Bayesian Algorithm for Retrosynthesis.

Area of Science:

• Organic Chemistry
• Computational Chemistry
• Artificial Intelligence

Background:

• Retrosynthetic planning is traditionally time-consuming and relies heavily on expert knowledge.
• Limited exploration of the vast chemical reaction space hinders efficient synthesis discovery.
• Machine learning offers a transformative approach to chemical retrosynthesis.

Purpose of the Study:

• To develop a machine learning-driven method for discovering synthetic routes.
• To efficiently explore the combinatorial complexity of retrosynthetic planning.
• To identify diverse and probable reaction pathways from target molecules to commercially available compounds.

Main Methods:

• Utilized a deep neural network for forward prediction of reaction products from reactants.
• Inverted the forward model into a backward model using Bayesian inference.
• Employed a Monte Carlo search algorithm to explore diverse reaction sequences.

Main Results:

• Achieved approximately 87% accuracy in forward model predictions.
• Successfully rediscovered 81.8% of known one-step and 33.3% of two-step synthetic routes with top-10 accuracy.
• Generated hundreds of diverse reaction routes for synthetic targets.

Conclusions:

• The Bayesian retrosynthesis algorithm significantly enhances the efficiency of synthetic route discovery.
• The method provides a diverse set of potential reaction pathways, aiding in complex synthesis planning.
• This approach integrates computational methods with expert chemical knowledge for improved retrosynthesis.