Chemical structure-aware molecular image representation learning.

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Molecular image-based drug discovery faces challenges with unlabeled data and extracting chemical structures from images.
• Current methods struggle to bridge the gap between implicit image information and explicit chemical structure encoding.

Purpose of the Study:

• To develop a novel framework for molecular representation learning that effectively transfers chemical knowledge from molecular graphs to images.
• To enable molecular image encoders to perceive and interpret chemical structures without explicit labels.

Main Methods:

• Proposed a Contrastive Graph-Image Pre-training (CGIP) framework utilizing self-supervised contrastive learning.
• Employed intra- and inter-modal contrastive learning to extract explicit graph information and implicit image information from large-scale unlabeled molecules.
• Leveraged molecular graphs encoding explicit chemical structures (e.g., benzene rings, double bonds).

Main Results:

• Achieved state-of-the-art performance across 12 benchmark datasets in molecular property prediction, cross-modal retrieval, and distribution similarity tasks.
• Demonstrated successful transfer of chemical knowledge from graphs to images, allowing image encoders to recognize chemical structures.
• Validated the framework's effectiveness in learning from unlabeled molecular data.

Conclusions:

• The Contrastive Graph-Image Pre-training (CGIP) framework effectively bridges the gap between molecular graphs and images for representation learning.
• CGIP enables image encoders to interpret chemical structures, advancing molecular image-based drug discovery.
• This approach highlights the potential of leveraging molecular images for enhanced representation learning in chemistry.