UCLA

Phylogenetic comparative methods seek to untangle the complex web of selective pressures driving biological evolution. These methods seek to identify associations between different biological traits over evolutionary history. Statistical models of phenotypic evolution need to account for the shared evolutionary history between different species, and accounting for this non-independence poses computational challenges. These challenges are compounded by missing observations, high-dimensional traits and highly-structured data. Here, I develop computational and modeling approaches that dramatically improve the computational efficiency and scalability of these models to enable Bayesian phylogenetic comparative analysis of unprecedentedly large data sets. First, I develop an algorithm that analytically marginalizes missing observations in a (relatively) simple model of phenotypic evolution. This algorithm is broadly applicable beyond this simple model and allows scalable inference under a variety of model extensions. These extensions include models that accommodate residual variance, allowing measurement of phylogenetic heritability, and linear dimension reduction, allowing phylogenetic comparative analyses for high-dimensional traits. I combine this work into a generalizable modeling framework that allows researchers to build flexible, highly structured models that remain scalable for both large number of taxa and many observations per taxon. This work achieves increases in computation speed by more than two orders of magnitude across several contexts, bringing computation time down from weeks or months to minutes or hours in multiple real-world applications.