Self-Supervised Molecular Representation Learning With Topology and Geometry.

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Molecular representation learning is crucial for drug analysis, with self-supervised pre-training addressing limited labeled data.
• Current methods integrate 2D and 3D structures but lack hierarchical learning for inter- and intra-molecule correlations.
• Existing approaches often use separate 2D or 3D encoders, leaving fusion potential unexplored.

Purpose of the Study:

• To propose a novel Multi-View Molecular Representation Learning (MVMRL) method for enhanced molecular property prediction.
• To develop hierarchical pre-training tasks capturing both 2D graph and 3D geometric information.
• To introduce a motif-level fusion strategy for combining multi-view molecular representations.

Main Methods:

• Designed hierarchical pre-training tasks: fine-grained atom-level for 2D graphs and coarse-grained molecule-level for 3D graphs.
• Implemented a motif-level fusion pattern during fine-tuning to integrate complementary 2D and 3D molecular features.
• Evaluated MVMRL against state-of-the-art methods on molecular property prediction tasks.

Main Results:

• The proposed MVMRL method demonstrated superior performance compared to existing baseline methods.
• Hierarchical pre-training effectively captured both inter- and intra-molecule correlations.
• Motif-level fusion significantly enhanced molecular property prediction accuracy.

Conclusions:

• MVMRL offers a powerful approach for molecular property prediction by effectively leveraging multi-view molecular data.
• The hierarchical and fusion strategies address limitations of previous representation learning methods.
• This work advances self-supervised learning in cheminformatics for drug discovery applications.