Using phylogeographic link-prediction in primates to prioritize human parasite screening

Abstract

Objectives.

The ongoing risk of emerging infectious disease has renewed calls for understanding the origins of zoonoses and identifying future zoonotic disease threats. Given their close phylogenetic relatedness and geographic overlap with humans, non-human primates (NHPs) have been the source of many infectious diseases throughout human evolution. NHPs harbor diverse parasites, with some infecting only a single host species while others infect species from multiple families.

Materials and Methods.

We applied a novel link-prediction method to predict undocumented instances of parasite sharing between humans and NHPs. Our model makes predictions based on phylogenetic distances and geographic overlap among NHPs and humans in six countries with high NHP diversity: Columbia, Brazil, Democratic Republic of Congo, Madagascar, China and Indonesia.

Results.

Of the 899 human parasites documented in the Global Infectious Diseases and Epidemiology Network (GIDEON) database for these countries, 12% were shared with at least one other NHP species. The link prediction model identified an additional 54 parasites that are likely to infect humans but were not reported in GIDEON. These parasites were mostly host generalists, yet their phylogenetic host breadth varied substantially.

Discussion.

As human activities and populations encroach on NHP habitats, opportunities for parasite sharing between human and non-human primates will continue to increase. Our study identifies specific infectious organisms to monitor in countries with high NHP diversity, while the comparative analysis of host generalism, parasite taxonomy, and transmission mode provides insights to types of parasites that represent high zoonotic risk.

Introduction

Infectious diseases have been a major force in shaping human history (Snowden, 2019), inequality and health disparities (Bonds, Keenan, Rohani, & Sachs, 2010), and the human genome (Fumagalli et al., 2011; Karlsson, Kwiatkowski, & Sabeti, 2014). These effects are not simply a part of our past, with outbreaks over the past 20 years renewing interest in disease ecology and evolution and catalyzing discussions on best practices for managing future disease threats (Petrovan et al., 2021). Global connectivity facilitates the rapid spread of highly transmissible diseases across international borders, and the international community has demonstrated low preparedness in the face of pandemic threats (Coccia, 2021). Controlling disease outbreaks at the local level before they escalate to a pandemic is a major challenge that requires an intimate understanding of risk factors that promote disease emergence.

Most major pandemics and epidemics in humans originated as zoonoses, and many of these can be traced back to one or a small number of zoonotic spillover events (Devaux, Mediannikov, Medkour, & Raoult, 2019; Ellwanger & Chies, 2021; Pépin, 2021; Piret & Boivin, 2020; Plowright et al., 2017). A successful cross-species spillover requires the alignment of ecological, physiological, and immunological conditions involving the host(s) and a parasite, which we define ecologically as any organism that benefits from living within or on a host at the host’s expense (Plowright et al., 2017). In a viral spillover, for example, a person must come into contact with infectious material from another host species, which requires ecological overlap of humans and infected animals in space and time. The virus must also be capable of entering the new host’s cells, replicating in those cells, and then successfully transmitting to new hosts, all before the host immune system clears the infection. Greater phylogenetic relatedness between a pair of hosts increases the probability of cross-species transmission (Cooper, Griffin, Franz, Omotayo, & Nunn, 2012; Davies & Pedersen, 2008; Poulin, 2010), while a higher mutation rate provides more opportunities for a parasite to adapt to a novel host (Loverdo & Lloyd-Smith, 2013; Woolhouse, Haydon, & Antia, 2005). The multiple factors that lead to spillover and then sustained transmission in human populations are complex, making it extremely challenging to predict zoonotic outbreaks (Plowright et al., 2017). However, their frequency is increasing as global anthropogenic change exerts mounting pressure on wildlife (Daszak, Cunningham, & Hyatt, 2001; Marani, Katul, Pan, & Parolari, 2021).

Disease surveillance and risk mitigation at the expanding human-wildlife interface are essential for early-stage control of zoonotic pathogens, yet resources for such efforts are limited. Prioritizing hosts and parasites for targeted sampling remains a major challenge. Existing data on host-parasite interactions can inform surveillance priorities, but a paucity of data on the vast majority of parasite species leads to uncertainty in inferences drawn from current datasets (Carlson, Dallas, Alexander, Phelan, & Phillips, 2020; Gibb et al., 2021; Poulin & Keeney, 2008).

A variety of modeling approaches have emerged in recent years to predict the zoonotic risk and epidemic potential of understudied parasites. These approaches, termed “zoonotic risk technology” by Carlson et al. (2021), include link prediction models—a suite of network imputation methods that can be used to identify likely host-parasite interactions that are absent from current datasets (Bartomeus et al., 2016; Dallas, Park, & Drake, 2017; Elmasri, Farrell, Davies, & Stephens, 2020; Farrell, Elmasri, Stephens, & Davies, 2022; Morales-Castilla, Matias, Gravel, & Araújo, 2015). Such absences are often thought to occur when a parasite (i.e., a link) is undocumented due to insufficient sampling (Nunn, Altizer, Jones, & Sechrest, 2003; Teitelbaum et al., 2020). Link prediction methods can help fill in these absences and highlight host-parasite combinations that have a high probability of occurring in the future. This framework could predict missing links to humans, with links representing infectious organisms in wildlife that are currently undocumented in humans or have not yet successfully transmitted to humans with onward spread in human populations.

Link prediction models can derive probabilities of parasite occurrence in a new species based on various predictors. The structure of the network itself can inform the addition of missing host-parasite links, but this method can exacerbate sampling bias if network properties are influenced by uneven sampling (Becker et al., 2022; Elmasri et al., 2020; Farrell et al., 2022). Alternatively, link prediction models can use ecological traits as predictors, based on the assumption that ecological similarity of host species increases the likelihood of transmission between them or via shared habitats or behaviors (Becker et al., 2022; Cooper, Kamilar, & Nunn, 2012; Dallas et al., 2017). Such models have shown high accuracy in imputing parasites of small mammals (Dallas et al., 2017) and bats (Becker et al., 2022), but they depend on the availability of extensive ecological trait data. In the absence of ecological data, the evolutionary relatedness of hosts, measured by phylogenetic distance, has strong predictive power in models of host-parasite interactions (Elmasri et al., 2020; Farrell et al., 2022; Morales-Castilla et al., 2015). Geographic overlap has also been shown to predict parasite sharing, as expected because spillover requires spatial overlap of organisms (Davies & Pedersen, 2008). Though the covariates on which they rely may vary, all link prediction models are affected by data limitations, and evaluation of prediction results is necessary to arrive at relevant zoonotic risk assessments that can inform surveillance priorities (Carlson et al., 2021).

Here, we apply a novel link-prediction method to identify undocumented parasites shared by humans and non-human primates (NHPs). In addition to rodents and bats, NHPs are at high risk for harboring emerging diseases that may infect humans (Alexander et al., 2018; Devaux et al., 2019). Environments that facilitate consistent spatiotemporal overlap between humans and NHPs, including urban areas and tourist destinations, can increase opportunities for disease-relevant contact (Ahmed et al., 2019; Devaux et al., 2019). In addition, communities in some parts of rural Africa, Asia, and South America traditionally hunt NHPs, which provides an essential source of protein but places these human populations on the front lines of zoonotic disease transmission (Booth et al., 2021; Wolfe, Daszak, Kilpatrick, & Burke, 2005). Notable instances of disease transmission from NHPs to humans include yellow fever virus, which is maintained in sylvatic and urban NHP populations and transmitted via the vector Aedes aegypti (Carrington & Auguste, 2013; Hanley et al., 2013), and the human immunodeficiency viruses (HIV-1 Group M and HIV-2), which evolved from simian immunodeficiency viruses (SIVs) found in Pan troglodytes and Cercocebus atys, respectively (Gao et al., 1999; Hahn, Shaw, De Cock, & Sharp, 2000; Pépin, 2021).

The novel link-prediction method presented here builds upon the model presented by Elmasri et al. (2020) and Farrell et al. (2022) by incorporating both phylogenetic and geographic predictors to estimate undocumented NHP-parasite interactions relevant to human health. We apply this model at a regional scale to bipartite networks consisting of (1) humans and NHP species known to have high rates contact with humans due to hunting, tourism, and urban cohabitation, and (2) their parasites, where parasite refers to both micro-parasites, such as viruses, bacteria, fungi and protozoa, and macro-parasites, such as helminths and arthropods. Farrell et al. (2022) applied phylogeographic link prediction to a global network, but we focus on regional networks to obtain a manageable list of missing links, allowing us to evaluate each imputed parasite by searching the literature for existing evidence of zoonotic and epidemic risk. Finally, we investigate taxonomic and phylogenetic host generalism, parasite type, and transmission mode of these parasites to better understand traits that increase risk of spillover to humans.

We predict that the majority of parasites imputed to infect humans through link prediction will be generalists that are observed in multiple host families or broader (a taxonomically defined measure of host generalism) and infect more distantly-related species than expected by a null distribution (a phylogenetically defined measure of generalism). Previous studies in NHPs have found that their viruses infect a wider range of hosts than other groups of parasites, such as protozoa or helminths (Pedersen et al. 2005; Cooper et al. 2012). Thus, we predict that humans and NHPs will be more likely to share viruses. We also expect that vector-transmission will characterize parasites imputed to infect humans, again based on previous findings that these vector-transmission modes predict a wider host range in parasites of NHPs (Pedersen, Altizer, Poss, Cunningham, & Nunn, 2005) and of mammals (Park et al., 2018). Conversely, we predict a lower likelihood of shared helminth parasites and parasites transmitted via intermediate hosts due to the relatively host-specific nature of these groups (Cooper, Griffin, et al., 2012; Pedersen et al., 2005).

Methods

Parasite Data Collection

NHP parasite data were downloaded from the Global Mammal Parasite Database (GMPD), a resource of parasites documented in wild NHP hosts (Nunn & Altizer, 2005; Stephens et al., 2017). We focused on NHPs with the highest risk of transmitting disease to humans by including NHP species that met one of three criteria. First, all great apes were included due to their close phylogenetic relationship with humans and popularity in tourist attractions (Sharp, Rayner, & Hahn, 2013). Second, we included all NHP species that dwell in urban habitats alongside dense populations of humans (Santini et al., 2019). Third, we included all NHP species subject to routine hunting and trapping pressure according to the IUCN (2021), which we identified by searching iucnredlist.org for NHP species under the threat category “Hunting and trapping of terrestrial animals” and the use and trade categories “Food” and “Medicine.” We chose to focus on NHP species subject to the highest anthropogenic pressure in order to prioritize the parasites of hosts for which spillover to humans is a more imminent risk.

Parasite data for high-risk NHPs were included if the NHP and parasite were identified to the species level. We ensured that NHP species names in the GMPD matched the IUCN taxonomy reported by Estrada et al. (2017). We filtered out any records in which a parasite was searched for but not found (i.e., prevalence = 0). Parasite attribute data on taxonomic host specificity, type, and transmission mode were also downloaded from the GMPD. Parasites were classified to broad taxonomic categories of bacteria, protozoa, viruses, arthropods, fungi, or helminths. Though we included all parasite types in our analysis, bacteria, arthropods, and fungi are severely under sampled in primates (Cooper & Nunn, 2013); this should be considered when interpreting results for these parasite groups.

Transmission mode was classified as close contact, sexual contact, non-close (environmental), vector, and/or via an intermediate host (Pedersen et al., 2005). Multiple transmission modes were possible for each parasite species. Close contact transmission describes parasites that require close proximity for transmission, including through grooming, biting, scratching, or other physical contact. Sexual transmission describes close-contact transmission that specifically involves sexual contact. Non-close transmission involves transmission through shared environments, including by fomites, water, or soil. Vector-borne parasites are transmitted through biting arthropods that actively search for hosts, such as mosquitoes, fleas, lice, or ticks. Transmission by intermediate hosts describes parasites with complex life cycles (excluding vector-borne parasites) or trophic transmission.

Host generality, defined as the breadth of hosts that a parasite infects, can be measured in multiple ways (Park et al., 2018). We used three metrics to investigate the host generality of parasites found in NHPs. First, we used taxonomic host generality scores from the GMPD that were developed by Pedersen et al. (2005). These integer scores represent the maximum taxonomic breadth of all vertebrate hosts in which evidence of a parasite was detected and range from one through five, representing whether a parasite is documented in (1) a single host species, (2) hosts in one genus, (3) hosts in one family, (4) hosts in one order, or (5) hosts in multiple mammalian orders and possibly non-mammals. Generality scores are regularly revised to incorporate new data on a parasite’s host range, with current scores assigned in 2020 or later.

As a second measure of host generality, we used the full GMPD dataset to calculate the mean phylogenetic distance between all NHP hosts that the parasite infects (MPD). The MPD of each parasite represents the average phylogenetic “distance” between its hosts in millions of years, with smaller MPD values characterizing parasites that mostly infect closely-related hosts; this measure avoids problems with variation in the phylogenetic ages of different parasite taxonomic groups. To calculate MPD, we downloaded a posterior distribution of 100 phylogenetic trees containing all NHP species from VertLife, which provides a Bayesian inference of phylogeny for multiple groups of vertebrates (Upham, Esselstyn, & Jetz, 2019). From this posterior distribution of phylogenies we used the phytools package to obtain a consensus phylogeny by computing the mean length of an edge across all trees, with absent edges receiving a value of zero (Revell, 2012). We then calculated the MPD of each parasite using the picante package (Kembel et al., 2010) in R.

As a final measure of host generality, we controlled for varying host abundances in the GMPD and calculated a standardized metric to enable comparison of MPD values across parasites. We used picante to calculate the standard effect sizes for the mean phylogenetic distances (SES.MPD; Kembel et al., 2010). SES.MPD values were obtained by comparing the raw PD values between host pairs to the expected PDs obtained from a null distribution, which was generated by randomizing the host-parasite interaction matrix 1000 times with an independent swap algorithm (Gotelli, 2000). Host and parasite occurrence frequencies were maintained in the null distribution, and the PD of each host pair was weighted by their relative frequencies in the GMPD. The resulting z-scores indicate whether a parasite’s MPD is significantly different from the null expectation, in which case the parasite is considered a generalist (z > 1.96) or a specialist (z < −1.96).

Human parasite data were compiled from the Global Infectious Disease and Epidemiology Network (GIDEON), a database for infectious diseases in humans at a global scale (Yu & Edberg, 2005). Zoonotic link-predictions are often conducted on global host-parasite datasets, yet effective prevention and containment of emerging infectious disease should occur at the regional level, ideally before an outbreak reaches epidemic or pandemic status. In order to make targeted predictions relevant to smaller-scale efforts, we identified six countries on multiple continents that contain a high diversity of NHP species by overlaying a map of World Countries (Esri 2019) with NHP geographic range polygons (IUCN 2021) in ArcMap 10.7 and determining total NHP species richness within each country. These countries included Madagascar (54 NHP spp.), Brazil (39 NHP spp.), Indonesia (30 NHP spp.), Democratic Republic of the Congo (DRC, 25 NHP spp.), China (17 NHP spp.), and Colombia (17 NHP spp.). We examined all references under the “Relevant Diseases’’ section of the GIDEON profile for each country. GIDEON organizes data by disease rather than the organism that causes disease. Thus, we excluded disease-only data and retained only records in which the parasite causing disease was identified to the species-level. All parasite names were corrected to match those in the GMPD. If a parasite was not found in the GMPD, we used the NCBI taxonomy browser (Sayers et al., 2019; Schoch et al., 2020) and the ICTV taxonomy for viruses (Lefkowitz et al., 2017) to assign the most widely accepted species name.

Link Prediction Model

To predict undocumented parasite sharing between human and NHPs, we constructed a Bayesian phylogeographic link-prediction model using the HPpredictions package in R (see Elmasri et al. 2020). The phylogeny-only version of this model takes a binary host-parasite interaction matrix and a phylogenetic distance matrix and uses phylogenetic relatedness to predict parasite sharing. We used an extended version of this model that adds a geographic distance matrix and predicts host-parasite sharing with phylogenetic relatedness and geographic overlap (Figure 1; S1), both of which are known predictors of parasite sharing in NHPs (Davies & Pedersen, 2008). We did not include ecological traits as predictors in the model because ecological trait data are subject to uneven sampling in NHP datasets, and key traits, such as population density, are difficult to generalize to humans.

Figure 1. Imputing missing interactions using link prediction.

Our updated Elmasri et al. (2020) phylogeographic model takes (a) a binary matrix of observed host-parasite interactions and uses (b) phylogenetic relatedness and (c) geographic range overlap to identify the probability of undocumented interactions. Depicted here are the parasites, phylogeny, and geographic ranges of humans (indicated in red) and NHPs found in Colombia. The human range is limited to Columbia, while the NHP ranges often include neighboring countries. Geographic overlap with humans is the area of the NHP range that overlaps the country of interest, while overlap amongst NHPs makes use of their entire ranges, including outside the country

We obtained pairwise phylogenetic distances from the consensus phylogeny of humans and NHPs, as described above (Upham et al., 2019). Geographic range shapefiles for each NHP were downloaded from the IUCN spatial database (IUCN 2021). The “geographic range” for humans was defined by the boundaries of each country (Esri). We used the sp package in R to project NHP ranges into the EPSG 3035 projection, which preserves polygon areas across the globe (Bivand, Pebesma, & Gomez-Rubio, 2013). We then created a geographic overlap matrix for each country by calculating the area of intersection between that country and the ranges of all NHPs found within its boundaries, as well as the intersection between the entire range of each pair of NHPs. By incorporating range overlap into our phylogeographic model, we assume that parasites found in a host are evenly distributed throughout its range. Phylogenetic distances and area of geographic overlap occur on different scales; thus, to incorporate them simultaneously into the prediction model, we standardized the values in each matrix to z-scores with a mean of zero and standard deviation of one. Area of geographic overlap is a measure of similarity, but the Elmasri et al. (2020) model requires a matrix of non-negative distances. We transformed the similarity values into distances by subtracting the z-score of each value from one. To meet the requirements of having only positive elements in the distance matrix, we added a constant to ensure all values remained positive, with a minimum value of 0.01.

We scaled the phylogenetic distance matrix to account for branch-length uncertainty by applying a single-parameter, early-burst evolutionary model (Elmasri et al., 2020; Harmon et al., 2010). Model performance was measured by the area under the receiver operating characteristic (ROC) curve (AUC; Elmasri et al. 2020). We ran the network imputation model with 2,000 iterations and implemented five-fold cross validation to test the posterior predictive strength of the model. We obtained a ROC curve by averaging the true positive and false positive rates across all folds (Elmasri et al., 2020). We also derived the ranked posterior distributions for the relative probabilities of each host-parasite interaction. The model assigned a cutoff relative probability threshold that maximized the AUC and produced an output dataset containing all imputed edges with a probability value above this threshold.

From the imputation results list, we pulled out parasites shared by humans and NHPs. To corroborate model performance and further evaluate risk of individual parasites to human populations within each country, we searched the literature for existing evidence beyond the GIDEON database on the zoonotic threat level of each parasite predicted to infect humans. We conducted a topic search of the PubMed and ScienceDirect for publications from 1900 to the present. We used a combination of key words that included the genus and species of each parasite (including synonyms), the disease name, and “human infection OR primate infection”. We also consulted https://www.cdc.gov/ Center for Disease Control and Prevention (CDC) and Beran et al. (2019) for additional information.

To understand spillover risk more broadly, we investigated whether certain traits were more likely to occur among NHP parasites predicted to infect humans. We examined the taxonomic and phylogenetic host generality of predicted parasites with the expectation that parasites with wider host ranges would represent a greater spillover risk to humans. We also investigated the associations between parasite taxonomic categories and transmission modes to determine whether certain parasite types (e.g. viruses) and transmission modes (e.g. vector-borne) were more likely to occur among model predictions.

Results

We compiled data on 98 high-risk primate species and 1,131 parasite species in the six countries of interest. As expected given greater effort to characterize parasites of humans, most of these parasites (899) were documented in humans. Of these human parasites, 12.2% were shared with at least one other NHP species. Humans shared the highest number of parasite species with NHPs in the DRC (46 parasites) and the lowest in Madagascar (5 parasites).

The link prediction model imputed a total of 54 parasites that are undocumented in GIDEON in the six countries. Performance of the imputation model varied across datasets for the six countries, as did the number of imputed interactions above the threshold cutoff values (Table 1). DRC had the highest number of parasites imputed to infect humans (25 parasites), followed next by China (18 parasites). The AUC values indicated high model accuracy for all countries except China (Table 1). Only 3.6% of the 560 parasites reported in China were observed in more than one host species, and this lack of observed parasite sharing resulted in lower predictive power. The model output for China indicates the limitations of applying link prediction to such sparsely sampled datasets; therefore, the results for China should be interpreted with caution.

Table 1.

The maximum area under the ROC curve (AUC) for each imputation model, calculated via five-fold cross-validation, along with the number of imputed edges above the probability threshold (for all primates) and a list of the subset of those imputed to humans.

country	max AUC	number of observed interactions	number of imputed interactions	parasites imputed to humans
Brazil	0.99	616	387	Trypanosoma minasense Dipetalonema gracile Prosthenorchis elegans
China	0.57	582	4,458	Plasmodium inui Plasmodium eylesi Plasmodium jefferyi Plasmodium youngi Strongyloides fuelleborni Colobenterobius presbytis Cytomegalovirus Macacine betaherpesvirus 3 Enterobius macaci Dirofilaria macacae Nochtia nochti Oesophagostomum bifurcum Plasmodium cynomolgi Polyomavirus Simian agent virus 40 Rubulavirus Parainfluenza virus 5 Simplexvirus Cercopithecine herpesvirus 1 Spumavirus Simian foamy virus Sandnema digitatum
Colombia	0.98	295	96	Trypanosoma minasense Mansonella marmosetae
Congo DRC	0.93	463	636	Strongylouides fuelleborni Hepatocystis kochi Alphavirus Semliki forest virus Alphavirus Chikungunya virus Staphylococcus aureus Dientamoeba fragilis Trypanosoma primatum Lentivirus Simian immunodeficiency virus Deltaretrovirus Primate T-lymphotrophic virus 3 Oesophagostomum bifurcum Oesophagostomum stephanostomum Encephalitozoon hellem Trypanosoma vivax Balantidium coli Physaloptera caucasia Enterobius antrhopopitheci Mansonella rodhaini Mastadenovirus Human mastadenovirus E Pediculus schaeffi Plasmodium gaboni Entamoeba chattoni Bertiella studeri Necator americanus Orthobunyavirus Bunyamwera orthobunyavirus Spumavirus Simian foamy virus
Madagascar	0.95	243	358	Haemaphysalis lemuris Trichuris lemuris Trichophilopterus babakotophilus Liponysella madagascariensis Dipetalonema petteri Lemuralges intermedius Ixodes lemuris Pararhabdonema longistriata
Indonesia	0.96	253	182	Oesophagostomum aculeatum Plasmodium inui Tanjong Rabok orthobunyavirus

Parasites predicted to infect humans were mostly taxonomic generalists, with 71% observed to infect hosts across taxonomic families or broader (Figure 2a). Viruses comprised the majority of imputed parasites infecting the broadest taxonomic host range (taxonomic generality score = 5, i.e. documented in multiple taxonomic orders), while helminths dominated the imputed parasites infecting NHP hosts from multiple families (generality score = 4).

Figure 2. Host specificity of NHP parasites predicted to infect humans.

(a) Each imputed parasite is plotted along two metrics of host specificity. Parasites in the GMPD were assigned a taxonomic generality score indicating whether it is (1) species-specific, (2) documented to infect members of a single genus, (3) documented to infect members of a single family, (4) documented to infect members of a single order, or (5) documented in multiple orders of mammals and potentially some non-mammals. The phylogenetic generality of each parasite represents the mean phylogenetic distance (MPD), in millions of years, between all hosts that a parasite infects. (b) Standard effect sizes of MPD (SES.MPD) were calculated relative to a null model. Parasites with significant, negative SES.MPD values (z score < −1.96, indicated by the dotted line) are considered phylogenetic host specialists.

The mean phylogenetic distance between all NHP hosts of a given parasite (MPD) varied among the generalist parasites (Figure 2a). The average MPD of all imputed parasites was 20.7 mya, which roughly corresponds to the phylogenetic distance between the primate genera Pan and Pongo or Eulemur and Hapalemur (Upham et al., 2019). In contrast to the taxonomic scores, the SES.MPD values indicated that none of the parasites predicted to infect humans were considered generalists when compared to a null model (Figure 2b). The average SES.MPD of all imputed parasites was −1.84. According to this metric, the host specificity of most parasites did not differ from the null expectation, though 36% could be considered phylogenetic specialists. Most phylogenetic specialists were protozoa and helminths but also included two sexually transmitted viruses.

Of the observed and imputed parasites shared between humans and NHPs, the majority were viruses (29%), helminths (27%) and protozoa (25%; Figure 3a). Compared to parasites transmitted sexually and through intermediate hosts, parasites transmitted through close contact, the environment, and vectors were more commonly observed and imputed among humans and NHPs.

Figure 3. (a) Parasite type and (b) transmission mode of observed and imputed parasites shared by NHPs and humans.

Of the 54 NHP parasites predicted to have high spillover risk, we found evidence from other sources that 47% have naturally infected humans (S2). Some of these parasites pose known threats to human health but have not yet been detected in the country of interest, such as Chikungunya virus in DRC. Some known zoonotic parasites are widespread but neglected and underreported (e.g. Dientamoeba fragilis), and others have been documented in humans only rarely (e.g. Trypanosoma vivax). In addition, some zoonotic viruses have been documented in humans, but evidence of disease is lacking (e.g. Simian foamy virus and Primate T-lymphotropic virus 3).

We did not find evidence of natural human infection for 53% of the predicted parasites. However, experimental studies indicate that several of these parasites can infect human cell lines, including Parainfluenza virus 5 and Macacine betaherpesvirus 3. Other predicted parasites are not confirmed to infect humans in any setting and may pose varying levels of spillover and disease risk. These include ectoparasites of lemurs (e.g. Haemaphysalis lemuris), pinworms of Old World monkeys and apes (e.g. Enterobius anthropopitheci) and viruses that circulate in Old World monkeys (e.g. Tanjong Rabok virus).

Discussion

The threat of infectious disease transmission between humans and other mammals underscores the need for effective predictive frameworks to guide disease surveillance at the growing human-animal interface. This is especially true for NHPs, which have been the source of many novel parasites throughout human evolution due to their close phylogenetic relatedness to humans, coupled with geographic and ecological overlap that provides opportunities for spillover. To investigate these threats, we used a phylogeographic model to identify previously undocumented host-parasite interactions between humans and NHPs in six countries with high primate diversity. We found that imputations varied by country, with the DRC, China, and Indonesia having the highest percentage of imputed edges between a known parasite in NHPs and humans. Overall, the models suggest that 54 unique parasites in NHPs represent threats to humans. These findings therefore provide country-specific targets for surveillance efforts aimed at preventing zoonotic spillover events, including surveillance in NHPs in close proximity to human populations.

Literature searches supported model predictions by revealing evidence that transmission of certain NHP parasites to humans is both biologically possible (Cormier, 2016) and spatially likely (Barratt, Harkness, Marriott, Ellis, & Stark, 2011; Zhang et al., 2016). In fact, several of the predicted parasites have already been documented to naturally infect humans outside of our countries of interest (Beran, 2019; Kotepui, Masangkay, Kotepui, & Milanez, 2021). For instance, the model identified Plasmodium cynomolgi as a threat to humans in China and Plasmodium inui in China and Indonesia. Macaque-to-human and human-to-human transmission of P. cynomolgi and P. inui have been demonstrated via accidental laboratory exposures and experiments performed on imprisoned populations (Cormier, 2016), and the first natural infections were recently reported in Malaysia (Liew et al., 2021; Ta et al., 2014). P. cynomolgi and P. inui are present at high prevalence in macaques known to share space with humans (Zhang et al., 2016) and represent continued risks for spillover into human populations (Kotepui et al., 2021; Yap et al., 2021). Also included among predictions is Dientamoeba fragilis, a known enteric protozoan with a global distribution (Barratt et al., 2011). D. fragilis remains neglected due to its historically debated pathogenicity, though studies have linked it to gastrointestinal illness (Norberg, Nord, & Evengård, 2003; Stark et al., 2010). This protozoan parasite was reported in humans in all countries except the DRC, where the model filled in the undocumented interaction. Such examples corroborate model performance while calling attention to underreported parasites with known zoonotic risk.

Our findings also provide more general insights to the nature of zoonotic threats beyond the six focal countries investigated here. One important general conclusion is that taxonomic generalism is a stronger predictor of spillover likelihood than parasite type or transmission mode. Thus, as expected, taxonomic generality scores revealed that parasites predicted to infect humans were mostly generalists that infect hosts of different families, orders, or even classes. We expected that phylogenetic generaltiy metrics would correlate with taxonomic scores, but mean phylogenetic distances (MPD) between NHP hosts were highly variable. The standardized effect sizes of the MPD values (SES.MPD) indicated that even parasites with a wide taxonomic host breadth (including non-primates) were, on average, constrained to closely-related NHP hosts. Large discrepancies between taxonomic and phylogenetic generality may highlight the opposing biases of both metrics; taxonomic generality scores can exaggerate host generalism given their simplicity and emphasis on maximum host breadth, while MPD and SES.MPD can overstate host specialism in sparsely sampled datasets like the GMPD (Park et al., 2018). Examining both phylogenetic and taxonomic generality metrics together presents a more complete picture of a parasite’s host range and its potential to infect additional hosts.

We found less support for the effects of parasite taxonomy or transmission mode. We hypothesized that parasites predicted to occur in humans would be dominated by viruses and by vector-transmitted parasites, as previous work based on taxonomic specificity indicated that these traits predict greater generalism among primate hosts (Cooper, Griffin, et al., 2012; Pedersen et al., 2005). However, we found that viruses, protozoa, and helminths were evenly represented among both the NHP parasites observed in humans and those predicted to infect humans. In addition, rather than a predominance of vector transmission in the observed and imputed parasites, we found evidence for a wider range of transmission modes, including close contact, environmental, and vector transmission.

Although roughly two-thirds of the parasites predicted to infect humans were protozoans and helminths, host immunity (i.e. to Trypanosoma spp.; Vanhollebeke, Lecordier, Perez-Morga, Amiguet-Vercher, Pays, 2007) and low human-to-human transmission potential (i.e. for Physaloptera spp.; Makki et al., 2017) may limit the likelihood of widespread outbreaks originating from these parasites. The presence of helminths in the imputation results was particularly surprising given that prior studies suggest that helminth parasites of NHPs are more likely to be host specialists (Cooper, Griffin, et al., 2012; Pedersen et al., 2005). Most of the specialist helminths in the GMPD have complex life-cycles, which limits the ecological likelihood that they will adapt to new hosts (Wells, Gibson, & Clark, 2019). Consistent with this association, we found that parasites requiring an intermediate host were least likely to occur among the imputed human-parasite interactions, while the majority of predicted helminths were transmitted environmentally or through vectors. Our results thus suggest that taxonomic generalism is a more consistent indicator of spillover likelihood than parasite type or transmission mode.

The rapid generation time of viruses, especially RNA viruses, allows them to adapt quickly to novel hosts, successfully transmit between hosts, and evade natural or vaccine-induced immunity (Duffy, 2018). Hence, global surveillance efforts aimed at curbing the next pandemic focus heavily on novel viruses (Carroll et al., 2018). Our model identified undocumented or underreported viruses that may present a significant threat to communities at the human-wildlife interface and beyond (S2). For example, Simian Foamy Virus (SFV) was predicted to occur in humans in DRC and China, yet has not been documented in GIDEON. A study of 305 people living in South and Southeast Asia detected SFV in 2.6% of participants, most of whom reported bites or scratches from macaques in urban and tourist settings (Jones-Engel et al., 2008). Although no evidence of disease in humans has been reported, the evolution of a pathogenic strain and subsequent outbreak remains a possibility (Khan, 2009; Pinto-Santini, Stenbak, & Linial, 2017). By identifying potential pathways of viral spillover, the phylogeographic link-prediction model can guide targeted surveillance efforts at the human-NHP interface.

Some viruses predicted by the phylogeographic model have not been documented in humans, yet evidence suggests that they pose significant spillover risk (S2). For example, Tanjong Rabok virus (TRV) of the Bakau bunyavirus group is predicted to infect humans in Indonesia. TRV has been isolated from Culex mosquitoes in Malaysia, and antibodies have been detected in a broad range of hosts (Beran, 1994). Though it has been suspected of causing disease in Southeast Asia, a naturally-acquired TRV infection has never been confirmed in humans (Beran, 1994). The Peribunyaviridae family, of which TRV is part, contains many emergent and understudied RNA viruses (Contigiani, Diaz, & Tauro, 2017). Emerging bunyaviruses may circulate asymptomatically in susceptible communities or cause symptoms indistinguishable from those of other etiological agents (Wolfe et al., 2005). The lack of evidence for human infection by TRVs and other predicted viruses does not necessarily indicate low spillover risk; rather, surveillance initiatives and experimental studies should prioritize these viruses to further assess their zoonotic potential.

The use of a phylogeographic model allowed us to better account for the dynamic processes that influence parasite sharing across the differing primate assemblages in each country. The DRC is home to the closest evolutionary relatives of humans – Pan troglodytes and Pan paniscus – and the model predicted a high degree of parasite sharing between these species and humans. However, in the case of China, the majority of parasites predicted to infect humans were imputed from Macaca mulatta rather than their closest relatives (the lesser apes). This likely reflects the large geographic range of Macaca mulatta, resulting in greater ecological opportunities for spillover. Thus, both phylogeny and geography contributed to the predictions.

We also found cases in which phylogeny and range overlap did not adequately capture risk of transmission to humans, such as in Madagascar. The primates of Madagascar are all lemurs, and as a distinct radiation from monkeys and apes, all lemurs are equally related to humans. They are also all endemic to Madagascar, so every NHP species overlapped with humans along its entire range. Parasites of lemurs are sparsely sampled compared to those of other NHPs, resulting in lower predictive power for this group. These factors resulted in some unexpected predictions for parasites from NHP in Madagascar to infect humans. Consider, for example, the high probabilities assigned to mites and lice, which are some of the only widely-sampled parasites among lemurs. Mites and lice tend to coevolve with their hosts, so a successful host switch to humans seems improbable (Hafner, Demastes, Spradling, & Reed, 2003). Systematic sampling for parasites in understudied NHPs is needed before their role in zoonotic transmission can be accurately assessed.

Variation in sampling effort has long been recognized as a challenge to understanding parasitism (Gregory, 1990; Walther, Cotgreave, Price, Gregory, & Clayton, 1995), including in NHPs (Nunn et al., 2003; Teitelbaum et al., 2020). Cooper and Nunn (2013) identified the scale of sampling gaps and the factors that drive sampling biases, finding that even for the best studied NHPs, nearly one-half of their parasites have yet to be sampled. Bacterial parasites in particular are underrepresented, making it difficult to draw general conclusions about bacteria host range and transmission risk in NHPs. In contrast, humans are much better sampled for parasites, with an estimated 2107 species, which is likely an underestimate (Dunn, Nunn, & Horvath, 2017). This estimate for humans is two orders of magnitude higher than the number of parasites documented in other primates, suggesting that humans might be super-parasitized due to epidemiological transitions associated with the agricultural transition, the rise of high-density cities, and globalization (Barrett, Kuzawa, McDade, & Armelagos, 1998). A recent comparative study (Amoroso & Nunn, 2021) failed to find strong support for this hypothesis of super-parasitism within individual countries; when pooled across countries, however, humans do appear to have more viruses than expected based on comparative patterns in NHPs. Thus, even the exceptional number of parasites in humans is largely due to more intensive sampling, rather than ecological factors.

Uneven sampling in the GMPD may give rise to biases in the link prediction results presented here. Relatively well-studied species in the GMPD tend to be those that have the largest geographic ranges and closest phylogenetic relatedness to humans, as expected given that traits of these hosts, including larger body size, terrestriality, and the ability to inhabit human-altered environments, are also conducive to accessibility for sampling (Cooper & Nunn, 2013). The elevated risk of zoonotic spillover from species like Pan troglodytes and Macaca mulatta may be exaggerated due to the higher number parasites observed in these hosts. In addition, link prediction cannot impute parasites that have yet to be documented, which is a major limitation given that only a fraction of parasite vertebrate diversity has been described (Carlson et al., 2020). We cannot expect the model to make ecologically relevant predictions of zoonotic spillover from lemurs, for instance, when entire parasite groups (i.e. viruses) are not described for lemurs in the GMPD.

The method presented here could theoretically be used to impute parasites of humans that may be undocumented in wild NHPs, thus providing useful information for NHP health and conservation. However, the vast difference in the number of reported parasites between humans and NHPs limits the feasibility of applying our method for this purpose; humans are so “oversampled” that their parasites would automatically comprise the majority of those imputed in NHPs. Link prediction models can serve as a valuable tool for illuminating aspects of zoonotic risk and host-parasite ecology, but their limitations highlight the continued need for systematic sampling of parasites in wildlife hosts.

In conclusion, we used a novel link-prediction approach to identify undocumented zoonotic parasites and reveal gaps in global preparedness for emerging disease. We applied this method to a narrowed dataset of high-risk NHP species in selective diversity hotspots, allowing us to manually evaluate all results and make targeted, relevant predictions for zoonotic and epidemic risk. Efforts to predict NHP parasites with high spillover risk must be continuously updated to reflect both ongoing parasite discovery and global change. The ranges of NHP species and their parasites will continue to shift under the synergistic effects of habitat destruction, climate change, and biodiversity loss (Carlson et al., 2017; Estrada et al., 2017). Models that predict future parasite-sharing risk should consider shifts in geographic ranges, as demonstrated by Morales-Castilla et al. (2021) in ungulate hosts, and future extinction scenarios, as explored by Herrera et al. (2021) in NHPs. In addition, the use of link-prediction models to identify human parasites that may spillover to NHPs should be explored as documentation of NHP parasite communities improves. Future link-prediction models applied at regional scales could incorporate more comprehensive data on parasite distributions, host behavior patterns, and contact probabilities between humans and wildlife, and should eventually move beyond binary, presence-absence predictions to better quantify parasite prevalence and transmission risk among at-risk communities. Finally, the application of link-prediction models to larger-scale host-parasite networks remains an important strategy for investigating broader patterns of parasite sharing, and will become increasingly useful as gaps in host-parasite databases are filled.