Nobel laureate David Baker’s lab unveiled an AI that designs proteins atom by atom

David Baker’s laboratory at the University of Washington has published the Science paper on RFdiffusion3, a generative AI system that designs proteins and nucleic acids by modeling individual atoms, and its authors report that it produced functional enzymes from scratch. The paper appeared in Science on October 15, 2026, about ten months after the lab first released the tool’s code.

Tech news in one quick read: Get the most interesting stories in startups, tech and programming, each with a short TLDR, in one free daily email from TLDR.

Subscribe to TLDR free →

We may earn a commission if you sign up.

Baker shared the 2024 Nobel Prize in Chemistry for computational protein design, and the new model extends the line of work that earned the award. Where earlier systems reasoned mostly about amino acids, this one treats atoms as the basic design unit, which matters most when a protein has to hold a small molecule or a catalytic group in exactly the right place.

RFdiffusion3 and atom-level protein design

The study, titled “Atomic functional site scaffolding in biopolymers with RFdiffusion3,” is credited to Jasper Butcher and colleagues, and a AAAS summary on EurekAlert points to Rohith Krishna and Baker of the university’s Institute for Protein Design as expert contacts. The paper’s identifier is DOI 10.1126/science.aeh2982. According to the summary, earlier methods worked mainly at the amino-acid level, or modeled atoms but struggled with interactions involving small-molecule ligands, nucleic acids or catalytic groups.

Trained on a broad collection of biological structures, the model can build complex functional sites, design RNA, and create enzymes that coordinate several catalytic components. In computational tests, the authors report it outperforms previous approaches in designing protein structures and organizing catalytic sites. That finer control over the molecular interactions that determine function is the point of working atom by atom. Designing RNA and other nucleic acids with the same system, rather than a separate tool for each molecule type, is part of why the authors present it as a general platform rather than a single-purpose enzyme generator. GeekWire described the model as general purpose, able to create proteins that bind DNA, other proteins and small molecules while also supporting enzyme-related functions, which is the breadth the Science paper now documents in more depth for catalytic sites.

Enzymes built from scratch

The laboratory experiments are where the claim moves from software to chemistry. The authors produced functional enzymes from scratch, including enzymes that break down paraoxon, a toxic organophosphate, and a protease engineered to cleave a peptide associated with transthyretin amyloidosis. Both targets illustrate the kind of precise, multi-part active site that the atom-level approach is meant to handle.

The release does not quantify how efficient those enzymes are or how many designs were tested, and the summary gives no direct quotes from the researchers, so the strength of the activity is best judged from the paper itself. What the summary does establish is the category of result: catalysts that did not exist in nature, specified by a model and then made and tested in the lab.

From a December 2025 code release to a Science paper

The tool was not a surprise to the field. GeekWire reported on December 3, 2025 that the lab was publicly releasing RFdiffusion3 alongside an update to its enzyme-design model, RFdiffusion2. Krishna, a postdoctoral fellow and lead developer, said the goal was to combine everything into one foundational model, and that the team was excited to see what other researchers build on it.

GEN, in its own account of the release, described RFdiffusion3 as open source and noted that it treats individual atoms as design units, whereas RFdiffusion2 used a hybrid of atoms and amino acid residues. The article reported that the new model runs about ten times faster than its predecessor, and quoted Krishna describing the aim of giving more knobs for generality and controllability. At that stage the work rested on a preprint that had not been peer reviewed.

That preprint also carried experimental proof of concept for designing proteins that recognize specific DNA sequences, which GEN said could serve as a basis for synthetic transcription factors in gene therapy. Krishna cautioned that the model struggles with biomolecules absent from its training data, such as non-canonical amino acids, a limit that matters for chemists who want to push beyond the twenty standard building blocks. Even so, the combination of an open code release, a peer-reviewed paper and laboratory-made enzymes gives other groups a concrete starting point for their own experiments.

Baker’s Nobel recognition, announced October 9, 2024 by the Royal Swedish Academy of Sciences, was shared with Demis Hassabis and John Jumper of DeepMind, who were honored for protein structure prediction, according to the University of Washington’s bioengineering department. Baker directs the Institute for Protein Design and holds the Henrietta and Aubrey Davis Endowed Professorship in Biochemistry. The department’s page identifies him as a professor at the University of Washington, which is the institution behind both the original RFdiffusion model and this third generation of the software. The peer-reviewed Science paper gives the atom-level approach its formal scientific record, and with the code already public, outside groups can test whether the enzyme results hold up in their own hands.

This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.