A generative AI system called RFdiffusion3 that can design proteins and nucleic acids at the level of individual atoms has had its complete design scheme officially made public recently. Even more remarkable, proteins newly designed with it have already delivered their expected functions in real chemical reactions.
An international team composed of the University of Washington in the US, the University of Oxford in the UK, NVIDIA and other institutions built this system, which can model individual atoms and is specifically used to design proteins and nucleic acids; the related findings were published in the international journal Science on the 16th local time. In fact, the team had already open-sourced RFdiffusion3 as early as December 3, 2025, and this paper lays out the design principles in full.
Nobel laureate at the helm, ten times faster than the previous generation
The corresponding author of this study is David Baker, a professor at the University of Washington — he won the 2024 Nobel Prize in Chemistry for using computers to predict protein structures, and one of his landmark achievements is the protein structure prediction system RosettaFold. RFdiffusion1 (2023) and RFdiffusion2 (2025) were both built on RosettaFold's structure prediction network; RFdiffusion3, however, instead draws on AlphaFold3's network architecture, characterizing a single amino acid with up to 14 atoms and shaping new structures by directly adjusting the three-dimensional coordinates of each atom.
Its skill lies in "designing with constraints." If researchers specify not only where the atoms needed for a reaction should land, but also set hydrogen bond formation conditions and the distribution ranges of particular atoms inside the protein, the system can produce structures that satisfy all the constraints. The team compared it against 41 active sites of existing enzymes, and in 37 cases RFdiffusion3 outperformed RFdiffusion2, with a generation speed about ten times that of the previous generation. The paraoxon-degrading enzyme and peptide-cleaving enzyme it designed both successfully completed their expected reactions in real experiments. Going from "predicting structures" to "building structures by function," this atom-by-atom sculpting AI is pushing protein engineering toward unprecedented precision.