Experiments on the host responses to infection have generated a wealth of data over the past decade. A number of these have focused on respiratory viruses including several which resulted in lethality in experimental animal models. What distinguished these infections from others where experimental animals recovered was the subject of our analysis. To this end we assembled a compendium of transcriptome measurements from the lungs of mice and used meta-analysis to determine gene signatures that distinguished high- from low-pathogenicity infections. To our knowledge ours is the first meta-analysis study to focus on high-pathogenicity infections (HPIs). Previous studies have applied meta-analysis to infection data but focused on more general features of disease. For example, Jenner and Young analyzed the transcriptomes of various host species (mice, macaques, and humans) infected by bacteria or viruses and identified a signature of approximately 500 genes . This signature showed that TLRs and pathogen-mediated signalling were broadly activated during infection but did not identify features resulting in animal lethality. Likewise Pennings et al. examined the transcriptional profiles of lung inflammation (in mice and macaques) due to various factors (including viruses, bacteria, chemicals, and allergens) and derived a 383-gene signature up-regulated during inflammation . This signature focused on the interferon response and immune signalling, but likewise no correlate of pathogenicity was identified. In addition both studies relied mainly on hierarchical clustering to determine signatures of interest. While hierarchical clustering provides a useful overview of high-dimensional data, it has the disadvantage of lacking a strong statistical basis (i.e., hypothesis testing) and its results (gene clusters) can be difficult to relate to specific outcomes.
In contrast to previous studies, our goal was to determine a gene signature specific for HPIs. We therefore applied an ensemble of meta-analysis methods and derived multiple signatures which we compared by multiple criteria including the capacity to predict the outcome of a test data set. Each method had a different statistical rationale and could be expected to identify different features of the compendium. We had no a priori expectation of which signature would produce the best outcome in terms of these criteria.
The 74-gene signature derived using Fisher's summary statistic showed a modest capacity to separate samples into pathogenicity groups and to predict test set pathogenicity. A number of the genes in this signature have known connections to the immune response and the outcome of respiratory infections, e.g., genes for the chemokines osteopontin (SPP1) and RANTES (CCL5) and the chemokine receptor CCR2. In previous studies mice lacking SPP1 and CCL5 were found to clear influenza infection with no adverse effects [10, 11], while mice lacking CCR2 survived infection by a mouse-adapted influenza A virus that killed wild-type mice . These studies suggest that at least some of the genes up-regulated during HPIs are non-essential to resolving influenza infection and that dysregulated activation may even be detrimental to the host.
The best performing signature, however, comprised 57 genes derived using the fold change-based z-test. Strikingly the majority of the genes in this signature were expressed at levels that corresponded with pathogenicity. For the majority of these genes, expression was lowest in LPI, highest in HPI, and intermediate in MPI (which had not been used in the derivation of the signature). In this case the signature appeared to provide a continuously variable signal that matched with output, a characteristic of analog signals. This finding also suggests that high- and low-pathogenic infections may result in the expression of the same key genes but with different kinetics. In particular HPIs may result in increased expression of signature gene products beyond the capacity of the host to cope, resulting in irreversible damage.
Genes of the analog signature largely differed from those in the digital signatures but overlapped at the pathway level. For example, chemokine genes were present in both Fisher's statistic-derived and analog signatures. However, the analog signature displayed additional coherence, encoding multiple chemokines for the same receptor, namely MIG (CXCL9), IP-10 (CXCL10), and I-TAC (CXCL11) which all serve as ligands for the receptor CXCR3 . CXCR3 is expressed on the surface of Th1 cells as well as NK and NKT cells and regulates the migration of these cells to sites of infection. Recent evidence indicates that CXCR3 engagement may drive further recruitment and inflammation, resulting in a positive feedback loop that may contribute to pathogenicity . The analog signature also included genes for chemokines that bind the CCR3 receptor, specifically MCP-3 (CCL7) and MCP-4 (CCL13). CCR3 is the major receptor expressed on eosinophils and has previously been shown to have a role in the promotion of lung inflammation . Interestingly CXCR3 ligands have been postulated to be antagonists for CCR3 , and the expression of both sets of chemokines may reflect a high degree of dysregulation during HPIs and the recruitment of multiple immune cell types that may not normally co-localize during a controlled infection.
In addition our signatures also identified genes that were down-regulated during HPI, relative to either mock infection or LPI. For example the Fisher's statistic-derived signature included several genes whose products may normally help to resolve infection. Hepsin (HPN) has been found to cleave influenza hemagglutinin directly resulting in non-infectious virus particles . Likewise surfactant-associated protein A1 (SFTPA1) and surfactant-associated protein D (SFTPD) maintain pulmonary structure but have also been found to inhibit infectivity by mechanisms that remain to be elucidated [17, 18]. The analog signature contained a similar set of genes but expressed at lower levels in HPIs compared to LPIs rather than to mock infections. Among these genes were several members of the secretoglobin family including SCGB1A1, SCGB3A1, and SCGB3A3. Interestingly many secretoglobins are expressed specifically in the lung epithelium and may contribute to lung repair following damage [19, 20]. For example, uteroglobin/CC16 (SCGB1A1) is secreted by bronchiolar Clara cells and postulated to have a role in reducing airway inflammation, though its exact function remains to be elucidated .