DeepRetro is an innovative open-source retrosynthetic framework for discovering synthetic routes to complex natural products. Transcending the limitations of existing methods, it integrates large-scale language models (LLMs), conventional retrosynthetic engines, and expert feedback in an iterative design loop. It combines the accuracy of template-based methods with the generative flexibility of LLMs, enabling rigorous chemical feasibility testing and recursive refinement. An interactive user interface dynamically explores and refines synthetic routes with algorithmic validation and expert feedback. It excels on standard retrosynthetic benchmarks and is particularly strong in proposing novel synthetic routes to highly complex natural products, previously challenging with automated planning. Detailed case studies demonstrate how it can be used to propose novel routes for total synthesis and foster human-machine collaboration in organic chemistry. Beyond retrosynthesis, it presents a practical model for leveraging LLMs in scientific discovery. It is open-sourced, with a transparent description of the system design, algorithms, and human feedback loop, enabling its application across a wide range of scientific disciplines.
