Abstract
Predicting which novel microorganisms may spill over from animals to humans has become a major priority in infectious disease biology. However, there are few tools to help assess the zoonotic potential of the enormous number of potential pathogens, the majority of which are undiscovered or unclassified and may be unlikely to infect or cause disease in humans. We adapt a new biological machine learning technique - phylofactorization - to partition viruses into clades based on their non-human host range and whether or not there exist evidence they have infected humans. Our cladistic analyses identify clades of viruses with common within-clade patterns - unusually high or low propensity for spillover. Phylofactorization by spillover yields many clades of viruses containing few to no representatives that have spilled over to humans, including the families Papillomaviridae and Herpesviridae, and the genus Parvovirus. Removal of these non-zoonotic clades from previous trait-based analyses changed the relative significance of traits determining spillover due to strong associations of traits with non-zoonotic clades. Phylofactorization by host breadth yielded clades with unusually high host breadth, including the family Togaviridae. We identify putative life-history traits differentiating clades’ host breadth and propensities for zoonosis, and discuss how these results can prioritize sequencing-based surveillance of emerging infectious diseases.