AI-predicted spatial transcriptomics unlocks breast cancer biomarkers from pathology.

Spatial transcriptomics (ST) assays are transforming our understanding of tumor heterogeneity, but their high cost limits their application in large-scale biomarker discovery. Here, we present "Path2Space," a deep-learning model that predicts spatial gene expression directly from histopathology slides. Trained on extensive breast cancer ST data, Path2Space robustly predicts the spatial expression of thousands of genes, outperforming 21 established methods. Charting the tumor microenvironment (TME) of 976 breast cancer TCGA (The Cancer Genome Atlas) tumors, it accurately infers cell-type abundances and identifies three spatially defined breast cancer subgroups with distinct survival outcomes. Notably, the derived low-cost spatial TME landscapes enable more accurate predictions of patient response to chemotherapy and trastuzumab compared with costly conventional bulk-sequencing-based biomarkers. Path2Space thus offers a scalable, fast, and cost-effective alternative to molecular assays. It opens avenues for large cohort treatment biomarker discovery and translationally relevant insights into tumor biology, with potential applicability across many cancer indications.