Hypergraph Models for Identifying Co-Regulatory Genomic Interactions

Abstract

A comprehensive understanding of biological systems requires the analysis of higher-order interactions among many genomic factors. Various genomic factors cooperate to affect biological processes including cancer occurrence, progression and metastasis. However, the complexity of genomic interactions presents a major barrier to identifying their co-regulatory roles and functional effects. Thus, this dissertation addresses the problem of analyzing complex relationships among many genomic factors in biological processes including cancers. We propose a hypergraph approach for modeling, learning and extracting: explicitly modeling higher-order genomic interactions, efficiently learning based on evolutionary methods, and effectively extracting biological knowledge from the model. A hypergraph model is a higher-order graphical model explicitly representing complex relationships among many variables from high-dimensional data. This property allows the proposed model to be suitable for the analysis of biological and medical phenomena characterizing higher-order interactions between various genomic factors. This dissertation proposes the advanced hypergraph-based models in terms of the learning methods and the model structures to analyze large-scale biological data focusing on identifying co-regulatory genomic interactions on a genome-wide level. We introduce an evolutionary approach based on information-theoretic criteria into the learning mechanisms for efficiently searching a huge problem space reflecting higher-order interactions between factors. This evolutionary learning is explained from the perspective of a sequential Bayesian sampling framework. Also, a hierarchy is introduced into the hypergraph model for modeling hierarchical genomic relationships. This hierarchical structure allows the hypergraph model to explicitly represent gene regulatory circuits as functional blocks or groups across the level of epigenetic, transcriptional, and post-transcriptional regulation. Moreover, the proposed graph-analyzing method is able to grasp the global structures of biological systems such as genomic modules and regulatory networks by analyzing the learned model structures. The proposed model is applied to analyzing cancer genomics considered as a major topic in current biology and medicine. We show that the performance of our model competes with or outperforms state-of-the-art models on multiple cancer genomic data. Furthermore, the propose model is capable of discovering new or hidden patterns as candidates of potential gene regulatory circuits such as gene modules, miRNA-mRNA networks, and multiple genomic interactions, associated with the specific cancer. The results of these analysis can provide several crucial evidences that can pave the way for identifying unknown functions in the cancer system. The proposed hypergraph model will contribute to elucidating core regulatory mechanisms and to comprehensive understanding of biological processes including cancers.