OpenAI and Molecule.one Report a Near-Autonomous AI Chemist That Improved a Stubborn Drug-Making Reaction Across 10,080 Experiments
GPT-5.4 paired with Molecule.one's Maria platform proposed, ran, and analyzed a 10,080-reaction campaign that lifted yields of a hard sulfonamide coupling using the additive TEMPO.
Overview
OpenAI and the chemistry startup Molecule.one have reported a system in which one of OpenAI’s frontier models, GPT-5.4, paired with Molecule.one’s agentic chemistry platform Maria and an automated high-throughput lab, improved a reaction that medicinal chemists rely on but have long struggled to run reliably, according to R&D World. The target was the primary sulfonamide Chan-Lam coupling, described by R&D World as “a copper-mediated reaction used to attach aryl groups to sulfonamide motifs common in drug-like molecules.”
The two organizations framed the work as the first documented case of an AI system independently improving an open-ended organic chemistry problem end-to-end, from literature review through experimental confirmation, according to AI Weekly. The underlying results are described in a preprint titled “TEMPO Improves Generality and Decreases Oxidative Deboronation in Chan–Lam Couplings of Primary Sulfonamides”, co-authored by researchers at Molecule.one and OpenAI.
What We Know
The chemistry problem. Primary sulfonamides are a common building block in drug molecules, but coupling them to aromatic groups through the Chan-Lam reaction has historically returned low yields, limiting the method’s usefulness. The preprint notes that primary sulfonamides “remain challenging substrates for direct Chan–Lam N-arylation” because “the strongly electron-withdrawing sulfonyl group renders them weak and highly polar nitrogen nucleophiles,” according to the preprint.
The scale of the campaign. The team ran two microscale screening campaigns “comprising 10,080 reactions,” evaluating oxidant identity, oxidant loading, copper source and loading, base, solvent, temperature, and substrate structure, per the preprint. The screen covered 12 primary sulfonamides and 8 boronic acids, giving 96 unique substrate pairs, and the authors describe it as “the largest Chan–Lam high-throughput experimentation screen reported to date,” according to the preprint. R&D World reports that “Maria ran 10,080 reactions in total,” as noted by R&D World.
The finding. The key result was that the stable radical additive TEMPO improved the reaction. Under the optimized condition, using 2 equivalents of TEMPO and 20 mol% copper acetate, “the mean estimated product yield increased to 25.2% (from 16.6%), and the fraction of reactions exceeding 30% yield increased over twofold to 37.5% (from 15.6%),” according to the preprint. Across two rounds, R&D World reports that “yields improved for 88% of the boronic acids and 83% of the sulfonamides tested,” per R&D World. The preprint attributes the benefit to TEMPO suppressing “oxidative deboronation,” a side reaction that degrades the boron-containing reagent, according to the preprint.
The authors also report that a related additive, 4-hydroxy-TEMPO, “maintained comparable performance, offering a potentially lower-cost and more readily removable alternative to TEMPO,” noting that its polar hydroxyl group means “it can be easily and cleanly washed out” during workup, per the preprint. Separate bench-scale validation confirmed the beneficial effect of TEMPO “in eleven of fourteen” cases tested, according to the preprint.
How autonomous it was. The companies described the process as near-autonomous rather than fully autonomous, with humans remaining in the loop at several stages. Piotr Byrski, co-founder and CEO of Molecule.one, said that “GPT-5.4 proposed the research topic: primary sulfonamide Chan-Lam coupling, and hypothesis: TEMPO-like mild oxidants will help with yield and side products,” according to R&D World. He added that “Maria turned selected ideas into HTE workflows and ran/analyzed them, GPT then reanalyzed the data and proposed a 2nd cycle,” per R&D World.
Byrski also described the human role: “Scientists using Maria wrote the prompts,” “Humans gave the final go-ahead on experiments, making minor corrections to experimental procedures,” and “At the end, humans ran manual validation and wrote up the manuscript with results,” according to R&D World. AI Weekly characterized the overall division of labor by saying “Human chemists guided and validated but did not drive the research,” per AI Weekly.
Ahmed El-Kishky, an OpenAI researcher, framed the demonstration as a window into “what becomes possible when frontier intelligence is paired with purpose-built scientific agents, automated laboratory infrastructure, and expert chemists,” according to R&D World.
Maria is Molecule.one’s own platform, which the company says “unites AI, Lab, and Data to power near-autonomous research for drug discovery & manufacturing,” according to Molecule.one. The company describes the collaboration as one in which “GPT-5.4 & Maria AI picked the research area, generated proposals, rated them, and ran the experiments in the Maria Lab,” per Molecule.one.
What We Don’t Know
The absolute yields remain modest: even after optimization, the mean estimated product yield was 25.2%, leaving most of the reaction mass unconverted, according to the preprint. How the optimized conditions transfer to substrates outside the 96 tested pairs is not established by the screen alone. Byrski said that while “all protocols, condition tables, and bench-scale experiments are publicly shared,” “there is no plan currently to release raw HTE reaction data,” according to R&D World, which limits independent replication of the full dataset.
Analysis
The headline claim here is narrower than “AI discovers chemistry”: the system did not invent a new reaction, but optimized a known one by re-examining an additive — TEMPO — that the preprint notes was present in the original Chan-Lam work yet largely set aside in later method development. What is notable is the workflow. Rather than a model suggesting a single idea for humans to test, GPT-5.4 set the research direction, Maria translated it into a high-throughput campaign of more than ten thousand reactions, and the model reanalyzed the output to propose a second round, with chemists approving and validating at each gate, according to R&D World.
The practical hook is the TEMPOL finding: a cheaper, easier-to-remove variant performing comparably matters more for process chemistry than a marginal yield bump on a benchtop, per the preprint. The caveat is equally clear from the companies’ own framing: this was a tightly scoped, human-gated campaign on one reaction class, and the partners’ decision to publish protocols but withhold the raw screening data, as reported by R&D World, means the broader claim of an autonomous AI chemist will be tested by whether others can reproduce the loop, not just the result.