Deep learning-based analysis of routine histopathology slides predicts response to PARP inhibitors in high grade serous ovarian cancer

Abstract 138 PARP inhibitors (PARPi) have transformed the treatment of high-grade serous tubo-ovarian cancer 139 (HGSOC), but current predictive biomarkers lack accuracy, scalability, and availability. Here, we 140 present an AI framework that infers PARPi response directly from routine hematoxylin and eosin 141 (H&E)-stained slides. Given the limited availability of large PARPi-treated datasets, we use platinum– 142 taxane chemotherapy response as a surrogate endpoint, training on 519 platinum–taxane-treated patients 143 and validating in two independent cohorts comprising 3,905 patients. Across these cohorts, the model 144 stratifies patients into risk groups based on overall survival, despite the variabilities across centers. 145 Applied to an independent cohort of 118 BRCA-wildtype patients receiving first-line niraparib, the 146 model’s risk predictions are associated with progression-free and overall survival, independent of 147 clinical covariates. To our knowledge, this is the first demonstration that PARPi response in HGSOC 148 can be derived directly from routinely collected H&E slides, supporting more precise and accessible 149 patient selection for PARPi therapy. 150 151