Retro-Expert: Collaborative Reasoning for Interpretable Retrosynthesis

Retro-Expert is a new interpretable framework for retrosynthesis prediction that integrates Large Language Models (LLMs) with specialized models through reinforcement learning to improve logical decision-making in chemical synthesis. It features three components: chemical knowledge distillation, LLM-driven reasoning for predictions, and knowledge-grounded policy optimization, which collectively enhance interpretability and trustworthiness in outputs. Experimental results indicate that Retro-Expert outperforms existing LLM and specialized models, providing chemically grounded explanations that can aid practitioners in making more informed decisions in chemical synthesis.