How Well Can LLMs Synthesize Molecules? An LLM-Powered Framework for Multi-Step Retrosynthesis

ICLR 2025 Conference Submission13841 Authors

Keywords: Chemistry, Retrosynthesis Planning, LLM

Abstract:

Predicting retrosynthesis routes is a fundamental challenge in chemistry, involving the design of a sequence of chemical reactions to synthesize a target molecule from commercially available starting materials. With a rapidly
growing interest in using large language models for planning, this work introduces an LLM-powered framework for multi-step retrosynthesis. Our framework employs molecular-similarity-based retrieval-augmented generation (RAG) to generate an initial retrosynthesis route, which is then iteratively refined through expert feedback. The use of molecular-similarity-based RAG improves reaction round-trip validity from 24.42\% to 51.64\% compared to GPT-4 with representative routes. With further refinement, the validity increases to 89.81\%, resulting in an overall route validity of 79.5\% with a perfect query success rate, comparable to traditional methods. Our framework offers a flexible, customizable approach to retrosynthesis, and we present a comprehensive analysis of the generated routes along with promising future research directions in LLM-driven multi-step retrosynthesis.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 13841