In situ adaptation and ecological release facilitate the occupied niche expansion of a non‐native Madagascan day gecko in Florida

Abstract Aim To investigate whether the frequently advocated climate‐matching species distribution modeling approach could predict the well‐characterized colonization of Florida by the Madagascar giant day gecko Phelsuma grandis. Location Madagascar and Florida, USA. Methods To determine the climatic conditions associated with the native range of P. grandis, we used native‐range presence‐only records and Bioclim climatic data to build a Maxent species distribution model and projected the climatic thresholds of the native range onto Florida. We then built an analogous model using Florida presence‐only data and projected it onto Madagascar. We constructed a third model using native‐range presences for both P. grandis and the closely related parapatric species P. kochi. Results Despite performing well within the native range, our Madagascar Bioclim model failed to identify suitable climatic habitat currently occupied by P. grandis in Florida. The model constructed using Florida presences also failed to reflect the distribution in Madagascar by overpredicting distribution, especially in western areas occupied by P. kochi. The model built using the combined P. kochi/P. grandis dataset modestly improved the prediction of the range of P. grandis in Florida, thereby implying competitive exclusion of P. grandis by P. kochi from habitat within the former's fundamental niche. These findings thus suggest ecological release of P. grandis in Florida. However, because ecological release cannot fully explain the divergent occupied niches of P. grandis in Madagascar versus Florida, our findings also demonstrate some degree of in situ adaptation in Florida. Main conclusions Our models suggest that the discrepancy between the predicted and observed range of P. grandis in Florida is attributable to either in situ adaptation by P. grandis within Florida, or a combination of such in situ adaptation and competition with P. kochi in Madagascar. Our study demonstrates that climate‐matching species distribution models can severely underpredict the establishment risk posed by non‐native herpetofauna.

Despite their widespread use, some species distribution modeling approaches have been criticized for their "ecologically untenable" assumptions (Dormann, 2007:387) and inability to capture and characterize environmental heterogeneity at biologically relevant spatial scales (Sears & Angilletta, 2015). Also problematic is the fact that SDMs are often constructed using observed realized niche data (Pearson & Dawson, 2003;Veloz et al., 2012), when in fact the focal taxon's fundamental niche may be significantly larger, but constrained by factors including dispersal limitations and biotic interactions (Boulangeat et al., 2012;Li et al., 2014;Pearson & Dawson, 2003;Rodriguez-Cabal et al., 2012). In addition, SDM predictions can vary dramatically according to the data and assumptions on which they are built (e.g., Anderson & Raza, 2010;Dowell et al., 2016;Pearson et al., 2006;Pyron et al., 2008;Radosavljevic & Anderson, 2014), leading to uncertainty when interpreting their outputs. Considerable shortcomings such as these have led some to conclude that climate-matching SDMs may not be warranted as a risk assessment tool for non-native herpetofauna (Li et al., 2014).
In this study, we tested the predictive accuracy of the climatematching species distribution modeling approach using range data for the Madagascar giant day gecko Phelsuma grandis Gray 1870 ( Figure 1). In both its native and colonized range, P. grandis can be found in a variety of habitat types, including primary forests, orchards, highly degraded forests, and anthropogenic habitats (Blumgart et al., 2017;D'Cruze & Kumar, 2011;Dubos et al., 2014;Krysko et al., 2019;Krysko et al., 2003;Raselimanana et al., 2000;Raxworthy & Nussbaum, 1994); we therefore expected that habitat variables would be poor predictors species distribution models can severely underpredict the establishment risk posed by non-native herpetofauna.

K E Y W O R D S
competitive exclusion, ecological niche modeling, fundamental niche, herpetofauna, nonnative species, Phelsuma grandis, Phelsuma kochi, realized niche, reptiles, species distribution modeling F I G U R E 1 Non-native Madagascar giant day gecko (Phelsuma grandis) in situ on a non-native coconut palm (Cocos nucifera) on Grassy Key, Monroe County, Florida, USA. Photograph (UF-Herpetology photographic voucher 170124) by Kenneth L. Krysko of P. grandis occupancy and thus concluded that a climate-matching SDM approach was desirable. Using native-range presence-only data, we generated a predicted distribution for P. grandis in Florida-a region in which the species is well established and its range well documented (Fieldsend & Krysko, 2019b)-which we then compared with the observed distribution. We also built an analogous model using Florida P. grandis presence-only data, which we projected onto both Florida and Madagascar, allowing us to check the degree of agreement between the outputs of the two models. Finally, we built a model combining native-range presence data for P. grandis and the parapatric , closely related species P. kochi Mertens 1954, to test whether our climate-matching SDM approach provided evidence for the competitive exclusion of P. grandis from a portion of its fundamental niche by P. kochi.

| ME THODS
We compiled a dataset of 71 georeferenced native-range Phelsuma grandis presence points (Appendix S1). Only confirmed observations from peer-reviewed literature were included in the dataset.
Points were checked against the species' known native distribution in northern Madagascar  to confirm their accuracy. Global Biodiversity Information Center (GBIF) data are often used in the construction of SDMs (e.g., Mothes et al., 2019;Nania et al., 2020;Suzuki-Ohno et al., 2017;Weterings & Vetter, 2018), but were not included in this list of presences as they include iNaturalist "Research Grade" observations (Boone & Basille, 2019) of P. grandis, many of which are actually misidentifications of P. kochi or P. madagascariensis Gray 1831 (pers. obs.). A preliminary projection of the 71 presence points onto the native range showed that nine were located fractionally offshore due to either recording or projection errors, and exacerbated by the partially coastal distribution of the species (see Appendix S1), leaving a total of 62 presence records, well above the minimum number required to develop an adequate Maxent model van Proosdij et al., 2016).
We collated 239 georeferenced observations of P. grandis from southern Florida from the Florida Museum of Natural History's Division of Herpetology records and from verified personal observations by the authors (Appendix S2). Two data points were removed as they were known to represent either singleton records or nowextirpated populations. Duplicate coordinates were then also removed, resulting in 115 unique records. Due to the coarse resolution of the spatial data relative to the small size of some of the Florida Keys, only 70 of these 115 records were categorized as being on land, with the remaining 45 points being omitted from the final dataset.
We combined twenty-one georeferenced native-range P. kochi observations taken from the Supplementary data of Raxworthy et al. (2007) with the aforementioned native-range P. grandis presence records to produce a P. kochi/P. grandis dataset (Appendix S3).
Nineteen Bioclim variables were downloaded for both Madagascar and Florida from the WorldClim database (http://www. world clim.org/) (Hijmans et al., 2005) at 30 arc-second resolution (~1 km 2 ) for use as predictor variables in the models. Bioclim variables were selected as they are the most commonly used environmental variables in species distribution modeling (Booth et al., 2014), thus making them the ideal data with which to test the validity of the climate-matching SDM approach. A detailed explanation of the creation and interpretation of these variables is given in O' Donnell and Ignizio (2012).
We used the Madagascar P. grandis presence records and Bioclim variables to develop a P. grandis native-range SDM (the "Madagascar model") trained on the whole of Madagascar using the Maxent algorithm (Phillips et al., 2006) via the "dismo" package (Hijmans et al., 2017) in R version 3.5.3 (R Core Team, 2013). The assumptions of the Maxent algorithm are discussed in great detail elsewhere (Elith et al., 2011;Merow et al., 2013). The model incorporated a targetgroup background (Phillips et al., 2009)  Model pre-evaluation included fivefold cross-validation, executed using the ENMevaluate function in "ENMeval," which also incorporates the Maxent algorithm (Muscarella et al., 2014). The regularization multiplier was set to 3 to reduce the risk of overfitting and smooth model output (Elith et al., 2011;Merow et al., 2013;Mutascio et al., 2018;Radosavljevic & Anderson, 2014); all other model parameters were run as default, with all Maxent feature classes allowed. The optimum Maxent feature class/class combination was determined to be that which returned the lowest average AUC DIFF (a measure of model overfitting; see Warren & Seifert, 2011) while also having an associated training AUC ≥ 0.9 (thus indicating excellent model performance; Swets, 1988). Thereafter, presence records were randomly partitioned 2:1 for use as training and validation datasets, respectively, with two thirds of the data used to build the model proper using the parameters described above, and the remaining third withheld to assess the model performance. The model proper was projected onto Florida and the Caribbean to determine which areas would be deemed bioclimatically suitable for P. grandis, and was validated using AUC and AUC DIFF (i.e., AUC TRAIN -AUC TEST ; Muscarella et al., 2014).
The process of creation, projection, and analysis of a second, combined-species "kochi/grandis model" was identical to that of the Madagascar model, except that P. grandis-only presence data were substituted with the P. kochi/P. grandis combined dataset previously described. Similarly, a third "Florida model" was preevaluated, trained, and validated using presence/background data for the whole of Florida-with a target-group background generated using 26,037 georeferenced Florida presence records for the Order Squamata from the Florida Museum of Natural History's Division of Herpetology records (09 March 2021) (Appendix S5)-instead of Madagascar, but was otherwise identical in construction. We projected the Florida model onto Florida to assess its predictive performance in the invasive range, and also projected it onto Madagascar to assess its ability to predict the native range.

| RE SULTS
Model parameters and performance statistics are summarized in Table 1. Fivefold cross-validation in "ENMeval" returned training AUC values ≥0.9193 for all three models-indicating very high predictive performance (Swets, 1988)-and low average AUC DIFF (≤0.0206) in all cases, confirming that overfitting was not occurring (Warren & Seifert, 2011). AUC and AUC DIFF results were similar for the models proper (AUC ≥ 0.939; AUC DIFF ≤ 0.014), again indicating satisfactory performance. Visual inspection of the Madagascar model projection (Figure 2) confirms that areas of predicted bioclimatic suitability closely match the known native range . Predicted suitability values for the validation data points

| D ISCUSS I ON
Our study tested the predictive accuracy of the widely advocated climate-matching species distribution modeling approach by using Maxent, Bioclim variables, and native-range presence-only data to identify areas of potential bioclimatic suitability for Phelsuma grandis in Florida, USA, and then comparing these predictions with the species' known distribution in the state. Interestingly, our model did not identify any of the already-colonized habitat as potentially suitable for P. grandis, demonstrating that climate-matching SDMs can severely underpredict the establishment risk posed by non-native herpetofauna.
It seems probable that the discrepancy between the predicted and observed distribution stems from the inherent assumptions of many SDMs, namely that 1) the observed native range of a taxon represents its fundamental bioclimatic niche and 2) adaptation to novel bioclimatic conditions will be trivial or nonexistent (Dormann, 2007;Uden et al., 2015). While these assumptions hold true to an extent for groups such as terrestrial plants (Petitpierre et al., 2012) The lack of evolutionary conservatism in the critical thermal minima (CT min ) (Brown, 1996) of lizards (Grigg & Buckley, 2013) suggests that P. grandis could potentially adapt physiologically to colder ambient temperatures, for instance if dispersing northward through Florida. Rapid in situ physiological adaptation of this nature has already been reported for several non-native lizard species in Florida (Stroud et al., 2020). Furthermore, since its initial establishment in Florida in the 1990s, P. grandis has been exposed to extreme coldweather events that have caused substantial cold-induced mortality in multiple non-native squamate species (Campbell, 2011;Fieldsend & Krysko, 2019a;Mazzotti et al., 2011Mazzotti et al., , 2016, illustrating how powerful selective forces might drive rapid population-level adaptation. Unlike most geckos, P. grandis is diurnal (Dubos, 2013), thereby allowing it greater scope for behavioral thermoregulation than nocturnal gekkotans (Brown, 1996). The species is also synanthropic Dubos, 2013;Dubos et al., 2014;Krysko et al., 2003), and so likely benefits from both the urban heat island effect (Campbell-Staton et al., 2020) and access to the warmer microhabitats associated with some anthropogenic structures (Hulbert et al., 2020;Lapwong et al., 2020;Sievert & Hutchison, 1988), which may also help to explain the low predictive power of minimum temperature.
It is thus possible that both behavioral and physiological adaptations contribute to the observed ability of P. grandis to endure brief periods of extreme cold in southern Florida (Fieldsend & Krysko, 2019a).
Given its tropical native range, it seems likely that the intensity and frequency of extreme cold events must ultimately limit the northward expansion of P. grandis (e.g., Warner et al., 2021). However, the high permutation importance of Temperature Seasonality-along with the lack of overlap in values for Temperature Seasonality between the Madagascar and Florida ranges (Figure 7a)-suggests that exposure to extended periods of suboptimal temperatures probably F I G U R E 4 "Madagascar" Maxent model showing predicted habitat suitability for Phelsuma grandis in Florida and the Caribbean. The scale bar to the right indicates the degree of predicted habitat suitability, with higher scores representing predicted higher suitability, with the range of possible values 0-1 also plays an important limiting role (e.g., Battles & Kolbe, 2019;Nania et al., 2020).
In Madagascar, the range of P. grandis is parapatric with the distribution of the closely related species P. kochi and P. madagascariensis, with little or no spatial overlap .
The projection of the Florida model onto Madagascar prima facie suggests that the distribution of P. grandis in Madagascar could be limited by the presence of P. kochi, which is acting as a competitor, and thus excluding P. grandis from occupying areas within its fundamental niche (as identified by the Florida model) in western, southwestern, and southern Madagascar (Figure 5b,c). There are no obvious geographic barriers that are otherwise preventing P. grandis from occupying these areas, and the extreme ecological flexibility Overlapping fundamental niches of recently speciated sister species is a prediction of ecological speciation on environmental gradients (ecotones), and the Phelsuma madagascariensis complex-comprising P. grandis, P. kochi, and P. madagascariensishas been considered as a strong candidate for ecological speciation . Interestingly, the Florida model-when projected onto Madagascar-provides no evidence for extensive overlap in the fundamental niches of P. grandis and P. madagascariensis (Figure 5b,c). Despite the morphological similarity of all three species, P. madagascariensis is more distantly related to P. kochi and P. grandis than they are to one another (Rocha et al., 2010) and may have evolved a fundamental niche quite distinct from that of either P. grandis or P. kochi. If true, this would further support the claim that ecological speciation has occurred within this species complex. However, a lack of climate analogues between Florida and the native range of P. madagascariensis could also lead to a similar prediction; in this case, P. grandis would by definition be unable to establish in such areas in Florida, and model output would consequently be biased against them.
Our results provide some evidence that the colonization of Florida by P. grandis may have been facilitated by ecological release (Kohn, 1972), in this case, from interspecific competition with P.
kochi. However, given that the projection of the Florida model onto Madagascar (Figure 5b) identifies more suitable native-range habitat for P. grandis than is identified even by the combined-species kochi/grandis model (Figure 6a), we suggest that some degree of in situ adaptation has almost certainly occurred during this colonization event, as a result of which the P. grandis population of Florida has expanded its occupied niche. Since the kochi/grandis model did not predict the observed successful colonization of Florida by P.
grandis with high accuracy (Figure 6b), the degree to which competition with P. kochi restricts the distribution of P. grandis to northern Florida, such as P. laticauda (Fieldsend et al., 2020).
In summary, our study adds weight to the argument that climatematching SDMs generated from native-range distributional data may not alone be appropriate tools for predicting the establishment risk of non-native herpetofauna (Li et al., 2014). In particular, our results highlight an example of an invasive species whose occupied nativerange niche is much smaller than its non-native-range niche, due to in situ adaptation in the non-native range, and potentially also competition with a closely related species within the native range.

ACK N OWLED G M ENTS
The authors wish to extend their thanks to FIU Institute of

CO N FLI C T O F I NTE R E S T
The authors state that there is no conflict of interest.

DATA AVA I L A B I L I T Y S TAT E M E N T
A ZIP file containing (a) the datafiles used in this study, (b) the R script used to generate, project, and analyze the models presented in this paper, and (c) the output files associated with the models is available via Dryad, DOI https://doi.org/10.5061/dryad.m905q fv1c.