Abstract 6728: Predicting gastric cancer response to anti-PD-1 first-line treatment based on multi-modal data: A multi-center study.

Abstract Objectives: Excepted for PD-L1 expression, no predictive maker for immunotherapy has been identified in gastric cancer. Our study aimed to develop a multimodal deep learning model by integrating clinical features, CT-derived radiomics, and H&E-based pathomics to accurately predict the efficacy of first-line immunotherapy in gastric cancer. Methods: This multicenter diagnostic study retrospectively enrolled 477 gastric cancer patients from 5 centers. Whole-slide images (WSIs) underwent a weakly supervised computational pathology pipeline: CLAM extracted patch-level representations, which were organized into spatially coherent structures for histological pattern aggregation; handcrafted pathomics features (morphological, textural, color) were also extracted to complement deep learning pathology embeddings. Contrast-enhanced CT images were segmented via a nnU-Net cascade. The nnU-Net encoder generated deep radiological embeddings, while handcrafted radiomics features (first-order statistics, texture matrices, shape descriptors) were computed from segmented regions to quantify tumor heterogeneity. A Transformer-based architecture integrated the above pathology/imaging features and patient metastatic information via cross-modal self-attention, enabling synergistic multimodal learning. Results: The dataset was allocated into training (n = 322, Center 1) and validation cohort (n = 115, Center 2-5). The integrated clinical-radio-pathomic (CRP) model demonstrated strong performance (Validation AUC = 0.857, Test set AUC=0.917, C-index = 0.746 for PFS prediction and C-index = 0.864 for OS prediction) and emerged as an independent prognostic biomarker for PFS (hazard ratio [HR] = 4.29, 95% confidence interval [CI],1.20-15.35, p = 0.025) and OS (HR = 8.39, 95% CI, 1.47-47.75, p = 0.017) controlling for other clinical variables (i.e., age, sex and PD-L1 CPS). Importantly, the CRP model outperformed unimodal models derived from radiology (Test AUC = 0.7583,Validation AUC = 0.759) or pathology (Test AUC = 0.83;Validation AUC = 0.796) alone, as well as PD-L1 CPS. Conclusion: Our findings highlight the promise of multimodal deep learning in precisely identifying patients who are most likely to benefit from first-line immunotherapy. Citation Format: Jingyuan Wang, Peng Kuang, Yufu Lin, Yutong Liu, Kaiyue Zheng, Yu Jiang, Guangjun Yu, Ying Ding, Yingyong Hou, Tianshu Liu. Predicting gastric cancer response to anti-PD-1 first-line treatment based on multi-modal data: A multi-center study [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 6728.