Retrosynthesis Zero: Self-Improving Global Synthesis Planning Using Reinforcement Learning.

Area of Science:

• Computational Chemistry
• Artificial Intelligence in Chemistry
• Chemical Synthesis Design

Background:

• Existing computer-aided synthesis planning (CASP) programs often require extensive datasets for neural network training, limiting their effectiveness due to data quality and reliance on prior chemical knowledge.
• Current single-step reaction template-based CASP methods face challenges in generating comprehensive synthesis routes.

Purpose of the Study:

• To introduce Retrosynthesis Zero (ReSynZ), a novel reaction template-based CASP method.
• To overcome the data dependency and prior knowledge limitations of current CASP programs.
• To develop a system capable of generating multiple synthesis pathways and suggesting reaction conditions for complex molecules.

Main Methods:

• ReSynZ employs a reaction template-based approach combined with Monte Carlo Tree Search and reinforcement learning, inspired by AlphaGo Zero.
• It utilizes complete synthesis paths derived from reaction rules as input for neural network training.
• The method is trained on relatively small reaction datasets (tens of thousands of data points).

Main Results:

• ReSynZ successfully trains neural networks with limited reaction data, generating multiple synthesis pathways for target molecules.
• The system demonstrates excellent predictive performance on various molecular retrosynthesis datasets, outperforming existing algorithms.
• ReSynZ can suggest appropriate reaction conditions for the proposed synthesis routes.

Conclusions:

• ReSynZ offers a significant advancement in computer-aided synthesis planning, overcoming key limitations of previous methods.
• Its self-improving model, flexible reward settings, and ability to generate diverse synthesis routes make it a powerful tool for chemical synthesis design.
• The approach has the potential to surpass human limitations in planning complex chemical synthesis routes.