Accurate prediction of Sites of Metabolism (SoMs) is vital for drug design and optimization, as it directly impacts both efficacy and safety. In this study, we introduce CYP-MAP, a pioneering multi-level Graph Neural Network (GNN) model for SoM prediction with a focus on Cytochrome P450 (CYP) enzymes, which play a crucial role in drug metabolism. We have developed the most comprehensive CYP-mediated SoM database to date, integrating existing datasets such as EBoMD and DrugBank with extensive additional data points gathered through rigorous literature searches. Leveraging this unique dataset, CYP-MAP incorporates atom-level, bond-level, and whole-molecule representations to capture complex CYP-mediated metabolic reaction points and metabolites. The model not only predicts SoMs but also forecasts the reaction types, offering a comprehensive view of potential metabolites. Compared to existing methods, CYP-MAP demonstrates substantial improvements in prediction accuracy, with an Area Under the Curve (AUC) of 0.91 This advancement, coupled with our extensive CYP-mediated SoM database, offers invaluable resources for accelerating the discovery of safer and more effective pharmaceuticals.