Phylogeny-informed machine learning for comparative genomics

The diversity of life on Earth is the result of a multi-billion year natural experiment in evolution, which has generated the genomes of millions of species, including ours. This extant biological and genetic diversity provides us with a catalog of phenotype-genotype combinations that differ from humans to varying degrees but are all possible within the constraints of biology, chemistry, and physics. It can therefore inform us about our own genome and its influences on our traits, physiology and health. Comparative genomics can allow us to uncover these connections by linking shared evolutionarily conserved regions to shared traits. When traits have evolved repeatedly in independent lineages, it is often possible to narrow the search to a smaller set of shared phenotypes and similar genetic elements. However, computational methods designed to identify shared molecular convergence have faced considerable difficulty due to the pervasive background of apparent convergence that originates from neutral evolution and shared ancestry. This dissertation introduces the Evolutionary Sparse Learning with Paired Species Contrast (ESL-PSC) method, which builds predictive genetic models of a trait of interest using the genetic data from phylogenetically close pairs of phenotype-contrasting species as input features. In this way, ESL-PSC models treat shared inherited variation as equally represented in both classes, and thus it is canceled out, leaving only variation shared within a class, for example, due to convergence. Following a review of the literature on computational methods for identifying molecular convergence, I demonstrate the effectiveness of the ESL-PSC method on empirical and simulated datasets and develop a new graphical software environment to make the method broadly accessible to users. Finally, because phylogenetic topology is key to such comparative methods for studying convergence, I developed Treemble, a graphical software tool to enable the extraction of machine-readable phylogenetic tree data from published tree diagrams. Together, this work advances both the methodology and infrastructure needed to study convergent evolution through comparative genomics at increasing scale across the tree of life.