OpenAI’s GPT-5.4-powered near-autonomous AI chemist lifts mean Chan-Lam coupling yields for primary sulfonamides from a baseline of 16.6% to 25.2% across 88% of tested substrates, per official collaboration results with Molecule.one validated at standard bench scale.
How the GPT-5.4-powered near-autonomous AI chemist lifts Chan-Lam coupling yields
Unlike narrow AI tools limited to predicting reaction outcomes, the end-to-end system operated across the full experimental research cycle with minimal human intervention. Human chemists first provided open-ended steering prompts focused on improving Chan-Lam coupling for process chemistry, with no pre-specified substrate class or additive target. This unconstrained framing allowed the model to explore a far wider range of potential improvements than a narrowly scoped prompt would enable.
The system operated across two full sequential experimental cycles, with GPT-5.4 generating and ranking thousands of potential research proposals. Human chemists selected four of these proposals for laboratory testing, per OpenAI’s official collaboration announcement.
The top-performing proposal, coded OAI-M1-03, independently identified primary sulfonamides as a high-value, underperforming substrate class. It proposed mild oxidants as a yield-boosting intervention, a suggestion human chemists described as both surprising and interesting, as Chan-Lam coupling is traditionally run under inert, oxygen-free conditions.
Molecule.one’s Maria AI platform translated the high-level proposal into detailed lab protocols, then ran thousands of high-throughput microliter-scale experiments. It analyzed raw yield data and fed structured results back to GPT-5.4 to inform follow-up iterations across the two full experimental cycles. Humans remained in the loop for prompt design, proposal selection, limited experimental plan corrections, basic lab operations, and independent final validation of results.
Chan-Lam coupling yield bottleneck for primary sulfonamides
Chan-Lam coupling is a widely used reaction in medicinal chemistry for forming carbon-nitrogen bonds, a structural motif present in the majority of approved small-molecule drugs. Prior to the AI system’s intervention, mean yields for Chan-Lam coupling of primary sulfonamides with boronic acids sat at 16.6%, per OpenAI’s research announcement. Primary sulfonamides are a chemical group found in treatments for oncology, infectious disease, and diuretic therapies.
Low yields for this substrate class force medicinal chemists to either abandon promising sulfonamide-containing candidate molecules or invest extra time developing alternative, less efficient synthesis routes. This creates a synthesis bottleneck in early-stage drug discovery, as scientists can only test molecules they can reliably produce at scale. Improving the reliability of this specific reaction class would expand the set of potentially therapeutic molecules medicinal chemists can explore, removing a key constraint for sulfonamide-containing drug candidates.
Bench-scale validation of AI-identified yield improvements
A key test of the microliter-scale results was whether they would hold up in the standard bench-scale workflows used by medicinal chemists daily, rather than only in specialized high-throughput screening setups. Many high-throughput screening results fail to translate to practical lab workflows, so this validation step is critical for real-world impact.
Human chemists repeated 14 representative substrate pairs at bench scale, and found higher yields for 11 of the 14 pairs, per OpenAI’s official notes. Most tested combinations saw more than a twofold increase in yield at bench scale, confirming the microliter-scale results are directly applicable to existing lab practices used in drug discovery programs.
The final mean yield of 25.2% for primary sulfonamides represents a 52% relative increase over the baseline 16.6% yield, with yields improving for 88% of tested boronic acids and 83% of tested sulfonamides. The share of reactions exceeding 30% yield more than doubled from 15.6% to 37.5% across the two experimental cycles.
Significance for agentic AI in life sciences research
This result marks a concrete step forward for agentic AI in the life sciences, moving beyond narrow prediction or literature summarization tools to systems that generate novel, experimentally testable hypotheses and iterate on them with real-world lab data. The model’s proposal of mild oxidants for a reaction class traditionally run under inert, oxygen-free conditions is counterintuitive.
This highlights the value of open-ended AI proposal generation in exploring understudied areas of reaction design that human chemists may overlook due to entrenched conventional practices.
This work builds on OpenAI’s earlier life sciences-focused model, GPT-Rosalind, which was designed specifically to support drug discovery workflows. It also follows prior documented cases of GPT-5 delivering cross-disciplinary research advances: these include novel mathematical findings related to the unit distance problem, new gluon amplitude calculations in theoretical physics, and automated biology research that reduced the cost of cell-free protein synthesis.
OpenAI frames the Chan-Lam coupling improvement as a concrete proof of concept for its broader vision of AI as a cross-disciplinary research partner, one that can operate across the full research loop from literature review to experimental validation.
