Drug discovery is like molecular Tetris. Chemists put atoms together, adjust the pieces until everything fits together, and suddenly a molecule produces a promising new medicine. Typically, creating better molecules consumes enormous amounts of time and money.

In a new study, researchers used machine learning to build a smarter prediction system that could speed up the process at a fraction of the cost.

“We sometimes use sophisticated physics-based computational chemistry tools to understand novel reactions. However, these tools are too expensive to make predictions about thousands of potential new molecules,” said Simone Gallarati, co-lead author of the study and joint postdoctoral researcher at the University of Utah and the University of California, Los Angeles. “We wanted to train statistical models that were ‘smart’ enough to make accurate predictions about untested reactions, but also as cheap as possible.”

Molecules can exist as mirror images, a property known as “laterality.” Left-handed versus right-handed ways are crucial; one can heal, the other can harm. Chemists need to find the right set of tools—catalysts, ligands, and substrates—to ensure they build the right version.

The new system acts as a high-tech filter that can examine tens of thousands of chemical structures to predict how pieces will fit together to produce a “hand” of one molecule on top of another. The workflow provides a cost-effective way to convert reaction components into numerical data that a computer can analyze, creating the framework for machine learning predictions.

With surprisingly little information, the model reliably predicted how components would behave, reducing the time, energy and expense spent testing reactions in the lab.

“Most AI requires huge amounts of data to train models. That’s a problem in chemistry where getting large, high-quality data sets from experimental work is very expensive and time-consuming,” he said. Mateo SigmanU chemist and co-author of the study. “The best thing about this tool is that it allows you to collect smaller chunks of data, build reasonably good models and make accurate predictions for known reactions, and also transfer predictions to reactions that the models have not yet seen.”

The study was published as an accelerated preview in the nature magazine on February 11, 2026.

High-tech filter

The researchers focused the workflow on asymmetric cross-coupling reactions, a powerful toolkit for drug development. The reactions join two carbon-based molecular fragments, using a metal catalyst to form more complex compounds. The reactions are called asymmetric because they are designed to favor a “manipulated” version of the molecule. Chemists often produce both versions, but without guidance, experiments will yield a 50/50 split. In contrast, asymmetric reactions produce, say, 95% of the desired shape and only 5% of the unwanted mirror image.

Asymmetric cross-coupling reactions generally require at least three elements: a metal, a ligand, and substrates. The metal catalyst does the heavy lifting by linking carbon-based molecules to build the product. A ligand binds to the metal, controlling which side of the molecule reacts, influencing the three-dimensional orientation of the product. The ligand is arguably the most important element in controlling the laterality of a molecule.

To train their model, Gallarati and the team identified four academic papers on asymmetric reactions (including previous work by co-author Abigail Doyle and Sigman) that used nickel-based catalysts with different ligands. Those results were the only training data for the workflow. The team then asked the system to predict the results of hypothetical components not included in the training data. They added a series of increasingly challenging tasks that forced the algorithm to make predictions with increasingly different materials from the original training data. The team tested the prediction in the Doyle lab, an effort led by Erin Bucci, co-lead author of the study and a doctoral student at UCLA.

“As a laboratory chemist, this tool is extremely valuable in saving time spent performing experiments,” Bucci said. “For example, instead of running 50 to 60 reactions, we can now run 5 to 10, potentially saving weeks or months. Each reaction component we test in the lab must be purchased or made from scratch; this tool greatly reduces the amount of money you would normally spend on materials.”

While the authors tested the tool in the context of new nickel-based reactions, the workflow can be applied across fields and even deepen our understanding of chemistry itself.

“One of the nice things about workflow is that it’s not a black box,” he said. Abigail Doylechemist at UCLA and co-author of the study. “We can learn something about chemistry from the predictions, even if they are wrong. We apply our chemistry expertise to help learn something we wouldn’t have learned without the tool.”

The pharmaceutical industry would immediately benefit from a tool like this, Sigman added. Let’s say a company needs to deliver large quantities of a compound for a clinical trial and wants to apply a reaction that already appears in the literature. But it has never been done with its specific composite objective.

“This is where this tool could be highly applicable,” he said. “Optimizing a reaction and time cost is the value proposition when building a drug. This simplified process could make the difference when they need to take a molecule from phase one to phase two.”

****

Gallarati, S. and others, Transferable enantioselectivity models from sparse data. Nature (2026). https://doi.org/10.1038/s41586-026-10239-7

The work was supported by the Swiss National Science Foundation (#222115), the US National Science Foundation (CHE-2202693 and CHE-1048804), the National Institutes of Health (S10OD028644), and the Center for High Performance Computing at the University of Utah.