Network-based prioritization of cancer biomarkers by phenotype-driven module detection and ranking

This paper describes an ensemble method with supervised module detection and further module prioritization for reliable network-based biomarker discovery. We design a module detection and ranking method called mRank to discover reliable network modules as cancer diagnostic biomarkers, with two procedures: (1) an iterative supervised module detection guided by phenotypic states in a specific network, (2) a block-based module ranking locally and globally via network topological centrality. We validate its effectiveness and efficiency by identifying hepatocellular carcinoma (HCC) network modules on a comprehensive gene regulatory network with specifying gene interactions by HCC RNA-seq data from the Cancer Genome Atlas (TCGA). These top-ranked modules by mRank get a mean AUC of 0.995 on TCGA HCC dataset with 371 tumor samples and 50 controls by cross-validation SVM. Based on the prior knowledge of cancer dysfunctions enriched in top-ranked modules, 69 genes are identified as HCC candidate biomarkers. They are further validated in independent cohorts with a classifier trained on TCGA HCC dataset. A mean AUC of 0.846 is achieved in distinguishing 976 disease samples from 827 controls. Moreover, some known HCC signatures such as AFP and SPP1 are also included in our identified biomarkers. mRank enables us to find more reliable network modules for cancer diagnosis. For a proof-of-concept study, we validate it in identifying HCC network biomarkers and it is generalizable to other cancers or complex disease. The overall results have demonstrated that mRank can find effective network biomarkers for cancer diagnosis which result in less false positives.