Gene- and Pathway-Based Deep Neural Network for Multi-omics Data Integration to Predict Cancer Survival Outcomes

Abstract

Data integration of multi-platform based omics data from biospecimen holds promise of improving survival prediction and personalized therapies in cancer. Multi-omics data provide comprehensive descriptions of human genomes regulated by complex interactions of multiple biological processes such as genetic, epigenetic, and transcriptional regulation. Therefore, the integration of multi-omics data is essential to decipher complex mechanisms of human diseases and to enhance treatments based on genetic understanding of each patient in precision medicine. In this paper, we propose a gene- and pathway-based deep neural network for multi-omics data integration (MiNet) to predict cancer survival outcomes. MiNet introduces a multi-omics layer that represents multi-layered biological processes of genetic, epigenetic, and transcriptional regulation, in the gene- and pathway-based neural network. MiNet captures nonlinear effects of multi-omics data to survival outcomes via a neural network framework, while allowing one to biologically interpret the model. In the extensive experiments with multi-omics data of Gliblastoma multiforme (GBM) patients, MiNet outperformed the current cutting-edge methods including SurvivalNet and Cox-nnet. Moreover, MiNet’s model showed the capability to interpret a multi-layered biological system. A number of biological literature in GBM supported the biological interpretation of MiNet. The open-source software of MiNet in PyTorch is publicly available at https://github.com/DataX-JieHao/MiNet.