Computational analysis of complex genetic interactions

Cold Spring Harbor Laboratory, Cold Spg Hbr NY

Investigators

Linked publications, trials & patents

Abstract

Project Summary / Abstract How does the DNA sequence of an organism (genotype) determine its form and function (phenotype)? New technologies such as massively parallel reporter assays (MPRAs), deep mutational scanning, and combinatorial CRISPR screens have the potential to expose the genotype-phenotype relationship at an unprecedented level of detail by measuring phenotypes for tens of thousands to millions of genotypes in a single experiment. However, interpreting the results of these experiments is difficult because the space of genotypes is intrinsically high-dimensional and combinations of mutations often interact in complicated ways. My research program is focused on developing new computational tools to analyze data from these high-throughput experiments, with the goals of (1) identifying the major qualitative features of the genotype-phenotype relationship in specific biological systems, (2) explaining how these qualitative features arise from underlying developmental, cell biological and biophysical mechanisms, (3) being able to accurately predict the phenotypes of unmeasured genotypes, and (4) quantifying the uncertainty in these predictions. During the previous funding period, our group developed a suite of new computational techniques for inferring the structure of genotype-phenotype relationships, where these techniques are based around a class of highly flexible statistical models known as Gaussian processes that we have customized for use in analyzing high-throughput combinatorial mutagenesis experiments. While these methods have been very successful at analyzing complex genetic interactions, they are still currently limited to the analysis of relatively short sequences. Here we propose to scale these methods up to work on whole proteins or even genome-wide by leveraging recent breakthroughs in the ability of GPUs to accelerate Gaussian process based inference. At the same time, we will develop two new Gaussian process based techniques, one that extends our existing methods to address diploid genotypes, and one that can better infer and then exploit knowledge of which sites and mutations in a functional genetic element are most important. Expressive computational modeling and high-throughput measurements have the potential to transform molecular biology by allowing accurate prediction of the effects of mutations, both singly and in combination. Important applications include mapping adaptive paths to immune escape and drug resistance variants in infectious disease and cancer, designing improved antibodies and enzymes, and genomic variant interpretation. Development of the computational tools proposed here will further these goals by providing a principled, practical framework for understanding the complex genetic interactions revealed in contemporary high-throughput experiments.