e20024 Background: Neoadjuvant immunotherapy has significantly improved outcomes in patients with resectable NSCLC; however, substantial heterogeneity in therapeutic response remains, and robust predictive biomarkers are lacking. Spatial pathology enables high-resolution characterization of the tumor immune microenvironment beyond conventional cell density metrics. Integrating spatial cellular organization with graph-based learning may provide a scalable and generalizable strategy for predicting response to neoadjuvant immunotherapy. Methods: This multicenter retrospective cohort study included NSCLC patients who received neoadjuvant immunotherapy after diagnostic biopsy between 2019 and 2025 across four institutions(GDPH,SYSUCC, GYFYY,FAHZZU). Pretreatment biopsy H&E whole-slide images were analyzed using a spatial pathology pipeline incorporating cell instance segmentation, spatial organization modeling, and graph neural network learning. The model was developed to predict major pathological response (MPR) and was evaluated in independent external validation cohorts. Results: A total of 1,002 NSCLC patients receiving neoadjuvant immunotherapy from four centers were included, all with pretreatment biopsy H&E whole-slide images available. The spatial pathology–based graph neural network model showed stable performance across multicenter cohorts. In independent external validation, the model achieved an area under the receiver operating characteristic curve (AUC) of 0.72 for predicting major pathological response, demonstrating good generalizability across institutions. Conclusions: This spatial pathology-based graph neural network provides a scalable and generalizable approach for predicting response to neoadjuvant immunotherapy in NSCLC using routine biopsy specimens, with potential clinical utility for preoperative risk stratification and treatment decision-making.
Cai et al. (Thu,) studied this question.