Enhancing CYP3A4 Inhibition Prediction Using a Hybrid GNN-ML Model with Data Augmentation

Area of Science:

• Computational Chemistry
• Pharmacology
• Artificial Intelligence in Drug Discovery

Background:

• Cytochrome P450 3A4 (CYP3A4) is a key enzyme in drug metabolism, affecting 30-50% of marketed drugs.
• Accurate prediction of CYP3A4 inhibition is vital for early identification of potential drug-drug interaction (DDI) risks and toxicity.
• Existing prediction methods require enhancement for improved accuracy and interpretability in drug development pipelines.

Purpose of the Study:

• To evaluate an integrated artificial intelligence (AI) framework for predicting CYP3A4 inhibition percentage.
• To assess the performance of various machine learning (ML) and graph neural network (GNN) models.
• To establish a robust and interpretable framework for metabolic screening and DDI risk assessment.

Main Methods:

• Compiled a large dataset of 23,713 compounds from diverse chemical and public databases.
• Evaluated vector-based ML models (LightGBM, XGBoost, CatBoost, weighted ensemble) and GNN models (O-GNN + CL + Mixup, D-MPNN, GINE, GATv2).
• Employed manifold mixup for GNN training and SMILES enumeration-based test-time augmentation for inference, integrating best models via a weighted ensemble.

Main Results:

• The weighted ML ensemble achieved the highest performance among ML models (RMSE=19.1031, PCC=0.7566).
• The O-GNN + CL + Mixup model showed the best performance among GNNs (RMSE=20.1002, PCC=0.7265).
• The integrated hybrid model demonstrated superior predictive accuracy (RMSE=19.0784, PCC=0.7570) and generalizability upon external validation.

Conclusions:

• Integrating ML and GNN models with data augmentation strategies significantly enhances the robustness and interpretability of CYP3A4 inhibition predictions.
• The developed AI framework provides a practical and effective tool for early-stage metabolic screening.
• This approach facilitates proactive assessment of DDI risks, contributing to safer drug development.