Integrating statistical and mechanistic modeling to analyze disease omic data

The advent of high throughput technologies has enabled large-scale measurements of the genome, transcriptome, proteome and metabolome of tissues samples, serum and even single cells. Additionally, prior biological knowledge is increasingly curated into accessible databases and reconstructed into computable models. My research aims to integrate high throughput data and prior knowledge to improve disease diagnosis and our understanding of biological systems, by leveraging the power of both statistical learning and mechanistic modeling approaches. The first part of my Ph.D. work is to apply increasingly mechanistic biological constraints in in the analysis of high throughput gene expression data to identify molecular signatures of disease phenotypes. Chapter 2 discusses the statistical issues and recommended steps to generate accurate and reproducible molecular signatures. Chapter 3 presents a new computational method that uses the relative expression level of interacting gene pairs as accurate molecular signatures. By incorporating prior knowledge about the relations between genes, this method increases molecular signature reproducibility compared with previous methods. Metabolic networks reconstructed from known reaction stoichiometry and gene-protein-reaction associations provide a mechanistic context to analyze gene expression data. In Chapter 4, I developed a new analysis pipeline that identified perturbations at metabolic branch points (i.e., structures where two reactions consume the same metabolite). Different phenotypes (e.g., cancer v.s. normal) can be accurately distinguished by transcriptional changes at metabolic branch points. Combining reaction expression state (high/low), mass conservation and thermodynamic constraints, I identified additional perturbed branch point reaction pairs that are not apparent from expression data alone. The second part of my PhD work is to contextualize and refine prior knowledge by integration with context-specific high throughput data. In Chapter 5, I developed a novel computational method mCADRE to reconstruct tissue-specific metabolic models. This method can use transcriptomic, proteomic and metabolomics data to infer the metabolic network of a given tissue or cell type. This iii method can be viewed as using tissue-specific omic data to refine and contextualize prior knowledge of metabolism. Using this new method, I reconstructed genome-scale metabolic models for 126 human tissues, providing a tissue-specific encyclopedia of metabolism. In Chapter 6, I applied mCADRE to reconstruct metabolic networks of commonly used breast cancer cell lines. Systematic comparison of model prediction and experimental results revealed different types of inconsistencies that call for further model curation and the development of new modeling approaches