Integration of heterogeneous data sources for identification of disease genes using computational techniques

Abstract

Genes related to causing some disease are called disease-causing genes or disease genes. In wet-lab experiments, disease genes are identified by mutation analysis, which is expensive and labor extensive. In this thesis, we propose novel computational techniques to predict disease genes. In the first part of this thesis, we proposed five novel topological features obtained from the Protein-Protein Interaction (PPI) network. We applied Support Vector Machine (SVM) and Multi-Layer Perceptron (MLP) networks to predict new cancer genes using these features. We found that SVM performed slightly better than MLP. We also found that the feature, named 2N-index, is the most discriminative feature between cancer genes and other genes. With the availability of various data sources related to genes and disease phenotype, accurate prediction of disease genes is possible by integrating the information available from multiple data sources. We propose several novel computational models to integrate multiple data sources for the identification of disease genes. These models are proposed to prioritize set of candidate disease genes, based on their functional similarity to known disease genes.