AI-Guided Spatial Proteomics Reveals Breast Cancer Recurrence Niches.

This study presents an innovative outcome-informed spatial pathology framework for triple-negative breast cancer (TNBC). It combines deep learning models, which predict cancer recurrence from standard H&E stained slides, with mass spectrometry-based spatial proteomics. The goal is to uncover the localized molecular states that underpin these AI predictions, which are typically opaque. The framework generates recurrence risk heatmaps from H&E slides, achieving strong predictive performance in an independent test cohort. These heatmaps then serve as precise guides to physically isolate and profile specific AI-defined tumor regions. Spatial proteomic analysis of these regions revealed distinct molecular contrasts: high-risk areas were enriched with mitotic programs, while low-risk areas showed enrichment in immune and antigen presentation programs. This integration of AI-generated morphological insights with localized molecular profiling defines a new role for AI models as experimental guides. It facilitates the discovery of biologically grounded, multiscale biomarkers, improving the prediction of recurrence-free survival when combined with the H&E-derived risk score.