Information-sharing models for computational genetics

Abstract

Modern genetics has been transformed by a dramatic explosion of data. As sample sizes and the number of measured data types grow, the need for computational methods tailored to deal with these noisy and complex datasets increases. In this thesis, we develop and apply integrated computational and biological approaches for two genetic problems. First, we build a statistical model for genetic mapping using pooled sequencing, a powerful and efficient technique for rapidly unraveling the genetic basis of complex traits. Our approach explicitly models the pooling process and genetic parameters underlying the noisy observed data, and we use it to calculate accurate intervals that contain the targeted regions of interest. We show that our model outperforms simpler alternatives that do not use all available marker data in a principled way. We apply this model to study several phenotypes in yeast, including the genetic basis of the surprising phenomenon of strain-specific essential genes. We demonstrate the complex genetic basis of many of these strain-specific viability phenotypes and uncover the influence of an inherited virus in modifying their effects. Second, we design a statistical model that uses additional functional information describing large sets of genetic variants in order to predict which variants are likely to cause phenotypic changes. Our technique is able to learn complicated relationships between candidate features and can accommodate the additional noise introduced by training on groups of candidate variants, instead of single labeled variants. We apply this model to a large genetic mapping study in yeast by collecting multiple genome-wide functional measurements. By using our model, we demonstrate the importance of several molecular phenotypes in predicting genetic impact. The common themes in this thesis are the development of computational models that accurately reflect the underlying biological processes and the integration of carefully controlled biological experiments to test and utilize our new models.