Molecular Property Prediction Based on Improved Graph Transformer Network and Multitask Joint Learning Strategy

Molecular property prediction is of great significance in drug design and materials science. However, due to the complexity and diversity of molecular structures, existing methods often struggle to simultaneously capture both the local chemical environments and the global structural characteristics of molecules, and they lack generalization ability when dealing with multiple data sets. To address these challenges, this paper proposes a molecular property prediction approach based on an improved Graph Transformer network combined with a multitask joint learning strategy. Specifically, we enhance the attention mechanism by integrating atomic relative position encoding and bond information encoding, thereby explicitly incorporating spatial structure and chemical bond features into the model. Meanwhile, we construct a hierarchical feature extraction architecture by alternately stacking local message-passing layers and global attention layers, and we adopt a mixture-of-experts mechanism to achieve collaborative representation of both local molecular features and global structure. In addition, we design a multitask joint learning strategy that leverages alternating training on multiple tasks and dynamic weighting adjustments to significantly improve the model's generalization performance across diverse data sources. Experimental results show that our method achieves higher prediction accuracy on multiple classification and regression data sets, with an average improvement of 6.4% and 16.7% over baseline methods. Compared with single-data set training, our multitask joint learning strategy further boosts the prediction accuracy by an average of 2.8% and 6.2%. These findings indicate that the proposed approach is highly effective in predicting a wide range of molecular properties.