Colorectal cancer survival analysis using deep learning based multi-omics data integration

Colorectal cancer (CRC) is the third most common type of cancer, yet multi-omics-based survival analysis studies for CRC are relatively limited compared to other malignancies. Accurate survival analysis of cancer patients is crucial for personalized treatment strategies. While multi-omics data provide a comprehensive molecular profile of tumours, their high dimensionality poses challenges. Most existing studies have focused on integrating single, two, or three omics data types. In this study, we incorporate four distinct multi-omics data types, RNA sequencing (RNA-seq), microRNA (miRNA), DNA methylation (DNA), and copy number variation (CNV) from The Cancer Genome Atlas (TCGA) colorectal cancer cohort to enhance survival prediction. To handle high-dimensional data, we employ autoencoders to learn compressed feature representations. The majority of previous multi-omics-based survival analysis research has relied on the Cox proportional hazards (Cox-PH) model, which assumes a linear relationship between covariates and the hazard function. In contrast, we employ DeepSurv, a deep neural network-based extension of Cox-PH, capable of capturing non-linear interactions among multi-omics features. Specifically, we modify DeepSurv by incorporating batch normalization and dropout layers, improving training stability, and preventing overfitting. Our results demonstrate that integrating multiple omics data enhances predictive performance over single-omics data, and our modified DeepSurv based model outperforms the traditional Cox-PH model, highlighting the advantages of deep learning for CRC survival analysis. These findings emphasize the importance of leveraging deep learning and comprehensive multi-omics integration to improve colorectal cancer survival analysis.