Deep learning to develop transcriptomic model for survival prediction in TCGA patients with hepatocellular carcinoma.

Abstract

Background: This study aimed to investigate the prognostic value of transcriptome and clinical data of Hepatocellular carcinoma (HCC) patients for overall survival (OS) by deep learning method. Methods: A total of 371 HCC patients with 20530 level three RNA-sequencing data were from The Cancer Genome Atlas (TCGA). Cox-nnet model, a deep learning model through an artificial neural network extension of the Cox regression model, was used for OS prediction. The patients were randomly split into train-set and test-set (7:3). In train-set, the significant genes associated with OS under univariate Cox regression were considered for modeling. Clinical parameters, including age, gender, pathologic stage, child pugh classification, creatinine level etc. were also considered. The Cox-nnet model was developed by cross-validation. Its discrimination was determined by the concordance index (CI) in the independent test-set and compared with multivariable Cox regression. The clustering method Uniform Manifold Approximation and Projection (UMAP) was used for revealing biological information from the hidden layer in the model. Results: In the train-set (n = 259), 1505 genes and two clinical variables (child pugh score and creatinine level) were significantly associated with OS (adjusted P-value < 0.05). To avoid overfitting, only 40 most significant genes were included in the Cox-nnet model. In the test-set (n = 112), the CI of Cox-nnet (0.76, se = 0.04) is better than the CI of multivariable Cox regression (0.71, se = 0.05). The difference between good or poor survival subgroups classified by Cox-nnet was remarkably significant (P-value = 1e-4, median OS: 80.7 vs. 25.1 months). In the Cox-nnet model with all significant variables, the weights in the hidden layer were clustered by UMAP into 3 positive clusters and 2 negative clusters, which are enriched in GO/KEGG. The “cell cycle” and “complement and coagulation cascades” are the most important signal pathways in positive and negative clusters, respectively. Conclusions: Combining transcriptomic and clinical data, and with deep learning algorithm, we built and validated a robust model for survival prediction in HCC patients. Our study would be useful to explore the clinical implications in survival prediction and corresponding genetic mechanisms.