What is de novo protein design?

How it works

Structure prediction starts with a sequence of amino acids, the building blocks of proteins. It asks what shape the sequence will fold into. De novo design runs that problem in reverse. It starts with a desired structure and computes a sequence that should produce it. That is how the Royal Swedish Academy of Sciences describes the approach David Baker's group at the University of Washington took with its Rosetta software.

The group's result, published in 2003, was Top7, a 93-amino-acid protein with a structure not found in nature. To test it, the team had bacteria produce the protein and then determined its structure by X-ray crystallography. The academy says the result almost exactly matched the design. Baker was awarded half of the 2024 Nobel Prize in Chemistry for computational protein design.

The AI tools

Deep-learning models have since joined the toolkit. In July 2023, Baker's lab and collaborators published RFdiffusion in Nature. The model was made by fine-tuning the RoseTTAFold structure prediction network to remove noise from protein structures. That is the diffusion approach also used in image generators. The authors reported lab tests of hundreds of designs, including binders. Binders are proteins made to attach to a chosen target.

In November 2025 the group and collaborators reported in Nature antibodies designed from scratch to bind chosen sites on their targets. They used a fine-tuned RFdiffusion together with lab screening. The authors wrote that the initial designs bound with modest strength. They said a further lab step, affinity maturation, was used to improve it.

Where things stand in 2026

SKYCovione is a COVID-19 vaccine developed by SK bioscience. Its underlying nanoparticle platform was developed at the University of Washington's Institute for Protein Design. In 2022 SKYCovione became the first computer-designed medicine to receive regulatory approval, the institute says.

GBP511 is a newer vaccine on the same platform that SK bioscience is taking into trials. It is built around a computer-designed protein particle with no natural counterpart. The institute says GBP511's Phase 1/2 trial began enrolling in January 2026 and is to cover about 368 healthy adults in Australia. Results are expected by 2028, it says.

Antibody drugs are another focus. Nature reported in December 2025 that scientists saw clinical trials of AI-designed antibodies on the horizon. It said the first proof-of-principle designs, made a year earlier, lacked the potency of commercial antibody drugs.