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

Area of Science:

• Pharmacology and Toxicology
• Computational Chemistry
• Drug Discovery

Background:

• P-glycoprotein (P-gp) is a crucial efflux transporter affecting drug absorption and distribution.
• Experimental P-gp substrate identification is challenging, time-consuming, and difficult to scale.
• Existing in silico models often suffer from limited and heterogeneous datasets.

Purpose of the Study:

• To develop a robust and scalable in silico model for predicting P-gp substrates.
• To leverage large public datasets and advanced machine learning for improved prediction accuracy.
• To provide an interpretable tool for early ADME risk assessment in drug design.

Main Methods:

• Integrated large public cytotoxicity assay data (PubChem AID 1346986/1346987) with KEGG BRITE and FDA substrates.
• Developed a graph neural network (GNN) classifier using a hybrid TransformerConv/NNConv ensemble.
• Employed rigorous cross-validation schemes including stratified, scaffold-based, and Butina leave-cluster-out methods.

Main Results:

• The GNN ensemble (MDR1-M4-HYB-v1) achieved high performance on internal and external datasets, with test ROC-AUC values up to 0.88.
• The model demonstrated strong predictive power on an independent external set (ROC-AUC 0.899).
• SHAP and surrogate analyses identified key molecular features influencing P-gp substrate prediction, such as lipophilicity and hydrogen-bonding capacity.

Conclusions:

• The public data-driven GNN ensemble offers a high-performing and interpretable in silico tool for P-gp substrate prediction.
• This model can support early ADME risk assessment and transporter-aware drug design.
• The approach facilitates prioritization of drug candidates with reduced P-gp interaction risk.