Predicting the global mammalian viral sharing network using phylogeography

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

org/10.1101/732255 doi: bioRxiv preprint many of which are mammal-virus interactions 29 . Pairs of species that share viruses in EID2, 165 but which were not in our training dataset (see Methods), had a much higher mean sharing 166 probability in our predicted network (20% versus 5%; Figure 2A ). In addition, more central 167 species in the predicted network were more likely to have been observed with a virus, 168 whether zoonotic ( Figure 2B ) or non-zoonotic ( Figure 2C ), implying that the predicted 169 network accurately captured realised potential for viral sharing and zoonotic spillover. This The high predicted centrality of known hosts may be due partly to selective sampling (i.e., 189 viral researchers are more likely to sample wide-ranging and common host species that also 190 share viruses with many other species 10,20 ). This possibility is supported by the increased 191 degree centrality for species that appear in both EID2 and our dataset rather than in only one 192 of the two, as these species are presumably more well-known ( Figure 2C ). Similarly, while 193 we believe that our model was successful at accounting for variation in host-level diversity 194 and study effort that influences network topology (see above; Figure The copyright holder for this preprint (which was not peer-reviewed) is the . https://doi.org/10.1101/732255 doi: bioRxiv preprint viruses, artificially increasing the likelihood of detecting these viruses in the same region 199 compared to a geographically random sampling regime. Moreover, when a mammal species 200 (e.g., a bat) is found with a focal virus (e.g., an ebolavirus), it is logical for researchers to then 201 investigate similar, closely related species in nearby locales 33 . These sampling approaches 202 could disproportionately weight the network towards finding phylogeographic effects on viral 203 sharing probability. However, it is highly encouraging that our model predicted patterns in 204 the external EID2 dataset, which was constructed using different data compilation methods 205 but also comprises global data covering several decades of research 29 . In sum, we believe that 206 our approach is a conservative method for minimising the biases inherent in the data. The 207 knowledge that the observed mammalian virome is biased ultimately calls for more uniform 208 viral sampling across the mammal class and increased coverage of rarely-sampled groups, 209 lending support to ongoing efforts to systematically catalogue mammalian viral diversity 3 . predicted viral sharing probability from our model were more likely to be observed sharing a virus in 216 the independent EID2 dataset. This comparison excludes species pairs that were also present in our 217 training data. B: species that hosted a zoonotic virus in our dataset had more viral sharing links in the 218 predicted all-mammal network than those without zoonotic viruses. C: species that had never been 219 observed with a virus have fewer links in the predicted network than species that hosted viruses in the 220 EID2 dataset only, in our training data only, or in both. The y axis represents viral sharing link 221 number, scaled to have a mean of 0 and a standard deviation of 1 within each order for clarity. Figure 222 SI4 displays these same data without the within-order scaling. 223 The copyright holder for this preprint (which was not peer-reviewed) is the . https://doi.org/10.1101/732255 doi: bioRxiv preprint interactions in addition