Machine learning is a powerful tool for linking phenotype and genotype in Saccharomycotina yeasts

Abstract

The recently characterized genomes, isolation environments, and presence/absence of growth on 122 substrates and conditions from 1,154 (nearly all known) yeast species in the subphylum Saccharomycotina provide a powerful but complex dataset for studying the evolution of the genotype-phenotype map. Using a random forest algorithm on this dataset elucidated an alternate galactose-degrading pathway in yeasts, as well as linked relevant traits and genes to the ecological niches of cactophily, the floral-bee niche, and generalism/specialism. For the final part of my thesis, I used a random forest trained on these datasets as well as relevant gene sequences to predict antifungal resistance in yeasts and found that the genes and variants associated with resistance were different than those found previously clinical samples. I conclude that machine learning is a powerful tool for investigating the macroevolution of the genotype-phenotype map in fungi.