Biocatalysed synthesis planning using data-driven learning.

Area of Science:

• Biocatalysis and Green Chemistry
• Artificial Intelligence in Chemistry
• Computational Chemistry

Background:

• Enzyme catalysts are crucial for sustainable chemical synthesis but predicting their activity and selectivity is challenging.
• Current retrosynthetic planning tools for biocatalysis are limited to rule-based systems.
• Data-driven approaches have primarily focused on forward reaction predictions.

Purpose of the Study:

• To extend data-driven prediction models for both forward reactions and retrosynthetic pathways in biocatalysis.
• To integrate enzymatic knowledge into the Molecular Transformer architecture.
• To facilitate the use of enzymatic catalysis in designing greener chemical processes.

Main Methods:

• Utilized the Molecular Transformer architecture for data-driven prediction of forward reactions and retrosynthetic pathways.
• Learned enzymatic knowledge from a large dataset of biochemical reactions.
• Implemented a novel class token scheme based on enzyme commission classification numbers to capture enzyme hierarchies and catalysis patterns.

Main Results:

• Achieved a top-1 accuracy of 49.6% for the forward reaction prediction model.
• Obtained a top-1 single-step round-trip accuracy of 39.6% for the retrosynthetic pathway prediction.
• Developed and curated a dataset for biocatalysis prediction, made publicly available.

Conclusions:

• The developed models successfully extend data-driven approaches to biocatalytic retrosynthetic planning.
• The publicly available dataset and models will aid researchers in adopting enzymatic catalysis for greener chemistry.
• This work bridges the gap between computational prediction and practical application of biocatalysis.