New RFdiffusion3 AI Speeds Up Protein Design 10 Times
The open-source system enables researchers to design custom enzymes with higher structural accuracy for chemical catalysis.
Updated on Oct. 16, 2026 in Chemistry
Live Poll
Is the rapid development of AI for biological and medical research a net benefit for society?
An international research team has released the open-source RFdiffusion3 AI system, designed to generate custom protein structures for chemical catalysis. The system, detailed in the journal Science, demonstrates significant gains in speed and structural accuracy over its predecessor.
Why it matters
This development accelerates the field of protein engineering by enabling the precise construction of enzyme active sites. By improving the speed and capability of designing complex molecular structures, researchers can more efficiently create synthetic proteins for targeted chemical reactions.
RFdiffusion3 models proteins by representing single amino acids with up to 14 atoms, outperforming the previous iteration in 37 of 41 active site design cases. The system also achieves a 10-fold increase in generation speed compared to RFdiffusion2.
The players
University of Washington
A research university recognized for its advanced computational protein design work, including the development of the RFdiffusion series.
University of Oxford
A leading research institution involved in the international collaboration to refine protein structure modeling.
NVIDIA
A technology company specializing in high-performance computing hardware and AI research that collaborated on the system design.
David Baker
A researcher who received the 2024 Nobel Prize in Chemistry for his contributions to computational protein structure prediction.
The details
RFdiffusion3 utilizes a diffusion model—a machine learning architecture that learns to generate complex data by reversing a process of adding noise to structured patterns. The system integrates constraints for specific atom positioning and hydrogen bond formation to ensure protein structures are chemically viable. By combining the positions of individual atoms and amino acids, the model allows for the high-precision design of enzyme active sites—the specific regions where enzymes catalyze chemical reactions.
Timeline
2023: Release of the original RFdiffusion1 system.
2024: David Baker receives the Nobel Prize in Chemistry.
2025: Release of RFdiffusion2.
December 3, 2025: Public open-source release of RFdiffusion3.
October 16, 2026: Official publication of the research in the journal Science.
The Tech Race
This release follows the Nobel-recognized work of David Baker and builds upon the rapid evolution from the 2023 RFdiffusion1 milestone. It represents the latest step in a competitive landscape focused on increasing the structural complexity and generation speed of synthetic biomolecules.
The system is currently available as open-source software for researchers and synthetic biology developers. Users can utilize the platform to design and verify protein structures that meet specific hydrogen bond and atom positioning constraints.
The takeaway
The evolution of the RFdiffusion series confirms that diffusion-based models are becoming the standard for rapid, high-precision protein engineering. Researchers should monitor the subsequent publication of validation studies to see how these designs perform in wet-lab experimental settings.
Further reading
For more information on the evolving tools shaping molecular design, see Chemistry.
Source note: This article includes information reported by Dongascience.
Live Poll
Is the rapid development of AI for biological and medical research a net benefit for society?
