Improved and Interpretable\nPrediction of Cytochrome\nP450-Mediated Metabolism by Molecule-Level Graph Modeling and Subgraph\nInformation Bottlenecks

Accurately\nidentifying sites of metabolism (SoM) mediated by cytochrome\nP450 (CYP) enzymes, which are responsible for drug metabolism in the\nbody, is critical in the early stage of drug discovery and development.\nCurrent computational methods for CYP-mediated SoM prediction face\nseveral challenges, including limitations to traditional machine learning\nmodels at the atomic level, heavy reliance on complex feature engineering,\nand the lack of interpretability relevant to medicinal chemistry.\nHere, we propose GraphCySoM, a novel molecule-level modeling approach\nbased on graph neural networks, utilizing lightweight features and\ninterpretable annotations on substructures, to effectively and interpretably\npredict CYP-mediated SoM. Unlike computationally expensive atomic\ndescriptors derived from resource-intensive chemistry or even quantum\nchemistry calculations, we emphasize that graph-based molecular modeling\ninitialized solely with lightweight features enables the adaptive\nlearning of molecular topology through message-passing mechanisms\ncombined with various aggregation kernels. Extensive ablation experiments\ndemonstrate that GraphCySoM significantly outperforms baseline models\nand achieves superior performance compared with competing methods\nwhile exhibiting advantages in computational efficiency. Moreover,\nthe attention mechanism and subgraph information bottlenecks are incorporated\nto analyze node importance and feature significance, resulting in\nmining substructures associated with the SoM. To the best of our knowledge,\nthis is the first comprehensive study of CYP-mediated SoM using molecule-level\nmodeling and interpretable technology. Our method achieves new state-of-the-art\nperformance and provides potential insights into the molecular and\npharmacological mechanisms underlying drug metabolism catalyzed by\nCYP enzymes. All source files and trained models are freely available\nat https://github.com/liyigerry/GraphCySoM.