A robust and interpretable graph neural network-based protocol for predicting p-glycoprotein substrates.

Area of Science:

• Pharmacology
• Computational Chemistry
• Biotechnology

Background:

• P-glycoprotein (P-gp) is an ATP-binding cassette transporter vital for drug absorption and distribution.
• P-gp's presence at the blood-brain barrier necessitates understanding drug interactions for CNS penetration.
• Current P-gp research often emphasizes inhibitors over substrates, highlighting a gap in substrate prediction.

Purpose of the Study:

• To develop a robust computational model for predicting P-glycoprotein (P-gp) substrates.
• To enhance the evaluation of drug candidates' ability to penetrate the central nervous system.
• To identify key molecular substructures associated with P-gp substrate activity.

Main Methods:

• Utilized a graph neural network approach, including graph convolutional networks and AttentiveFP.
• Trained and validated the model on a dataset of 1995 drug molecules (1202 substrates, 793 nonsubstrates).
• Employed integrated gradient analysis to interpret model predictions and identify critical substructures.

Main Results:

• The AttentiveFP model achieved an ROC-AUC of 0.848 and an accuracy of 0.815, outperforming traditional methods.
• Identified 20 key substructures significantly associated with P-gp substrate classification.
• Discovered four substructures that confer a >70% probability of substrate classification, enabling rapid assessment.

Conclusions:

• The developed graph neural network framework provides an effective and interpretable method for predicting P-gp substrates.
• This approach can significantly aid drug development by improving the assessment of CNS penetration.
• The identification of key substructures offers a valuable tool for early-stage drug screening and design.