Data-driven predictions of potential Leishmania vectors in the Americas

Abstract

The incidence of vector-borne diseases is rising as deforestation, climate change, and globalization bring humans in contact with arthropods that can transmit pathogens. In particular, incidence of American Cutaneous Leishmaniasis (ACL), a disease caused by parasites transmitted by sandflies, is increasing as previously intact habitats are cleared for agriculture and urban areas, potentially bringing people into contact with vectors and reservoir hosts. Previous evidence has identified dozens of sandfly species that have been infected with and/or transmit Leishmania parasites. However, there is an incomplete understanding of which sandfly species transmit the parasite, complicating efforts to limit disease spread. Here, we apply machine learning models (boosted regression trees) to leverage biological and geographical traits of known sandfly vectors to predict potential vectors. Additionally, we generate trait profiles of confirmed vectors and identify important factors in transmission. Our model performed well with an average out of sample accuracy of 86%. The models predict that synanthropic sandflies living in areas with greater canopy height, less human modification, and within an optimal range of rainfall are more likely to be Leishmania vectors. We also observed that generalist sandflies that are able to inhabit many different ecoregions are more likely to transmit the parasites. Our results suggest that Psychodopygus amazonensis and Nyssomia antunesi are unidentified potential vectors, and should be the focus of sampling and research efforts. Overall, we found that our machine learning approach provides valuable information for Leishmania surveillance and management in an otherwise complex and data sparse system.

Fig 1. ACL life cycle.

Female sandfly vectors pick up Leishmania parasites during their blood meal from reservoir hosts. Spillover events occur when a sandfly with Leishmania parasites in its salivary glands takes a blood meal from a human and infects the human with the parasite. Identity of all reservoir hosts and sandfly vectors are still unknown, which makes it hard to model and prevent transmission. Created with BioRender.com.

Fig 2. The model accurately classifies known vectors and identifies relatively few species of unknown status as likely vectors.

A distribution of predicted probabilities of sandfly species separated by vector status, and scaled by percentage. Red bars indicate the proportion of confirmed vectors that were predicted at that probability, while beige bars indicate the proportion of sandfly species not previously identified as vectors that were predicted at that probability.

Fig 3

Confirmed (A) and newly-predicted (B) vectors occur throughout the Americas. (A) Observed occurrences of confirmed vectors of Leishmania spp. that cause ACL, taken from GBIF and plotted in ArcGIS (Esri, USGS | Esri, Garmin, FAO, NOAA, USGS) [78,79]. (B) Observed occurrences of sandflies of unknown vector status that our models assigned a predicted probability above 0.5. Most predicted vectors are in Brazil due to more extensive survey efforts and availability of public data [78,79]. Maps showing species richness and vector distribution for each species of Leishmania spp. can be found in S4 and S5 Figs.

Fig 4. Human biting, study effort, and canopy height were the most important features for predicting vector status.

(A) Partial dependence plots of the top three variables from the BRT analysis showing the marginal dependence of each trait (shown in order of importance) on the probability of being a vector of ACL. The variable along with its average importance (on a scale of 0–1) are above each plot, the trait value is shown on the x axis, and the effect on probability is shown on the y-axis. The colored lines represent the marginal dependence of the trait from the 100 BRT models, while the solid black line represents the average dependence. The definition of each variable can be found in S1 Table. (B) Variable importance, scaled from 0–1, for the top 10 most important variables with 95% confidence intervals. Points represent mean gain value across 100 iterations. The importances for binary variables were summed up to obtain a single value for the entire categorical variable.