Generative AI system RFdiffusion3 has officially released its complete design specifications, demonstrating that proteins built at the atomic level can perform real chemical reactions in laboratory tests.
Global Collaboration Publishes Atomic-Level Architecture
The international research effort behind RFdiffusion3 brings together the University of Washington in the United States, the University of Oxford in the United Kingdom, and NVIDIA. Their findings were published on the 16th local time in the international journal Science, detailing the underlying design principles of a system that models individual atoms for protein and nucleic acid engineering. The team previously open-sourced RFdiffusion3 on December 3, 2025, ahead of this comprehensive paper publication.
De novo design of allosterically controlled enzymes🚀 New preprint from David Baker!🚀1. The study presents a general de novo strategy to build compact enzymes whose catalytic activity is switched on or off by binding of a separately designed protein effector at a distal… pic.twitter.com/ph8ZWkdbge
Professor David Baker of the University of Washington serves as the corresponding author of the study. Baker received the 2024 Nobel Prize in Chemistry for using computers to predict protein structures, with the RosettaFold system standing as one of his landmark achievements.
How AlphaFold3 Architecture Drives Faster Protein Generation
While earlier iterations including RFdiffusion1 (2023) and RFdiffusion2 (2025) relied on RosettaFold’s structure prediction network, RFdiffusion3 shifts to the network architecture of AlphaFold3. This updated framework represents each amino acid with up to 14 atoms, generating new structures by directly manipulating the three-dimensional coordinates of every atom.
The primary advantage of the new system lies in constraint-based design. Researchers can specify exact placement targets for required atoms while setting conditions for hydrogen bond formation and spatial distribution. Across evaluations utilizing 41 active sites of existing enzymes, RFdiffusion3 outperformed RFdiffusion2 in 37 instances. Furthermore, the system operates at roughly ten times the speed of its predecessor.
Laboratory Results Confirm Functional Enzyme Design
Real-world experiments validate the shift from merely predicting structures to engineering functional designs. The organophosphate-degrading enzyme and the peptide-cleaving enzyme built by RFdiffusion3 successfully carried out their intended reactions during laboratory testing. This atom-by-atom generative approach advances protein engineering into a phase defined by functional predictability and precision.
