DeepRetro discovers retrosynthetic pathways through iterative large language model reasoning.

Synthesizing complex natural products is a grand challenge in organic chemistry. We present DeepRetro, a significant advancement in computational retrosynthesis that discovers viable synthetic routes for molecules previously considered too complex for automated methods. DeepRetro is a novel, open-source framework that tightly integrates large language models (LLMs), traditional retrosynthetic engines, and expert human feedback into an iterative design loop. Unlike prior approaches that rely on either template-based methods or unconstrained LLMs, our hybrid system combines the precision of templates with the generative flexibility of LLMs, governed by rigorous chemical validity checks and recursive refinement. This system dynamically explores and revises synthetic pathways, guided by algorithmic checks and expert input through an interactive interface. While DeepRetro shows strong performance on standard benchmarks, its main strength is its ability to propose novel, viable pathways for highly complex natural products. Through case studies, we demonstrate how this approach facilitates new total synthesis routes and enhances human-machine collaboration. DeepRetro serves as a working model for applying LLMs to scientific discovery, and we release it as an open-source tool to accelerate progress in drug discovery and materials design.