AI is entering various fields, even working in laboratories.Author and source: AIBase
A generative AI system called RFdiffusion3, capable of designing proteins and nucleic acids at the atomic level, has officially released its complete design specifications. Even more impressively, proteins newly designed using this system have already demonstrated their intended functions in real chemical reactions.
An international team comprising institutions such as the University of Washington in the United States, the University of Oxford in the United Kingdom, and NVIDIA developed this system capable of modeling individual atoms, specifically designed for protein and nucleic acid design. The findings were published on the 16th local time in the international journal Science. In fact, the team had already open-sourced RFdiffusion3 on December 3, 2025; this paper now thoroughly explains the underlying design principles.
Crafted by a Nobel laureate, ten times faster than the previous generation.
The corresponding author of this study is Professor David Baker of the University of Washington, who was awarded the 2024 Nobel Prize in Chemistry for using computers to predict protein structures; one of his landmark achievements is the protein structure prediction system RosettaFold. Both RFdiffusion1 (2023) and RFdiffusion2 (2025) are built upon RosettaFold’s structure prediction network, while RFdiffusion3 adopts the network architecture of AlphaFold3, representing each amino acid with up to 14 atoms and generating new structures by directly adjusting the 3D coordinates of each atom.
Its strength lies in "design with constraints." If researchers specify not only where the required atoms should land but also set conditions for hydrogen bond formation and the spatial distribution of specific atoms within the protein, the system can generate structures that satisfy all constraints. In comparisons using 41 active sites of existing enzymes, RFdiffusion3 outperformed RFdiffusion2 in 37 cases and operates at roughly ten times the speed of its predecessor. The organophosphate-degrading enzyme and peptide-cleaving enzyme designed by RFdiffusion3 successfully performed their intended reactions in real-world experiments. Moving from "predicting structures" to "designing structures by function," this atom-by-atom AI is pushing protein engineering to unprecedented levels of precision.