Hierarchical Molecular Graph Self-Supervised Learning for property prediction.

Area of Science:

• Computational chemistry
• Machine learning in chemistry

Background:

• Molecular graph representation learning is crucial for drug discovery.
• Self-supervised learning is increasingly vital due to scarce labeled molecular data.
• Existing Graph Neural Networks (GNNs) overlook crucial chemical motif information.

Purpose of the Study:

• To introduce Hierarchical Molecular Graph Self-supervised Learning (HiMol) for improved molecular representation.
• To enhance molecular property prediction accuracy using a novel pre-training framework.
• To capture hierarchical chemical structures and semantic information within molecules.

Main Methods:

• Developed a Hierarchical Molecular Graph Neural Network (HMGNN) to encode motif structures.
• Implemented Multi-level Self-supervised Pre-training (MSP) with generative and predictive tasks.
• Learned node-motif-graph hierarchical representations for molecules.

Main Results:

• Achieved superior performance on molecular property prediction tasks (classification and regression).
• Demonstrated HiMol's effectiveness in capturing chemical semantic information.
• Visualization confirmed the model's ability to learn meaningful molecular representations.

Conclusions:

• HiMol effectively learns hierarchical molecular representations by integrating motif information.
• The proposed self-supervised pre-training framework significantly boosts molecular property prediction.
• HiMol offers a promising approach for advancing drug discovery and molecular analysis.