Predicting genetic biodiversity in salamanders using geographic, climatic, and life history traits

affect levels of biodiversity by influencing species biology, anthropogenic factors can influence the way we find and describe this diversity (e.g., increased sampling in wealthier, more popu- lated locations). We extracted species specific data from several GIS layers, including anthro- pogenic biome [65], human population density [66], and gross domestic product [67] in order to evaluate how anthropogenic factors impact undescribed diversity. We utilized the R packages ’raster’ [68], ’rgdal’ [69], ’geosphere’ [70], and ’plyr’ [71] to extract species specific information from each layer using geographic occurrence records obtained from phylogatR. To represent the environmental variation within the occupied range of each species, we extracted the value of each environmental layer for each GPS coordinate associated with each species. We then took the mean and standard deviation for each environmental variable. To obtain species specific data related to geographic distribution we extracted the minimum, maximum, mean, and length of latitude and longitude from the GPS points of each species. We used the R package ‘mice’ [72] to impute trait values missing from our dataset (see S1 Fig in S1 File for distribution of missing data and specific trait values imputed). The imputa- tion method ‘pmm’ was used for all numeric variables and ‘polyreg’ was used for categorical variables (i.e., reproductive strategy and habitat). We ran the imputation 15 times (S2 Fig in S1 File) and then pooled the iterations to generate the final imputed values. The final database

containing all trait values (both imputed and original) is available in S4 Data. Predictive modeling We used the R package ’caret’ [73] to generate a random forest classification model [34] based on our previously generated database of predictor variables and a consensus of our species delimitation results. Two separate sets of consensus models were generated to assess the role of geography, environment, and life history traits on the presence of hidden diversity (Fig 1A). The first model (all agree) represents a strict consensus of delimitation results from species in which results from all methods of species delimitation agree (Fig 1B). Any species with con- flicting delimitation results were excluded from analysis. The second model (majority rules) represents a majority rule consensus in which species are assigned to a response category based on relative support of delimitation results (Fig 1C). For each model, we used 70% of the data to train the model and the remaining 30% was set aside as a test set. Models were gener- ated using 10-fold cross validation with five repeats to tune the parameter ’mtry’, the number of variables randomly sampled at each split, and optimize the area under the receiver operating characteristic curve, ROC. After training, we extracted the variable importance measures mean decrease accuracy (MDA) and Gini impurity (Gini) from the final models. We then used the final models on the test set data to evaluate model performance. Model performance was evaluated across a variety of metrics including model accuracy, which reflects how well the

predicted classifications agree with the observed classifications, and both positive and negative predictive value, which indicate the how the model performs on observations from each class. Additionally, we calculated the no information rate (NIR), the proportion of observations that fall into the majority class, and the p-value [Accuracy>NIR], to test for model significance. The top important predictor variables from our best model were compared using a Kruskal- Wallis test to determine if these variables are significantly different between species that do or do not contain hidden diversity. Results Genetic and geographic dataset Our final dataset consisted of 1676 DNA barcoding sequences (Fig 2). Of these, 768 sequences were from the Cytochrome oxidase I gene (COI), and 908 sequences were from the PLOS ONE Predicting hidden lineages in salamanders PLOS ONE | https://doi.org/10.1371/journal.pone.0310932 October 18, 2024 5 / 20