An interpretable deep learning biomarker for prognostication and prediction of adjuvant chemotherapy benefit in gastric cancer

Prognostic stratification in gastric cancer (GC) currently relies on the tumour-node-metastasis (TNM) staging system, which incompletely captures tumour heterogeneity. Routine haematoxylin and eosin (H&E)-stained whole-slide images (WSIs) contain additional prognostic information that is not routinely quantified. We developed an interpretable deep learning framework using a weakly supervised Transformer to derive a pathological risk score (TPRS) from WSIs for overall survival (OS) stratification and adjuvant chemotherapy benefit prediction. TPRS was developed on HMU-GC (n = 2876) and validated internally (n = 288) and on TCGA-STAD (n = 355). It achieved a mean 10-fold cross-validation C-index of 0.765 ± 0.003 internally and 0.621 ± 0.005 externally, and was an independent prognostic factor. Stage III patients with high TPRS showed significant survival benefit from adjuvant chemotherapy. Mediation analysis of differentially expressed genes (DEGs) and cellular features in high-attention patches supported a 'Gene → Cellular Features → TPRS' relationship, linking transcriptomics to cellular features and TPRS.