This AI 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 knowing how to connect atoms – it’s knowing how to do it properly, when to protect the dangerous parts of a molecule, and how to avoid dead ends that can waste months of lab work.

Traditionally, that knowledge resides in the heads of medical professionals. Now, the EPFL team wants to put it in the language of the language.

Researchers led by Philippe Schwaller published a paper this week in Matter to describe Synthegy, a framework that uses large-scale speech models as reasoning engines for chemical synthesis planning. The key insight is subtle but fundamental: instead of asking AI to create molecules, the team uses AI to test combinations that traditional software already creates.

Here’s how it works: A chemist puts a goal in plain English, like “make the pyrimidine ring first.” Existing retrosynthesis software, which works by breaking down target molecules into simpler pieces, then creates several or hundreds of steps.

Synthegy converts each method into text and submits it to LLM, which evaluates each method as it matches the chemist’s instructions. The best ones float to the top, and I’ll explain why.

“In the design of machine tools, the user interface is very important, and previous tools relied on filters and complex rules,” said Andres M. Bran, the lead author of the study. words from EPFL.

The system was confirmed in a double-blind study involving 36 independent experts who reviewed 368 methods. Their decisions corresponded to Synthegy’s 71.2% of the time, a figure that is in line with the consensus of medical experts. Senior researchers (professors and research scientists) collaborated with Synthegy more often than PhD students, meaning that the system also captures the processes that come with the experience.

The researchers tested several AI models, including GPT-4o, Claude, and DeepSeek-r1. AI has been around intervention in drug discovery for many years, but most methods focus on well-trained models for specific tasks. Synthegy is designed to be modular – it can connect any retrosynthesis engine on the back, and any LLM capable on the logic side. Gemini-2.5-pro performed very well in the benchmark, while DeepSeek-r1 appears to be an open source solution that can run locally.

This framework also has a second problem: the analysis of processes. This is the question of why chemical reactions occur—what electron transport occurs at each step. Synthegy divides the reactions into basic components and enables the LLM to analyze each component in order to present it as a product. For simple reactions such as nucleophilic substitution, the best models that have been achieved are almost correct.

The cases in which it can be used are broad. Access to drugs is obvious. AI already has it showing promise predicting the outcome of cancer treatment, but the same approach applies wherever scientists need to develop new tools or optimize industrial processes. One useful detail: testing 60 methods with Synthegy takes about 12 minutes and costs about $2–3 in API fees.

This paper acknowledges the present limitations. LLMs sometimes misjudge how they respond to their pitches, resulting in the wrong call. Small samples do no better than random guesswork. Steps longer than 20 steps are difficult to follow coherently.

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