Deep Learning for Multi-Omics Integration in Cancer: From Molecular Mechanisms to Clinical Translation

Cancer is a highly heterogeneous disease driven by complex interactions across multiple molecular layers, including genomic, epigenomic, transcriptomic, proteomic, metabolomic, and microbiomic alterations. While high-throughput technologies have enabled large-scale profiling of these omics layers, extracting coherent biological and clinical insights from such high-dimensional and heterogeneous data remains a fundamental challenge. Deep learning has emerged as a powerful paradigm for multi-omics data integration, enabling the discovery of nonlinear, hierarchical relationships that underlie cancer initiation, progression, and therapeutic response. Unlike traditional statistical and shallow machine learning approaches, deep learning architectures can jointly model diverse data modalities, accommodate missing or noisy measurements, and learn shared latent representations that capture cross-omics dependencies. This article reviews recent advances in deep learning–based multi-omics integration for cancer research, spanning tumor classification, biomarker discovery, pathway inference, drug response prediction, and precision oncology. Emphasis is placed on innovative integrative architectures, including autoencoders, graph neural networks, attention-based transformers, and multimodal foundation models, rather than classical single-omics analyses. Key translational challenges related to model interpretability, data bias, scalability, and clinical deployment are critically discussed. Collectively, this review highlights how deep learning–driven multi-omics integration is reshaping cancer research and accelerating the transition from molecular profiling to clinically actionable precision oncology.