GTransCYPs: an improved graph transformer neural network with attention pooling for reliably predicting CYP450...

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Predicting Cytochrome P450 (CYP450) enzyme inhibitors is crucial for early-stage drug discovery.
• Current machine learning (ML) methods for in silico prediction of CYP450 inhibitors face challenges in accuracy.

Purpose of the Study:

• To introduce GTransCYPs, an enhanced graph neural network (GNN) model with a transformer mechanism.
• To improve the prediction accuracy of CYP450 inhibitors, particularly for major isozymes (1A2, 2C9, 2C19, 2D6, and 3A4).

Main Methods:

• Developed GTransCYPs, a deep learning architecture integrating GNNs with a transformer mechanism and attention pooling.
• Utilized transformer convolution layers for feature processing and global attention-pooling for graph-level information synthesis.
• Tested four GTransCYPs variations with different pooling techniques on CYP450 prediction tasks.

Main Results:

• GTransCYPs demonstrated superior performance in discriminating CYP450 inhibitors from non-inhibitors compared to existing state-of-the-art methods.
• The graph transformer with attention pooling algorithm achieved the best predictive performance.
• Experimental and ablation studies confirmed the model's efficacy.

Conclusions:

• GTransCYPs offers a cost-effective and highly efficient computational approach for predicting CYP450 inhibition.
• The proposed deep learning model significantly advances the accuracy of in silico CYP450 inhibitor prediction.
• The source code for GTransCYPs is publicly available for further research and application.