A machine-learning system developed by researchers at the University of Utah and collaborators can predict how drug molecules form their left- or right-handed versions, slashing the number of lab experiments from roughly 50 to 60 reactions down to just 5 to 10. Published in Nature in March 2026, the study demonstrates a data-efficient approach to enantioselective nickel-catalyzed couplings that cuts months of lab work to days and dramatically lowers reagent costs.
Why Molecular Handedness Matters for Drug Safety
Many molecules used in medicines share the same atoms connected in the same order, but their three-dimensional arrangements are mirror images of each other, a property chemists call handedness or chirality. The distinction is far from academic. One “hand” of a molecule may treat a disease effectively, while its mirror image can be inert or even harmful. Thalidomide remains the most infamous example: one form treated morning sickness, while the other caused severe birth defects.
Getting the right-handed or left-handed version of a drug candidate in high purity typically requires extensive trial-and-error screening. Chemists test dozens of catalysts, ligands, solvents, and temperature combinations to find conditions that favor one mirror form over the other. That screening process eats through weeks of researcher time and thousands of dollars in reagents, creating a bottleneck that slows early-stage pharmaceutical development.
The new work underscores why that bottleneck is so costly. As one researcher explained in an interview about how mirror-image molecules behave differently in the body, even subtle changes in three-dimensional shape can alter how a drug binds to its biological target. That means chemists cannot simply rely on overall composition; they must control the exact spatial arrangement of atoms to ensure both safety and efficacy.
From 60 Reactions to Fewer Than 10
The new system, described in a Nature paper on transferable enantioselectivity models, tackles that bottleneck head-on. Trained on sparse experimental data rather than massive datasets, the machine-learning model predicts reaction outcomes for nickel-catalyzed C(sp3)-couplings and transfers those predictions to entirely new reactions. In wet-lab validation, the tool reduced experimental runs from roughly 50 to 60 reactions down to approximately 5 to 10, a reduction of nearly 90 percent in the number of trials needed to identify optimal conditions.
That efficiency gain translates directly into saved hours and saved money. “As a lab-based chemist, I find that this tool is extremely valuable for saving time spent running experiments,” said co-author Ryan Bucci of the University of Utah. The practical impact is straightforward: fewer reactions mean fewer purchased or synthesized reagents, less solvent waste, and less time tied up on a single optimization campaign.
Because the model can propose promising conditions before a chemist steps into the lab, it also changes how projects are prioritized. Instead of committing large chunks of time to explore whether a new coupling strategy might work, teams can quickly gauge feasibility and either move forward with confidence or pivot to alternatives, tightening the feedback loop between design and synthesis.
How Sparse Data Beats Brute Force
Most machine-learning tools in chemistry demand large training sets, which is precisely the resource that bench chemists lack when exploring a new reaction type. The Utah-led team flipped that constraint into a design principle. Their model learns from small batches of experimental results, builds predictive maps of how molecular structure influences handedness outcomes, and then generalizes across related reaction families. One researcher involved described the approach this way: “The coolest thing about this tool is that it allows someone to collect smaller bits of data, build reasonably good models and make predictions” about which catalyst will deliver the desired handed version of a molecule.
The system acts as a high-tech filter. Rather than forcing a chemist to run every plausible combination at the bench, it screens tens of thousands of chemical structures computationally, predicting how molecular pieces will assemble before any glassware is touched. Only the most promising candidates advance to physical testing. This reversal of the traditional workflow, where computation precedes experimentation rather than following it, is what compresses timelines from months to days.
Crucially, the model is designed to be transferable. It does not just memorize outcomes for a narrow set of substrates; it captures patterns in how ligands and reaction conditions steer enantioselectivity. That allows it to extrapolate to new coupling partners and substitution patterns, which is essential in medicinal chemistry, where each new project may feature distinct scaffolds and functional groups.
A Broader Shift Toward AI-Guided Synthesis
The Utah team’s work fits into a wider movement connecting machine learning with automated chemistry platforms. Peer-reviewed research on Bayesian optimization in robotic labs has shown that AI-driven optimization can shrink experimental campaigns by selecting the most informative next experiment at each step, rather than sweeping through a grid of conditions. In these systems, algorithms iteratively propose new experiments, analyze the results, and refine their models, steadily homing in on optimal conditions with far fewer trials than traditional design-of-experiments approaches.
Survey-level evidence from a separate review on self-driving laboratories documents cases where optimization time dropped from weeks to hours and experimental costs fell substantially. These platforms combine robotics, in-line analytics, and machine learning to create closed-loop workflows that can run around the clock with minimal human intervention.
What distinguishes the new Nature study from these broader trends is its focus on transferability under data scarcity. Many optimization tools perform well within a single reaction class but fail when applied to a structurally different substrate. The sparse-data approach demonstrated here transfers predictions to new reactions without retraining from scratch, a feature that could make it useful across multiple drug programs rather than requiring bespoke calibration for each one.
In practical terms, that means a single modeling framework could guide enantioselective couplings across an entire portfolio of projects. Instead of building and maintaining separate models for each reaction family, chemists could rely on a common tool that learns as it goes, accumulating knowledge from one campaign and deploying it in the next.
What This Changes for Drug Development Economics
The pharmaceutical industry spends heavily on failed synthesis routes before a drug candidate ever reaches clinical trials. Each round of optimization that dead-ends wastes not just reagents but researcher salaries and equipment time. By compressing the screening phase, tools like the Utah system can reduce the number of dead ends that must be explored in depth.
For early-stage projects, where resources are tight and timelines are closely watched, being able to identify viable enantioselective routes in a handful of experiments can shift the economics of discovery. Compounds that might previously have been shelved because their stereoselective synthesis looked too complex or expensive could now be reconsidered. Conversely, candidates that prove stubborn even with AI guidance can be deprioritized earlier, before they consume disproportionate effort.
The approach also offers environmental and operational benefits. Fewer optimization experiments mean less solvent consumption, reduced chemical waste, and lower energy use for temperature-controlled reactions and purification. In an era when many companies are setting sustainability targets for their R&D operations, data-efficient optimization tools provide a direct lever for reducing the footprint of synthesis campaigns.
Looking ahead, the real impact may emerge as these models are woven into end-to-end digital workflows. A medicinal chemist could propose a new target structure, have an AI system suggest enantioselective routes with predicted outcomes, and then pass the top candidates to an automated platform for rapid validation. Human expertise would remain central (especially in interpreting results, assessing risks, and making strategic decisions), but the routine exploration of conditions would increasingly be handled by algorithms and robots.
For now, the Utah-led study offers a concrete demonstration that sophisticated models do not always require massive datasets. By embracing sparse data and focusing on transferability, the researchers have shown that machine learning can deliver practical gains in one of the most challenging corners of synthetic chemistry. If similar strategies are extended to other reaction classes, the long-standing trial-and-error culture of enantioselective synthesis may give way to a more predictive, data-informed era.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
