AI Chemist Optimizes a Complex Medicinal Chemistry Reaction

Date: June 17, 2026
Research Publication

In a collaborative effort with Molecule.one’s Maria, OpenAI's GPT-5.4 has successfully identified a surprising additive that enhances the yields of Chan-Lam Coupling for more than 80% of the substrates tested.

OpenAI’s venture into the sciences is driven by the conviction that sophisticated AI can serve as a potent collaborator for researchers—enabling them to bridge disparate concepts, explore a wider array of hypotheses, refine experimental design, and accelerate breakthroughs for the benefit of society.

A Trajectory of Scientific AI

This achievement is part of a broader pattern of AI-driven scientific contributions, including:

Mathematics: Advancements regarding the unit distance problem.
Theoretical Physics: New insights into gluon amplitudes.
Biology: Utilizing GPT-5 to reduce the costs associated with cell-free protein synthesis in automated environments.
Life Sciences: The creation of GPT-Rosalind, a model tailored specifically for drug discovery and biological research.

This latest project moves beyond theoretical reasoning into the realm of medicinal chemistry, where success is measured by tangible results in a physical lab, accounting for experimental noise and real-world molecular behavior.

The Integration: GPT-5.4 $\times$ Maria

To achieve this, OpenAI linked GPT-5.4 with Maria, an agentic AI integrated with a high-throughput laboratory. The goal was open-ended: improve a critical class of chemical reactions.

The Autonomous Workflow

The system operated through a continuous loop of hypothesis and testing:

AI Responsibilities:

Literature review and hypothesis generation.
Experimental design and execution.
Data analysis.
Iterative refinement of proposals.

Human Responsibilities:

Designing steering and grading prompts.
Selecting the most viable proposals for testing.
Providing minor corrections to experimental plans.
Assisting with basic lab operations.
Independent validation of final results.

Solving the Chan-Lam Coupling Challenge

The most successful proposal, identified as OAI-M1-03, targeted a specific, difficult version of the Chan-Lam coupling—a reaction essential for creating carbon-nitrogen (C-N) bonds.

Why This Matters

Organic chemistry is the foundation of small-molecule drugs, electronics, and agricultural products. However, synthesis is frequently the primary bottleneck in drug discovery.

Feature	Impact on Drug Discovery
Reliability	High-yield reactions allow scientists to test more diverse molecules.
Efficiency	Reducing byproducts saves time and resources.
Accessibility	Reliable C-N bonding opens new paths for molecular exploration.

The specific challenge involved primary sulfonamides. These are critical components in:

Anticancer medications
Diuretics
Antimicrobials

Historically, the coupling of primary sulfonamides with boronic acids has been plagued by ~~low yields~~ inefficient results.

The Breakthrough

GPT-5.4 independently hypothesized that mild oxidants, specifically TEMPO, could boost the reaction. After two cycles of experimentation in the Maria Lab (which processed a total of $10,080$ reactions for OAI-M1-03), the results were significant.

Yield Improvements:

Boronic Acids: $\text{Improved for } 88\%$ of substrates.
Sulfonamides: $\text{Improved for } 83\%$ of substrates.
Mean Yield: $\text{Increased from } 16.6\% \rightarrow 25.2\%$ .
High-Yield Success: $\text{Reactions } > 30\% \text{ yield rose from } 15.6\% \rightarrow 37.5\%$ .

Chart comparing TEMPO, 4-hydroxy-TEMPO, 4-oxo-TEMPO, and PMP performance with chemical structures.

Validation and Conclusion

To ensure these results weren't merely an artifact of micro-scale screening, human chemists performed bench-scale validations.

Bench-Scale Results:

11 out of 14 substrate pairs showed higher yields.
Most cases exhibited a more than twofold increase in yield.

Labeled glass reaction vials from Molecule.one bench-scale validation experiments.

This project demonstrates a glimpse into the future of scientific research: AI systems that don't just process data, but act as partners throughout the entire research loop—from reviewing literature and proposing unexpected ideas to designing experiments and arriving at verifiable scientific findings.