Learning symmetry-aware atom mapping in chemical reactions through deep graph matching

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Chemical Informatics

Background:

• Accurate atom mapping is essential for understanding chemical reactions.
• Existing methods may struggle with complex molecular structures and symmetry.
• A comprehensive approach mapping all atoms is needed.

Purpose of the Study:

• To present AMNet, a novel end-to-end deep graph matching model for atom mapping.
• To improve the accuracy and efficiency of atom correspondence prediction in chemical reactions.
• To develop a model that considers molecular symmetry and maps entire reaction atom sets.

Main Methods:

• Formulating atom mapping as a deep graph matching task using molecular graph representations.
• Employing graph neural networks with atom and bond features to capture molecular structures.
• Integrating the Weisfeiler-Lehman isomorphism test for symmetry identification to refine predictions.

Main Results:

• AMNet achieves an average accuracy of 97.3% on mapped atoms.
• 99.7% of reactions are correctly mapped when the correct atom is within the top 10 predictions.
• The model demonstrates enhanced accuracy and reduced computational complexity, especially with symmetry consideration.

Conclusions:

• AMNet provides a novel and effective deep graph matching approach for chemical reaction atom mapping.
• The integration of molecular symmetry detection significantly improves prediction accuracy and efficiency.
• This comprehensive mapping of all reaction atoms offers a more complete analysis than previous methods.