IHGAMP: Pan-Cancer HRD Prediction From Routine H&E Whole-Slide Images Using Foundation Models

Homologous recombination deficiency (HRD) guides PARP inhibitor and platinum-based therapy selection but relies on costly next-generation sequencing with limited accessibility. Here, we show that IHGAMP, a computational framework using vision transformer foundation models, predicts HRD from routine H&E whole-slide images with AUROC 0.766 (95% CI 0.727-0.803) using OpenCLIP embeddings on the TCGA held-out test set (n = 8109 patients, 31 cancer types), improving to 0.812 with histopathology-pretrained OpenSlideFM. External evaluation on 927 patients from seven independent cohorts confirmed generalization in adenocarcinoma/serous histologies (CPTAC-LUAD AUROC 0.723; PTRC-HGSOC platinum resistance AUROC 0.673), with attenuation in squamous histologies. True external validation using the frozen TCGA-trained model on CPTAC-LUAD yielded AUROC 0.671, confirming cross-institutional generalization. Pathway specificity analysis demonstrated that the apparent TP53 correlation was mediated through HRD (partial r = 0.112 after controlling for scarHRD), and evaluation on MMR-deficient endometrial carcinoma yielded below-random prediction (AUC = 0.445), ruling out general genomic instability capture. Sensitivity analysis across label definitions (scarHRD ≥ 33 and ≥ 42) showed robust performance, and a BRCA1/2 mutation-trained comparator performed near-randomly (AUC = 0.486), confirming that learned features reflect HRD biology beyond single-gene effects. IHGAMP supports a screening/triage role to prioritize confirmatory molecular testing where appropriate.