Cluster-Validated Graph Neural Network for P-gp Substrate Prediction from Public Data

P-glycoprotein (P-gp, MDR1/ABCB1) is a key efflux transporter that limits oral absorption and brain penetration and contributes to clinically relevant drug–drug interactions, yet experimental identification of P-gp substrates is labor-intensive, assay-dependent, and difficult to scale, and many existing in silico models rely on relatively small and heterogeneous data sets. We developed a graph neural network (GNN)-based classifier for P-gp substrate prediction by integrating large public quantitative high-throughput screening cytotoxicity assays in KB-3–1 and KB-8–5–11 cell lines (PubChem AID 1346986/1346987) with KEGG BRITE and FDA exemplar substrates, yielding 7,321 RDKit-derived molecular graphs (1,524 substrates, 20.8%) and training a hybrid TransformerConv/NNConv ensemble under stratified, scaffold-based, and advanced Butina leave-cluster-out cross-validation schemes. Across stringent cluster-based folds, the final ensemble (MDR1-M4-HYB-v1) achieved test ROC-AUC values of 0.82–0.88, PR-AUC values of 0.57–0.64, and F1 scores of 0.50–0.60, while on the full internal data set, ROC-AUC and PR-AUC were 0.987 and 0.945, respectively, and on an independent external set (n = 1,475; 1.5% substrates), the model maintained ROC-AUC 0.899 and PR-AUC 0.280 with strong top-ranked enrichment of true substrates. SHAP and surrogate analyses highlighted aromatic ring count, fraction of sp3 carbons, lipophilicity, and hydrogen-bonding capacity as major global determinants, indicating that this public data–driven GNN ensemble provides a high-performing and interpretable in silico tool for P-gp substrate prediction that may support early ADME risk assessment, prioritization of nonsubstrate chemotypes, and transporter-aware drug design.