A Machine Learning approach to study plant functional trait divergence

Abstract

Premise of the study: Plant functional traits are often used to describe spectra of ecological strategies among species. Here we demonstrate a machine learning approach for identifying the traits that contribute most to interspecific phenotypic divergence in multivariate trait space. Methods: Descriptive and predictive machine learning approaches were applied to trait data for the genus Helianthus, including Random Forest and Gradient Boosting Machine classifiers, Recursive Feature Elimination, and the Boruta algorithm. These approaches were applied at the genus level as well as within each of the three major clades within the genus to examine the variability in major axes of trait divergence in three independent species radiations. Key Results: Machine learning models were able to predict species identity from functional traits with high accuracy, and differences in functional trait importance were observed between the genus level and clade levels indicating different axes of phenotypic divergence. Conclusions: Applying machine-learning approaches to identify divergent traits can provide insights into the predictability or repeatability of evolution through comparison of parallel diversification of clades within a genus. These approaches can be implemented in a range of contexts across basic and applied plant science from interspecific divergence to intraspecific variation across time, space, and environmental conditions.