LB 13-123 - Predicting host-virus compatibility with a trait-based machine learning model

Most emerging infectious diseases are caused by zoonotic pathogens transmitted from animals to humans. Predicting which host species could be infected by specific pathogens is an important component of monitoring and preparing for potential cross-species transmission events that can seed outbreaks. Machine learning models have demonstrated promise in prioritizing sampling for known and novel viruses, facilitating surveillance efforts by predicting species with high probabilities of hosting pathogens. Many such predictive models have been trained on host traits including phylogeny, biogeography, and life history traits, but often lack key viral traits such as genomic features that also influence potential host range. To identify the contribution of viral genomic traits to predicting host-virus associations or ‘links’, we extend the application of boosted regression trees trained on host trait data to include viral genomic features, using orthopoxviruses as a case study. This group of closely related viruses includes generalist and specialist pathogens frequently attributed to the variable repertoire of accessory genes which antagonize the host innate immune system. We incorporate principal components of the presence-absence of orthopoxvirus accessory genes as viral traits in our link prediction model. We compare performance of our link prediction model against a model based on only host traits. We identify taxonomic patterns in predictions; visualize the spatial distribution of predicted mammal genera; and rank model features by their relative importance.The link prediction model informed by both host and virus features outperformed the host-only model (AUC = 0.98 vs. 0.886). Both model types displayed strong phylogenetic signal in their predictions, and although they differed in some of the host clades likely to harbor novel associations, they both predicted the family Felidae to be more likely to host orthopoxviruses. Results of the link prediction model estimated more than a four-fold increase in the number of mammal genera as potential orthopoxvirus hosts and revealed more distinct areas of high overlap in the distribution of predicted host species compared to the host-only model. Finally, within the link prediction model, several viral predictors constructed from principal components were highly influential, among which genes with the greatest contributions had numerous predicted functional roles including encoding of membrane-associated proteins and virulence factors that target innate immune pathways to mitigate the antiviral effects of cytokines and chemokines. Our findings illustrate how including viral genomic sequence data in trait-based models can inform host predictions and provide molecular insights into host-virus compatibility to aid wildlife sampling efforts.