Chemical reaction prediction is a vital application of artificial intelligence. While Transformer models are widely used for this task, they often overlook deeper-level semantic information. In addition, the traditional Transformer model suffers from a decline in prediction performance and shows poor generalization when faced with different representations of the same molecule. To address these challenges, we propose a dual encoder-based reaction prediction method tailored for multilevel organic chemistry. Our approach began with the introduction of synergistic dual-encoder architecture: The atomic encoder focused on inter-atomic attention weights. In contrast, the molecular encoder employed a molecular maximum dimension reduction algorithm to identify key chemical features. We then performed multilevel feature fusion by combining the outputs from both the atomic and molecular encoders. Finally, we applied an optimized contrast loss to enhance the model’s robustness. The results indicated that this method outperformed existing models across all four datasets, significantly improving generalization performance and contributing to advancements in artificial intelligence-driven drug development and research.