Predicting bladder cancer prognosis by integrating multi-omics data through a transfer learning-based Cox proportional hazards network

Abstract

Predicting bladder cancer outcomes is important for patients’ treatments, and it’s common to predict the outcomes from omics data. However, using a single type of omics data suffers from data noise since individual omics type represents only one single view of bladder cancer patients. In this study, we have estimated bladder cancer prognosis by integrating multi-omics data including RNA-seq, miRNA-seq, DNA methylation, and copy number variation data. To effectively integrate the multi-omics data, we have developed a transfer-learning based Cox proportional hazards network (TCAP) by utilizing an integrated loss function consisted of two modules: the data reconstruction module to ensure learning a representative hidden layer for the input data, and the proportional hazard module to estimate patients’ risks. The experiments on 336 patients from The Cancer Genome Atlas (TCGA) showed that our method achieved a concordance index (C-index) of 0.665, which is higher than previous methods. In consideration of the expense to obtain multi-omics data in clinics, we fitted the risks estimated from TCAP by training an XGboost model based on mRNA data only. The model achieved a reasonable C-index of 0.621, and independent tests on three additional datasets achieved an average C-index of 0.637 ± 0.047. The essentially same result as the one achieved on TCGA dataset indicates the robustness of our model. Based on the risk subgroups divided by TCAP, we identified 12 candidate genes that affected the survival of bladder cancer patients, among which 7 genes (58.3%) have been proved to associate with bladder cancer through literature review. In summary, the results indicated that we have constructed an accurate and robust model for predicting bladder cancer outcomes.