Interpretable machine learning and graph attention network based model for predicting PAMPA permeability.

Area of Science:

• Computational chemistry and cheminformatics
• Drug discovery and development
• Pharmacokinetics and drug metabolism

Background:

• Parallel Artificial Membrane Permeability Assay (PAMPA) is a high-throughput method for assessing drug permeability.
• Accurate prediction of PAMPA permeability is crucial for efficient early-stage drug discovery.
• Existing methods require optimization for broader chemical space coverage.

Purpose of the Study:

• To develop and evaluate machine learning (ML) and deep learning (DL) models for predicting PAMPA permeability.
• To compare the predictive performance of Random Forest (RF), Explainable Boosting Machine (EBM), Adaboost, and Graph Attention Network (GAT) models.
• To assess model generalizability using both internal and external datasets.

Main Methods:

• Utilized a curated dataset of 5447 compounds with PAMPA permeability scores.
• Trained and validated RF, EBM, Adaboost, and GAT models.
• Evaluated model performance on an independent external dataset.

Main Results:

• During internal validation, RF achieved 81% accuracy and EBM achieved 80% accuracy.
• Adaboost and GAT showed accuracies of 76% and 74%, respectively.
• On the external dataset, RF, EBM, and Adaboost achieved accuracies of 91%, 90%, and 89%, while GAT reached 86%.

Conclusions:

• All developed models demonstrated robust prediction of PAMPA permeability, outperforming benchmarks.
• RF and EBM models showed particularly strong performance across both internal and external validation datasets.
• The study provides valuable computational tools for predicting drug permeability, supporting efficient drug discovery pipelines.