Retro-Expert: Collaborative Reasoning for Interpretable Retrosynthesis

ACL ARR 2025 May Submission2771 Authors

Abstract: Retrosynthesis prediction aims to infer the reactant molecule based on a given product molecule, which is a fundamental task in chemical synthesis. The development of interpretable retrosynthesis models is crucial for chemist's decision by providing meaningful explanation. Building on this, we propose Retro-Expert, an interpretable retrosynthesis reasoning framework that combines domain-specific small models with large language models (LLMs) to generate human-readable reasoning alongside predictions via reinforcement learning. Unlike black-box data-driven models, Retro-Expert outputs natural language explanations grounded in chemical logic (e.g., reaction rules, principles) through three components: (1) specialized small models ensuring chemically valid candidates for reasoning, (2) LLM-driven reasoning to synthesize a decision-making pathway, and (3) reinforcement learning optimizing interpretable decision policy. Experiments show Retro-Expert achieves higher accuracy than single models while producing expert-aligned explanations, bridging AI predictions with actionable chemical insights.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: NLP Applications

Contribution Types: Model analysis & interpretability

Languages Studied: English

Keywords: NLP Applications

Submission Number: 2771