Predicting the global mammalian viral sharing network using phylogeography

Gregory F Albery; Evan A Eskew; Noam Ross; Kevin J Olival

to searching for species-level traits associated with high viral 257 diversity. 258 259 Encouragingly, our network showed predictable scaling laws similar to those of other known 260 ecological networks 37 . Viral link numbers in within-order subnetworks (e.g., between 261 different bat species) correlated strongly with species diversity within each order (R 2 =~0.85), 262 following a power law with a Z value of ~0.8 ( Figure SI6 ). Similarly, out-of-order links (e.g., 263 between a bat and a rodent) scaled linearly with the product of the species richness of both 264 orders ( Figure SI7 ). We investigated the predictive potential of our model by iteratively selecting all but one of 318 the known hosts for a given virus, then using the predicted sharing patterns of the remaining 319 hosts to identify how the focal (removed) host was ranked in terms of its sharing probability. 320 In practical terms, these species-level rankings could set sampling priorities for public health 321 efforts seeking to identify hosts of a novel zoonotic virus, where one or more hosts are 322 already known. Across all 250 viruses, the median ranking of the left-out host was 72 out of a 323 potential 4196 mammals (i.e., in the top 1.7% of potential hosts). To compare this ranking to 324 alternative heuristics, we examined how high the focal host would be ranked using simple 325 ranked phylogenetic relatedness or spatial overlap values alone (i.e., the most closely-related, 326 followed by the second-most-related, etc.). Using this method, the focal host was ranked The copyright holder for this preprint (which was not peer-reviewed) is the . https://doi.org/10.1101/732255 doi: bioRxiv preprint 337 We observed substantial variation in our model's ability to predict known hosts among 338 different viruses. For example, the correct host was predicted first in every iteration for 7 339 viruses and in the top 10 hosts for 42 viruses. Results for 128 viruses had the focal host 340 falling within the top 100 guesses, and for only 6 viruses were the model-based host searches 341 worse than chance (focal host ranked lower than 50% of all mammals in terms of sharing 342 probability). We used this measure of viral sharing "predictability" to investigate whether 343 certain viral traits affected the ease with which phylogeography predicted their hosts. Viruses 344 with broad host phylogenetic ranges, most notably Ebola virus, challenge reservoir prediction 345 efforts since many more species must often be sampled before identifying the correct host(s). 346 To investigate whether the predictive strength of our model was limited for viruses with 347 broad host ranges and/or other viral traits, we fitted a linear mixed model (LMM) which 348 showed a strong negative association between viruses' known phylogenetic host breadth and 349 the predictability of focal hosts (model R 2 =0.70; host breadth R 2 =0.67; Figure SI9 ). This 350 association demonstrates, unsurprisingly, that predicting the hosts of generalist viruses is 351 intrinsically difficult using our method. This adds a potential limitation to the applicability of 352 our network approach, given that zoonotic viruses commonly exhibit wide host ranges 2,5 . A In summary, we present a simple, highly interpretable model that predicted a substantial 366 proportion of viral sharing across mammals and is capable of identifying species-level 367 sampling priorities for viral surveillance and discovery. It is worth noting th