A near-autonomous AI chemist improves a challenging reaction in medicinal chemistry

OpenAI and Molecule.one have demonstrated a near-autonomous AI chemist utilizing GPT-5.4 to improve a challenging drug-making reaction in medicinal chemistry. Working in conjunction with Maria, an agentic chemistry AI integrated into a high-throughput laboratory, the system was given the open-ended goal of improving important reaction classes. The project focused on the Chan-Lam coupling reaction, which forms carbon-nitrogen bonds but historically yields low results when pairing primary sulfonamides with boronic acids. Through this collaboration, GPT-5.4 independently identified primary sulfonamides as a challenging substrate class and proposed that mild oxidants, including TEMPO, could improve the reaction. Over two cycles of experimentation in the Maria Lab involving 10,080 reactions, the optimized conditions raised the mean yield from 16.6% to 25.2% and increased the share of reactions above 30% yield from 15.6% to 37.5%. Human chemists remained involved by designing steering prompts, selecting proposals, making limited corrections, assisting with lab operations, and independently validating the final results at bench scale, where representative reactions showed higher yields for 11 of 14 substrate pairs. The full process took three months, showcasing how advanced language models can review literature, propose unexpected ideas, analyze experimental data, and help drive concrete discoveries that human researchers can evaluate and build upon.