Incorporating topological information for predicting robust cancer subnetwork markers in human protein-protein interaction network

Abstract

Background: Discovering robust markers for cancer prognosis based on gene expression data is an important yet challenging problem in translational bioinformatics. By integrating additional information in biological pathways or a protein-protein interaction (PPI) network, we can find better biomarkers that lead to more accurate and reproducible prognostic predictions. In fact, recent studies have shown that, "modular markers," that integrate multiple genes with potential interactions can improve disease classification and also provide better understanding of the disease mechanisms. Results: In this work, we propose a novel algorithm for finding robust and effective subnetwork markers that can accurately predict cancer prognosis. To simultaneously discover multiple synergistic subnetwork markers in a human PPI network, we build on our previous work that uses affinity propagation, an efficient clustering algorithm based on a message-passing scheme. Using affinity propagation, we identify potential subnetwork markers that consist of discriminative genes that display coherent expression patterns and whose protein products are closely located on the PPI network. Furthermore, we incorporate the topological information from the PPI network to evaluate the potential of a given set of proteins to be involved in a functional module. Primarily, we adopt widely made assumptions that densely connected subnetworks may likely be potential functional modules and that proteins that are not directly connected but interact with similar sets of other proteins may share similar functionalities. Conclusions: Incorporating topological attributes based on these assumptions can enhance the prediction of potential subnetwork markers. We evaluate the performance of the proposed subnetwork marker identification method by performing classification experiments using multiple independent breast cancer gene expression datasets and PPI networks. We show that our method leads to the discovery of robust subnetwork markers that can improve cancer classification.