Drug-drug interactions (DDIs) can lead to unexpected pharmacological effects, underscoring their importance in identifying active sites of drug-metabolizing enzymes for drug discovery and clinical applications. While many deep learning models perform well in DDI prediction, their interpretability in revealing molecular mechanisms related to enzyme inhibition or activation remains limited. Here, we present DrugDAGT, a dual-attention graph transformer framework with contrastive learning, designed to predict various types of metabolic DDIs. By incorporating attention mechanisms at both the bond and atomic levels, the model captures short- and long-range dependencies, allowing it to identify key structures essential for DDI metabolism sites. Furthermore, DrugDAGT applies graph contrastive learning to enhance representation similarity across views, improving molecular structure discrimination. DrugDAGT achieved competitive performance compared to state-of-the-art models. Attention map visualizations reveal key substructures in enzyme-mediated DDIs and SARS-CoV-2 drug combinations, providing interpretable insights instead of black-box results. All data and code of our DrugDAGT can be found at https://github.com/codejiajia/DrugDAGT.