An effective self-supervised framework for learning expressive molecular global representations to drug discovery.

Area of Science:

• Artificial intelligence in drug discovery
• Machine learning for molecular modeling

Background:

• Developing expressive molecular representations is crucial for AI-driven drug discovery.
• Graph neural networks (GNNs) are powerful for molecular data, but supervised methods face data limitations.
• Existing approaches often struggle with data scarcity and poor generalization.

Purpose of the Study:

• To propose a novel molecular pre-training framework (MPG) for learning representations from large unlabeled molecular datasets.
• To introduce MolGNet, a powerful GNN for molecular graph modeling within the MPG framework.
• To enable state-of-the-art performance on diverse drug discovery tasks through self-supervised pre-training.

Main Methods:

• Developed MPG, a graph-based deep learning framework for molecular representation learning.
• Proposed MolGNet, a GNN architecture for molecular graph modeling.
• Implemented a self-supervised strategy for pre-training MolGNet at node and graph levels on 11 million unlabeled molecules.

Main Results:

• Pre-trained MolGNet captures valuable chemical insights and produces interpretable representations.
• Fine-tuning the pre-trained MolGNet achieved state-of-the-art results on 14 benchmark datasets.
• Demonstrated effectiveness across molecular properties prediction, drug-drug interaction, and drug-target interaction tasks.

Conclusions:

• The MPG framework and pre-trained MolGNet offer a powerful approach for molecular representation learning.
• MPG addresses limitations of supervised methods by leveraging large unlabeled datasets.
• Pre-trained MolGNet shows significant potential as an advanced molecular encoder in drug discovery pipelines.