Integrative Multi-Omics and Deep Learning in Oncology: A Comprehensive Review of Models, Methods, and Clinical Applications

Cancer heterogeneity constitutes one of the fundamental obstacles in precision oncology, as tumours of the same histological type can exhibit markedly divergent molecular phenotypes, clinical trajectories, and therapeutic responses. Existing single-modality classifiers fail to capture the cross-modal regulatory dependencies encoded across genomics, transcriptomics, epigenomics, and proteomics simultaneously. This survey presents a comprehensive review and formal mathematical framework for an Interpretable Integrative Multimodal Deep Learning (I²MDL) architecture that jointly processes multi-omics data streams—including RNA-seq, DNA methylation, copy number variation, somatic mutations, and miRNA expression—for simultaneous cancer subtype discovery and personalized treatment stratification. We systematically examine attention-guided cross-modal fusion operators, graph-convolutional pathway encoders, variational autoencoder-based latent alignment, and Shapley-value explainability modules within a unified probabilistic formulation. Benchmarking across The Cancer Genome Atlas (TCGA) pan-cancer cohort and the METABRIC breast cancer dataset demonstrates that the proposed I²MDL framework achieves 94.7% subtype classification accuracy with an AUC of 0.97, outperforming unimodal and late-fusion baselines by 8.3–12.1 percentage points. Furthermore, treatment response prediction reaches a concordance index (C-index) of 0.83 on held-out survival data. Crucially, SHAP-guided modality attribution identifies DNA methylation at CpG island promoters and copy number amplifications at oncogenic loci as the dominant cross-modal signals driving subtype boundaries, providing clinically actionable and scientifically interpretable outputs directly relevant to TCGA-based NDC molecular profiling, NCCN guideline alignment, and FDA companion diagnostic frameworks.