ReLMole: Molecular Representation Learning Based on Two-Level Graph Similarities.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Bioinformatics

Background:

• Graph neural networks (GNNs) are key for molecular representation learning.
• Limited task-specific labels in biomedicine hinder GNN performance.
• Existing GNN pretraining methods overlook unique molecular data properties.

Purpose of the Study:

• To develop a novel representation learning method for molecular graphs.
• To address the limitations of GNNs due to label scarcity in biomedical domains.
• To improve predictions for molecular properties and drug-drug interactions.

Main Methods:

• Proposed ReLMole, a hierarchical graph modeling approach for molecules.
• Implemented a contrastive learning scheme based on two-level graph similarities.
• Assessed performance on molecular property (MP) and drug-drug interaction (DDI) prediction tasks.

Main Results:

• ReLMole achieved superior performance on both MP and DDI prediction tasks.
• Outperformed baseline models by over 2.6% in ROC-AUC for six MP prediction tasks.
• Improved F1 value by 7-18% in DDI prediction for unseen drugs compared to other self-supervised models.

Conclusions:

• ReLMole offers a powerful self-supervised pretraining strategy for molecular graph representation.
• The method effectively handles label scarcity and improves predictive accuracy in drug discovery.
• ReLMole demonstrates significant potential for advancing computational drug design and biomedical predictions.