Identification of Relevant Protein-Gene Associations by Integrating Gene Expression Data and Transcriptional Regulatory Networks.

Abstract

One challenge in systems biology is integrating different biological data types to more accurately describe how a biological system functions. If networks describing a pathway or a particular regulatory activity is merged with gene expression data, the specific regulator-gene portions of the pathway responsible for changes in gene expression could be identified. In this thesis, I hypothesize that merging gene expression data with transcriptional network information will allow me to identify possibly regulatory mechanisms that govern the observed gene expression patterns. I developed a computational approach to merge these data types and demonstrated that the method can identify which regulator-gene associations better explain the gene expression patterns even when the activities of the regulators are not observed. Due to the complex interplay of different regulatory proteins during mRNA regulation, the individual activity of these proteins often can’t be measured directly. Previously described methods of identifying protein-gene associations have two main limitations: (1) failing in identifying combinatoric relationships and (2) prediction of inactive regulatory associations. The methods I developed model a regulatory network as a bipartite network with a top layer of unobserved regulators (protein activities) connected to a lower level of observed variables (mRNA expression values). This bipartite approach has been used in the past to study regulatory networks but assuming a linear mixing model. In contrast, I use a multinomial model that better captures the nonlinear patterns seen in gene regulation networks: Bayesian networks. I tested the developed tools using synthetic, E. coli, and human expression data. The synthetic data results show that the method is capable of identifying relevant connections. When using E.coli and human gene expression data, the method identified a simplified regulatory network that is both mechanistically sound and maximally consistent with the expression data. By identifying regulatory relationships that are apparently active given a set of gene expression data, this thesis provides a new lens to view gene expression data in general. The methods developed here are directly applicable to large transcriptional networks of any species and provide the foundation for a new branch of bioinformatics analysis.