GraphGIM: rethinking molecular graph contrastive learning via geometry image modeling
Abstract Background Learning molecular representations is crucial for accurate drug discovery. Using graphs to represent molecules is a popular solution, and many researchers have used contrastive learning to improve the generalization of molecular graph representations. Results In this work, we rev...
Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
BMC
2025-07-01
|
| Series: | BMC Biology |
| Subjects: | |
| Online Access: | https://doi.org/10.1186/s12915-025-02249-0 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract Background Learning molecular representations is crucial for accurate drug discovery. Using graphs to represent molecules is a popular solution, and many researchers have used contrastive learning to improve the generalization of molecular graph representations. Results In this work, we revisit existing graph-based contrastive methods and find that these methods have limited diversity in the constructed sample pairs, resulting in insufficient performance gains. To alleviate the above challenge, we propose a novel molecular graph contrastive learning method via geometry image modeling, called GraphGIM, which enhances the diversity between sample pairs. GraphGIM is pre-trained on 2 million 2D graphs and multi-view 3D geometry images through contrastive learning. Furthermore, we find that as the convolutional layers process the image becomes deeper, the information of feature maps gradually changes from global molecular-level information (molecular scaffolds) to local atomic-level information (molecular atoms and functional groups), which provides chemical information at different scales. Therefore, we propose two variants of GraphGIM, called GraphGIM-M and GraphGIM-P, which fuse feature maps of different scales in the image using a weighted strategy and a prompt-based strategy, respectively. Conclusions Extensive experiments show that GraphGIM and its two variants outperform state-of-the-art graph contrastive learning methods on eight molecular property prediction benchmarks from MoleculeNet and achieve competitive results with state-of-the-art methods. The code is available at https://github.com/cyli029/GraphGIM . |
|---|---|
| ISSN: | 1741-7007 |