Developing Computational Methods for Analyzing Metabolomics Data and Applications to Study Obesity

Abstract

Metabolomics is a powerful approach for discovering biomarkers and for characterizing the biochemical consequences of genetic variation. This dissertation consists of several projects that focused on developing computational approaches to analyze both targeted and untargeted metabolite profiling data and applying them to study obesity. First, we developed PAIRUP-MS, a suite of methods for analyzing unknown, unidentified metabolite signals across multiple mass spectrometry-based untargeted profiling datasets. PAIRUP-MS contains an imputation-based approach for matching unknown signals across datasets, allowing for meta-analysis of matched signals across studies that would otherwise be incompatible. It also offers a pathway annotation and enrichment analysis framework that links signals to plausible biological functions without needing to confirm their chemical identities. We validated PAIRUP-MS using genetic data and showed that it enables the discovery of many more biologically relevant signals and pathways compared to the standard practice of only analyzing known metabolites. Next, we combined PAIRUP-MS with Mendelian randomization approaches to dissect the causal relationships between body mass index (BMI) and BMI-associated signals in untargeted metabolomics data. We identified known and unknown signals that are likely to be the cause or the effect of obesity and performed pathway analyses to compare these signals. We found causal signals to be enriched in the “glutathione-mediated detoxification” pathway, whereas effect signals were enriched in many more pathways including “lysine catabolism”, “dopamine metabolism”, and “signaling by GPCR”. Lastly, we constructed and validated a metabolite risk score (MRS) for predicting future weight gain using targeted profiling data. We showed that none of the available anthropometric, lifestyle, and glycemic risk factors fully account for the MRS prediction of weight gain, even though the score was strongly correlated with baseline insulin sensitivity. We also identified nine genetic loci associated (p < 1e-6) with the MRS, including two loci (FADS1/2 and PNPLA3) previously linked to obesity phenotypes. Therefore, the MRS captures a composite biological picture related to weight gain. Overall, the computational approaches described in this dissertation enable analyses of both known and unknown metabolite signals in robust, biologically meaningful manners and provide a path towards more comprehensive, well-powered studies of metabolomics data.