Objective Drug-drug interaction DDI prediction can assist in determining effective and safe combinations of drugs for treatment. However previous prediction methods typically model drugs based on their two-dimensional molecular structures without considering the three-dimensional geometric structure of drugs. Moreover they overlook the diversity of drug functional substructures and a single feature-fitting path may limit the predictive ability of the model. Methods Therefore a DDI prediction model based on a graph isomorphic network and attention mechanism was proposed. The model constructed a molecular graph feature generation module using the graph isomorphic network to generate drug molecular graph features in a three-dimensional coordinate system. On this basis a self-attention mechanism was used to build a substructure feature extraction module to extract spatial substructures of drugs from multiple dimensions. Finally a DDI triplet prediction module was used to identify potential DDIs. The entire model adopted a dual-layer mutually independent network to extract different acting substructures. Results Experimental results demonstrated that the model achieved significant advantages over previous methods on the DrugBank dataset particularly in predicting interactions between new drugs. It achieved improvements of 7. 59% in accuracy ACC 9. 45% in area under the receiver operating characteristic curve AUC and 12. 95% in F1 score compared with the best baseline demonstrating excellent generalization performance. Conclusion Ablation experiments prove that the substructure feature extraction module can effectively extract spatial substructures and the dual-layer network makes the model obtain better prediction performance.