Generative AI Enables Breast Cancer Genomic Subtype Prediction from Histology Images

Abstract

Breast cancer subtyping is essential for precision oncology, influencing prognosis, treatment selection, and clinical trial design. The Integrative Subtype Classification (IC) categorizes breast tumors into groups with distinct long-term outcomes based on genomic and correlated transcriptomic features. This method relies on sequencing data, which, despite decreasing costs, is not always available in research or clinical settings. Here we introduce PATH-IC, a computational pathology model that predicts ER+ breast cancer IC subtype risk of relapse categories from routine histology data. We enhance the current state-of-the-art computational pathology approach with BERGERON, which leverages generative AI to correct class imbalance and reduce overfitting, showing that synthetic data improves PATH-ICs performance by the equivalent of 41% more real training samples. PATH-IC achieves a testing AUROC of 0.814, with predictions correlating to Oncotype DX scores and long-term relapse risk. Using attention-based model interpretation and CRAWFORD, a novel embedding-to-image foundation model, we demonstrate that PATH-IC identifies expected tumor microenvironment patterns for IC subtypes and highlights heterochromatin condensation as a key feature of high-risk tumors. Matched single-cell spatial transcriptomics confirm IC subtype-specific gene expression patterns identified by PATH-IC, including active metabolic, proliferative, and proteostasis pathways in the high-risk group. PATH-IC advances computational pathology through generative AI, enabling subtype inference from histopathology data.