Harnessing LLM Reasoning for Multi-Step Retrosynthesis

Establish methods that effectively harness the reasoning capabilities of large language models (LLMs) to perform multi-step retrosynthetic planning, enabling reliable discovery of complete synthetic pathways from target molecules.

Background

Retrosynthetic planning seeks to decompose complex target molecules into simpler precursors through a sequence of validated reaction steps. While traditional computer-aided synthesis planning tools rely on templates and heuristic search, they can struggle to propose novel or non-obvious routes, especially in multi-step settings.

LLMs have recently shown promise in chemistry tasks, including single-step retrosynthesis. However, translating their general reasoning abilities into robust, controllable multi-step synthesis plans presents unique challenges, such as error propagation and hallucinations. The paper introduces DeepRetro as an iterative, validated integration of LLM suggestions with conventional tools, highlighting that determining how to systematically and reliably harness LLM reasoning for multi-step planning remains a broader open question.