Geometry-Aware Line Graph Transformer Pretraining for Molecular Property Prediction.

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Deep learning for molecular property prediction is advancing.
• Self-supervised learning (SSL) addresses data scarcity by learning from unlabeled molecules.
• Integrating 3-D molecular geometry with 2-D representations can improve learning.

Purpose of the Study:

• To propose a novel SSL framework, Galformer, for enhanced molecular representation learning.
• To address challenges in dual-modality (2-D/3-D) pretraining for molecules.
• To improve the accuracy of molecular property prediction.

Main Methods:

• Developed a Geometry-aware line graph transformer (Galformer) backbone.
• Designed dual-modality pretraining tasks to capture inter- and intramodality relations.
• Utilized a dual-modality line graph transformer to encode 2-D topological and 3-D geometric information.

Main Results:

• Galformer consistently outperformed ten state-of-the-art baselines on 15 property prediction benchmarks.
• The framework demonstrated effectiveness in both classification and regression tasks.
• The proposed method successfully extracts discriminative 2-D and 3-D knowledge from unlabeled molecules.

Conclusions:

• Galformer provides a robust framework for molecular representation learning by effectively integrating 2-D and 3-D information.
• The dual-modality approach significantly enhances performance in downstream molecular property prediction tasks.
• This work offers a promising direction for leveraging geometric and topological data in cheminformatics.