Deep Learning-Based Inference of Mutational Signature Exposures from Histology Across Tumour Types

Background/Objectives: Mutational signatures reflect the processes that shape cancer genomes. We investigated whether continuous mutational-signature exposures could be inferred from routine histology across tumour types. Methods: We developed whole-slide image regression models for 30 single-base substitution (SBS) signatures in 2741 The Cancer Genome Atlas (TCGA) patients across seven tumour types and evaluated them in 580 Clinical Proteomic Tumor Analysis Consortium (CPTAC) patients. RetCCL mean-feature ridge regression and RetCCL/CTransPath attention-based multiple-instance learning were combined by fixed unweighted averaging. Final ensemble outputs were standardised SBS scores. Patient-disjoint evaluation, full-panel false discovery rate (FDR) adjustment and tumour-labelled slide sensitivity were performed. Results: Among 185 pairs with correlations available in both cohorts, 73 showed positive correlations in both cohorts. Fourteen had nominal positive-direction p < 0.05 in both, including breast cancer–SBS3 (TCGA r = 0.289, CPTAC r = 0.215) and lung adenocarcinoma–SBS4 (r = 0.128 and 0.263). Cohort-specific FDR-adjusted values are provided for all 210 pairs. All 14 associations remained positive after restricting inputs to tumour-labelled slides. The SBS3 score correlated with continuous homologous recombination deficiency (r = 0.332). Exploratory TCGA SBS3, SBS6 and SBS10 scores yielded areas under the receiver operating characteristic curve of 0.693, 0.619 and 0.628 for homologous recombination deficiency-high, microsatellite instability and qualifying POLE/POLD1 mutations. External MSI and polymerase AUCs were 0.573 and 0.680, respectively. Conclusions: Histology-derived scores were associated with selected genomic SBS exposures in both cohorts. Further studies are needed to identify the tissue features underlying these associations and assess their clinical relevance.