Chemical reasoning in LLMs unlocks strategy-aware synthesis planning and reaction mechanism elucidation

TL;DR

The paper reframes chemical reasoning as a problem of guiding traditional search with LLM-based strategic evaluation, rather than direct structure generation. It introduces Synthegy, a framework that couples AiZynthfinder-style retrosynthesis with LLMs to satisfy strategic constraints and to elucidate reaction mechanisms via guided search. Results show that state-of-the-art (and open) LLMs can meaningfully assess route feasibility and propose plausible mechanisms, with larger models achieving higher fidelity and explainability. The work highlights both the promise of human-like chemical reasoning in automation and current limitations (long sequences, input formatting, and model biases) while outlining a path toward more intuitive, capable computer-aided chemistry systems.

Abstract

While automated chemical tools excel at specific tasks, they have struggled to capture the strategic thinking that characterizes expert chemical reasoning. Here we demonstrate that large language models (LLMs) can serve as powerful tools enabling chemical analysis. When integrated with traditional search algorithms, they enable a new approach to computer-aided synthesis 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 system allows chemists to specify desired synthetic strategies in natural language -- from protecting group strategies to global feasibility assessment -- and uses traditional or LLM-guided Monte Carlo Tree Search to find routes that satisfy these constraints. In mechanism elucidation, LLMs guide the search for plausible reaction mechanisms by combining chemical principles with systematic exploration. This approach shows strong performance across diverse chemical tasks, with newer and 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 automation systems.

Figures (4)

• 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) Analytical capabilities go beyond single reactions, being able to analyze a route in global terms. LLMs are capable of detecting when protecting groups are unnecessarily proposed as part of a synthetic route. d) 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. e) Application to synthesis planning: LLMs assess candidate routes based on guidance queries specifying strategic requirements (e.g., "break pyrimidine in the early stage"). This yields strategically relevant synthetic pathways with detailed rationales for synthetic choices. f) Application to mechanism elucidation: LLMs guide search through possible reaction mechanisms by evaluating the plausibility of elementary electron-pushing steps. The system efficiently identifies correct mechanistic pathways while providing chemically meaningful justifications. This approach combines the strategic understanding of LLMs with the precision of traditional chemical search algorithms.
• Figure 2: Performance of the system for strategy-aware synthesis planning presented in this work. a Performance of multiple LLMs of different sizes and providers across all the tasks in the benchmark. The tasks are grouped by synthetic target, and each column specifies a prompt as specified in SI-\ref{['si:steering-prompts']}. The y axis displays the correlation between LLM-produced scores and those computed as specified in the benchmark. Each data point represents a separate execution of the benchmark, with the number of repetitions (n) varying across LLMs based on execution time and resource constraints as specified in the legend. Error bars denote confidence intervals (95%). b Example of an LLM's analysis of a synthetic route, where it provides a justification why a specific route received a high score. The example illustrates that the LLM analyses each reaction (exemplified with step 4), and then provides an overall analysis where it highlights the alignment with the user's query. c Average performance of multiple LLMs over time. Shown are only models that are the state of the art for their time, or close to it. The plot illustrates the rapidly evolving advancements in capability. d Illustrates the task of strategy-aware synthesis planning: a user specifies a target molecule along with a query in natural language, which specifies desired features in the route. The proposed solutions are given together with scores that signify their alignment with the query given by the user.
• Figure 3: Selecting highly feasible routes.a Distribution of feasibility scores as determined by Synthegy, for the four targets in SI-\ref{['si:steering-prompts']}. Results have been computed for solutions coming from 3 different retrosynthesis engines and one experimentally validated route, for each target. b Example analysis of a feasible synthetic route. The model determines that the two reactions are feasible and the full plan is consistent. c Example of an unfeasible synthetic route. Synthegy correctly identifies 2 key flaws with the presented plan: namely an unfeasible last step, and an inefficient and illogical reaction sequence in preparation for a coupling reaction.
• Figure 4: Mechanisms elucidation a) Requirements and impact of mechanistic elucidation. b) Example of actions in our mechanism framework displayed on example structures, with a post-processing interpretation. c) Example task broken down into moves compliant with our mechanism game framework. d) Model separation performance, averaged across the full mechanism in each case, with n=5 repetitions. e) Comparison of global performance with and without guidance prompts. f) All starting reactants and goal products of our 12 tasks, grouped by category.