The full design of RFdiffusion3 is now public: the generative AI system designs proteins and nucleic acids by positioning individual atoms, and an international team from the University of Washington, the University of Oxford, NVIDIA and other institutions published the work in Science on October 16, 2026. Details disclosed on October 19 show that two enzymes built to its specifications have already carried out their intended functions in real chemical reactions. For protein design, the significance is not that there is one more generative model, but that design precision has reached the atomic level where function is decided.
From predicting structures to arranging atoms one by one
Corresponding author David Baker leads the Institute for Protein Design at the University of Washington and received the 2024 Nobel Prize in Chemistry for work on protein structure prediction. His team's earlier RosettaFold learned spatial relationships between amino acids to predict what existing proteins look like; RFdiffusion in 2023 and RFdiffusion2 in 2025 began designing new proteins instead, but still worked at the level of amino-acid residues, so the individual atoms that decide binding and catalysis could only be inferred indirectly.
RFdiffusion3 takes a different approach. Drawing on the network architecture of AlphaFold3, it represents each amino acid with up to 14 atoms, the diffusion process adjusts the three-dimensional coordinates of every atom directly, and it shares no code with its two predecessors. Researchers can state conditions very concretely: which atoms must sit in which positions, how hydrogen bonds should form, and how far particular atoms may spread inside the protein, and the model then generates a structure satisfying all of those constraints at once. Designing the chemistry first and growing the whole molecule from it is exactly where it pulls away from residue-level tools.
Ahead on 37 of 41 active sites, and it worked in the lab
Judging a design tool comes down to two things: whether its calculations hold up, and whether what it produces actually works. The team ran comparative designs on 41 active sites of existing enzymes, and RFdiffusion3 beat RFdiffusion2 on 37 of them, at roughly ten times the generation speed of the previous generation. More important is the experimental check: a phosphotriesterase and a peptide-cleaving enzyme designed with it both catalyzed their intended chemical reactions in the laboratory, rather than stopping at a computer score.
The same results also mark the limits. What has been validated so far is tasks like enzyme active sites, where the goal is clear and the constraints can be specified precisely; a design that catalyzes a reaction is still separated from becoming a drug or therapy by expression, stability, safety and manufacturing at scale. Calling RFdiffusion3 a tool that can design any protein would go beyond the evidence in the paper; the more accurate reading is that the more precisely a task's constraints are written, the clearer its advantage becomes.
What is open is the whole method, not just a model file
RFdiffusion3 was already released as open source on December 3, 2025, with code and training code available on GitHub through the Rosetta Commons Foundry; what the Science paper adds is the complete design rationale and the experimental validation. Directions the team lists include enzymes that break down microplastics, synthetic transcription factors for gene therapy and biosensors for environmental monitoring, applications whose common thread is interactions with DNA and small molecules that must be precise to the atom, squarely in the new model's strongest territory.
Among related work, DeepMind's watermarking of AI-designed proteins addresses how design outputs can be marked and traced, which pairs with this push toward atomic-level design precision as a before-and-after relationship: every step forward in generative capability makes verification and governance harder to skip. For ordinary laboratories, RFdiffusion3 can be downloaded and tried today; the real test over the coming months is whether independent teams can reproduce paper-level design success rates with it.
