Predicting enzymatic reactions with a molecular transformer.

Area of Science:

• Biocatalysis
• Machine Learning in Chemistry
• Organic Synthesis

Background:

• Enzymes offer selective and economical routes for organic synthesis.
• Predicting enzyme-catalyzed reactions remains a significant challenge in chemistry.

Purpose of the Study:

• To develop a predictive model for enzyme transformations.
• To accurately predict the structure and stereochemistry of enzymatic reaction products.

Main Methods:

• Utilized multi-task transfer learning to train a molecular transformer model.
• Combined one million chemical reactions from the USPTO database with 32,181 enzymatic transformations.
• Integrated reaction SMILES (Simplified Molecular Input Line Entry System) with natural language descriptions of enzymes.

Main Results:

• The enzymatic transformer model demonstrated remarkable accuracy in predicting reaction products.
• The model successfully interpreted both SMILES notation and human language descriptions of enzymes.
• Achieved accurate prediction of product structure and stereochemistry for enzyme-catalyzed reactions.

Conclusions:

• The developed enzymatic transformer advances the prediction of biocatalytic processes.
• This approach enhances the utility of enzymes in organic synthesis by improving predictability.
• Combines chemical language (SMILES) with expert enzyme descriptions for enhanced machine learning performance.