Preprint
Reinforcement Learning

Strategy-first synthesis planning for complex natural products

Daniel Armstrong, Xuan-Vu Nguyen, Octavian Susanu, Gabriel Gibberd, Théo A. Neukomm, Taddäus Strunden, Dan Forster, Morgane Delattre, Shawn Teh, Clément Rols, John Federice, Hayden Leatherwood, M. Lavelle Barnes, Maarten R. Dobbelaere, Peter Wipf, Jon T. Njardarson, Jieping Zhu, Philippe Schwaller
August 7, 2026
arXivPDF

0

Citations

0

Influential Citations

—

Venue

2026

Year

Abstract

The total synthesis of a complex molecule is among the most demanding intellectual and experimental feats in chemistry: a chemist must plan many steps ahead for how to assemble simple building blocks into an intricate target, devise backup strategies, and anticipate procedural challenges. It is also a profoundly creative activity. For half a century, efforts to automate the retrosynthetic design of natural products and other complex molecules have drawn on catalogued reactions, and the resulting tools now report near-complete success on benchmarks built from that same source. But these tools were shaped to fit benchmarked chemistry, and they falter on many natural products, the frontier of the field, whose densely functionalized, polycyclic architectures demand precisely the inventive chemistry the record contains least. Whether a machine could reasonably design such syntheses like an expert chemist does has remained unclear. Here, we show that SynthEx, an agentic framework built on large language models, plans routes to complex natural products that lie beyond the reach of conventional design algorithms. SynthEx proposes competing strategies, assembles a sequence of routine and key steps into a cohesive route, and critiques and improves its own design; the chemistry it favours is more convergent than existing tools produce, and spans a region of reaction space that catalogue-based tools cannot match. Most notably, in blinded assessments, expert chemists judged its key steps comparable to those of published human syntheses and engaged with them as genuine synthesis plans, a response algorithmic route prediction has not previously accomplished. We release routes to more than a thousand natural products as SynthAtlas, an open, interactive database, and anticipate it will become a shared resource for a collection of complex target molecules that lack existing literature routes.

Analysis

Why This Paper Matters

For decades, computer-aided retrosynthesis has been benchmarked on known chemistry, with tools achieving near-perfect performance on datasets derived from the same reaction databases they were trained on. However, these tools fail on complex natural products—the frontier of synthetic chemistry—whose densely functionalized, polycyclic architectures demand inventive, non-catalogued reactions. This paper addresses that gap by introducing SynthEx, an agentic framework built on large language models (LLMs) that plans routes to complex natural products beyond the reach of conventional algorithms.

The significance lies in the blinded expert evaluation: for the first time, expert chemists judged algorithmic key steps as comparable to published human syntheses and engaged with them as genuine synthesis plans. This suggests that LLM-based agents can capture the creative, strategic aspects of retrosynthetic design, not just pattern matching from known reactions. The release of SynthAtlas, an open database of routes to over a thousand natural products without existing literature routes, provides a valuable resource for the community.

Technical Contributions

  • Agentic LLM framework: SynthEx uses LLMs to propose competing strategies, assemble sequences of routine and key steps, and critique/improve its own designs, mimicking an expert chemist's iterative process.
  • Beyond catalogue-based tools: By leveraging LLMs, SynthEx explores a broader reaction space, including inventive chemistry not present in traditional reaction databases.
  • Convergent synthesis design: The routes generated are more convergent than existing tools, indicating better strategic planning.
  • SynthAtlas database: An open, interactive database of routes to over a thousand natural products, filling a gap for targets lacking literature routes.

Results

The paper reports that in blinded assessments, expert chemists judged SynthEx's key steps comparable to those of published human syntheses. This is a qualitative result, but it marks a significant milestone in algorithmic route prediction. Additionally, SynthEx's chemistry is more convergent than existing tools and spans a region of reaction space that catalogue-based tools cannot match. The release of SynthAtlas provides routes to more than a thousand natural products, though no quantitative success rates or experimental validations are provided.

Significance

This work demonstrates that LLM-based agents can achieve a level of creativity and strategic planning in retrosynthesis that was previously thought to be uniquely human. It opens new avenues for AI-driven drug discovery and natural product synthesis, potentially reducing the time and cost of developing synthetic routes. The SynthAtlas database will serve as a shared resource for researchers, and the methodology could be extended to other domains requiring multi-step planning and creative problem-solving. However, the lack of experimental validation and quantitative benchmarks means further work is needed to establish practical utility.

large-language-modelsreinforcement-learningretrosynthesischemistryagentic-frameworknatural-products
Large Language ModelsBenchmarks

Related Papers