GenoGlyph: Pan-cancer genomic mutation inference and risk stratification from diagnostic histopathology slides

Genomic alterations drive therapeutic decisions in solid tumors, yet next-generation sequencing remains inaccessible in a substantial fraction of clinical settings worldwide. We present GenoGlyph, an interpretable deep learning framework designed to decode the grammar of histopathology for pan-cancer genomic mutational inference and clinical risk stratification. Across 6,391 whole-slide images spanning 14 solid tumor types, GenoGlyph robustly predicted actionable alterations (e.g., TP53, KRAS, PIK3CA, APC), yielding AUCs of 0.63-0.98 and outperforming existing pan-cancer benchmarks. To determine whether these morphologically inferred genomic signatures capture underlying tumor biology, we integrated mutation predictions with matched transcriptomic profiles and observed concordant activation of canonical pathways, such as elevated mTORC1 signaling in PIK3CA-mutated tumors. Furthermore, we demonstrate GenoGlyph's utility as a functional genomics tool by showing that mutations predicted by the model, including sequencing-discordant (false-positive) cases, exhibit pathway perturbations consistent with the inferred genotype, such as suppression of DNA repair programs in TP53-mutated tumors. Finally, we validate the clinical utility of GenoGlyph's learned representations by demonstrating independent prognostic value for overall, disease-free, recurrence-free, and progression-free survival across four independent external cohorts (n=982). GenoGlyph establishes that complex genotype-phenotype relationships are systematically legible from standard H&E sections. Rather than treating mutation prediction as an endpoint, our framework leverages these genomic signatures to learn latent features encapsulating hidden tumor-intrinsic and microenvironmental states. This provides a scalable, biologically grounded paradigm for democratized precision oncology, functional variant interpretation, and survival stratification.