Abstract PS5-03-17: Ai-derived Morphometric And Transcriptomic Biomarkers From H&E-stained Images Predict Response To Durvalumab And Olaparib In Metastatic Triple Negative Breast Cancer

Abstract Introduction: PARP inhibitors (PARPi) have demonstrated the potential to enhance tumor immunogenicity and sensitize cancer cells to immune checkpoint blockade, as shown in the AMTEC clinical trial. However, identifying PARPi responders and understanding the tumor's adaptive response remain significant clinical challenges. Advances in multi-omic profiling offer unprecedented resolution into these dynamics, yet the complexity and volume of such data demand sophisticated analytical approaches. By applying AI to histopathology images, we aim to uncover predictive morphogenomic biomarkers of PARPi response, potentially enabling more precise treatment selection and to reveal novel mechanisms of therapeutic adaptation in the tumor and its immune microenvironment. Method: We applied two complementary approaches to identify predictive morphometric and morphogenomic biomarkers in biopsy samples collected during the AMTEC trial. Using the PATHOMIQ Phenotype Atlas, we quantified fold changes in individual morphological phenotypes between matched pre-treatment and on-treatment biopsies to identify morphology-based predictors of treatment response. Using the PATHOMIQ Genoscope, a deep learning model that infers RNA expression from H&E-stained slides, we estimated gene expression profiles directly from histology images. Genes that were significant in univariate Cox modeling were evaluated in multivariate models to identify transcriptomic predictors. Result: To assess phenotypic changes across serial biopsies, we computed fold changes in the prevalence of individual morphological phenotypes between pre-treatment and on-treatment samples. Specific morphometric changes related to desmoplasty and necrosis were significantly associated with progression-free survival (PFS). Their emergence in on-treatment biopsies enabled clear patient stratification, as confirmed by a log-rank test (p < 0.001). Separately, univariate Cox proportional hazards modeling identified 15 genes, predicted directly from H&E images via Genoscope, as significantly associated with PFS. A multivariate model incorporating the top-ranked genes along with FOXC1 demonstrated robust stratification of patient outcomes. Kaplan-Meier estimation using a median-expression cutoff revealed a significant survival difference (p < 0.01, log-rank test). Discussion: These results show that AI-based analysis of H&E slides can identify both morphological and gene expression markers predictive of PARP inhibitor response. The identification of these desmoplastic changes during treatment may signal early resistance and these necrotic changes signal early response, while inferred expression of genes such as FOXC1, a known biomarker, was linked to poor progression-free survival. This morphogenomic approach enables non-destructive, scalable patient stratification and may support real-time treatment adaptation in future studies. Citation Format: H. Muhammad, S. S. Chavan, C. Feng, SMMART Clinical Trials Program, H. S. Basu, W. Huang, R. Rajat, G. Wilding, S. Kummar, G. B. Mills. Ai-derived Morphometric And Transcriptomic Biomarkers From H&E-stained Images Predict Response To Durvalumab And Olaparib In Metastatic Triple Negative Breast Cancer [abstract]. In: Proceedings of the San Antonio Breast Cancer Symposium 2025; 2025 Dec 9-12; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2026;32(4 Suppl):Abstract nr PS5-03-17.