T-MGCL: Molecule Graph Contrastive Learning Based on Transformer for Molecular Property Prediction

Area of Science:

• Computational Chemistry
• Machine Learning
• Drug Discovery

Background:

• Machine learning is increasingly vital for molecular studies, aiding tasks like property prediction and drug design.
• A key challenge is developing models that leverage unlabeled data for training while maintaining high performance.

Purpose of the Study:

• To propose a novel neural network architecture, Molecule Graph Contrastive Learning based on the Transformer framework (T-MGCL).
• To address the need for effective utilization of extensive unlabeled molecular data in machine learning models.

Main Methods:

• Developed a Molecule Graph Contrastive Learning (T-MGCL) approach utilizing the Transformer framework.
• Employed unsupervised molecular graph expansions and a contrast estimator for consistency.
• Incorporated atom distances and molecular graph attributes within the Transformer to capture structural information.

Main Results:

• The T-MGCL model demonstrated superior performance across multiple molecular property prediction tasks compared to existing models.
• Attention weights learned by T-MGCL were found to be chemically interpretable.

Conclusions:

• T-MGCL offers a powerful approach for molecular property prediction using unlabeled data.
• The model's ability to interpret attention weights provides chemical insights, enhancing its utility in drug design and molecular studies.