Finding robust subnetwork markers that improve cross-dataset performance of cancer classification

Recent studies have shown that the utilization of additional biological information, such as pathway knowledge or protein-protein interaction data, can improve cancer classification in terms of prediction accuracy and reproducibility of the obtained biomarkers. In this study, we propose a method for identifying subnetwork markers from a human PPI network, which can be used to predict breast cancer prognosis. The proposed method utilizes a clustering algorithm based on a message passing scheme. Our experiments using two large-scale breast cancer datasets show that the identified subnetwork markers are more reliable and reproducible across datasets compared to those identified by an existing method, hence they may ultimately lead to more effective cancer classifiers.