New records and modelling the impacts of climate change on the black-tailed marmosets

In a similar manner, the ten edaphic and topographic variables were summarized on PCA axes, using a Broken-Stick model to select the number of axes to be included in each model. According to the premises of Broken-Stick modeling, the PCA axes selected, were the two that explained more of the variation than the null model produced by the analysis. Since there are no temporal projections of edaphic and topographic variables, the two more informative com- ponents selected by the broken-stick criterion were used in all climate scenarios, since we do not expect geological processes to act in such a short time. On the other hand, such variables are important in modeling [36], and we chose to include them, but used the same edaphic and topographic PCAs axes created with the present-day variables to build the projections of envi- ronmental suitability for all present and future changes scenarios used in this study [37]. Modeling. For modeling the following algorithms were used using the R biomod2 pack- age, following [38]. We used nine algorithms: Generalized Boosted Models (GBM), Classifica- tory Tree Analysis (CTA), Random Forest (RF), Generalized Linear Model (GLM), Generalized Additive Model (GAM), Artificial Neural Networks (ANN), Flexible Discrimi- nant Analysis (FDA), Multiple Regression Adapted for Splines (MARS) and Maximum Entropy (MAXENT). During modeling, M. nigriceps occurrence records were tallied using a gradient system based on cells 10km X 10km per side. We did not obtain a minimum of 25 spatially unique

occurrence records, that is, a single record in a cell of 100km2, which would add at least 25 cells to the grid with an occurrence record in each. Due to the small amount of data available, we used the Jackknife approach for training and evaluating model performances [39, 40]. In this approach, an occurrence record is removed from the modeling and the capacity of the model to predict the existence of the removed locality is tested. We repeated the procedure of removing an occurrence record until all points had been excluded from the modeling at least once, and then evaluated performance with a binomial test, which showed whether our model was better than one produced at random [40]. Most modeling algorithms require data with presence and absence records for the modeled species. However, confirmed absence data is rare for most species, and in these cases we used pseudo-absences (PAs). PAs are random points selected within the study area background that serve to represent absences when calibrating and evaluating models. We calibrated each algo- rithm using the biomod2 package default configuration, as this modifies only pseudo-absence numbers (PAs). In such cases, the PAs in GBM, RF and CTA were ranked based on the number of occurrences, as recommended by [41]. For GLM, GAM, ANN, FDA, MARS and MAXENT models, 10000 PAs points were randomly selected, following [42]. In addition, PAs were gener- ated using the two sets of algorithms via a disk arrangement, which defined the minimum and

maximum distances at all points of occurrence available for the species, with PAs sampled within these minimum and maximum limits, that is, inside the “disk” [38]. A set of 10 PAs was run in 10 modeling sequences for each set of PAs, equivalent to 100 models for each algorithm [39]. After removing an occurrence record, producing the models and obtaining the maps with the continuous projections of each algorithm, we joined the projections of all algorithms in a single continuous averaged ensemble map. We performed this procedure every time we removed an occurrence record. After obtaining the continuous modeling maps, we built binary representation maps, where 0 indicates absence and 1 presence of the species. To produce such binary maps, we used the threshold that maximizes the sum of sensitivity and specificity, com- monly called MaxSpecSens or ROC Threshold [42], so that all cells in the study area that had suitability values above the threshold were considered as presence (value 1), while all study area values with suitability less than the threshold were marked as species absent (value 0). Finally, we combined all the binary maps that matched the points left out of the modeling into a final consensus map that represented the most suitable conditions for M. nigriceps. We PLOS ONE Modeling of Mico nigriceps PLOS ONE | https://doi.org/10.1371/journal.pone.0256270 September 7, 2021 6 / 14