Interpretable machine learning for predicting immunotherapy response in gastric cancer via serum cytokine profiling and organoid co-culture platforms

Immunotherapy has become a promising strategy for gastric cancer, yet predicting treatment response remains challenging. We developed an exploratory, biologically informed machine learning framework integrating peri-treatment serum biomarkers and clinical variables, with candidate cytokines informed by an in vitro patient-derived organoids (PDOs)–cancer-associated fibroblasts (CAFs)–T cells co-culture platform, to assess pathological response to PD-1 inhibitor (PD-1i)-based therapy in gastric cancer. Data from 125 patients receiving PD-1i were analyzed. Serum samples collected before and after immunotherapy were tested for cytokines. Eighteen successfully established PDOs-CAFs-T cells co-culture models were used for cytokine profiling and to identify seven candidate cytokines. These cytokines were subsequently measured in paired serum samples from the 125-patient clinical cohort and combined with clinicopathological variables for machine learning model development. Tumor Regression Grade (TRG) served as the binary outcome variable, representing patients with good versus poor/no tumor regression. Experimental and clinical data were used to train predictive models with multiple algorithms, including CatBoost and several other machine learning methods. Model performance was assessed by tenfold cross-validation, ROC, precision-recall (PR) Curve, and decision curve analyses. Feature importance was interpreted using Shapley additive explanations (SHAP) and visualized using decision tree. The PDOs-CAFs-T cells co-culture platform showed cytokine-change trends that were generally concordant with those observed in patient serum after treatment. 10 key predictors were identified, including pre-/post-treatment maximum diameter, ycT/ycN stages, and pre- or post-treatment cytokines (IFN-α2, IL-2, IL-6, IL-8, MCP-1). In the internal testing set, the tuned CatBoost model achieved an AUC of 0.845. SHAP analysis revealed post-treatment maximum diameter and cytokine profiles as the most influential features. Decision trees clearly and interpretably present the model. This study presents an exploratory, biologically informed machine learning framework for peri-treatment assessment of pathological response in gastric cancer. The PDOs-CAFs-T cells co-culture platform provided biological context for candidate cytokine selection.