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

Area of Science:

• Computational chemistry
• Cheminformatics
• Materials science

Background:

• Molecular property prediction is crucial for drug design and materials science.
• Existing methods struggle to capture both local and global molecular features, limiting generalization.
• Challenges include handling complex molecular structures and diverse datasets.

Purpose of the Study:

• To develop a novel molecular property prediction approach.
• To improve the accuracy and generalization ability of predictive models.
• To address limitations of existing methods in capturing molecular complexity.

Main Methods:

• An improved Graph Transformer network incorporating atomic relative position and bond information encoding.
• A hierarchical feature extraction architecture combining local message-passing and global attention layers.
• A mixture-of-experts mechanism for collaborative local and global feature representation.
• A multitask joint learning strategy with alternating training and dynamic weighting.

Main Results:

• The proposed method achieved higher prediction accuracy on multiple classification and regression datasets, outperforming baseline methods by 6.4% and 16.7% on average.
• The multitask joint learning strategy improved prediction accuracy by an average of 2.8% and 6.2% compared to single-dataset training.
• Demonstrated significant improvements in generalization performance across diverse data sources.

Conclusions:

• The enhanced Graph Transformer network effectively predicts molecular properties by integrating spatial and chemical bond information.
• The multitask joint learning strategy significantly enhances model generalization across various datasets.
• The proposed approach offers a robust and effective solution for molecular property prediction in drug design and materials science.