Associations and Confounding in High-Dimensional Genomics

Abstract

Understanding the genetic basis of complex traits has the promise of curing many human diseases. We are now closer to fulfilling this goal by pairing our understanding of the effects of genetic mutations with the promise of gene editing techniques such as CRISPR [132, 81]. This promise has been advanced by genetic association studies, which yield insight into the genetic architecture of human disease. Interpreting the functional significance of genetic association results is dependent upon developing robust and statistically powerful methods that integrate sources of data from heterogeneous and sample-limited experiments. This thesis leverages dimensionality reduction techniques to detect statistical associations in high-dimensional and noisy datasets that result from biological experiments.

The following chapters demonstrate how developments in probabilistic methods and numerical techniques can translate to advancements in genetic association studies. A primary theme in this research is developing computational and statistical methodologies to perform associations of genetic data to high-dimensional study data (e.g., disease state, level of mRNA expression, and cell morphology) while controlling for confounding factors.

In order to fulfill the promise of curing human disease, it is imperative to understand how mutations in the human genome affect the biological hierarchy and ultimately lead to disease. Estimating the effect of genetic mutations requires taking into account the uncertainty of biological assays and the dynamic nature of biological systems. The methods I have developed employ linear and non-linear dimensionality reduction to provide robust and accurate estimates of genetic associations. This thesis facilitates insight into the link between genetic mutations and disease by integrating heterogeneous biological data, controlling for confounding, and maintaining interpretability.