News

AI Chemist Points to a More Automated Drug Discovery Loop, With Humans Still in Control

• By John K. Waters
• 06/22/2026

OpenAI and Molecule.one reported a step toward automated drug-discovery research this week, saying an AI system connected to a high-throughput chemistry laboratory improved a difficult reaction used in medicinal chemistry.

The companies described the system as a “near-autonomous AI chemist,” pairing OpenAI’s GPT-5.4 model with Molecule.one’s Maria, an agentic chemistry AI integrated with Maria Lab. OpenAI said the system reviewed literature, generated and ranked research proposals, designed and ran experiments, analyzed data, and suggested follow-up work.

The experiment focused on Chan-Lam coupling, a reaction used to form carbon-nitrogen bonds. OpenAI said GPT-5.4 identified primary sulfonamides as a challenging substrate class and proposed using mild oxidants, including TEMPO, to improve yields.

According to OpenAI, Maria Lab ran 10,080 reactions across two experimental cycles. Under optimized conditions, measured yields improved for 88% of tested boronic acids and 83% of tested sulfonamides. The mean yield rose from 16.6% to 25.2%, and the share of reactions with yields above 30% increased from 15.6% to 37.5%.

Human chemists then repeated representative reactions at bench scale. OpenAI said those experiments showed higher yields for 11 of 14 substrate pairs, with more than a twofold increase in eight cases.

The result is potentially significant because chemical synthesis remains a practical bottleneck in small-molecule drug discovery. Researchers can only test molecules they can synthesize or obtain, and low-yielding reactions can limit the compounds explored. OpenAI said the sulfonamide group appears in medicines across oncology, infectious disease, and other therapeutic areas.

But the company also cautioned against overstating the result. OpenAI said the workflow was “near-autonomous, not fully autonomous,” because human chemists still made important decisions. Those roles included writing, steering, and grading prompts; selecting proposals for laboratory testing; correcting some experimental details; helping with lab operations; and validating results by hand.

The distinction matters. The work does not show that AI can run a chemistry research program end-to-end. OpenAI said the workflow depended on specialized high-throughput infrastructure, expert oversight, and constrained experiments. It also said the result does not establish that the method will generalize to other coupling reactions, other substrate classes, or manufacturing conditions.

OpenAI said the project took three months, from the first prompt on March 4 to sharing results with independent experts on June 4. Four external chemistry experts reviewed the preprint, according to the company. Tim Cernak, associate professor of medicinal chemistry at the University of Michigan, said in the OpenAI post that the combination of high-throughput experimentation and modern AI represents “a new frontier of scientific discovery.”

OpenAI also discussed safety limits around the work. The company said the project was scoped to a legitimate medicinal-chemistry problem and did not involve toxins, chemical weapons, or requests to design harmful compounds. It said GPT-5.4 had undergone relevant evaluations with the UK AI Security Institute, and that human chemists retained control over which proposals entered the lab and over the physical infrastructure.

For AI drug-discovery companies and pharmaceutical researchers, the larger question is whether systems like Maria can shorten the cycle from hypothesis to experiment, analysis, and follow-up testing. The work by OpenAI and Molecule.one points toward that possibility, but it also shows why experimental chemistry is harder to automate than information work. Molecules must still be made, measured, reproduced, and judged by scientists.

OpenAI said the next steps include testing a broader range of starting materials, investigating why the additives improved the reaction, mapping where the effect works and fails, and supporting independent replication.

The practical takeaway is not that AI has replaced chemists, but that it may be starting to handle more of the experimental loop under expert supervision. In drug discovery, that could matter if it helps researchers explore additional synthetic routes, eliminate weaker ideas more quickly, and surface useful reactions that might otherwise be overlooked.