Using Large Scale Genomics Data To Understand The Genetic Basis Of Complex Traits

With the arrival of big data in genetics in the past decade, the field has experienced drastic changes. One game-changing breakthrough in genetics was the invention of genotyping and sequencing technology that allows researchers to examining single nucleotide polymorphisms (SNPs) across the entire genome. The other major breakthrough was the identification of haplotypes of common alleles in major human populations, which permitted the design of genotyping assays that effectively cover entire human genomes at a resolution appropriate for genetic mapping. Together, these technology breakthroughs have permitted researchers to carry out Genome Wide Association Studies (GWAS) on a wide range of traits including, for example, height and disease status. With GWAS, causal SNPs have been identified for some Mendelian traits, but for more complex genetic traits, the genetic heritability explained by the associated SNPs are low. In addition, high-throughput technologies to generate other types of -omics data such as gene expression, DNA methylation, and protein levels data have also emerged recently. How to best utilize the SNP data and other multi-omics data to understand genetic traits is one of the most important questions in the field today. With the increasing prevalence of multi-omics data, new types of analysis schemes and tools are needed to handle the additional complexity of the data. In particular, two areas of method development are in great need. First, statistical methods employed by GWAS do not consider the potential interacting relationships among genetic loci. Thus, methods that can explore the joint effect between multiple genetic loci or genetic factors could unveil new associations. Second, different types of omics data may give distinctive representations of the overall biological system. By combining multi-omics data, we could potentially aggregate non-overlapping information from each individual data types. Thus, the focus of this dissertation is on developing and improving computational methods that can jointly model multiple types of genomics data. First, an evaluation of an existing method, grammatical evolution neural network, was conducted to identify the optimal algorithm settings for the detection of genetic associations. It was found that under certain algorithm settings, the neural networks have been restricted to one-layer simple network. Using a parameter sweep approach, the analysis identified optimal settings that allow for building more flexible network structures. Then, the algorithm was applied to integrate multi-omics data to model drug-induced cytotoxicity for a number of cancer drugs. By combining different types of omics data including SNPs, gene expression and methylation levels, we were able to model a higher portion of the observed variability than any individual data type alone. However, one drawback of the existing neural network approach is the limited interpretability. To this end, a new algorithm based on Bayesian Networks was created. One novelty of the approach is the ability to independently fit a distinct Bayesian Network for each categories of a phenotype. This allows for identifying category specific interactions as well as common interactions across different categories. Analysis using simulated SNP data has shown that the Bayesian Network approach outperformed the Neural Network approach in many settings, particularly in situation where the data contains multiple interacting loci. When applied to a type 2 diabetes dataset, the algorithm was able to identify distinctive interaction patterns between cases and controls. Ultimately, the goal of this dissertation has been to fully take advantage of the newly available data to understand the genetic basis of complex traits.