Inferring sparse latent structure from genotype-phenotype maps
Authors: Samantha Petti; Gautam Reddy; Michael Desai
Affiliation: Harvard University, Cambridge, MA
Keywords: Theory & Method Development
Inferring the structure of a genotype-phenotype map is a long-standing problem in quantitative genetics. Covariation across phenotypes and covariation of genetic effects on the phenotypes contain useful statistical information about phenotype-determining pathways that share the same genes. Structure discovery therefore benefits from measuring the effects of many genetic perturbations on a large number of phenotypes. We develop a conceptual framework and an accompanying analysis pipeline for joint QTL mapping and structure discovery. First, we use a penalized regression framework to jointly map causal loci and their effects across phenotypes. Second, we develop statistical methods to test for the presence of sparse, lower-dimensional “core phenotypes” that explain covariation patterns. Finally, we show that a penalized matrix decomposition framework can be used to identify this sparse structure. We apply our methods on a variety of fitness-based datasets, including genotype-fitness measurements of 100,000 budding yeast offspring from an F1 cross, fitness effects in diverse environments of adaptive mutations in yeast, genotoxic fitness screens from human cell lines and a large-scale yeast chemogenomics assay. We find that individual genes often affect only a sparse subset of core phenotypes, which however can influence fitness in diverse contexts. The extent of pleiotropy varies across genes and depends in general on the nature of the genetic perturbation. Covarying patterns of pleiotropy allow for clustering genes into putative pathways, which we compare to existing annotations.