The persistent advantage of model‐based phylogenetic methods for single‐trait prediction

Abstract Reliable predictions of biological traits support a wide range of applications, from bioprospecting to informing conservation priorities. Given the complexity and diversity of trait evolution, robust methods for trait prediction are essential for drawing meaningful evolutionary inferences from phylogenetic data. This study benchmarks traditional phylogenetic comparative methods against emerging machine learning (ML) approaches for predicting traits of taxa in phylogenetic trees. We simulated extensive datasets under diverse evolutionary scenarios, including models of discrete and continuous trait evolution, alongside real trait data for a large order of flowering plants. Traditional prediction tools included corHMM (binary traits) and phylopars (continuous traits), while ML approaches leveraged dimensionality reduction (phylogenetic eigenvectors, UMAP, autoencoders) combined with logistic and linear regression models, XGBoost and a novel phylogenetic k‐nearest neighbours method (phylokNN). Results show that methods based on explicit evolutionary models—specifically corHMM for binary traits and phylopars for continuous traits—consistently outperformed ML methods across simulations and real data. Although ML approaches achieved competitive accuracy, they rarely surpassed conventional models. The study concludes that phylogenetic comparative methods that incorporate explicit evolutionary models retain significant power in predicting trait values when a single trait is analysed, even as trait evolutionary processes violate model assumptions.