<p>Graph-based molecular representation learning plays a crucial role in retrosynthesis prediction. The Directed Message Passing Neural Network (D-MPNN) has been widely used for encoding molecular structures but mainly focuses on bond features and fails to fully consider the synergy between atom features and bond features. Atom features are the core basis for defining the contextual environment and chemical significance of bonds, and without localized information of nodes, the feature propagation of bonds will lose their due chemical context constraints, making it difficult for models to accurately grasp the local features of molecular structures. To address this limitation, we incorporate Edge-Gated Graph Attention (EGAT) into the D-MPNN framework, leveraging its ability to refine bond features and enhance message passing process. We introduce a reverse bond redundant subtractors (RBRS) to remove the interference of reverse bonds and Atom-Bond Attention-enhanced Directed Message Passing Modules (ABA-dMPMs) to enhance information interaction in graph encoder. Our model effectively integrates atom and bond features, leading to a more expressive molecular representation. Experimental results demonstrate that our method significantly outperforms traditional D-MPNN models in retrosynthesis prediction. In the USPTO-50k benchmark dataset, our model achieves a top-1 accuracy of 55.0% even when the reaction type is unknown, surpassing existing graph-based models in the field.</p>