In this paper, we present DeepRetro, an open-source iterative hybrid retrosynthetic framework that integrates existing template-based/Monte Carlo tree search tools with the ability to generate large-scale language models (LLMs) to solve the retrosynthesis problem essential for complex molecular synthesis. DeepRetro first attempts a synthetic plan with a template-based engine, and if it fails, the LLM proposes a single-step retrosynthetic disjunction. The proposed disjunction is then tested for validity, stability, and hallucination, and the resulting precursors are recursively fed back into the pipeline for further evaluation. This iterative improvement allows dynamic path exploration and modification. Through benchmark evaluations and case studies, we demonstrate its ability to identify feasible and novel retrosynthetic pathways for complex natural product compounds, and in particular, we demonstrate the potential of LLM inference by developing an interactive graphical user interface that allows human-loop feedback from expert chemists.
