HARNESSING MULTI-MODAL ARTIFICIAL INTELLIGENCE FOR GENOMIC PATHOLOGY: A COMPREHENSIVE REVIEW OF ARCHITECTURES, CLINICAL FEASIBILITY, AND TRANSLATIONAL PROSPECTS

The exponential accumulation of biomedical data has fundamentally transformed modern healthcare, yet millions of archived genomic samples and histopathological specimens remain systematically underexploited. This review examines the convergence of artificial intelligence, genomics, and digital pathology—particularly through initiatives like the Silent Genome project—to unlock latent information stored within historical biological archives. We synthesize evidence across five critical research domains: neural network architectures for reconstructing degraded genetic profiles, feasibility of retrospective genomic-pathologic reanalysis in clinical laboratories, prediction of mutational signatures from histology alone, multimodal fusion strategies for cancer subtype classification and drug response prediction, and mechanisms by which hybrid systems accelerate target identification. Analysis of over 40 peer-reviewed studies reveals that while technical feasibility has been firmly established across multiple cancer types, clinical deployment remains hindered by domain shift, interpretability challenges, and insufficient external validation. Multimodal approaches consistently outperform unimodal baselines for prognosis and subtype classification (c-indices 0.60–0.85), though drug response prediction remains nascent and robustness to site-specific biases requires systematic characterization. We argue that translation to clinical practice requires not sophistication alone, but rigorous cross-institutional validation, transparent interpretability frameworks, and equity-centered study design. This article provides a critical synthesis of opportunities and persistent gaps, offering a roadmap for responsible integration of these systems into routine pathology workflows.