Reasoning-Driven Retrosynthesis Prediction with Large Language Models via Reinforcement Learning

2025-07-23

AI Summary

Existing retrosynthetic prediction methods suffer from poor generalizability, limited interpretability, and insufficient explicit integration of chemical knowledge. To address these limitations, we propose RetroDFM-R—the first large language model that synergistically integrates an explicit multi-step reasoning mechanism with chemical-rule-guided reinforcement learning. It employs a verifiable reward function to optimize the generation of synthetically viable pathways, enabling high-accuracy and high-fidelity retrosynthetic planning. RetroDFM-R supports end-to-end multi-step pathway deconstruction and achieves a state-of-the-art 65.0% top-1 accuracy on USPTO-50K. Double-blind expert evaluation confirms its chemical plausibility, and it successfully reproduces real-world multi-step syntheses of diverse pharmaceuticals and perovskite materials. This work advances both theoretical foundations—through principled incorporation of domain-specific chemical constraints—and practical applicability—by delivering robust, interpretable, and experimentally validated synthetic routes.

Technology Category

• Planning, Routing, and Scheduling: Planning with Language Models
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Code Generation / Program Synthesis from Natural Language

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Retrosynthesis planning, essential in organic synthesis and drug discovery, has greatly benefited from recent AI-driven advancements. Nevertheless, existing methods frequently face limitations in both applicability and explainability. Traditional graph-based and sequence-to-sequence models often lack generalized chemical knowledge, leading to predictions that are neither consistently accurate nor easily explainable. To address these challenges, we introduce RetroDFM-R, a reasoning-based large language model (LLM) designed specifically for chemical retrosynthesis. Leveraging large-scale reinforcement learning guided by chemically verifiable rewards, RetroDFM-R significantly enhances prediction accuracy and explainability. Comprehensive evaluations demonstrate that RetroDFM-R significantly outperforms state-of-the-art methods, achieving a top-1 accuracy of 65.0% on the USPTO-50K benchmark. Double-blind human assessments further validate the chemical plausibility and practical utility of RetroDFM-R's predictions. RetroDFM-R also accurately predicts multistep retrosynthetic routes reported in the literature for both real-world drug molecules and perovskite materials. Crucially, the model's explicit reasoning process provides human-interpretable insights, thereby enhancing trust and practical value in real-world retrosynthesis applications.

Problem

• Improving accuracy in retrosynthesis prediction using LLMs
• Enhancing explainability of AI-driven chemical synthesis plans
• Generalizing chemical knowledge for broader applicability

Innovation

• LLM with reinforcement learning for retrosynthesis
• Chemically verifiable rewards enhance accuracy
• Explicit reasoning provides interpretable insights