Discriminating models of trait evolution

Abstract

AO_SCPLOWBSTRACTC_SCPLOWA central challenge in comparative biology is linking present-day trait variation across species with unobserved evolutionary processes that occurred in the past. In this endeavor, phylogenetic comparative methods are invaluable for fitting, comparing, and selecting evolutionary models of varying complexity and biological meaning. Traditionally, evolutionary studies have relied on conventional statistical approaches to assess model fit and identify the one that best explains variation in a given trait. Here we explore an alternative strategy by applying supervised learning to predict evolutionary models via discriminant analysis. We formally introduce Evolutionary Discriminant Analysis (EvoDA) as an addition to the biologists toolkit, offering a suite of new methods for studying trait evolution. We evaluate the performance of EvoDA alongside conventional model selection through a series of fungal phylogeny case studies, each targeting increasingly challenging analytical tasks. These results showcase the strengths of EvoDA, with substantial improvements over conventional approaches when studying traits subject to measurement error, which likely reflect realistic conditions in empirical datasets. To complement our simulation-based benchmarking, we explore the application of EvoDA for tackling a notoriously difficult task: predicting the mode and tempo of gene expression evolution. This empirical analysis suggests that stabilizing selection acts on a majority of genes, with bursts of expression evolution in a handful of genes related to stress, cellular transportation, and transcription regulation. Collectively, our findings illustrate the promise of EvoDA for predicting trait models across a range of evolutionary and experimental contexts, establishing a new methodological framework for the next era of comparative research. SO_SCPLOWIGNIFICANCEC_SCPLOW SO_SCPLOWTATEMENTC_SCPLOWTo make sense of biodiversity, evolutionary studies have historically relied on conventional statistical procedures to evaluate competing hypotheses about the mode and tempo of trait evolution. Here, we introduce new supervised learning methods that substantially outperform traditional techniques for correctly assigning trait models across a range of evolutionary and experimental conditions. We find that these methods are highly robust to measurement noise expected from realistic trait data and offer new insights into a central question in comparative genomics: what are the evolutionary forces shaping variation in gene expression across species?