From pixels to spatial microenvironments: AI-powered CAF-Epi niche prediction on H&E slides

Routine H&E slides remain the most accessible and rapid diagnostic material in clinical practice, particularly for gastrointestinal cancers. However, although H&E effectively distinguishes cancer from non-cancer, it provides insufficient information to predict treatment response. Spatial niches can be conceptualized as multicellular microanatomical units-defined by characteristic patterns of cellular adjacency, interaction and coordinated signaling. Using spatial transcriptomics across multistage esophageal squamous cell carcinoma (ESCC), we identified a CAF-epithelial (CAF-Epi) niche in which cancer-associated fibroblasts and epithelial cells cooperatively shield tumor cells from immune surveillance. In a neoadjuvant immunotherapy cohort, we found CAF-Epi score derived from this niche robustly discriminated responders from non-responders and stratified long-term survival, indicating that this functional niche, rather than subtle changes in cell proportions or morphology, more directly governs treatment outcome. To enable scalable CAF-Epi assessment on routine histopathology, we developed SHEEN-CAF (Spatial H&E-based CAF-Epi niche model), a deep-learning model that identifies ST-defined CAF-Epi niches directly from H&E slides. We assembled a paired H&E-ST atlas of 198 multistage ESCC specimens, totaling 9.6 million cell annotations and 275,753 CAF-Epi-labeled FOVs. SHEEN-CAF learns multi-scale morphological representations and reconstructs whole-slide CAF-Epi distributions. In the test set, it achieved an AUC of 0.91 with cross-center accuracies of 87-94%. In an independent external validation cohort with paired H&E and multiplex immunofluorescence, SHEEN-CAF-derived CAF-Epi scores correlated with immunofluorescence-defined CAF-Epi signals (r = 0.81, P < 0.01) and discriminated high- versus low-CAF-Epi burden lesions (AUC = 0.89). In an ESCC cohort treated with immune checkpoint blockade, non-responders exhibited significantly higher CAF-Epi scores than responders (P < 0.05), linking CAF-Epi burden on routine H&E to immunotherapy resistance. In summary, SHEEN-CAF converts CAF-Epi niche burden-previously measurable only with advanced spatial-omics-into an automated, quantitative biomarker derived solely from routine diagnostic H&E slides. By requiring no additional tissue or sequencing, it provides an efficient and scalable measure of the spatial microenvironment, with potential to refine immunotherapy response assessment and risk stratification in ESCC.