IT之家 · 智能时代 · 10/19/2026, 07:22:56
RFdiffusion3 Achieves Atomic-Level Protein Design Validated in Science
The University of Washington and NVIDIA team published RFdiffusion3 in Science, enabling atomic-level modeling of amino acid coordinates to design functional proteins. Experimental validation confirms that enzymes designed by the system exhibit expected catalytic activity in real reactions, with generation speed approximately ten times faster than previous versions.
SOURCE COVERAGEFull Original Report
October 19 — The detailed design specifications for RFdiffusion3, a generative AI system capable of designing proteins and nucleic acids at the level of individual atomic combinations, have been officially released. Proteins newly designed using RFdiffusion3 have demonstrated their intended functions in real-world chemical reactions.
An international collaborative research team comprising institutions such as the University of Washington (USA), the University of Oxford (UK), and NVIDIA has developed this system, which models individual atoms to design proteins and nucleic acids. The related research findings were published on October 16 (local time) in the international journal Science.
IT Home noted that the research team open-sourced RFdiffusion3 on December 3, 2025. The paper published in Science provides a detailed explanation of RFdiffusion3's design principles.
AI-based protein research has evolved from predicting the three-dimensional structures of existing proteins to designing novel proteins with specific target structures and functions.
David Baker, a professor at the University of Washington, is the corresponding author of this RFdiffusion3 study. He was awarded the 2024 Nobel Prize in Chemistry for his contributions to computer-based prediction of protein structures. One of Baker’s landmark achievements is the development of RosettaFold, an AI system for protein structure prediction.
RosettaFold learns the spatial relationships between amino acids that constitute proteins and uses this information to predict their three-dimensional structures. Both RFdiffusion1, released in 2023, and RFdiffusion2, released in 2025, were developed based on the structure prediction networks used by RosettaFold.
RFdiffusion utilizes diffusion models to determine the positions of amino acids, thereby completing protein design. Diffusion models are a class of AI models that learn the distribution patterns of existing data to generate new data.
RFdiffusion3 was developed by referencing the network architecture of AlphaFold3, another protein structure prediction program. RFdiffusion3 can represent a single amino acid using up to 14 atoms, designing novel protein structures by directly adjusting the three-dimensional coordinates of each atom.
The research team achieved precise design of enzyme active sites using this system. Enzymes catalyze various chemical reactions; they are proteins that act as biological catalysts, accelerating the rate of various chemical reactions within organisms.
When researchers specify not only the positions of atoms required for a reaction but also conditions for hydrogen bond formation and the distribution range of specific atoms within the protein, RFdiffusion3 generates protein structures that satisfy all constraints.
Researchers compared the design performance against 41 active sites of existing enzymes. The results showed that RFdiffusion3 outperformed RFdiffusion2 in 37 of the test cases. It generates protein structures approximately ten times faster than RFdiffusion2.
Paraoxon-degrading enzymes and peptide-cleaving enzymes designed by RFdiffusion3 successfully completed their intended chemical reactions in real-world experiments. The research team stated: "In the future, RFdiffusion3 will be widely applied to various biomolecular design tasks and related application fields."
References: