MetaCYP: a unified framework for prediction of cytochrome P450 metabolic sites and reaction types via multimodal deep learning

Introduction Cytochrome P450 (CYP) enzymes are the predominant drug-metabolizing proteins in humans, collectively governing the structural transformation of drugs and xenobiotics while directly shaping their pharmacological activity and toxicological profiles. Accurate prediction of CYP–substrate reaction sites and reaction types is therefore central to drug discovery and metabolic risk assessment. Existing computational models, however, largely depend on intrinsic molecular properties or hardcoded reaction rules, constraining their generalization across CYP isoforms. Methods Here we present MetaCYP, a multimodal deep learning framework that predicts bonds of metabolism (BoMs) and reaction types in CYP-mediated biotransformation. MetaCYP encodes CYP amino acid sequences with the protein language model ESM-2 and extracts bond-level substrate features using Uni-Mol, integrating both modalities through an attention-based cross-modal fusion mechanism that captures enzyme–substrate interactions. Results This architecture enables a single unified model to resolve isoform-specific catalytic selectivity for identical substrates. MetaCYP achieves state-of-the-art performance in BoM prediction (MCC: 0.741; ROC-AUC: 0.956) and reaction type prediction (MCC: 0.796; ROC-AUC: 0.946), outperforming current benchmarks. Discussion MetaCYP establishes a unified deep learning framework that models reaction sites and reaction types from enzyme–substrate information, offering a mechanistically grounded and interpretable tool for elucidating CYP catalytic selectivity. Its capacity to improve the accuracy of ADME property predictions, alongside its scalability across isoforms, positions it as a practical resource for early-stage drug screening, metabolic risk assessment, and rational drug design.