CypEGAT: A Deep Learning Framework Integrating Protein Language Model and Graph Attention Networks for Enhanced CYP450s Substrate Prediction

Human Cytochrome P450 enzymes (CYP450s) are respon- sible for metabolizing 70–80% of clinically used drugs. The develop- ment of computational tools to accurately predict CYP450 enzyme- substrate interactions is crucial for drug discovery and chemical tox- icology studies. In this work, we introduce CypEGAT, a deep learn- ing framework designed to enhance prediction performance by integrat- ing protein embeddings of CYP450s (extracted using the pre-trained ESM-2 Transformer model) with molecular embeddings generated by our fine-tuned Graph Attention Network (GAT). The CypEGAT model was trained end-to-end on two large-scale experimental enzyme-substrate datasets and our CYP450s dataset, which comprises 51,753 CYP450 enzyme-substrate pairs and 27,857 enzyme-nonsubstrate pairs. Focusing on five major human CYP450 isoforms (CYP1A2, CYP2C9, CYP2C19, CYP2D6, and CYP3A4), CypEGAT achieves an overall predictive accu- racy of 0.882 and an AUROC of 0.928. The model demonstrates robust generalizability to novel chemical compounds across different CYP450 isoforms, underscoring its potential as a powerful tool for drug metabolism studies.