Synthegy AI matches chemists 71.2%

Chemistry software is already good at generating lots of possible synthesis routes. The hard part is picking the one a real chemist would actually want to try. That gap matters because early drug and materials work is full of strategic choices — when to build a ring, what to protect, what to avoid, what will probably fail in the lab. EPFL’s new Synthegy system is interesting because it aims at that judgment layer, not just raw route generation. ### What is Synthegy actually doing? Synthegy is a framework from Philippe Schwaller’s group at EPFL. A chemist gives it a target molecule and a plain-English instruction — things like form a certain ring early or avoid unnecessary protecting groups. Standard retrosynthesis software then proposes many candidate routes. Synthegy turns those routes into text, uses a language model to judge how well each one fits the chemist’s strategy, and then ranks them with an explanation. (actu.epfl.ch) Basically, the LLM is acting more like a senior reviewer than a molecule generator. ### Why is that a different bet from most AI chemistry tools? A lot of AI chemistry work tries to predict the next reaction step or generate brand-new structures directly. Synthegy takes a narrower path. It leaves the heavy search to existing chemistry engines and uses the model to interpret intent and score options. That sounds less flashy, but turns out it solves a very real problem — chemists often do not need 500 more routes, they need help narrowing 500 routes to the five that fit an actual lab strategy. (actu.epfl.ch) ### Where does the 71.2% number come from? The headline result comes from a double-blind expert study. Thirty-six chemists produced 368 valid evaluations of route pairs, and their choices agreed with Synthegy’s rankings 71.2% of the time on average. The important nuance is that this was not framed as “AI beats chemists.” It was framed as alignment with expert judgment — whether the system tends to prefer the same routes humans prefer when given the same strategic goal. (actu.epfl.ch) ### Is 71.2% actually good? For this kind of task, yes — because route selection is not a clean right-or-wrong exam. Expert chemists often disagree with each other too. That means the benchmark is closer to taste plus experience than to arithmetic. So a system landing in the same neighborhood as inter-expert agreement is a stronger result than the raw percentage first suggests. It is less “the AI solved chemistry” and more “the AI is making recognizably chemist-like tradeoffs.” (actu.epfl.ch) ### Why does plain language matter so much? Because chemistry planning tools usually make people express strategy through rigid filters, templates, or handcrafted rules. Synthegy lets chemists state intent directly. That changes the interface from parameter-tuning to conversation. If that holds up in practice, it could make these systems more usable for working chemists, especially in early-stage design where the goal is often fuzzy and changes fast. (decrypt.co) ### Does it only work on retrosynthesis? No — the same paper says the framework also helps with reaction mechanisms. It breaks reactions into elementary electron-movement steps, explores possibilities, and uses language-model reasoning to score which mechanisms make sense. That broadens the story. This is not just a route-ranking widget; it is an attempt to use language models as a chemistry reasoning layer across multiple planning tasks. (actu.epfl.ch) ### What’s the catch? The catch is that Synthegy is still a decision-support tool, not an autonomous chemist. It depends on existing search systems to generate candidates in the first place, and its usefulness rises or falls with the quality of those candidates and the prompts guiding evaluation. Also, matching expert preference is not the same as proving a route will be cheapest, safest, or highest-yielding in every real lab. (actu.epfl.ch) ### Bottom line The real advance here is not that an AI found one magic synthesis route. It is that EPFL showed a language model can sit on top of chemistry software and make route choices that look a lot like expert judgment. If that scales, the bottleneck in synthesis planning may shift from generating options to deciding which human goals matter most. (actu.epfl.ch)