Machine Learning Guided Atom Mapping of Metabolic Reactions.

Area of Science:

• Computational chemistry
• Chemical informatics
• Drug discovery

Background:

• Atom mapping is crucial for understanding chemical reaction mechanisms in computational drug design.
• Current graph-based methods for atom mapping lose vital chemical information.
• Enhancements using experimental bond stabilities exist but require manual input.

Purpose of the Study:

• To develop a fully automated, optimization-based approach for chemical reaction atom mapping.
• To improve the accuracy and robustness of automatic atom mapping techniques.
• To integrate machine learning for estimating bond stabilities.

Main Methods:

• Developed Automated Machine Learning Guided Atom Mapping (AMLGAM).
• Employed machine learning to estimate bond stabilities based on chemical environments.
• Utilized an optimization method favoring breakage/formation of less stable bonds.

Main Results:

• Evaluated AMLGAM on datasets of 382 and 7400 chemical reactions.
• Demonstrated improved accuracy over existing atom mapping techniques on a common dataset.
• Showcased capability in handling unbalanced chemical reactions.

Conclusions:

• AMLGAM offers a significant advancement in automated atom mapping.
• The method enhances the reliability of computational studies in drug design.
• This approach provides a more chemically informed graph representation for reactions.