Defense titled "Bayesian factorization framework for high-dimensional metabolomic–genomic association analysis".

Metabolomics is the high-throughput study of small molecule metabolites in one's tissues or bodily fluids and has shown remarkable potential to explain variation in human phenotypes and elucidate the etiology of disease. In the last decades, researchers have combined metabolomics and genome-wide association studies to explore the impact of genetics on metabolite levels, as it can help predict novel phenotypes, understand pleiotropy, and infer the causal effects of metabolites on disease phenotypes. However, the tremendous complexity of these data requires new methods that leverage our prior knowledge of the relationships between metabolites. To address the challenges in metabolomics studies and metabolite genome-wide association studies (mtGWAS), I propose to develop a comprehensive framework to analyze high-throughput metabolomics data and mtGWAS data. My framework consists of three main parts: a pathway-guided prior based on the Hierarchical Dirichlet Process for metabolite-specific loadings, a Bayesian factor model to estimate and perform inference on indirect and direct genetic effects, a scalable multi-trait fine-mapping algorithm, a metabolite-annotation method based on the pathway-guided prior. We run thorough and sufficient simulations to justify the accuracy and power of our algorithms in estimating the number of latent factors in mtGWAS data, inferring genetic effects, and drawing biological conclusions at the pathway level. To demonstrate the power of our methods, we applied our framework to real-world datasets and justified the biological conclusions obtained using our methods through extensive literature reviews.