OpenAI says AI chemist improved a difficult medicinal chemistry reaction

OpenAI said an AI chemist improved a difficult medicinal chemistry reaction by pushing a stubborn Chan-Lam coupling past a long-standing yield barrier, then holding up under human validation. Across two microscale screening campaigns totaling 10,080 reactions, the system raised mean estimated product yield from 16.6% to 25.2% and increased the share of reactions above 30% yield from 15.6% to 37.5%.

The work paired GPT-5.4 with Maria, Molecule.one’s agentic chemistry system connected to a high-throughput laboratory in Maria Lab. OpenAI said the system did more than suggest ideas: it generated research proposals, designed and ran experiments, analyzed the data and proposed follow-up tests. Human chemists remained in the loop, steering the project, choosing among proposals and validating the final result at the bench.

The reaction chosen for the demo was a challenging version of Chan-Lam coupling, a carbon-heteroatom bond-forming method that is useful in medicinal chemistry but notoriously sensitive to substrate choice. The specific target was direct N-arylation of primary sulfonamides, which the paper described as a difficult disconnection because these compounds are low in nucleophilicity and the reaction can be prone to over-arylation and boronic-acid degradation. OpenAI said the system identified primary sulfonamides as a high-value but difficult substrate class and suggested mild oxidants such as TEMPO as a path to better performance.

That hunch paid off in the screen. The optimized condition used 2 equivalents of TEMPO and 20 mol% Cu(OAc)2, while 4-hydroxy-TEMPO performed similarly and may offer a lower-cost, easier-to-remove alternative. The paper said yields improved for 88% of the boronic acids and 83% of the sulfonamides tested. When human chemists repeated representative reactions at bench scale, they confirmed higher yields for 11 of 14 substrate pairs, with most showing more than a twofold increase.

The result matters because medicinal chemistry still depends on reactions that work reliably across many substrates, not just in carefully chosen examples. OpenAI has argued in earlier science-focused writing that even when the right idea exists, moving from concept to tested result can take years. This collaboration was aimed at that bottleneck: turning model-generated hypotheses into actual chemistry, then checking whether the chemistry survived outside the screening plate.

The joint paper listed a long author team spanning OpenAI and Molecule.one, including Jan Rzymkowski, Shuyuan Zhang, Artur Chołuj, Aleksander Szkółka, Mateja Dud, Mateusz Bruno-Kamiński, Jan Busz and Michał Sadowski. For OpenAI and Molecule.one, the claim was not that AI replaced chemists, but that it narrowed the gap between proposing a reaction and proving one in the lab.