ReMol: A Chemical Reaction Knowledge-guided Self-supervised Molecular Image Representation Learning Framework

Molecular representation learning (MRL) is critical in computational chemistry and drug discovery, paving the way for efficient molecular properties and biological activity prediction. However, existing sequence-based or graph-based MRL methods emphasize static and intrinsic molecular topological features while ignoring dynamic and interactive chemical knowledge, resulting in insufficient generalization ability. MolR meets this challenge by leveraging the equivalence of molecules participating in chemical reactions in embedding space to assist in learning molecular representations. However, it can suffer from unsatisfactory performance because it lacks reaction center information and the complex relationship between reactions, which provides a deeper understanding of the chemical processes. We propose ReMol, an elaborate chemical reaction knowledge-guided self-supervised molecular image representation learning framework to address this issue. The ReMol framework integrates comprehensive reaction inductive biases, including reaction templates and consistency, and diversity in chemical reactions. Experimental results demonstrate that our framework achieves state-of-the-art results compared with cutting-edge methods on various challenging downstream tasks, such as chemical reaction and molecular property prediction tasks. Overall, our work offers a robust tool for advancing chemistry research, with the potential to make significant contributions to both molecular representation learning and drug discovery.