Autonomous AI chemist improves a notoriously difficult medicinal chemistry reaction

OpenAI carried out a medicinal chemistry research project by connecting GPT-5.4 to Maria, Molecule.one's autonomous experimentation robot. The system built an autonomous research loop that handles everything from proposing research directions to designing experiments, executing them, analyzing data, and proposing follow-up experiments.

The core achievement of this project was improving the yield of the Chan–Lam coupling reaction. During the experimental process, GPT-5.4 formed an unexpected hypothesis that mild oxidants including TEMPO could promote the reaction, leading to the following results.

• Yield improvements were confirmed in 88% of tested boronic acids and 83% of sulfonamides
• Average yield rose from 16.6% to 25.2%
• The proportion of reactions achieving yields of 30% or higher increased from 15.6% to 37.5%

When human chemists conducted reproduction experiments at bench scale, yields more than doubled in most cases, demonstrating that results from microliter-scale experiments hold up in real research settings as well. This case shows that AI can become a powerful partner in compound synthesis, a major bottleneck in the drug discovery process.