Climate change in Europe between 90 and 50 kyr BP and Neanderthal territorial habitability

In order to reduce the bias of model estimators and to validate the model by verifying the stability of its predictions, we implemented a bootstrap with 1000 replicates, each replicated sample obtained by sampling with replacement from the presence points, with the number of samples equaling the total number of presence points. Bootstrapping also allows us to compute the average Area Under the ROC (Receiver Operating Characteristic) Curve (AUC) and its standard deviation across models. AUC, applicable to any species modelling method, measures the quality of ranking of sites [27]. In the case of presence-only data, it can be interpreted as the probability that a randomly chosen presence site is ranked above a random background site [24]. Used as an indicator of good predictivity, models with values above 0.85 are consid- ered potentially useful [28]. To evaluate the contribution of each environmental variable to the model, we used the jack- knife test performed by Maxent. It gives the importance of each environmental variable com- puting the regularized gain (measure related to deviance) by training with each environmental variable first omitted, then used in isolation. Spatial comparison In order to compare the probability maps obtained and to study the evolution in suitability for Neanderthal occupation of areas in Europe during the 5 periods, we used a Structural Similar- ity (SSIM) index [29]. This index compares continuous-valued spatial data and allows us to simultaneously consider the local magnitude and spatial structure. SSIM is bounded by (-1, 1),

-1 indicating complete dissimilarity between the spatial structure of the compared maps, and 1 indicating that maps are identical. Due to the range and meaning of SSIM, we chose the arbi- trary threshold of 0.2. Below this value, we consider the maps of the periods being compared to be dissimilar. This analysis was conducted using R-4.2.1 software [30]. Results ROC and AUC To assess the accuracy of our models we perform a ROC analysis (Curves and AUC Values) (S File) The ROC curves are similar for the five periods and AUC values are always bigger than 0.92 showing a high level of confidence in the model predictions (see S4 File). For each period, data concerning temperature is always the environmental variable which contributes the most to the Maxent model (see S5 File). In each period the temperature is the most effective variable for predicting the suitability of each site for Neanderthal occupation: temperature appears to have the most useful information by itself and it is also the variable that decreases the gain the most when it is omitted. Precipitation is the other important vari- able while friction and carrying capacity bring little additional information. The environmen- tal variables chosen are not orthogonal: temperature and precipitation are themselves strong controls on the biome reconstruction and thus the friction and carrying capacity terms we use [18]. Maxent maps Figs 2–6 show the results of the 1000 bootstraps carried out for each period on the output files concerning the estimation of probability of environmental suitability for Neanderthals. There

are clearly many similarities among the maps representing the five successive periods extend- ing from 90 to 50 kyr BP. The five heat maps identify regions over time that had a very low (less than 0.2) or even zero probability of providing suitable ecological niches for Neanderthals. These areas are the PLOS ONE Neanderthal habitability PLOS ONE | https://doi.org/10.1371/journal.pone.0308690 February 26, 2025 5 / 17