Improvement of cancer subtype prediction by incorporating transcriptome expression data and heterogeneous biological networks

Abstract

Background: Identification of cancer subtypes is of great importance to facilitate cancer diagnosis and therapy. A number of methods have been proposed to integrate multi-sources data to identify cancer subtypes in recent years. However, few of them consider the regulatory associations between genome features and the contribution weights of different data-views in data integration. It is widely accepted that the regulatory associations between features play important roles in cancer subtype studies. In addition, different data-views may have different contributions in data integration for cancer subtype prediction. Results: In this paper, we propose a method, CSPRV, to improve the cancer subtype prediction by incorporating multi-sources transcriptome expression data and heterogeneous biological networks. We extract multiple expression features of each genome element based on the regulatory associations in the heterogeneous biological networks and use a generalized matrix correlation method (RV 2 ) to predict the similarities between samples in each view of expression data. We fuse the similarity information in multiple data-views according to different integration weights. Based on the integrated similarities between samples, we cluster samples into different subtype groups. Comprehensive experiments on TCGA cancer datasets demonstrate that the proposed method can identify more clinically meaningful cancer subtypes comparing with most existing methods. Conclusions: The consideration of regulatory associations between biological features and data-views contribution is important to improve the understanding of cancer subtypes. The proposed method provides an open framework to incorporate transcriptome expression data and biological regulation network to predict cancer subtypes.