An AI-integrated patient-derived organoid platform to enable high-throughput drug response prediction in glioma.

Abstract

e14086 Background: Glioma exhibits pronounced heterogeneity in both the tumor microenvironment and molecular subtypes, leading to substantial inter-individual variability in drug response and posing major challenges for therapeutic prediction and clinical decision-making. However, current in vitro glioma models remain insufficient for rapid, cost-effective, and biologically faithful high-throughput drug screening. To address these limitations, we integrated patient-derived glioma organoids with artificial intelligence–based modeling to establish an improved framework for treatment response evaluation. Methods: An AI agent was used to systematically curate transcriptomic and pharmacological data from publicly available literature and databases. We developed AI4Med, an integrative translational framework that combines patient-derived glioma organoids with AI-based predictive modeling. Multiple complementary algorithms—including gradient boosting methods (LightGBM, XGBoost, CatBoost), tree-based ensembles (Random Forest, Extra Trees), K-Nearest Neighbors, and Neural Networks—were integrated to capture diverse transcriptomic patterns. For each drug, an independent regression model was trained to predict IC50 values. Feature selection was performed by ranking genes according to predictive importance, retaining the top 100 genes per drug to reduce dimensionality and improve generalization. In parallel, a high-throughput organoid-based drug screening system was established and validated against matched native tumors for molecular fidelity and drug response consistency. Results: Model training was conducted using curated transcriptomic and pharmacological data comprising 1,406 cancer cell lines, 481 chemical compounds, and 860 cancer cell line models. Transcriptomic features were represented at the pathway level, with drug-specific pathway weighting applied during model optimization. Using an ensemble-based strategy, AI4Med demonstrated robust predictive performance, which was further refined and independently validated using patient-derived glioma organoid drug screening assays. In organoid-based validation, the model achieved a top-5 drug sensitivity prediction accuracy of 97% and a top-10 accuracy of 85%, enabling early identification of patient-specific therapeutic vulnerabilities. Conclusions: We established an AI-integrated, high-throughput drug screening platform for glioma, supported by real-world experimental validation using patient-derived organoids. This biologically grounded and scalable framework enables precision drug selection in glioma and may be extended to other heterogeneous malignancies, highlighting its broad translational potential.