Human-guided synthesis planning via prompting

Area of Science:

• Computational chemistry
• Organic synthesis planning
• Artificial intelligence in chemistry

Background:

• Current AI tools like AiZynthFinder excel at retrosynthesis but lack human expert knowledge integration.
• Incorporating chemist's insights on bond disconnections is crucial for planning efficient, joint synthesis routes for molecule sets.

Purpose of the Study:

• To develop a novel human-guided multistep retrosynthesis strategy within AiZynthFinder.
• To enable chemists to input prior knowledge, such as specific bond constraints, into the AI planning process.

Main Methods:

• Implemented a prompting strategy for human guidance in AiZynthFinder.
• Developed a filter to discard reactions violating user-defined bond constraints.
• Benchmarked disconnection strategies, combining a disconnection-aware transformer with multi-objective search.

Main Results:

• The novel approach significantly improved the success rate of generating routes satisfying bond constraints (75.57% vs. 54.80% on the PaRoutes dataset).
• Successfully applied the strategy to a set of drug molecules, generating a joint synthesis route for eight out of ten targets.

Conclusions:

• Human-guided retrosynthesis via prompting enhances AI planning capabilities.
• The developed strategy effectively incorporates expert chemical knowledge, leading to more successful and constrained synthetic route generation.