Supraspecific units in correlative niche modeling improves the prediction of geographic potential of biological invasions

Background Biological invasions rank among the most significant threats to biodiversity and ecosystems. Correlative ecological niche modeling is among the most frequently used tools with which to estimate potential distributions of invasive species. However, when areas accessible to the species across its native distribution do not represent the full spectrum of environmental conditions that the species can tolerate, correlative studies often underestimate fundamental niches. Methods Here, we explore the utility of supraspecific modeling units to improve the predictive ability of models focused on biological invasions. Taking into account phylogenetic relationships in correlative ecological niche models, we studied the invasion patterns of three species (Aedes aegypti, Pterois volitans and Oreochromis mossambicus). Results Use of supraspecific modeling units improved the predictive ability of correlative niche models in anticipating potential distributions of three invasive species. We demonstrated that integrating data on closely related species allowed a more complete characterization of fundamental niches. This approach could be used to model species with invasive potential but that have not yet invaded new regions.


Maxent methods
Maxent models were calibrated across the M area designed for each modeling unit.
We partitioned presence records into two sets, training and evaluation, using the block function of ENMeval (Muscarella et al., 2014). Settings for the model construction were: the crossvalidation/replicated functionality, three features (l=linear, q=quadratic and p=product), regularization multipliers from 1 up to 4, and a "logistic" output. We chose the previous features trying to emulate mathematical functions fitted in MVEs, in order to minimize differences between algorithms given by factors other than the presence data. We also set aside 50% of presences as a test percentage and conducted five replicates analysis to take into account the variances due to specific calibration data sets on model outputs. Finally, the models were projected worldwide applying truncation as transfer procedure.

Maxent results
In all three species, the AUC ratio of the partial ROC test increased when occurrences of closely related species were included (Fig. S3). This increase was similar to that observed in the MVEs, where it was clearer in Ae. aegypti and P.
volitans whereas in O. mossambicus the greatest increase in performance was reached up to U5.
For Ae. aegypti, thresholded models of U0 failed to predict the invasion in different regions that are predicted by U1, such as in the Florida Peninsula, central Mexico, the Atlantic Forest and eastern Australia ( Fig. S4B and Fig. S5A and C).
In P. volitans, unlike U0, the U1 model predicts invasion in the Red Sea, the Adriatic Seas and in the Persian Gulf ( Fig. S4D and Fig. S5D and E).    Potential distribution models obtained with Maxent using the black X's as input presence data (light blue) (i.e., U0), and potential distribution models obtained with Maxent using the black X's + red X's as input presence data (dark blue) (i.e., U1) for Ae. aegypti (B), P. volitans (D) and O. mossambicus (F).