Skip to content

RetroReasoner: A Reasoning LLM for Strategic Retrosynthesis Prediction

Mar 2026 · arXiv.org · Vol abs/2603.12666 · 1 citation · 67 references
Computer Science

TL;DR

Experimental results show that RetroReasoner outperforms prior baselines, including not only molecular LLMs but also retrosynthesis-specific expert models, and generates a broader range of feasible reactant proposals, especially for challenging reaction instances.

Abstract

Retrosynthesis prediction aims to identify reactants that can synthesize a given product molecule. Although molecular large language models (LLMs) have recently shown promising results, most existing methods either generate reactants directly or provide only generic product-level analysis, without explicitly reasoning about bond-disconnection strategies that justify specific reactant choices. This paper proposes RetroReasoner, a retrosynthetic reasoning model that captures chemists'strategic disconnection-based thinking. RetroReasoner is trained with supervised fine-tuning and reinforcement learning. For supervised fine-tuning, SyntheticRetro generates structured disconnection rationales paired with reactant predictions. For reinforcement learning, a round-trip reward evaluates predicted reactants by passing them through a forward synthesis model and rewarding predictions that reconstruct the original product. RetroReasoner can also be applied to multi-step retrosynthetic planning by incorporating it into a parallelized Monte Carlo tree search framework, reducing search time while increasing the number and diversity of valid synthetic pathways. Experimental results show that RetroReasoner outperforms prior baselines, including not only molecular LLMs but also retrosynthesis-specific expert models, and generates a broader range of feasible reactant proposals, especially for challenging reaction instances. The code is available at https://github.com/KU-AGI/RetroReasoner.

View source

Similar papers

Sep 2026

Integrated Retrosynthesis and Large Language Modeling for Chemistry-Informed Circular Chemical Reaction Networks

The proposed framework demonstrates that combining language-based knowledge extraction with structure-informed retrosynthetic reasoning enables scalable construction of CCRNs and supports the systematic exploration of candidate circular chemical pathways for subsequent economic and environmental evaluation.

Sunghoon Kim, Avan Kumar, Hariprasad Kodamana et al. · 0 citations
Conference Open access Sep 2026

Toward Synthesizability-Aware Multi-Step Retrosynthetic Planning

This work proposes GuideRetro, a synthesizability-aware framework for multi-step retrosynthetic planning that integrates global syn-thesizability knowledge into step-wise retrosyn-thetic prediction and improves planning accuracy and search efficiency under realistic retrosynthetic settings.

Yu-Jie Chen, Ajie Lin, Tengfei Ma et al. · 0 citations

RetroGEF: Dynamic Graph Edit Flow for Single-Step Retrosynthesis

RetroGEF, a flow-based generative model for single-step retrosynthesis that learns this process directly from product--reactant pairs without requiring a prescribed edit order, achieves state-of-the-art performance.

Xiao-Zhuang Song, Xue-Min Chen, Xinjian Zhao et al. · 0 citations
#artificial intelligence Preprint Aug 2026

Training Chemical Plausibility-Aware Large Language Models for Single-Step Retrosynthesis

This work introduces Top-K prompting as a robust training and inference paradigm to better capture diverse, plausible reaction predictions and establishes Top-K, plausibility-aware training as a practical new direction for robust future LLM-based synthesis planning.

B. Zagribelnyy, Ivan D. Ilin, N. Bondarev et al. · 1 citation
Sep 2026

Adaptive prioritized expansion for cost-conscious retrosynthetic route planning.

Results indicate that incorporating molecular cost information into heuristic search can improve the practicality and economic efficiency of retrosynthetic planning.

Shuan Liu, Jing-Wen Wang, Shao-Ye Zhang et al. · 0 citations
Open access Aug 2026

Leveraging the Condensed Graph of Reaction for Clustering Retrosynthetic Pathways.

This work proposes a two-level clustering strategy in which routes are first grouped by shared SBPs, ensuring high interpretability based on key retrosynthetic disconnections, and then further differentiated using RouteCGR similarity to capture variations in starting materials and auxiliary transformations.

Almaz Gilmullin, T. Akhmetshin, D. Zankov et al. · 0 citations

Related blog posts

MIT News · Artificial Intelligence Sep 29, 2026

Who we become when we talk to machines

Professor Sherry Turkle’s new book, “Artificial Intimacy,” offers a withering critique of chatbots and the antisocial dynamics she believes they encourage.

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.