Introduction

Leishmaniasis, a debilitating and sometimes fatal parasitic disease, is one of the 20 neglected tropical diseases targeted for control and elimination by the United Nations by 2030 [1]. However, leishmaniasis remains a persistent threat throughout the tropics and subtropics. In the Americas, leishmaniasis cases occur from the southern United States to Argentina, including the Caribbean, with an estimated 58,500 new cases each year [2]. Further, incidence is rapidly shifting in geographic extent with land-use and climate change [3–7]. For instance, hotspots of leishmaniasis have significantly expanded throughout Costa Rica in the last two decades, in concert with agricultural intensification, deforestation, and reforestation [3].

Treatment for leishmaniasis has many side effects and are extremely expensive [2,8], thus control and elimination hinges on limiting transmission. Leishmania parasites are transmitted by sandflies (Order: Diptera; Family: Psychodidae) and maintained in the environment via a multitude of sylvatic and domestic mammalian hosts. As such, controlling transmission requires ecological interventions targeting either component of the transmission cycles. For instance, clearing vector habitat proximate to households, supporting populations of natural enemies that regulate populations of vectors or small mammals (e.g., fish and mosquito larvae), and/or modifying distribution of wildlife food sources are sustainable and effective interventions in preventing transmission of flaviviruses, henipaviruses, Plasmodium spp. (malaria), and Brucella abortus (brucellosis) [9,10]. Similar interventions could be employed against Leishmania, however, successful application requires species identification of vectors, hosts, and their corresponding interactions with their biotic and abiotic environment.

Leishmaniasis was first diagnosed by Western medicine in the Americas in 1909, at the time dogs and foxes were believed to be the zoonotic hosts [11]. Continued research throughout the century hypothesized that rodents, and in some cases opossums and sloths, were the primary zoonotic hosts of these pathogens [11]. However, especially with the advancement of molecular diagnostic techniques, the number and taxonomic diversity of hosts has since expeditiously increased. In the Americas, there are now 137 species of wildlife that are recognized as potential hosts (i.e., animals that are naturally exposed and become infected once exposed), of which around 60 may act as competent hosts (i.e., animals can become infectious once infected and may maintain the pathogen in the environment as reservoir hosts) [12,13]. Wild hosts are now known to include eight orders of mammals including Chiropetra (bats), Carnivora, Cingulata (armadillos), Didelphimorphia (opossums), Lagomorpha (rabbits), Pilosa (sloths and anteaters), Primates, and Rodentia. However, even with progress in host discovery, the full suite of Leishmania hosts is not yet described: Leishmania spillover occurs in absence of known hosts and new hosts are discovered almost annually [14–17].

Limited knowledge on the transmission cycle makes it difficult to understand the mechanisms driving Leishmania distribution and spillover. As previously stated, unknown transmission cycles impede effective ecological interventions to reduce spillover risk to people. Additionally, global change can alter host dynamics to amplify spillover risk. For example, land-use change may increase host densities and/or push hosts into closer proximity to anthropophilic vectors and humans [9]. Climate warming may also shrink or expand the range of hosts, further changing the rate of contact among hosts, vectors, and humans [18]. As such, the incomplete classification of possible Leishmania hosts could lead to unexpected emergence and hotspots of human leishmaniasis.

Study designs optimal for detecting Leishmania hosts, such as systematic sampling of large populations of a diversity of species through space and time, are limited by logistics and cost of sampling. Leishmania host discovery could benefit from strong a priori hypotheses about which mammals to target for sampling effort; taxa specific sampling could help to efficiently use time and monetary resources, ultimately expediting the full classification of the Leishmania transmission cycle. Through the last decade, machine learning approaches have been used to generate predictions of hosts of a multitude of different pathogens, including betacoronaviruses [19], flaviviruses [20], Rabies virus [21], and Borrelia burgdoferi (causative agent of Lyme disease) [22]. These methods have facilitated more targeted laboratory and field work of novel pathogens and have great potential to improve our understanding of diseases of poverty with broad host and vector ranges.

Host status for any parasite or pathogen is driven by two processes: exposure and physiological susceptibility [23,24]. Exposure depends on the environmental conditions needed to support pathogen transmission and, in the case of vector-borne disease, vector reproduction, contact with competent hosts, and vector survival [25]. Following exposure, physiological susceptibility then depends on pathogen interaction with host cells to gain entry and, in some cases, avoid immune attack and replicate [25]. Host traits related to exposure and/or susceptibility likely interact in non-linear and higher-order combinations to delineate zoonotic Leishmania hosts from non-hosts. In brief, the effect of a trait on host status may vary in magnitude or direction depending on trait value and traits may become more or less important under different conditions. Hypothetically, for example, the average temperature of a species range may have a positive effect on host status between 25–30°C, but a negative effect outside of this range. Depending on biting activity of vector species, animals that forage on the ground may be more likely to be hosts if they are also nocturnal, whereas animals that are arboreal may be more likely to be hosts if they are crepuscular but less likely to be hosts if they are nocturnal. Thus, when viewing the zoonotic Leishmania system in one or two dimensions (e.g., only considering phylogeny or habitat use), it is difficult to determine unifying traits useful for identifying hosts. To date, it has been challenging to combine these processes across scales from physiological to evolutionary to ecological to biogeographic to predict their overall impact on infection potential. Machine learning offers a set of new tools that can be used to incorporate and dissect this complexity by increasing the dimensionality in which the Leishmania system can be studied, allowing for multiple predictors that may have nonlinear and interactive effects on host status. The flexibility of these models allows for pinpointing combinations of mammalian traits unique to the confluence of Leishmania exposure and susceptibility by learning patterns of known hosts. With this information, we can identify likely host species as targets for Leishmania surveillance efforts and the key traits that distinguish host species from non-hosts.

Here we use tree-based machine learning (extreme gradient boosted regression, ‘Xgboost’) and mammalian traits, including but not limited to biogeographical, phylogenetic, and life-history features, to: (i) describe trait profiles of hosts for zoonotic Leishmania (Viannia) and Leishmania (Leishmania) parasites; and (ii) use the trait profiles to predict unrecognized wild hosts of zoonotic Leishmania within the subgenera Leishmania (Viannia) and Leishmania (Leishmania).

We aim for this analysis to identify animal species that are likely to be exposed and infected in the wild; due to extreme data sparsity, we do not employ our methods to identify competent hosts (i.e., animals that become infectious once infected and may act as reservoirs given the density and distribution of their populations). Our findings can be used to identify hosts that can then be surveyed at larger spatio-temporal scales, investigated for competence, or used as sentinel species. As such, while our modeling approach does not necessarily identify competent hosts that can transmit zoonotic Leishmania on to sandflies and eventually humans, it does identify animals likely to be exposed to and susceptible to zoonotic Leishmania infection, facilitating the first step of determining epidemiological importance and fine-tuning surveillance efforts.

Methods

Zoonotic Leishmania host status

We gathered Leishmania host status from all wild, endemic, and invasive terrestrial mammals with ranges in Mexico, Central America, and South America using recent reviews [12,13] in addition to a Web of Science query on February 12, 2021 and February 22, 2021 (see S1 Text for the specific search strings). Host status was defined as a binary trait: an animal has been naturally infected by one of the 14 Leishmania species known to cause leishmaniasis in humans (1) or there is no record of infection by a Leishmania species known to cause leishmaniasis in humans (0). The latter category (“non-positives”) includes animals that have been tested for Leishmania and have not been found to be infected as well as those that have not been tested for Leishmania infection. Ideally, the analysis would include true-positives (1) and true negatives (0). However, due to the intensive, systematic sampling required to declare an animal as a true-negative (longitudinal sampling across large geographic scales), there is not enough data to conduct the analysis with true-negatives. Our analysis (using true-positives and non-positives) still enables us to identify species that have not yet been tested but have a high probability of being a host. Infection was determined via detection of Leishmania species specific antigens or genetic material. Species identification was performed either using animal tissue samples directly or using live parasites cultured from animal tissue. To ensure that all relevant mammals were included in our study, a list of species names of endemic Latin American mammals was retrieved from the International Union for Conservation of Nature (IUCN) Red List Inventory [27] and a list of invasive mammals was retrieved from the Global Invasive Species Database [28]. We cross-referenced our host status data table with the Global Infectious Disease and Epidemiology Network [29]. We then limited our analysis to wild mammals that had at least a 10% range overlap with reported human cases of leishmaniasis for inclusion in our analysis. In total, 86.64% (1460/1685) of species were retained for the L. Viannia analysis and 87.24% (1470/1685) of species were retained for the L. Leishmania analysis. Zoonotic Leishmania ranges were constructed by outlining a concave polygon around cases of disease occurrence in humans [30] (Fig 1A and 1C). We conducted all data filtering and downstream analyses separately for each subgenus.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. The ranges of zoonotic Leishmania and their known and predicted hosts extend throughout the Americas.

The number of overlapping ranges of known Leishmania hosts are depicted in (a & c), with the range of human cases of L. (Viannia) and L. (Leishmania) outlined in grey. The number of overlapping ranges of newly predicted hosts are outlined in (b & d). Hylaeamys megacephalus and Calomys callosus are unrecognized hosts with the highest and second highest predictions for L. (Viannia) according to Shapley values. Their ranges are outlined in dark green and light green, respectively, in (b). Dasypus novemcinctus and Leopardus wiedii are hosts with the highest and second highest Shapley values for L. (Leishmania). Their ranges are outlined in dark blue and light blue, respectively, in (d). An exhaustive list of newly predicted hosts is listed in S2 Table. Base maps were created in R using open source shapefiles from Natural Earth (admin-0 countries: https://www.naturalearthdata.com/downloads/10m-cultural-vectors/; https://www.naturalearthdata.com/about/terms-of-use/) [51]; species ranges were created in R using open source shape files from IUCN (https://www.iucnredlist.org/resources/spatial-data-download; https://www.iucnredlist.org/terms/terms-of-use) [27].

https://doi.org/10.1371/journal.pntd.0010879.g001

Results

Leishmania (Viannia)

Through our literature search, we found 96 known L. (Viannia) hosts out of 1460 mammals that occur within the geographic range of L. (Viannia) parasites (Figs 1A and 2A). After accounting for target shuffling, our model performed moderately well with an average out of sample AUC of 0.81 (95% CI: 0.80–0.82) and in sample AUC of 0.94.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2.

The number of known and predicted hosts per order for L. (Viannia) (a) and L. (Leishmania) (b). Grey bars represent the number of known hosts within each order, colored bars represent the number of newly predicted hosts per order. Animals were classified as newly predicted if > 95% of mean SHAP values were greater than 0 (a-b). Bars in (c) represent the proportion of newly classified hosts within each order; colors match (a-b).

https://doi.org/10.1371/journal.pntd.0010879.g002

L. (Viannia) trait profiles

After removing highly correlated traits, 30 features were used to predict L. (Viannia) hosts. A mix of biogeography, life-history, phylogeny, and study effort covariates were among the top important model features when estimating global feature contribution (Fig 3). Phylogenetic distances, reduced into six axes of a PCoA ordination, were, on average, the most important trait for Leishmania (Viannia) hosts. Specifically, animals were more likely to be hosts if they were rodents and opossums, and less likely to be hosts if they were non-human primates, carnivores, and even toed ungulates (S1 and S2 Figs). Similar to other zoonotic pathogen systems, L. (Viannia) hosts live at high population densities and demonstrate signatures of fast-paced life-histories, indicated by short gestation periods (Fig 4A). Further, L. (Viannia) host ranges are large, have high proportions of crop and urban land cover, and, on average, peak at ~28°C during the warmest quarter of the year (Fig 4A). Geographic coordinates of the species’ native ranges were also an important predictors: animals were more likely to be hosts if the minimum longitude of their native range was around western Central America. Notably, although study effort (number of PubMed citations, normalized with a Box-Cox transformation) was among the most important traits, our supplementary analyses indicated that the host models were not simply predicting well-studied hosts. Specifically, the supplementary model predicting study effort (number of PubMed citations) performed poorly (average out-of-sample R² = 0.19). Further, traits that predict citation count, and the response curves of those traits, greatly differed from traits that predict L. (Viannia) host status (S3–S4 Figs). As such, while L. (Viannia) hosts are likely understudied, we are not simply reporting the trait profiles of well-studied mammals.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Biogeography, life-history, and phylogenetic traits all significantly contributed to model predictions for host status.

Biogeographical traits (minimum longitude of the range, maximum longitude of the range, % cover of land-use/land cover in the species range, average temperature in the warmest quarter in the species range, range area) are colored in blue, life-history traits (population density, gestation length, litter size) are colored in red, phylogenetic distance (location along PCoA ordination axes) is colored in orange, and study effort is colored in grey. Points are the absolute value of the mean Shapley importance (mean |SHAP value|) for the trait across all mammals (i.e., global feature contribution), bars represent the absolute values of the 0.05–0.95 percentiles. Only features with 0.05 percentiles > 0 are shown.

https://doi.org/10.1371/journal.pntd.0010879.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Hosts have fast-paced life-histories and live in proximity to humans.

Shapley partial dependence plots showing the effect of each feature on L. (Viannia) (a) and L. (Leishmania) (b) host status after accounting for the average effect of the other features in the model. Colored lines represent the average effect across model iterations, while grey lines show each individual model iteration (model fit with 80% of data) (blue = biogeographical traits; red = life-history). Features with global mean feature contribution scores > 0 for > 95% of model iterations are shown. Rug plots on the x-axis indicate the distribution of the data.

https://doi.org/10.1371/journal.pntd.0010879.g004

Discussion

Our literature search identified 137 wild mammals in the Americas as known zoonotic Leishmania hosts and our model predicted there to be an additional 136 likely hosts (animals that are likely to be exposed and infected in the wild, but not necessarily infectious upon infection). The majority of the predicted hosts for both subgenera fell within the Rodentia order (L. (Leishmania) N = 27; L. (Viannia) N: = 34), with a similar number of predicted Leishmania (Leishmania) hosts within the Chiroptera order (N = 32). While their contribution to model predictions varied between subgenera, key traits included phylogenetic distance, human associated land-use, climate, population density, and study effort.

For both Leishmania subgenera, biogeographical features encompassing descriptors of species ranges were among the top five contributors to host predictive accuracy. Notably, zoonotic L. (Leishmania) hosts are more likely to be animals with larger ranges than non-hosts. Geographic range area was also reported to be a shared predictive feature among zoonotic hosts and reservoirs in other systems [21,53]. Further, ranges of L. (Leishmania) and L. (Viannia) hosts are made up of higher proportions of crop cover and urban cover than ranges of mammals that are not Leishmania hosts. These results are in accordance with previous findings that, in general, zoonotic host richness and relative abundance increases with anthropogenic pressure [54]. Further, our results align with recent literature showing that risk of leishmaniasis is associated with agricultural intensification and urbanization [4,55]. L. (Viannia) hosts are marginally more likely to have ranges with moderate forest cover than non-hosts whereas there is little effect of forest cover on L. (Leishmania) host status. Vectors of cutaneous leishmaniasis, which is caused by all species of zoonotic Leishmania except L. infantum, are more likely to live in forests with high integrity and low human modification [56], while hosts are more common in disturbed landscapes [57]. Our analysis adds support to the latter finding, highlighting that transmission of parasites associated with cutaneous leishmaniasis likely occurs at forest-human settlement interfaces. The incomplete overlap between vector ranges and host ranges may explain the highly localized dynamics of cutaneous leishmaniasis spillover (i.e., transmission from wildlife to humans via the sandfly vector) [6,58]. Spillover occurs when reservoir hosts, vectors, and humans overlap in space and time so that forward transmission occurs from animal to sandfly to human. If vector and host distributions only overlap at habitat interfaces, area of reservoir host, sandfly, and human co-occurrence may be low and variable through time. Finally, average temperature of the warmest quarter was an important predictor of host status, with the feature contribution to model predictions peaking around 28°C. In the case of zoonotic leishmaniasis, temperature is likely a predictor of Leishmania exposure as this trait profile reflects temperatures associated with vector survival and parasite development [59,60].

The profiles of important life history traits for Leishmania hosts resembled those observed in other comparative analyses of zoonotic diseases. Specifically, for both Leishmania subgenera, the probability of host status peaked at low to median lengths of gestation, a pattern similar to that of rodent reservoirs of zoonotic pathogens and competent hosts of zoonotic tick-borne pathogens [53,61]. Further, the probability of zoonotic L. (Leishmania) host status increased with litter size, a trend that parallels that of both rodents reservoirs of zoonotic pathogens and bat and carnivore reservoirs of rabies virus [21,53]. Altogether, our findings suggest that, like other zoonotic reservoirs and hosts, zoonotic Leishmania hosts may have fast-paced life-histories. Animals with fast-paced life-histories are believed to maximize fitness by expeditiously investing in short-term reproductive effort—such as through younger ages at sexual maturity, short gestation lengths, greater litter sizes—at the cost of long-term survival [62]. As a result, animals with fast-paced life histories are thought to divert energetic resources away from effective adaptive immunity [63], or utilize responses conducive to tolerance but not resistance [64], allowing them to allocate maximal resources to short term reproductive effort. As such, animals with fast-paced life-histories may be more susceptible to infections.

However, L. (Viannia) hosts only had one trait (gestation length) clearly associated with a fast-paced life history, suggesting weaker evidence for the “fast-paced” host hypothesis for this subgenus. Phylogenetic distance was the most important trait for L. (Viannia) hosts. We hypothesize that phylogenetic relationships may better account for host physiological defenses against Leishmania than life-history traits. If this is the case, inclusion of specific genetic markers associated with Leishmania-host cell interactions may improve model predictions.

Notably, study effort significantly contributed to host predictions for both subgenera. While our follow up analyses do not indicate that model results are biased by uneven study effort, this finding may suggest that zoonotic Leishmania hosts are generally under sampled. Though leishmaniasis is a neglected tropical disease primarily impacting resource-poor communities, state-of-the-art zoonotic Leishmania diagnostic and surveillance techniques are available [16,57,65–69]. As such, under sampling Leishmania hosts primarily stems from both the lack of funding to employ these tools during large scale surveillance efforts and the large number of animal species that are potentially hosts. The main leishmaniasis research funding mechanism within the USA, the National Institute of Health (NIH), primarily focuses on the molecular biology of leishmaniasis work and does not prioritize research elucidating the ecology of the sandfly-wildlife transmission cycle. In fact, a simple search for “leishmania” on the NIH and National Science Foundation grants database yields a combined total of three awards aiming to elucidate transmission cycles of Leishmania in the Americas (searched October 10, 2022). Researchers across Latin America are faced with similar, if not more severe, funding constraints. For example, in Brazil, a major hotspot of leishmaniasis, the federal administration from 2019–2022 cut 90% of federal science funding [70]. While our model results facilitate more targeted field sampling for particular species, optimized surveillance should consist of an iterative feedback between local knowledge and model predictions to guide field work and updating the model with new results from field sampling to guide ongoing surveillance and field studies [16,19]. Importantly, to reduce study effort bias going forward, care should be taken to estimate and survey under studied areas and species in addition to newly predicted species, particularly in areas with high rates of human spillover with poorly known hosts. A lack of investment in understanding the basic ecology of Leishmania slows research progress by obscuring the major drivers of disease dynamics and spillover transmission to humans. Leishmaniasis surveillance and management would benefit from larger funding mechanisms prioritizing understanding the ecology of this group of pathogens.

In some cases, our host predictions are supported by multiple independent but complementary lines of evidence (summarized below and in S3 Table). Non-specific Leishmania antigens, antibodies, or DNA has been detected in the crab-eating racoon (Procyon cancrivorus), neotropical river otter (Lontra longicaudis), six-banded armadillo (Euphractus sexcinctus), coyote (Canis latrans), and the grey short-tailed opposum (Monodelphis domestica) [12,68,71,72]. As tests were not parasite species specific, it is unconfirmed if they were infected with zoonotic Leishmania species, but our model suggests they may be. Further, Leishmania-like flagellates were discovered in Derby’s wooly opossum (Caluromys derbianus) [71]; however, to our knowledge this species has not been tested for zoonotic Leishmania species.

A number of animals predicted to be hosts for L. (Leishmania) were also found to co-occur with vectors of L. (L.) mexicana in Mexico: the hairy big-eyed bat (Chiroderma villosum), Brazilian brown bat (Eptesicus brasiliensis), great sac-winged bat (Saccopteryx bilineata), tent-making bat (Uroderma bilobatum), Mexican woodrat (Neotoma mexicana), black footed pygmy rice rat (Oligoryzomys nigripes), Coues’s rice rat (Oryzomys couesi), fulvous harvest mouse (Reithrodontomys fulvescens), nine-banded armadillo (Dasypus novemcinctus), lowland paca (Cuniculus paca), Virginia oppossum (Didelphis virginiana), kinkajou (Potos flavus), Northern racoon (Procyon lotor), white-nose coati (Nasua narica), and the margay (Leopardus wiedii) [16,73]. Notably, the nine-banded armadillo and margay had the two highest summed Shapley scores as potential L. (Leishmania) hosts. The nine-banded armadillo is a known host of L. (Viannia) as well as other zoonotic pathogens including the causative agent of leprosy [74].

Host competence (i.e., the ability to become infected and infectious) has been demonstrated in the lab for four of the predicted species: the Virginia opossum (Didlephis virginiana), the large vesper mouse (Calomys callosus), spix’s yellow toothed cavy (Galea spixii), and the Mexican free-tailed bat (Tadarida brasiliensis) [71,75–77]. Although host competence for these animals has been tested in the laboratory, infection/exposure has not been tested in the wild or animals have not tested positive in the wild (possibly due to minimal sampling of the species). Our model suggests that these animals are also likely to be naturally infected in the wild and thus warrant further investigation. The large vesper mouse is one of the top predicted species for L. (Viannia). Surprisingly, we found no peer-reviewed field studies testing wild large vesper mice for zoonotic Leishmania infection. Finally, our model suggests that a handful of species naturally infected with L. infantum in Europe are also likely to be exposed to and susceptible to zoonotic Leishmania in the Americas: the red fox (Vulpes vulpes), the European hare (Lepus europaeus), and the European rabbit (Oryctolagus cuniculus) [78–80].

Importantly, while there are 14 species of zoonotic Leishmania parasites in the Americas, our analysis divides these species into two subgenera but does not identify Leishmania species specific hosts. Future work could leverage the information presented by the current analysis to identify Leishmania species specific hosts. Additionally, our analysis identifies animals that are likely to be infected in the wild but does not narrow the host pool down to competent hosts (animals that are likely to be both infected and infectious in the wild). Next steps of model development could identify trait profiles of competent hosts that delineate them from non-competent hosts. Notably, [56] used machine learning to predict unrecognized vectors of leishmaniasis. Metabarcoding of sandflies and their blood meals [66,81] may be useful in linking model predictions of vectors and hosts to local transmission cycles.

Overall, models for both subgenera identified a number of new hosts, including but not limited to the nine-banded armadillo, the margay, the large vesper mouse, and the large headed rice rat, that can be targeted for future surveillance and research to determine their ecological significance in leishmaniaisis transmission and spillover to humans. Next steps should include ground truthing the role of newly identified potential hosts in the zoonotic Leishmania transmission cycle through field sampling and laboratory experiments, through engagement with local communities, and iteratively updating the current analysis as new information is collected, taking careful consideration to not perpetuate the effects of study bias. Although a cautious first step in host discovery work, our study can immediately inform surveillance efforts and study design. For instance, host predictions may provide a starting point for implication of hosts during local outbreaks of human leishmaniasis by providing a targeted list of likely suspects to sample. While the study of leishmaniasis has historically focused on rodents and small marsupials, our study suggests that capture methods should more commonly be supplemented by bat mist-netting and/or techniques to capture small to medium sized carnivores. Finally, many newly predicted hosts have ranges that extend throughout the southeastern USA. Ecological niche models suggest that climate warming may increase the range of Leishmania vectors as far north as Canada and leishmaniasis endemicity within the US was recently confirmed [82–84]. Similarly, the range of predicted hosts for L. (Viannia) extends farther south through Argentina than currently known hosts. Our analysis provides tools for monitoring leishmaniasis risk under global change as these hosts could be used as sentinel species for tracking leishmaniasis range expansion. Testing these hosts for leishmaniasis infection beyond the current putative boundaries of zoonotic Leishmania ranges would help to evaluate if range expansion has already occurred as well as provide an updated baseline for future monitoring.

Identifying local transmission cycles is key for reducing leishmaniasis spillover and disease burden in people, especially in the face of land-use and climate change, which likely influence the geographic distribution and extent of vector—host—human contact and were strongly associated with novel host predictions in our models. We estimate that there are likely many unknown hosts of leishmaniasis throughout the Americas, highlighting the complexity of the transmission cycle and a need to increase study effort on the ecology and epidemiology of this system.

Supporting information

Acknowledgments

We thank the Mordecai Lab for providing feedback on manuscript drafts.