Using ecological niche modeling to investigate why the pronghorn ( Antilocapra ) did not disperse during the Great American Biotic Interchange

We examine whether ecological niche modeling (ENM) can account for the fact that some taxa did not disperse in the Great American Biotic Interchange (GABI) during the Plio-Pleistocene (5.3 million to 14,000 years ago). It has been hypothesized that the expansion of savannas throughout the Americas enabled dispersal for savanna-adapted taxa during GABI. One notable exception is the pronghorn ( Antilocapra ), which did not disperse to South America despite thriving in North American grasslands. We used ENM to test whether this lack of dispersal can be accounted for by the absence of appropriate climatic conditions connecting North and South America, or whether other factors need to be invoked to explain its lack of dispersal into South America. The analysis suggests that Antilocapra probably did not have suitable connected habitat in Central and South America during the time periods analyzed, and thus that the presence of savanna in of itself is too simple a characterization of at least the pronghorn’s favored habitat, or that other factors prevented it from dispering. However, given the uncertainty of the models, it is possible the pronghorn could have dispersed and then gone extinct. The niche models also predict that if Antilocapra was able to reach South America, it would have likely had suitable habitat in modern day Venezuela, Guyana, Suriname, and French Guiana.

The Great American Biotic Interchange (GABI) was a series of dispersals between North and South America throughout the Plio-Pleistocene. These dispersals utilized the newly emergent Isthmus of Panama as a bridge between the two continents. The timing of the Isthmus formation is uncertain and hypotheses have ranged from 23 to 2.8 Ma ago (Bacon et al., 2015; Montes et al., 2015; O’Dea et al., 2016; Jaramillo et al., 2017). The South American mammal fauna, characterized by marsupials and xenarthrans, had developed in relative isolation since the Early Eocene (Leigh et al., 2014), and the first fossil evidence of mammals moving from one continent to another are the South American megalonychid sloths, which appeared in North America around 9 Ma ago (Morgan, 2008). Conversely, North America had multiple rounds of dispersal given its periodic connections with Eurasia since the Mesozoic (Marincovich and Gladenkov, 1999), and probably as a consequence of this, had more genera than South America (Marshall et al., 1982). Regardless of when the Panamanian land bridge first formed, dispersals via Central America appear to have accelerated for mammals around 2.7 Ma ago (O’Dea et al., 2016).

Various ideas have been put forth to explain why certain animals dispersed during the Interchange, while others did not. Simpson (1950) was the first to propose a Central American faunal filter for GABI, but he did not discuss the nature of the filter other than to note it was ecological. Since then, it has been proposed that the expansion of savanna habitats may have led to selective dispersal across the Isthmus of Panama (Webb, 1978, 1991; Marshall, 1988; Vrba, 1992). Marshall (1988) also suggested distance and competition as possible individual barriers. It has been noted that niche breadth may have also been a determinant in which groups were able to disperse (Montoya Valencia, 2018).

There are many taxonomic lists of the mammalian groups that participated in the Interchange (Patterson and Pascual, 1968; Webb, 1976, 1978, 1991, 2006; Vrba, 1992; Woodburne et al., 2006; Freitas-Oliveira et al., 2025), but only a few studies (Montoya Valencia, 2018; Jackson, 2021) have analyzed groups that did not disperse during this period. According to Marshall et al. (1982), 11% or less of mammal genera from North America and 7% or less from South America participated in the interchange. Recently, studies have begun to use Ecological Niche Modeling (ENM) to assess the causes of and barriers to dispersal during GABI (Freitas-Oliveira et al., 2024; Magoulick et al., 2025). In Magoulick et al. (2025) we used ENM to assess dispersal pathways for Glyptotherium, a GABI participant. Here, we focus on a genus of non-dispersers from North America — Antilocapra, the pronghorn. Janis and Manning (1998) noted that the fact that Antilocapra stayed in North America is at odds with the hypothesis that savanna-adapted taxa dispersed during GABI.

Antilocapra americana (Ord, 1815), the only living member of the family Antilocapridae, currently ranges from southern Canada to Mexico (Figure 1, O’Gara, 1978), although the range is reduced from its historical extent (Yoakum and O’Gara, 2000). Antilocapridae arrived in North America from Eurasia by at least the early Miocene (21–23 Ma; Beatty and Martin, 2009), and during the Plio-Pleistocene many genera inhabited the continent including Capromeryx, Tetrameryx, and Stockoceros (White and Morgan, 2024). In addition to Antilocapra americana, the genus also includes the extinct A. pacifica (Richards and McCrossin, 1991) and possibly the extinct ? A. garciae (Webb, 1973) although there is disagreement over this last taxon with some arguing that it belongs in the genus Sphenophalos (Davis, 2007). These species exist throughout many North American Land Mammal Ages including the Hemphillian (9.4-4.7 million years ago), Blancan (4.7-1.4 million years ago), Irvingtonian (1.4 million-240,000 years ago), and Rancholabrean (240,000-14,000 years ago). ? A. garciae dates to the late Hemphillian (∼5.9-4.6 Ma), A. pacifica ranges from the Irvingtonian to the Rancholabrean (∼1.8 Ma–14,000 years), and A. americana continues from the Rancholabrean (240,000 years) to the present (Davis, 2007).

Maxent niche models for the extant Antilocapra americana have been produced for various parts of their range (Poor et al., 2012; Duncan et al., 2016; Zeller et al., 2021; Bean et al., 2024b), but there have been no attempts to use ENM to model past ranges of Antilocapra. Here we model past habitat suitability for Antilocapra to test the hypothesis that the expansion of savanna habitat was the primary driver of dispersal during GABI. Antilocapra americana appears to have maintained its niche over time (Pineda-Munoz et al., 2021; Smith et al., 2022), and so the genus is a good candidate for niche modeling. We modeled Antilocapra in the Rancholabrean (0.014–0.240 Ma), and then projected the model to the Late Blancan (1.80–2.588 Ma) as we wanted to examine the time during and just before the temporal range of A. pacifica and A. americana. We hypothesize that Antilocapra did not have continuous suitable conditions in between North and South America, and thus that the savanna expansion hypothesis is too simplistic to explain GABI patterns. We also compare the extent of suitable conditions for Antilocapra to our previous work using ENM on the taxon Glyptotherium, which did disperse during GABI.

Fossil locality data

We wanted to model pronghorn in the past rather than in the present because there is evidence that they have shifted their range and behavior as a result of human activities (Beckmann et al., 2012; Faurby and Araújo, 2018). Fossil occurrence data for Antilocapra were downloaded from the Global Biodiversity Information Facility (GBIF; accessed 03/14/2024; https://doi.org/10.15468/dl.rqfpry), the Paleobiology Database (PBDB; accessed 04/18/2024) (Uhen et al., 2023), and Neotoma (accessed 04/7/2025). These records were supplemented from the literature and data from White and Morgan (2024). We included all occurrences assigned to the genus Antilocapra including Antilocapra cf. americana, Antilocapra sp., and Antilocapra cf. Antilocapra sp. We opted to perform the analysis at the genus level because there were too few occurrences identified to the species level for the Rancholabrean. Taxonomic lumping is an established practice for paleoENMs (Varela et al., 2011; Smith et al., 2019; Blois et al., 2025). In the case of Antilocapra, lumping is supported by evidence showing that niches function and are conserved at the genus level for many North American mammals over the past 130,000 years (Hadly et al., 2009). Data and code can be accessed at (https://doi.org/10.5281/zenodo.21985290).

Following Magoulick et al. (2025), localities were sorted into the two time bins based on the time periods of the paleoclimate model: the Rancholabrean (0.014–0.240 Ma; n = 52) and the Late Blancan (1.80–2.588 Ma; n = 1). Occurrence data were rotated to their corresponding paleocoordinates in the Rancholabrean (112 ka) and the Gelasian (2.194 Ma; corresponding to the Late Blancan) using Scotese’s (2016) PaleoAtlas and GPlates ver.2.3.0 (Müller et al., 2018). Rancholabrean occurrences were subsequently filtered environmentally to reduce sampling bias (Varela et al., 2014). We filtered out occurrences that were not unique at a climate resolution of 2 units (i.e., combinations of 2 mm of precipitation and 2 °C), leaving 35 unique Rancholabrean occurrences for Antilocapra.

Paleoclimate model and environmental variables

To characterize the abiotic niche of Antilocapra, we used the climatic data from Magoulick et al. (2025). This data is from the HadCM3 paleoclimate model (HadCM3L-M2.1D) and has a resolution of 3.75° longitude (approximately ∼300 km in the tropics) and 2.5° latitude for the atmospheric and oceanic components of the model (Valdes et al., 2017, 2021). We used climate variables at two time slices, 112 ka ago and 1.80 - 2.588 Ma ago, which had already been bilinearly interpolated to 1° (Magoulick et al., 2025). These time slices were selected because the genus Antilocapra ranges from at least the Irvingtonian (1.8 million – 240 ka years ago) to the present (Davis, 2007). Precipitation and temperature have been noted to impact pronghorn populations (Gedir et al., 2015; Bean et al., 2024a). So, for climatic variables we chose annual mean temperature, annual precipitation, minimum temperature of the coldest month, maximum temperature of the warmest month, precipitation of the driest month, and precipitation of the wettest month as our potential climatic predictors for use in the ENMs. These variables were also limited by paleoclimatic data availability. We excluded annual mean temperature and annual precipitation which had variance inflation factors (VIFs) greater than 10 (Montgomery et al., 2021), and thus were collinear (Magoulick et al., 2025).

Ecological niche modeling

We modeled environmental suitability using MaxEnt ver.3.4 (Phillips et al., 2017). MaxEnt is a niche modeling software that uses maximum entropy modeling via machine learning. We chose MaxEnt because it is commonly used for paleontological studies, as it does not require absence data which is notoriously difficult to obtain for the fossil record (Myers et al., 2015). However, MaxEnt can sometimes have issues with overfitting and transferability (Peterson et al., 2007). We used the aggregate function in the terra package ver.1.8.54 (Hijmans, 2025) to create a 750 km buffer around occurrences in order to select at least 1000 background points following Magoulick et al. (2025).

Niche modeling was performed in R ver.4.5.0 using kuenm ver.1.1.10 (Cobos et al., 2019). We produced 2,030 candidate models using combinations of feature classes, which determine the shape of the response curves, regularization multipliers, which act as a penalty to avoid overfitting, and environmental variables. For feature classes we used all combinations (l =linear, q =quadratic, p =product, t =threshold and h =hinge), excluding p by itself and t by itself as they require 80 samples (Elith et al., 2011). Models were created with a minimum of three of the four environmental variables per model and regularization multipliers from 0.1 to 1.0 with intervals of 0.1, and from 2.0 to 5.0, with intervals of 1.0.

Models were evaluated using five-fold cross-validation and 1,000 background points. We used free extrapolation so the model could “extrapolate” to new environmental conditions, but not clamping, which “clamps” (or holds) the projection to the calibration area values (Owens et al., 2013).

We used kuenm to select candidate models based on partial receiver operating characteristic (partial ROC; Peterson et al., 2008), omission rates (E = 5%, Anderson et al., 2003), and ΔAICc (up to 2, Burnham and Anderson, 2002). This resulted in 58 candidate models, which we then evaluated using area under the curve (AUC) of the partial ROC (Peterson et al., 2008), the continuous Boyce index (CBI; Boyce et al., 2002, Hirzel et al., 2006), the true skill statistic (TSS; Allouche et al., 2006), and kappa (Cohen, 1960). We also varied E to 15% which resulted in the same 58 candidate models. We selected models with AUC > 0.70 and TSS > 0.4 following Zhang et al. (2015), and kappa > 0.1 of which there were only two (Table 1). We calculated CBI using ecospat.boyce in ecospat ver.4.1.2 (di Cola et al., 2017) and TSS and kappa using (https://github.com/KarlssonCatharina/MaxEnt_TSS_Calculations). These two models were then projected to the Late Blancan.

We used terra ver.1.8.54 (Hijmans, 2025) to calculate mean models from the median of the five cross-validation runs (output by Maxent) for each model. We converted the mean suitability maps into binary suitability maps by maximizing the sum of sensitivity and specificity (maxSSS threshold, Manel et al., 2001) and 95% least training presence (LTP, Pearson et al., 2007). For binary suitability, everything above the threshold is considered to be suitable conditions (equal to 1) and everything under is considered unsuitable conditions (equal to 0). With these binary suitability maps we again used terra to calculate the consensus of the mean models (i.e. areas where all binary suitability maps were equal to 1). To compare the models to the present-day climatic niche of Antilocapra, we extracted the maximum and minimum values for each of the environmental variables from the binary suitability maps. To calculate the corresponding modern climatic ranges we used the equivalent WorldClim 2.1 bioclimatic variables at a 10 minute resolution (Fick and Hijmans, 2017) and bilinearly interpolated them to 1° to match the paleoclimate data. We used the modern range of Antilocapra americana based on the IUCN range map (NatureServe, 2008) (Figure 1).

Fragmentation statistics

Finally, we compared the suitability results to those of Glyptotherium from Magoulick et al. (2025). We calculated a buffer of 250 kilometers around the countries of Mexico, Guatemala, Belize, El Salvador, Honduras, Nicaragua, Costa Rica, and Panama. We examined Glyptotherium and Antilocapra median consensus maps in the Rancholabrean. We used terra ver.1.8.54 (Hijmans, 2025) to calculate the proportion of suitable cells in the buffer area. Then buffer area maps were converted to Albers Equal Area (EPSG:9822) and saved as GEOtiff files. These files were uploaded to FRAGSTATS ver.4.2 (McGarigal et al., 2023) to calculate the cohesion (class level) for suitable area. Cohesion or patch cohesion index is an aggregation metric that ranges from 0 to 100, increasing as the patch (in this case suitable habitat) becomes more aggregated or connected (McGarigal et al., 2023).

The final two models included three of the four climatic predictors considered: minimum temperature of the coldest month and precipitation of the wettest and driest months. Their AUC, TSS, kappa, and CBI values are reported in Table 1. The AUC and TSS values indicate moderate model performance, although the kappa values are poor (Zhang et al., 2015). The CBI values suggest the models have good transferability to different regions.

In the Late Blancan, both LTP and maxSSS suitability maps show no direct connection between suitable climatic conditions in the potential dispersal corridor in Central America for Antilocapra, however the LTP map has large patches of suitable habitat close to each other so dispersal may have been possible. Both LTP and maxSSS maps predict the suitable habitat in the location of the Late Blancan occurrence of Antilocapra (Figure 2), which increases confidence in the model. The maxSSS map for the Rancholabrean also shows little connection between North and South America (Figure 2).

The Rancholabrean LTP map indicates possible connections between suitable areas with some exceptions (possibly due to resolution dependency) around the Isthmus and in southern Mexico (Figure 2), but no continuous connection all the way from North America to South America.

Proportions of total suitable area in Central America and Mexico during the Rancholabrean, representing the amount of suitable conditions in the potential dispersal corridor, were 38.9% (maxSSS) and 71.4% (LTP) for Gyptotherium and 38.6% (maxSSS) and 75.8% (LTP) for Antilocapra. Cohesion was 95.9 (maxSSS) and 97.8 (LTP) for Glyptotherium and 91.8 (maxSSS) and 98.5 (LTP) for Antilocapra indicating that the suitable conditions were similarly aggregated for both genera.

We compared the model outputs to the present-day climatic niche of Antilocapra americana. The LTP and maxSSS precipitation values bracket the modern day values, but the temperature values differ (Table 2, Figure 3).

We evaluated suitable climate conditions for Antilocapra using ENM in order to examine whether climatic factors, and by extension savanna habitat, inhibited their dispersal into South America during GABI. We found that although niche models predict discontinuity in suitable conditions for Antilocapra in Central America, proportion of suitable area and cohesion were not that different between Antilocapra and Glyptotherium, which did disperse during the Interchange. In the Rancholabrean, Glyptotherium and Antilocapra had similar proportions of suitable area, and the LTP thresholds show similar cohesion, yet visually Glyptotherium appears to have more continuous suitable conditions (Figure 2). The maxSSS thresholds did diverge with Glyptotherium having slightly higher cohesion. Thus, it is ambiguous whether or not Glyptotherium or Antilocapra could have dispersed during the Rancholabrean. In the case of Glyptotherium, the analysis Magoulick et al. (2025) shows that there are corridor conditions during the Middle Blancan. Perhaps this corridor was gone or was patchier by the time A. americana evolved – certainly the data show highly fragmented patches of suitable habitat in Central America in the Late Blancan for Antilocapra (Figure 2), in contrast to the almost continuous habitat for Glyptotherium.

The extracted upper and lower bounds for the climatic niche of Antilocapra do not align with the modern day range of A. americana, but they do bracket the living species’ current niche. The minimum temperature of the coldest month has a smaller range in the present day than either LTP or maxSSS, and the maximum value is much closer to zero (Table 2, Figure 3). For the precipitation of the wettest and driest months, the modern range of A. americana falls between LTP and maxSSS (Table 2, Figure 2), indicating that perhaps habitat suitability also falls somewhere between the two estimates. It is important to note that Pronghorn exhibit seasonal migration both now (Sawyer et al., 2005; Xu et al., 2023) and in the past (Fraser et al., 2021) and so utilize different parts of their niche at different times in the year. With present day data, this seasonal migration is reflected in ENMs. One such study found that A. americana had a niche with winter mean minimum temperatures warmer than -10 °C (Bean et al., 2024b), which is similar to the lower end of the maxSSS range (Table 2). The wide range of the LTP threshold could mean that the models have calculated too much suitable habitat, or that the niche of pronghorn has changed in the last few thousand years.

We acknowledge the AUCs of the models are relatively low, but even some niche models of Antilocapra made with modern data have relatively low AUCs (Poor et al., 2012; Duncan et al., 2016; Zeller et al., 2021). This may mean that it is difficult to reliably capture the niche of pronghorn via ENM. Since the range of Antilocapra has been impacted by humans, we opted not to model the present-day niche of Antilocapra, and instead compared the climatic variable values from our model to values from the modern range. However, as with the models of Glyptotherium, there are many sources of bias in our models of Antilocapra that could impact the results. These include coarse spatial resolution of the climate model, fluctuating climate on finer timescales not captured by the model, and low sample size for the fossil occurrences (Magoulick et al., 2025). For example, the shift between glacials and interglacials caused vegetation changes (Correa-Metrio et al., 2012; Cárdenes-Sandí et al., 2019; Lozano-García et al., 2022) which could in turn lead to niche changes that are currently on too fine a scale for the models to pick up. In our case, the low number of occurrences for Antilocapra and the transferability of MaxEnt models also add uncertainty. Some concerns have also been raised by researchers that Antilocapra fossils may not all belong to the genus (Díaz Sibaja, 2018; White and Morgan, 2024). This may mean that the model is based on occurrences that include other genera, and thus represents a larger climatic niche than corresponds to Antilocapra. We have also assumed that the niche is conserved on the genus level for Antilocapra and through time from the Rancholabrean to the Late Blancan.

Overall, further work is needed to fully explain why Antilocapra (and other groups) did not disperse during GABI, but, the fossil record and models, despite their limitations do provide a reasonable explanation for why Antilocapra did not make it to South America — the dispersal corridor did not have continuous climatically suitable conditions.

As noted by Janis and Manning (1998), the lack of fossil occurrences of Antilocapra in South America calls into question Webb’s hypothesis that the expansion of savannas (during glacial time periods) was what enabled dispersal during GABI. The results suggest that there may not, in fact, have been a climatically suitable corridor for Antilocapra, indicating that savanna habitats alone are probably not sufficient for dispersal. Further, recent palynological analyses in Panama have not found evidence of savanna environments in the last 20 million years (Jaramillo et al., 2014), indicating that at least the area around the isthmus may not have been a savanna corridor. Likewise, Correa-Metrio et al. (2012) found that in the Yucatan savanna expansion occurred but was not correlated with glaciation, calling into question Webb’s hypothesis.

Nonetheless, it is also possible that pronghorns did cross the Panamanian Isthmus but their fossils have not yet been found in South America. Historically, the Amazon Rainforest (which covers much of the northern part of the continent) was seen as a poor area for finding vertebrate fossils (Ranzi, 2000) given the dense vegetation and acidic soil. So it is not implausible that the genus once lived there, but has not yet been seen in the fossil record. However, recently, an increasing number of mammalian fossils have been found in the region (Asevedo et al., 2021), but as of yet there is no evidence of Antilocapra. If members of Antilocapra did disperse to South America, the models suggest that Rancholabrean fossils could be present in modern day Venezuela, Guyana, Suriname, and French Guiana (Figure 2). In North America, pronghorns have a stable population (assessed as Least Concern (IUCN SSC Antelope Specialist Group, 2016)), so their current absence in South America is more stark. Thus, overall, it seems unlikely that Antilocapra ever dispersed to South America.

Interestingly, the North American Cheetah (Miracinonyx), one of the primary predators of Antilocapra, similarly did not disperse to South America (Higgins et al., 2023). It is possible that they may have had similar niche requirements or may have had their dispersal inhibited for similar reasons. Nascimento and Pires (2025) suggest that predation by felids may have led to the antilocaprid extinctions in North America (except the extant pronghorn), and so it is possible that the presence of other felids during GABI prevented Antilocapra, and the other antilocaprids, from dispersing.

Although the results do not show continuous suitable climatic conditions for Antilocapra in Central America, consistent with it not dispersing, the possibility that they did disperse given shifting climatic conditions over short time scales that may not be captured by the models cannot be ruled out. Further work is needed to refine the understanding of the nature of the Central American Filter during GABI.