Artificial intelligence-based prediction of claudin 18.2 expression and immune phenotype to guide treatment decisions in patients with gastric cancer.

Abstract

4047 Background: With the increasing availability of immune checkpoint inhibitor (ICI)- and targeted agent-based treatments for gastric cancer, evaluating predictive biomarkers to guide first-line treatment has become increasingly complex. Claudin 18.2 (CLDN18.2) is a clinically relevant target for zolbetuximab-based chemotherapy, typically assessed by immunohistochemistry. However, limitations such as inadequate tumor specimens, high costs, and long turnaround times pose a practical challenge. Methods: An artificial intelligence (AI) model was developed to predict CLDN18.2 expression based on hematoxylin and eosin (H&E) slides from 459 patients with gastric cancer (derivation cohort). CLDN18.2 positivity was defined as moderate-to-strong expression in ≥75% of tumor cells. The AI model utilized a Vision Transformer-based architecture with a multiple-instance learning framework and was trained using five-fold cross-validation. Model performance was validated in two independent cohorts: an internal cohort of 381 patients treated with first-line ICI plus chemotherapy (ICI-Chemo) or chemotherapy alone (Chemo-only) and an external cohort of 100 patients from diverse ethnic backgrounds. Immune phenotypes (IPs) were assessed using an AI-powered whole slide image analyzer to further stratify patient outcomes. Results: The prevalence of CLDN18.2 positivity was 43.4% (derivation), 37.3% (internal validation), and 26.0% (external validation). The model achieved an AUROC of 0.753 in the derivation cohort, with sensitivity and specificity of 0.638 and 0.723, respectively. In the internal and external validation cohorts, AUROCs were 0.752 and 0.746, with similar sensitivity and specificity levels. Among the internal validation cohort, patients were stratified into subgroups based on predicted CLDN18.2 positivity and IP status. Among these, the subgroup predicted to be CLDN18.2-negative and inflamed IP demonstrated the most significant benefit from ICI-Chemo compared to the Chemo-only group. Conclusions: The AI model reliably predicted CLDN18.2 expression from H&E slides and exhibited reliable performance. The differential survival outcomes observed in subgroups stratified by AI-predicted CLDN18.2 expression and IP suggest its potential utility in guiding first-line treatment decisions for gastric cancer patients. Hazard ratio (HR) for progression-free survival (PFS) and overall survival (OS). Groups HR for PFS (95% CI) HR for OS (95% CI) CLND18.2-negative and inflamed IP 0.37 (0.17–0.78, p=0.009) 0.41 (0.19–0.88, p=0.021) CLDN18.2-negative and non-inflamed IP 0.75 (0.54–1.05, p=0.095) 0.81 (0.59–1.13, p=0.224) CLDN18.2-positive and inflamed IP 0.67 (0.40–1.12, p=0.128) 0.62 (0.37–1.04, p=0.073) CLDN18.2-positive and non-inflamed IP 1.60 (0.74–3.46, p=0.229) 1.23 (0.59–2.58, p=0.579) reference: Chemo-only.