Explainable artificial intelligence for multi-modal cancer analysis: From genomics to immunology.

Cancer remains one of the most biologically intricate and heterogeneous diseases, with variability spanning genetic, epigenetic, metabolic, spatial, and immunological dimensions. This multifactorial complexity is shaped by tumor-microenvironment interactions, immune modulation, clonal evolution, and therapeutic pressures, creating profound challenges for accurate prognosis, treatment selection, and durable response prediction. Conventional single-modality analytical approaches often capture only isolated features of this landscape, limiting their ability to generalize across patient populations and clinical settings. Selective multimodal deep learning (MDL) provides a transformative framework by mechanism-informed integration of complementary biomedical data-including genomics, transcriptomics, histopathology, medical imaging, proteomics, metabolomics, electronic health records (EHRs), and immunological profiles into unified models capable of capturing cross-scale dependencies and emergent patterns, while mitigating interference through targeted fusion strategies. The interpretability of these models is enhanced by explainable artificial intelligence (XAI), which translates complex computational predictions into transparent, biologically grounded explanations that align with known oncological and immunological processes. A distinctive focus of this review is immunology-informed integration, which connects predictive signatures to tumor-immune microenvironment dynamics, enhancing biomarker discovery, immunotherapy stratification, and hypothesis generation. We review key foundations, fusion strategies, and explainability methods in precision oncology, emphasizing rigorous validation frameworks that include statistical metrics and biological plausibility assessments. We address challenges such as data harmonization and modality imbalance, and discuss criteria for selecting modalities to ensure scientific validity. Finally, we outline future directions like federated learning for privacy, causal inference-driven explainability, and patient-specific digital twins, all aimed at advancing personalized precision oncology.