Digital Pathology Predicts Immunotherapy Response and Maps Spatial Biomarkers in Gastric Cancer

Background Biomarkers have limited reliability, making predictions of response to immune checkpoint inhibitors (ICIs) challenging in gastric cancer. We developed and validated an artificial intelligence (AI)-based model using whole slide images (WSIs) to predict ICI response and assess its prognostic and biological relevance. Methods In this multicenter retrospective study, we developed an AI model using a development cohort (internal) consisting of 272 slides from 107 patients and validated it externally on two independent cohorts comprising 373 slides from 127 patients and 25 slides from 18 patients, respectively. Three histopathology foundation models and eight multiple-instance learning algorithms were compared to optimize prediction performance. Spatial transcriptomics (Xenium 5K) data from 14 ICI-treated patients were used to identify spatial biomarkers by differential expression analysis. Gene ontology analysis was performed on transcriptomic data from TCGA-STAD to characterize the genes associated with the predictions of the AI model. Results The combination of the UNI foundation model with the CLAM algorithm achieved the best performance for the internal test cohort (AUROC = 0.844). High-attention patches revealed signet ring cells and fibrosis in non-responders versus nuclear hyperchromasia and enlarged nuclei in responders. The external validation revealed an AUROC of 0.7151 and 0.7545, respectively. Survival analysis revealed a significant association between the internal test cohort and progression-free and overall survival ( p  < 0.001), which was also observed in the external cohort ( p  < 0.05). Genes associated with AI predictions in TCGA showed immune activation in responders and increased cell cycle activity in non-responders. Spatial transcriptomics confirmed multi-compartmental FGFR2 overexpression in non-responders and strong immune pathway enrichment in responders, particularly in densely infiltrated tumor regions. Conclusion Our results indicate that AI-powered histopathologic slide analysis represents a robust biomarker for predicting immunotherapy response and patient outcomes in gastric cancer. Integration with spatial and bulk transcriptomic data confirms its biological validity, thus supporting AI-guided pathology as a scalable approach for informed immunotherapy in gastric cancer.