Enhancing molecular property prediction of transformer models with dual graph representation

Area of Science:

• Computational chemistry and cheminformatics.
• Machine learning applications in scientific discovery.

Background:

• Accurate prediction of molecular properties is crucial for advancing chemistry, materials science, and drug discovery.
• Effective molecular representations that capture topology and structure are essential for machine learning on molecular graphs.

Purpose of the Study:

• To propose the dual graph transformer (DGT), a novel self-attention architecture for comprehensive molecular encodings.
• To improve the accuracy and interpretability of molecular property prediction using machine learning.

Main Methods:

• Developed DGT, a self-attention architecture that jointly models atom and bond graphs.
• Fused atom and bond features, graph topology, structure, and stereogeometric information within the self-attention module.
• Benchmarked DGT on diverse datasets for molecular property prediction.

Main Results:

• DGT significantly outperforms the current state-of-the-art in molecular property prediction.
• Demonstrated performance contributions from dual graph representation, positional/structural encodings, and stereogeometric information.
• Showcased DGT's interpretability at the molecular structural level.

Conclusions:

• DGT provides a powerful and interpretable approach for molecular machine learning.
• The dual graph representation and integrated features enhance prediction accuracy.
• DGT is poised to advance the field of molecular property prediction and analysis.