Abstract 2473: Generative AI improves breast cancer genomic subtype prediction from histology

Abstract Despite advancements in prognostic biomarker identification in cancer, the clinical infeasibility of testing for many biomarkers limits their adoption into clinical practice. Computational pathology has demonstrated remarkable success in efficiently predicting a variety of biomarkers from routine histology images that would otherwise require expensive sequencing or spatial profiling. However, current approaches are limited for data-poor and challenging tasks such as the prediction of genomic subtype in breast cancer. Here, we show that generative AI improves state-of-the-art computational pathology methods by leveraging the power of synthetic data in the task of predicting breast cancer genomic subtype from histology images. Incorporating generative AI to enrich the training dataset improves model AUC, overall accuracy, and accuracy on the underrepresented high risk subtype. Finally, we show that model interpretation using explainable AI techniques on the tile and pixel levels corroborates expected phenotypes as well as identifies previously unknown histologic features associated with genomic subtype. Our results present an improved framework for computational pathology for the enhanced clinical detection and translation of biomarkers. Citation Format: Brennan G. Simon, Clemens L. Weiss, Christina Curtis. Generative AI improves breast cancer genomic subtype prediction from histology [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 2473.