Attention-wise masked graph contrastive learning for predicting molecular property.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery

Background:

• Accurate molecular property prediction is crucial for drug research.
• Supervised learning methods for molecular representation are limited by data availability and generalizability.

Purpose of the Study:

• To propose a self-supervised learning method, ATMOL, for improved molecular representation and property prediction.
• To address the limitations of supervised learning in exploring chemical space and ensuring generalizability.

Main Methods:

• Developed a novel attention-wise graph masking strategy for molecular graph augmentation.
• Utilized a graph attention network as the molecular graph encoder.
• Employed contrastive learning by minimizing loss between original and augmented molecular graphs.

Main Results:

• ATMOL achieved state-of-the-art performance in downstream molecular property prediction tasks.
• Pretraining on unlabeled data improved the generalization of learned molecular representations.
• Attention heatmaps revealed meaningful patterns of important atoms and groups for specific properties.

Conclusions:

• Self-supervised learning with attention-wise graph masking is effective for molecular representation.
• ATMOL enhances the ability to capture crucial structural and semantic information in molecules.
• The method shows promise for advancing drug discovery through improved molecular property prediction.