Automated Retrosynthesis Planning of Macromolecules Using Large Language Models and Knowledge Graphs

Area of Science:

• Materials Chemistry
• Polymer Science
• Computational Chemistry

Background:

• Macromolecular nomenclature complexity hinders reliable synthesis pathway identification in polymer science.
• Existing methods for retrosynthesis planning are often limited and not fully automated.

Purpose of the Study:

• To develop a fully automated agent for retrosynthesis planning of macromolecules using large language models (LLMs) and knowledge graphs.
• To address the challenges posed by complex and nonunique nomenclature in polymer synthesis.

Main Methods:

• Integration of LLMs for chemical name recognition and data extraction.
• Utilization of knowledge graphs for structured data storage and retrieval.
• Development of a novel Multi-branched Reaction Pathway Search Algorithm (MBRPS) to identify all valid multi-branched reaction pathways.
• Automated literature retrieval, reaction data extraction, database querying, and pathway construction.

Main Results:

• The proposed agent successfully automates literature retrieval, reaction data extraction, and retrosynthesis pathway construction.
• The novel MBRPS algorithm identifies multi-branched reaction pathways, overcoming limitations of previous single-intermediate approaches.
• Application to polyimide synthesis generated a retrosynthetic pathway tree with hundreds of pathways, recommending optimized routes.

Conclusions:

• This work presents the first fully automated retrosynthesis planning agent for macromolecules powered by LLMs.
• The developed system enhances the efficiency and scope of identifying reliable synthesis pathways in polymer chemistry.
• The approach facilitates the discovery of both known and novel synthetic routes for complex polymers.