330P Open the black box: Transparent prognostic prediction and discovery with AI de novo designed spatial biomarkers

Despite the impressive prognostic performance of end-to-end deep learning (DL) in computational pathology, most DL models remain black boxes. The vast visual complexity of whole-slide images (WSIs) and the high dimensionality of model parameters obscure the underlying predictive features. This lack of transparency limits biological insight, clinical trust, and hypothesis-driven discovery.