Treatment-aware deep learning enables counterfactual prediction of individual benefit from PARP inhibitors in ovarian cancer.

Homologous recombination deficiency (HRD) assays are used to select patients with ovarian cancer for PARP inhibitors, but they do not fully capture the heterogeneity of treatment benefit. Using data from the phase III PAOLA-1 randomized trial, we developed a deep learning model based on diagnostic hematoxylin and eosin whole-slide images from treatment-naive primary tumors to estimate individual benefit from maintenance olaparib plus bevacizumab versus bevacizumab alone. The model combined image-derived features with randomized treatment assignment and progression-free survival data to estimate, for each patient, the risk of progression under both treatments. The difference between these risks defined a continuous Estimated Treatment Improvement (ETI) score, which quantifies patient-level treatment effect. In 421 patients with available pre-treatment slides, ETI showed a strong treatment-biomarker interaction for first progression-free survival (interaction HR 0.36, 95% CI 0.22-0.59; p < 0.005), while HRD status showed a similar interaction magnitude (HR 0.41, 95% CI 0.25-0.64; p < 0.005). ETI remained independently associated with outcome after adjustment for clinical covariates and HRD status. Subgroup and STEPP analyses indicated that ETI further stratified benefit within both HRD-positive and HRD-negative tumors. Attention maps suggested an association between multifocal tumor-infiltrating lymphocytes and predicted benefit from olaparib. These findings suggest that histology-based modeling of heterogeneity of treatment effects may complement HRD testing for treatment decision-making, although external validation remains necessary.