Abstract PS3-06-22: Artificial Intelligence Predicts OncotypeDX Recurrence Scores Directly from H&E-Stained Whole Slide Images of ER+/HER2- Node-Negative Breast Cancer Surgical Sections

Background: Oncotype DX (ODX) is a widely used genomic assay for risk stratification and guiding adjuvant treatment decision making in early stage ER+/HER2- node-negative breast cancer. However, its cost and turnaround time can significantly limit its use, particularly in low-resource settings. We developed a deep learning model to predict the ODX recurrence score directly from H&E-stained whole-slide images (WSIs) of surgically resected ER+/HER2- node-negative invasive breast tumors and externally validated its performance using retrospective real-world clinical cases. Methods: We developed our model using publicly available datasets consisting of WSIs with associated RNA-Seq data. The ODX score was estimated from the expression levels of individual genes included in the ODX assay. A two-step deep learning framework was utilized to predict the estimated recurrence score from WSIs of H&E stained tissue sections of surgically resected breast tumor. First, our internal Phenotype Atlas, a library of pan-cancer morphological features, was used to identify distinct morphological phenotypes across slides. Second, these phenotypes were mapped to the estimated ODX score using weakly supervised learning to associate ODX scores to morphological features, thus generating a slide-level prediction. For external validation, we used WSIs of an independent cohort of 300 early stage ER+/HER2- node-negative breast cancers with available ODX scores and over 10 years of follow-up. Model performance was assessed using the area under the receiver operating characteristic curve (AUC) for high and low ODX categories as defined by thresholds set in the TAILORx clinical trial (High: ODX>25, Low: ODX≤25 when age is greater than 50). In addition, Kaplan-Meier analysis was used to compare patient outcomes based on treatment decisions guided by either ODX or our model. Results: The median age at diagnosis of the 300 patients in our external validation cohort was 53 years (range: 26–82). The ODX recurrence scores ranged from 0 to 71, with 150 patients having an ODX score below 26. Ninety-five patients (32%) were premenopausal. Our AI model predicted ODX scores directly from H&E WSI alone with an AUC of 0.82. When the AI-predicted ODX score was combined with clinical variables including age, tumor size, tumor grade in a linear regression model, performance improved to an AUC of 0.86. Furthermore, we assessed the performance of the model on premenopausal patients and achieved an AUC of 0.91. No significant difference was found in prediction of patient risk of distant metastasis using Oncotype score in comparison to our AI model (log-rank test p=0.9). Conclusions: Our AI model can estimate ODX recurrence scores directly from H&E WSIs with high accuracy. This model has the potential to serve as a low-cost substitute for ODX with a rapid turnaround time that can be utilized for guiding adjuvant therapy decisions in early stage ER+/HER2- node-negative breast cancer patients. Citation Format: C. Feng, H. Muhammad, S. S. Chavan, D. K. Almaraz, H. Basu, W. Huang, R. Roy, G. Wilding, S. Kummar, S. Krishnamurthy. Artificial Intelligence Predicts OncotypeDX Recurrence Scores Directly from H&E-Stained Whole Slide Images of ER+/HER2- Node-Negative Breast Cancer Surgical Sections [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 PS3-06-22.