Decoding Genomic Risk from Tumor Morphology

Genomic assays such as Oncotype DX play a central role in breast cancer risk stratification and treatment decision-making but remain costly and variably accessible. In contrast, hematoxylin and eosin (H&E) whole-slide images are routinely available and capture morphologic patterns that reflect underlying tumor biology. This talk explores the extent to which genomic recurrence risk is encoded in tumor morphology and can be inferred using artificial intelligence. I will present AI models trained on large, multi-institutional cohorts to predict Oncotype DX recurrence risk directly from whole-slide histopathology. While these models achieve strong average predictive performance, we identify a critical limitation: substantial prediction variability driven by intra-tumoral heterogeneity and slide selection. Analyses of patients with multiple tumor slides show that predictions can vary widely depending on which slide is used, even when overall calibration appears robust. These findings underscore both the promise and the constraints of histopathology-based genomic surrogates and motivate multi-slide, uncertainty-aware approaches for reliable clinical deployment.