Viral zoonotic risk is homogenous among taxonomic orders of mammalian and avian reservoir hosts

We used literature searches to expand an existing dataset of the maintenance cycles of single-stranded RNA viruses (27) to include all 35 RNA and DNA virus families infecting birds and mammals, the hosts that account for the vast majority of zoonoses (8). From a total of 673 virus species, 415 species from 30 families had compelling evidence for transmission by one or more of the 11 taxonomic orders of nonhuman reservoirs (three avian, eight mammalian) which met our inclusion criteria (Materials and Methods). When possible, viral species linked to more than one reservoir group were subdivided into reservoir order-associated viral clades (n = 11 species), and otherwise excluded (n = 2), creating a final dataset of 429 reservoir–virus associations. Data on zoonotic status (i.e., whether a virus had been reported to transmit from nonhumans to humans) were obtained from Olival et al. (9), Woolhouse and Brierley (29), and primary literature searches. We found dramatic variation in the number and diversity of viruses across reservoir groups (Fig. 1). Cetartiodactyls and rodents were the most common reservoirs in our dataset, together accounting for half of the included viruses (50.6%), while anseriforms (waterfowl) were the most poorly represented, with only 1.4% of viruses (Fig. 1A). We quantified the degree of skew in viral diversity among groups of reservoirs using a sliding metric which places varying amounts of emphasis on rare viral lineages (30), here defined as those virus taxonomic lineages that were rarely found in the focal reservoir order. Despite harboring substantial viral species richness, the taxonomic distribution of viruses was notably skewed for some host groups (Fig. 1B). The majority of primate and rodent viruses fell in two to three viral families and the remaining viruses were scattered across families that contained few other species infecting these host groups (on average, two species; Fig. 1 A and B). Other groups including carnivores and odd-toed ungulates had smaller, but more balanced, viral communities (Fig. 1B). Both the number and the proportion of zoonotic virus species appeared to differ among reservoir groups, although the 95% CI of proportions overlapped in most cases (Fig. 1C). The taxonomic distribution of zoonoses within host groups was highly skewed (Fig. 1A). For example, 50% of zoonotic bat viruses were rhabdoviruses and 53.7% of zoonotic rodent viruses were hantaviruses or arenaviruses, reinforcing the need to account for virus taxonomy to avoid detecting effects of reservoirs on zoonotic status that would be unlikely to generalize across bat or rodent viruses (SI Appendix, Fig. S1).

We next used generalized additive mixed models with zoonotic status as a binary response variable to identify and rank the host and/or viral traits that influenced the zoonotic status of viruses. When ranking models using the Akaike information criterion (AIC), the top model explained a total of 52.3% of deviance in zoonotic status (Fig. 2A). This model contained several previously reported effects of viral biology, including transmission by arthropod vectors and exclusively cytoplasm-based replication, which increased the odds of being zoonotic 4.1-fold (log odds = 1.40) and 15.7-fold (log odds = 2.75), respectively (Fig. 2C). However, neither explained >2.2% of the deviance (Fig. 2A). Zoonotic viruses also tended to be more studied, with the number of publications associated with each virus species explaining 15.2% of the deviance (Fig. 2B).

Among effects related to reservoir hosts, zoonotic risk declined with greater phylogenetic distance from reservoirs to primates, but this effect explained only 0.7% of the deviance in the top model and was absent from other competitive models (Fig. 2 A and E). Effects of reservoir order or phylogeny (either alone or crossed with an effect of virus family-level taxonomy) did not occur until the 23rd-ranked model, where they explained <0.01% of the deviance (ΔAIC from top model >16.9; Fig. 2A). The only strongly supported models containing an effect of reservoir host restrained the effect to certain viral families (i.e., a random effect of reservoir nested within virus family, here termed the “virus family-specific reservoir” effect). This effect explained 20.7% of the deviance in zoonotic status in the top model. In particular, primate adenoviruses and retroviruses, bat rhabdoviruses, and rodent picornaviruses were more likely to be zoonotic relative to viruses from the same families associated with other reservoirs (Fig. 2F). Critically however, none of these reservoir groups increased zoonotic propensity across additional viral families, suggesting the absence of generalizable effects of reservoir group on the likelihood of zoonotic transmission (Fig. 2A and SI Appendix, Fig. S2).

We further evaluated whether effects of reservoir hosts on zoonotic status might have been missed due to the inclusion of potentially collinear variables or insufficient statistical power to fit complex models. First, given that study effort understandably increased for zoonotic viruses, including this factor might have disguised true reservoir effects. However, models which did not adjust for virus-related publication counts failed to support reservoir effects; the best of such models achieved similar AIC values with and without reservoir effects (models 177 to 182, Fig. 2A). Second, including virus taxonomy might have reduced the power of our models. However, reservoir effects remained negligible or absent in models which excluded viral taxonomy (SI Appendix, Fig. S3). In contrast, viral family effects were robust to inclusion of general reservoir effects (i.e., those testing for consistent reservoir effects across viral families). When the family-specific reservoir effect was not present, viral rather than reservoir taxonomy was the stronger predictor of zoonotic status (models 23 to 28 in Fig. 2A and SI Appendix, Fig. S3). Thus, our results provide no support for the special reservoir hypothesis, which predicted that viruses associated with certain reservoir groups would be more likely to be zoonotic.

Instead, both the total number of viruses associated with each reservoir group (“viral richness”) and the number of zoonoses matched expectations under the reservoir richness hypothesis. The absence of reservoir effects meant that viral richness was strongly correlated with the number of zoonotic species (R² = 0.88, P < 0.001; Fig. 3), consistent with a similar per-virus zoonotic risk across reservoirs. Notably, a similar correlation was previously observed when measuring viral richness at the host species level (9). A series of negative-binomial generalized linear models showed that—after controlling for greater viral richness in mammals compared to birds—reservoir species richness was correlated with both the number of zoonotic viruses (65.5% of deviance) and the viral richness (56.4% of deviance) of different reservoir orders (Fig. 4 and SI Appendix, Fig. S4). These models outcompeted those including previously identified effects of virus-related research effort and the phylogenetic distance of each reservoir group from primates (9), which when included explained a maximum of 3.1% additional deviance (ΔAIC ≥ 1.09; models 2 to 4 in Fig. 4A). Exchanging the reservoir class effect (i.e., birds versus mammals) for phylogenetic distance to primates performed considerably worse than the top model (ΔAIC = 4.72; model 5 in Fig. 4A). Similar results were observed for models predicting total viral richness (SI Appendix, Fig. S4).

We next assessed how well the predictions of the reservoir richness model matched observations in individual reservoir groups. Highly recognized mammalian reservoir groups including bats and rodents hosted close to the number of zoonoses expected from their species richness (bats: 22 observed and 28 predicted [95% CI: 10.7 to 45.7]; rodents: 41 observed and 42 predicted [95% CI: 10.2 to 72.9]; Fig. 4E). In contrast, lagomorphs (rabbits, hares, and pikas) and diprotodonts (an order of marsupials) hosted both fewer viruses and fewer zoonoses than predicted; however, these were among the least-studied animal groups (Fig. 4E and SI Appendix, Fig. S4E). The only potentially “special” reservoirs identified by this analysis were the cetartiodactyls (even-toed ungulates and whales) and primates, which hosted more zoonoses than would be predicted for mammalian groups with their species richness (Fig. 4E). In the case of primates, this likely reflects the higher zoonotic propensity of primate-borne adeno- and retroviruses identified above (24 zoonotic viruses observed, of which 3 are adenoviruses and 7 are retroviruses; 15 predicted [95% CI: 7.8 to 22.4]). The mismatch was greater for cetartiodactyls (28 observed, 12 predicted [95% CI: 6.1 to 17.3]). However, perhaps due to their close association with humans as domestic livestock, this group also had considerably higher total viral richness than predicted (112 observed and 45 predicted [95% CI: 25.3 to 65.5]; SI Appendix, Fig. S4E). Consequently, the proportion of cetartiodactyl viruses which were zoonotic was unexceptional, inconsistent with the special reservoir hypothesis (Figs. 1 and 3). Strikingly, although mammals had elevated viral and zoonosis richness compared to birds, the proportion of zoonotic viruses was similar between classes, with 9/30 of avian and 115/387 of mammalian virus species being zoonotic (30% and 28.8%, respectively), emphasizing that—regardless of the cause—it is the underlying number of virus species which differs between mammals and birds, not the zoonotic component.

Animal groups have been proposed to differ in their propensity to transmit viruses to humans as a consequence of variation in their life history, physiology, or ecology. By combining a large dataset of reservoir host–virus associations with records describing their histories of human infection, our analysis found no evidence that the taxonomic identity of reservoirs affects the probability that the viruses they harbor are zoonotic. Instead, variation in the number of zoonoses maintained by each reservoir group was consistent with a largely host-neutral model, whereby more species-rich reservoir groups host more virus species and therefore a larger number of zoonotic species. These findings imply the need to reconsider some current approaches to virus discovery, surveillance, and research.

The absence of effects of reservoir host associations on the zoonotic status of viruses is at odds with the idea that conserved traits of certain host groups alter the zoonotic potential of their viruses. If host traits predictably altered zoonotic risk, these effects would be expected to act on multiple viral groups. Instead, we found that reservoir effects were isolated within individual virus families, such that no reservoir group altered the zoonotic risk of viruses across a broad range of viral families. For example, bats are widely considered special reservoirs due to their association with several high-profile zoonoses, including Severe acute respiratory syndrome-related coronavirus (Coronaviridae), Nipah henipavirus (Paramyxoviridae), and Ebola viruses (Filoviridae) (17, 19, 31). However, virus species within those families with a bat reservoir were no more likely to be zoonotic than those transmitted by other hosts, although we acknowledge that this effect would have been difficult to detect for filoviruses given the small number of virus species in this family. While our results do not dispute the existence of distinct features of bat immunity or life history which have been hypothesized to influence viral communities in bats (16, 18, 19, 32, 33), they provide no compelling evidence that these traits translate into an increased probability of bat-associated viruses infecting humans.

Idiosyncratic family-specific reservoir effects could reflect especially zoonotic clades within viral families that are strongly host-associated, sampling biases, or genuine interactions between the biological features of specific viral groups and specific reservoirs that increase zoonotic capability (e.g., differences in viral shedding or tissue tropism). Among rhabdoviruses, a significantly elevated proportion of bat-associated species were zoonotic, but this effect was driven by the rabies-causing genus Lyssavirus, which contained 73% of the 25 zoonotic rhabdovirus species and is strongly bat-associated (35). However, Rabies lyssavirus is also zoonotic when associated with carnivores and bat-associated Rhabdoviruses outside of the Lyssaviruses are not known to be zoonotic, suggesting that traits of this viral genus rather than its reservoirs are the underlying driver of the detected bat effect (SI Appendix, Fig. S3). The putatively heightened frequency of zoonotic rodent-associated picornaviruses might conceivably arise if transmission by rodents selects for viral traits that disproportionately promote zoonotic capability or if picornaviruses have similar zoonotic capability across reservoir groups but heightened ecological overlap favors zoonotic transmission from rodents. Yet, given that our dataset contained only three rodent-associated picornaviruses (all zoonotic) and primates and cetartiodactyls also hosted zoonotic picornaviruses (three out of four and one out of six zoonotic, respectively), a statistical artifact of low sample size cannot be ruled out. In contrast, effects of nonhuman primate reservoirs on the zoonotic potential of adenoviruses and retroviruses may have a biological basis. All primate-associated adenoviruses (three out of three) and 7 out of 12 primate-associated retroviruses were zoonotic, while none of the 15 adenoviruses and 16 retroviruses associated with other reservoirs have been reported to infect humans. This suggests that evolutionarily conserved similarities between nonhuman primates and humans may facilitate zoonotic transmission in these viral families. Yet, our results also highlight limitations of host phylogeny as a generalizable predictor of zoonotic capability since primate associations did not affect the zoonotic status in most of the viral families maintained by nonhuman primates. We hypothesize that the advantages conferred by phylogenetic relatedness are crucial for viral families that are inherently host-restricted (e.g., retroviruses, which must integrate into host genomes) but are less influential for viruses with less specialized infection and replication mechanisms. More generally, the relatively poor performance of phylogenetic distance in predicting either the proportion or the number of zoonoses from different reservoir groups (Figs. 2E and 4 A and B) suggests that the evolutionarily conserved factors that facilitate cross-species transmission within animal orders (25, 36, 37) are unlikely to extend over deeper evolutionary distances.

Our modeling framework allowed us to distinguish the relative contribution of host and virus features to zoonotic status. Consistent effects of viral taxonomy support the conclusion that zoonotic ability is a feature of viral clades rather than host associations (9, 28); however, the underlying viral trait determinants remain unresolved. As shown here (Fig. 2) and elsewhere, transmission by arthropod vectors and replication in the cytoplasm of host cells were linked to zoonotic transmission (9, 38). These findings are potentially explainable by selection for viral generalism arising from the need to replicate in both vertebrate and invertebrate hosts and the avoidance of the additional barriers associated with entering and replicating in the nucleus in novel host species, respectively. Yet, the site of viral replication within cells is invariant within viral families and arthropod-borne transmission is also strongly evolutionarily conserved (27, 39). Therefore, these traits have limited utility for explaining large variation in zoonotic risk within viral families. Identifying whether factors beyond ecological opportunity predict zoonotic status is vitally important to anticipate zoonotic transmission and to estimate the number of viruses which may emerge in the future.

The number of zoonoses scaled positively with the host species richness of animal orders (Fig. 4). Since reservoir species richness did not affect the probability that viruses were zoonotic (Fig. 2A), direct effects of species richness—which might occur if viral maintenance by more species rich host orders facilitates zoonotic transmission by selecting for broader host range—are unlikely (2). Instead, our results support a numerical, host species-neutral explanation: more species-rich animal orders maintain more viruses and hence more zoonoses. Indeed, after controlling for diminished viral richness in birds compared to mammals, key groups including bats and rodents harbored close to the number of zoonoses expected from their species richness (Fig. 4). These results mirror—and potentially explain—observations that the risk of zoonotic emergence is elevated in regions with high species richness (3, 40), since more virus species and hence zoonoses would be expected in species-rich habitats. More generally, our results suggest that the processes which shape viral richness (i.e., extinction, codivergence, and host shifts) do not vary enough among animal orders to create significant differences in the average number of viruses per species across reservoir orders. One implication of the relationship between reservoir species richness and the number of zoonoses is that surveillance efforts aimed at finding potential zoonoses should scale with the species richness of reservoir groups. This recommendation differs from the current practice which is based on the a priori expectation that certain groups (e.g., bats, rodents, and primates) are more likely to maintain zoonotic viruses (7). Analogously, our results underscore the challenges of identifying unknown reservoirs for viruses of importance to human health, since sampling effort would need to scale with local biodiversity, which itself may be uncertain.

Our dataset and analytical approach differed in several epidemiologically important ways from previous studies which suggested variation in viral zoonotic propensity across animal groups (9, 12). First, restricting virus–reservoir associations to the order level allowed us to test hypotheses on the role of broad animal groups on zoonotic origins in the face of widespread uncertainty in the identity of viral reservoirs at the host species level (22, 23, 27). This meant excluding human viruses that occasionally infect nonhumans and all associations between viruses and nonhuman hosts which are not currently considered to be important in natural transmission cycles. While it is conceivable that some excluded host–virus associations have an unrecognized role in transmission, adding 302 previously reported associations which did not meet our criteria did not qualitatively change our results (SI Appendix, Fig. S6). Second, we considered a single reservoir order for most viruses rather than modeling every association of each virus with each infected host species within that order independently. This was critical to avoid potentially spurious effects driven by surveillance effort and multiple observations of the same virus in closely related species. For example, Dengue virus (DENV) has been detected in at least 20 bat species, which comprise 76% of DENV–host associations (9, 12), but nonhuman primates, not bats, are the currently accepted reservoirs of zoonotic DENV outbreaks (41–43). Our dataset therefore included a single primate reservoir since the majority of DENV–host associations would have obscured conclusions on zoonotic origins. Finally, unlike previous studies, we included avian viruses; however, restricting our analysis to mammalian viruses failed to recover reservoir host effects on zoonotic risk (SI Appendix, Fig. S7). One possible explanation for the apparent discrepancy between our findings and those based on shared virus detections could be that heightened surveillance for or susceptibility to zoonoses has led to elevated detections of spillover infections in previously identified “special hosts.” Our results highlight the importance of separating exposure and infection from transmission in future studies which involve reservoirs of infection.

A clear limitation of our approach was that we were unable to consider traits that vary within reservoir groups (e.g., reproductive rates, population size, and geographic range size) which may moderate the baseline zoonotic risk by altering viral richness, transmission dynamics, or ecological contacts of hosts with humans (44). In addition, even at the broad level of host taxonomy we used, knowledge of virus–reservoir relationships and zoonotic capability is incomplete and unevenly distributed among host groups (45). Analogous analyses of the capability of viruses to infect nonhuman hosts could also enable broader understanding of the determinants of cross-species transmission but will require constructing comprehensive datasets of infection histories in these alternative species. Finally, we evaluated only viral richness and whether viruses were reported to infect humans. Whether viruses from different reservoirs differ systematically in their pathogenicity, capacity to transmit among humans, or in their frequency of zoonotic transmission cannot be assessed from our data.

In summary, our analysis suggests that variation in the frequency of zoonoses among major bird and mammal reservoir groups is an emergent property of variation in host and virus species richness. We find no evidence that intrinsic or ecological differences among animal groups increases the number of viruses they maintain or the likelihood that any given virus is zoonotic. Basing public health surveillance and research strategies aiming to identify high-risk viruses on the assumption that some taxonomic orders of hosts are disproportionate sources of zoonoses risks missing important zoonotic viruses while simultaneously reenforcing patterns that may reflect detection biases rather than zoonotic risk.

We thank Dan Haydon and Roman Biek for helpful comments on earlier versions of this manuscript, as well as two anonymous reviewers, whose suggestions greatly improved the clarity of this publication. D.G.S. was supported by a Sir Henry Dale Fellowship, jointly funded by the Wellcome Trust and Royal Society (102507/Z/13/Z) and a Wellcome Senior Research Fellowship (217221/Z/19/Z). Additional funding was provided by the Medical Research Council through programme grants MC_UU_12014/8 and MC_UU_12014/12. Reservoir silhouettes were obtained from phylopic.org and were created by Matt Martyniuk (Anseriformes), Brian Gratwicke and T. Michael Keesey (Carnivora), and Sarah Werning (Lagomorpha), used with permission (https://creativecommons.org/licenses/by/3.0). Additional silhouettes by Andrew Butko (Passeriformes; https://creativecommons.org/licenses/by-sa/3.0) and Tamara L. Clark (Perissodactyla; https://creativecommons.org/licenses/by-nc-sa/3.0). All other silhouettes are in the public domain.