Ticks vector the most diverse range of zoonotic pathogens and are responsible for the majority of vector-borne diseases in the United States (Eisen et al. 2017). Among ticks, the best studied species belong to the family Ixodidae (Acari), responsible for vectoring 40% of the 131 emerging vector-borne zoonotic diseases, including Lyme disease, Rocky Mountain spotted fever, and ehrlichiosis, among many others (Swei et al. 2020). Incidences of these diseases are increasing; for example, the number of tick-borne human disease cases doubled between 2004 and 2016 in the United States, primarily due to Lyme disease (Rosenberg et al. 2018). In northern temperate regions in the United States, some tick species are expanding their geographic range northward and westward in part due to climate change, increasing the chances of zoonotic pathogen transmission to humans (Sonenshine 2018, Couper et al. 2021). While ticks and emerging tick-borne diseases are better studied in the United States, less information is available for Africa and South America (Swei et al. 2020). There is a high diversity of ticks in southeast Asia, yet relatively little is known about their ecology, distribution, and host–parasite relationships (Petney et al. 2019, Sharifah et al. 2020).

While similarly globally distributed and responsible for the transmission of multiple pathogens causing zoonotic disease in humans, ticks in the genus Dermacentor Koch 1844 have comparatively received less research attention than ticks in the genus Ixodes Latreille 1795 (Acari: Ixodidae), which over the past several decades have become infamous as the primary vectors of Lyme disease in eastern North America. While studies reporting humans being bitten by Dermacentor ticks are relatively few compared to Ixodes, multiple Dermacentor species have been reported to parasitize humans (Guglielmone et al. 2020, Mathison and Sapp 2021). Diseases confirmed to be vectored by Dermacentor ticks include Rocky Mountain spotted fever, other rickettsioses (spotted fevers), tick-borne encephalitis, and tularemia ( https://www.cdc.gov/ticks/). New Dermacentor tick species continue to be recognized (Apanaskevich et al. 2021, Lado et al. 2021).

The increasing incidence of tick-borne zoonoses worldwide has prompted questions about what causes some tick species to be more effective vectors of zoonotic pathogens compared to others. Identifying particular traits that confer this propensity may highlight mechanistic hypotheses about what drives vector competence across species and across different life stages within a species. To explore this, Yang and Han (2018) compiled data on the intrinsic traits of tick species in the genus Ixodes to predict their zoonotic vector status. The present work builds on that study by creating a global database of features for Dermacentor at the species level. We systematically compiled trait data, including morphological characteristics, host range, geographic distribution, and zoonotic diseases that they vector. We then trained a machine learning algorithm (generalized boosted regression) via the R package gbm (Greenwell et al. 2020, R Core Team 2022) on these species-level traits to predict zoonotic vector status. This modeling approach identified specific traits that had the highest influence on vector status, and also predicted additional Dermacentor species that may be undiscovered zoonotic vectors. Connecting the traits of individual tick species to their zoonotic vector status has implications for more targeted surveillance measures based on disease risk from vector species, contributing to disease prevention, mitigation, and management.

Materials and Methods

Data Analysis

To identify which tick traits are the most important predictors of zoonotic disease status and, based on those traits, which Dermacentor species are most likely to vector disease, we used a machine learning technique, generalized boosted regression (package gbm in R version 4.1.3; Elith et al. 2008, Greenwell et al. 2020, Ridgeway 2020, R Core Team 2022). We chose this method following Yang and Han (2018), which utilized this technique for a similar model of Ixodes traits and which allowed us to directly compare model results for these two tick genera. Generalized boosted regression iteratively fits thousands of classification or regression trees by incorporating boosting, which combines multiple weakly predictive trees into an ensemble with superior predictive accuracy. This approach accommodates multiple data types, hidden and nonlinear interactions among multiple variables, and can also handle missing data by treating ‘missingness’ as a value for splitting decisions (Elith et al. 2008). The fitted model can be used to identify which variables are most important for accurate classification and to make predictions about novel vector species as new host and geographic data records become available. Our model classified tick species as vectors or nonvectors, a binary outcome.

To decrease the chance of overfitting on a small data set, we reduced the number of predictor variables and also employed 4-fold cross-validation within the gbm and generated a null model against which to compare gbm performance. To prune the number of predictor variables we excluded variables with near zero variance and/or less than 33% coverage across species. We then performed a grid search to tune the input parameters. The parameter values giving the highest training accuracy and stable deviance curve were learning rate = 0.0001, maximum tree depth = 3, minimum number of observations in node = 1, with a maximum number of trees = 80,000 and 70/30 split between training and test data. Boosted regression tree models often split data into a larger training dataset and smaller test set (Leathwick et al. 2006), especially when the dataset is small as in our case. Using these parameter values, we bootstrapped our gbm model (n = 50 runs), in each run using a randomly selected 70% of the data for training. To evaluate model performance, we computed the area under the receiver operating characteristic curve (AUC) for predictions made on the test data set (remaining 30% of data) for each bootstrap model run. To generate null models using this bootstrapping procedure (50 runs), binary labels (vector status) were randomly reassigned in each run (target shuffling, as in Fischhoff et al. 2021). We then corrected our mean test AUC by subtracting the difference between the mean test AUC of null models and the expected value for a model performing no better than chance (AUC = 0.5). Bootstrapped models give mean and standard error of the relative influence of each predictor variable and were also used to generate partial dependence plots that depict the importance of each feature for classification accuracy.

Fig. 1.

Mean relative influence of the top predictor variables. Relative influence indicates the importance of each variable in reducing prediction error. Error lines represent ± 1.5 × IQR, where IQR is the interquartile range between the first and third quartiles, generated from 50 bootstrap runs of the generalized boosted regression model. Morphological measurements are given in millimeters. Predictor variables are defined in Supp.  Table S1 (online only).

To assess the possible effects of sampling bias on classifying vector status, we also ran a gbm model with the number of citations for each species (citation count) as the response variable. This model enabled us to determine whether study effort is predictive of vector status among Dermacentor species, and if so, whether our models of vector status and study effort reflect similar trait profiles.

We further investigated the relationships between mammalian host range (number of host families, number of host orders, and host phylogenetic diversity) and geographic range (number of biomes, ecoregions, and climate zones) using scatterplots and Spearman's rank correlation tests for each pair of variables in R version 4.1.3 (R Core Team 2022).

Results

In sum, we collected data for 44 species of Dermacentor ticks for a total of 87 variables (  Supp Table S2 (tjac125_suppl_supplementary_table_s2.xlsx) [online only]). The following species were excluded from our gbm analysis due to lack of sufficient accessible data: D. asper, D. dispar, and D. pomerantzevi. A full list of predictor variables, their definitions, and coverage is included in   Supp Table S1 (tjac125_suppl_supplementary_table_s1.docx) (online only). Of the 41 species included in our analysis, 10 species (24%) were reported to be zoonotic vectors by GIDEON and/or Mathison and Sapp (2021;   Supp Table S3 (tjac125_suppl_supplementary_table_s3.docx) [online only]).

Our compilation of Dermacentor species traits encompassed 76 morphological characteristics for all life stages (larvae, nymphs, and both male and female adults), such as body size, length and width of the capitulum (mouthparts), hypostome (used to anchor the tick during feeding), and tarsi (leg segment that bears a sensory organ used for locating hosts; Sonenshine and Roe 2013;   Supp Table S1 (tjac125_suppl_supplementary_table_s1.docx) [online only]). The number of mammalian host families and orders for individual tick species ranged from 1 to 21 (families) and 1 to 8 (orders); host families and orders for each species are provided in   Supp Table S5 (tjac125_suppl_supplementary_table_s5.xlsx) (online only). We compiled a total of 6,786 georeferenced occurrence points in our geographic database, of which 3,969 were used for tallying the number of biomes, ecoregions, and climate zones for each species (  Supp Table S6 (tjac125_suppl_supplementary_table_s6.xlsx) [online only]). We excluded geographic occurrence points from counts when they could not reliably be assigned to an ecoregion or climate zone.

Our corrected mean test AUC was 0.84 (corrected test AUC = mean test AUC [0.95] – (mean test AUC for the null model [0.61] – 0.5)), indicating that our models trained on trait data correctly classified vector status among Dermacentor ticks with ∼84% accuracy and a minimal degree of overfitting (correction factor = 0.11) lending confidence that variable pruning, cross-validation, and correction via target shuffling combined to effectively address the learning constraints of a small data set.

We found that host range (number of mammalian host families) had the highest relative influence on predicting vector status, followed by hypostome width of nymphal ticks and body length of larval ticks (Fig. 1). Additional variables with low relative importance (but >1%) included host phylogenetic diversity, scutum width of larval ticks, geographic range (number of ecoregions), citation count, body length of (engorged) female ticks, and several other morphological characteristics.

Some of the top predictors indicated by our model, such as the number of host families, subsume sampling and publication bias inherent in most ecological data sets. For instance, more hosts are documented for tick species that are better studied, and fewer hosts are known for ticks that are understudied. Other predictors, such as morphological characteristics, however, are not always subject to this same bias. The results of our model using traits to predict the number of citations for a given Dermacentor species indicated that tick traits poorly predict study effort (the number of citations), with a pseudo-R² of 0.05. These results suggest that while study effort does vary widely across this group, our model is learning and generating predictions using the features of zoonotic vectors rather than the features of species that are simply better studied.

In order to better understand the relationships between the top predictor variables and vector status, we generated partial dependence plots, which show the effect of a given trait on prediction accuracy while controlling for the effects of all other predictor variables in the model. We found that, while most tick species have a small number of host families (gray bars, Fig. 2), those that infest more than 10 host families are more likely to be zoonotic disease vectors (black lines, Fig. 2). Dermacentor species with larger hypostomes during their nymphal stage and species with smaller-bodied larvae were also more likely to be vectors, as were species found in a high number of ecoregions.

Our model correctly classified nine of the ten confirmed Dermacentor vector species, as determined by having over 50% probability of being a vector using predictions made with the full data set. Although D. similis, a known vector species described in 2021 by Lado et al., was not classified as a vector species by our model, it was close to our 50% probability cutoff (49% likely to be a vector). Other species near the cutoff include D. niveus, which our model assigned a 31% probability of being a vector based on its traits profile; and D. parumapertus (22% probability). Both species are not currently known to be vectors.

We used our geographic database (  Supp Table S6 (tjac125_suppl_supplementary_table_s6.xlsx) [online only]) to generate a global map of Dermacentor species distributions (Fig. 3). We also quantified Dermacentor species richness in three areas of interest—North America, Europe, and Asia—and found that areas with the greatest species richness occur in Asia, despite the relative scarcity of trait data for species in this region (Fig. 4).

We consistently found significantly positive relationships between mammalian host range (number of host families, number of host orders, and host phylogenetic diversity) and geographic range (number of biomes, ecoregions, and climate zones); all p-values were of <0.01. Results of Spearman's rank correlation tests are provided in   Supp Table S7 (tjac125_suppl_supplementary_table_s7.docx) (online only), and these relationships are visualized in  Supp Fig. S1 (tjac125_suppl_supplementary_figure.docx) (online only).

Discussion

We systematically reviewed the literature to compile a comprehensive, first of its kind data set for the genus Dermacentor, an understudied group of zoonotic pathogen-vectoring ticks. In particular, we collated and standardized detailed information on species' morphological characteristics, the diversity of host families they infest, and their geographic distribution, which represents a valuable ecoinformatic contribution to the knowledge base about this group. Our analyses suggest that Dermacentor species that are the most likely to be zoonotic vectors are broad ranging, both in terms of the range of hosts they infest and the range of ecoregions across which they are found, and also tend to have large hypostomes to anchor themselves to their hosts and be small-bodied as immatures.

The most important predictor of zoonotic vector status was host range, specifically, the number of mammalian host families a tick species is known to infest. Host phylogenetic diversity was also among the top predictor variables. Five mammalian families host at least half of all Dermacentor species (Bovidae, Suidae, Canidae, Muridae, and Leporidae). Bovidae, Suidae, and Canidae include domestic species (e.g., cattle, pigs, and dogs), and Muridae (rodents) are often human-associated species. A promising area for future research would be to compare the host families of vector and nonvector tick species by life stage. For pathogens in general, the ability to infect a broad range of hosts is associated with a high risk of emergence (Cleaveland et al. 2001, Woolhouse and Gowtage-Sequeria 2005). We postulate that ticks with a broad host range are also more likely to encounter a high diversity of pathogens, including zoonotic pathogens (Estrada-Peña et al. 2015), or they are more likely to encounter a single broadly distributed hyper-reservoir species (e.g., Han et al. 2015) that carries numerous zoonotic pathogens to which the tick is susceptible or for which it is a competent vector.

Mirroring broad host range, Dermacentor ticks found in a wide variety of ecoregions were more likely to be zoonotic vectors. We also found a positive correlation between host range and geographic range for Dermacentor ticks ( Supp Fig. S1 (tjac125_suppl_supplementary_figure.docx) [online only],   Supp Table S7 (tjac125_suppl_supplementary_table_s7.docx) [online only]). While species with broader host and geographic ranges tend to be those that are better studied, our analyses showed that this study bias did not extend to the intrinsic morphological traits distinguishing Dermacentor vectors from nonvectors in our study. An outstanding area for future work is to determine overlapping ranges of these zoonotic vectors with potential reservoir host species. To this end, our high-resolution geographic location database could be used to support species distribution modeling for Dermacentor ticks, which has thus far been limited to a few well-known zoonotic vector species on regional (Široký et al. 2011, Walter et al. 2016, Huercha et al. 2020) or continental (Minigan et al. 2018, Boorgula et al. 2020) scales. On a global scale, VectorMap ( http://vectormap.si.edu/), one of the sources we used for our geographic data set, has mapped the distributions of and utilized predictive climate suitability models for several Dermacentor tick species. The global scope of our geographic data set enables augmentation with climatic data from sources such as TerraClimate (Abatzoglou et al. 2018) to facilitate more comprehensive species distribution models, generating suitability maps of Dermacentor species under the present and future climate scenarios. Taking seasonality into account, climate optima for different Dermacentor species could be determined, as has been done with thermal optima for mosquito vectors, to postulate the resulting implications for human disease incidence with a changing climate (Mordecai et al. 2019).

For Dermacentor tick nymphs, species with wider hypostomes were more likely to be zoonotic vectors. The hypostome is inserted into the host and serves to anchor the tick in place during blood feeding (Richter et al. 2013, Sonenshine and Roe 2013). Larger hypostomes could form a more secure anchor (Richter et al. 2013), enabling a longer period of attachment and increasing the probability of pathogen transmission (Kaufman 2010). We also found that smaller-bodied immature Dermacentor ticks were more likely to be vectors. It is possible that the results for body size of immature ticks was driven by the frequency distribution of this trait. In other words, the tendency to be classified as a vector species (Fig. 2, black line) tracks with the frequency distribution of this trait across Dermacentor species (Fig. 2, gray bars). Less information is available about immature Dermacentor ticks and although there were few records of human infestation by immature Dermacentor ticks (  Supp Table S5 (tjac125_suppl_supplementary_table_s5.xlsx) [online only]), we postulate that immature ticks could be acquiring pathogens from wildlife (e.g., rodents) and transmitting them to humans as adults. Finally, we found that species with longer body sizes as engorged females were more likely to be zoonotic vectors, similar to Yang and Han (2018). This may serve as a rough proxy for fecundity and population abundance, if body length varies reliably with the maximum size of egg clutches.

Fig. 2.

Partial dependence plots for a selection of top predictor variables from the generalized boosted regression model used to predict the vector status of Dermacentor ticks. The black line represents the average marginal effect (y-axis, left) of a given trait (x-axis) on vector status after accounting for the average effect of all other predictor variables in the model. Gray bands represent 95% CI. The histograms show the relative frequency (y-axis, right) of tick species with a given value of each trait. Morphological measurements are given in millimeters.

Fig. 3.

Global distribution of Dermacentor tick species.

Fig. 4.

Species richness of Dermacentor ticks in North America (A), Europe (B), and Asia (C). Species richness ranges from a single species (lighter colors) to eight species (darker colors).

Among known vector species, our model correctly classified all but one species, D. similis, and identified two additional potential vector species. Based on molecular and morphological traits, Lado et al. (2021) determined that the population of D. variabilis in the western United States constituted a new species, D. similis. We suspect that because it was recently documented, the diversity of hosts as well as the geographic distribution of this species may be underdescribed; for our data set, we excluded records where it could not be determined if a specimen was D. variabilis or D. similis. Our model identified D. niveus and D. parumapertus, based on trait similarity to confirmed vector species, to be the highest probability species not currently classified as vectors. Dermacentor niveus has a broad host range (12 host families) and is known to infest humans on rare occasions (Guglielmone et al. 2020). The known distribution of this species is in mountainous regions of central Asia across six ecoregions. Crimean-Congo hemorrhagic fever virus was detected in D. niveus in Kazakhstan (Onishchenko et al. 2005) but definitive evidence that it vectors this virus to humans is currently lacking. Dermacentor parumapertus is found in North America across 16 ecoregions and is known to parasitize 10 host families. Rickettsia parkeri was detected in D. parumapertus ticks from black-tailed jackrabbits in Mexico (Sánchez-Montes et al. 2018) and the western United States (Paddock et al. 2017). Although D. niveus and D. parumapertus are known to parasitize humans, they do so rarely (Guglielmone et al. 2020). This suggests that while they share trait similarities with other known zoonotic vectors, host preference in these species may confer some protection against transmission of zoonotic pathogens to humans, or that they are found in sparsely populated areas.

Mapping the distributions of known and predicted pathogen-vectoring ticks is a necessary step toward assessing high-risk areas for tick-borne disease (Eisen and Paddock 2020). Our finely resolved geographic location database that encompasses the entire genus of Dermacentor ticks has the potential to fill in key gaps in our knowledge of tick distributions for species whose geographic ranges have not been fully delineated. Taking advantage of this potential could generate critical predictions of high-risk areas in understudied regions such as southeast Asia and South America as anthropogenic climate change continues to shift the distributions of hosts and vectors. In particular, our global map of Dermacentor species richness underscores the need for additional research in Asia, a research gap that was also recognized in a recent review of mapping tick-borne disease (Lippi et al. 2021).

During a time when anthropogenic impacts such as climate and land use change appear to be favoring the continued spread of tickborne diseases, knowledge about these vector species is even more crucial (Swei et al. 2020). Our database of Dermacentor species morphological characteristics, host range, and geographic distribution has relevance beyond the scope of our study. The information we compiled can be used for public health applications, such as species distribution modeling, to determine areas deserving increased surveillance measures.