Chemists can just talk, iterate faster, navigate complex ideas
For as long as chemistry has been practiced, the translation of human intention into molecular reality has demanded years of hard-won intuition. A new system called Synthegy, developed at EPFL, now allows chemists to describe their goals in plain language and receive AI-ranked synthesis pathways in return — not to replace the expert, but to serve as a bridge between human reasoning and computational possibility. In a field where a single misstep in planning can consume months of labor, the significance lies less in what the machine does alone and more in what it makes possible together with the scientist.
- Designing molecules has always required a rare fusion of computational power and chemical intuition that few tools have managed to honor simultaneously.
- Synthegy disrupts the old paradigm by letting chemists issue instructions in everyday language — 'avoid protecting groups,' 'keep temperatures low' — rather than wrestling with rigid rule-based filters.
- The system converts candidate synthesis pathways into text, scores them against the chemist's stated goals, and returns a ranked list with transparent reasoning — turning opacity into dialogue.
- In a double-blind validation with 36 chemists evaluating 368 synthesis plans, Synthegy's rankings matched expert assessments 71.2% of the time, a result that signals genuine utility without overstating autonomy.
- The trajectory points toward faster drug discovery and a lowering of barriers for less experienced researchers, with the deeper implication that language models may reshape AI's role across many technical disciplines.
For decades, designing a new molecule has been a craft built on years of apprenticeship. A chemist works backward from a target compound, asking which simpler pieces can be assembled, which bonds should form first, and where vulnerable parts of a structure need protection. Wrong choices early can waste months. Now a system called Synthegy is letting chemists describe what they want in plain English and have an AI evaluate the options.
The challenge Synthegy addresses is fundamental. Retrosynthesis — decomposing a target molecule into buildable precursors — requires both computational speed and chemical intuition. A computer can enumerate thousands of pathways in seconds but cannot judge which ones a real chemist would actually attempt, nor weigh the cost of protecting sensitive molecular groups against the elegance of an early-stage ring formation. Understanding reaction mechanisms poses a related problem: current tools, for all their speed, often miss the chemical logic that guides an expert toward the most plausible path.
Researchers at EPFL, led by Philippe Schwaller, built Synthegy by deploying large language models not as generators of chemical structures — a task at which they are unreliable — but as evaluators and reasoners. A chemist types an instruction in everyday language. Standard retrosynthesis software generates hundreds of candidate pathways. Synthegy converts each into text, asks the language model to score how well it matches the chemist's goals, and returns a ranked list with justifications. First author Andres M Bran notes that the interface itself matters: previous tools demanded navigation of cumbersome filters, while Synthegy lets chemists simply talk and iterate faster through complex synthetic ideas.
Validation came from a double-blind study in which 36 chemists evaluated 368 synthesis plans. Their assessments agreed with Synthegy's rankings 71.2% of the time — a result that demonstrates the system can identify unnecessary steps, judge feasibility, and prioritize efficient solutions, with larger language models outperforming smaller ones.
What makes the work significant is its restraint. Synthegy does not replace the chemist or generate novel molecules from scratch. It positions the language model as a guide — a translator between human intention and algorithmic output, with the chemist remaining the strategist. The implications extend from accelerating drug discovery to making molecular design more accessible to less experienced scientists, and to a broader insight: that language models can reason meaningfully about complex technical domains when given the right role.
For decades, designing a new molecule has been a craft learned through years of apprenticeship. A chemist stares at a target compound—the drug candidate, the material, the thing that doesn't yet exist—and works backward, asking: What simpler pieces could I snap together to build this? Which bonds should form first? Where do I need to shield vulnerable parts of the structure from unwanted reactions? These decisions cascade. A wrong choice early on can waste months of lab work. Now a system called Synthegy is letting chemists describe what they want in plain English and have an AI system evaluate the options for them.
The problem Synthegy addresses is old and fundamental. Retrosynthesis—the art of decomposing a target molecule into buildable precursors—requires both computational power and chemical intuition. A computer can enumerate thousands of possible reaction pathways in seconds. But it cannot easily judge which ones a real chemist would actually attempt. It cannot weigh the cost of protecting sensitive molecular groups against the time saved by a clever early-stage ring formation. It cannot read between the lines of a synthetic strategy the way an experienced researcher can.
There is a second, related challenge: understanding reaction mechanisms. These are the step-by-step electron movements that explain how a reaction actually proceeds. Knowing the mechanism lets a chemist predict whether a reaction will work, optimize it, or even discover entirely new transformations. Yet current computational tools, for all their speed, often miss the chemical logic that would guide a human expert toward the most plausible pathway.
Researchers at EPFL, led by Philippe Schwaller, built Synthegy by treating large language models not as generators of chemical structures—a task at which they are unreliable—but as evaluators and reasoners. The system works like this: a chemist types an instruction in everyday language. "Form this ring early." "Avoid protecting groups where possible." "Keep the reaction temperature below 50 degrees Celsius." Standard retrosynthesis software then generates hundreds of candidate pathways. Synthegy converts each one into text and asks the language model to score how well it matches the chemist's stated goals. The model explains its reasoning. The chemist sees a ranked list of options, each with a justification.
Andres M Bran, the first author of the paper published in Matter, emphasizes that the interface itself matters enormously. Previous computational tools for chemistry demanded that users navigate cumbersome filters and rule systems. "With Synthegy, we're giving chemists the power to just talk," Bran says. "They can iterate much faster and navigate more complex synthetic ideas." The system applies the same logic to reaction mechanisms, breaking them into electron movements, exploring possibilities, and steering the search toward pathways that make chemical sense.
Validation came from a double-blind study in which 36 chemists evaluated 368 different synthesis plans. On average, the chemists' assessments agreed with Synthegy's rankings 71.2 percent of the time. The system could identify unnecessary protecting steps, judge reaction feasibility, and prioritize efficient solutions. Larger language models performed better than smaller ones, suggesting that scale matters for chemical reasoning.
What makes this work significant is what it does not claim to do. Synthegy does not replace the chemist. It does not generate novel molecules from scratch. Instead, it positions the language model as a guide—a tool that helps interpret and refine what computational chemistry already produces. The chemist remains the strategist. The AI becomes a translator between human intention and algorithmic output.
The implications ripple outward. Drug discovery could accelerate if researchers can explore synthetic routes faster. Molecular design tools could become more accessible to scientists without decades of experience. And the underlying insight—that language models can reason about complex domains when given the right role—may reshape how AI supports other technical fields. Bran notes that synthesis planning and mechanism discovery have always been connected: mechanisms reveal new reactions, which enable new syntheses. "Our work is bridging that gap computationally through a unified natural language interface."
Notable Quotes
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, first author of the Synthegy paper
Our work is bridging that gap computationally through a unified natural language interface, connecting synthesis planning and mechanism discovery.— Andres M Bran
The Hearth Conversation Another angle on the story
Why does a chemist need to work backward from the molecule they want? Why not just build forward?
Because forward synthesis is combinatorially explosive. There are too many possible starting materials and reaction sequences. Working backward from the target constrains the problem—you're asking, what are the fewest steps to get here? But even that is hard because each step opens multiple branches.
And the language model is just... scoring the branches?
Yes, but that's the key insight. The computer can generate thousands of pathways. The language model reads each one as a narrative and judges whether it aligns with what the chemist actually wants. It's not generating chemistry; it's understanding strategy.
The 71.2 percent agreement rate—is that good?
It's surprisingly good for a first system. You're comparing an AI to expert human judgment on complex, subjective decisions. Perfect agreement would be suspicious. What matters is that the system is reliable enough to narrow the search space and save the chemist time.
Could this eventually replace chemists?
Not in the way people worry about. The chemist still decides what to make and what constraints matter. Synthegy just helps them navigate the vast space of how to make it. If anything, it democratizes the craft—a less experienced chemist can now access the kind of strategic reasoning that used to require years of training.
What happens when the language model gets it wrong?
The chemist sees the reasoning. They can push back, refine their instructions, and try again. It's iterative. The human stays in the loop, which is exactly where they should be in chemistry.