LLMs Achieve Retrosynthesis via Atom-Anchored Reasoning

A new framework enables general-purpose large language models (LLMs) to perform single-step retrosynthesis without task-specific training. The method anchors chain-of-thought reasoning to molecular structure using unique atomic identifiers. In a zero-shot task, the LLM identifies relevant fragments and chemical labels; an optional few-shot step uses class examples to predict the transformation. This approach overcomes prior LLM underperformance in retrosynthesis, validated on academic benchmarks and expert-validated drug discovery molecules. The work addresses the scarcity of labeled chemical data by leveraging LLMs' reasoning capabilities.