Centrality in primate–parasite networks reveals the potential for the transmission of emerging infectious diseases to humans

Results and Discussion

We obtained 6,304 records of parasite and pathogen from the Global Mammal Parasite Database, representing 300 parasite species (including viruses, bacteria, helminths, protozoa, arthropods, and fungi) infecting 140 primate hosts (23). We built a presence/absence bipartite network linking each primate species with their parasites. We then obtained the weighted unipartite projection, in which each node was a primate species that was connected to other primates on the basis of the number of parasite species they shared. Centrality was estimated using five metrics: strength, Opsahl degree, eigenvector, betweenness, and closeness (14, 24). Each of these metrics captures different and complementary aspects of centrality (SI Appendix, SI Materials and Methods). Whereas strength, Opsahl degree, and eigenvector centrality are more related to the pattern of direct cosharing of parasites with the rest of the hosts in the network, betweenness and closeness capture aspects of indirect sharing through other species across the entire network (8, 12, 14, 15).

Centrality may also be affected by the number of studies that have been conducted on each primate species (i.e., sampling effort), because more heavily sampled primates may seem to have more parasites (7, 19). Thus, to get an accurate estimate of centrality requires controlling for variation in sampling effort among primates. We used the number of citations (=number of studies) as an estimate of sampling effort for each primate (25). At present, different methods of controlling for sampling effort are available, and it is unclear which method should be used. To ensure that our findings are robust, we therefore used four different approaches (I–IV) of controlling for variation in sampling effort: I, including sampling effort in the computation of centrality estimates by up-weighting those primates that have more parasites than expected given their sampling effort; II, including sampling effort in the computation of centrality estimates by up-weighting the least sampled primates and down-weighting the most sampled primates; III, obtaining the residuals by regressing centrality on measures of sampling effort (25); and IV, using only the subset of 38 well-sampled primates from ref. 7. A full explanation of the four weighting methods can be found in SI Appendix, SI Materials and Methods. In addition, we ran all analyses in methods I–III including sampling effort as a covariate. Results were consistent across the methods. For the sake of brevity, we provide results from method I in the main text (with other results provided in SI Appendix, SI Results).

All five centrality indices showed positive correlations (0.311 < r < 0.998, P < 0.0001 in all cases, n = 140 primates; SI Appendix, Tables S1, S6, and S13), indicating that they detected similar primate species as most central. Betweenness was the centrality index least correlated with the other indices (0.311 < r < 0.690, P < 0.0001). A main characteristic of betweenness is that, irrespective of the network, many nodes attain a score of zero, resulting in a zero-inflated distribution. This characteristic is thus responsible for the weaker correlation between betweenness and the other centrality metrics. To get a clearer picture of the effect of centrality on primate transmission ability, we obtained a composite index that integrates the different and complementary aspects of the five centrality metrics by performing a principal component analysis on the centrality index correlations. This analysis found a single factor explaining between 77% and 84% of the variance of the indices, depending on the weighting method (SI Appendix, Tables S2, S7, and S14). This factor, hereafter called “centrality,” is used in the subsequent analyses (results for different measures of centrality are provided in SI Appendix, Results). The top 10 most central primates were, in descending order, Papio ursinus, Saimiri sciureus, Alouatta seniculus, Chlorocebus aethiops, Piliocolobus badius, Papio cynocephalus, Papio anubis, Gorilla gorilla, Pan troglodytes, and Macaca sinica (Fig. 1A).

Several traits have been found to affect parasitism, such as population density, geographic range size, group size, body mass, and diet (19). We explored whether these traits also covary with centrality of the primates using phylogenetic generalized least squares (PGLS) models (26) with the primate consensus dated tree from the 10kTrees Project (27). In a model with multiple predictor variables, we found that primates with denser populations living in larger groups and having broad distributions were more central [population density: estimate ± 1 SE = 0.32 ± 0.15, t = 2.10, P = 0.04; geographic range size: estimate ± 1 SE = 0.28 ± 0.14, t = 1.95, P = 0.05; group size: estimate ± 1 SE = 0.36 ± 0.17, t = 2.08, P = 0.04, R² = 0.490; PGLS with phylogenetic signal (λ) = 0.138, P = 0.06 compared to a model with λ = 0, n = 122 primates]. This outcome was similar for each of the centrality indices considered independently (SI Appendix, Table S3). We suggest that primates with a large geographic range size and denser populations will have a higher probability of interspecific encounters with other primate species, enhancing the conditions for parasite sharing.

Inferring when, where, and to whom parasites are transmitted are key questions in disease ecology (28). Risk of transmission to humans has been explored by determining the parasites shared both by humans and wild primates (19). Those primate species sharing more pathogens with humans have a higher probability of behaving as zoonotic reservoirs owing to host shifting observed in most primate pathogens (29). To infer the potential ability of different primate species to act as zoonotic reservoirs, we determined the similarity between the parasite communities harbored by primates and humans. Human parasite community and EIDs were obtained from Taylor et al.’s database (30), comprising 1,415 zoonotic and nonzoonotic species. Primates showed variation in the degree to which their parasite community corresponds to that of humans. Some apes, such as G. gorilla (Bray index = 0.16), P. troglodytes (0.15), and Pongo pygmaeus (0.12), had parasite communities similar to those in humans. This outcome is expected owing to the significant phylogenetic conservatism observed in ecological interactions in general (31). Nevertheless, several Old World monkeys also showed similar levels of sharing by this measure, including C. aethiops (0.18), P. cynocephalus (0.13), and P. ursinus (0.11). This finding indicates that in addition to apes, other primates may act as sources of human diseases (7), especially Old World monkeys that show ecological overlap with humans, such as terrestrial substrate use. We found that more central primates had parasite communities that were more similar to humans (estimate ± 1 SE = 0.011 ± 0.002, t = 5.60, P < 0.0001, R² = 0.750; PGLS with λ = 0.150, P = 0.001, n = 129 primates; Fig. 1B). Again this outcome was consistent across different centrality indices and for most ways of controlling for sampling effort (SI Appendix, Tables S4, S8, and S10). We found the same significant relationship when using only the well-sampled primate species (Fig. 2A and SI Appendix, Table S15).

Identifying hosts most frequently associated with EIDs may predict potential pathogen reservoirs and prevent future disease outbreaks (4, 32). We found that, after controlling for sampling effort and phylogeny, the number of infectious organisms found in primates and identified as emerging in humans (EID richness) was significantly related to the centrality of the primates in a multivariate model (Fig. 1C and Table 1). The association with EID richness was significant and positive for each of the five centrality indices, irrespective of the methodology used to control for variation in sampling effort (SI Appendix, Tables S5, S9 , S11, and S16), and maintained when using only the 38 well-sampled primates (Fig. 2B). To assess how phylogenetic uncertainty affects these results, we repeated the analysis using 100 different trees obtained from 10kTrees (27) and found that the association between EID richness and centrality was consistent across all of the phylogenetic trees (SI Appendix, Table S17). In addition, as noted above, centrality is predicted by several ecological and phenotypic traits, such as geographic range size and population density. However, we found that only parasite richness and geographic range size covary with EID richness (Table 1).

Collectively, these analyses suggest that nonhuman primates that are more central in the primate–parasite network also have a higher probability of sharing infectious diseases with humans, including those identified as EIDs. As expected, apes were found among these most-central primates. They are obvious candidates to share infectious organisms with us owing to their phylogenetic proximity, and genetic evidence suggests that spillover has occurred from these species to humans [e.g., falciparum malaria from gorillas (33) and SIV/HIV from chimpanzees (34)]. Nevertheless, we also found that other central non-ape primates, such as baboons and vervet monkeys, are infected by many parasites identified as EIDs in humans.

Our findings suggest that centrality may help to detect risks that might otherwise go unnoticed, and thus to predict disease emergence in advance of outbreaks—an important goal for stemming future zoonotic disease risks (6). For example, with information on the parasites of different host species, effort could focus on sampling individuals of more central species for viruses with higher mutation rates—such as RNA viruses—that have greater potential for spillover to humans and other wildlife. It is important to emphasize that this approach focuses on all parasites and pathogens, regardless of whether they occur in humans; simply having the ecological network is what matters, rather than details on parasites identified as “emerging” or shared with humans. By demonstrating that ecological network characteristics from primates also predict the presence of EIDs in humans, our results validate the use of similar network approaches for predicting future disease emergence in other animal groups that are important sources of EIDs, such as ungulates, rodents, bats, and carnivores.

In conclusion, we provide an approach based on network theory to detect the probability of EID transmission from wildlife to humans. Centrality could therefore become a useful tool for allocating resources to prevent future emerging diseases.

Materials and Methods

Detailed methods, R scripts, and dataset are in SI Appendix, SI Materials and Methods and Dataset S1.

Supplementary Material