DETECTING CANCER-RELATED GENES AND GENE-GENE INTERACTIONS BY MACHINE LEARNING METHODS

Abstract

To understand the underlying molecular mechanisms of cancer and therefore to improve pathogenesis, prevention, diagnosis and treatment of cancer, it is necessary to explore the activities of cancer-related genes and the interactions among these genes. In this dissertation, I use machine learning and computational methods to identify differential gene relations and detect gene-gene interactions. To identify gene pairs that have different relationships in normal versus cancer tissues, I develop an integrative method based on the bootstrapping K-S test to evaluate a large number of microarray datasets. The experimental results demonstrate that my method can find meaningful alterations in gene relations. For gene-gene interaction detection, I propose to use two Bayesian Network based methods: DASSO-MB (Detection of ASSOciations using Markov Blanket) and EpiBN (Epistatic interaction detection using Bayesian Network model) to address the two critical challenges: searching and scoring. DASSO-MB is based on the concept of Markov Blanket in Bayesian Networks. In EpiBN, I develop a new scoring function, which can reflect higher-order gene-gene interactions and detect the true number of disease markers, and apply a fast Branch-and-Bound (B&B) algorithm to learn the structure of Bayesian Network. Both DASSO-MB and EpiBN outperform some other commonly-used methods and are scalable to genome-wide data.