Chemists can articulate their objectives in plain language and obtain solutions that resonate with their strategic vision.
For generations, the art of chemical synthesis has lived in the hands and minds of patient experts who could sense, almost intuitively, which molecular pathways were worth pursuing. Researchers at EPFL have now built a system called Synthegy that honors that intuition rather than displacing it — using large language models to translate a chemist's strategic reasoning into computational guidance, achieving agreement with expert judgment nearly three-quarters of the time. The development suggests that the most productive role for artificial intelligence in science may not be to think for us, but to help us think more clearly about what we already know we want.
- The gap between a computer's ability to generate thousands of chemical pathways and a chemist's ability to recognize which ones are actually worth pursuing has long stalled the promise of AI-assisted synthesis.
- Existing software floods researchers with plausible-but-undifferentiated options, leaving the hardest intellectual work — distinguishing the promising from the merely possible — entirely to human experts already stretched thin.
- EPFL's Synthegy reframes the AI's role entirely: instead of generating chemistry, it listens to the chemist's plain-language instructions and scores computational outputs against those stated priorities.
- A double-blind validation with 36 chemists evaluating 368 synthesis routes found Synthegy aligned with expert judgment 71.2% of the time — a result strong enough to signal genuine chemical reasoning, not pattern mimicry.
- The system now points toward a future where drug discovery and materials research accelerate not because machines work harder, but because humans can articulate their intentions more precisely and be heard.
Chemistry has always demanded a particular kind of patience — the ability to envision a finished molecule and then reason backward through every step required to build it. This process, called retrosynthesis, draws on years of accumulated intuition: which building blocks will cooperate, which parts of a molecule need protection, which routes are elegant and which are merely possible. Computers can search vast libraries of reactions, but they have long lacked the chemist's sense of what is actually worth pursuing.
At EPFL in Switzerland, Philippe Schwaller's team decided to work with that intuition rather than around it. Their system, Synthegy, published in the journal Matter, uses large language models not to generate chemical structures but to evaluate them — acting as a translator between what a chemist wants and what the computational search produces. The insight is straightforward: chemists think and decide in language, so an AI that understands language can help bridge strategy and search.
In practice, a chemist specifies a target molecule and adds plain-English instructions — form this ring early, skip the protecting groups here. Traditional software generates the candidate routes; Synthegy converts each into text, scores it against the chemist's stated preferences, and explains its reasoning. The system also handles reaction mechanism analysis, breaking reactions into their electron-movement fundamentals and steering the search toward chemically sound options.
Validation came through a double-blind study in which 36 chemists assessed 368 synthesis routes. Synthegy's judgments aligned with theirs 71.2% of the time — a strong signal that the system was capturing something genuine about chemical reasoning rather than approximating it superficially. Larger language models outperformed smaller ones, suggesting that nuanced scientific judgment rewards scale.
What distinguishes Synthegy is its restraint. It does not attempt to replace the chemist; it amplifies the chemist's capacity to navigate complexity. By unifying synthesis planning and reaction mechanism analysis through a single natural language interface, it points toward a future where the limiting factor is no longer computational search, but the human ability to clearly articulate what matters — and to be understood.
Chemistry has always been a discipline of patience and accumulated wisdom. A chemist tasked with synthesizing a new drug or material must envision the final molecule, then reverse-engineer the path to get there—a process called retrosynthesis. It demands years of training, intuition about which building blocks will cooperate, and judgment about when to protect vulnerable parts of a molecule from unwanted reactions. Computers can now sift through vast libraries of possible chemical reactions, but they lack something harder to quantify: the seasoned chemist's sense of which pathways are actually worth pursuing.
This gap between computational power and chemical intuition has long frustrated researchers. A second, related problem compounds the difficulty. Understanding how reactions actually work—the precise choreography of electrons moving from one atom to another—allows scientists to predict new reactions and avoid expensive dead ends. Yet existing software can generate hundreds of plausible pathways without the deeper reasoning needed to distinguish the promising ones from the merely possible.
At EPFL in Switzerland, a team led by Philippe Schwaller decided to approach the problem differently. Rather than asking artificial intelligence to generate chemical structures from scratch, they built a system that uses large language models as evaluators—tools that could understand what a chemist actually wanted and help filter the noise. The result, published in the journal Matter, is called Synthegy. The key insight is simple but powerful: chemists think in language. They make decisions based on strategic reasoning that can be expressed in words. Why not let an AI system that understands language help translate those decisions into chemical action?
Andres M Bran, the lead author on the Synthegy paper, describes the interface this way: a chemist specifies a target molecule and adds instructions in plain English. Form this ring early. Skip the protecting groups here. The traditional retrosynthesis software then generates dozens or hundreds of possible routes. Synthegy translates each one into text, scores how well it matches the chemist's stated preferences, and explains its reasoning. The chemist can then rank and filter the results, iterating much faster than before. The system also works backward from reaction mechanisms, breaking reactions into their fundamental electron movements and using the language model to steer the search toward chemically sound possibilities.
The validation came through a double-blind study. Thirty-six chemists evaluated 368 different synthesis routes. On average, Synthegy's assessments agreed with the chemists' judgments 71.2 percent of the time—a strong signal that the system was capturing something real about chemical reasoning. Larger language models outperformed smaller ones, suggesting that scale matters when the task involves nuanced judgment across a complex domain.
What makes Synthegy notable is what it does not do. It does not replace the chemist. Instead, it amplifies the chemist's ability to think through complex problems by serving as a translator between human strategy and computational search. A researcher can say, in effect, "Here's what I'm trying to achieve, here's what I want to avoid," and the system helps navigate the vast space of possibilities toward solutions that align with that vision. The implications ripple outward: faster drug discovery, better material design, and tools that become more useful rather than less as they grow more sophisticated.
Bran notes that the work bridges two traditionally separate domains—synthesis planning and reaction mechanism analysis—through a single natural language interface. That bridge matters because understanding how reactions work is how chemists discover new reactions, which in turn enables new molecules. By unifying those two processes computationally, Synthegy points toward a future where the bottleneck is no longer the computer's ability to search, but the human's ability to articulate what matters.
Citações Notáveis
With Synthegy, we're giving chemists the power to just talk, allowing them to iterate much faster and navigate more complex synthetic ideas.— Andres M Bran, lead author of Synthegy publication
Our work is bridging that gap computationally through a unified natural language interface.— Andres M Bran, on connecting synthesis planning and reaction mechanisms
A Conversa do Hearth Outra perspectiva sobre a história
So the system doesn't actually design molecules. It just talks to chemists?
Right. It listens to what a chemist wants and helps them navigate the options that already exist. The real synthesis software still does the heavy lifting—generating pathways. Synthegy just makes sense of them.
Why does that matter? Couldn't a chemist just read through the pathways themselves?
Hundreds of them? In a reasonable timeframe? The system doesn't just rank them—it explains why each one does or doesn't fit what you asked for. It's like having someone who understands both chemistry and your specific constraints working through the options with you.
The 71 percent agreement with chemists—is that good?
It's solid. It's not perfect, but it's high enough that chemists trust it as a thinking partner. And remember, those chemists were evaluating in a blind study. They weren't biased toward agreeing with the system.
What happens when the system disagrees with a chemist?
That's actually useful. It forces the chemist to think about why. Maybe the system caught something they missed, or maybe they have domain knowledge the system doesn't. Either way, it's a conversation, not a verdict.
Does this speed things up?
Significantly. Instead of spending weeks or months exploring dead ends, a chemist can iterate through strategic ideas in hours. They're not waiting for trial-and-error experiments to fail. They're thinking faster.