Phylogenetic transferability of ecological reconstruction

Phylogenetic transferability of ecological reconstruction Because closely related species may share similar trait values owing to common evolutionary history, random partitioning of species between training and test sets can lead to overly optimistic estimates of predictive performance [1]. Phylogenetically structured validation, in which species within a defined phylogenetic distance from the test species are excluded from model training, provides a means of evaluating model transferability to increasingly phylogenetically distinct taxa [2]. Phylogenetic-distance exclusion validation was additionally performed to determine whether ecological information contained in chlorophyll a fluorescence signatures could be transferred beyond closely related species. Pairwise phylogenetic distances among species were calculated as cophenetic distances from the reconstructed phylogenetic tree [3]. Each species was treated independently as a test species, while the remaining species available for model training were progressively restricted according to their phylogenetic distance from the focal species. Three levels of phylogenetic exclusion from the training set were considered, representing progressively increasing minimum phylogenetic separation between each focal test species and the species retained for model training. This design follows the general principle of phylogenetically buffered leave-one-out cross-validation, in which species occurring within a predefined phylogenetic-distance buffer around the focal species are excluded from model training [2]. Under the None level, only the focal species was excluded from training, providing a leave-one-species-out baseline. The Moderate level excluded all species separated from the focal species by a phylogenetic distance less than or equal to the 25th percentile of the distribution of pairwise distances. The Strong level was defined as the largest observed phylogenetic-distance threshold for which every focal species retained at least 100 potential training species. In the present dataset, these criteria corresponded to distance thresholds of approximately 237 and 247 for the Moderate and Strong levels, respectively. The resulting median numbers of available training species were 262, 193, and 149 for the None, Moderate, and Strong levels, respectively. At each exclusion level, Ridge regression, Elastic Net, partial least squares (PLS) regression, and Random Forest models were fitted separately for each ecological characteristic using the 17 OJIP parameters as predictors. The same model specifications and tuning procedures used in the repeated 75/25 training–test validation were applied at each level of phylogenetic exclusion. Predictions were generated independently for each focal species using only species satisfying the corresponding phylogenetic-distance criterion. Predictions obtained across all focal species were then pooled within each combination of ecological characteristic, model, and exclusion level. Predictive performance was quantified using the out-of-sample coefficient of determination (predictive R2), together with RMSE, MAE, and the correlation between observed and predicted values [4,5]. Phylogenetic transferability was evaluated by comparing predictive R2 across progressively increasing phylogenetic-exclusion distances between focal test species and species retained for model training, following the general framework of phylogenetically buffered cross-validation [2]. References 1. Schweizer, D.; Machado, R.; Durigan, G.; Brancalion, P.H.S. Phylogenetic Patterns of Atlantic Forest Restoration Communities Are Mainly Driven by Stochastic, Dispersal Related Factors. For. Ecol. Manage. 2015, 354, 300–308, doi:10.1016/j.foreco.2015.05.026. 2. Roberts, D.R.; Bahn, V.; Ciuti, S.; Boyce, M.S.; Elith, J.; Guillera‐Arroita, G.; Hauenstein, S.; Lahoz‐Monfort, J.J.; Schröder, B.; Thuiller, W.; et al. Cross‐validation Strategies for Data with Temporal, Spatial, Hierarchical, or Phylogenetic Structure. Ecography (Cop.). 2017, 40, 913–929, doi:10.1111/ecog.02881. 3. Cardona, G.; Mir, A.; Rosselló, F.; Rotger, L.; Sánchez, D. Cophenetic Metrics for Phylogenetic Trees, after Sokal and Rohlf. BMC Bioinformatics 2013, 14, 3, doi:10.1186/1471-2105-14-3. 4. Willmott, C.; Matsuura, K. Advantages of the Mean Absolute Error (MAE) over the Root Mean Square Error (RMSE) in Assessing Average Model Performance. Clim. Res. 2005, 30, 79–82, doi:10.3354/cr030079. 5. Chicco, D.; Warrens, M.J.; Jurman, G. The Coefficient of Determination R-Squared Is More Informative than SMAPE, MAE, MAPE, MSE and RMSE in Regression Analysis Evaluation. PeerJ Comput. Sci. 2021, 7, e623, doi:10.7717/peerj-cs.623.