RLSynC: Offline-Online Reinforcement Learning for Synthon Completion.

Area of Science:

• Computational Chemistry
• Artificial Intelligence in Chemistry
• Organic Synthesis Planning

Background:

• Retrosynthesis is crucial for planning chemical synthesis.
• Semitemplate-based methods predict reaction centers and complete synthons.
• Existing methods face challenges in efficient synthon completion.

Purpose of the Study:

• To introduce RLSynC, an offline-online reinforcement learning method for synthon completion.
• To improve the accuracy and efficiency of semitemplate-based retrosynthesis.
• To enable exploration of novel reaction pathways in synthesis planning.

Main Methods:

• RLSynC employs multiple agents, one per synthon, for synchronized, step-by-step completion.
• The method combines offline training data with online interaction for policy learning.
• A standalone forward synthesis model guides action selection by evaluating reactant likelihood.

Main Results:

• RLSynC significantly outperforms current state-of-the-art synthon completion techniques.
• Improvements in performance reached up to 14.9% compared to existing methods.
• The synchronized agent approach enhances the exploration of chemical reaction spaces.

Conclusions:

• RLSynC offers a powerful new approach for automated synthesis planning.
• The method demonstrates the potential of reinforcement learning in advancing retrosynthesis.
• RLSynC paves the way for more efficient and comprehensive chemical synthesis design.