Design and implementation of a cancer prognosis prediction system for multi-omics data

Abstract

Cancer prognosis prediction plays a crucial role in guiding personalized treatment decisions. This thesis presents a deep learning-based system that integrates multi-omics data—including RNA expression, DNA methylation, and somatic mutations—to improve survival prediction and provide personalized drug recommendations. The system employs modality-specific encoders to extract features from each omics layer, followed by a fusion network that generates unified patient embeddings. A Cox proportional hazards model is applied to estimate survival risks, while drug recommendations are generated through embedding-based similarity matching between patient profiles and drug molecular fingerprints. The system is evaluated on the TCGA-BRCA dataset with 800 breast cancer patients. The proposed model achieves a concordance index (C-index) of 0.72 and an integrated Brier score of 0.18, outperforming several baseline models such as CoxPH and DeepSurv. Subtype classification reaches an accuracy of 87% and an F1 score of 0.85, demonstrating the biological relevance of the learned embeddings. An interactive visualization interface is also developed to support clinical interpretability by enabling users to explore patient risk scores, survival curves, and recommended drugs. While the system demonstrates strong performance, limitations include the use of retrospective data and the need for prospective clinical validation. This study highlights the potential of multi-omics integration and deep learning to support precision oncology.