OpenAI describes a near-autonomous workflow in which GPT‑5.4 worked with Molecule.one’s Maria agent and a high-throughput laboratory to improve a challenging Chan–Lam coupling used to form carbon–nitrogen bonds. The project focused on coupling primary sulfonamides with boronic acids, a substrate class that has historically given low yields.
How the system worked
Scientists used prompts and a harness to have GPT‑5.4 generate and rank research proposals. Humans selected four top proposals for testing. Maria translated the chosen plans into lab instructions, ran experiments at microliter scale, and returned structured data that GPT‑5.4 analyzed to propose follow-up experiments. OpenAI describes the workflow as near-autonomous: human chemists provided high-level steering, corrected experimental details, prepared reagents, and repeated key experiments by hand.
Results
The proposal labeled OAI-M1-03 identified TEMPO as a mild oxidant that could improve the primary sulfonamide Chan–Lam coupling. Across two cycles Maria Lab ran 10,080 reactions. Under the optimized conditions, yields improved for 88% of the boronic acids and 83% of the sulfonamides tested. The mean yield increased from 16.6% to 25.2%, and the proportion of reactions above 30% yield rose from 15.6% to 37.5%.
OpenAI reports that human chemists validated representative reactions at bench scale: 11 of 14 substrate pairs showed higher yields, and most of those increases were more than twofold. A follow-up finding was that TEMPO could be substituted with a cheaper analog, 4-hydroxy-TEMPO, with little loss in performance. External reviewers assessed the preprint as novel and worth sharing.
Limitations and next steps
OpenAI notes several limits: the result does not demonstrate that AI can run an entire chemistry research program independently, the workflow depended on specialized high-throughput infrastructure, and the method is not yet shown to generalize to other reactions, substrate classes, or manufacturing conditions. Yield estimates derive from a microliter-scale platform and bench validation covered 14 substrate pairs.
The reported timeline ran from March 4 to June 4. OpenAI states the immediate next steps are broader substrate testing, investigation of the additive’s mechanism, mapping where the effect succeeds or fails, and independent replication. Tim Cernak, Associate Professor of Medicinal Chemistry, University of Michigan, described the work as a demonstration of combining high-throughput experimentation and modern AI for reaction discovery.
Original source: OpenAI News