GraphGIM: rethinking molecular graph contrastive learning via geometry image modeling.

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Molecular representation learning is vital for drug discovery.
• Graph-based methods and contrastive learning are used to improve molecular graph representations.
• Existing methods have limited sample diversity, hindering performance.

Purpose of the Study:

• To address the limited sample diversity in existing graph-based contrastive methods.
• To propose a novel molecular graph contrastive learning method, GraphGIM, enhancing sample pair diversity.
• To improve the generalization and performance of molecular representations for drug discovery.

Main Methods:

• Developed GraphGIM, a novel molecular graph contrastive learning method utilizing geometry image modeling.
• Pre-trained GraphGIM on 2 million 2D graphs and multi-view 3D geometry images.
• Introduced two variants, GraphGIM-M and GraphGIM-P, by fusing feature maps at different scales using weighted and prompt-based strategies.

Main Results:

• GraphGIM enhances diversity between sample pairs, overcoming limitations of existing methods.
• Analysis revealed that convolutional layers extract both global (scaffold) and local (atom, functional group) chemical information at different scales.
• GraphGIM and its variants demonstrated superior performance on molecular property prediction benchmarks.

Conclusions:

• GraphGIM and its variants outperform state-of-the-art graph contrastive learning methods on MoleculeNet benchmarks.
• The proposed method achieves competitive results in molecular property prediction.
• The study provides a novel approach to enhance molecular representations for drug discovery.