Interpretable graph learning of spatial TME patterns in digital pathology for gastric cancer prognosis

Current gastric cancer (GC) risk stratification, based on TNM staging and molecular subtypes, often fails to capture intratumoral heterogeneity driving variable outcomes. Although the tumor microenvironment (TME) spatial architecture within histopathological images holds critical prognostic value, quantifying these interactions remains challenging. We present the Graph-based Tumor Microenvironment Aggregation (GTMA) framework, an interpretable deep learning approach modeling whole-slide images as spatial graphs to decipher tissue interactions. Validated in a large cohort ( n = 1871) and TCGA-STAD dataset ( n = 213), the GTMA-derived Risk Score (RS) provided prognostic information complementary to TNM staging and clinical models. Notably, the RS stratified patients into distinct risk groups within identical TNM stages and molecular subtypes, identifying high-risk patients traditionally overlooked. Integrative analysis revealed high-RS tumors feature a fibroblast-dominant, immune-excluded phenotype and distinct stromal spatial patterns. By linking digital pathology to biological TME patterns, GTMA offers a robust, interpretable AI tool for individualized prognostic risk stratification and precision oncology in GC.