aims at comparing series of (R-Ahx-R) 4 -PMO conjugates analogs differing in Arg spacer length, in hydrophobicity of the spacer and in stereochemistry of Arg. Criteria for the comparative evaluation of these conjugates include cellular uptake, splicing correction efficiency, affinity for heparin, hydrophobicity and synthetic membrane-destabilizing potential. The antisense PMO (CCT CTT ACC TCA GTT ACA) was synthesized as described (13, 14) . CPPs were synthesized using Fmoc chemistry and purified to >95% as determined by high-pressure liquid chromatograph and MALDI-TOF mass spectrometry analysis. Conjugation, purification and analysis of CPP-PMO conjugates were described previously (3, 15) . HeLa pLuc705 cells were cultured as exponentially growing subconfluent monolayers in DMEM medium (Gibco) supplemented with 10% fetal bovine serum (FBS), 1 mM Na pyruvate and nonessential amino acids. To analyze (R-X-R) 4 -PMO conjugates cell internalization, exponentially growing HeLa pLuc705 cells (1.75 Â 10 5 cells seeded and grown overnight in 24-well plates) were incubated in OptiMEM with FAM-labeled (R-X-R) 4 -PMO. Cells were then washed twice with PBS, detached by incubating for 5 min at 378C with 0.5 mg/ml trypsin per 0.35 mM, EDTA.4Na and washed by centrifugation (5 min, 900g) in ice-cold PBS containing 5% FBS. The resulting cell pellet was resuspended in ice-cold PBS containing 0.5% FBS and 0.05 mg/ml propidium iodide (PI) (Molecular Probes, Eugene, OR, USA). Fluorescence analysis was performed with a BD FacsCanto flow cytometer (BD Biosciences, San Jose, CA, USA). Cells stained with PI were excluded from further analysis. A minimum of 20 000 events per sample was analyzed. (R-X-R) 4 Three micrograms of each (R-X-R) 4 -PMO conjugate were injected in triplicate on a 1 ml HiTrap Sepharose/heparin column (Amersham Biosciences, Freiburg, Germany) fitted on a Beckman-Gold HPLC chromatograph (Beckman Coulter, Fullerton, CA, USA). The conjugates were eluted in 30 min at a flow rate of 1 ml/min of 2.5 mM phosphate buffer (pH 7) by a linear gradient of NaCl from 70 to 970 mM. Elution of the conjugates was followed by UV absorption at 260 nm. Results are presented as eluting NaCl concentrations and expressed as the mean and standard deviation of triplicate measurements. Total 0.1 mg of each (R-X-R) 4 -PMO conjugate were injected in triplicate on a C18 Waters Symmetry Shield 4.6 Â 250 mm column and fitted on a Beckman-Gold HPLC chromatograph. The conjugates were eluted at a flow rate of 1 ml/min of H 2 O/0.1% TFA by a linear gradient of acetonitrile from 5% to 95% in 30 min. Elution of the conjugates was followed by UV absorption at 260 nm. Results are presented as eluting acetonitrile concentrations and expressed as the mean and standard deviation of triplicate measurements. Exponentially growing HeLa pLuc705 cells (1.75 Â 10 5 cells seeded and grown overnight in 24-well plates) were co-incubated with the (R-Ahx-R) 4 -PMO conjugates and with 20 mg/ml saponin for 30 min. The conjugates were removed and the cells were washed twice with PBS and incubation continued for 24 h in complete medium (DMEM plus 10% FBS). Cells were washed twice with ice-cold PBS, lysed with Reporter Lysis Buffer and processed as described above. To analyze (R-X-R) 4 -PMO conjugates intracellular distribution, exponentially growing HeLa pLuc 705 cells (3.5 Â 10 4 cells seeded and grown overnight in 2 ml culture dishes) were washed with OptiMEM and incubated with 2 mM FAM-labeled (R-X-R) 4 -PMO in the absence or in the presence of 20 mg/ml saponin for 30 min in OptiMEM medium. Cells were then washed with PBS prior a co-incubation step with 10 mg/ml Transferrin-Alexa 546 (red fluorescence) and Hoechst 33342 dye (blue fluorescence) for 10 min in order to stain endosomes and nuclei, respectively. The distribution of fluorescence in live unfixed cells was analyzed on Zeiss Axiovert 200M fluorescence microscope (Carl Zeiss, Obercochen, Germany). Large unilamellar vesicles (LUV) were prepared as described previously (16) . In short, lipids dissolved in benzene/methanol (95:5) were freeze-dried overnight and the resulting dry lipid powder was hydrated in a buffer containing the ANTS fluorescent dye 8-aminonaphthalene-1,3,6-trisulfonic acid, disodium salt (Invitrogen, Carlsbad, CA, USA) together with a DPX quencher p-xylene-bispyridinium bromide (Invitrogen, Carlsbad, CA, USA) at a final lipid concentration of 10 mM. The suspension was vigorously agitated with a Vortex, freeze-thawed 10 times and then extruded 10 times through two stacked 100 nm polycarbonate filters (Nucleopore, Whatman). Free dye and quencher were then removed by gel filtration on a PD-10 desalting column (Amersham Biosciences, Piscataway, NJ, USA). To mimic the lipid composition of late endosomes we used the following lipid mixture: dioleoylphosphatidylcholine (DOPC)/dioleoyl-phosphatidylethanolamine (DOPE)/phosphatidylinositol from soybean (PI)/bis(monooleoylglycero) phosphate (LBPA) (5:2:1:2) (23). All lipids were purchased from Avanti Polar Lipids Inc., Alabaster, AL. Leakage of ANTS/DPX from the vesicles was measured as an increase in fluorescence intensity of ANTS upon addition of the CPP-PMO conjugates (5 mM final concentration) to 2 ml of vesicles (25 mM) (17) . Infinite dilution of the probe used to determine fluorescence of the completely unquenched probe was achieved by solubilizing the membranes with 0.1% (v/v) Triton X-100. Most studies on basic amino-acids-rich CPPs emphasized the importance of the guanidinium side chains of arginines and of the spacing between the charged groups. Studies by Rothbard et al. (18) in particular have shown that a 6-carbon 6-aminohexanoic acid linker seemed optimal for cellular uptake as measured by the whole cell fluorescence but no data concerning efficiency in terms of cytoplasmic or nuclear delivery of a biologically functional payload was provided. We therefore designed a series of (R-X-R) 4 -PMO conjugates with X varying from 2 to 8 carbons (compounds 1-7 in Figure 1A ). The present study revealed a dependence of charge spacing with an optimum for (R-Ahx-R) 4 (in which X = 6) in terms of nuclear delivery of the PMO payload as illustrated below. Based on this first set of data, we designed a series of C6 linked-Arg peptides differing in terms of hydrophobicity (compounds 8-11 in Figure 2A ). Since metabolic stability has often been proposed as a factor governing CPP efficiency, the D-Arg modified (R-Ahx-R) 4 , (r-Ahx-R) 4 (compound 13 in Figure 3A ), has been included. Finally, we evaluated the splice correcting ability of (R-X-R) n -PMO conjugates with n < 3 as a possible strategy to reduce cytotoxicity. It is now well admitted that basic CPPs interact with heparan sulfate-rich cell surface glycosaminoglycans before being internalized by endocytosis (19) . A sufficient affinity for these negatively charged proteoglycans is required for cell binding and for subsequent cellular uptake. On the other hand, too much affinity for heparan sulfate might be detrimental for the release of CPP-ON conjugates from endocytic vesicles as hypothesized in our previous publications (3, 10) . (R-X-R) 4 -PMO conjugates with X spacers of increasing lengths (from 2 to 8 atoms) ( Figure 1A ) have thus been compared in terms of affinity for a model heparan sulfate on a Hi-trap Heparin column ( Figure 1B ). Increasing spacer length leads to decreased affinity as monitored by the NaCl concentration required for elution in keeping with published data (20) . Conjugates in this series were then compared for their ability to promote splicing correction in dose-response experiments ( Figure 1C and data not shown). Increasing the length of the spacer led to increased luciferase expression with an optimum for the C5-linked material. Increasing the affinity for heparan sulfates thus appears being detrimental for splicing correction efficiency. Along the same lines, (Arg) 9 -PMO has an even higher affinity for heparan sulfate than (R-G-R) 4 -PMO and is less active in splicing correction [(4) and data not shown]. Compound 7 was thus expected to be more active in splicing correction than compound 5 which was not observed ( Figure 1C ). However, increasing the hydrocarbon spacer length also increases hydrophobicity which could itself be promoting unfavorable membrane interactions (see below). Increased hydrophobicity has indeed been verified by C18-column chromatography for compound 7 ( Figure 2B) . Figure 1C . (D) Heparin affinity chromatography. Samples were treated and data were processed as described in the legend of Figure 1B . ÃÃÃ and ÃÃ Indicate statistically significant differences; NS indicates that the difference is not statistically significant. PI uptake has been monitored in parallel as an index of cell membrane integrity. No significant PI uptake was seen at doses up to 2.5 mM for any one of these compounds except for compound 7, which leads to a concentrationdependent membrane destabilization ( Figure 4B ). This might also contribute to its lower splicing correction potential. Since the hydrophobicity of the linker appeared to influence splicing correction efficiency, we have compared a series of PMO conjugates (compounds 8-11 in Figure 2A ) with the same spacing (a 6-carbon atom spacer as in (R-Ahx-R) 4 -PMO) but with varying sidechain hydrophobicities. Figure 1C . 5, (R-Ahx-R) 4 -PMO; 13, (r-Ahx-R) 4 -PMO. (C) Hydrophobicity chromatography. Samples were treated and data were processed as described in the legend of Figure 2B . (D) Heparin affinity chromatography. Samples were treated and data were processed as described in the legend of Figure 1B . ÃÃÃ and ÃÃ Indicate statistically significant differences; NS indicates that the difference is not statistically significant. Some compounds in this series (11 in Figure 2A ) have hydrophobicities comparable to the parent (R-Ahx-R) 4 -PMO (compound 5) taken as a reference while others (compounds 8-10) have a significantly higher hydrophobicity, as monitored by C18-column chromatography ( Figure 2B ). These conjugates were analyzed for splicing correction efficiency at various concentrations ( Figure 2C and data not shown). Splicing correction efficiency is lower for the more hydrophobic conjugates (compounds 8-10) and remains the most active for compound 5. As expected, compounds 8-11 had similar affinities for heparin ( Figure 2D ). Therefore, differences in splicing correction in this series were largely influenced by hydrophobicity. We cannot explain why compound 11 has a lower splicing correction activity than compound 5 as their hydrophobicity and heparin affinity are similar. Increased metabolic stability should improve biological efficiency and could in part explain the higher efficacy of (R-Ahx-R) 4 -PMO as compared to (Arg) n -PMO and Tat 48-60 -PMO, as discussed previously (3) . However, the (R-Ahx-R) 4 portion of (R-Ahx-R) 4 -PMO was found to be degraded in intact cells (11) . We therefore synthesized (r-Ahx-R) 4 -PMO (compound 13 in Figure 3A ) in which one of the two L-Arg residues in each R-Ahx-R repeat was replaced by a D-Arg (r) residue. Unexpectedly, (r-Ahx-R) 4 -PMO was significantly less efficient in splicing correction than (R-Ahx-R) 4 -PMO ( Figure 3B ). Both L-and D-Arg-containing peptides have similar hydrophobicities ( Figure 3C ). Interestingly, (r-Ahx-R) 4 -PMO has a signficantly higher affinity for heparan sulfate than the parent (R-Ahx-R) 4 -PMO ( Figure 3D ), thus pointing again to the role played by this parameter in splicing correction efficiency. As already mentioned, (R-Ahx-R) 4 -PMO conjugates become cytotoxic in murine models at high doses. The cytotoxicity of nonviral delivery vectors for nucleic acids is generally associated to their resulting cationic charge. We therefore investigated whether reducing the number n of repeats in (R-X-R) n -PMO conjugates could be possible. A significant loss of splicing correction efficiency was found with shorter versions of these PMO conjugates as shown for (R-AbuL-R) n -PMO conjugates ( Figure 5 ). Most (R-X-R) 4 -PMO conjugates were synthesized as fluorescent FAM conjugates to allow assessment of celular uptake by FACS analysis and by fluorescence microscopy. As seen in Figure 4 , there is no correlation between cellular uptake and splicing correction activity. Increasing the spacing between arginine residues (compounds 1-7) leads to decreased cellular uptake in parallel to heparin affinity but on the contrary leads to increased splicing correction. Remarkably, (R-Ahx-R) 4 -PMO which was the most active conjugate in this series in terms of splicing correction turned out the less efficient in terms of cellular uptake. In addition, changing the hydrophobicity of the spacer (compounds 8-11) or modifying the stereochemistry of Arg (compounds 5 and 13) had no significant impact on cellular uptake. We next examined whether differences in splicing correction activity could be explained by differences in endosomal escape. All (R-X-R) 4 -PMO conjugates have therefore been synthesized as FAM-labeled derivatives and their intracellular distribution has been analyzed by fluorescence microscopy on live cells to avoid artefactual redistribution upon cell fixation. As shown in Figure 6A for the parent (R-Ahx-R) 4 -PMO-FAM conjugate, most of the material was distributed as punctate cytoplasmic material and none was detected in the nuclei. Splicing correction is probably due to the small amount of material which has escaped from the endocytic vesicles and remains undetectable by fluorescence microscopy analysis. Not surprisingly, a similar situation has been observed for other (R-X-R) 4 -PMO-FAM conjugates from our SAR studies and no concluding data have been provided by fluorescence microscopy comparative analysis (data not shown). Likewise, previous work from several groups including our own one had documented an increase in splicing correction upon treatment with endosomolytic agents such as chloroquine or calcium ions (3) . However, splicing correction never reached levels achieved with 2-OMe ON analogs transfected as lipoplexes and accordingly redistribution of the endosome-entrapped material could not be documented (21) . We now capitalize on a saponin treatment protocol which allows to gently permeabilize the plasma membrane. It was shown to open transient holes in the plasma membrane and to allow the passage of macromolecules while not damaging membranes from intracellular organelles (22) . As shown in Figure 6B , the (R-Ahx-R) 4 -PMO-FAM conjugate was widely distributed within the cell with a clear accumulation in nuclei in saponin-permeabilized cells. These low molecular mass conjugates are indeed expected to diffuse freely and rapidly from the cytoplasm to the nuclei through the nuclear pores. In a different protocol, cells were loaded with Alexa-labeled Transferrin (a known marker of endosomes) and then treated with saponin. At variance with the wide distribution of the CPP-PMO-FAM conjugates, Transferrin-associated red fluorescence remained punctate in keeping with the reported minimal effects of saponin on intracellular architecture ( Figure 6C ). Along the same lines, we have verified that saponin did not lead to a significant release of (R-Ahx-R) 4 -PMO conjugate preloaded in endocytic vesicles (data not shown). We next compared luciferase expression in doseresponse experiments in saponin-treated and untreated cells. As shown in Figure 7 for the (R-Ahx-R) 4 -PMO conjugate, splicing correction was more efficient in saponin-treated than in untreated cells. Significant luciferase expression could already be detected upon 30 min incubation in saponin-treated cells and increased to a much higher level than in untreated cells. In addition, splicing correction in saponin-treated cells reached similar levels for conjugates found much less active in nonpermeabilized cells than (R-Ahx-R) 4 -PMO ( Figure 7B) . These data were expected if saponin permeabilization of the plasma membrane allows to bypass endocytosis and as a consequence endosome segregation. Differences in splicing correction efficiency between (R-X-R) 4 -PMO analogs were also expected to be largely abolished if caused by differences in trafficking efficiency. Our data have confirmed that cellular uptake is not the limiting factor in the efficiency of splicing correction by (R-X-R) 4 -PMO conjugates and therefore intracellular trafficking and endosomal escape likely to be major limiting factors. To evaluate the ability of CPP-PMO conjugates to escape from endosomes we employed a liposome leakage assay. Late endosomes are characterized by a rather unusual lipid composition enriched in LBPA (23) and have a pH 5.5 lumen (24) . We therefore prepared liposomes from a lipid mixture mimicking the lipid composition of late endosomes DOPC/DOPE/PI/LBPA (5:2:1:2) and monitored the effect of low pH on the CPP-PMO induced leakage of a fluorescent dye entrapped in the lipid vesicles. We compared conjugates 1, 5, 9 and (Arg) 8 -PMO. All conjugates induced a fairly modest leakage that was strongly promoted at pH 5.5. In correlation with the data on splicing activity, (R-Ahx-R) 4 -PMO conjugate was by far the most active in this group, followed by (R-G-R) 4 endosomal escape is a major contributing factor for the efficient nuclear delivery of these CPP-PMO conjugates. New arginine-rich CPPs have recently been proposed for the nuclear delivery of neutral ON analogs as PMO (3) or PNA (10) . They represent a significant improvement over first generation CPPs as Tat, Pen, oligoarginine or oligolysine since they allow a sequence-specific splicing correction at lower concentrations (EC 50 ranging between 0.5 and 2.0 mM) which do not lead to membrane permeabilization and, importantly, do not require endomosomolytic drugs or treatments. Importantly as well, (R-Ahx-R) 4 -PMO conjugates lead to a sustained expression of dystrophin in skeletal muscles when injected intraperitoneally (5 mg/kg) in DMD mice (25, 26) . These encouraging data should however be tempered since these may be still high doses too close to the toxic doses found in other murine models (6) . The present SAR study has therefore been initiated in order to delineate step(s) limiting the splice correcting activity of (R-Ahx-R) 4 -PMO as well as important molecular features of the (R-Ahx-R) 4 -CPP moiety. Spacing of the guanidinium charged groups in argininerich CPPs has been extensively studied by Wender et al. (20) and found to be a key determinant of their cellular uptake. A series of (R-X-R) 4 -PMO conjugates with a X linker extending from two to eight atoms have been compared in terms of cellular uptake, splicing correction efficiency and affinity for heparin ( Figure 1B and R. Abes and H. Moulton). As expected, heparin (chosen as a model heparan sulfate) affinity decreased significantly with an increase in X linker length and with a decrease in cationic charges density ( Figure 1B ). In keeping with this observation, cellular uptake as monitored by FACS analysis decreased in parallel ( Figure 4) . However, the ranking of these (R-X-R) 4 -PMO conjugates in terms of splicing correction efficiency had a bell-shaped profile with (R-Ahx-R) 4 -PMO being significantly more efficient than (R-G-R) 4 -PMO ( Figure 1C and data not shown). These observations do strongly suggest that another step than cellular uptake is responsible for differences in splicing correction efficiency among these (R-X-R) 4 -PMO conjugates. Whether too much affinity for heparan sulfates could be detrimental for dissociation of the heparan-bound material in endocytic vesicles and for endosomal escape is a possibilty but it is unfortunately not amenable to direct demonstration. It is worth pointing out that similar conclusions could be drawn from our previous comparison of (Arg) 9 -PMO, Tat-PMO and (R-Ahx-R) 4 -PMO conjugates. (R-Ahx-R) 4 -PMO was found more active in the splicing correction assay despite binding less efficiently to heparin and being taken up less well than the parent (Arg) 9 -PMO and than Tat-PMO (3). The (R-Acy-R) 4 -PMO conjugate (compound 7 with a C8 linker) did not follow the ranking observed for other conjugates in this series since it had lower heparin-binding affinity but corrected splicing less efficiently than (R-Ahx-R) 4 -PMO. This might be explained by the higher hydrophobicity of its longer spacer as evidenced by an increased retention on a C18-affinity column ( Figure 2B ). In keeping with this hypothesis, increasing the hydrophobicity of the X linker above a threshold value ( Figure 2B ) while maintaining charge spacing in a series of (R-X-R) 4 -PMO analogs ( Figure 2B ) had little impact on cellular uptake (Figure 4 ) but decreased significantly splicing correction efficiency ( Figure 2C ). Being too hydrophobic might conceivably lead to entrapment into membranes and as a consequence might be detrimental to endosomal release. The parent and most active (R-Ahx-R) 4 -PMO conjugate was rather resistant to proteolytic degradation in serum but was still cleaved by cellular proteases (11) . It was thus anticipated that the (r-Ahx-R) 4 -PMO in which some L-Arg residues have been replaced by their D-analog would become more protease-resistant and as a consequence more active in the splicing correction assay. Unexpectedly, (r-Ahx-R) 4 -PMO was significantly less active than (R-Ahx-R) 4 -PMO thus indicating that metabolic stability is not a limiting factor at least in these in vitro experiments. Whether (r-Ahx-R) 4 -PMO might be of interest for in vivo applications will have to be evaluated using transgenic murine models for splicing correction. Whether the lower biological activity of (r-Ahx-R) 4 -PMO could be due to its increased affinity for heparin is a possibility. Alternatively, earlier work from our group (11) has shown that the peptide part of (R-Ahx-R) 4 -PMO was rapidly degraded in cells thus releasing free PMOs. It is fully possible that the more stable CPP entity may decrease the rate of endosomal release of PMO. Along the same lines, linking a splice correcting PNA and a CPP (R 6 Pen in this case) by a stable linker gave rise to a lower efficiency than in the case of a reducible disulfide linker (27) . Altogether these SAR studies have pointed to the influence of heparin affinity and hydrophobicity on the splice correcting activity of CPP-PMO conjugates. Remarkably, relatively small changes in these parameters had a rather significant impact on biological activity. Quite clearly, cellular uptake could not be an explanation since, on the contrary, we have observed in some instances an inverse correlation between cellular uptake and biological activity. In order to determine whether differences in biological activity could be explained by differences in intracellular trafficking, we deliberately permeabilized cells by a brief treatment with saponin, using conditions known to have little impact on the internal cellular architecture (22) . Saponin treatment clearly lead to a complete re-localization of FAM-labeled conjugates ( Figure 6 and data not shown) from a dotted cytoplasmic to an homogeneous nuclear distribution in keeping with a direct membrane translocation in the presence of saponin and an endocytic process in its absence. As expected from these data, splicing correction efficiency was largely increased in saponin-treated cells for all (R-X-R) 4 -PMO (Figure 7 and data not shown) thus indicating that retention within cytoplasmic vesicles remains a major road-block even for the most active of our conjugates. We thus tentatively conclude at this stage that (R-X-R) 4 -PMO accumulate in cytoplasmic vesicles after binding to cell surface glycosaminoglycans and endocytosis. Differences in splicing correction might thus be primarily due to the efficiency with which these conjugates escape from endocytic vesicles and reach the cytoplasm. Since a direct evaluation of endosome leakage in intact cells could not be easily monitored, we have capitalized on synthetic lipid vesicles with a lipid composition mimicking that of the late endosomal membrane. We observed a modest but significant fluorescent dye release in the presence of (R-X-R) 4 -PMO. Importantly, the release of the probe was strongly promoted at pH 5.5 that is the characteristic pH for late endosomes. Although preliminary, these studies do indicate that more active (R-X-R) 4 -PMO conjugates in this series destabilize more efficiently these lipid vesicles than less active ones. In conclusion, our studies support the now wellaccepted scheme of cellular internalization involving initial binding to cell surface glycosaminoglycans, endocytosis and entrapment within cytoplasmic vesicles. Most of the material unfortunately remains segregated in endocytic vesicles even for the most active of our conjugates and efforts should now be geared at improving endosomal escape. Model systems described here might turn rather useful in future SAR studies if pH-dependent liposome destabilization can be correlated with splicing correction efficiency. The most active conjugates will also have to be monitored for their biodisponibility, metabolic stability and biological activity in animal models.

Immune Mechanisms Responsible for Vaccination against and Clearance of Mucosal and Lymphatic Norovirus Infection

Two cardinal manifestations of viral immunity are efficient clearance of acute infection and the capacity to vaccinate against secondary viral exposure. For noroviruses, the contributions of T cells to viral clearance and vaccination have not been elucidated. We report here that both CD4 and CD8 T cells are required for efficient clearance of primary murine norovirus (MNV) infection from the intestine and intestinal lymph nodes. Further, long-lasting protective immunity was generated by oral live virus vaccination. Systemic vaccination with the MNV capsid protein also effectively protected against mucosal challenge, while vaccination with the capsid protein of the distantly related human Lordsdale virus provided partial protection. Fully effective vaccination required a broad immune response including CD4 T cells, CD8 T cells, and B cells, but the importance of specific immune cell types varied between the intestine and intestinal lymph nodes. Perforin, but not interferon gamma, was required for clearance of MNV infection by adoptively transferred T lymphocytes from vaccinated hosts. These studies prove the feasibility of both mucosal and systemic vaccination against mucosal norovirus infection, demonstrate tissue specificity of norovirus immune cells, and indicate that efficient vaccination strategies should induce potent CD4 and CD8 T cell responses.

More than 90% of epidemic nonbacterial gastroenteritis worldwide can be attributed to human noroviruses (HuNV) [1] [2] [3] . Infection is transmitted fecal-orally, and symptomatic infection is characterized by nausea, vomiting and/or diarrhea lasting 24-48 hours within 24 hours of exposure [4] . Despite the significant costs and morbidity of HuNV infections, no vaccine is currently available. The elderly and individuals in long-term care facilities may be more susceptible to either norovirus infection or norovirusinduced disease [5] and would be an important target population for a norovirus vaccine. The reasons for increased incidence and/ or susceptibility to HuNV disease are unknown. This is due in part to our incomplete understanding of norovirus immunity. The potential to vaccinate against these and related viruses has been demonstrated in gnotobiotic piglets, cats and rabbits [6] [7] [8] , but the immune mechanisms responsible have not been identified. The challenges for vaccine efficacy may be very different between different caliciviruses. For example, variation in MNV strains is significantly less than between HuNV strains [9] . Human volunteer studies demonstrate short-term, but not long-term, protection against homologous, but not heterologous, viral challenge [10] [11] [12] . Since HuNV belong to 3 genogroups (GI, GII and GIV) with many strains in each genogroup [4] , this lack of cross-protection is a challenge for vaccine development. Frequent exposure to noroviruses within short time periods stimulates sustained immunity and resistance to norovirus induced illness [13, 14] . Serum antibody levels in adults reflect susceptibility to infection and do not always correlate with protection [13, 14] . In children, however, serum antibody levels correlate with protection, likely reflecting short-term immunity and recent exposure [15] [16] [17] . A nonfunctional fucosyl transferase gene (FUT2) accounts for a significant proportion, though not all, of resistance to Norwalk virus infection, suggesting that other factors, yet undiscovered, may contribute to norovirus resistance [18, 19] . In the absence of a cell culture system for HuNV, virus like particles (VLPs) that assemble when the viral capsid protein is expressed have been important for evaluating norovirus immune responses [20] [21] [22] [23] . Studies using Norwalk Virus (GI), Snow Mountain Virus (GII) and HuNoV-HS66 (GII) VLPs to evaluate immunity after infection with live virus or immunization with VLPs orally show production of T cell effector cytokines such as IL-2 and interferon c (IFN-c) and proliferation of norovirus specific T cells after in vitro restimulation with VLPs [24] [25] [26] . These studies show that T cell responses develop, but do not define their role in either clearance of primary infection or resistance to re-challenge. Together, they suggest the potential for vaccination, but leave open important questions about the effectiveness and longevity of vaccine immune responses, mechanisms of vaccination, the viral protein targets for protective responses, and the potential for cross-protection between distantly related noroviruses. The identification of the first murine norovirus, MNV, and its propagation in cultured cells provides a facile animal model for studies of norovirus immunity and pathogenesis [27, 28] . MNV, an enteric virus that infects tissues of the gastrointestinal tract, is spread by the fecal-oral route ( [27] and unpublished studies). The MNV genome encodes four open reading frames. ORF1 encodes a polyprotein that is cleaved into individual non-structural proteins similar to the polyprotein of HuNV [29] . ORF2 encodes the major capsid protein VP1 and ORF3 encodes a minor capsid protein. The existence of a protein product for ORF4 has not been confirmed. In the MNV virion structure, the capsid, like that of human noroviruses, consists of 90 dimers of VP1 [30] . There are differences between the MNV virion and previously reported VLP structures. The MNV protruding domain is lifted off the shell domain by approximately 16 Angstroms and rotated approximately 40 degrees in a clockwise fashion, forming interactions at the P1 base in an infectious virion that have not been observed previously. The existence of these novel aspects of the structure are consistent with the hypothesis that MNV may undergo a capsid maturation process [30] . Studies of MNV pathogenesis reveal an important role for interferon (IFN) and STAT-1 mediated innate immunity in resistance to infection and MNV induced lethality [27, 31] . The importance of adaptive immunity in control of MNV infection is indicated by the observation that RAG1-/-mice develop persistent MNV infection while wild type (WT) mice can clear infection with some strains of MNV [9, 27, 31] . While MNV is an efficient enteric virus that infects many mice in research mouse colonies around the world, diarrhea has not been reported after MNV infection. Thus, MNV provides an infection only model for HuNV infection. Viral titers in tissues of infected mice have not been reported to exceed 10 6 PFU/ml, and this highest level of viral titer is obtained after infection of highly susceptible STAT1-/-mice [31] . In RAG1-/-mice and WT mice, viral titers of 10 2 to 10 4 PFU/ml are routinely observed [9, 31] . The availability of a plaque assay for MNV allows the analysis of MNV infection despite these low titers. Some MNV strains persist at a low level in WT mice, while others are cleared from intestine, spleen, liver, mesenteric lymph nodes (MLN) and feces within 7 days of infection [9, 27, 31] . Additionally, in wild type C57BL6/J mice MNV replicates maximally in the distal ileum [9] , in comparison to wild type 129S6/SvEvTac mice where replication occurs in the proximal intestine [31] . The significance of these differences is not known. Studies of norovirus infection in human volunteers have not specifically investigated whether the infection spreads beyond the intestine to the local lymph nodes, however, it is possible that systemic invasion occurs in humans with chronic conditions or immunosuppressed hosts [32] [33] [34] [35] [36] . Additionally, viremia has been reported in infections of gnotobiotic pigs and calves [26, 37, 38] . Thus, the ability of MNV to spread to tissues other than the intestine after oral infection may not be unique, but the relationship of this aspect of MNV pathogenesis to human infection is not clear. The availability of strains that can be cleared from WT mice, such as MNV1.CW3, provides an opportunity to define the mechanisms responsible for two cardinal aspects of viral immunity: the capacity to effectively clear acute infection and the immune mechanisms responsible for effective vaccination. B cells and MNV specific antibody are important in the clearance of primary MNV infection [39] , but the role of T cells in clearance and the potential and mechanisms of vaccination against mucosal norovirus challenge are unknown. We show here that vaccination with either live MNV or Venezuelan Equine Encephalitis replicon particles (VRPs) expressing the MNV capsid protein VP1 protect the intestine against re-challenge for at least six months. Live virus was more effective than VRP-mediated vaccination. There was partial cross protection against MNV infection after vaccination with a HuNV capsid protein. We found that both the clearance of primary infection and vaccination require the concerted efforts of CD4 T cells, CD8 T cells, B cells, and that T cells required the effector molecule perforin for maximal impact on MNV infection. The effects of specific immune cell types were tissue specific, differing between ileum and mesenteric lymph nodes. These are the first studies to demonstrate immune mechanisms responsible for norovirus clearance and vaccination. We first determined whether we could detect short-term immunity to homologous MNV challenge and whether proteins encoded by specific MNV ORFs could elicit effective immunity. VRPs expressing ORF1, ORF2 and ORF3 of MNV1.CW3 and ORF2 of the HuNV Lordsdale (genogroup GII.4) and Chiba (genogroup GI.4) were produced for vaccination experiments. Western blots of VRP-infected cell lysates revealed proteins of appropriate sizes [29, 40] and additionally showed that hyperimmune polyclonal rabbit antisera to MNV [28] cross-reacted at low levels with VLPs from Chiba virus and Lordsdale virus ( Figure S1A ). WT mice were vaccinated and boosted three weeks later. Two weeks after boosting, mice were challenged with MNV1.CW3 and organs titered for MNV three days later ( Figure 1A ). In these WT mice, maximal MNV replication in the intestinal tract occurs in the distal ileum [9] and viral titers could not be detected in duodenum/ jejunum (data not shown). After oral inoculation with MNV1.CW3, WT mice exhibit detectable viral titers in the distal ileum and the MLN three to five days post-infection [9, 31] . Prior infection with Human noroviruses are the most common cause of epidemic nonbacterial gastroenteritis in the world. Despite their importance as human pathogens, little is known about how the immune system controls and clears norovirus infection, and the potential and mechanisms of vaccination remain unclear. Here, we used norovirus infection of mice to show that vaccination can provide long-lasting immunity against mucosal norovirus challenge and to identify the types of immune cells that are important in vaccination against norovirus infection. Similarly, we identified the types of immune T cells that are important for clearance of acute infection. Efficient vaccination required all three major arms of adaptive immunity: CD4 T cells, CD8 T cell, and B cells. Importantly, protective vaccination against mucosal challenge was observed after either mucosal or systemic norovirus antigen exposure. The pore-forming molecule perforin was important for T cell-mediated control of norovirus infection. Our study has important implications for understanding adaptive immunity to norovirus infection, and may provide insight into the directions to take in developing a human norovirus vaccine. either MNV1.CW1 (p = 0.0002) or MNV1.CW3 (p = 0.0009) significantly decreased MNV1.CW3 replication in the distal ileum compared to control mice infected with reovirus ( Figure 1B ). Similar decreases were observed in the MLN after vaccination with MNV1.CW1 (p = 0.0001) or MNV1.CW3 (p = 0.0003) ( Figure 1C ) compared to the reovirus controls. Similar results were observed in the spleen (data not shown). There was no statistically significant difference between vaccination with MNV1.CW1 or MNV1.CW3. This demonstrates that a protective secondary immune response develops after clearance of primary MNV infection. ORF2 VRPs protected against MNV1.CW3 in both distal ileum (p = 0.005) and MLN (p = 0.02) compared to control VRPs expressing hemagglutinin (HA) from a mouse adapted influenza A virus [41] (HA VRP control group). Controls for VRP vaccination also included PBS. HA VRP controls were not significantly different from PBS controls across all experiments and statistical comparisons for VRP vaccination are therefore shown to HA VRP controls. ORF1 VRPs alone in the distal ileum, or in both the distal ileum and MLN when combined with ORF3 VRPs, had a small but statistically significant effect on MNV1.CW3 levels ( Figure 1B and 1C). ORF3 VRPs alone did not confer significant protection ( Figure 1B and 1C). Together these data show that vaccination with either live virus or ORF2 VRPs can confer shortterm protection against MNV challenge. We next assessed vaccination with heterologous ORF2 proteins. Mice were vaccinated and boosted with VRPs expressing ORF2 from Chiba Virus or Lordsdale virus and challenged with MNV1.CW3. Vaccination with Lordsdale virus capsid led to statistically significant protection against MNV infection in the distal ileum, (p = 0.0007, Figure 1B ) but not the MLN ( Figure 1C ). No significant reduction in MNV titers was seen after immunization with Chiba virus capsid ( Figure 1B and 1C ). Protection after Lordsdale ORF2 VRP vaccination did not correlate with generation of cross-reactive serum IgG in these mice, measured by ELISA, despite the potential for such cross-reactivity revealed by western blot ( Figure S1B ). Fecal extracts from immunized mice yielded no measurable homotypic or heterotypic IgG or IgA (data not shown). Taken together, these data show that there is measurable functional immunologic cross protection between Lordsdale virus and MNV in the distal ileum. The lack of a correlation between serum or fecal antibody responses and protection suggested that protection may be T cell mediated. Since older adults may be more susceptible than younger adults to norovirus infection or disease [5] , we determined whether increased age altered vaccine efficacy. Prior work has shown that mice older than 1 year of age have diminished vaccine responses to SARS virus antigens [42] . We therefore compared vaccine efficacy in adult (8 week old) and aged (14 month old) mice. Adult and aged mice were vaccinated and challenged as before. In contrast to studies using SARS virus antigens [42] , aged mice responded as well as adult mice to MNV ORF2 vaccination in both the distal ileum and MLN ( Figure 1D and 1E). Despite this protective effect, sera from vaccinated aged mice had significantly lower anti-MNV ORF2 IgG compared to adult mice ( Figure S1C ). These data indicated that protection against MNV infection occurred in the absence of robust serologic responses, again raising the possibility that T cells play a fundamentally important role in vaccination against MNV. We next determined whether protection conferred by MNV1.CW3 or MNV ORF2 VRPs was long lived. WT mice were primed and boosted as shown in Figure 2A with MNV1.CW3 or MNV ORF2 VRPs. Mice were then challenged with MNV1.CW3 two, four, 14, or 24 weeks later and MNV titers measured three days post-challenge. Two weeks post-boost, we observed complete protection against ileal MNV1.CW3 infection after vaccination with either MNV1.CW3 (p = 0.0001) or ORF2 VRPs (p,0.0001) compared to reovirus or HA VRP controls ( Figure 2B ). At two weeks, while vaccination with either MNV1.CW3 or ORF2 VRPs limited MNV1.CW3 replication in MLN, live virus vaccination was more effective (p,0.0001) ( Figure 2C ). Live virus vaccination conferred full protection against MNV1.CW3 replication in both the distal ileum and the MLN at four, 14 and 24 weeks after vaccine boost. Vaccination with ORF2 VRPs was also protective, albeit less effective than vaccination with MNV1.CW3 ( Figure 2B and 2C). Thus both live virus and subunit vaccine induce long-term protection against MNV infection, with live virus vaccination providing more complete protection. We next determined the mechanism(s) responsible for effective vaccination. We vaccinated mice lacking both major histocompatibility complex (MHC) Class I and b2 microglobulin (b2M) [43] (CD8 T cells deficiency [44] ), MHC Class II (CD4 T cells deficiency [45] ), or B cell deficient mice [46] ( Figure 3A) . These experiments were conducted concurrently with the experiments in Figure 2 above, as such the data from WT mice are repeated in the figure for comparison. Live MNV vaccination induced significant protection against MNV challenge in both the distal ileum and the MLN of B cell-/-, MHC Class II-/-and MHC Class I6b2M-/-mice (p,0.05 in all cases, Figure 3B and 3C). However, there was considerable variation in the efficacy of vaccination in distal ileum and MLN between different immunodeficient strains. In B cell-/-mice, after vaccination with live virus, only 2 out of 15 mice had any titer (and those two mice had less than 100 PFU of MNV) and in MHC Class I6b2M-/-mice, similar vaccination led to undetectable viral titers in the distal ileum ( Figure 3B ) but detectable titers in the MLN ( Figure 3C ). In MHC Class II-/-mice, there were detectable titers in both tissues ( Figure 3B and 3C). Results for ORF2 vaccination showed that protection required the activity of all major aspects of the adaptive immune response ( Figure 3B and 3C). Moreover, there was no protection elicited by ORF2 vaccination in either intestine or MLN tissue after vaccination of MHC Class I6b2M-/-mice with ORF2 VRPs (Figure 3B and 3C) indicating that protection by VRPs critically depends on CD8 T cells. These data demonstrated that complete protection in all tissues after vaccination with live virus required the concerted actions of B cells, MHC Class II, MHC Class I and b2M. Further, the results were consistent with tissue specific roles for B cells, CD4 T cells and CD8 T cells in the development of complete protection against MNV infection. We next determined whether the same cell types that were required for vaccination were also required for efficient clearance of acute infection. We focused on the role of T cells in clearance since the role of B cells in clearance has already been demonstrated [39] . To determine the role of T cells in clearance of acute MNV infection we inoculated WT, MHC Class II-/-, and MHC Class I6b2M-/-mice orally with MNV1.CW3 and measured viral titers in the distal ileum and MLN three, five, seven and 21 days post-infection ( Figure 4B-4E) . There was no significant difference in viral titer between MHC Class I6b2M-/-mice and WT mice at three and five days postinfection, indicating that MHC Class I and b2M were not required in MNV infection at early time points ( Figure 4B II-/-compared to WT mice at seven days post-infection (p = 0.04, Figure 4E ). By eight days post-infection, MLN infection was cleared. Together these data indicated that MHC Class II, and by inference CD4 T cells, were necessary for control of acute MNV infection but are not required for eventual clearance of MNV infection. To exclude the possibility that the phenotypes we observed in MHC Class I6b2M-/-and MHC Class II-/-mice were due to abnormal immune ontogeny in knockout mice, we determined the requirement for CD4 and CD8 T cells in the clearance of primary MNV infection in WT mice depleted of CD4 and CD8 T cells. Depletion of CD4 and CD8 T cells was at least 90% effective as assessed by flow cytometry of isolated splenocytes ( Figure 5A ) and this depletion protocol is effective at depleting T cells in secondary lymphoid organs and the intestine [47, 48] . In comparison to control antibody, depletion of CD4 T cells, led to a significant increase in MNV titers in the distal ileum (p = 0.0053, Figure 5B ), but not the MLN ( Figure 5C ). In contrast, depletion of CD8 T cells led to an increase in MNV titers in both the distal ileum (p = 0.004, Figure 5B ), and the MLN (p = 0.0025, Figure 5C ). Together, these data from primary challenges of non-immune mice lacking antigen presenting molecules or depleted of specific T cell subsets demonstrated that CD4 T cells are important for efficient MNV clearance in the distal ileum especially at days three We next determined whether CD4 and CD8 T cells from vaccinated mice can, alone or in combination, clear MNV infection from mucosal sites. We have previously shown that MNV infected RAG1-/-mice have high levels of viral RNA present in multiple tissues up to 90 days post-infection [27] . We therefore determined MNV viral titers in RAG1-/-mice. By 42 days postinfection, all RAG1-/-mice had consistent, high levels of MNV in both duodenum/jejunum and distal ileum ( Figure 6A and 6B) , as well as several other tissues (data not shown). These data confirmed that mice lacking adaptive immunity fail to clear MNV infection [27] . The availability of persistently infected RAG1-/-mice allowed us to determine the role of CD4 and CD8 T cells in clearance of MNV infection using adoptive transfer of splenocytes from MNV immune WT mice into persistently infected RAG1-/-mice. Transfer of immune, but not non-immune, splenocytes signifi- To define which cells were required for MNV clearance, CD4 or CD8 T cells were depleted from splenocytes transferred into RAG1-/-recipients. Anti-T cell antibodies effectively depleted the appropriate T cell populations, as measured six days post-transfer by flow cytometry (Figure 7A ). Depletion of either CD4 or CD8 T cells individually led to a significant increase in MNV titers in duodenum/jejunum compared to control depletion ( Figure 7B , CD4 depletion p = 0.0042; CD8 depletion p = 0.0002). Depletion of both CD4 and CD8 T cells from transferred immune splenocytes caused a significant additional increase in MNV titers when compared to either CD4 depletion alone (p = 0.02) or CD8 depletion alone (p = 0.03). In the distal ileum, depletion of either CD4 T cells (p = 0.0003) or CD8 T cells (p,0.0001) led to a significant increase in MNV titers ( Figure 7C ). These data demonstrated that both immune CD4 and CD8 T cells were necessary for clearance of persistent MNV infection from the intestine. Two major effector mechanisms for the antiviral effects of T cells are the production of IFNc and perforin mediated cytolysis [49] . We therefore adoptively transferred immune splenocytes from IFNc-/-or perforin-/-mice into persistently infected RAG1-/-mice and determined their capacity to clear intestinal MNV infection. Immune splenocytes from IFNc-/-mice were as effective as those from WT mice ( Figure 7B and 7C) . However, immune splenocytes from perforin-/-mice were less effective at clearing MNV infection from the duodenum/jejunum (p = 0.0003, Figure 7C ) or distal ileum (p = 0.0075, Figure 7B ) than cells from either WT or IFNc-/-mice, but more effective compared to transfer of non-immune cells in duodenum/jejunum (p = 0.0086) or distal ileum (p = 0.0001). Thus, while perforin was critical for efficient clearance of MNV infection from the intestine, it was not the only relevant effector mechanism. In this paper we define the mechanisms of immunity to norovirus infection as measured by both vaccination against, and clearance of, mucosal infection. We found that it is possible to generate highly effective, and remarkably long lasting, immunity to norovirus infection by oral exposure to live virus. Further, systemic exposure to the viral capsid protein expressed in a vaccine vector resulted in effective immunity, albeit not as effective as that observed after live virus vaccination. Importantly, this shows that the MNV VP1 protein contains relevant B cell, CD4 T cell and CD8 T cell epitopes. Vaccination was effective in aged mice. Additionally, vaccination in adult mice required the concerted action of CD4 T cells, CD8 T cell, and B cells to be completely protective in the tissues surveyed. Interestingly, the activities of different components of the adaptive immune system in clearance and vaccination were tissue specific, with different cells playing roles in the intestine itself compared to the draining lymph nodes. Perforin was an important effector molecule. These data have important implications for understanding adaptive immunity to an animal norovirus, representative of a genus that causes significant disease in humans. HuNV infection and disease is rapid, with symptoms developing within 24-48 hours of infection and lasting for a few days. Thus, we selected three days after challenge as a readout for infection in our studies, since relevant vaccine-generated immune responses would have to act very early after challenge. Lack of any of the three components of the adaptive response: B cells, CD4 T cells, or CD8 T cells significantly diminished vaccine effects generated by either live virus or VP1 capsid protein immunization, and delayed viral clearance during primary infection. This indicates that VP1 has antibody epitopes as well as MHC H-2b restricted CD4 and CD8 T cell epitopes. These data suggest that it may be necessary to engage the concerted actions of an intact immune response including both MHC Class I and MHC Class II restricted T cells and antibody responses to efficiently vaccinate against HuNV infection. The protection against MNV infection in aged mice in the absence of robust generation of anti-MNV antibodies raised the possibility that an important component of the vaccine response is T cell dependent, a hypothesis borne out in adoptive transfer studies. Importantly, the antiviral effector perforin is important in the clearance of MNV from the intestine, suggesting that the cytotoxic T cell response is a key effector mechanism. It is possible that other cell types such as NK cells might also use perforin as a mechanism to control MNV infection. Our data do not rule out a role for IFNc in clearance of MNV infection since NK cells in recipient RAG1-/-mice can make IFNc, but do suggest that T cell derived IFNc plays at most a minor role in effector T cell function in the ileum. This argues that classical CTL assays may be a good surrogate for the development of effective vaccine-generated immune responses to HuNV. Live virus vaccination was more effective than VRP based vaccination. The lower level of protection that we observed with ORF2 VRPs in contrast to MNV1.CW3 may be due to many factors, and this study does not provide mechanistic insights into this difference. In comparison to VLPs, VRPs may have advantages in systemic vaccination including targeting dendritic cells and intrinsic adjuvant activities [50] . These properties of VRPs may be responsible for the effectiveness of systemic single protein subunit vaccination against mucosal viral challenge in this case. However, it may be that because VRPs undergo a single round of replication at the site of inoculation they cannot generate the same breadth of immunity that is generated by live replicating virus. While VRP vaccination clearly induces some relevant effector and memory cell responses, vaccination with capsid alone may not sufficient to generate the complete antigenic repertoire required for effective immunity. Interestingly, we found some protection with the non-structural ORF1 polyprotein, suggesting that protective epitopes exist outside of the capsid protein. As the ORF1 polyprotein is expressed early after infection, it may be that these epitopes would be valuable targets for generating an efficient immune response. Of note, vaccination with VP1 via the subcutaneous route provided significant protection despite the fact that the vaccination occurred systemically, while protection was read out at a mucosal site. This indicates that an active systemic immune response can provide protection against norovirus infection, and a mucosal vaccine may not be necessary to vaccinate against norovirus infection. Importantly, systemic vaccination was dependent on T cells, indicating that the relevant cells can traffic to the intestine after peripheral VRP-based vaccination. These studies leave several important questions unanswered. Firstly, we used a homologous virus challenge. In nature, it is likely that hosts are repeatedly challenged with antigenically distinct noroviruses. However, the mouse norovirus strains identified so far fall into a single genogroup, GV, which likely represents a single serotype [9] . In this way murine noroviruses identified to date may present less of a challenge for the immune system than HuNV, which are distributed across 3 genogroups and appear to evolve under antibody selection [51] . In addition, we selected a strain of MNV that is cleared by WT mice. Other strains persist for prolonged periods of up to 35 days [9] . It remains to be determined whether vaccination will be effective against persistent MNV strains. It is interesting that human noroviruses can persist beyond the time frame of usual clinical symptoms [52] [53] [54] [55] . Longterm persistence might contribute to explaining the sporadic epidemics of infection in the absence of an animal reservoir. Antigenic and pathogenetic complexity will likely be a major issue for the development of norovirus vaccines. The lack of comparable variation in MNV strains limits the utility of the MNV model for assessing immunity to antigenically distinct strains. Perhaps this limitation will be overcome as additional strains of MNV are identified, sequenced, and studied. However, the fact that we observed partial cross protection between MNV and one HuNV, and the demonstration that vaccination with many different VLPs can enhance generation of cross reactive antibodies [56] provide some encouragement. There are two ways in which murine norovirus infection may not represent the same biology as HuNV infection. The first is the lack of diarrhea in mice infected with the strains of MNV used here. It is possible that the adaptive responses that clear MNV from the intestine demonstrated here are irrelevant to the responses that may prevent human disease. In this regard, it is important to note that studies of adult mice with rotaviruses (also an infection only model), have been important to our consider-ations of rotavirus vaccines [57] . Importantly, human studies may not reveal the mechanisms of effective immunity and are based on surrogate assays of immunity, since invasive sampling of tissues may be technically difficult. Studies in piglets may be revealing since piglets develop diarrhea when infected with the HuNV strain, HuNoV-HS66 [37] . However, it is more difficult to study immune mechanisms in this system. Thus, we are left with several imperfect systems for considering what one should seek in a HuNV vaccine. Our studies in mice argue for a vaccine that induces all aspects of the adaptive immune response, and that assays for cytotoxic lymphocyte responses to HuNV infection may be an important surrogate assay for protection. The second aspect of murine norovirus infection that is of unknown relevance to human infection is the impressive capacity of MNV to infect lymph nodes draining the intestine (this paper and [31, 58, 59] ). This may be related to the tropism of MNV for dendritic cells and macrophages [28, 59] and likely reflects spread of MNV directly from the intestine, but may also reflect seeding of the MLN from systemic sites. Considering the distal ileum alone, B cells and MHC Class I and b2M were not required for live virus vaccination, and there was significant, but incomplete, protection in MHC Class II-/-mice ( Figure 3B and 3C). Consistent with this, studies of primary clearance showed that any single arm of the adaptive response was dispensable for ultimate control of primary infection in the intestine. However, vaccination-mediated control of infection in the MLN, and clearance of primary infection from the MLN [39] , required B cells. This differential requirement for components of the immune response in different organs raises an important question about norovirus pathogenesis and lymphoid infection: are the cells infected in intestine and MLN the same? Differences in viral tropism in the two tissues might explain the differential requirement for B cells between ileum and MLN, indicating the importance of future studies on the role of immunity in norovirus cell and organ tropism. Viruses, Viral Stocks, VRPs, Plaque Assays MNV strains MNV1.CW3 or MNV1.CW1 were used in all virus infections [9, 28, 31] . Two mutations (that result in changes in the encoded amino acids) distinguish the genomes of MNV1.CW3 and MNV1.CW1 [28] . To generate a concentrated virus stock, RAW 264.7 cells (ATCC, Manassas, VA) were infected in VP-SFM media (Gibco, Carlsbad, CA) for 2 days at a multiplicity of infection (MOI) of 0.05. Supernatants were clarified by low-speed centrifugation for 20 min at 3,000 rpm. Virus was concentrated by centrifugation at 4uC for 3 h at 27,000 rpm (90,000 g) in a SW32 rotor. Viral pellets were resuspended in PBS and titered on RAW 264.7 cells as previously described [28] . Type I Lang reovirus was kindly provided by Dr. Terrence S. Dermody (Vanderbilt University, Nashville, TN). Plaque assays were performed as previously described [28] with the following modifications. Tissues were harvested into sterile, screw-top 2-ml tubes containing 500 ml of 1-mm zirconia/silica beads (BioSpec Products, Bartlesville, OK) and stored at 280uC. To obtain viral titers in these tissues 1 ml of complete DMEM was added to each sample on ice and homogenized using a MagNA Lyser (Roche Applied Science, Hague Road, IN) prior to plaque assay. The limit of detection was 20 plaque forming units (PFU)/ml. All VRPs were produced as previously described [60] . Briefly, ORFs 1, 2 and 3 from MNV1.CW3 and ORF2 from Lordsdale virus and Chiba virus were each cloned into VRP expression vectors. Following infection of BHK cells with VRPs for 24 h, culture supernatants were harvested and cells lysed. Proteins were separated by SDS-PAGE and analyzed by western blot with polyclonal rabbit anti-MNV serum [28] . VRP titers and efficient expression of recombinant protein were determined by immunofluorescence assay using mouse antisera generated from inoculation with respective antigens. Cell lysates from MNV ORF2, Chiba virus and Lordsdale virus VRP-infected cell cultures were further purified to obtain VLPs [56] . RAW 264.7 cells were maintained as previously described [28] . Monoclonal antibodies (MAbs) specific to CD4 (YTS191.1 [61] ), CD8 (H35 [62] ) and SFR3-DR5 (ATCC HB-151 [63] ) were produced from hybridoma cell lines in INTEGRA Celline CL1000 flasks (Integra Biosciences, Ijamsville, MD) using CD Hybridoma media (Gibco, Carlsbad, CA) as previously described [64] . All mice (or cage sentinel mice for mice deficient in antibody production) were tested by ELISA for the presence of MNV antibody prior to experiments [27] . All mice used in these studies were seronegative at the initiation of experiments. Mice used in vaccination studies were immunized with 3610 7 PFU of MNV1.CW1 [28] , MNV1.CW3 [31] , or control Type I Lang reovirus per orally (p.o.) in 25 ml of DMEM containing 10% fetal bovine serum (Hyclone, Logan, UT) (cDMEM). VRP immunizations were with 2.5610 6 infectious units (IU) of each VRP expressing MNV1.CW3 ORF1, ORF2, or ORF3 individually or in groups of 2-3 VRPs; Chiba virus ORF2 or Lordsdale virus ORF2 in 10 ml or 50 ml volume by footpad inoculation (into the subcutaneous space) [65] on day 0 and boosted on day 21. HA VRP and PBS immunizations in 10 ml or 50 ml volume by footpad inoculation [65] on day 0 and boosted on day 21 served as controls for all VRP immunizations. Mice were challenged with 3610 7 PFU of MNV1.CW3 at specified times after boost and tissues harvested three days post-challenge. Controls for VRP vaccination included PBS or VRPs expressing hemagglutinin (HA) from a mouse adapted influenza A virus [41] . PBS served as a control for HA VRP in these experiments in the event that HA VRP had a significant effect on MNV replication. HA VRP controls were not significantly different from PBS controls in all experiments and both are presented in all figures for completeness. RAG1-/-and all splenocyte donor mice were infected with 3610 6 PFU of MNV1.CW3 p.o. in 25 ml of cDMEM. All other mice were infected with 3610 7 PFU MNV1.CW3 p.o. In RAG1-/-mice two segments of the small intestine were harvested: a one inch section of the small intestine immediately distal to the pylorus of the stomach, (designated the duodenum/jejunum), and a one inch section of the small intestine immediately proximal to the cecum (designated the distal ileum). In all other mice the distal ileum and three mesenteric lymph nodes (MLN) were harvested. With the exception of RAG1-/-mice (inoculated at 4-6 weeks of age) and aged WT mice (inoculated at 14 months of age), all other mice were inoculated at 6-10 weeks of age. Spleens were harvested from mice and single cell suspensions were generated [65] . Cells were counted and diluted in RPMI-1640 media (Sigma, Saint Louis, MO) supplemented with 10% fetal calf serum (HyClone, Logan, UT), 100 U penicillin/ml, 100 mg/ml streptomycin, 10 mM HEPES (N-2-hydroxyethylpiperazine-N9-2-ethanesulfonic acid), 1mM sodium pyruvate, 50 mM 2-mercaptoethanol and 2 mM L-glutamine (cRPMI). In all adoptive transfer experiments, 1610 7 cells were injected into persistently infected RAG1-/-mice by intraperitoneal (i.p.) injection in 0.5ml cRPMI. For depletions in WT mice, 500 mg of lymphocyte-depleting antibody or an isotype-matched control antibody [SFR3-DR5, IgG2b] was administered i.p. one day prior and one day after infection. For depletions in adoptive transfer experiments, depleting antibodies were administered to RAG1-/-recipients as described above with one dose one day prior to splenocyte transfer and a second dose on the day of trnsfer. The efficacy of lymphocyte depletion in both sets of depletion experiments was monitored by flow cytometric analysis of splenocytes at the end of the experiment. ELISA to detect binding of polyclonal anti-serum or fecal extract-derived antibody to purified MNV virions or MNV VLPs was performed as previously described [27, 56] . All data were analyzed using GraphPad Prism software (GraphPad Software, San Diego, CA). Viral titer data were analyzed with the nonparametric Mann-Whitney test. All differences not specifically stated to be significant were insignificant (p.0.05).

Self-Interest versus Group-Interest in Antiviral Control

Antiviral agents have been hailed to hold considerable promise for the treatment and prevention of emerging viral diseases like H5N1 avian influenza and SARS. However, antiviral drugs are not completely harmless, and the conditions under which individuals are willing to participate in a large-scale antiviral drug treatment program are as yet unknown. We provide population dynamical and game theoretical analyses of large-scale prophylactic antiviral treatment programs. Throughout we compare the antiviral control strategy that is optimal from the public health perspective with the control strategy that would evolve if individuals make their own, rational decisions. To this end we investigate the conditions under which a large-scale antiviral control program can prevent an epidemic, and we analyze at what point in an unfolding epidemic the risk of infection starts to outweigh the cost of antiviral treatment. This enables investigation of how the optimal control strategy is moulded by the efficacy of antiviral drugs, the risk of mortality by antiviral prophylaxis, and the transmissibility of the pathogen. Our analyses show that there can be a strong incentive for an individual to take less antiviral drugs than is optimal from the public health perspective. In particular, when public health asks for early and aggressive control to prevent or curb an emerging pathogen, for the individual antiviral drug treatment is attractive only when the risk of infection has become non-negligible. It is even possible that from a public health perspective a situation in which everybody takes antiviral drugs is optimal, while the process of individual choice leads to a situation where nobody is willing to take antiviral drugs.

Recent outbreaks of SARS and H5N1 influenza have underlined the threat that viruses in the animal reservoir pose to the human population. Fortunately, neither SARS nor H5N1 influenza have become endemic in humans. Nevertheless, these and other events have stressed the importance of being prepared for emerging and reemerging infectious diseases. For most infectious diseases vaccination is the preferred control measure. Indeed, vaccines generally have proven highly efficacious, providing strong and long-lasting immunity against infection, disease, and transmission. However, in case of a previously unknown infectious disease a vaccine may not be readily available. For such emerging infectious diseases the control options are limited. This is especially so for viral pathogens that cannot be treated by effective antimicrobial drugs. For these pathogens the control options are restricted to case isolation and contact tracing, promotion of changes in behavior, vaccination using vaccines with poor efficacy, and antiviral drugs. Although the efficacy of currently available antiviral drugs is also far from perfect and although antiviral drugs provide protection for a short amount of time only, an advantage of antiviral drugs over vaccines is their broad spectrum of action [1] [2] [3] [4] [5] . For a newly arising viral infectious disease it may be possible to contain an outbreak at an early stage by means of a large-scale antiviral control program if the control program is started early and has high compliance rates, if the efficacy of antiviral drugs is sufficiently high, and if the transmissibility of the pathogen is sufficiently low [6] [7] [8] . Hence, it seems logical that all efforts should be geared towards early control of an outbreak. However, whether people will cooperate with such a containment strategy is not known. Probably, the willingness to participate in a control program depends on the (perceived) risk of infection and the consequences of the subsequent disease as compared to the (perceived) risk of taking antiviral drugs. If there are adverse effects of taking antiviral drugs it may well be that people will only be inclined to start taking antiviral drugs once the risk of infection becomes non-negligible. In this paper we employ population dynamical and game theoretical analyses to investigate (i) under which conditions an antiviral control program can prevent an epidemic, and (ii) when people should consider taking antiviral drugs. With regard to the latter question we take two perspectives. First, we focus on the public health officer whose goal it is to minimize the total amount of damage caused by both infection and prophylactic antiviral treatment. In a second step we then compare the strategy that is optimal from the point of view of the population as a whole with the strategy that would evolve if individuals pursue their own interest. The dilemma that an individual faces is the following. Should you take your chances and refuse antiviral prophylaxis? The price that you may have to pay upon infection may be high (death in its most extreme consequence). The potential reward is that once you have successfully recovered from infection you also reap the benefit of long-lasting immunity. The alternative is that you take antiviral drugs and so avoid the potentially high cost of infection. The drawback of this option is that you may have to take antivirals for a prolonged period of time. This has negative side-effects in the short term [9] , and may in the long run also not be harmless. The situation is complicated by the fact that an individual's risk of infection does not only depend on whether or not the individual itself decides to take antiviral drugs but also on the decision of others. In the context of vaccination it is well known that in such a situation there can be a trend of decreasing willingness to participate in a control program, which will lead to strategies that are not optimal from the population perspective [10] [11] [12] [13] [14] . Here we ask whether similar phenomena occur in case of antiviral prophylaxis. While vaccination usually provides long-lasting and strong immunity after one or a few vaccination bouts, antiviral prophylactic therapy relies on the continuous administration of drugs. This implies that, in contrast with vaccination, individuals have more opportunities to adjust their actions to the situation in which they face themselves. Further, while the earlier studies focus on the relative perceived risk of infection as compared to vaccination, we consider infections and antiviral drugs that induce a real, albeit possibly small, risk of death. Throughout, our aim is to decipher how the optimal antiviral prophylactic control strategy is moulded by the transmissibility of the pathogen, by the risk associated with antiviral prophylaxis, and by the efficacy of antiviral drugs in reducing susceptibility, infectiousness, and mortality. We would like to stress from the onset that we strive more for conceptual clarification than for the most precise representation of a specific system. In particular, all our analyses are based on the simplifying assumptions that individuals act rationally, have perfect information and foresight while they do not engage in projecting the epidemic, and that the details of population structure play a minor role. We are aware that these simplifying assumptions cannot be neglected in the real world, and we therefore do not believe that our model is suited to make quantitative predictions for any specific emerging infectious disease. Rather, our models serve a purpose by laying out, in an idealized context, the key factors shaping the interests of individuals and public health officers. In case of influenza vaccination others have investigated models with added layers of complexity with the goal to make quantitative predictions [13] . Stochastic and deterministic SIRV-type epidemic models in which individuals are susceptible (S), infected and infectious (I 1 or I 2 ), recovered and immune (R), or (partially) protected against infection by antiviral prophylactic treatment (V) form the basis of the analyses. Figure 1 shows a schematic of the model. Details are given in Text S1. Throughout susceptible individuals enter the population by birth. The (natural) death rate of susceptible and recovered individuals is denoted by m, and the excess mortality while on antiviral treatment is given by c. Hence, life expectancy is m 21 in the absence of infection and antiviral control, and (m+c) 21 while on antiviral drugs. In the infected classes (I 1 and I 2 ) the excess death rates resulting from infection are given by n and n(12AVE I ), where AVE I is the antiviral efficacy for infectiousness and virulence. In the following, c and n will be referred to succinctly as the cost of antiviral prophylaxis and infection. From the susceptible class individuals move to the protected and infected classes at rates s and l. The parameter l is colloquially called the force of infection, and it depends on the prevalence of infection (Text S1). Individuals in class V are infected at a reduced rate l(12AVE S ), where AVE S denotes the antiviral efficacy for susceptibility. This implies that individuals in class V cannot be infected at all if AVE S = 1, while the antiviral drug provides no protection against infection if AVE S = 0. Finally, the rates of recovery and non-compliance are given by a and r, respectively. An overview of the model parameters and their default values is given in Table 1 . Details of the model analyses are provided in Text S1. When will a prophylactic antiviral control program be able to prevent an epidemic? Several studies have addressed this question using simulations of metapopulation models that include household structure and other population and pathogen details [6] [7] [8] . Here we focus on a basic model for a large well-mixed population. To evaluate whether successful invasion of the pathogen is possible we calculate the (basic) reproduction number (denoted by R 0 ), which gives the number of new infections caused by a single infected individual in a population of uninfected individuals in the early stages of an outbreak [15] . If R 0 .1 the pathogen can invade a population in which it is not yet present, while it cannot if R 0 ,1. In the case that the pathogen-induced mortality is such that it hardly affects the infectious period, the reproduction number is given by (see Text S1 for a derivation). The first factor in equation ( Equation (1) shows that the pathogen cannot invade the population if the rate of antiviral prophylaxis exceeds a certain critical rate of antiviral prophylactic therapy s c , which is given by the solution of the equation R 0 = 1. For the default parameter values (Table 1 ) it turns out that the critical rate of antiviral prophylaxis is s c = 1.96 (yr 21 ) if antiviral drugs provide complete protection against infection (AVE S = 1). This implies that approximately two-thirds of the population needs to be protected against infection by antiviral prophylaxis in order to prevent an epidemic. This fraction increases if antiviral drugs provide partial protection against infection and subsequent transmission. Is it possible to prevent a major epidemic with an antiviral response that is started quickly after an introduction of the pathogen? To answer this question we performed stochastic simulations in which an antiviral response is initiated after a certain number of individuals are infected (see Text S1 for details). Figure 2 shows three representative simulation runs of an epidemic in a large but finite population (10 6 individuals). If no control measures are put into place (top panel), epidemiological theory informs that a large epidemic will unfold with probability 1{R {1 0 &0:67, and the fraction of individuals that is infected once the epidemic has taken off is roughly given by the solution of the final size equation (log(12x) = 2R 0 x) [15] . For the default parameter values this means that 94% of the population will be infected, of which some 2% will die from the sequelae of infection. In a population of one million individuals this implies that more than 18,000 individuals will die during the course of an epidemic. The situation is completely different if a large-scale antiviral prophylactic control program is initiated once a certain number of individuals is infected. The middle and bottom panels of Figure 2 show simulations in case of a perfect and imperfect antiviral drug, respectively. In both panels we assume that antiviral control is started once 100 individuals are infected, and that no individuals are exempted from antiviral drug treatment. The middle panel shows that even though the cost of antiviral prophylaxis is much smaller than the cost of infection ( Table 1 ) the number of individuals that has died by antiviral treatment at the end of the epidemic equals the number that has died from infection (8 individuals) . If the antiviral drug is imperfect (bottom panel), the duration of the epidemic is considerably longer, and many more individuals will have died from antiviral treatment than from infection (22 versus 6) . Still, the total number of deaths is orders of magnitude smaller than in the case that no antiviral control program is effective. The simulations of Figure 2 illustrate two general phenomena. First, the number of infections and deaths is reduced dramatically by an antiviral control program that is able to successfully contain an epidemic [6] [7] [8] . Second, while the number of deaths caused by infection is proportional to the number of infected individuals (which is relatively small at an early stage of an epidemic), the number of deaths related to antiviral prophylaxis is proportional to the number of individuals that have taken antiviral drugs. The latter may be quite large, and is probably on the order of total population size. Hence, even though the individual risk of antiviral prophylaxis is small and large-scale antiviral prophylactic control appears to be the rational strategy, it may well be that the number of deaths induced by antiviral treatment exceeds the number of deaths induced by infection. Motivated by these observations we investigate in the following (i) the conditions under which a largescale antiviral control program is able to halt an epidemic, and (ii) the conditions under which rational individuals are willing to take antiviral drugs. At what point in an unfolding epidemic does the risk of infection exceed the risk of antiviral treatment? This question is relevant because it determines the incentive for individuals to take antiviral drugs. We focus on the probability that an individual is alive after a certain time (the horizon) given that it is initially susceptible or (partially) protected against infection by antiviral control. If the probability to remain alive is larger when initially susceptible than when under antiviral treatment, the best option is not to take antiviral drugs. Otherwise the reverse is true. The formal analyses are given in Text S1. Here we summarize the main findings. Now let us suppose that attempts to control an outbreak at an early stage have been unsuccessful. In this case it is still of interest to determine whether and under which conditions antiviral prophylaxis should be part of a strategy aimed at pathogen eradication or containment. If antiviral treatment provides complete protection against infection (AVE S = 1) the equilibrium prevalence of infection decreases monotonically with increasing rate of antiviral control, up to the point where the pathogen cannot persist (Text S1). Furthermore, as long as the risk of antiviral prophylaxis remains small its precise value hardly affects the prevalence of infection. This is due to the fact that mortality related to antiviral treatment is negligible in comparison with natural mortality. i.e. if the cost of antiviral prophylaxis is small compared to the cost of infection. This implies that at equilibrium excess mortality is lowest at the point where the pathogen is just driven to extinction. If, on the other hand, inequality (2) is reversed, excess mortality increases with increasing rate of antiviral prophylaxis, so that excess mortality is lowest if no antiviral drugs are taken at all. For the default parameter values the right-hand side of inequality (2) equals 0.00025 (yr 21 ), while the cost of antiviral prophylaxis is c = 0.0001 (yr 21 ). Hence, in our default scenario excess mortality decreases with increasing rate of antiviral control up to the point where the pathogen is just unable to persist (s = 1.96 (yr 21 )). Next we turn our attention to the different perspectives of the individual versus the public health officer. Our focus is on the antiviral treatment rate that minimizes excess mortality. Minimiz- ) and high (r = 6 (yr 21 )) rates of non-compliance, respectively. Other parameters are as in Table 1 ing this quantity with respect to the antiviral treatment rate yields the strategy that is optimal from the population perspective. Throughout this section and the next we assume that the pathogen is endemically present at the population dynamical equilibrium. As we have argued above, the optimal population strategy is such that the pathogen is just unable to persist if the risk of antiviral prophylaxis is small (i.e. if (2) is satisfied). Otherwise, the optimal population strategy is not to take antiviral drugs at all (Text S1) In Text S1 we also show how excess mortality of a small group of individuals with antiviral treatment rate s y is calculated in a population where the majority of individuals take antiviral drugs at a rate s x . This allows one to determine the antiviral treatment strategy that will evolve at the population level by the process of individual choice. The optimal population and individual rates of antiviral treatment at the population level will be denoted by s à pop and s à ind , respectively. Figure 5 shows the results of a systematic investigation of the relation between the model parameters and the fractions of individuals taking antiviral drugs (which are determined by the antiviral control rates s à pop or s à ind ). The top panel shows the fraction of individuals taking antiviral drugs (black lines) and the associated excess mortality (grey lines) as a function of pathogen transmissibility. If transmissibility is low (b,51 (yr 21 )) the pathogen cannot persist, and there is no need to take antiviral drugs. If transmissibility is intermediate there is a positive population optimum (s à pop w0) which ensures eradication of the pathogen, while the individual optimum is still zero (s à ind~0 ). If transmissibility is high both the population and individual control rates are positive, although eradication is only achieved by the optimal population control rate. The bottom panel of Figure 5 illustrates how the fractions of individuals taking antiviral drugs depend on the antiviral death rate. Not surprisingly, if antiviral drugs incur no cost (c = 0) then both the population and individual optimal control rates are such that the pathogen is driven to extinction. For the default parameter values this is achieved if at least two-third of the population is on antiviral drug treatment (s$1.96). If there is a cost to antiviral treatment, then the best option is to drive the pathogen to extinction if one takes the population perspective, until the risk of antiviral prophylaxis exceeds the risk of infection at the endemic equilibrium with no antiviral control (c.0.00025 (yr 21 )). Alternatively, if all individuals are allowed to flexibly adjust their own strategy, the optimal rate of antiviral prophylaxis decreases gradually with increasing antiviral death rate. In this case the optimal rate of antiviral prophylaxis is zero if c.0.00016 (yr 21 ). Notice that for intermediate cost of antiviral treatment the public health officer favours a strategy that is aimed at eradicating the disease, while the process of individual choice leads to a situation where nobody is willing to take antiviral drugs. Unfortunately, to date there are no antiviral drugs that provide complete protection against infection and disease. For instance, an analysis of two recent trials with the antiviral drug oseltamivir shows that it provides little protection against infection with influenza, and moderate protection against subsequent shedding and disease [16] . Therefore, we will in this section study the consequences of antiviral prophylactic treatment with an imperfect antiviral drug. In the analyses below we take AVE S = 0.3 and AVE I = 0.8 as default parameter values [16] . The relation between the antiviral efficacies for susceptibility and infectiousness, and the optimal rates of antiviral control is investigated in Figure 6 . The top panel shows the relation between the antiviral efficacy for susceptibility and the optimal fraction of individuals taking antiviral drugs, assuming that antiviral efficacy for infectiousness and virulence is fixed at AVE I = 0.8. The fact that the optimal fractions taking antiviral drugs decrease with increasing antiviral efficacy for susceptibility can be understood as follows. A decrease in the antiviral efficacy for susceptibility renders the antiviral drug less effective. However, since the cost of antiviral treatment is low while the antiviral efficacy for infectiousness is relatively high the rational strategy is to eradicate the pathogen if one takes the population perspective. With decreasing antiviral efficacy for susceptibility this is achieved by increasing the rate of antiviral control. Interestingly, the top panel indicates that if one takes the individual perspective excess mortality is highest if antiviral drugs provide complete protection against infection, since then the optimal control rate is lowest. The bottom panel of Figure 6 shows the relation between the optimal fractions of individuals taking antiviral drugs as a function of the antiviral efficacy for infectiousness. The antiviral efficacy for susceptibility is fixed at AVE S = 0.3. The picture in this panel is more complicated than in the top panel. In particular, eradication of the pathogen is not feasible if the antiviral efficacy for infectiousness drops below a critical value (AVE I ,0.51). If the antiviral efficacy for infectiousness is just above this critical value it still is the best strategy to drive the pathogen to extinction if one If, on the other hand, the antiviral efficacy for infectiousness is low (AVE I ,0.07) it is better not to take antiviral drugs at all (s à pop~0 ) as the benefit of taking antiviral drugs do not outweigh the cost. In this region of parameter space both optimal control rates are zero. In the intermediate parameter regime (0.07,AVE I ,0.51) it is not possible to achieve eradication, but it may nevertheless be wise to take antiviral drugs. In fact, taking the population perspective, it is always better to be (partially) protected by antiviral drugs than to be fully susceptible in this region of parameter space (i.e. s à pop ??), even though eradication is not possible. Intuitively, it may seem that one should consider antiviral treatment when faced with a highly transmissible pathogen that can kill its host. However, this line of reasoning may be inaccurate. In particular, the notion that antiviral treatment is attractive because the drugs are relatively harmless and because a potentially large number of infection-induced deaths can be prevented is not necessarily true. Our analyses show that over the course of an epidemic the death toll can be considerable if no antiviral drugs are taken, and that the number of deaths is orders of magnitudes smaller if a large-scale antiviral control program is effective ( Figure 2) . However, the analyses also show that in an epidemic that is effectively controlled by a large-scale antiviral treatment program the majority of deaths result from the use of antiviral drugs. The intuitive explanation is that although the hazard of mortality by antiviral prophylaxis is small on the short-term individual level, the total death toll may be quite high as the number of individuals that must receive antiviral drugs for an effective control effort is probably on the order of total population size. Moreover, adding to this is the fact that in the face of an imminent threat it may prove necessary to continue taking antiviral drugs for a prolonged period. In the early stages of an unfolding epidemic the probability of infection is still small. Consequently, individuals may be tempted to put off antiviral drug treatment until the prevalence of infection and hence the probability of infection has become non-negligible. Our analyses have shown that the critical force of infection that determines at what point individuals should start taking antiviral drugs depends strongly on the adverse effects of antiviral drug use ( Figure 3 ). Hence, for purposes of successful prevention or early containment it is important that the adverse effects of antiviral drugs remain small. Our analyses have revealed that, taking the population perspective, the best option is either to provide antiviral drugs up to the point where the pathogen is unable to invade and persist, or not to provide antiviral drugs at all (Figures 5-6) , depending on the cost of antiviral treatment and the effectiveness of antiviral treatment. If, on the other hand, one takes the individual perspective there is an incentive to take less antiviral drugs than is optimal from the population perspective, and complete prevention or eradication of the disease is rarely possible. Interestingly, the conflict between the individual and the public health officer appears to be most pronounced when the cost of antiviral drug treatment or the effectiveness of antiviral drugs in reducing the adverse effects of infection are intermediate (Figures 5-6 ). In fact, if the cost of antiviral prophylaxis is intermediate it is possible that public health favours an aggressive containment strategy that aims at pathogen eradication, while the process of individual choice leads to a situation in which nobody is willing to take antiviral drugs. On the other hand, if the cost of antiviral treatment is high or if the effectiveness of antiviral drugs in preventing the adverse effects of infection is very low, then the conflict may become small or disappear at all (Figures 5-6) . One aspect of our model that deserves special attention is that we have throughout assumed that the cost of both infection and antiviral treatment are in terms of an increased risk of death. This is convenient because it enables a straightforward comparison of the positive and negative consequences of infection and antiviral prophylaxis. However, although there is no question that human infections with SARS and H5N1 avian influenza bring along a risk of death, it is as yet unclear how severe the adverse effects of antiviral drug treatment may be. This is especially so for rare but possibly severe adverse effects. For instance, it is well-documented that oseltamivir (the neuraminidase inhibitor currently used to treat and prevent influenza infections) frequently leads to nausea and a number of less frequent adverse effects such as hepatitis and skin reactions [9, 17] . Recently, there have been suggestions of more serious adverse effects, including neuropsychiatric syndromes that may have contributed to a number of suicide events in Japan [18] [19] . Table 1 While we have used mortality as the currency to compare the costs and benefits of antiviral drug use, previous game theoretical studies of vaccination focused on the relative perceived risk of vaccination as compared to infection, and thereby also introduced a common currency to compare the costs and benefits of vaccination [10] [11] [12] [13] [14] 20] . Using relative perceived risk of vaccination as the basis of comparison has the advantage that it can be easily modeled. However, this approach also has some disadvantages as it assumes that the payoff loss for individuals who choose to vaccinate is a fixed quantity (the relative perceived cost of vaccination) which is unrelated to the actual number of adverse events in the population, while the payoff loss for individuals who choose not to vaccinate is proportional to the prevalence of infection, and so does not take into account the discounting of different costs and benefits (e.g., individuals who successfully recover from infection reap the long-term benefit of prolonged immunity). Ultimately, we believe that game theoretical models such as the one we have analyzed here should be refined to include the dynamics of human risk perception. In such models the perceived risks of infection and antiviral treatment are not static (as in [10] [11] [12] [13] [14] ) but dynamically adjusted, being functions of the different types of adverse events (different types of morbidity, deaths) that actually occur in the population. Of course, how such dynamical human risk perceptions can or should be modeled is not straightforward, and would necessitate adding a fair bit of sociology to our epidemic-game theoretical model. Throughout this paper we have made the simplifying assumptions that individuals and public health officer's act rationally, have perfect information and foresight, and that the details of population structure and antiviral drug action play a minor role. These assumptions were made in order to be able to keep the analyses manageable, and to be able focus in detail on the conflict of interest. We are, of course, aware that in the real world a variety of complicating factors play a role. Therefore, our study is not intended nor suited to make quantitative predictions, but it serves to explore how the public and individual interests are shaped by pathogen transmissibility, cost of antiviral treatment, and antiviral efficacy for susceptibility and infectiousness. It would be interesting to extend the model in a number of directions in order to be able to make specific predictions for specific viral threats. For this purpose several steps should be taken. First, depending on the precise research question some form of population structure would probably need to be taken into account. For instance, if the goal were to decide how a limited supply of antiviral drugs is best distributed across different risk groups, the model would need to include different risk groups and take into account that the stockpile of antiviral drugs or vaccines is not infinitely large [21] [22] . Alternatively, if the goal were to investigate whether local containment is possible by means of a targeted antiviral drug treatment program, it would be necessary to include spatial structure, household structure, and possible also workplace structure [6] [7] [8] . Overall, however, we believe that the present state of knowledge just barely suffices to make realistic quantitative predictions as to how effective a large-scale prophylactic antiviral drug program will be, let alone that it will be possible to make quantitative predictions when the dynamics of human choice are taken into account. This, of course, is not tantamount to saying that individual choice is unimportant. Fortunately, none of the recent viral threats from the animal reservoir (avian influenza, SARS) has succeeded in getting a definitive foothold in the human population. As a consequence, the key epidemiological characteristics of the next emerging virus (transmissibility, infectious period, virulence) remain unknown. This is also largely true for rare but serious side-effects of antiviral drugs. This has rendered attempts to provide realistic predictions of the effectiveness of control measures such as antiviral treatment somewhat speculative [6] [7] [8] . Our model lacks much of the sophistication of the earlier models, and is not suited to make quantitative predictions. Rather, the analyses have laid out the principles guiding the decisions of rational individuals and public health officers when faced with an emerging viral threat for which antiviral drugs can be deployed as a first line of defense. Text S1 Model structure and details of the model analyses Found at: doi:10.1371/journal.pone.0001558.s001 (0.10 MB PDF)

Cross-subtype Immunity against Avian Influenza in Persons Recently Vaccinated for Influenza

Avian influenza virus (H5N1) can be transmitted to humans, resulting in a severe or fatal disease. The aim of this study was to evaluate the immune cross-reactivity between human and avian influenza (H5N1) strains in healthy donors vaccinated for seasonal influenza A (H1N1)/(H3N2). A small frequency of CD4 T cells specific for subtype H5N1 was detected in several persons at baseline, and seasonal vaccine administration enhanced the frequency of such reactive CD4 T cells. We also observed that seasonal vaccination is able to raise neutralizing immunity against influenza (H5N1) in a large number of donors. No correlation between influenza-specific CD4 T cells and humoral responses was observed. N1 may possibly be a target for both cellular and humoral cross-type immunity, but additional experiments are needed to clarify this point. These findings highlight the possibility of boosting cross-type cellular and humoral immunity against highly pathogenic avian influenza A virus subtype H5N1 by seasonal influenza vaccination.

I nfl uenza viruses are segmented, negative-sense RNA viruses belonging to the family Orthomyxoviridae. According to the antigenic differences in nucleoprotein and matrix proteins, 3 types of infl uenza viruses (A, B, and C) have been described. Infl uenza viruses A and B are associated with seasonal illness and death, whereas infl uenza virus C causes mild infections (1, 2) . Infl uenza A viruses are subtyped on the basis of the antigenic differences on external hemagglutinin (HA) and neuraminidase (NA) glycoproteins. Human type A infl uenza virus subtypes have been limited to H1, H2, and H3 and to N1 and N2 (3) . Several HAs and NAs have been isolated from avian hosts; occasionally, they have been associated with human outbreaks (4, 5) . Cytotoxic T lymphocytes play a central role in the clearance of primary infl uenza virus infection, peaking after 7-10 days; the peak in antibody titers occurs 4-7 weeks after primary infection (6) (7) (8) . Neutralizing antibodies are completely protective against secondary challenges only with closely related strains, but they are ineffective against viruses with major antigenic divergence. For this reason, current infl uenza vaccines are prepared annually on the basis of World Health Organization forecasts on the most probable infl uenza A and B virus strains thought to be circulating in the next seasonal outbreak (5, 7) . By contrast, cellular responses to cross-reactive epitopes may provide a substantial degree of protection against serologically distinct viruses (9) . The ability of infl uenza viruses to mutate and reassort their HA-NA genome segments between different animal species is a main concern because immunity generated by previous infections or vaccinations is unable to prevent infection by itself, although it may reduce virus replication and spread (8) (9) (10) . To date, 3 infl uenza subtypes have produced pandemic disease in humans: H1N1 in 1918, H2N2 in 1957, and H3N2 in 1968 (4,11,12) . In 1997, during the avian infl uenza (H5N1) outbreak in Hong Kong Special Administrative Region, People's Republic of China, a cross-reactive cellular immune response induced by infl uenza (H9N2) was able to protect chickens from infl uenza (H5N1) (13) . Moreover, adults living in the United States who were never exposed to H5N1 subtype have shown cross-type cellular immunity to infl uenza A virus strains derived from swine and avian species (including the H5N1 subtype isolated in Hong Kong) (14) . Thus, speculation that cross-reactive T cells may decrease illness and death by reducing the replication of the new infl uenza virus, even if elicited by a different strain, is reasonable. Avian infl uenza A viruses of the H5N1 subtype are currently causing widespread infections in bird populations. Numerous instances of transmission to humans have been recently reported in Asia and Africa, with the infection resulting in severe disease or death (>50% fatality rate). Hence, the aim of the present study was to evaluate the immune cross-reactivity between human and avian infl uenza (H5N1) strains in healthy donors recently vaccinated for seasonal infl uenza A (H1N1/H3N2). Our data indicate that infl uenza vaccination may boost cross-subtype immunity against infl uenza (H5N1), involving cellular or humoral responses or both. Healthcare workers wishing to receive seasonal infl uenza vaccination at the Spallanzani Institute (n = 42) were enrolled. The study was approved by the local Ethical Committee; all participants gave written informed consent. Baseline characteristics of the study population are reported in the Table. Blood samples were obtained before (t0) and 30 days after vaccination (t1). The vaccine formulation was Fluarix, an inactivated and purifi ed split infl uenza vaccine (GlaxoSmithKline, Verona, Italy). The antigen composition and strains were A/California/7/2004-H3N2; A/New Caledonia/20/99-H1N1; and B/Shanghai/361/2002. Each 0.5-mL vaccine dose contains 15 μg HA of each strain in phosphate-buffered saline and excipients. Vaccine was administered intramuscularly. Madin-Darby-canine kidney (MDCK) cells were maintained in Dulbecco modifi ed Eagle medium (DMEM) containing 10% fetal calf serum (FCS), and 2 mmol/L Lglutamine, at 37°C in a 5% CO 2 humidifi ed atmosphere. The infl uenza (H5N1) virus used was strain A/Hong Kong/156/97 (kindly provided by Paul Chan) (15) . The virus stock used as challenge antigen in the hemagglutination inhibition (HI) assay was propagated in the allantoic cavities of 10-day-old embryonated hen eggs. The allantoic fl uid was harvested 48 h postinoculation and clarifi ed by centrifugation. Virus concentration was determined by HA titration as previously described (16) , and the virus was stored at -80°C until used. The virus stock used in the microneutralization (NT) and in the cell-mediated immunity assays was propagated in MDCK cells, and the culture supernatants were collected 48 h postinoculation. The 50% tissue culture infectious dose (TCID 50 ), determined by titration in MDCK cells, was calculated by the Reed and Muench method (17) . Infl uenza vaccine, UV-inactivated MDCK-derived infl uenza virus (H5N1), or synthetic infl uenza (H5N1) oligo- Emerging Infectious Diseases • www.cdc.gov/eid • Vol. 14, No. 1, January 2008 peptides were used as antigens for cell-mediated immunity. Infl uenza virus (H5N1) was inactivated by exposure to UV light for 10 min, and complete inactivation of UV-exposed virus was checked by infecting MDCK monolayers with undiluted preparation and by back-titrating the infectivity after 5 days postinfection. Four synthetic peptides of the infl uenza (H5N1) were purchased from Biodesign International (Kennebunk, ME, USA). The sequence of these peptides is specifi c for H5-C-terminal (15 aa), H5-middle region (14 aa), and N1-C-terminal (15 aa) and for N1middle region (16 aa), with no cross-matching with other HA and NA sequences. These peptides can bind different HLA-DRB1 alleles, as established according to the SYF-PEITHI site (www.syfpeithi.de). Specifi cally, N1-specifi c peptides can bind the following HLA-DR alleles: HLA-DRB1*0101, B1*0301, B1*0401, B1*0701, B1*1101, B1*1501. The H5-specifi c peptide from the N-terminal region binds several HLA-DR alleles (HLA-DRB1*0101, B1*0301, B1*0401, B1*0701, B1*1101, B1*1501). In contrast, the H5-specifi c peptide from the middle region did not appear to bind any HLA-DR alleles. According to the HLA-DRB1 allele frequency in the local population, these peptides can be effi ciently presented by most (up to 84%) of study participants. Cell-mediated response was assessed by detecting intracellular interferon-gamma (IFN-γ) production by effector T cells, after antigen-specifi c stimulation in vitro to generate effector cells from memory cells (18) . Peripheral blood mononuclear cells (PBMC) were isolated by density gradient centrifugation (Ficoll-Hypaque, Pharmacia Biotech, Uppsala, Sweden) and frozen at -150°C. Thawed PBMC in culture medium (RPMI 1640, 10% FCS, 2 mmol/L L-glutamine) were stimulated with the infl uenza vaccine preparation (1.5 μg/mL), UV-inactivated infl uenza (H5N1) (MOI 0.1), or synthetic infl uenza (H5N1) peptides (NA and HA) (1 μg peptide/mL) for 3 days and expanded for 6 additional days in the presence of recombinant interleukin-2 (IL-2) (5 IU/mL, Boehringer-Mannheim, Mannheim, Germany). On day 9, cells were restimulated with the same antigens in the presence of 1 μg/mL αCD28 and αCD49d (immunoglobulin G1 [IgG 1 ], K clones CD28.2 and 9f10, respectively; Becton Dickinson, Mountain View, CA, USA) and of Brefeldin-A (10 μg/mL, Sigma, St. Louis, MO, USA). As a negative control, a mock virus preparation, obtained with uninfected MDCK cells, or irrelevant peptides were used. To control the spontaneous cytokine production, cells incubated with only αCD28 and αCD49d were included. In addition, the frequency of IFN-γ-producing CD4 T lymphocytes from each donor in the absence of any stimulation was used to calculate the background for each stimulation. The resulting background levels were very low in every experiment, and no differences were observed between samples obtained before (t0 0.03% ± 0.04%) and after vaccination (t1 0.01% ± 0.03%). The frequency of antigen-specifi c CD4 T cells for each study participant was calculated by subtracting the relative background levels at t0 and t1. Cell-mediated immunity was considered positive when the net increase was >0.2%. Although retesting samples on separate occasions gave reproducible results, t0 and t1 samples for each participant were tested simultaneously to further reduce test variability. Monoclonal antibodies coupled with phycoerythrin (PE), peridinin-chlorophyll protein (PerCP), allophycocyanin (APC), and phycoerythrin-Cy-7 (PE-Cy7) were combined for simultaneous staining. In this study the following were used: anti-CD4 PerCP (IgG1, clone SK3), anti-CD3 PE-cy7 (IgG2a, clone SK7), anti-IFN-γ APC (IgG1, clone B27), and anti-IL-2 PE (IgG1, clone 5344.111) (Becton Dickinson). Cells were stained as previously described (19) . Multiparametric fl ow cytometry was performed by using a FACSCanto fl ow cytometer (Becton Dickinson). A total of 300,000 live events were acquired, gated on small viable lymphocytes, and analyzed with FACSDiva software (Becton Dickinson). The instrument was routinely calibrated according to the manufacturer's instructions. The NT was performed according to a previously described procedure (20) , in agreement with indications from the World Health Organization (21) and the US Department of Health and Human Services (22) . Specifi cally, 2fold serial dilutions of heat-inactivated (30 min at 56°C) human sera were performed in 50 μL DMEM without FCS in 96-well microplates. An equal volume of infl uenza virus (H5N1) (10 3 TCID 50 /mL) was then added to each well. Uninfected-cell wells, incubated with each test serum, were included in each plate as negative controls. After 1 h incubation at 37°C, the mixtures were transferred on MDCK cell monolayers and adsorbed at 37°C for 1 h. After washing, DMEM was added, and the plates were incubated for 2 days at 37°C in 5% CO 2 . NT titer was assessed as the highest serum dilution in which no cytopathic effect was observed by light microscope inspection. All serum specimens were tested in duplicate, and t0, and t1 samples from each patient were assayed in the same plate at the same time. The results were scored by persons blinded to the study participant's identifi cation. The test results were reproducible because random replication of the assays on independent occasions gave consistent results. The antibody titer was also established by HI test, using for challenge either the seasonal vaccine or the egg-derived infl uenza (H5N1) preparation. HI assays were performed in V-bottom 96-well plates with 0.5% chicken erythrocytes, as described (16) . All experiments with live highly pathogenic avian infl uenza A virus (H5N1) were conducted by using Biosafety Level 3-plus (BSL3+) containment procedures (23) . All investigators were required to wear appropriate masks with HEPA fi lters. The frequency of circulating antigen-specifi c CD4 T cells in healthy donors enrolled in the study was analyzed by fl ow cytometry, by using intracellular cytokine staining assay after the in vitro expansion of effector cells. To generate effector cells from their memory precursors, PBMC were challenged with antigen in vitro for 3 days and expanded for 6 additional days in the presence of IL-2 (18) . Effector cells were characterized for their ability to release IFN-γ when cultured overnight in the presence of antigen. CD4 T cells were gated and analyzed for IFN-γ and IL-2 cytokine expression. A representative experiment with PBMC from a recently vaccinated healthy donor is shown in Figure 1 . Without stimuli, no cytokine production in CD4 T cells was detected (Figure 1, panel A) . However, the stimulation with the seasonal infl uenza vaccine preparation induced the production of IFN-γ by CD4 ef-fector T cells (Figure 1, panel B: 3 .2% of IFN-γ+ CD4+ T cells). Stimulation with inactivated infl uenza (H5N1) virus induced a CD4 T-cell response (Figure 1 , panel C: 1.0% of IFN-γ+ CD4+ T cells). Finally, some CD4 T cells specifi c for a pool of H5 and N1 (H5/N1) peptides were also generated in this donor (Figure 1 , panel D: 0.6% of IFN-γ+ CD4+ T cells). No IL-2 production was observed in these experimental conditions. When the extent of CD4 T-cell-mediated immunity before and after seasonal infl uenza vaccination was compared in the healthy donors enrolled in the study, a nonhomogeneous pattern of responses was detected (online Appendix Figure; available from www.cdc.gov/EID/content/14/1/121-appG.htm. After vaccination (t1), a 2-fold variation of the frequency of antigen-specifi c T cells higher than baseline was arbitrarily considered signifi cant. According to this threshold, an increased frequency of IFNγ-producing CD4 T cells specifi c for vaccine preparation was observed after vaccination in 5 (donors 8, 11, 17, 26, 42) of 21 donors (23.8%). A slight increase of frequency of the vaccine preparation-specifi c CD4 T cells was observed in 5 donors (donors 9, 12, 33, 36, 40; 23.8%); a mild-tosignifi cant decrease was observed in the remaining donors (n = 11; 52.3%). As shown in the online Appendix Figure, subtype (online Appendix Figure, panel B); among them, 3 were also showing an increase of the frequency of IFNγ-producing CD4 T cells specifi c for vaccine preparation (donors 11, 17, 42). Two of them, donor 11 and donor 42, had a signifi cant increase of IFN-γ-producing CD4 T cells specifi c for H5/N1 peptides (online Appendix Figure, panel C), which suggests that cross-type immunity may directly involve the HA/NA proteins. Furthermore, 3 other donors (donors 12, 16, 36) had an increased frequency of H5/N1 peptides-specifi c CD4 T cells, even if they were unable to respond to whole virus. Indeed, in some persons we also observed a signifi cant decrease at t1 in CD4 T cells specifi c for vaccine preparation (donors 2, 16, 23, 27, 30, 31, 35, 41) , specifi c for infl uenza (H5N1) (donors 4, 16, 27) , and specifi c for H5/N1 peptides (donors 2, 27, 34, 39, 41). Donors with a reduced specifi c response to vaccine preparation at t1 showed a higher frequency of specifi c CD4 T cells at t0 when compared to other donors (3.4% ± 0.88 vs. 1.29% ± 0.35, respectively, p = 0.013). Similar results were obtained when we observed the infl uenza virus (H5N1) (1.07% ± 0.47 vs. 0.14% ± 0.03, respectively, p = 0.0093) and H5/N1 peptides (1.19% ± 0.54 vs. 0.13% ± 0.07, respectively; p = 0.0018). Because some study participants were reactive to inactivated infl uenza virus (H5N1) as well as to a peptide pool composed of 2 peptides from H5 and 2 from N1 consensus sequences, we analyzed whether this reactivity was preferentially directed against HA or NA. As shown by PBMC from a representative donor in Figure 2 , the frequency of IFN-γ-producing CD4 effector T cells was appreciable after challenge with the inactivated infl uenza virus (H5N1) (Figure 2, No specifi c CD4 T cells producing interferon-gamma (IFN-γ) were observed after challenge with H5 peptides (D). As negative control, either mock-infected culture supernatants or irrelevant peptides were used, giving results very similar to unstimulated cultures (not shown). A similar pattern was observed in 4 other study participants, supporting the hypothesis that the actual target of cross-subtype immunity could be N1. pattern was observed in 4 other study participants, which supports the hypothesis that the target of cross-subtype immunity could actually be N1. Human sera from the same donors were tested for HI activity against both vaccine and infl uenza (H5N1) preparations and for neutralization activity against infl uenza (H5N1) virus. Individual titers are reported in Figure 3 . A 4-fold rise in HA antibody titer is considered noteworthy, and after vaccination most donors (28/38; 73.7%) showed a noteworthy rise of HI titers against vaccine preparation, as indicated by an asterisk (Figure 3, top panel, black bars) . HI titers against infl uenza virus (H5N1) remained at undetectable levels after seasonal vaccination (data not shown), but a rise of neutralization titer >20-fold over baseline was observed in 13 (34.2%) of 38 donors ( Figure 3 , bottom panel, asterisk). All but 1 study participant also responded to seasonal vaccination by a rise in HI titers against vaccine preparation. One donor (21) showed high titers against the H5N1 subtype in NT but a low HI titer against vaccine, a unique situation in the study population. However, antibodies to both anti-infl uenza (H5N1) and infl uenza vaccine are raised by vaccination. Our fi ndings indicate that seasonal vaccination can raise neutralizing immunity against infl uenza (H5N1), which shows the existence of an antibodydependent cross-type immunity. No correlation between infl uenza-specifi c CD4 T cells and humoral responses was observed, which suggests that this type of antibody response was mainly CD4 T-cell independent. We observed that infl uenza-specifi c CD4-effector T cells could be generated by long-term cultures in vitro and easily monitored by fl ow cytometry as IFN-γ-producing cells. When this approach was used, a small frequency of CD4 T cells specifi c for H5N1 subtype could be detected in several persons at baseline. Seasonal vaccine administration may enhance the frequency of reactive CD4 T cells, boosting the cross-subtype cellular immunity against avian infl uenza (H5N1). We also observed that seasonal vaccination raised neutralizing immunity against H5N1 subtype in a large number of donors, showing the existence of an antibody-dependent cross-type immunity. Thus, cross-reactive immunity may involve cellular and/or humoral responses, but the humoral response seems to be CD4 independent. From the present data, N1 appears to be 1 target for cross-type cellular immunity, although we could not rule out the involvement of different (i.e., internal) antigens as possible targets of immune recognition by effector CD4 T cells. Nevertheless, in animal models, cellular immunity (mainly CLT) targeting internal proteins (i.e., NP), partly responsible for heterosubtypic protection, was not induced effi ciently by inactivated vaccines (24) . We did not use live virus, only inactivated split vaccine, whole inactivated virus, or HA and NA peptides for the infl uenza (H5N1) A/ Hong-Kong/156/97 strain. From our data, discriminating between the CD4 T-cell response against external or internal antigens in the case of vaccine preparation was not possible. For H5N1 subtype response, we can presume that the response is against the external antigens and that the results against peptides point to a specifi c response against NA. Results obtained with the whole virus and those obtained with the H5 and N1 peptides are not in complete agreement (online Appendix Figure) . This fi nding can be explained on the basis of the substantial differences in the antigen presentation underlying the whole virus and peptides. Moreover, we observed that a high activation of specifi c cells at baseline (t0) was associated with a reduced specifi c response after vaccination (t1), which suggests that stimulation of pre-activated T cells with high dose of antigen could induce T-cell anergy (25) with consequent loss of immune response. Preliminary evidence also suggests that humoral crosstype immunity is targeting antigens differently from HA: sera from persons showing signifi cant neutralizing titers against infl uenza (H5N1) did not recognize insect cells expressing HA from the H5N1 subtype (not shown) and did not show HI activity against H5N1 subtype. N1 may possibly also be a target of humoral immunity, but additional experiments such as Western blot analysis or inhibition of NA activity (26) are needed to clarify this point. In animals, exposure to 1 specifi c subtype of infl uenza A virus can also induce protective immunity against challenges with other subtypes. This heterosubtypic or crossprotective immunity could represent a key mechanism for facing, and limiting, new infl uenza outbreaks. In 1997, during the Hong Kong infl uenza (H5N1) outbreak, an immune response induced by an infl uenza virus (H9N2), being T cells but not antibodies, protected chickens from lethal infl uenza (H5N1) (13) . Moreover, adults living in an urban area of the United States have been described as having infl uenza-specifi c memory T cells that recognize epitopes of infl uenza A virus strains derived from swine and avian species, including the infl uenza (H5N1) strain involved in the Hong Kong outbreak in humans (14) . Our data confi rm that persons who have never been exposed to H5N1 subtype may be able to generate a cell-mediated response against the Hong Kong infl uenza (H5N1) isolate. This cross-type response may be naturally occurring (probably as a consequence of exposure to seasonal infl uenza strains). In mice, both CD4 T-cell-independent and -dependent antibody responses contribute to the control of infl uenza virus infection (27, 28) . Although antibodies appear to facilitate the recovery from infl uenza infection, it is generally believed that B cells cannot produce neutralizing, isotypeswitched, infl uenza-specifi c antibodies in the absence of CD4 T-cell help (29, 30) . However, other data clearly demonstrate that B cells can also produce anti-infl uenza IgA, IgM, and IgG responses independent of CD4 helper T cells (27, 31) . A non-antigen-specifi c bystander response driven by activated CD4 T cells specifi c for heterologous antigen may contribute to so-called heterosubtypic immunity (8) (9) (10) 12) . However, the ability of infl uenza virus infection to promote B-cell activation and differentiation into shortlived, isotype-switched, antibody-secreting cells may result from a combination of B-cell receptor hypercross-linking, engagement of toll-like receptors, production of cytokines, as well as triggering of innate immunity. In our study, cellular and humoral cross-reactive immunity seemed to target antigens other than HA. Infl uenza (H5N1) cases occur mainly in young people (32) . This fi nding may be explained by hypothesizing that older people, although not previously exposed to H5N1 subtype, may have gained protective immunity by previous infections sustained by circulating infl uenza virus strains. It has also been shown that immunity to the N1 NA from the human infl uenza virus cross-reacts with the avian N1 NA virus and that this cross-reactivity protects mice against infection with the avian infl uenza virus (H5N1) (26) . All these fi ndings may be explained by hypothesizing that cross-reactive immunity is targeting the N1 NA antigen. However, whether cross-reactive antibodies to NA and CD4 T cells would be protective against illness and death, especially from infl uenza (H5N1) infection is not known. Further studies will be necessary to elucidate this point. In conclusion, we demonstrated that vaccination against seasonal infl uenza may boost a cross-reactive immunity against an unrelated strain responsible for deadly infections in humans, i.e., the avian infl uenza (H5N1) strain A/Hong Kong/156/97. These data, together with previous experimental results from mice studies and epidemiologic reports, indicate that cross-type immunity should be considered an important component of the immune response against novel infl uenza A infections.

Human Bocavirus Infections in Hospitalized Children and Adults

Studies have reported human bocavirus (HBoV) in children with respiratory tract infections (RTIs), but only occasionally in adults. We searched for HBoV DNA in nasopharyngeal aspirates (NPAs) from adults with exacerbations of chronic bronchitis or pneumonia, from children hospitalized for acute RTIs, and from asymptomatic children during the winter of 2002–2003 in Canada. HBoV was detected in NPAs of 1 (0.8%) of 126 symptomatic adults, 31 (13.8%) of 225 symptomatic children, and 43 (43%) of 100 asymptomatic children undergoing elective surgery. Another virus was detected in 22 (71%) of the 31 HBoV-positive NPAs from symptomatic children. Two clades of HBoV were identified. The pathogenic role of HBoV in RTIs is uncertain because it was frequently detected in symptomatic and asymptomatic children and was commonly found with other viruses in symptomatic children.

H uman bocavirus (HBoV) is a newly described human virus closely related to bovine parvovirus and canine minute virus. It is currently classifi ed in the genus Bocavirus within the family Parvoviridae. This virus was fi rst identifi ed in respiratory tract specimens from Swedish children with lower respiratory tract infections (RTIs) (1) . Nucleic acid amplifi cation has detected HBoV in respiratory samples of children with acute respiratory disease, with incidence rates ranging from 3% to 19% . However, the pathogenic role of HBoV is uncertain because other viruses have been frequently detected in HBoV-positive children with lower RTIs (range 37%-90%) (2, 3, 7, (9) (10) (11) (20) (21) (22) . The objective of this study was to describe the incidence and clinical manifestations of HBoV infections in children and adults with respiratory tract symptoms, including a control group of children without symptoms. Respiratory samples from adults were obtained from a previous study conducted from December 2002 to April 2003 at 3 university-affi liated hospitals in the province of Quebec, Canada (24) . Two groups of patients were enrolled: those >40 years of age with chronic obstructive pulmonary disease (COPD) who came to emergency departments with exacerbation of their illness (including patients with pneumonia), and those >18 years of age without COPD who were admitted to the hospital with a diagnosis of community-acquired pneumonia. Patients were excluded from the study if they came to the hospital >7 days after onset of symptoms. Respiratory samples from children were obtained from a case-control study, the results of which have been published (25) . Participants included children <3 years of age who were hospitalized from December 2002 to April 2003, at Laval University Hospital Center in Quebec City, Quebec, Canada. Case-patients were children admitted for an acute RTI (mostly bronchiolitis, pneumonitis, and laryngotracheobronchitis) who had a nasopharyngeal aspirate (NPA) collected as part of investigation of their illness. A specifi c questionnaire was completed at admission by a research nurse in the presence of the parents. At the end of hospitalization, charts of the children were reviewed to collect clinical and laboratory data. Eligible controls were children hospitalized during the same period for any elective surgery (ear, nose, and throat surgeries in 71% of the cases). These children had no concomitant respiratory symptoms or fever at admission. The study nurse obtained a signed consent from parents and an NPA was obtained during surgery. The original studies were reviewed and approved by the ethics committees of all participating healthcare centers. All pediatric (from case-patients and controls) and adult (case-patients only) NPA specimens were previously analyzed by using a multiplex real-time PCR assay for infl uenza A and B viruses, human respiratory syncytial virus (hRSV), and human metapneumovirus (hMPV) (24, 25) . For symptomatic children, viral cultures and antigen detection assays were performed upon request by the treating physician. Remaining specimens were frozen at -80°C until subsequent HBoV PCR studies. Nucleic acids were extracted from 200 μL of NPA by using the QIAamp viral RNA Mini Kit (QIAGEN, Inc., Mississauga, Ontario, Canada). A duplex HBoV PCR (TaqMan assay) was used to amplify conserved regions of NP-1 and NS-1 genes as described (14) , except that the NS-1 forward primer was replaced with primer 5′-TAG TTG TTT GGT GGG ARG A-3′. Probes were labeled with 6-carboxyfl uorescein (FAM) or tetrachloro-6-carboxyfl uorescein (TET) at the 5′ end and with a quencher at the 3′ end. Amplicons were 81 bp (NP-1) and 74 bp (NS-1), respectively. Duplex amplifi cation was conducted by using 1 μmol/L NS-1 forward primer and 0.4 μmol/L NS-1 reverse primer and the 2 NP-1 primers. Taqman probes were used at concentrations of 0.1 mmol/L for the NP-1 gene and 0.2 mmol/L for the NS-1 gene (14) . The amplifi cation master mixture consisted of 2.5 mmol/L MgCl 2 , 3.33 mg/mL bovine serum albumin, 0.2 mmol/L of each of the 4 deoxynucleotide triphosphates (Amersham Biosciences, Uppsala, Sweden), 10 mmol/L Tris-HCl, 50 mmol/L KCl, 0.625 U Promega Taq DNA polymerase (Fisher Scientifi c, Markham, Ontario, Canada) combined with TaqStart antibody (BD Biosciences Clontech, Palo Alto, CA, USA), and 3 μL DNA in a fi nal volume of 25 μL. PCR amplifi cation (180 s at 94°C and 45 cycles for 10 s at 95°C, 30 s at 58°C, and 30 s at 72°C) was performed in a Smart Cycler thermal cycler (Cepheid, Sunnyvale, CA, USA). A PCR extension step of 5 min at 72°C was performed at the end of the cycling protocol. An HBoV infection was defi ned by a positive PCR result for NP-1 and NS-1. The duplex assay had a sensitivity of 10 genome copies for NP-1 and NS-1 gene targets on the basis of quantifi cation analysis of positive control plasmids. Half of the HBoV-positive samples were randomly selected for phylogenetic analysis, which consisted of amplifying and sequencing a 842-bp region of the VP1/VP2 genes as described (6) . The VP1/VP2 nucleotide sequences from this study, as well as prototype sequence type (ST)1 and ST2 (1), were entered into a multiple alignment generated by ClustalW software version 1.83 (www.molecularevolution.org/software/clustalw) and corrected through fi nal visual inspection with the SeqLab application (Wisconsin package version 10.3; Accelrys, San Diego, CA, USA). Phylogenetic analyses were conducted with the MEGA version 3.1 software (26) by using the distance method and the neighbor-joining algorithm with Kimura-2 parameters. Topologic accuracy of the tree was evaluated by using 1,000 bootstrap replicates. Proportions of clinical characteristics in different groups of patients were compared by using the χ 2 test or the Fisher exact test. The Wilcoxon nonparametric test was used to compare age distribution and length of stay. Analyses were performed by using SAS software version 9.1 (SAS Institute, Inc., Cary, NC, USA). HBoV DNA was detected in NPA samples from 1 (0.8%) of 126 symptomatic adults (71 years of age) and from 31 (13.8%) of 225 symptomatic children (mean age 17 months, median age 15 months). However, HBoV was detected more frequently (43%, p<0.001) in the 100 asymptomatic control children (mean age 22 months, median age 23 months). Another virus was detected in 22 (71%) of 31 HBoV-positive NPAs from symptomatic children. The virus most commonly co-isolated with HBoV was hRSV (16/31, 52%), followed by infl uenza A/B (3 cases), hMPV (3 cases), adenovirus (1 case), and parainfl uenza virus (1 case). Two children were infected with 2 other viruses in addition to HBoV. The median age of symptomatic children with HBoV infection (15 months) was signifi cantly greater than that of symptomatic children without HBoV infection (8 months; p<0.0001). The hospital length of stay was similar for children positive for HBoV DNA (mean 5.1 days, median 4 days) and those negative for HBoV DNA (mean 6.6 days, median 3 days) (p = 0.9). Clinical characteristics of HBoV-positive children are summarized in the Table. There were signifi cantly fewer bronchiolitis episodes in children infected only with HBoV than in children infected only with hRSV (p<0.0001). None of the children with single HBoV infections and only 2 (6%) of all 31 HBoV-infected children were admitted to the intensive care unit. In the control group of asymptomatic children who underwent elective surgery, ear, nose, and throat surgery was more frequently performed in children with HBoV infections (36/43, 84%) than in children without HBoV infections (35/57, 61%) (p = 0.014). Ear, nose, and throat elective surgeries consisted mainly of myringotomies, adenoidectomies, and tonsillectomies. The 1 adult with an HBoV infection was a 71-year-old man (a smoker) who came to the hospital for a COPD exacerbation and was treated with systemic corticosteroids and antimicrobial drugs. No other microbiologic agents (bacteria or viruses) could be identifi ed in his sputum or NPA. He was hospitalized for 11 days. Sequence analysis of the HBoV VP1/VP2 genes performed on ≈50% of HBoV-positive specimens showed 2 distinct clades of viruses (Figure) . These genotypes clustered with the original strains described by Allander et al. (ST1, GenBank accession no. DQ000495, and ST2, Gen-Bank accession no. DQ000496) (1). There was no temporal link between the clades because both were equally distributed throughout the study period. No obvious relationship was found between clades and the presence or absence of symptoms. Results from our study indicate that HBoV was rarely detected in adults with respiratory symptoms but was frequently detected in symptomatic and asymptomatic children during the 2002-2003 winter season. HBoV was detected in NPA samples from 1 (0.8%) of 126 symptomatic adults, 31 (13.8%) of 225 symptomatic children, and 43 (43%) of 100 asymptomatic children. Another virus was detected in 22 (71%) of 31 HBoV-positive samples from symptomatic children. Overall, these data do not support a pathogenic role for HBoV in acute RTIs in children. The full spectrum of clinical diseases associated with HBoV infections and the epidemiology of this new parvovirus are not fully understood. This is particularly true for adult patients in whom few studies have been performed. Allander et al. (1) found no HBoV DNA in 112 culturenegative NPA samples from adults with respiratory symptoms. Bastien et al. (5) reported an overall rate of infection of 1.5% in respiratory samples negative for other viruses, with no differences between age groups. Maggi et al. (16) reported only 1 HBoV-positive sample from an adult with lymphoma in 62 bronchoalveolar lavages (BALs). These investigators also tested 22 nasal swabs from adults with persistent asthma symptoms and found no samples positive for HBoV. Fry et al. (10) identifi ed HBoV DNA in 1% of adults >20 years of age hospitalized with pneumonia in Thailand. Kupfer et al. (27) described a case of HBoV infection associated with severe atypical pneumonia in a patient with non-Hodgkin lymphoma who was also infected with cytomegalovirus in a BAL sample. We found 1 case of HBoV infection in an adult, which represented 0.8% of the tested population. The HBoV-positive adult did not show immunosuppression but was treated with corticosteroids for a COPD exacerbation. Overall, our results are consistent with those of previously described studies and support the fact that HBoV infection is rare in adults but may occur more frequently in those with other illnesses or immunosuppression. Studies have reported HBoV DNA in 3%-19% of children with RTIs. Rates of detection tend to be higher in children <1 year of age (4, 10) . The incidence of HBoV infections also tends to be higher in samples from the lower respiratory tract, such as NPA or BAL (4.4%-19%) (2, 7, 9, 11, 13, 19, 21, 22) , than in nasal swabs (1%-6%) (10, 15, 16, 18) . The percentage of co-pathogens in our HBoV-positive children (71%) was comparable with those reported in the literature, with rates of co-infections ranging from 35% to 90% (2, 3, 7, (9) (10) (11) (20) (21) (22) . Moreover, coinfecting viruses detected in conjunction with HBoV in our population were similar to those described in other studies, i.e., hRSV, infl uenza A virus, and adenovirus (11, 23) . The high frequency of HBoV detection (43%) in our asymptomatic children contrasts with the results of the few other studies that included a control group of asymptomatic children. Fry et al. (10) detected HBoV DNA in only 1% of nasal swabs from asymptomatic patients. Maggi et al. (16) did not detect HBoV DNA in nasal swabs from 51 asymptomatic children (including 30 healthy infants with a mean age of 6 months and 21 preadolescent healthy children with a mean age of 12.8 years). However, these studies analyzed nasal swabs instead of NPA or BAL samples for HBoV detection, which may result in lower rates of viral detection, as shown in symptomatic persons. Allander et al. (2) did not detect HBoV in any of 64 asymptomatic children (median age 4.1 years, range 5 months to 14 years) but used nasal swabs in asymptomatic patients and NPA samples in symptomatic patients. Furthermore, their control group was also older than our population (mean 18.6 months, median 18 months). Kesebir DNA in nasal washes from 96 asymptomatic children <2 years of age seen at a clinic compared with 22 (5.2%) of 425 various samples from symptomatic children sent to a hospital clinical laboratory. None of the previous studies used a control group consisting of children matched for age and week of admission and analyzed the same type of respiratory samples for cases and controls. Our positive results for HBoV were confi rmed by using 2 sets of PCR primers targeting different genes (NP1 and NS1) in a duplex PCR assay and by subsequent testing with a third set of primers (VP1/VP2) for sequencing. Also, sample preparation and PCR amplifi cation were performed in separate laboratory areas following the stringent quality control program of our institution. Thus, it is unlikely that our positive results were due to PCR cross-contamination. Our method was also very sensitive (detection limit = 10 genome copies), which probably enabled an increased in-fection rate compared with previous reports. We cannot exclude the possibility that prior RTIs (in the few weeks preceding sampling) occurred in our asymptomatic children hospitalized for an elective surgery or that HBoV could establish a prolonged infection in children compared with other respiratory viruses. However, the 3× higher detection rate in controls than in symptomatic children make these explanations unlikely. We did not quantify HBoV DNA load in samples from our study, which could have been different between asymptomatic and symptomatic children. Nevertheless, we detected hRSV, hMPV, and infl uenza virus RNA in <1% of the same NPA samples from those asymptomatic children compared with a rate of 43% for HBoV DNA (25) . At the very least, our results should raise concerns about the pathogenic role of HBoV in children. We detected 2 HBoV genotypes circulating at the same time in both symptomatic and asymptomatic children during the winter of 2002-2003 in Quebec. This result is consistent with fi ndings of other groups from North America and Europe during 2002-2004 and highlights the fact that HBoV lineages do not appear to be geographically clustered (1, 6, 9, 12) . The seasonality of HBoV infection is still a matter of debate, but it seems to involve primarily the colder months of the year (9, 20, 21) . However, most studies, including ours, were performed during the typical respiratory virus season, which may have introduced a bias. Additional studies are needed to address the prevalence of HBoV outside the respiratory virus season and its role in nonrespiratory syndromes. Moreover, the possibility that this virus might be transmitted and isolated in the respiratory tract, but could cause viremia and other clinical syndromes such as gastroenteritis, should be investigated. Vicente et al. analyzed 527 stool samples from children with gastroenteritis and no respiratory symptoms and found a positivity rate of 9.1% for HBoV (with a co-infection rate of 58%) (22) . In conclusion, our study shows that HBoV was frequently detected in both symptomatic and asymptomatic children during the winter of 2002-2003 in Quebec City. Conversely, this virus was rarely found in the adult population during the same period. Further studies are needed to establish whether this recently described parvovirus is pathogenic by using well-matched control groups and sequential samples to detect viral persistence. (1) . Numbers along branches are bootstrap values from 1,000 replicates. Scale bar shows 1 substitution for every 1,000 nucleic acid residues.

Pandemic Influenza Planning in the United States from a Health Disparities Perspective

We explored how different socioeconomic and racial/ethnic groups in the United States might fare in an influenza pandemic on the basis of social factors that shape exposure, vulnerability to influenza virus, and timeliness and adequacy of treatment. We discuss policies that might differentially affect social groups’ risk for illness or death. Our purpose is not to establish the precise magnitude of disparities likely to occur; rather, it is to call attention to avoidable disparities that can be expected in the absence of systematic attention to differential social risks in pandemic preparedness plans. Policy makers at the federal, state, and local levels should consider potential sources of socioeconomic and racial/ethnic disparities during a pandemic and formulate specific plans to minimize these disparities.

T he threat of pandemic infl uenza has generated concern among politicians, policy makers, healthcare professionals, and the general public. For the past several centuries, major infl uenza pandemics have occurred every 10 to 30 years (1); it is widely believed that a new pandemic is "inevitable" (2) . The possibility of an imminent infl uenza pandemic has been heightened by the appearance and spread of avian infl uenza A (H5N1), which has a case-fatality ratio of >50% (3) . Although the assumption has been that avian infl uenza viruses could not directly infect humans, the transmission of infl uenza virus (H5N1) directly from chickens to humans in 1997 caused experts to reconsider that assumption (4) . Genetic changes in infl uenza virus subtype H5N1 in 2003 resulted in a new strain of the virus, which spread to multiple countries in East and Southeast Asia (5) , as well as Europe and Africa. Whether the avian infl uenza virus (H5N1) develops human pandemic potential, its spread from birds to humans and the severity of the resulting disease have heightened concerns about a possible future infl uenza pandemic. Considerable fi nancial resources have been devoted to pandemic infl uenza preparedness planning at the federal and state levels (6, 7) ; however, resources at state and local levels may be inadequate to implement a robust preparedness plan (8, 9) . Past experience with natural disasters and current socioeconomic and racial/ethnic disparities in healthcare in the United States (10, 11) raise questions about the adequacy of plans to address the needs of disadvantaged populations. For example, in responding to Hurricane Katrina, planners apparently failed to consider that many low-income persons might lack private modes of transportation and would depend on institutional help for evacuation. Although the evacuation was successful overall (12), deaths, injuries, and illness occurred disproportionately among low-income persons in New Orleans because of economic and logistic constraints on their ability to respond to government recommendations to leave the city. Low-income and disadvantaged persons often suffer disproportionately during natural disasters and epidemics, and historical evidence demonstrates that low-income persons fared considerably worse than high-income persons during the 1918 pandemic in the United States (13) . In this article, we describe ways in which different socioeconomic and racial/ethnic groups might fare differently in an infl uenza pandemic, on the basis of current knowledge of social factors that shape exposure and vulnerability to infl uenza virus and that infl uence the timeliness and adequacy of treatment among those who become ill. We also discuss policy decisions, made either before or during a pandemic, which might differentially affect risk for illness or death for those of low income and of specifi c racial/ethnic groups. Our purposes are to 1) call attention to potentially major and avoidable social disparities in suffering and death during an infl uenza pandemic and 2) highlight the importance of including in pandemic preparedness plans targeted strategies for minimizing or avoiding these social disparities. The following discussion is not meant to be exhaustive; rather, it is meant to provoke refl ection about how potential disparities in the effects of an infl uenza pandemic might be reduced or eliminated through appropriate planning and implementation of clinical and public health activities. Using a conceptual framework adapted from Diderichsen et al. (14) , we systematically considered possible sources of disparities during an infl uenza pandemic by examining the following 3 levels at which underlying socioeconomic or racial/ethnic differences could lead to disparities in illness or death: 1) likelihood of being exposed to the infl uenza virus; 2) likelihood of contracting infl uenza disease, if exposed; and 3) likelihood of receiving timely and effective treatment after infl uenza disease has developed. To explore socioeconomic and racial/ethnic disparities at each level, we searched the literature for relevant fi ndings based on population-based national data (Figure, Table) . Regardless of which strain of infl uenza virus causes the next pandemic, it will be highly transmissible between humans. Transmission of infl uenza is primarily airborne, through aerosolized respiratory tract secretions expelled during coughing and sneezing, although transmission by direct and possibly indirect contact may occur. Transmission can be expected to occur in various settings, including homes, healthcare facilities, schools, work sites, public transportation, and other settings at which people gather for social, commercial, or entertainment purposes. Higher exposure risk among particular population groups as a result of factors such as crowding and occupation could contribute to health disparities among socioeconomic and racial/ ethnic groups during an infl uenza pandemic. Crowding, an established risk factor for many infectious diseases, can increase the likelihood of pathogen transmission. In the United States, urban poverty and Hispanic and Asian ethnicity are correlated with domestic crowding; even at higher income levels, Hispanic and Asian households are relatively more crowded than white and African-American households (15) . In addition, in the United States, low-income persons, African Americans, and nonwhite Hispanics are more likely than persons in other groups to obtain regular medical care at emergency departments and publicly funded clinics (10) , where airborne transmission of infectious agents has been documented. Because these locations typically do not segregate sick and well patients and are becoming increasingly crowded (16) , patients waiting for care in these settings are likely to have greater exposure to infl uenza viruses and other respiratory pathogens. Another source of increased exposure to infected persons is public transportation, where persons from low-income and minority households account for 63% of users (17) . Occupational factors are also likely to lead to differential exposure risk during an infl uenza pandemic, particularly in terms of adherence to strategies that aim to limit case-patient contact with others (18) . Staying home may not be economically feasible for persons in lower wage occupations; these persons are less able to afford losing income as a result of missed work and often lack the job fl exibility that would permit them to work at home. In addition, their jobs may be necessary because they provide essential goods and services. For these reasons, parents in lower wage/lower status occupations may be more likely to keep their children in communal childcare settings-where exposure risks are relatively high-during an infl uenza pandemic, placing everyone in the family at greater risk for exposure. Among persons who have been exposed to infl uenza virus, the likelihood of contracting disease may be modifi ed by underlying host factors and medical conditions, such as age, smoking status, nutritional status, stress levels, and cardiopulmonary disease. The infl uence that most host factors will have on the development of infl uenza during a future pandemic is uncertain; some evidence suggests that the factors affecting disease severity and death may differ from those typically observed during annual infl uenza epidemics (19) . However, given overwhelming evidence that low-income persons are generally more susceptible to infectious diseases, it is reasonable to plan on the basis of well-documented annual epidemic patterns, in which infl uenza disease development is infl uenced by factors that are differentially distributed across socioeconomic and racial/ ethnic groups. These patterns, as well as patterns of many other diseases, indicate that socially disadvantaged groups are likely to be at higher risk for infl uenza disease, particularly severe disease. The inability to predict which infl uenza virus will cause a future pandemic, together with the very limited national and global capacity to produce infl uenza vaccine in massive quantities in a short time, almost ensures that an effective vaccine will be unavailable to most or all of the population during the early stages of a pandemic and in very short supply thereafter. Even so, current plans assume that local and state public health agencies will have a primary role for distributing pandemic infl uenza vaccine. In general, however, these plans do not adequately address preventing or minimizing socioeconomic or racial/ethnic disparities in vaccine distribution and acceptance, despite evidence that such disparities have been the rule for the annual infl uenza vaccine, even among persons >65 years of age (20) . In the United States, routine use of annual infl uenza vaccine in preschool children has only recently been introduced; information focusing on school-age children is limited (21) . Nevertheless, African American/black children and children from lower income families, who are at higher risk of contracting infl uenza (22) in this country, are less likely to be up to date with other routine immunizations (23) . It is possible that, in the context of an infl uenza pandemic, vaccine-seeking and acceptance behavior and resultant coverage patterns may differ from those observed during routine vaccination efforts; however, the weight of available evidence indicates that social disparities in vac-cine coverage are likely to occur in the absence of careful planning to prevent them. Among those who contract infl uenza, subsequent illness and death may be infl uenced by underlying factors and conditions and by the timeliness and effectiveness of various treatment modalities. Most infl uenza illnesses are self-limiting, and most infected persons during both annual infl uenza epidemics and infl uenza pandemics (including that of 1918-19) recover with only supportive care in the community. Even so, current planning efforts recognize the potential importance of reducing disease during a pandemic, through early treatment with antiviral drugs and through other forms of treatment such as respiratory support and antimicrobial agents to treat secondary bacterial pneumonia, among those with more severe disease. In the United States, the likelihood of substantial disparities in access to timely and appropriate care under infl uenza pandemic conditions is high, given long-standing and persistent disparities in access to medical care. For example, persons with low income are ≈2× as likely as those with higher incomes to lack a usual source of healthcare (24) . Similarly, non-Hispanic black and Hispanic persons are signifi cantly less likely than non-Hispanic white persons to report having a usual primary care provider (10). Among persons who do report having a usual source of care, those who are poor or near poor and those who are non-Hispanic black or Hispanic are 2.5-4× as likely as their relatively higher income and white counterparts to rely on a hospitalbased source of primary care (24) . These same groups are also more likely to report having diffi culty obtaining timely appointments for illness or injury, which suggests problems with access to care even among those with a usual source of healthcare (10) . Language and cultural barriers to seeking and receiving medical care also may contribute to disparities. In emergency departments, for example, interpreters are frequently unavailable or underused, which has potentially adverse implications for patients' understanding of their disease or treatment and for clinical decision making Emerging Infectious Diseases • www.cdc.gov/eid • Vol. 14, No. 5, May 2008 711 Table. Factors that could contribute to health disparities among socioeconomic and racial/ethnic groups during an influenza pandemic Differences in exposure to influenza virus Crowding in households, medical facilities, public transportation Occupational factors such as inability to work from home, dependence on childcare outside of the home Differences in susceptibility to influenza disease, once exposed to the virus Host factors, including preexisting immunity, age, other underlying diseases or conditions, smoking, nutritional status, stress Vaccination status, reflecting differences in vaccine seeking and acceptance and in vaccine availability Differences in timely effective treatment, once influenza disease has developed Access to outpatient and inpatient medical care Care-seeking attitudes and behavior Financial obstacles, including lack of adequate insurance coverage Logistic obstacles, including transportation, language Quality of care Availability of antiviral treatments Appropriate inpatient treatment and quality of care (25) . In addition, the large numbers of persons who lack health insurance, as well as those who lack documentation of US citizenship, often delay seeking care because they are concerned about paying for the care or encountering legal diffi culties. Evidence from previous outbreaks suggests that antiviral drugs may be effective for treatment (26) and prevention (27) of pandemic infl uenza, and current antiviral drugs seem to be biologically effective against 1918 and 1918like viruses (28) . Because vaccine may not be available when a pandemic begins, experts have suggested that the antiviral drug oseltamivir should be stockpiled for use during a pandemic infl uenza outbreak. Recent models suggest that early use of oseltamivir may contain outbreaks if certain criteria regarding transmissibility and compliance are met (29) . However, experience with nonpandemic infl uenza indicates that oseltamivir must be given early during symptom development for it to have any substantial biological effect (30) ; modest delays may vitiate the treatment effectiveness (31) . Although plans for release and distribution of antiviral drugs are still being fi nalized, overcoming long-standing disparities in access to timely treatment by socioeconomic status, race/ethnicity, ability to speak English, and legal status will present numerous challenges to ensuring equal access to such drugs during a pandemic. Reasons for concern about disparities in the timeliness and appropriateness of the care received by infl uenza patients who might benefi t from in-hospital care are similar. Given the predicted insuffi cient supply of hospital beds and staff during a pandemic (32), a person's access to potentially lifesaving therapies such as respiratory support and antimicrobial treatment of secondary bacterial pneumonias in an inpatient setting is likely to depend on factors that include usual source of care, citizenship status, and ability to speak English. Disparities may also occur in the quality of care received by persons who are hospitalized. Earlier US studies of persons hospitalized for pneumonia have found that blacks and "other minorities" are 71% and 79% as likely, respectively, as non-Hispanic whites to receive antimicrobial agents within 8 hours of arrival at the hospital (33) and signifi cantly less likely to have blood cultures obtained before receiving antimicrobial therapy (10) . Such disparities in quality of care would likely persist during an infl uenza pandemic. Although reducing or eliminating socioeconomic and racial/ethnic disparities in health and healthcare has been an offi cial federal and state policy priority for 2 decades (34), such disparities remain prevalent and may inadvertently become wider when not explicitly addressed by policies designed to improve the health of the population as a whole and of disadvantaged persons in particular (35) . Given the current limitations of our public health infrastructure and the disparities in healthcare, a pandemic infl uenza outbreak in the United States is likely to disproportionately affect persons from socially disadvantaged groups. Explicit, systematic, and detailed plans are essential for overcoming the social barriers that are predicted to result in socioeconomic and racial/ethnic disparities in pandemic infl uenza illness and death. Saunders and Monet also have called for pandemic infl uenza planning that appropriately considers the needs of disadvantaged populations (36) . The Pandemic Infl uenza Plan of the US Department of Health and Human Services (HHS) (37) does not adequately address potential social disparities in exposure, vaccination, or treatment; the possible effects of such disparities; or strategies for minimizing or eliminating them. The HHS plan (37) , the federal guidance on vaccine allocation (38) , and the recent Centers for Disease Control and Prevention (CDC) guidelines for community-level mitigation strategies (18) should be credited for calling for community engagement and inclusion of a wide variety of stakeholders in planning at the local level. Outreach to providers, community leaders, and organizations, particularly in disadvantaged communities, will be an important component of any strategy for addressing disparities during a pandemic. However, the available versions of offi cial plans do not call attention to the need for special efforts to overcome the greater barriers likely to be faced by socially disadvantaged groups. On a US government website for pandemic infl uenza (www.pandemicfl u.gov), a question asks which groups would be especially vulnerable during an infl uenza pandemic. The answer notes that people may be vulnerable for a variety of reasons, including limited access to healthcare; limited profi ciency in English; or being disabled, homeless, economically disadvantaged, or a single parent. The response calls for faith-based and community-based organizations to develop plans "to care for dependent populations" and to "provide fi nancial aid to the poor who are unable to work and are in need of emergency income for housing, medicine, or other essential needs" (www.pandemicfl u.gov/faq/pandemicinfl uenza/pi-0001.html), which implies that attention to the needs of economically or socially vulnerable persons is not primarily a public-sector responsibility but is more a matter for private charity. The 2005 HHS plan (37) itself acknowledges that some groups may need fi nancial assistance if they are unable to work but does not indicate how that assistance would be provided or who would provide it. Those who are still formulating plans should consider likely differences in infl uenza exposure and identify potential strategies for mitigating such disparities. Mathematical models have demonstrated that community-based interventions, such as quarantine and individual isolation, may be important for reducing infl uenza attack rates and overall incidence (29) . Most pandemic plans call for limiting public gatherings and closing schools to slow the spread of infl uenza, without adequately taking into account how implementing these strategies could differentially affect disadvantaged groups. Recent recommendations from CDC go further in recognizing the differential effect of socialdistancing measures on vulnerable communities (18) . Although CDC advocates fl exible work arrangements, income replacement, and job security to minimize the negative effects of social-distancing measures, it pays inadequate attention to those whose jobs will not accommodate these interventions. More specifi c solutions should be outlined in pandemic preparedness plans to address the economic effects of quarantine on low-income persons, who by staying home may be at risk wage loss, job termination, or both. Job security and income replacement are key components to limiting the effects of potential quarantine measures on disadvantaged persons (39) and should be extended to all persons, regardless of their type of work. Important decisions also will need to be made concerning access to vaccination and treatment in the event of a pandemic. The federal government's Draft Guidance on Allocating and Targeting Pandemic Infl uenza Vaccine (38) provides a basic framework for allocating vaccine during the pandemic. An appendix to that document mentions (on p.17) that the principles of "fairness and equity (recognizing that all persons have equal value, and providing equal opportunity for vaccination among all persons in a priority group)" were considered when drafting the guidelines. Although the proposed schema very reasonably fi rst defi nes groups of different priority levels according to occupation and then, within the general population, according to age and pregnancy status, it does not provide explicit attention to groups who are vulnerable because of social disadvantage. Nor does it note the need for explicit attention to vulnerable social subgroups, for example, low-wage workers in prioritized occupational fi elds and low-income and minority pregnant women, infants, and toddlers. We are not questioning the rationality of defi ning major priority groups according to occupation or of using biological criteria to further prioritize within the general population. Rather, our concern is with the absence of attention to both biological and social risk factors, which must be addressed to overcome the many social barriers to equal opportunity for vaccination. Well-documented evidence of existing healthcare disparities suggests that during a pandemic shortages of infl uenza vaccine, antiviral drugs, inpatient services, and healthcare staff will disproportionately affect persons in socially disadvantaged groups. To limit the crowds that might occur at hospitals and clinics, plans for the release of stockpiles of vaccines, medications, or both could include distribu-tion from private pharmacies or doctors' offi ces. However, because private pharmacies and private practitioners are less likely to be located in lower income neighborhoods, plans to make access to potentially lifesaving vaccines and drugs speedier and more equitable might, in fact, exacerbate disparities. Distribution plans may need to include mobile community health centers (staffed by nurses and nurse practitioners) that can travel to low-income areas, along with a variety of community medical and other service providers and nontraditional sites like soup kitchens, sheltered workshops, and transit points, which have become popular places for administering yearly infl uenza vaccine (40) . Other factors, such as the availability of transportation to a hospital, might also become more important during a pandemic. Access to a private car may be a major determinant of who is able to obtain care, presenting constraints like those that led to disparities in evacuation from New Orleans before Hurricane Katrina. To ensure that disadvantaged communities are reached and that resources are equitably allocated during an infl uenza pandemic, preparedness plans can and should involve community-based providers and organizations that are familiar with vulnerable groups. Social group disparities in exposure, susceptibility, and access to timely and effective treatment for a variety of diseases have been well documented in the United States. Infl uenza pandemic preparedness plans that fail to explicitly provide guidelines on how to mitigate these issues could lead to decisions that may, on the surface, seem reasonable, but that are likely to exacerbate social group disparities in health outcomes. Given the existence of major disparities in health and healthcare, we cannot expect pandemic preparedness and response planning to eliminate the deep divides that exist between socioeconomic and racial/ethnic groups. These disparities can, however, be minimized through careful planning that considers and proactively addresses vulnerability at each level: exposure to disease, susceptibility to disease if exposed, and treatment of disease. Public offi cials should systematically consider the additional barriers faced by socially disadvantaged groups at each of these levels and then actively seek ways to address those barriers. Local service providers, leaders of community-based organizations and other organizations working with socially vulnerable groups, and leaders of labor unions representing low-wage service workers are likely to have valuable insights and should be included in the planning process. Plans calling for stakeholder involvement without explicitly emphasizing the need to involve representatives of socially disadvantaged groups are unlikely to be effective at minimizing social disparities during an infl uenza pandemic. We have focused here on the United States, but similar fundamental principles-the need for systematic and concrete planning to minimize the social disparities that can be expected to occur in the face of natural disasters such as an infl uenza pandemic-apply worldwide. Countries with universal fi nancial access to healthcare and strong social safety nets will be best positioned to minimize such disparities. Countries in which large proportions of the population are impoverished or otherwise socially excluded and countries that have more limited resources and weaker public health and social welfare infrastructures will face the greatest challenges. The framework used here-considering and proactively addressing social vulnerability in exposure to pathogens, susceptibility to disease once exposed, and consequences of illness-should be applicable across national and subnational settings.

Global Distribution of Novel Rhinovirus Genotype

Global surveillance for a novel rhinovirus genotype indicated its association with community outbreaks and pediatric respiratory disease in Africa, Asia, Australia, Europe, and North America. Molecular dating indicates that these viruses have been circulating for at least 250 years.

A cute respiratory illness (ARI) is the most frequent infectious disease of humans. Ordinary upper respiratory tract infections are usually self-limited; nevertheless, they result in major economic impact through loss of productivity and strain on healthcare systems. Lower respiratory tract infections (LRTIs) are among the leading causes of death in children <5 years of age worldwide, particularly in resource-poor regions (1) . Streptococcus pneumoniae and Haemophilus infl uenzae are important bacterial causes of ARI, although their impact is expected to decline with increasing vaccine coverage. Collectively, however, viruses dominate as causative agents in ARI. Viruses frequently implicated in ARI include infl uenza virus, respiratory syncytial virus, metapneumovirus, parainfl uenza virus, human enterovirus (HEV), and human rhinovirus (HRV). HRVs are grouped taxonomically into Human rhinovirus A (HRV-A) and Human rhinovirus B (HRV-B), 2 species within the family Picornaviridae (International Committee on Taxonomy of Viruses database [ICTVdb]; http:// phene.cpmc.columbia.edu). These nonenveloped, positivesense, single-stranded RNA viruses have been classifi ed serologically and on the basis of antiviral susceptibility profi le, nucleotide sequence relatedness, and receptor usage (2) . Phylogenetic analyses of viral protein VP4/VP2 and VP1 coding regions indicate the presence of 74 serotypes in genetic group A and 25 serotypes in genetic group B (2) . Isolated in the 1950s from persons with upper respiratory tract symptoms (2, 3) , HRVs have become known as the common cold virus because they are implicated in ≈50% of upper respiratory tract infections (4) . Large community surveys, including the Virus Watch studies of the 1960-1970s (5), have shed light on some aspects of HRV biology and epidemiology. HRVs were also observed in LRTIs soon after their recognition (3) , and data supporting a causative association have accumulated over the past decade (6, 7) . HRVs have also been implicated in exacerbations of asthma and chronic bronchitis and are increasingly reported in LRTIs of infants, elderly persons, and immunocompromised patients (4). The advent of broad-range molecular assays, including multiplex PCR and microarray systems, promises new insights into the epidemiology and pathogenesis of respiratory disease (8, 9) , given that a laboratory diagnosis is not routinely achieved for a substantial portion of respiratory specimens from symptomatic patients. We recently described the application of a multiplex PCR method for microbial surveillance wherein primers are attached to tags of varying mass that serve as digital signatures for their genetic targets. Tags are cleaved from primers and recorded by mass spectroscopy, enabling a sensitive, inexpensive, and highly multiplexed microbial detection. We used the multiplex MassTag PCR system (10) to investigate respiratory samples that had tested negative during routine diagnostic assessment. This previous study yielded pathogen candidates in approximately one third of cases, and in 8 cases identifi ed a novel genetic clade of picornaviruses divergent from the previously characterized clades, including HRV-A and-B (8) . To assess whether this novel clade cir- Figure) . Samples collected in Côte d'Ivoire, West Africa, were from symptomatic persons living in the vicinity of Taϊ National Park. This location was the most remote of our study; residents have limited contact with other human populations. In this location, 2 (10%) HRV-A were identifi ed in the 52 samples available for analysis (Table 1, Figure) . In Nepal, viruses of the novel genotype were identifi ed in specimens collected during ILI surveillance or outbreaks of respiratory disease. Samples from ILI surveillance activities were collected in Kathmandu and Bharatpur. Outbreak samples were collected in the summer months from camps of >100,000 refugees from Bhutan located in Jhapa, southeast Nepal. Samples represented all age groups and were collected from December 2005 through July 2006. The novel genotype was identifi ed by independent molecular typing in both laboratories in 4 (5%) samples ( Figure) . Additional sample sets were obtained through main diagnostic laboratories in Western Australia, Denmark, and Spain, representing random respiratory specimens submitted for laboratory analysis. In 1 sample available from Western Australia, the novel genotype was identifi ed in a preterm infant with undiagnosed, wheezy LRTI. The novel genotype was also found in 5 (7%) of 70 samples from Denmark and in 6 (43%) of 14 samples with previously diagnosed HRV infection from Spain (Table 1, Figure) . The 5% overall frequency of the novel genotype across our study samples, representing 34% of all detected picornavirus infections, and its observed global distribution, led us to analyze the accumulating sequence data for insights into their history. Rates of evolutionary change and the Time to the Most Recent Common Ancestor (TMRCA) of the novel clade were estimated by using the Bayesian Markov Chain Monte Carlo approach (BEAST package [14] ; ), applying a relaxed molecular clock with an uncorrelated lognormal distribution of rates, a GTR + I + Γ 4 model of nucleotide substitution (determined by MODELTEST [15] ), and exponential population growth. Statistical uncertainty in each parameter estimate is expressed as 95% highest probability density (HPD) values. The estimated mean rate of evolutionary change was 6.6 × 10 -4 substitutions/site/y (95% HPD = 0.3-14.6 × 10 -4 substitutions/ site/y; 38 dated samples collected over 32 mo (8,16) (S.R. Dominguez et al., unpub. data). Under this rate the mean TMRCA was estimated at 1,800 y, although with wide variance caused by the short sequence available (95% HPD = 279-5,201 y). Despite the inherent sampling error, this analysis suggests that this third clade of rhinovirus has been circulating for >250 years. The diversity observed within the novel clade and its genetic distance from other HRV/ HEV were comparable to those seen for HRV-A, -B, or the HEV species (Table 2) . A clade of picornaviruses recently discovered in New York State is globally distributed and is found in association with community outbreaks of ARI and severe LRTIs of infants. These viruses contribute both to a substantial proportion of previously undiagnosed respiratory illness and to diagnosed, but nontyped cases of HRV infection. Similar viruses were recently characterized also in Queensland, Australia (11); California, USA (12); Hong Kong Special Administrative Region, People's Republic of China (13) ; and Germany (16) . Our fi ndings indicate the need for further investigation into this third (HRV-C) group of rhinoviruses with emphasis on epidemiology, pathogenesis, and strategies to prevent and ameliorate disease caused by HRV infection.

Cost effective strategies for completing the Interactome

Comprehensive protein interaction mapping projects are underway for many model species and humans. A key step in these projects is estimating the time, cost, and personnel required for obtaining an accurate and complete map. Here, we model the cost of interaction map completion across a spectrum of experimental designs. We show that current efforts may require up to 20 independent tests covering each protein pair to approach completion. We explore designs for reducing this cost substantially, including prioritization of protein pairs, probability thresholding, and interaction prediction. The best designs lower cost by four-fold overall and >100-fold in early stages of mapping. We demonstrate the best strategy in an ongoing project in Drosophila, in which we map 450 high-confidence interactions using 47 microtiter plates, versus thousands of plates expected using current designs. This study provides a framework for assessing the feasibility of interaction mapping projects and for future efforts to increase their efficiency.

Mapping a complete gene or protein network evokes similar challenges and considerations as mapping a complete genome sequence. In the case of the human and model genome projects, large-scale sequencing efforts were accompanied by a series of feasibility studies27,28 which used mathematical formulations and pilot projects to explore strategies for genome assembly and the work required for each. In the case of interaction networks, similar feasibility studies are just beginning29-31. Some of the questions to be addressed are: What is the cost of completing an interactome map and what is the best strategy for minimizing that cost? Given practical cost constraints, how can the quality and coverage of interaction data be maximized? How many independent assay types are needed? How should direct pairwise tests for interaction be combined with pooled screening? What is the effect of the test sensitivity on the final cost? How should interaction predictions be incorporated, and what is their effect on the mapping cost? Which specific improvements to experimental and computational methods are likely to have the largest effect? To approach these questions, we modeled several standard and alternative strategies for using pairwise protein interaction experiments to determine the interactome of the fruit fly Drosophila melanogaster. Our analysis shows that completing the interactome using sequential pairwise or pooled screening is probably too costly to be practical. However, this cost can be reduced substantially using a strategy that combines pooling with prioritized testing and interaction prediction. We carry out several iterations of this strategy to efficiently map 450 new high confidence interactions in Drosophila. In contrast to a genome, the interactome has been more difficult to define. Some authors have argued32 that the "True Interactome" should be defined as all possible interactions encoded by a genome-i.e., the set of all pairwise protein interactions that occur in at least one biological condition or cell type. The assumption is that every true interaction will be detectable by some assay, and that given enough independent measurements, most of the interactome could be detected. Many assay types have been described for detecting proteinprotein interactions, a few of which have been adapted to large-scale screening1,32-34. On the other hand, some interactions may be immeasurable by any available assay, or will not arise in the conditions surveyed. Therefore, we use the term "Mappable Interactome" for the subset of true pairwise interactions that are reproducibly detectable by any given assay method or combination of methods. To define appropriate criteria for determining when an interactome map is "complete", we distinguish between the terms saturation and coverage. Saturation measures the percentage of true interactions that have been experimentally observed at least once. Coverage we define more strictly to mean the percentage of true interactions that have been experimentally validated with high confidence such that the percentage of false interactions (i.e., the False Discovery Rate or FDR) is kept below a predetermined threshold. We treat "completion" as achieving 95% coverage of the Mappable Interactome at 5% FDR, which requires that the map include at least 95% of all true interactions with no more than 5% of the reported interactions being false. We simulated a series of mapping strategies implementing a variety of basic and sophisticated features ( Fig. 1 ; Flowcharts of each strategy are provided in Supplementary Fig. 1 ). All strategies were evaluated using a statistical model based on naïve Bayes to estimate saturation and coverage of the Drosophila interactome as a function of the number of interaction tests. We programmed this model with the assumption that the fly interactome contains approximately 105 interactions overall, along with estimates for the false positive rate (FPR-the probability that a non-interacting protein pair is reported as interacting) and the false negative rate (FNR-the probability that an interacting pair is reported as noninteracting). Although the magnitudes of these errors are still under debate, previous studies2,5,29,35,36 have reported Y2H error rates of FPR < 1% with FNR in the range 50-80% (note that several of these studies erroneously refer to FDR as FPR). Here, we used 0.2% FPR and 66% FNR. Due to the high FNR of a particular assay, it becomes clear that multiple assay types will likely be needed to achieve complete coverage, and that these assays should be independent or at least only partially dependent. Although the precise correlations between different assay types have not been well studied, complementarity between assays has been widely assumed and occasionally demonstrated: For instance, protein interactions have been shown to be of substantially higher confidence if they are detected in different orientations (baitprey vs. prey-bait)2; in different Y2H screens3,8,35; by different types of Y2H system such as LexA-based vs. Gal4-based36; or by both Y2H and co-affinity purification29. We first simulated a "Basic serial" strategy, in which all pairs of proteins are tested for interaction sequentially. Under this formulation, achieving a saturation of 95% required eight comprehensive screens, in which each protein pair was tested by eight independent assays, equivalent to ~7×108 pairwise tests assuming a total of 13,600 protein-encoding genes in fly37 (Table 1 and Fig. 2a) . Moreover, 93% of all observed interactions in this network were false positives (including 99% of interactions observed exactly once and 21% of interactions observed twice- Fig. 2b ). The false-positives predominate because, although the 0.2% FPR seems low, the number of non-interacting protein pairs is far in excess of the number of true interactions. To ensure an overall FDR < 5%, we found that every interaction must be reported by at least three independent assays. After eight screens 55% of the interactome was covered under these conditions. The coverage goal of 95% was achieved only after 21 comprehensive pairwise screens (Fig. 2c) . This overall outcome-that the number of experiments required to reach full coverage is two to three times that required to reach saturation-was observed over a range of error parameters (Supplementary Table 1 ). Clearly, completing the interactome map under these conditions is impractical, as it would require testing 92 million protein pairs 21 separate times. To reduce the number of tests, assays such as Y2H typically use pooled screens in which a single protein "bait" is tested for interaction against pools of protein "preys" (phase I) 38 . For pools that test positive, pairwise tests are immediately conducted between the bait and each prey in the pool (phase II-this second phase can also be conducted by sequencing3,5). The benefit of pooling is that large numbers of potential interactions can be sampled at relatively low cost. This comes at the expense of FNR, as the chance a true interaction is missed in the pool is higher than the chance it would be missed by direct pairwise tests38. Through simulation, we found that basic two-phase pooling (Pooling strategy) does indeed achieve a four-to five-fold reduction in coverage cost over pairwise testing (~4 million plates for Pooling compared to ~20 million plates for Basic-serial, Table 1 ). However, assuming the rate of interaction screening of a typical laboratory (e.g., ~2400 plate-matings per person per year), pooled screens would still require approximately 1700 person-years to achieve completion of the Drosophila protein network. We next considered an interaction mapping strategy that, rather than treat all protein pairs equally, maintains a rank-ordered list of pairs according to their probabilities of interaction (Thresholding strategy, Table 1 ). All probabilities start at the background frequency of interaction for random protein pairs (as for Basic-serial and Pooling). Protein interactions are initially tested using pooled screening, and after each two-phase pooled experiment the probabilities increase for interactions that are observed and decrease for interactions that are tested but not observed. Unlike previous strategies, however, protein pairs with probability greater than an upper threshold (i.e., 95%) are declared to be definite "interactors" and are removed from subsequent testing (Fig. 1b) . Likewise, interactions with probability beneath a lower threshold are declared to be "non-interactors" and are also removed from further consideration. The motivation for thresholding is to more quickly exclude the overwhelming number of non-interacting protein pairs. Finally, candidate interactions are defined as those with probabilities between the upper threshold and background. When candidates are available they are always tested immediately using pairwise assays, before resorting to pooling, until their probabilities are pushed above the upper threshold or below background. The motivation for prioritizing candidate interactions is to more quickly cover the interactions likely to be positive. Overall, Thresholding resulted in more than a two-fold cost reduction compared to Pooling (Table 1 and Fig. 3a ). Lastly, we considered whether computational prediction of interactions, based on prior knowledge and data, could hasten the time to interactome completion. A variety of prediction methods have been proposed based on evolutionary conservation39-41-i.e., transfer of interaction measurements from one species to another-or based on integrating conservation with additional features such as co-expression and co-annotation42-47. Such predictions impact the experimental design by setting the prior probabilities of interaction for each protein pair in lieu of background probabilities. In the Prediction strategy, we explored the effect of setting these prior probabilities using theoretical prediction methods simulated over a range of predetermined prediction accuracies (a range of different values for FPR, FNR, and corresponding FDR of the predictions). We found that even predictors with very high FDRs could have a major impact on the mapping cost (Table 1 and Fig. 3b) . For example, a predictor with 92.2% FDR gave a four-fold reduction in cost over Pooling, with a >50-fold reduction in cost to achieve 50% coverage and a savings of hundreds of fold in the early stages of mapping. Moreover, the 92.2% FDR means that even a predictor that makes 12 false predictions for every true one can lead to a major reduction in the cost of interactome completion. The best prediction method required approximately 385 personyears to achieve 95% coverage of the Drosophila protein network and 12 person-years to achieve 50% coverage. Thus, while obtaining full coverage of an interactome map may still be some years away, a draft scaffold providing half coverage might be feasibly achieved by a team of ~12 technicians working over a period of one year. Given the high performance of the Prediction strategy in simulations, we explored an experimental implementation in which Drosophila protein interactions were predicted using the cross-species method of Sharan et al.39 (Fig. 4a ). According to this method, existing protein interaction networks in yeast, worm, and fly are aligned based on sequence similarity to identify conserved interaction clusters, and these alignments are used to transfer interactions that have been observed in some species but not yet in others (Fig. 4b) . A total of 1,294 interactions were predicted using this method, each of which was prioritized as a candidate with high prior probability (92.4%) based on the FDR reported by Sharan et al.39 (7.6%). Since this prior was much greater than the background probability of other protein pairs (0.1%), we began by using the pairwise Lex-A Y2H assay48 to test all 606 predictions for which sequence-verified clones were available. Of these, 136 tested positive and 470 negative. After each 96-well plate of tests (seven plates total), the interaction probabilities were updated resulting in an increase to >99.9% for pairs testing positive and a decrease to 90.5% for pairs testing negative. Since the 136 positives now had probability greater than the upper threshold (95%), all of these could be added to the interactome map and removed from further testing. Although the remaining 470 predictions had tested negative once, their high probability (90.5%) still prioritized them as candidate interactions. Therefore, as dictated by the Prediction strategy these pairs were tested again immediately using a second assay type. For this second assay, Lex-A Y2H was run in a "reverse" orientation in which the two proteins cloned as bait and prey, respectively, were exchanged as prey and bait. We tested 251 of the 470 predictions for which sequence-verified clones were available in the opposite orientation. This resulted in 35 positives, elevating these interactions to probability >99.9% and adding them to the map. The pairs that tested negative in the reverse orientation were downgraded to 88.1% probability. Overall, after performing Y2H in both forward and reverse orientations, 171 new interactions were identified out of 606 protein pairs for a success rate of 28%. Although we ended our experimentation at this point, the Prediction strategy could be continued by next testing the "double negatives" (pairs testing negative in both orientations of Lex-A Y2H) using a third type of assay such as Gal4-based Y2H. A means of predicting additional protein interactions is to probabilistically integrate many different lines of evidence into a single classifier42-47. Along these lines, we applied a machine-learning-based classifier for predicting interactions that combined many relevant features including gene expression, domain-domain interactions, conserved protein-protein interactions, genetic interactions, and shared gene annotations (Supplementary Methods). We used this approach to generate 24,798 high confidence predictions. We randomly selected 2,047 of these for testing using forward-orientation Y2H and, as above, retested the negative pairs using reverse-orientation Y2H (for which clones were available). In total, this procedure added 279 new high-confidence interactions to the map for a 13.6% success rate. Combined over both conservation-based and multiple-evidence-based predictions, 450 new protein-protein interactions were added to the Drosophila map using 47 96-well plates (Fig. 3a,b) . To establish the background rate of interaction, we also tested 2,354 randomly chosen pairs, 72 of which were positive yielding a 3% background rate (Fig. 4b) . These results show that both types of prediction are highly enriched for true interactions. Note that even if all predicted interactions were true, the expected confirmation rate would be limited by the false negative rate of the Y2H assay, equal to 1-FNR =33% in our model. An underlying assumption of our simulations is that different assay types are conditionally independent-i.e., given that a tested protein pair is known to be positive or negative, the result of one assay is uncorrelated with that of another. To examine the extent to which this assumption holds, we compared Y2H data for protein pairs tested in both forward and reverse orientations-the two assay types used in our study. Overall, we obtained Y2H tests in both orientations for 309 conservation-based predictions (including data reported above combined with additional tests; Supplementary Data). Of these, we observed 58 positives in the forward orientation and 50 positives in the reverse orientation, for an average positive rate of 17% [(58 + 50)/(309 * 2)]. Fifteen positives were found in both orientations, representing 4.9% of the tests. Assuming all predictions are true interactions, this percentage is very close to that predicted by conditional independence, for which 3.1% of tests are expected to be positive in both orientations [17% ^ 2]. If some predictions are not true as expected, the percentages come into even better agreement-e.g., a prediction FDR of 20% predicts that 4.8% positives would arise in both orientations. A similar analysis was performed on a set of 1,572 combined-evidence predictions that were tested in both orientations, leading to similar agreement with the conditional independence assumption. The interactions predicted by cross-species conservation were at least as accurate as we had assumed in our simulations. On the other hand, their power to prioritize interactions is dependent on the network coverage in other species, and the long-term viability of this approach will depend on obtaining greater numbers of predictions than the 1,294 that are currently available. As interactome maps progress across an ever-widening array of species, these maps might be dynamically cross-compared to continually generate sufficient numbers of candidate interactions for testing. The second set of predictions, made by integrating various lines of evidence, had a lower success rate than the predictions based solely on conservation. Their potential utility is higher, however, since the number of available predictions is nearly 20 times that of the conservation-based predictions and could be increased further by including lower confidence predictions. Even with a lower success rate, the performance of the integrated classifier was superior to the best theoretical predictor we simulated. Predictions lead to a lower interactome mapping cost for two reasons. First, predicted protein pairs are much more likely than arbitrary pairs to be true. Second, protein pairs with high prior probabilities do not require repeated positive measurements to confirm them as true interactions. Both effects underlie the finding that 450 new predicted interactions could be added to the interaction map using just 47 microtiter plates. In contrast, the Pooling strategy would require nearly 105 plates to add this number of interactions to the map. One might intuitively object that, rather than test predicted interactions, a better strategy would focus on the "novel" areas of the interactome that have never before been suggested by any species or data set. The problem with such an approach is that it would very quickly produce an interactome map with a very high error rate. Conversely, the rationale behind the Thresholding and Prediction strategies is that one should first clean up the map by validating predicted interactions using real experiments, and only then resort to testing random protein pairs in pools. A second objection might be that prioritizing candidate interactions requires the corresponding Y2H baits and preys to be rearrayed in microtiter plates in different orders over the course of an interaction mapping project. While the cost of rearraying was not included in our analysis, in our lab (Finley) these costs are greatly alleviated through robotic transfer systems. Certainly, failure to rearray leads to a ~4-fold increase in cost and a ~10fold increase in the early stages of mapping (compare Pooling versus Prediction in Table 1 ). Regardless, mapping the Interactome remains a daunting task. Our study makes it clear that achieving 95% coverage of an interactome requires many more screens than one pass through all pools or over all protein pairs. If complete coverage is to be obtained in the near future, it will be necessary to invoke better strategies for experimental design, technologies reporting fewer false negatives, or both. In terms of experimental design, we have shown that the cost of completion is reduced substantially by careful ordering of pooled screens. In terms of technology, our study underscores the importance of decreasing the FNR or of different assays that provide independent samples of a protein pair. Even if the error rates are lower than assumed here, advanced mapping strategies are still likely to be worthwhile (Suppl. Table 1 ). Here we have used two types of Y2H assay, forward and reverse orientations, to obtain multiple samples which appear largely independent. If the assays were partially dependent, multiple tests might still be worth the cost as long as they were not perfectly correlated (and the dependence could be handled quantitatively using a statistical model). In the present study, the conditional independence assumption leads to a "best-case scenario" or lower-bound on the number of interaction tests that will likely be required to achieve full coverage of an interactome. Further work will be needed to better characterize the relative dependencies among the wide range of other interaction assays that are currently available-if the current assays are highly dependent, then the required number of tests will be greater than was estimated here. "True" reference interactomes for fly and human were generated by random sampling of interactions from the set of all possible pairs of proteins using the interaction probabilities in the String database46. Protein pairs not included in the String database were sampled using a low background probability, such that the total number of interactions in the sampled interactomes agreed with current estimates of interactome sizes30 (~100,000 fly interactions and ~260,000 human interactions). The detectability of each protein pair was independently sampled for each new assay type (representing a new type of measurement technology or new bait/prey orientation) using a 66% FNR for true interactions and 0.2% FPR for false interactions (corresponding to 82% FDR). Once an interaction was defined as "Detectable/ Undetectable", direct pairwise experiments were assumed to be 100% reproducible for a given protein pair and assay. For pooled assays, each detectable interaction in the sample space of a pool was assumed to be observed in the pool with probability equal to the pooling sensitivity (Table 1) . Pools with at least one observed interaction were declared positive. For each strategy, after every 1000 experiments the mapped interactomes were compared to the "true" interactomes and the coverage and FDR were recorded. We used the LexA-based yeast two-hybrid mating assay48 using sequence-verified clones as previously described36 (Supplementary Methods). All new protein interactions have been submitted to the IMEx consortium (http://imex.sf.net) through IntAct49 and assigned the identifier IM-9552. The data are also available at DroID (www.droidb.org). Detailed descriptions of the interaction probability model, the combined-evidence method for interaction prediction, the computation of thresholds, and the yeast two-hybrid test protocol appear in the Supplementary Methods. Refer to Web version on PubMed Central for supplementary material. (a) At any given point in the project, every pair of proteins is assigned an interaction probability based on its experimental history (initially these probabilities are set to background or informed by predictions). The interaction probabilities and experimental history are used to design a 96-well plate Y2H experiment according to one of the strategies. The result of this experiment is simulated based on the detectability of the tested interactions (given the assay type) and the pooling sensitivity. The new experimental results are recorded in the history and also (b) used to update the interaction probabilities of the relevant protein pairs. The pyramid represents the ordered list of protein pairs ranked by probability. It is wider at the bottom than at the top to reflect that most pairs are negative-i.e., most pairs will have low probability and only a few pairs will percolate to the top of the list with high probability. Interactions with probability above an upper threshold are added to the mapped interactome, which is compared to the simulated "True Network" at intervals of 1,000 plates for reporting coverage and FDR. The false discovery rate (FDR) of interactions that are observed exactly once (orange), twice (purple), thrice (green), four times (yellow), and five times (cyan) as a function of the number of times they are tested with independent assays. To achieve FDR < 5% interactions should be observed at least twice when tested with < 5 independent assays, and at least three times when tested with 5-17 assays. (c) The effective coverage level at FDR < 5% is shown (red curve) by embedding the observation threshold from (b) into the curves of (a). While saturation is achieved after 8 screens, 21 screens are required for 95% coverage at FDR < 5%. * Interaction costs are given in units of total number of plates (K = Thousands, M = Millions) required for 50% or 95% coverage. When 95% coverage is achieved more than once, the greatest cost is presented.

Augmented Lung Inflammation Protects against Influenza A Pneumonia

BACKGROUND: Influenza pneumonia causes high mortality every year, and pandemic episodes kill millions of people. Influenza-related mortality has been variously ascribed to an ineffective host response that fails to limit viral replication, an excessive host inflammatory response that results in lung injury and impairment of gas exchange, or to bacterial superinfection. We sought to determine whether lung inflammation promoted or impaired host survival in influenza pneumonia. METHODS AND FINDINGS: To distinguish among these possible causes of influenza-related death, we induced robust lung inflammation by exposing mice to an aerosolized bacterial lysate prior to challenge with live virus. The treatment induced expression of the inflammatory cytokines IL-6 and TNF in bronchoalveolar lavage fluid 8- and 40-fold greater, respectively, than that caused by lethal influenza infection. Yet, this augmented inflammation was associated with striking resistance to host mortality (0% vs 90% survival, p = 0.0001) and reduced viral titers (p = 0.004). Bacterial superinfection of virus infected lungs was not observed. When mice were repeatedly exposed to the bacterial lysate, as would be clinically desirable during an influenza epidemic, there was no tachyphylaxis of the induced viral resistance. When the bacterial lysate was administered after the viral challenge, there was still some mortality benefit, and when ribavirin was added to the aerosolized bacterial lysate, host survival was synergistically improved (0% vs 93.3% survival, p<0.0001). CONCLUSIONS: Together, these data indicate that innate immune resistance to influenza can be effectively stimulated, and suggest that ineffective rather than excessive inflammation is the major cause of mortality in influenza pneumonia.

The annual worldwide mortality associated with pneumonia exceeds that of any other infection [1, 2, 3] . In particular, influenza pneumonia annually causes more than 40,000 deaths in the United States alone [4, 5] . Beyond the impact of seasonal influenza, episodes of pandemic influenza have accounted for as many as 50 million deaths [6] . H5N1 avian influenza has already caused more than 240 human deaths worldwide (http://www. who.int/csr/disease/avian_influenza/), and increased globalization since 1918 suggests that eventual human-to-human transmission of avian influenza may cause even greater lethality than the infamous ''Spanish flu'' [7] . Further, viral pathogens, including influenza, are considered potential agents of bioterror [8] . The mechanisms underlying influenza-related mortality remain controversial. Progressive pneumonia following an insufficient antiviral host response is one possible cause [9] . Virally-induced excessive and/or dysregulated lung inflammation is another potential mechanism [10, 11, 12, 13] . Secondary bacterial infections have also been proposed as the primary contributors to influenzarelated mortality, due to virally-injured epithelium or virusattenuated leukocyte responses [13, 14, 15] . We have recently reported that treatment with an aerosolized lysate of nontypeable Haemophilus influenzae (NTHi) induces profound inflammation in the lungs, yet it strongly protects mice against otherwise fatal bacterial pneumonia [16] . The induced protective phenomenon, known as Stimulated Innate Resistance (StIR), is maximal within 4 h of treatment and does not rely on recruited neutrophils or resident mast cells and alveolar macrophages. Given the profound induction of lung inflammation by this treatment, we perceived an opportunity to determine whether host inflammation contributed to or prevented mortality in influenza A pneumonia. We demonstrate that, despite inducing inflammation greater than that observed in lethally infected animals, the aerosolized NTHi lysate results in remarkable protection against otherwise lethal influenza pneumonia, and that it can be synergistically combined with antiviral medicine for post-infectious treatment. Together, these data suggest that ineffective rather than excessive inflammation is the major cause of mortality in influenza pneumonia, and indicate that innate immune resistance to viruses can be therapeutically stimulated to protect populations from pandemic influenza. Having shown that profound inflammatory lung responses to a bacterial lysate improved survival of mice infected with noncognate bacteria [16] , we investigated the effect of stimulation of lung mucosal innate immunity on survival of influenza A pneumonia. Mice challenged by aerosol with influenza A/Hong Kong/8/68 (H3N2) (A/HK) universally succumbed to hemorrhagic pneumonia unless pretreated with aerosolized NTHi lysate (Fig. 1A) . This treatment reduced mortality .90% if delivered on the day prior to infection, .50% if delivered 3 days prior to infection, and to a lesser degree if delivered one day after infection (Fig. 1B) . Mortality occurred within 10 days of the viral inoculation, and observation for 21 days after infection showed no subsequent mortality. The protection imparted by the aerosolized lysate was paralleled by the recovery of body weight lost early in the course of infection (Fig. 1C ). Lysate-induced survival was accompanied by a significant reduction in viral load on the fourth day after challenge (Fig. 1D ). Since prior reports have described poorer outcomes associated with virus-induced inflammation [10, 11] , we compared the inflammation induced by the NTHi lysate, inflammation induced by influenza infection, and inflammation induced by the combination of treatment and infection. Figure 2 shows that single or repeated treatment with NTHi lysate induces dramatic inflammation in the lungs, measured by cytokine levels ( Fig. 2A ) and cellular influx (Fig. 2D) . Remarkably, while treatment with NTHi lysate acutely induces 8-fold more IL-6 and almost 40-fold more TNF than infection alone, the treated mice actually have lower levels of inflammatory cytokines in their lungs by day 3 after the infection (Fig. 2B ). By that time after treatment, the NTHi lysate-induced rise in cytokines in uninfected mice has entirely resolved (data not shown), and the inflammatory cell infiltration has fallen by more than 80% [16] . Together, this suggests that a robust early inflammatory reaction may allow for more rapid resolution of the infection-induced inflammation. Notably, the intense inflammation observed in the lungs is not seen systemically. Despite increases in IL-6 and TNF of 700-and 900-fold in the lungs, respectively, these cytokines increase only minimally and transiently in the serum (Fig. 2C ). Since interferon signaling is essential to baseline host antiviral resistance in the lungs [17, 18] , we investigated whether treatment with NTHi lysate was capable of inducing an interferon response. In addition to the robust induction of inflammatory cytokines, we found that NTHi lysate treatment also induces significant increases in lung interferon c levels ( Fig. 2A) . However, unlike IL-6 and TNF, the fulminant influenza infections observed in untreated mice actually induce 4-fold higher levels of interferon c than the NTHi lysate treatment (Fig. 2B ). Presumably this reflects the ongoing antiviral response of the host, in contrast to the better contained infections of the treated mice. To further characterize the lung interferon response to the NTHi lysate, we assessed gene expression using whole genome microarray analysis. As maximal resistance against S. pneumoniae is achieved by 4 h post-treatment, we compared gene expression of sham treated mouse lungs to mouse lungs 2 and 4 h posttreatment. Pathway analysis revealed interferon signaling to be among the most upregulated events following NTHi lysate treatment (p,10 211 ). Interferon signaling pathway transcripts are reported in Table 1 , showing that treatment induces expression of numerous transcripts critical for both type I (interferon-a/b) and type II (interferon-c) signaling. Since a high level of protection against influenza virus only lasts for 3 days (Fig. 1) , prolonged protection would require repetitive dosing. To test whether tachyphylaxis to the protective effects of the aerosolized NTHi lysate occurs, we challenged mice with influenza in three different conditions: no lysate treatment, a single lysate treatment, or repetitive lysate treatment. We again observed 100% mortality in the untreated group, but found identical protection by NTHi lysate treatment whether given once one day prior to infection, or given three times (seven, four and one days prior to infection, Fig. 3A ). In both NTHi lysate treated groups, protection was associated with recovery of early weight loss by the fifth day after infection (Fig. 3B ). Observation for 21 days after the challenge showed no subsequent mortality. An alternative to repetitive stimulation of lung innate immunity for prophylaxis during an epidemic viral infection would be to treat with the pro-inflammatory aerosol after infection in combination with an antiviral drug. We recently found that treatment with high-dose aerosolized ribavirin after exposure to influenza virus provided some survival improvement in mice [19] . To test whether these two interventions might be effectively combined, mice were infected with influenza virus then treated with ribavirin alone, NTHi lysate alone, or a combination of the two. Ribavirin alone did not improve survival with the high level infectious challenge used in this study. However, as shown in Figure 4 , suspending ribavirin in a single dose of NTHi lysate on day 1 after infection improved survival significantly more than NTHi lysate alone (66.7% vs 20.0% survival, p = 0.013), and an additional NTHi lysate treatment on day 2 almost completely protected from mortality (93.3% survival, p,0.0001 vs. control). No untoward effects or additional protection were noted if a third NTHi lysate dose was added on day 3. Observations through day 21 showed no subsequent mortality. Lower respiratory tract infections are the leading cause of infectious death worldwide [1] . We have recently shown that Stimulated Innate Resistance (StIR) of the lungs induced by an aerosolized bacterial lysate can protect mice against otherwise fatal bacterial pneumonias [16] . We show here that the same treatment can prevent viral pneumonia, despite induction of lung inflam- mation of greater magnitude than that observed with lethal infection. Influenza-induced excessive and/or dysregulated lung inflammation has been recently described as a mechanism by which pandemic infections cause mortality [10, 11, 12, 13] . Consequent to these observations, considerable effort has been invested in attempts to improve outcomes of influenza-infected mice through genetic and pharmacologic suppression of inflammatory cytokine production [12, 20, 21, 22] . In contrast, our current and previous data demonstrate robust inflammation associated with the induction of protective host responses (Fig. 3a-d) . As we have previously shown that repetitive treatments result in diminishing inflammatory responses without loss of protection, it is possible that the cytokines themselves are not responsible for the protective effect. However, like our previous observations with protection against bacterial pathogens [16] , the pro-inflammatory pretreatment resulted in a significant reduction in pathogen burden, here represented by decreased viral titers (Fig. 1d) . As interferon signaling is associated with survival of viral and bacterial infections, the observation of augmented type I and II interferon signaling after NTHi lysate treatment ( Figure 3A -B, Table 1 ) supports our hypothesis that this is also an important element of anti-viral StIR, and we will explore this in the future. One manner in which the host response to NTHi lysate differs from reported pandemic virus-induced inflammation [11] is in its restriction to the lung. While documenting massive increases in indicators of lung inflammation following treatment, there is virtually no systemic inflammation noted (Fig 3c) . Similarly, treatment with NTHi lysate induces no systemic leukocytosis (data not shown), despite the profound cellular infiltration induced in the lungs. We presume this confinement of the response to the lungs explains the lack of morbidity we have observed when mice are treated acutely even with very high doses of NTHi lysate [16] . Based on these observations, we conclude that influenza pneumonia does not kill via excessive pulmonary inflammation, but progresses through a deficit of effective inflammation. In addition to exuberant lung inflammation, secondary bacterial infections are often considered important contributors to influenza-related mortality [13, 14, 15] . Interestingly, we here demonstrate that the host response to bacterial products actually enhances the clearance of a viral pathogen (Fig. 1D) , just as we have shown for bacteria [16] . So, while viral infections may impair antibacterial interferon-c responses [15] , we observe no evidence that the response to the bacterial lysate impedes either type I or type II interferon signaling. Rather, the response to the bacterial products seems to reinforce protective antiviral events via vigorous interferon gene expression. Epidemic respiratory infections, as previously observed with influenza [10, 11] and SARS [23] , result in high mortality, sometimes before the pathogen is identified, and often without effective post-exposure treatment options. Seasonal influenza, though the case-mortality rate is lower, still kills ten of thousands of Americans annually [4, 5] and clinicians are faced with the problems of ineffective vaccine strategies and declining effectiveness of neuraminidase inhibitors [24, 25] . In such situations, rapidly dispersed, broad protection would be highly advantageous. At present, the primary means to contain the spread of influenza and to prevent exposed individuals from developing disease is through the use of the trivalent hemagglutinin subunit vaccine or the live-attenuated trivalent vaccine coordinated by the United Stated Centers for Disease Control and Prevention. This strategy is limited by the fact that prevalent (potentially pandemic) strains may not be accurately predicted for inclusion in the annual vaccine, that normal adaptive immunity is required for protection against included strains, and that development of immunity takes weeks -potentially obviating the benefit in the setting of a rapidly spreading outbreak. In this study, we showed that a single treatment with NTHi lysate was sufficient to prevent otherwise fatal influenza pneumonia (Fig. 1) . The aerosolized lysate was also administered repetitively to model the sustained protection that would be clinically desirable during an influenza epidemic. No tachyphylaxis to the influenza-protective effect was observed with successive administrations (Fig. 2) , despite the fact that the rise in lung lavage neutrophils induced by the aerosolized lysate becomes less with repeated administration [26] . These findings are consistent with the development of tolerance to the inflammatory effects of innate immune stimulation without loss of antimicrobial effector function, as recently described [27] . Alternatively, the progressive reduction in inflammation could reflect the tempering of innate immune responses by adaptive immune cells that become engaged with repetitive stimulation [28] . In addition to the preventive benefits of NTHi lysate-induced StIR, we show that stimulation of lung innate immunity can be therapeutically combined with an antiviral medication after influenza infection to improve survival more than either treatment alone (Fig. 4) . In our current and prior work, we have reported only a modest survival benefit when the lysate is delivered 24 h after infection. However, we found that continuing induction of an antimicrobial inflammatory lung environment for up to 72 h potentiates the effects of high dose ribavirin and almost completely averts an otherwise uniformly fatal outcome. While Zheng and colleagues [22] recently reported improved influenza survival from 0% to 53.3% with the anti-viral/anti-inflammatory combination of zanamivir, celecoxib and mesalazine, we believe our current study to be the first to demonstrate such dramatic protection (improved from 0% to 93.3% survival) with an anti-viral combination known to induce inflammation. The seeming contradiction of these results may best be resolved by the hypothesis that the anti-inflammatory regimen diminished some toxicity associated with uncontrolled immune responses (i.e., reduced harmful inflammation) while our pro-inflammatory combination promoted the robust generation of defense effectors (i.e., enhanced protective inflammation). In summary, we have shown that stimulation of lung mucosal innate immunity with a complex bacterial lysate confers striking protection against a virulent viral pathogen. Resistance to influenza is stimulated despite profound induction of lungrestricted inflammation, suggesting that excessive pulmonary inflammation is not the primary cause of influenza-induced mortality. Further, we found that the same treatment can be combined in the post-exposure setting with antiviral medication to improve survival. Taken together, we infer that therapeutic manipulation of StIR may be possible to reduce the mortality burden associated with viral pneumonias, and potentially, to protect the population against bioterror attacks. Six to eight week old NIH Swiss-Webster mice (Charles River) were used for viral challenges. For protection studies, mice were divided into groups of 20 mice (5 for virus lung titers, 15 for survival). For bronchoalveolar lavage assays, an additional 5 mice were added for each group. Six to eight week old C57BL/6 mice were used for gene expression analysis, 6 mice per condition. Mice were handled in accordance with protocols approved by the Institutional Animal Care and Use Committees of Baylor College of Medicine and The University of Texas-M.D. Anderson Cancer Center, and euthanized if distressed. Non-typeable Haemophilus influenzae (NTHi) was stored, grown and harvested as described [16, 26] . The cell pellet was washed and resuspended in 20 ml 0.9% sodium chloride solution. This suspension was passed three times through an EmulsiFlex C5 cell disruptor (Avestin) at greater then 10,000 psi, then diluted to 4-5 mg/ml in 0.9% sodium chloride solution by bicinchoninic assay (Pierce) and centrifuged at 15,0006g for 10 min. The supernatant was collected, the protein concentration was adjusted to 2.5 mg/ml, and the lysate was sterilized by passage through a 0.2 mm filter and frozen in 8 ml aliquots at 280uC. Treatment of mice with aerosolized lysate was performed as described [16] , delivering 8 mL of suspension during each 30 min exposure. A clinical isolate of influenza A/Hong Kong/8/68 (H3N2) (A/ HK; Mouse Lung Pool 11-29-05) virus that had been passaged at least nine times through mice was stored as frozen stock (2.8610 7 TCID 50 /ml) in the supernatant of mouse lung homogenates [29] . Stock was diluted 1:300-1:1,000 in 0.05% gelatin in Eagle's minimal essential medium (Sigma-Aldrich) and aerosolized for 20 min to achieve LD 90 -LD 100 (target 100 TDIC 50 /mouse). Viral concentration in the nebulizer before and after aerosolization and in lung homogenates was determined by hemagglutination assay of infected MDCK cells [30] . In some experiments, 1 g of ribavirin was dissolved in 10 ml of NTHi lysate prior to aerosolization. Final ribavirin and lysate protein concentrations were 100 mg/ml and 2.5 mg/ml, respectively. Mice were challenged without pretreatment, following pretreatment on day 21, or following pretreatment on days 27, 24 and 21. On day 0, all groups of mice were exposed to the same viral aerosol. On day +4, 5 mice from each group were sacrificed and their lungs removed. Lungs were homogenized by beadbeating and the levels of virus determined. Remaining mice in each group were observed daily for up to 21 days for overt illness, morbidity and mortality. Mice were weighed on days 0 and +4, and three times weekly from day +7 until day +21. To characterize the inflammatory host response to treatment and/or infection lung lavage and cell counts were performed as described [16] . In order to assess the response to NTHi lysate treatment, lavage samples from unchallenged mice were obtained 4 h after the final (or only) NTHi lysate treatment. In order to collect samples during acute virus-induced illness, lavage samples were also collected on day +3 following infection with influenza. Multiplex ELISA cytokine analysis was performed by Searchlight Protein Array Analysis (Pierce Biotechnology). Total leukocyte count was determined with a hemacytometer, and differential count by cytocentrifugation of 300 ml of bronchoalveolar lavage fluid at 4506g for 5 min, followed by Wright-Giemsa staining. To better understand the gene expression response to the treatment, mice were treated with the aerosolized NTHi lysate, then euthanized after 2 h or 4 h for comparison to untreated mice. To reduce the lung leukocyte burden, the pulmonary vasculature was perfused and the airways lavaged with PBS. The lungs were mechanically homogenized, then total RNA was isolated from lung homogenates using the RNeasy system (Qiagen), and cRNA was synthesized and amplified from equal masses of total RNA using the Ilumina TotalPrep RNA amplification kit (Ambion). Amplified cRNA was hybridized and labeled on Sentrix Mouse-6 Expression BeadChips (Illumina), then scanned on a BeadStation 500 (Illumina). Primary microarray data were deposited at the NCBI Gene Expression Omnibus (http://www.ncbi.nlm.nih.gov/ geo/) consistent with MIAME standards (GEO Accession GSE13740). Primary signal intensity was normalized between and within samples, and differentially expressed genes were identified based on signal change and inter-sample variation. Gene ontology analysis was performed using the NIAID Database for Annotation, Visualization and Integrated Discovery (DAVID) and the KEGG Database (GenomeNet). Differentially expressed genes were mapped to signaling pathways using Ingenuity Pathways Analysis 5.0 (Ingenuity Systems), and the pathway nodules were individually reviewed. C57BL/6 mice were used for the gene expression studies as they demonstrate similar NTHi lysate induced protection against all investigated pathogens, and because the genetically manipulated mice the authors plan to use to dissect the mechanisms of StIR are primarily on C57BL/6 backgrounds. To characterize the interferon-related gene expression changes induced by NTHi, Table 1 presents a list of genes containing all transcripts from the Ingenuity Pathway Analysis canonical interferon signaling pathway, all detected interferon-related JAK-STAT transcripts in KEGG, and additional interferon related transcripts identified by the authors. Baseline signal intensity values of 1 were assigned to undetected control transcripts in order to avoid reporting infinite fold change values. Summary statistics for virus in lung tissue were compared using Student's t-test. Proportions of mice surviving pathogen challenges were compared using Fisher's exact text, and log-rank comparisons of survival distribution were performed using Kaplan-Meier estimation. All data shown are representative of at least two independent experiments, and were not combined for analysis because of modest differences in virus challenge doses. Analyses were performed using SAS/STAT (SAS Institute). For gene expression analyses, treated and untreated samples were compared to identify treatment-induced gene expression using an ANOVA-based scheme written in R (Free Software Foundation, Boston, MA), utilizing an Illumina library developed by Simon Lin, Northwestern University, Chicago, IL.

A Novel Peptide Enhances Therapeutic Efficacy of Liposomal Anti-Cancer Drugs in Mice Models of Human Lung Cancer

Lung cancer is the leading cause of cancer-related mortality worldwide. The lack of tumor specificity remains a major drawback for effective chemotherapies and results in dose-limiting toxicities. However, a ligand-mediated drug delivery system should be able to render chemotherapy more specific to tumor cells and less toxic to normal tissues. In this study, we isolated a novel peptide ligand from a phage-displayed peptide library that bound to non-small cell lung cancer (NSCLC) cell lines. The targeting phage bound to several NSCLC cell lines but not to normal cells. Both the targeting phage and the synthetic peptide recognized the surgical specimens of NSCLC with a positive rate of 75% (27 of 36 specimens). In severe combined immunodeficiency (SCID) mice bearing NSCLC xenografts, the targeting phage specifically bound to tumor masses. The tumor homing ability of the targeting phage was inhibited by the cognate synthetic peptide, but not by a control or a WTY-mutated peptide. When the targeting peptide was coupled to liposomes carrying doxorubicin or vinorelbine, the therapeutic index of the chemotherapeutic agents and the survival rates of mice with human lung cancer xenografts markedly increased. Furthermore, the targeting liposomes increased drug accumulation in tumor tissues by 5.7-fold compared with free drugs and enhanced cancer cell apoptosis resulting from a higher concentration of bioavailable doxorubicin. The current study suggests that this tumor-specific peptide may be used to create chemotherapies specifically targeting tumor cells in the treatment of NSCLC and to design targeted gene transfer vectors or it may be used one in the diagnosis of this malignancy.

Lung cancer is one of the most commonly diagnosed malignancies in developed countries and is a growing problem in developing countries [1] . There are two major types of lung cancer: non-small cell lung cancer (NSCLC) and small cell lung cancer (SCLC). NSCLC makes up approximately 80% of all lung cancer cases [2] and has a limited response rate to current chemotherapeutic agents, with tumor shrinkage in only 20% of patients and a two-year survival rate between 10% and 16% [3] . One major reason for this unsatisfactory outcome of chemotherapy is compromised drug delivery to the lung cancer tissues due to high interstitial fluid pressures (IFP) within the tumor [4] . Systemically administered chemotherapy cannot be adequately delivered into solid tumors because of the immature vasculature with abnormal architecture [5] and leaky, heterogeneous vessel walls [6] as well as the high IFP within tumor tissues [7, 8] . Furthermore, a lack of tumor specificity allows anti-cancer drugs to distribute indiscriminately to normal organs and tissues. Thus, cancer cells are exposed to a lower concentration of the drug than normal cells [9] , resulting in not only decreased effectiveness but also increased toxicity. Therefore, it is important to develop a strategy to enhance the amount of drugs delivered to tumor tissues in a targeted way while sparing normal tissues. Efforts are ongoing to improve the therapeutic index of anticancer agents, either by increasing the drug concentration inside the tumor or by decreasing it in normal host tissues [10] . Compared with conventionally administered chemotherapeutic agents, lipid-or polymer-based nano-medicine drug delivery systems have the advantage of improving the pharmacological and therapeutic properties of cytotoxic drugs [11] . Most smallmolecule chemotherapeutic agents have a large volume of distribution on intravenous administration [12] and a narrow therapeutic window due to serious toxicity to normal tissues. By encapsulating drugs in nano-particles such as liposomes, scientists can significantly reduce the volume of distribution and increase the concentration of active drug within the tumor [13] . PEGylated liposomal doxorubicin (with brand names of Doxil in the US and Caelyx in Europe) [14] has been shown to significantly improve the therapeutic index of doxorubicin both in preclinical [15] [16] [17] and clinical studies [18] [19] [20] [21] . Several drug delivery systems of this kind have been approved for marketing [22, 23] . Other than PEGylated liposomes, higher and more selective anti-cancer activity can be achieved through ligand-mediated targeting liposomes. In this novel drug delivery system, targeting moieties are coupled to the surface of liposomes to promote selective binding to tumor-specific antigens and facilitate the delivery of drug-containing liposomes to the intended cellular sites. This system has the advantages of a higher drug-to-carrier ratio than immunoconjugates and the multivalent presentation of ligands leading to increased binding avidity [24] . Researchers have already produced liposomes conjugated with various peptide ligands that specifically target certain tumor cells or tumor vasculature [25] [26] [27] [28] [29] . Because of the favorable selectivity and specificity, ligandconjugated liposomal anticancer drugs are a promising approach for new chemotherapy research. The use of peptides as ligands to direct liposomes to tumors represents a potentially feasible method for increasing the specificity and effectiveness of liposomecontaining drugs [26, 30] . Phage display is a technique of selecting targeting peptides, in which a peptide is expressed on the surface of a bacteriophage as a fusion-protein with one of the virion's own coat proteins [31] . Phage-displayed peptide libraries allows researchers to map protein-protein contacts such as B-cell epitopes [32] [33] [34] [35] and receptor-ligand interactions [36] . Such peptide libraries can also be used to identify organ-and cell-type-specific peptides [26, 27, [37] [38] [39] . In this study, we used a phage-displayed peptide library to identify a novel peptide that bound specifically to NSCLC cell lines and surgical specimens from lung cancer patients. Liposomal doxorubicin and vinorelbine conjugated with this targeting peptide demonstrated enhanced accumulation of the drugs in tumor tissues and improved therapeutic index for human lung cancer xenografts in SCID mice. We used a phage-displayed random peptide library to isolate phages that were able to bind to NSCLC CL1-5 cells. After five rounds of affinity selection (biopanning) with CL1-5, cells increased the titer of phage by 40-fold (Fig. 1A) . Enriched phages from the third to the fifth biopanning rounds were randomly selected. We then sequenced the phage clones with higher CL1-5binding activities. Using the Genetics Computer Group (GCG) software analysis, we found that these selected phages (PC3-1, PC4-1, PC4-5, PC5-2 and PC5-4) displayed the consensus motif, tryptophan (W)-threonine (T)/tyrosine (Y)-tyrosine (Y) ( Table 1) . Interestingly, the phage PC5-2 appeared in the third (PC3-1), fourth (PC4-1) and fifth (PC5-2) biopanning rounds. During the biopanning rounds, the frequency of PC5-2 increased from 20% (1/5) in the third cycle to 90% (27/30) in the fifth cycle (Table 1) . We chose to focus on the novel peptide displayed by PC5-2, TDSILRSYDWTY, for further study. In vitro phage display screening for peptides that bind to NSCLC. (A) A phage-displayed random peptide library was used to select phages that bind to the NSCLC cell line CL1-5. (B) Visualization of PC5-2 binding to CL1-5 and PC13 lung cancer cells (arrowheads) with immunohistochemical staining. The control phage did not bind to CL1-5 cells. Scale bar: 10 mm. (C) The FITC-labeled peptide SP5-2 bound to five NSCLC cell lines but not to NPC-TW01 cells as detected by immunofluorescent staining. Scale bar: 10 mm. (D) Representative photomicrographs of tumor sections from surgical specimens of human lung cancer were detected using both PC5-2 (a, arrowhead) and biotinylated SP5-2 (c, arrowhead), respectively. In comparison, the control phage or biotinylated control peptide could not bind to these surgical specimens (b and d). PC5-2 was competitively inhibited by the synthetic peptide SP5-2 (e). Mutated peptide, MP5-2, lost this competition ability (f). Scale bar: 25 mm. doi:10.1371/journal.pone.0004171.g001 Identification of phage clones specifically binding to NSCLC cells To investigate whether PC5-2 would bind to NSCLC cells, we used immunohistochemistry to locate the phage particles in different cell types. Our results showed that PC5-2 bound specifically to NSCLC cell lines including CL1-5 and PC13 (Fig. 1B, arrowheads) . The control helper phage did not bind to CL1-5 cells. PC5-2 bound neither to other cancer cell lines, including oral cancer (SAS) and nasopharyngeal carcinoma cells (NPC-TW01), nor to normal epithelial cells (NNM) from nasal mucosa (Fig. 1B) . The CL1-5-binding ability of PC5-2 was further confirmed by a peptide competitive inhibition experiment using immunofluorescent staining. The results showed that the binding activity of the PC5-2 phage to CL1-5 cells was inhibited by the synthetic peptide SP5-2 in a dose-dependent manner. At a concentration of 27 mg/ml, SP5-2 completely inhibited the binding activity of PC5-2 (Fig. S1 ). The control phage did not bind to CL1-5 cells, and PC5-2 did not bind to NPC-TW01 in this assay ( Fig. S1 and Text S2). To further verify that the PC5-2 phage would bind to a target molecule expressed on the surface of CL1-5 cells, we measured PC5-2-bound cells using flow cytometry ( Fig. S2 and Text S1, S2). A control phage was used to estimate non-specific background binding (Fig. S2b) . The results showed that 42.6% of CL1-5 cells were bound by PC5-2 ( Fig. S2c) , and this binding was completely inhibited by 27 mg/ml of SP5-2 peptide (Fig. S2d ). PC5-2 did not bind to SAS or NNM ( To determine whether the peptide sequences displayed on PC5-2 would actually interact with NSCLC cells, we used fluorescein isothiocyanate (FITC)-labeled SP5-2 peptide (FITC-SP5-2) in place of the PC5-2 phage for a peptide-binding assay through immunofluorescent staining. FITC-SP5-2 specifically bound to all of the NSCLC cell lines we tested, including CL1-5, H460, A549, PC13 and H23, but did not bind to NNM. The same concentration of FITC-labeled control peptide (FITC-Con-P) revealed no such binding activity (Fig. 1C) . We also evaluated the magnitude and specificity of SP5-2 binding using flow cytometry. The proportions of CL1-5, H460, A549, PC13 and H23 cells bound by SP5-2 were 43.0%, 45.8%, 44.3%, 20.1% and 44.0%, respectively ( Fig. S3 and Text S2). To determine whether this targeting ligand had an affinity for human lung cancer surgical specimens, we tested the reactivity of PC5-2 and SP5-2 with pulmonary adenocarcinoma cells using immunohistochemistry. Both PC5-2 and biotin-labeled SP5-2 (B-SP5-2) recognized the tumor cells of NSCLC surgical specimens (Fig. 1D a and c, arrowheads) , and the control phage and biotinlabeled control peptide (B-Con-P) did not (Fig. 1D b and d) . SP5-2 (TDSILRSYDWTY) competed with PC5-2 for binding to surgical specimens of pulmonary adenocarcinoma (Fig. 1De ), but the same concentration of a mutated peptide, MP5-2 (TDSILRSYDGGG) did not (Fig. 1Df) . Seventy-five percent (27/36) of the pulmonary adenocarcinoma specimens from 36 patients expressed a target molecule that was recognized by this peptide (Table S1 ). These data indicated that SP5-2 could recognize unidentified molecules expressed on NSCLC cell lines and actual cancer cells from the surgical specimens of lung cancer. To investigate the targeting ability of the PC5-2 phage in vivo, we injected phages into the tail vein of mice bearing CL1-5-derived tumors and then recovered them after perfusion. We determined the titers of the phage in tumor masses and normal control organs (brain, heart and lungs) [26, 30] . PC5-2 showed specific homing to tumor masses with concentrations 15-fold higher than its concentration in the control organs ( Fig. 2A) . Control helper phages did not show any specific targeting to tumor tissues ( Fig. 2A) . The tumor-homing ability of PC5-2 was further confirmed by a peptide competitive inhibition experiment, in which synthetic peptide SP5-2, injected together with PC5-2, markedly inhibited the recovery of the phage from tumor masses (Fig. 2B) . One hundred micrograms of SP5-2 inhibited 92% of PC5-2 binding to tumor masses, but the same concentration of a control peptide had no such inhibitory effect (Fig. 2B) . From in vitro phage display screening, we identified two clones (PC5-2 and PC5-4) with a consensus motif of W-T/Y-Y (Table 1) . Like PC5-2, the tumor-homing ability of PC5-4 was also competitively inhibited by SP5-2 ( Fig. S4 and Text S2), suggesting that these two phages may bind through this motif to the same target molecule on the plasma membrane of lung cancer cells. We proposed that these three amino acid residues might play a crucial role in homing to tumor tissues. To test this hypothesis, we changed these three amino acid residues in SP5-2 (TDSILR-SYDWTY) to GGG in a mutant peptide, MP5-2 (TDSILR-SYDGGG). Although the tumor-homing ability of PC5-2 had been markedly inhibited by the peptide SP5-2, this competitive inhibition was lost in MP5-2, which contains the GGG residue instead of the WTY residues in SP5-2 (Fig. 2B) . These data indicate that the WTY residues were important for the binding ability of SP5-2 to NSCLC cells. The tissue distribution of PC5-2 was also studied using immunostaining. We injected SCID mice bearing NSCLC xenografts with PC5-2 and then removed and fixed the tumor and control organs for localization of the phage particles. PC5-2 was found to localize in tumor tissues ( Fig. 2Cd) . At a higher magnification, the immunoreactivity of the phage was detected on the plasma membrane with some diffusion in the surrounding cytoplasm of tumor cells (Fig. 2Ce ). There was no reaction product detected on normal organs such as brain, heart and lung tissues ( Fig. 2C a-c), nor on tumor tissues treated by the control phage (Fig. 2Ci) . The specific targeting of PC5-2 to the NSCLC xenograft was inhibited by the synthetic peptide SP5-2 in the in vivo homing experiment (Fig. 2Cj ). To determine whether the lung cancer-targeting peptide SP5-2 could be used to improve the chemotherapeutic efficacy of cancer treatment, we coupled the peptide to liposomes containing anticancer drugs. SCID mice bearing size-matched, CL1-5-derived xenografts were treated with (1) SP5-2-conjugated liposomal doxorubicin (SP5-2-LD), (2) mutant peptide-conjugated liposomal doxorubicin (MP5-2-LD), (3) non-targeted liposomal doxorubicin (LD), (4) free doxorubicin (FD), or (5) equivalent volumes of phosphate-buffered saline (PBS). All the formulations were injected intravenously (i.v.) at a total doxorubicin dosage of 8 mg/kg (1 mg/kg twice a week for a total of eight injections). The tumors in mice that received SP5-2-LD ( Fig. 3A group a) were significantly smaller than those in the MP5-2-LD, LD, FD, and PBS groups (P,0.01) ( Fig. 3A group b-e). The tumor sizes in the LD and MP5-2-LD groups were 2.1 and 2.6 times larger than those in the SP5-2-LD group, respectively. The tumor sizes in the FD and control PBS groups were 7.0 and 8.3 times larger than that in the SP5-2-LD group, respectively (Fig. 3A) . Interestingly, the greater therapeutic efficacy of SP5-2-LD was lost when the WTY motif in the peptide had been changed to GGG in MP5-2-LD. Free doxorubicin exhibited little therapeutic efficacy at this concentration, as the tumor size in this group was only 16% smaller than that in the PBS group (Fig. 3A) . To verify whether large xenografts would respond to SP5-2-LD treatment, mice bearing large lung cancer xenografts (500 mm 3 ) were assigned to three treatment groups. After a course of doxorubicin treatment with a total dosage of 16 mg/kg (2 mg/kg twice a week for eight injections), the tumor sizes in the control PBS and LD groups gradually increased to 3.3 and 2.1 times the tumor size in the SP5-2-LD group (P,0.01) (Fig. 3B) . These results revealed that SP5-2-LD also increased the therapeutic efficacy of doxorubicin in SCID mice bearing large lung cancer xenografts. To further test whether SP5-2 would optimize the therapeutic index in lung cancer treatment, SP5-2-LD was used to treat a different type of lung cancer xenograft (H460-derived tumor). SCID mice bearing size-matched, H460-derived xenografts were treated with SP5-2-LD, LD, FD, or equivalent volumes of PBS, through i.v. injection at a total doxorubicin dosage of 8 mg/kg (1 mg/kg twice a week for eight injections). The final tumor size in the SP5-2-LD group was significantly smaller than those in the LD, FD and PBS groups (P,0.05). Mice in the LD, FD and PBS groups had tumors with sizes 2.0-, 8.6-and 10.0-fold larger than that in the SP5-52-LD group (Fig. 3C) . Free doxorubicin at this concentration reduced the tumor size by only 14% compared with PBS groups. Not only was tumor growth markedly suppressed in the SP5-2-LD group (Fig. 3C) , the body weight of the mice in this group increased by 10.3% (2.38 g) at the end of the treatment period. In contrast the LD-treated mice had a smaller increase in body weight of 2.4% (0.59 g) (Fig. 3D) . To further confirm SP5-2 could increase the therapeutic index for lung cancer, we linked the SP5-2 peptide to another anticancer drug, liposomal vinorelbine (SP5-2-LV) and tested its efficacy against lung cancer xenografts. SCID mice bearing sizematched CL1-5-derived xenografts were given i.v. injections of SP5-2-LV, liposomal vinorelbine (LV), or equivalent volumes of PBS at a total vinorelbine dose of 24 mg/kg (2 mg/kg twice a week for twelve injections). The tumor-bearing mice treated with SP5-2-LV (Fig. 3E , group a) had significantly smaller tumors than the LV and PBS groups (P,0.01) (Fig. 3E, group b and c) . The tumor size in the LV group was 6.75-fold larger than the SP5-2-LV group. The average tumor size in the control PBS group was 25-fold larger than the SP5-2-LV group (Fig. 3E ). To assess side effects of the treatments, the mice were weighed twice a week. The body weight of mice increased by 5.7% (1.37 g) in the SP5-2-LV group and by 2.4% (0.58 g) in the LV group at the end of the treatment period (data not shown). Finally, we compared the survival rates of tumor-bearing mice after treatment with SP5-2-LV, LV, or PBS over 102 days. All five animals in the PBS-treated group died (survival rate 0%). Three mice died in the LV-treated group (survival rate 40%). In the SP5-2-LV-treated group, however, the survival rate was 80%, significantly higher than the other two groups (Fig. 3F) . These experiments demonstrate that SP5-2 increased the therapeutic efficiency of liposome-encapsulated doxorubicin and vinorelbine with less toxicity. The biodistribution and tumor localization of SP5-2-LD, MP5-2-LD, LD, and FD were estimated by measuring the intrinsic auto-fluorescence signal of doxorubicin in mice with NSCLC xenografts. Doxorubicin (M r 543.54), a small-molecule chemotherapeutic agent as its M r is ,1000, has a poor pharmacokinetic profile, and its blood concentration drops to background level within one hour after administration ( Fig. S5A and Text S2). The pharmacokinetic profile of various liposomal doxorubicin formulations, including SP5-2-LD, MP5-2-LD, and LD, were markedly greater than FD (Fig. S5A) . The area under the concentrationtime curve (AUC 0-48 hours ) of doxorubicin in tumor tissues was 10.4 mg?hr/g, 31.5 mg?hr/g, 29.0 mg?hr/g, and 59.9 mg?hr/g in the FD, LD, MP5-2-LD, and SP5-2-LD groups, respectively (Table S2 ). The mean intra-tumor doxorubicin concentration in the SP5-2-LD-treated group was 5.7-, 1.9-and 2.1-fold higher than the intra-tumor doxorubicin concentration in the FD, LD, and MP5-2-LD groups (Fig. 4A and Table S2 ). To assess the bioavailability of the liposomal drugs, we used nuclear doxorubicin accumulation as an indicator of drug cytotoxicity [40] . The AUC 0-48 hours of bioavailable doxorubicin (i.e., bound to nuclei) for FD, LD, MP5-2-LD, and SP5-2-LD was 3.7 mg?hr/g, 6.7 mg?hr/g, 7.5 mg?hr/g, and 17.7 mg?hr/g, respectively (Table S2 ). The intra-tumor nuclear doxorubicin concentration in the SP5-2-LD group was 4.8-, 2.6-and 2.4-fold higher than the nuclear doxorubicin concentration in the FD, LD, and MP5-2-LD groups ( Fig. 4B and Table S2 ). Doxorubicin concentration inside tumor tissues and nucleus was not significantly different between the LD and MP5-2-LD groups ( Fig. 4 and Table S2 ). To compare the drug delivery profile of the four doxorubicin formulations, we tried to detect the drug in tumor tissues using the fluorescence microscope. Images from all tumors showed that doxorubicin was visualized in the tumor nuclei one hour after SP5-2-LD was administered, but not after other formulations were injected (Fig. 5A) . Over time, areas of tumor sections with detectable doxorubicin increased. The areas with detectable doxorubicin were significantly larger in SP5-2-LD-treated tumors than that in FD, LD, and MP5-2-LD-treated tumors at each time point (Fig. 5A) . The MP5-2-LD formulation had the same distribution pattern as LD, but the regions of FD-treated tumors showed no detectable doxorubicin (Fig. 5A) . When the tumor tissues in each treatment group (Fig. 3C) were examined by H&E staining, markedly disseminated necrotic/ apoptotic areas were present throughout the sections of SP5-2-LDtreated xenografts (Fig. 5B ). In addition, terminal deoxynucleotidyl transferase-mediated dUTP nick end labeling (TUNEL) was used to identify apoptotic cells and tomato lectin was applied to detect tumor blood vessels. The tumors had larger apoptotic areas and a The greater accumulation of anti-cancer drugs in tumor tissues and more bioavailable doxorubicin in cancer cell nuclei from the ligand-conjugated liposomes further demonstrated that the SP5-2 peptide recognized the target molecule on the surface of lung cancer cells and thus increased direct drug delivery to the tumor. The therapeutic efficacy of anticancer drugs can be achieved by enhancing the drug formulation with molecules that preferentially bind to tumor tissues [22, 26, 37, 41] . Phage display biopanning on intact cell allows for the isolation of highly specific peptides that target tumor-associated antigens. Anti-cancer regimens armed with these peptides can be used as ''cruise missiles'' that are precisely guided to the cancer cells and deliver high enough doses to kill these cells with minimal damage to normal tissues. In this study, we identified a NSCLC-targeting peptide and demonstrated its improved therapeutic efficiency in animal models. Specifically, we found a phage clone PC5-2 that made up 90% of the selected phages binding to CL1-5 cells after five rounds of biopanning (Table 1) . Immunohistochemistry and flow cytometry assays confirmed that PC5-2 bound to NSCLC cell surface. The same binding results were observed in the cognate synthetic peptide SP5-2, which replaced PC5-2 ( Fig. 1 and Figs. S1, S2, S3). These experiments supported that this targeting peptide can specifically bind to the cell surface of NSCLC cell lines. In vivo experiments also consolidated the homing ability of the peptide. The expression of the SP5-2 peptide (TDSILR-SYDWTY) guides the phage to accumulate in NSCLC xenografts but not in normal organs (Fig. 2) . The binding activity of PC5-2 to tumor tissues was inhibited by the synthetic peptide SP5-2 (Fig. 2B) , indicating that PC5-2 interacted with NSCLC cells by its displayed peptide and not by another part of the phage particle. That the mutated synthetic peptide, MP5-2, did not inhibit the PC5-2 binding demonstrated the importance of the WTY motif in the binding activity (Fig. 2B) . Moreover, this phenomenon was observed in immunohistochemical staining of surgical specimens from human lung cancer (Fig. 1D e and f) , in animal models for ligand-targeted chemotherapy (Fig. 3A) and in tumor localization of doxorubicin delivered by SP5-2-conjugated targeting liposomes (Figs. 4 and 5A) . Immunohistochemistry assessments on pulmonary adenocarcinoma surgical specimens demonstrated that SP5-2 has a clinical potential as imaging probes to identify NSCLC or drug delivery agents for the treatment of NSCLC. Both PC5-2 and biotinlabeled SP5-2 detected pulmonary adenocarcinoma surgical specimens in our experiments (Fig. 1D) . Seventy-five percent (27/36) of pulmonary adenocarcinoma specimens from 36 patients expressed a target molecule that was recognized by the peptide (Table S1 ). SP5-2, but not MP5-2, competed with PC5-2 for binding to pulmonary adenocarcinoma surgical specimens (Fig. 1D) , which further confirmed the specificity of this targeting ligand. In previous studies, doxorubicin has largely been used in cancer treatment because of its broad spectrum of antitumor activity. However, the efficacy of doxorubicin in the treatment of NSCLC remains unsatisfactory with a response rate of 15% [42] . This is in part a result of suboptimal doses within the tumor due to indiscriminate drug distribution throughout the body and severe toxicity to normal tissues and organs. Liposomal encapsulation with a targeting ligand may be an effective strategy to deliver the drug directly to tumor cells. Our data revealed that this peptide markedly increased the therapeutic efficacy of liposomal chemotherapies and resulted in higher survival rates in mice with human lung cancer xenografts, and produced limited side effects on the animals (Fig. 3) . This peptide increased the therapeutic index of not only doxorubicin but also of vinorelbine, a vinca alkaloid used to treat advanced NSCLC [3] . Furthermore, we observed decreased vessel density and substantially increased cell apoptosis in tumor tissues after the targeting liposome treatment (Figs. 5B and S6). This ligand-mediated liposomal formulation is potentially superior to conventional anti-cancer therapy for NSCLC. Our in vivo pharmacokinetic studies in mice showed that liposomal doxorubicin dramatically changed the transportation and distribution of doxorubicin in the heart, lungs, kidneys, and liver in mice (Figs. 4 and S5) . These results echo other similar investigations [43, 44] . With prolonged presence of liposomes in circulation, more doxorubicin was taken up by tumor cells than conventional doxorubicin administration (Fig. 4) ; this finding was further confirmed in fluorescence signaling of doxorubicin in tumor tissues (Fig. 5A) . This passive targeting phenomenon of non-ligand conjugated liposomes is called the ''enhanced permeability and retention effect'' [45, 46] . In our study, the doxorubicin concentration in the LD-treated group was three times that of the (Table S2) . However, the tumors in the SP5-2-LD group had even higher doxorubicin concentration that was 1.9-and 2.1fold higher than those in the LD and MP5-2-LD groups (Table S2 ). The bioavailable drug in the nuclei of cancer cells in the SP5-2-LD was also 2.6-and 2.4-fold higher than those in the LD and MP5-2-LD groups (Table S2 ). These results indicated that this peptide directly delivered the chemotherapeutic drug to intended targets. Enhancement of drug accumulation in tumor tissues correlated with the increased therapeutic efficacy (Figs. 3, 4 and 5) . The peptide-functionalized liposomes were found to have important clinical potential in a targeted drug delivery system. In order to for them to be used clinically, however, a final targeting liposome construct will need to be selected after each component, i.e. peptide ligands, conjugation methodologies, and liposomal drugs, have been optimized. In conclusion, using phage display peptide libraries to screen for peptides that bind to NSCLC cells, we identified several novel peptides including SP5-2 that specifically bound to the cell surface of NSCLC cells both in vitro and in vivo. Linking SP5-2 to liposomes containing doxorubicin and vinorelbine increased the therapeutic efficacy and survival rates in mice with human NSCLC xenografts because of enhanced tumor apoptosis and decreased tumor angiogenesis. Quantitation and visualization of doxorubicin levels also showed increased drug concentration in tumor tissues in this formulation, highlighting the enhancement in both the delivery and penetration of doxorubicin into the tumor. Our results indicate that the SP5-2 tumor targeting peptide may be used as imaging probes of NSCLC and targeting ligands for liposomal delivery systems to increase the efficacy of chemotherapy for NSCLC. Lung cancer cell lines (A549, CL1-5, H23, H460, and PC13), a nasopharyngeal carcinoma cell line (NPC-TW01), an oral cancer cell line (SAS), and human normal nasal mucosal epithelial (NNM) cells were used in this study. The NNM cells were a primary culture derived from a nasal polyp [27] . The A549, H23, and H460 lines were obtained from the American Type Culture Collection. The CL1-5 line was established by Chu et al. [47] . The A549, CL1-5, H23 and H460 cells were grown in RPMI 1640 (Gibco, CA, USA) containing 10% fetal bovine serum (FBS, Gibco, CA, USA) at 37uC in a 5% CO 2 incubator. The NPC-TW01, PC13, SAS, and NNM cells were grown in DMEM (Gibco, CA, USA) containing 10% FBS at 37uC in a 10% CO 2 incubator. Phage display biopanning procedures CL1-5 cells were first incubated with UV-treated inactive control helper phage (insertless phage). The phage-displayed peptide library (New England BioLabs, MA, USA), which initially contained 5610 10 plaque-forming units (pfu), was then added. The bound phages were eluted with a lysis buffer (150 mM NaCl, 50 mM Tris-HCl, 1 mM EDTA, 1% NP-40, 0.5% sodium deoxycholate, 0.1% SDS, pH 7.4) on ice. This eluted phage pool was amplified and titered in an Escherichia coli ER2738 culture (New England BioLabs, MA, USA). Recovered phages were used as input for the next round of panning as described previously [26, 27] . Ninty-six-well ELISA plates (Falcon, CA, USA) were seeded with either cancer or control cells. Individual phage particles were added to the cell-coated plates and incubated, followed by incubation with horseradish peroxidase (HRP)-conjugated mouse anti-M13 monoclonal antibody (mAb) (Pharmacia, Uppsala, Sweden) and subsequently with the peroxidase substrate ophenylenediamine dihydrochloride (Sigma, MO, USA). The reaction was stopped and absorbance was measured at 490 nm using an ELISA reader. The selected phage clones were further analyzed using DNA sequencing. The sequencing was performed with the primer 59-CCCTCATAGTTAGCGTAACG-39 corresponding to the pIII gene sequence. The phage-displayed peptide sequences were translated and aligned using GCG program. The synthetic targeting peptide SP5-2 (TDSILRSYDWTY), mutant peptide MP5-2 (TDSILRSYDGGG), and control peptide (RLLDTNRPLLPY) [27] were synthesized (Invitrogene, Inc., CA, USA) and purified using reverse-phase high-performance liquid chromatography to .95% purity. Conjugation of these peptides with FITC or biotin was performed through the addition of FITC or biotin to the peptide amino terminus by the same company. Cells were plated and grown to about 80% confluence on cover slips. The cover slips were treated with 3% hydrogen peroxide plus 0.1% NaN 3 to block endogenous peroxidase activity, and then incubated with phages. After the cover slips had been washed and fixed with 3% paraformaldehyde, they were incubated with HRPlabeled mouse anti-M13 mAb and treated with peroxidase substrate. For the peptide binding assays, 30 mg/ml FITC-labeled SP5-2 or control peptide was added on each cover slip and incubated. They were counterstained with Hoechst 33258 (Molecular Probes, OR, USA) and mounted with a mounting solution (Vector, CA, USA). The cells were then examined under a Leica universal microscope. The images were merged using the SimplePCI software (C-IMAGING, PA, USA). For localization of peptide binding on lung cancer tissues, frozen sections of NSCLC tissues were prepared and incubated with phage clones or biotin-labeled peptides. For the peptide competitive inhibition assay, phages were mixed with the synthetic targeting or mutant peptide. The slides were subjected to routine immunohistochemical staining [26, 30] . All surgical specimens were obtained from the tissue bank of National Taiwan University Hospital (NTUH) with approval from the Institutional Review Board in NTUH (IRB9461702021). In vivo homing experiments and tissue distribution of phages SCID mice were injected subcutaneously (s.c.) in the dorsolateral flank with 1610 7 human NSCLC cells. The mice bearing size-matched lung cancer xenografts (approximately 500 mm 3 ) were injected i.v. with 10 9 pfu of the targeting or control phage. After perfusion, xenograft tumors and mouse organs were removed and homogenized. The phages bound to each tissue sample were recovered through the addition of ER2738 bacteria and titered on IPTG/X-Gal agar plates. In the peptide competitive inhibition experiments, the phages were injected along with 100 mg synthetic targeting peptide, control peptide or mutant peptide. The organs and tumor masses were fixed in Bouin's solution (Sigma, MO, USA). After fixation, the samples were embedded in paraffin blocks. The paraffin sections were deparaffinized, rehydrated, and subjected to immunostaining using the mouse M13 mAb as described above. Peptide-conjugated liposomes containing doxorubicin or vinorelbine were prepared as described in other studies [26, 27] . Briefly, the peptide was coupled to NHS-PEG-DSPE [N-hydroxysuccinimido-carboxyl-polyethylene glycol (MW, 3400)-derived distearoylphosphatidyl ethanolamine] (NOF Corporation, Japan) in a 1:1.5 molar ratio. The reaction was completed and confirmed by quantitation of the remaining amino groups using TNBS (Trinitrobenzenesulfonate) reagent (Sigma, MO, USA). Doxorubicin and vinorelbine were encapsulated in liposomes through a remote loading method at a concentration of 1 mg of drug per 10 mmol phospholipids. Peptidyl-PEG-DSPE was transferred to pre-formed liposomes after co-incubation at a transition temperature of the lipid bilayer. There were 500 peptide molecules per liposome as described previously [48] . SCID mice bearing NSCLC xenografts (,500 mm 3 ), were injected in the tail vein with various formulations of liposomal doxorubicin (SP5-2-LD, MP5-2-LD, and LD) and free doxorubicin at a dose of 2 mg/kg. At selected time points, three mice in each group were anaesthetized and sacrificed. Blood samples were collected through submaxillary punctures, and plasma samples were prepared. After perfusion, xenograft tumors and mouse organs were removed and homogenized. Procedures for isolating tumor cell nuclei and extracting nuclear doxorubicin were carried out according to previous reports [40, 49] . Total doxorubicin concentration was measured using a method described by Mayer et al. [49] . Total doxorubicin was quantified using spectrofluorometry at l ex 485/20 nm and l em 645/40 nm (Synergy HT Multi-Detection Microplate Reader, BioTek Instruments, Winooski, VT 05404 USA). To correct for background fluorescence, a standard curve was obtained by spiking tissue extracts derived from mice that did not receive doxorubicin. Tissue concentrations of doxorubicin were expressed as microequivalents per milliliter of plasma or per gram of tissue. A standard doxorubicin curve was prepared in control homogenates following the same extraction procedure as above. Drug levels were estimated on the basis of doxorubicin fluorescent equivalents. To determine the presence of the drug localized in tumor tissues, doxorubicin autofluorescence was detected using a Zeiss Axiovert 200 M inverted microscope with a 100 W HBO mercury light source equipped with a 546/12 nm excitation and a 590 nm emission filter set. Tissue sections were imaged with a FLUAR 106/0.50 NA lens and captured with a Roper Scientific CoolSnap HQ CCD camera. All images were captured in 8-bit signal depth and subsequently pseudo-colored. Animal model for the study of ligand-targeted therapy Mice 4-6 weeks of age were injected s.c. in the dorsolateral flank with human NSCLC cells. Mice with size-matched tumors (approximately 50-100 mm 3 ) were then randomly assigned to different treatment groups and injected with SP5-2-LD or SP5-2-LV, or LD or LV through the tail vein. The dosage of SP-5-2-LD was 1 mg/kg injected twice a week for four weeks, and that of SP5-2-LV was 2 mg/kg injected twice a week for six weeks. Mice bearing large tumors (approximately 500 mm 3 ) were treated with SP5-2-LD for eight times (2 mg/kg, twice a week for four weeks). Mouse body weights and tumor sizes were measured twice a week. Tumor volumes were calculated using the equation: length6(width) 2 60.52. Animal care was carried out in accordance with guidelines of Academia Sinica, Taiwan. Terminal deoxynucleotidyl transferase-mediated dUTP nick end labeling (TUNEL) staining The frozen tumor tissue sections were incubated with terminal deoxynucleotidyl transferase-mediated dUTP nick end labeling reaction mixture (Roche Diagnostics) at 37uC for an hour. The slides were counterstained with mounting medium with DAPI (Vector Laboratories). The slides were then visualized under a fluorescent microscope. Tissues were removed from treated mice, fixed with 4% paraformaldehyde and embedded with paraffin. Blood vessels were detected by staining of Lycopersicon esculentum (tomato) lectin conjugated to biotin (Vector, CA, USA). The biotinylated lectin was visualized with streptavidin-conjugated rhodamine (Pierce, IL, USA). We analyzed the data of phage titer, tumor volume, body weight, and doxorubicin concentration using two-sided unpaired Student's t-test. We considered a P value below 0.05 as significant for all analyses. All values are represented as mean6standard deviation. Text S1 Supporting Information and figure legends Figure S1 Identification of PC5-2 binding to NSCLC cells. Representative microscopy images of CL1-5 cells stained with propidium iodide (a-f, red), anti-M13 monoclonal antibody (g-l, green), and merge (m-r). The binding of PC5-2 to CL1-5 cells (a, g, m) was inhibited by 3 mg/ml (b, h, n), 9 mg/ml (c, i, o), and 27 mg/ml (d, j, p) of SP5-2 in a dose-dependent manner. The control phage and PC5-2 did not bind to CL1-5 cells (e, k, q) and NPC-TW01 cells, respectively. Scale bar: 10 mm. Figure S4 Tumor homing ability of PC5-4 phage. SCID mice bearing NSCLC xenografts were injected i.v. with PC5-4, and phage was recovered after perfusion. Recovery of PC5-4 from the tumor was higher than from control organs. Targeting activity of PC5-4 to tumor tissues was inhibited by SP5-2.

IL-1β, IL-6, and RANTES as Biomarkers of Chikungunya Severity

BACKGROUND: Little is known about the immunopathogenesis of Chikungunya virus. Circulating levels of immune mediators and growth factors were analyzed from patients infected during the first Singaporean Chikungunya fever outbreak in early 2008 to establish biomarkers associated with infection and/or disease severity. METHODS AND FINDINGS: Adult patients with laboratory-confirmed Chikungunya fever infection, who were referred to the Communicable Disease Centre/Tan Tock Seng Hospital during the period from January to February 2008, were included in this retrospective study. Plasma fractions were analyzed using a multiplex-microbead immunoassay. Among the patients, the most common clinical features were fever (100%), arthralgia (90%), rash (50%) and conjunctivitis (40%). Profiles of 30 cytokines, chemokines, and growth factors were able to discriminate the clinical forms of Chikungunya from healthy controls, with patients classified as non-severe and severe disease. Levels of 8 plasma cytokines and 4 growth factors were significantly elevated. Statistical analysis showed that an increase in IL-1β, IL-6 and a decrease in RANTES were associated with disease severity. CONCLUSIONS: This is the first comprehensive report on the production of cytokines, chemokines, and growth factors during acute Chikungunya virus infection. Using these biomarkers, we were able to distinguish between mild disease and more severe forms of Chikungunya fever, thus enabling the identification of patients with poor prognosis and monitoring of the disease.

In the recent years, emerging and re-emerging tropical infectious diseases have been shown to cause high social and economic impact. Vector-borne infectious diseases such as Dengue, West Nile have been resurging largely due to the spread of insecticide resistance, to socio-demographic changes, and to genetic mutations in the pathogens. More recently, chikungungya fever (CHIKF) has now emerged as the next important infection in South-East Asia, the Pacific region and Europe [1] [2] [3] [4] [5] , making it a major threat that requires immediate attention. Recent epidemic resurgence of CHIKF in several African and Asian countries demonstrated that infection can spread alarmingly rapidly [6] [7] [8] [9] from limited early transmission that then developed into an unprecedented and unexpected epidemic, infecting 38% of the population as occurred in Reunion island [6, 8] . The appearance of cases in Europe, the United States and other countries by travelers returning from known outbreak areas underscores the contributory factors of increased human mobility, tourism, global climate change, and increases in insecticide resistance [10] [11] [12] [13] [14] [15] . In this era of globalization, the threat of such disease epidemics should not be underestimated as such public health events could cripple public health systems and economies. Chikungunya virus (CHIKV), which causes CHIKF, is an alphavirus of the Togaviridae family, with a 12,000-nucleotides linear, positive-sense, single-stranded RNA genome containing two large open reading frames (ORF). The first, ORF1, encodes 4 non-structural proteins (nsP1, nsP2, nsP3 and nsP4) while ORF2 encodes structural proteins that include 1 capsid protein (C), 2 major envelope surface glycoproteins (E1, E2) and 2 small proteins (E3, 6K) [8, 9] . CHIKV is transmitted by Aedes mosquitoes (mainly A. albopictus and A. aegypti). CHIKF is an acute illness with abrupt fever, skin rash, arthralgia, and occasional involvement of the nervous system, heart and liver. Prolonged incapacitating arthralgia has sometimes been reported to persist for years [8, 9, 13] . It is of concern that the re-emerged CHIKV has caused considerable morbidity and some fatalities, whereas previously CHIKF was considered as relatively benign. Despite the fact that the clinical features of recent acute CHIKV infections from several countries have been described [16] [17] [18] , little is known about the long-term sequelae or the pathogenesis of arthropathy, and the acquisition of protective immunity remains unexplored. It has been proposed that CHIKVinduced arthritis or arthralgia is of immunopathologic origin [19, 20] . At present, there is no specific or effective treatment for CHIKF, and patient management is largely symptomatic relief and primarily anti-inflammatory drugs [8] . Given the expanding geographic range of CHIKV and its potential to rapidly cause large scale epidemics, it has become important to understand the immune and pathogenic mechanisms active during CHIKV infections in order to guide the development of targeted and effective control and treatment strategies. Cytokines and chemokines are thought to play an important role in viral immunopathology. Although IL-2. IL-10 and IFN-c have been implicated in the pathogenesis of CHIKF [21] , a global analysis of their specific involvement with disease severity has not yet been defined. Growth factors are usually produced in response to injury. Viral infections such as CHIKV, induce cellular damage which may lead to secretion of these factors; however, limited studies have been conducted [21] . In this study, we took the opportunity to conduct a detailed study on the patients from the first outbreak of CHIKF in Singapore [22, 23] . We measured circulating levels of a wide range of cytokines, chemokines, and growth factors in 10 laboratory confirmed cases of CHIKF, and compared them with healthy individuals. We next determined which biomarker was associated with infection and/or severity. We showed for the first time that CHIKV infection induced a wide range of cytokines, chemokines, and growth factors. We subsequently found that 3 specific biomarkers, namely IL-1b, IL-6, and RANTES levels, were associated with severe CHIKF. All 10 patients included in this study were males. Their age ranged from 22 to 65 years (median, 35 years). All except one were foreign nationals. Half of our patients were classified as severe CHIKF. We defined severe CHIKF as having a temperature of .38.5uC or pulse rate .100/min, or platelet count ,100610 9 g/ L based on studies defining severe diseases [24] [25] [26] [27] [28] . Except for the Singapore resident, none had any pre-existing medical condition. Despite being previously healthy, four non-residents developed severe illness. Fever lasted 2-10 days, and fever duration was not significantly different between those who had more severe illness and those who had not (mean, 6.6 days vs. mean, 3.8 days; P.0.05). Two patients reported persistent arthralgia lasting more than two weeks. Demographic and clinical details of the 10 CHIKF patients are summarized in Table 1 . Among our patients, the most common clinical features were fever (100%), arthralgia (90%), rash (50%), and conjunctivitis (40%) ( Table 2 ). Gastrointestinal and constitutional symptoms were less prominent. Arthritis was observed in only one patient, who had an effusion on the right knee. None had neurologic involvement or hemorrhagic manifestation. Table 3 presents a summary of the key laboratory findings among our patients throughout the course of their illness. White cell count, hemoglobin, hematocrit, platelet count, erythrocyte sedimentation rate for most patients were within the normal range. The mean nadir platelet count (6SD) was 1996115610 9 /L. Only one patient had severe thrombocytopenia (nadir platelet count, ,100610 9 /L) during the course of his illness. Elevated C-reactive protein levels (CRP, .10.0 mg/L) were observed in 60% of patients, but the peak C-reactive protein level was not significantly different between those who classified as severely ill and those who were not (mean, 40.3 mg/L vs. mean, 9.9 mg/L; P = 0.195). The mean peak alanine and aspartate transaminases (ALT and AST) (6SD) were 58636 U/L and 50625 U/L respectively. Both ALT and AST were 2-fold greater than the upper limit of normal in one patient, who had pre-existing liver cirrhosis. None of the patients had a clinically abnormal total protein, urea or creatinine level. Among our patients, the mean nadir protein level (6SD) was 6765 g/dL, and the mean peak urea and creatinine levels (6SD) were 5.161.7 mmol/L and 101616 mmol/L respectively. Lactate dehydrogenase level was the only laboratory parameter that was significantly higher in severely ill patients, compared to those who were not (mean, 732 U/L vs. mean, 525 U/L; P = 0.047). Profiles of 30 cytokines, chemokines and growth factors were determined by a multiplex-microbead immunoassay on acute blood samples collected upon hospitalization. The samples collected ranged from day 2 to day 19 of illness (median, day 4.5). To characterize the overall patterns, a two-way hierarchical clustering analysis was done to allow the classification of individuals according to disease severity based on the clinical features ( Figure 1 ). Evidently, this had the power to discriminate the clinical forms of CHIKF in the samples in this study from the healthy controls, with patients classified as non-severe and severe disease segregating perfectly. The levels of 8 plasma cytokines (IL-2R, IL-5, IL-6, IL-7, IL-8, IL-10, IL-15 and IFN-a) were observed to be most significantly elevated (Figure 2a ) in CHIKF patients compared to uninfected subjects (P,0.05). Among these was proinflammatory cytokine, IL-6 which was very significant. Interestingly, another proinflammatory cytokine, IL-8 was down-regulated in these patients. Anti-inflammatory cytokine, IL-10 was found to be significantly raised in most of the patients (P,0.05). The plasma concentrations of IL-2R and IL-5 were found to be increased in all patients. Levels of IFNa and IL-7 were elevated in all patients. The levels of other cytokines such as IL-2, IL-4, IL-12, IL-13, IL-17, IFN-c, and TNF-a, were only marginally increased in the CHIKF group compared with those in the uninfected group ( Figure S1 ). Profiles of chemokines, IP-10 and MIG were shown to be significantly elevated, while Eotaxin was suppressed (Figure 2b ). There was no difference in the levels of other chemokines namely, MCP-1, MIP-1a, MIP-1b and RANTES ( Figure S1 ). Interestingly, the levels of 4 growth factors were found to be significant in the patients, with up-regulation of HGF, FGF-basic and VEGF, with the exception of EGF which was almost totally suppressed ( Figure 2c) . It was observed that the CHIKF patients exhibited low levels of GM-CSF and G-CSF ( Figure S1 ). Finally, in an effort to identify cytokine, chemokine, and growth factor plasmatic levels associated with severity, statistical analyses were performed after stratification of the CHIKF patients according to severity. It was observed that an increase in levels of IL-1b and IL-6, and a decrease in RANTES respectively were associated with disease severity (Figure 3a) . The levels of all other markers were not significantly different ( Figure S2 and S3). CHIKF, an emerging arboviral infection, which induces high fever, has only been recently reported in Singapore. Up to Dec 2007, all CHIKF patients had contracted the infection overseas [22] . The first local outbreak of CHIKF occurred in Jan 2008. More than 2,500 people who lived or worked in the outbreak area were screened and a total of 13 PCR-confirmed cases were identified [22, 24] . All confirmed CHIKF cases were referred to the CDC/TTSH. Our report included 10 patients who participated in this study. Phylogenetic analysis of the viral sequences of our patients has revealed that the circulating strains were of the Indian Ocean genotype and closely related to those from the 2006 outbreak in India [29] but without the A226V mutation, further emphasizing how remarkably rapid the disease could spread with the right environmental conditions. The attack rate in our outbreak was 0.5%, much lower than the 34% reported in Reunion Island and the 5.4% observed in Italy [13, 30] . This could be attributed to the rapid removal of human reservoirs through isolation, enhanced vector control, or the circulation of a virus strain of lower epidemic potential. Clinical features of our patients were similar to those reported in recent outbreaks [6] [7] [8] [9] [11] [12] [13] [14] [16] [17] [18] 21, 30] , indicating that although people are genetically diverse response to diseases is homogeneous across people in non-homogeneous populations. The majority of our patients was #45 years and had no premorbid condition. Unlike patients reported in the Reunion Island outbreak [6] , where the patients' underlying medical conditions could have contributed to the observed morbidities, our patients were younger and healthier. Furthermore, none of our patients was co-infected with dengue, as confirmed by RT-PCR and dengue enzyme-linked immunoabsorbent assay (ELISA)-IgM and IgG [31] . Hence, our immunologic observations can be largely attributed to acute CHIKV infection itself. In recent years, most of the studies on CHIKF have been addressed with the clinical description of the disease [4] [5] [6] [7] [8] [9] 11, 17, 18, 21] , the molecular nature of the virus [9, 10, 19] and diagnostics methods [8] [9] [10] 24] , and the interactions of the virus with its mosquito vector, Aedes [1] [2] [3] [11] [12] [13] [14] [15] . Here, we describe for the first time the comprehensive systemic production of cytokines, chemokines, and growth factors during acute CHIKV infection which may light the path ahead in understanding the innate response to the infection. We first showed that a wide range of cytokines such as IFN-a, IL-5, IL-6, IL-7, IL-10, IL-15 were produced in response to CHIKV infection. IFN-a is a potent antiviral cytokine and has been shown to strongly inhibit CHIKV in vitro [32] . The high levels of IFN-a that we detected provide a logical explanation for how the body rapidly brings CHIKV viremia under control [8, 11] . It has been shown that the main producers of IFN-a are plasmocytoid dendritic cells [33] and monocytes [34] . The profile of circulating cytokines revealed a predominance of Type 2 cytokines. Mainly IL-5, IL-6 and IL-10 levels were increased and those of IFN-c or TNF-a were unchanged as compared to noninfected controls. This suggests that acute CHIKV infection tilts the cytokine profile to anti-inflammatory response, which would argue against the common understanding of CHIKV infection which does not really support the common description of the CHIKV infection as an inflammatory disease [8] . Alternatively, it is possible that an inflammatory response might occur earlier when the virus is actively replicating, and then gets down-regulated by a counter-anti-inflammatory response when the virus is being eliminated from circulation. High levels of anti-inflammatory IL-10 and the presence of high levels of chemokines IP-10 and MIG (ligands of CXCR3 associated with Th1-type reactions) [35] , detected here would support this hypothesis. Further studies would be needed to clarify this issue. The Type 2 cytokines detected are also important mediators of B cell growth and maturation, and thus may allow the production of high levels of persisting anti-CHIKV IgG [5] . The detection of high levels of circulating IL-15 is of interest, since this cytokine has been shown to be a major stimulator of NK cells [36] and T cells [37] . Thus our data suggest that these lymphocytes population might be activated during acute infection and may also contribute to viral control during the acute phase of CHIKV infection. Detection of soluble IL-2R in the plasma suggests T cell activation since this molecule is secreted by activated T cells [38] . Experiments are planned to study the activation phenotype of T and NK cell subsets in acutely infected patients. The detection of IL-7 and IL-15 is significantly interesting with regards to the immunopathology of CHIKF since CHIKV infection has been shown to induce rapidly developing and persisting arthralgia [6] . Here, 9 of the 10 patients manifested this pathology. IL-7 is known to have an important role in the development of rheumatoid arthritis [39] , while IL-15 has been associated with the development of joint inflammation [40] . It has been proposed that expansion of a particular IL15-induced NK cell subsets was responsible for this phenomenon [41] . The role of IL-15 and NK cells in the development of CHIKV arthralgia would definitively be worth investigating. We did not detect TNF-a in the plasma of the patients with acute CHIKV infection. This is surprising since this cytokine has been detected repeatedly in the blood of patients suffering from other arthritides such as rheumatoid arthritis and is known to be involved in the pathogenesis of these entities [42] . Thus, it is possible that CHIKV-induced arthralgia does not depend on TNF-a. Alternatively, TNF-a might be produced only locally. Analysis of synovial fluid or joint tissue immunohistochemistry would be necessary to provide important information on the role of TNF-a and other mediators. Chemokines are crucial mediators of innate and adaptive immunity against various viral infections [43] . IP-10, and MIG had increased plasma levels during CHIKV infection. These two chemokines signal through the same receptor CXCR3 and thus might activate and direct migration of this T cell subset [35] . IL-8 and Eotaxin levels were lower than those of naive controls. Defining the exact contributions of these different chemokines will require further studies. We also tested the presence of growth factors in the plasma of CHIKV-infected patients. HGF, FGF-basic and VEGF were produced at high levels and may reflect a physiological response to tissue destruction resulting for the viral infection. Interestingly, EGF levels were lower than in healthy controls. The low levels of EGF might be due to the concomitant decrease of platelets observed in infected patients since previous studies have shown that plasma levels of EGF are associated with circulating platelets [44] . Although limited, we had access to sufficient patients to perform data analysis in relation to the severity of the disease (severe illness was defined as fever .38.5uC, or maximum pulse rate .100 beats/minute, or nadir platelet count ,100610 9 /L). Using this definition, we observed that higher disease severity was associated with increased plasma levels of IL-1b and IL-6 and a decreased level in RANTES (Figure 3b ). IL-1b and IL-6, whose levels are already high in the CHIKV infected patients, are potent endogenous pyrogens [45] [46] [47] [48] . Therefore, elevations of IL-1b and IL-6 might account for the high fever in the severe cases. The increase production of IL-1b might also mediate the development of abrupt and persistent arthralgia since this cytokine is involved in the immunopathogenesis of many arthritic pathologies such as rheumatoid arthritis [49] . On the contrary, T cell chemokine RANTES levels were significantly suppressed in severe CHIKF patients. Platelets are a major reservoir of RANTES in the peripheral circulation [50] , and severe CHIKF was characterized by thrombocytopenia. Thus, as mentioned above for EGF, thrombocytopenia can also reduce levels of circulating RANTES. Low levels of RANTES correlate with disease severity and mortality in individuals with severe malaria, who were also correspondingly thrombocytopenic [51] . Interestingly, it was observed in other studies that RANTES levels were up-regulated in dengue [48] , except for one single report from Cuba [52] . Since the symptoms of CHIKF mimic those of dengue fever, results obtained from this study strongly suggest that RANTES could be a potential biomarker that differentiates between these 2 clinically very similar diseases. One limitation of this study is in the classification of disease severity as none of our patients developed neurologic or hemorrhagic complications previously reported in CHIKF patients. Nonetheless, our definition of severe illness would have included patients with sepsis, a serious form of infection commonly associated with a temperature of .38uC and heart rate of .90/ minute [24] . Furthermore, we included thrombocytopenia of ,100610 9 /L as a criteria for severe CHIKF. Marked thrombocytopenia is a common feature of sepsis [25] and has been identified as a predictor of mortality [26, 27] . The degree of thrombocytopenia is a determinant of survival and once the platelet count decreases below 100610 9 /L, mortality continues to increase, even though the risk of bleeding does not [28] . A wide spectrum of disease has been reported in CHIKF ranging from asymptomatic infections, to self-limiting febrile illness [8] , to neurologic complications, and death [17] . The ''severe illness'' cohort in our study possibly represents a more severe form of selflimiting febrile illness, an intermediate group with higher levels of viremia (data not shown) and distinctly more severe clinical features (i.e. high temperature, tachycardia, and severe thrombocytopenia). Using this clinical phenotype, we have shown in this study that immune mediators are able to distinguish very mild disease from more severe forms of CHIKF disease at the acute stage. Follow-up studies will be required to determine if long-term sequelae are indeed different between non-severe and severe clinical presentations. Elucidating the association of disease severity with two cytokines and one chemokine can be useful in order to provide early identification and monitoring of patients with severe disease. Although this study is limited by the size of the outbreak, nevertheless, based on these observations, measurement of immune mediators could be helpful for the management of future outbreaks. This study strongly suggests these biomarkers be used for measuring disease severity and be tested in outbreaks in different populations and different strains. Once confirmed, they will be useful for follow-up studies, association studies, and prognosis for public health management. More importantly, these biomarkers can potentially lead to the development of modulators to reduce disease severity and to halt disease progression. Ten patients who presented with acute CHIKF to the Communicable Disease Centre at Tan Tock Seng Hospital (CDC/TTSH), the national infectious disease referral centre in Singapore, during the outbreak period from January to February 2008, were included in this study. An acute case of CHIKF was defined as any case with clinical features consistent with CHIKF, and had CHIKV infection confirmed by either reverse transcriptionpolymerase chain reaction (RT-PCR) or virus isolation [53, 54] . The study was approved by the institution's domain-specific ethics review board (DSRB Reference No. B/08/026). Written consent was obtained from each patient and healthy control subject. Plasma samples were obtained from patients during the acute phase of their illness. Data on demographic characteristics, premorbid conditions, clinical features, and routine hematological and biochemical laboratory test findings (i.e. full blood count, renal and liver function tests, C-reactive protein) were also collected. All symptomatic patients were isolated at CDC/TTSH until the febrile illness resolved and a negative CHIKV RT-PCR test was obtained. During the hospital stay, daily monitoring of body temperature, vital parameters, and blood counts were carried out. A patient was defined as having severe illness, if he had either a maximum temperature of more than 38.5uC, or a maximum pulse rate of more than 100 beats/minute, or a nadir platelet count of less than 100610 9 /L. Laboratory results were expressed as mean6SD. In addition, plasma or serum samples from 9 healthy volunteers (who did not have a febrile illness in the preceding week and were not epidemiologically-linked to the outbreak) were also included as controls in our study. Plasma samples collected as described above, were aliquoted and stored at 280uC until analyses were done. A multiplex biometric immunoassay, containing fluorescent dyed microspheres conjugated with a monoclonal antibody specific for a target protein, was used for cytokine measurement according to manufacturer's instructions (Biosource Human Cytokine 30-plex Assay, Invitrogen). The following groups of cytokines were: inflammatory (GM-CSF, IL-1beta, IL-1RA, IL-6, IL-8, TNF-alpha); Th1/Th2 (IFN-gamma, IL-2, IL-2R, IL-4, IL-5, IL-10); Cytokine II (IFN-alpha, IL-7, IL-12p40/p70, IL-13, IL-15, IL-17); Chemokines (Eotaxin, IP-10, MCP-1, MIG, MIP-1alpha, MIP-1beta, RANTES); and Growth Factors (EGF, HGF, FGF-basic, G-CSF, VEGF). Briefly, 25 ul of plasma samples were diluted in 1:2 and incubated with antibody coupled beads for 2 h at RT uC. Complexes were washed twice with the use of a vacuum manifold before incubation with biotinylated detector antibody for 1 h at RT uC. Complexes were then washed twice followed by incubation with Streptavidin-phycoeryrhrin (RPE) for 30 min at RT uC. Complexes were washed thrice and incubated with wash buffer for another 3 min before detection in the Luminex 200 TM instrument. Results were acquired by the IS 2.3 software and the standard curves were plotted through a fiveparameter logistic curve setting. To compare and analyze the expression profiles across the results, the raw cytokine values were normalized using z-score conversion based on the formula: where x is the raw value to be converted, m is the mean of the population and s is the standard deviation (SD) of the population. The transformed value is denoted by z and exhibits positive value when the raw value is above mean and vice versa. Further, to examine the nuances and correlations masked in the full set of data, the z values are subjected to cluster analysis [55] to yield an ordered NumOfRows.x NumOfCols expression level matrix, E el . Hierarchical clustering is applied on the columns which represent the myriad of cytokine levels measurements (e.g.) based on McQuitty's or WPGMA method [56] where the distance between a pair of groups A and B is measured using the weighted arithmetic mean of all the pairwise distances between the data points in A and B. The rows which represent the suspected CHIKF patients and healthy individuals are left untouched. Seriation [57] is performed following the clustering approach to re-order the clustered data points using the minimum path length algorithm to minimize the sum of all the distances between adjacent columns. Euclidean distance is used in both the clustering and seriation phases to measure the difference or dissimilarity, d, between data points (x 1 , y 1 ) and (x 2 , y 2 ) given by the equation: d~ffi ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Comparisons between groups were calculated by Mann-Whitney rank sum. Further statistical analyses were done by Kruskal-Wallis test followed by Dunn's multiple comparison tests. P values of ,0.05 were considered to be statistically significant. Figure S1 Complete profile of the levels (pg/ml) of cytokines, chemokines and growth factors determined by multiplex-bead arrays from blood samples collected from CHIKV-infected patients and healthy control subjects.