Development and validation of an AI model for predicting germline BRCA1/2 mutations from HR+/HER2- breast cancer histology images.

Abstract

e13611 Background: Pathogenic germline BRCA1/2 mutations (g BRCA mut) are critical biomarkers in breast cancer (BC), guiding prevention strategies. However, testing is mainly limited to clinical context of suspected germline predisposition (family history, young age), leaving up to 50% of g BRCA mut carriers undiagnosed. BRCA status is increasingly required for therapeutic decisions. Despite breakthroughs in capacities of analysis, molecular testing is time consuming and requires dedicated facilities and remains expensive, limiting its access. Leveraging histological data presents an opportunity to address this gap through a scalable screening approach. As part of Owkin’s collaboration with Gustave Roussy (GR) and Centre Léon Bérard (CLB) through PortrAIt, a French consortium aimed at advancing precision medicine, we developed a proof-of-concept (PoC) AI model to predict the likelihood of g BRCA mut directly from standard stained histology images of early BC HR+/HER2- patients, paving the way for further development and validation. Methods: We developed a PoC multiple instance learning model trained on GR data (n = 671, 50% gBRCAmut) to predict gBRCA status from histology features. The slides were scanned using Olympus VS200 and stained using Hematoxylin and Eosin Saffron (HES). External validation was performed on three independent cohorts (CLB, US, UK), including out-of-domain scanners and staining types: Hematoxylin and Eosin (H&E), Hematoxylin Phloxine Saffron (HPS). Robustness was assessed by computing the Pearson correlation (R) and Concordance Correlation Coefficient (CCC) between predictions obtained for samples from the same tumor blocks, prepared with different stains and scanners. Results: Our PoC model shows AUC performances ranging from 0.72 [0.64, 0.8] to 0.89 [0.71, 1.0] for resections. It also shows strong robustness performances across stains and scanners with an R of 0.95 (0.026) and a CCC of 0.87 (0.047). Conclusions: Our PoC AI model achieves promising performance across diverse datasets with varying preparation protocols and scanner models, with AUCs ranging from 0.72 to 0.89 and strong robustness (R = 0.95, CCC = 0.87). This approach shows potential for scalable screening of gBRCA1/2 mutations, enabling an expanded testing pool and improved testing efficiency. Future work aims to optimize model performance on both resection and biopsy specimens, notably by advancing to multi-cohort training, and conduct further validation studies. Cohort Staining Scanner N (% gBRCAmut) AUC (95% CI) Specificity at 90% Sensitivity CLB H&E Leica GT450 135 (25.9%) 0.79 [0.71, 0.86] 46% CLB H&E Leica AT2 134 (25.4%) 0.78 [0.71, 0.86] 46% CLB HPS Leica GT450 145 (24.8%) 0.74 [0.67, 0.81] 36% CLB HPS Leica AT2 139 (25.9%) 0.72 [0.64, 0.8] 39% US H&E Pramana 48 (6.2%) 0.89 [0.71, 1.0] 71% US H&E Huron 35 (2.9%) 0.85 [0.76, 0.94] 85% UK H&E Hamamatsu S60 26 (15.4%) 0.83 [0.65, 0.97] 59%