Chemical reasoning in LLMs unlocks steerable synthesis planning and reaction mechanism elucidation

Andres M Bran, Théo A. Neukomm, Daniel Armstrong, Zlatko Jončev, Philippe Schwaller · 2025

Read Chemical reasoning in LLMs unlocks steerable synthesis planning and reaction mechanism elucidation by Andres M Bran, Théo A. Neukomm, Daniel Armstrong, Zlatko Jončev, Philippe Schwaller (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-08537.txt.

Introduction

Andres Bran and co-authors show how large language models can guide the planning of chemical work, in a 2025 paper. Software tools for chemistry are strong at narrow tasks but weak at the strategic thinking of an expert. The team pairs language models with classic search methods instead of letting the models move molecules directly. The models judge strategies, and search does the precise building.

Two hard tasks test the idea. In planning how to build a target molecule, chemists can state a strategy in plain words, and the system hunts for routes that fit it. In explaining how a reaction proceeds, the models steer a search for plausible step-by-step paths using chemical principles. Newer, larger models showed richer reasoning across tasks. The authors frame the mix as a more natural and powerful form of computer-aided chemistry.

Catalog record

CREATORAndres M Bran, Théo A. Neukomm, Daniel Armstrong, Zlatko Jončev, Philippe Schwaller
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE51 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-08537

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Chemical reasoning in LLMs unlocks steerable synthesis planning and reaction mechanism elucidation Andres M Bran, Théo A. Neukomm, Daniel Armstrong, Zlatko Jončev, Philippe Schwaller Abstract While machine learning algorithms have been shown to excel at specific chemical tasks, they have struggled to capture the strategic thinking that characterizes expert chemical reasoning, limiting their widespread adoption. Here we demonstrate that large language models (LLMs) can serve as powerful chemical reasoning engines when integrated with traditional search algorithms, enabling a new approach to computer-aided chemistry that mirrors human expert thinking. Rather than using LLMs to directly manipulate chemical structures, we leverage their ability to evaluate chemical strategies and guide search algorithms toward chemically meaningful solutions. We demonstrate this paradigm through two fundamental challenges: strategy-aware retrosynthetic planning and mechanism elucidation. In retrosynthetic planning, our method allows chemists to specify desired synthetic strategies in natural language to find routes that satisfy these constraints in vast searches. In mechanism elucidation, LLMs guide the search for plausible reaction mechanisms by combining chemical principles with systematic exploration. Our approach shows strong performance across diverse chemical tasks, with larger models demonstrating increasingly sophisticated chemical reasoning. Our approach establishes a new paradigm for computer-aided chemistry that combines the strategic understanding of LLMs with the precision of traditional chemical tools, opening possibilities for more intuitive and powerful chemical reasoning systems. 1 Introduction The automation of chemical reasoning has been a long-standing goal in many areas of chemistry, promising to accelerate drug discovery, retrosynthesis 1 , 2 , 3 and our understanding of chemical reactivity 4 . Traditional machine learning and computational approaches have focused on specialized algorithms for specific tasks – predicting properties 5 , planning syntheses 1 , 6 , or proposing reaction mechanisms 7 , 8 , 9 , 10 , 11 , 12 , 13 , 14 . While successful in narrow domains, these systems lack the flexible reasoning and strategic multi-step thinking that characterize expert chemical problem-solving 15 , 16 , 17 . A synthetic chemist planning the synthesis of a complex molecule, for instance, must simultaneously consider multiple strategic factors: which rings to form first, when to install sensitive functional groups and protective groups, and how to leverage available starting materials 18 . Additionally, this logical deconstruction of molecules must be supported by mechanistic reasoning which requires extrapolation of chemical principles of elementary steps and reactive intermediates, to newly proposed molecules, usually not observed before 4 , 19 . The exponential growth in capabilities of Large Language Models (LLMs) has sparked a revolution across scientific disciplines 20 , 21 , 22 , 3 , with applications spanning from automated literature analysis to hypothesis generation 23 , 24 , 25 . These models have demonstrated an unprecedented ability to understand and reason about chemical concepts at multiple scales - from individual functional groups to complete synthetic pathways 3 , 26 , 27 . Most remarkably, they exhibit reasoning patterns that mirror human chemical intuition rather than traditional computational approaches 28 , 21 , showing particular promise in analyzing strategic elements of synthesis such as protecting group patterns and ring construction timing 18 . However, a fundamental limitation persists: while LLMs excel at analyzing chemical concepts and strategies, they struggle to generate valid chemical representations, particularly SMILES strings 29 , 30 , 31 , limiting their direct application in critical chemical tasks. Here, we present a paradigm shift in how LLMs can advance chemical science: rather than attempting to generate chemical structures directly, we position these models as sophisticated reasoning engines that guide traditional search algorithms toward chemically meaningful solutions. This approach combines LLMs’ ability to understand and evaluate complex chemical strategies with the systematic exploration capabilities of established search methods. Through systematic evaluation, we first demonstrate that LLMs can effectively analyze chemical entities and strategic patterns across multiple scales. We then show how these reasoning capabilities can be exploited in two challenging applications: strategy-aware retrosynthetic planning and mechanism elucidation. Retrosynthetic planning —the process of systematically breaking down complex target molecules into simpler, commercially available starting materials— represents one of the most crucial and intellectually demanding tasks in organic chemistry 1 , 32 . Current computational methods employ sophisticated search algorithms guided by either carefully designed or learned heuristics 33 , 34 , 35 . While these systems excel at finding routes that end in commercially available materials 36 , 37 , they often struggle to incorporate strategic considerations commonly used by expert chemists, such as optimal timing for ring construction, protecting group introduction or incompatible transformations 18 . Recent specialized systems have addressed specific constraints like starting material availability 38 , 39 and bond preservation 40 , 41 , but a framework for arbitrary strategy-aware synthesis planning has remained elusive. Our approach allows chemists to specify strategic requirements in natural language, with LLM-guided search identifying synthetic routes that satisfy these complex constraints. We demonstrate these capabilities on a challenging benchmark along with case studies including the synthetic routes of atorvastatin and strychnine. Similarly, mechanism elucidation —understanding the step-by-step electron movements that transform reactants into products— is fundamental to both chemical understanding and reaction optimization 4 , 19 , 42 . The power of mechanistic understanding lies not only in explaining individual reactions but in their potential to generalize patterns of chemical reactivity that can be applied to previously unseen molecules 43 . Existing computational methods can enumerate possible reaction paths 7 , 44 but often lack the chemical intuition needed to identify plausible mechanisms. While specialized approaches combining quantum calculations with search algorithms show promise 10 , they struggle to scale for complex systems and rely on predefined templates or atom-mapping, limiting their applicability. In our approach, basic electron-pushing steps are evaluated using LLMs’ understanding of chemical principles in the context of a search algorithm, allowing the LLM to guide the search towards reasonable mechanisms, while potentially considering diverse forms of experimental evidence and practical constraints. Our results demonstrate that LLMs can effectively guide search processes and select optimal solutions, while providing chemically meaningful rationales for their decisions. Furthermore, we provide insights into how different models’ capabilities —related to pretraining 45 , post-training 46 , 47 , 28 and inference-time scaling 48 , 28 — affect solution quality, establishing crucial practical considerations for deploying such systems. This integration represents a significant step toward computational chemistry systems that can reason strategically about complex synthesis challenges while maintaining the precision of traditional computational tools. Figure 1: LLMs as chemical reasoning engines for synthesis planning and mechanism elucidation. a) Current state of LLMs in chemistry, highlighting strengths in property prediction, multiple choice questions, and agentic workflows, alongside limitations in structure generation tasks. b) LLMs demonstrate sophisticated chemical reasoning capabilities, providing detailed analyses of reaction mechanisms and functional group transformations. c) Our key insight: positioning LLMs as strategic evaluators within chemical search frameworks. Rather than generating structures directly, LLMs guide traditional search algorithms toward chemically meaningful solutions. d) Application to synthesis planning: LLMs assess candidate routes based on expert queries specifying strategic requirements (e.g., "break pyrimidine in the early stage"). This yields strategically relevant synthetic pathways with detailed rationales for synthetic choices. e) Application to mechanism elucidation: LLMs guide search through possible reaction mechanisms by evaluating the plausibility of elementary

Read the complete text