Chemists Can Now Design Molecules by Describing Them to AI
Researchers at EPFL have developed a system that lets chemists use plain language to guide artificial intelligence through the process of designing new molecules. The tool, called Synthegy, combines traditional chemical software with large language models to help scientists plan synthesis routes and understand reaction mechanisms.
Designing new molecules has always required years of expertise. Chemists must work backward from a target compound-a process called retrosynthesis-to identify simpler starting materials and viable reaction pathways. The strategic choices involved, such as when to form rings or protect sensitive parts of a molecule, demand the kind of judgment that computers have struggled to replicate.
How Synthegy Works
Rather than generating chemical structures directly, Synthegy uses language models as evaluators that guide existing computational systems. A chemist describes what they want in everyday language-for example, "form this ring early" or "avoid unnecessary protecting groups." Standard retrosynthesis software then generates possible pathways.
Synthegy converts each pathway into text and scores how well it matches the chemist's instructions. The language model explains its reasoning, making it easier to rank and filter the best routes. This approach lets researchers iterate faster and explore more complex ideas than traditional tools with rigid filters and rules.
The system applies the same method to reaction mechanisms, which describe how reactions proceed through the movement of electrons. Synthegy breaks reactions into basic steps and steers the search toward pathways that make chemical sense.
Validation From Working Chemists
In a double-blind study, 36 chemists reviewed 368 different synthesis pathways. Their assessments agreed with Synthegy's results 71.2% of the time on average. The system accurately flagged unnecessary protecting steps, judged reaction feasibility, and prioritized efficient solutions.
Larger language models performed better than smaller ones, suggesting that model size matters for this type of chemical reasoning.
A Different Role for AI in Science
Synthegy represents a shift in how AI supports chemistry. Instead of replacing human decision-making, the system interprets and refines computational results based on what chemists actually want to do. This approach could accelerate drug discovery, improve reaction design, and make advanced tools more accessible to researchers who lack deep expertise in retrosynthesis planning.
The work also bridges two areas of chemistry that researchers usually treat separately: synthesis planning and reaction mechanisms. By using a unified natural language interface, Synthegy connects these domains computationally.
Learn more about how large language models work as reasoning tools or explore AI applications in scientific research.
Your membership also unlocks:
