A team of researchers at the Korea Advanced Institute of Science and Technology (KAIST) has used artificial intelligence to build proteins that do not exist in nature and shown that these synthetic molecules can detect six specific compounds, including the stress hormone cortisol and the blood-thinning drug apixaban. The results, published in Nature Communications in April 2026, represent one of the clearest demonstrations yet that computers can design molecular sensors from a blank slate, potentially replacing the slow, expensive antibodies that underpin most diagnostic tests today.
“We wanted to show that de novo protein design can go beyond proof-of-principle and actually produce binders with the sensitivity needed for real diagnostics,” Jaekyoung Son, a lead researcher in KAIST’s InnoCORE Research Group, said in a university announcement. The group collaborated with Nobel Laureate David Baker, whose laboratory at the University of Washington has pioneered the use of AI tools such as RFdiffusion and Rosetta for protein engineering.
What the study found
The researchers developed a computational pipeline that combines AI-driven structure prediction with physics-based energy calculations to generate proteins tailored to grip a chosen small molecule. Starting from the chemical shape of each target, the system proposes thousands of candidate protein architectures, scores them for binding strength and stability, and filters the results down to a handful of designs worth testing in the lab.
Applied to six chemically distinct compounds, the pipeline produced proteins with binding affinities in the nanomolar to low micromolar range. That bracket places them in the same sensitivity zone as many antibodies used in clinical blood tests. The authors note that the computational design stage can be completed far more quickly than the months typically required to raise and validate an antibody, though direct timing comparisons depend on the complexity of each target.
Three of the designs were confirmed at atomic resolution through X-ray crystallography, with structures deposited in the Protein Data Bank for anyone to inspect. The cortisol-binding protein, cataloged as PDB entry 8UQF, shows a binding pocket that matches the computational prediction almost exactly. Two additional structures for apixaban-binding designs (PDB entries 8VFQ and 8VEZ) demonstrate that the method works across targets with very different chemical properties, not just a single favorable case.
The team went a step further with cortisol, translating one design into a working biosensor. The device uses a technique called chemically induced dimerization: when cortisol enters the system, it causes two protein components to lock together, producing a measurable signal. That signal could, in principle, be read by a handheld device, though the current prototype has only been tested under controlled laboratory conditions.
What has not been proven yet
The distance between a lab biosensor and a device that works on a drop of blood or a sample of river water is significant. No clinical or field validation of the cortisol sensor has been reported. All binding measurements were performed in purified buffer solutions, and performance in complex biological fluids, where thousands of competing molecules could interfere, remains untested. Secondary coverage, including a Phys.org summary, highlights potential applications in diagnostics and environmental monitoring, but those remain projections.
The study also builds on earlier work. The authors cite a 2024 paper by An and colleagues on pseudocycle-based scaffolds for small-molecule binding, as well as a separate study demonstrating sub-angstrom accuracy in de novo protein design. The new paper claims to overcome prior limitations in generating diverse pocket geometries, but independent benchmarking by groups outside the Baker and KAIST collaboration has not yet appeared. Whether the pipeline reliably produces high-affinity binders for arbitrary new targets, or whether the six demonstrated compounds represent favorable test cases, is an open question.
KAIST has filed a U.S. provisional patent related to the technology, though the specific claims have not been made public. Provisional patents establish a priority date but do not guarantee that a full patent will be granted.
Why it matters for diagnostics
Most rapid diagnostic tests, from home pregnancy kits to hospital cortisol assays, rely on antibodies harvested from animal immune systems or engineered in cell cultures. That process is proven but expensive, time-consuming, and limited by biology: some molecules are poor at triggering an immune response, making it difficult to generate antibodies against them. A computational approach that can design a binder for virtually any small molecule, given only its chemical structure, would sidestep those constraints.
The practical appeal is speed. If a new drug enters the market or a novel environmental contaminant is identified, a lab using this pipeline could, in theory, have a candidate sensor protein within days rather than waiting months for antibody development. The KAIST team’s decision to archive its design scripts on Zenodo (DOI: 10.5281/zenodo.17847477) and deposit deep-sequencing data in NCBI BioProject PRJNA1356499 lowers the barrier for other laboratories to test and extend the method.
Still, several practical hurdles stand between this proof of concept and a commercial product. Each new sensor protein must be expressed, purified, and characterized in a wet lab, which requires infrastructure and expertise. If extensive experimental tuning turns out to be necessary for every new target, the speed advantage over antibody development could narrow. Off-target binding, long-term protein stability, and manufacturability at scale have not been addressed.
Regulatory pathways add another layer of uncertainty. Any diagnostic device that uses these novel proteins in contact with patient samples would need to clear safety and performance evaluations. Because de novo proteins have no history of human exposure, regulators may require additional toxicology and immunogenicity testing before approving their use in clinical settings. For environmental monitoring, certification standards vary by jurisdiction and have not been evaluated for sensors based on computationally designed proteins.
The bottom line
The KAIST and Baker lab collaboration has produced one of the most concrete demonstrations to date that AI can design functional molecular sensors from scratch. The crystal structures confirm that the proteins fold as predicted and bind their targets with the geometry the computer specified. The cortisol biosensor shows that binding can be converted into a detectable signal. Those are real, verifiable results, backed by openly available data and code.
What the study does not yet prove is that this approach can scale to dozens or hundreds of new targets with predictable success, or that the resulting sensors will perform reliably outside a controlled laboratory. Future studies that test the pipeline against more chemically diverse compounds, validate sensors in blood or saliva, and include independent replication by outside groups will determine whether this milestone becomes the foundation for a new generation of programmable biosensors or remains a technically impressive but specialized achievement.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
