Mining Biological Networks towards Protein-Complex Detection and Gene-Disease Association

Eileen Marie Hanna


Large amounts of biological data are continuously generated nowadays, thanks to the advancements of high-throughput experimental techniques. Mining valuable knowledge from such data still motivates the design of suitable computational methods, to complement the experimental work which is often bound by considerable time and cost requirements. Protein complexes, or groups of interacting proteins, are key players in most cellular events. The identification of complexes not only allows to better understand normal biological processes but also to uncover disease-triggering malfunctions. Ultimately, findings in this research branch can highly enhance the design of effective medical treatments. The aim of this research is to detect protein complexes in protein-protein interaction networks and to associate the detected entities to diseases. The work is divided into three main objectives: first, develop a suitable method for the identification of protein complexes in static interaction networks; second, model the dynamic aspect of protein interaction networks and detect complexes accordingly; and third, design a learning model to link proteins, and subsequently protein complexes, to diseases. In response to these objectives, we present, ProRank+, a novel complex-detection approach based on a ranking algorithm and a merging procedure. Then, we introduce DyCluster, which uses gene expression data, to model the dynamics of the interaction networks, and we adapt the detection algorithm accordingly. Finally, we integrate network topology attributes and several biological features of proteins to form a classification model for gene-disease association. The reliability of the proposed methods is supported by various experimental studies conducted to compare them with existing approaches. ProRank+ detects more protein complexes than other state-of-the-art methods. DyCluster goes a step further and achieves a better performance than similar techniques. Then, our learning model shows that combining topological and biological features can greatly enhance the gene-disease association process. Finally, we present a comprehensive case study of breast cancer in which we pinpoint disease gene using our learning model; subsequently, we detect favorable groupings of those genes in a protein interaction network using the ProRank+ algorithm.