Abstract PS3-06-24: Transcriptomics-guided AI outperforms and expands on OncotypeDX to predict Hormonal therapy response and Chemotherapy benefit from H&E-stained Whole Slide images in ER+/HER2- Breast Cancer

Abstract Purpose Oncotype DX (ODX) is one of the most widely used genomic tests for risk stratification and prediction of response to hormonal therapy (HT) of early stage ER+/HER2- node-negative breast cancer. However, this test does not provide guidance on the extent of chemotherapy (CT) benefit. We developed an AI-based model that predicts response to HT and failure of HT+CT directly from H&E-stained whole slide images (WSI) of the breast tumor in surgical resections through inferring underlying genomic and signaling pathway alterations directly from morphology. Our model was developed to improve upon the performance of ODX in risk stratification and prediction of distant metastasis in HT treated patients but also extend its clinical utility by predicting HT+CT failure to enable consideration of other therapy options. Methods Our AI technology leverages cross-modality transfer learning to infer transcriptomic signals from morphological features. We first developed a biomarker for HT-response and combined HT+CT-failure using only genomic data by mapping RNA-Seq to metastasis outcomes. These biomarkers integrated thousands of gene expressions and signaling pathway activation. After validating the predictive accuracy of these genomic biomarkers, we trained a deep learning model to infer the biomarkers directly from H&E WSIs of the invasive breast tumor. For external validation, we used our AI model to predict on H&E-stained WSIs from 300 early stage ER+/HER2- node-negative breast cancer patients including 150 treated with HT and 150 with HT+CT, based on ODX recommendations. For each WSI, we predicted the presence or absence of our HT-response and HT+CT-failure biomarkers, enabling further stratification of patients within their respective groups. Kaplan-Meier survival estimation and log-rank statistical testing were used to measure the magnitude and significance of stratification in the HT and HT+CT treated groups. Results The 150 ER+/HER2-node negative breast cancer patients who were treated with HT alone based on ODX scores ranging from 0 to 25 were stratified with our HT-response biomarker with a hazard ratio (HR) of 0.28 (p < 0.002). Fifteen patients with ODX<20 treated with HT alone metastasized and our AI model identified 11/15 (73%) of these patients as HT-failures. The 150 patients treated with HT+CT based on ODX scores ranging from 26 to 71 were stratified using our CT-failure biomarker with an HR of 2.42 (p < 0.02), identifying 22/35 (65%) patients who metastasized after being treated with HT+CT. In a multivariate analysis, ODX score added prognostic information to the AI- based HT-response biomarker (p=0.04) and age added prognostic information to the AI- based HT+CT-failure biomarker (p=0.03). By integrating signaling pathway alterations together with changes in gene expressions, we were able to utilize these enriched genomic changes using our AI model to identify responders and non-responders to HT and HT+CT. Conclusion To our knowledge, this is the first AI model that can predict response to HT and HT+CT in early stage ER+/HER2- node-negative breast cancer. Our AI model by itself identified 73% patients with low ODX scores who developed metastasis following HT alone and 63% patients with high ODX who did not benefit from HT+CT. Combining ODX recurrence scores with the output of our AI model improved the prediction of response in HT treated patients. Citation Format: H. Muhammad, S. S. Chavan, C. Feng, D. K. Almaraz, H. Basu, W. Huang, R. Roy, G. Wilding, G. Mills, S. Kummar, S. Krishnamurthy. Transcriptomics-guided AI outperforms and expands on OncotypeDX to predict Hormonal therapy response and Chemotherapy benefit from H&E-stained Whole Slide images in ER+/HER2- 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 PS3-06-24.