Improved Prediction of Cancer Outcome Using Graph-Embedded Generative Adversarial Networks

Abstract

Precise prognosis of cancer patients is important because it is associated with suggesting appropriate therapeutic strategies. Several computational and statistical methods have been proposed, but further improvement of these methods in terms of prediction accuracy is required. This paper presents a deep learning-based method for learning networks of prognostic genes, instead of only individual biomarker genes, for more accurate cancer prognosis prediction. This method utilizes generative adversarial networks, where the generator uses a biological network instead of a traditional fully connected network to learn the distributions of gene expression (mRNA), copy number variation, single nucleotide polymorphism, and DNA methylation data from cancer patients. The proposed model was applied to seven cancer types and exhibited higher prediction accuracy as compared to the existing state-of-the-art methods. On average, the area under the curve (AUC) was improved by 4% compared to the best performing existing methods for seven cancer types. In particular, for pancreatic adenocarcinoma, AUC was improved by 27.9%. The identified prognostic genes were reproducible and functionally meaningful. To the best of our knowledge, the proposed method represents the first attempt to learn genetic networks from multi-omics data.