GPT-5.4 et Maria AI Lift Chan-Lam rendent plus de 88% des substrates

GPT-5.4 raised Chan-Lam coupling yields from an average of 16.6% to 25.2% across 88% of tested boronic acids by proposing a mild oxidant, TEMPO, that no human chemist had considered for this specific reaction. That improvement came from a near-autonomous loop connecting a language model to a high-throughput lab that ran 10,080 reactions without a scientist touching a pipette.

Why Primary Sulfonamides Were the Bottleneck

Chan-Lam coupling forms carbon-nitrogen bonds, a staple in drug molecules. Primary sulfonamides appear in anticancer drugs, antimicrobials, and diuretics, but their coupling with boronic acids has historically given miserable yields. Mean yields sat at 16.6%, with only 15.6% of reactions breaking 30% yield. Medicinal chemists often abandon promising sulfonamide-containing candidates because they cannot make enough material to test.

How GPT-5.4 and Maria Ran 10,080 Reactions

OpenAI paired GPT-5.4 with Molecule.one's Maria, an agentic chemistry AI integrated with a high-throughput laboratory. Scientists wrote steering prompts, then GPT-5.4 generated and ranked thousands of research proposals. Human chemists selected four for testing. Maria AI translated those plans into detailed lab instructions and executed two cycles of experimentation. One proposal, OAI-M1-03, zeroed in on TEMPO as an additive.

Under the optimized conditions, 88% of boronic acids and 83% of sulfonamides produced higher yields. The share of reactions above 30% yield jumped from 15.6% to 37.5%. Maria ran all 10,080 reactions autonomously, including data analysis and structured output back to GPT-5.4 for follow-up proposals.

Validation at Bench Scale Confirms the Microliter Results

Human chemists repeated representative reactions at bench scale. 11 of 14 substrate pairs showed higher yields, most with more than a twofold increase. That gap between microliter screening and practical lab scale often kills automation discoveries, but here the improvement survived the translation.

For medicinal chemists stuck on a low-yielding sulfonamide coupling, this AI-derived protocol now offers a practical alternative. The bigger takeaway: AI systems that combine literature review, hypothesis generation, and autonomous experimentation are no longer a demo - they are running 10,000-reaction campaigns and delivering validated results.

Source: A near-autonomous AI chemist improves a challenging reaction in medicinal chemistry
Domain: openai.com