Completing and Balancing Database Excerpted Chemical Reactions with a Hybrid Mechanistic-Machine Learning Approach

Area of Science:

• Chemical Informatics
• Computational Chemistry
• Machine Learning in Chemistry

Background:

• Computer-aided synthesis planning (CASP) relies on complete reaction data, but current databases often lack essential reaction coparticipants.
• Existing reaction prediction and atom mapping tools struggle to identify missing molecules due to reliance on incomplete training datasets.

Purpose of the Study:

• To develop a robust workflow for completing imbalanced chemical reactions and predicting missing reaction participants.
• To enhance the completeness of chemical reaction databases for improved CASP.

Main Methods:

• A heuristic-based method was employed to identify and balance reactions within existing databases by adding candidate molecules.
• A machine learning masked language model (MLM) was trained on simplified molecular input line entry system (SMILES) strings of completed reactions.
• The MLM was utilized to predict missing molecules for incomplete reaction records, analogous to predicting missing words in text.

Main Results:

• The developed workflow successfully identified and completed a significant portion of imbalanced reactions.
• The machine learning model demonstrated promise in predicting small- and middle-sized missing molecules.
• The combined heuristic and machine learning approach addressed over half of the reaction space requiring completion.

Conclusions:

• The integrated workflow effectively tackles the challenge of incomplete reaction data in chemical databases.
• This method significantly advances the potential for accurate computer-aided synthesis planning by providing more complete reaction information.
• The approach offers a scalable solution for improving the quality and utility of chemical reaction datasets.