Learning Retrosynthetic Planning through Simulated Experience.

Area of Science:

• Computational Chemistry
• Artificial Intelligence in Chemistry
• Retrosynthetic Analysis

Background:

• Retrosynthetic planning involves working backward from a target molecule to starting materials.
• The combinatorial complexity and uncertain reaction outcomes make this process challenging.
• Existing methods often rely on heuristics with limitations in exploring the vast chemical space.

Purpose of the Study:

• To develop an artificial intelligence (AI) approach for optimizing retrosynthetic planning.
• To train a model that predicts the cost and feasibility of chemical synthesis steps.
• To identify optimal reaction pathways using deep reinforcement learning (DRL).

Main Methods:

• Framing retrosynthetic planning as a one-player game solvable with DRL.
• Training a neural network on simulated synthesis data to estimate molecular synthesis cost.
• Developing learned policies for making optimal reaction choices based on predicted molecular value.

Main Results:

• The DRL-based policies demonstrated superior performance compared to heuristic approaches.
• The learned policies effectively identified synthesis routes with fewer reactions for unfamiliar molecules.
• The trained neural network accurately estimated the expected cost of synthesizing molecules.

Conclusions:

• Deep reinforcement learning offers a powerful framework for automating and optimizing retrosynthetic planning.
• The developed value network and learned policies can be integrated into existing chemical synthesis tools.
• The approach is adaptable to different cost metrics and changes in starting material availability.