This artificial intelligence reads your chemistry instructions and finds the best way to build a molecule for you

Designing a molecule from scratch is one of the most difficult problems in chemistry. It’s not just about knowing which atoms to connect, it’s about knowing the correct order of reactions, when to protect sensitive parts of the molecule, and how to avoid dead-ends that can ruin months of laboratory work.

Traditionally, this knowledge lives in the heads of experienced chemists. Now, a team from EPFL wants to put it into a linguistic model.

Researchers led by Philip Schwaler Published paper This week in Matter describes Synthegy, a framework that uses large language models as reasoning engines to plan chemical synthesis. The basic idea is subtle but important: Instead of asking AI to create molecules, the team is using AI to evaluate synthesis pathways already produced by traditional software.

Here’s how it works: A chemist writes an objective in plain English, such as “early-stage pyrimidine ring formation.” Current retrosynthesis software, which works by breaking down target molecules into simpler parts, then generates dozens or hundreds of potential synthesis routes.

Synthegy converts each track to text and delivers it to LLM, which scores each track according to how well it matches the chemist’s instructions. The best ones float to the top, with written explanations of why.

Andres M. said: Brann, lead author of the study, said in his article: “When making tools for chemists, the user interface is of great importance, and previous tools relied on cumbersome filters and rules.” statement From EPFL.

The system was validated in a double-blind study involving 36 independent chemists who reviewed 368 pairs of pathways. Their choices matched Synthegy 71.2% of the time, a number roughly in line with the number of times expert chemists agree with each other. Senior researchers (professors and research scientists) agreed with Synthegy more often than doctoral students, suggesting that the system captures the same strategic intuition that comes with experience.

The researchers tested several AI models, including GPT-4o, Claude, and DeepSeek-r1. It was artificial intelligence Achieving successes has been used in drug discovery for years, but most approaches focus on models narrowly trained for specific tasks. Synthegy is designed to be modular, connecting to any recompiler engine on the backend, and any LLM capable engine on the logic side. Gemini-2.5-pro scored high in the benchmark, while DeepSeek-r1 appears to be a solid open source alternative that can run locally.

The framework also addresses a second problem: clarifying the reaction mechanism. This is the question of why a chemical reaction occurs, i.e. the electron movements that occur in each step. Synthegy breaks down reactions into elementary kinetics and LLM evaluates each candidate step for chemical plausibility. In simple interactions such as nucleophilic substituents, the best models achieved near-perfect accuracy.

The potential use cases are wide-ranging. Drug discovery is the obvious thing. Amnesty International has already done so The clear promise predicting cancer treatment outcomes, but the same approach applies anywhere chemists need to design new materials or improve industrial reactions. One practical detail: Evaluating 60 candidate paths with Synthegy takes about 12 minutes and costs about $2-3 in API fees.

The paper acknowledges current limitations. Sometimes, MBAs misread the reaction trend into its text representation, leading to wrong feasibility conclusions. Smaller models perform no better than random guessing. Routes longer than 20 steps are difficult to follow coherently.

The code and standards are publicly available at github.com/schwallergroup/steer.