Designing a molecule from scratch is one of chemistry’s toughest engineering problems: it’s not just which atoms to join, but the sequence of reactions, when to shield fragile groups, and how to avoid dead ends that can waste months. A team at EPFL wants to outsource that strategic know-how to a language model.
This week in Matter, Philippe Schwaller and colleagues unveiled Synthegy, a framework that uses large language models (LLMs) as reasoning engines for synthesis planning. The twist: instead of asking AI to invent molecules, Synthegy asks AI to judge and rank the candidate synthesis routes that conventional retrosynthesis software already proposes.
How it works
- A chemist types a plain-English constraint (for example, “form the pyrimidine ring in the early stages”).
- Retrosynthesis tools generate dozens or hundreds of possible routes by breaking the target molecule into simpler building blocks.
- Synthegy translates each route into text and hands those descriptions to an LLM, which scores every route against the chemist’s instruction and produces human-readable explanations.
The result: the most strategically sensible routes float to the top, plus a rationale for each choice. “When making tools for chemists, the user interface matters a lot, and previous tools relied on cumbersome filters and rules,” says lead author Andres M. Bran.
Validation and benchmarks
Synthegy was validated in a double-blind study with 36 independent chemists who reviewed 368 route pairs. The chemists chose the same routes Synthegy ranked highest 71.2% of the time — roughly on par with inter-expert agreement. Senior researchers matched the system more often than PhD students, suggesting the model captures high-level strategic instincts that come with experience.
The team trialed multiple models, including GPT-4o, Claude, and the open-source DeepSeek-r1. Gemini-2.5-pro topped the benchmark, while DeepSeek-r1 emerged as a strong local, open alternative. Synthegy is deliberately modular: any retrosynthesis engine can sit on the backend, and any capable LLM can power the reasoning layer.
Beyond route ranking: mechanism checking
Synthegy also tackles reaction mechanism elucidation. It breaks reactions into elementary electron-movement steps and asks the LLM to assess each step’s chemical plausibility. On straightforward reaction types like nucleophilic substitutions, top models achieved near-perfect accuracy.
Practical numbers and caveats
- Speed/cost: evaluating ~60 candidate routes takes about 12 minutes and costs roughly $2–3 in API fees.
- Limits: LLMs sometimes misread reaction direction from text, leading to wrong feasibility calls; small models perform like random guessing; and routes longer than ~20 steps become hard to track coherently.
Open science and implications
The code and benchmarks are publicly available at github.com/schwallergroup/steer. With its composable design and explainable outputs, Synthegy could accelerate drug discovery, materials design, and industrial process optimization — essentially serving as an “oracle” that ranks feasible chemical paths and explains why, reducing trial-and-error in the lab. For a field where intuition often lives only in senior scientists’ heads, that’s a practical power-up.
Read more AI-generated news on: undefined/news
