Skip to content
ContentLora

    Tip: press / anywhere to search.

    technology

    RFdiffusion

    Also known as RoseTTAFold diffusion, RFdiffusion2, RFdiffusion3, RFD3

    RFdiffusion is a family of generative models from the University of Washington's Institute for Protein Design that invent new proteins by refining random noise into designed structures.[1] The third version, released open source in December 2025, designs at the level of individual atoms and can target DNA and small molecules.[2][3]

    Editor reviewedUpdated AI for scienceLife sciencesArtificial intelligence
    Key facts

    RFdiffusion is an open tool for designing proteins from scratch, released by the University of Washington’s Institute for Protein Design.[2] The field it belongs to, computational protein design, won David Baker half of the 2024 Nobel Prize in Chemistry.[4][5]

    How it works

    The team took RoseTTAFold, a structure prediction network, and fine-tuned it to remove noise from protein structures. Run step by step from pure noise, the network produces new protein backbones, much as image generators produce new pictures.[1] Designers can condition the process on a target: a surface to bind, a symmetry to adopt, or an enzyme active site to hold in place.[6]

    Evidence it works

    The July 2023 Nature paper reported strong results across binder design, symmetric oligomers, enzyme active-site scaffolding and motif scaffolding, and characterised hundreds of designs in the lab.[6] A cryo-electron microscopy structure of one designed binder, bound to influenza haemagglutinin, was nearly identical to the computer model.[7]

    RFdiffusion3

    Released in December 2025, RFdiffusion3 treats individual atoms as the units it designs, rather than amino-acid residues. It can generate proteins that bind DNA, other proteins and small molecules, and it runs about ten times faster than RFdiffusion2.[2][3] Training code and weights are public through the Rosetta Commons Foundry.[2] It is a new codebase rather than an update, sharing no code with RFdiffusion or RFdiffusion2.[8] Its developers point to uses such as enzymes that break down microplastics, synthetic transcription factors for gene therapy and environmental biosensors.[9]

    Part of a wider toolkit

    Open design models are increasingly run by automated agents. In August 2026 Anthropic reported that Claude models, chaining open-source design tools, produced binders for 14 of 15 targets in lab tests run by Adaptyv Bio and Twist Bioscience.[10] Anthropic reported hit rates of 22–35%, against roughly 10–15% it describes as typical, though success ranged from 90% to 0% depending on the target.[11][12] These are company-reported results, not peer-reviewed.[11]

    The wider design field

    RFdiffusion is one of several generative design tools. Chai Discovery’s Chai-2, described in a July 2025 preprint, targets antibodies: the company reported a 16% hit rate for fully de novo designs and at least one binder for half of 52 targets, testing 20 or fewer designs per target.[13] Design also depends on scoring how well a candidate binds. The open Boltz-2 model predicts structure and binding affinity together, and its authors report accuracy approaching physics-based free-energy calculations at a fraction of the computing cost.[14]

    Enzymes are a demanding case because they must position atoms precisely to make and break chemical bonds, which is the problem RFdiffusion3’s atom-level design targets.[8] In October 2026 the US Genesis Mission gave a Phase II award to a University of Washington project to build tools that predict and design functional enzymes.[15] The RFdiffusion paper itself had reported enzyme active-site scaffolding as one of its tasks.[6]

    Why it matters

    RFdiffusion and alphafold together show the two halves of modern protein AI: predicting nature’s structures and writing new ones. Prediction models such as AlphaFold learned from the experimental archive in the protein-data-bank.[16]

    Questions readers ask

    What can RFdiffusion design?

    The original model handled binders, symmetric assemblies, metal-binding proteins and enzyme active-site scaffolds; RFdiffusion3 can design proteins that interact with almost any molecule found in cells.[6][2]

    Do RFdiffusion designs work in the lab?

    The 2023 paper characterised hundreds of designs experimentally, and a cryo-EM structure of one binder matched its design model almost exactly.[6][7]

    Is RFdiffusion open source?

    Yes. RFdiffusion3 training code and weights are on GitHub through the Rosetta Commons Foundry.[2]

    Sources

    Each numbered claim is a statement we checked against the sources listed with it. Status shows how well established it is.

    1. [1]

      RFdiffusion, published in Nature in July 2023, was made by fine-tuning the RoseTTAFold structure prediction network on structure denoising, producing a generative model of protein backbones. confirmedas of 2023-07-11

    2. [2]

      In December 2025 the Institute for Protein Design released RFdiffusion3, an all-atom model that designs proteins interacting with any type of molecule commonly found in cells, with training code and weights available through the Rosetta Commons Foundry. confirmedas of 2025-12-03

    3. [3]

      The Institute for Protein Design says RFdiffusion3 treats individual atoms as the units being designed and runs ten times faster than RFdiffusion2. confirmedas of 2025-12-03

    4. [4]

      The 2024 Nobel Prize in Chemistry went half to David Baker for computational protein design and half jointly to Demis Hassabis and John Jumper for protein structure prediction. confirmedas of 2024-10-09

    5. [5]

      The Nobel committee credited David Baker with building entirely new kinds of proteins. confirmedas of 2024-10-09

    6. [6]

      The RFdiffusion paper reported strong results on binder design, symmetric oligomer design, enzyme active-site scaffolding and motif scaffolding, with hundreds of designs characterized experimentally. confirmedas of 2023-07-11

    7. [7]

      A cryo-electron microscopy structure of an RFdiffusion-designed binder bound to influenza haemagglutinin was nearly identical to the design model. confirmedas of 2023-07-11

    8. [8]

      The Institute for Protein Design says DNA-binding proteins and enzymes have long required designers to specify atom positions precisely, and that RFdiffusion3 shares no code with RFdiffusion or RFdiffusion2. confirmedas of 2025-12-03

    9. [9]

      The Institute for Protein Design says RFdiffusion3 opens possibilities such as enzymes that break down microplastics, synthetic transcription factors for gene therapy and environmental biosensors. confirmedas of 2025-12-03

    10. [10]

      In August 2026 Anthropic reported that Claude models, using open-source protein design tools, designed binders against 14 of 15 targets, with designs produced and tested by Adaptyv Bio and Twist Bioscience. reportedas of 2026-08-18

    11. [11]

      Anthropic reported hit rates of 22% to 35% for Claude-led binder design depending on setup, against the 10–15% it describes as typical for protein design campaigns, with 354 binders from 1,320 designs. reportedas of 2026-08-18

    12. [12]

      In Anthropic's campaign, hit rates for individual targets ranged from 90% to 0%. reportedas of 2026-08-18

    13. [13]

      In a preprint posted in July 2025, Chai Discovery reported that its Chai-2 model achieved a 16% hit rate in fully de novo antibody design and found at least one binder for 50% of 52 targets, testing 20 or fewer designs per target, none of which had an existing antibody binder in the Protein Data Bank. confirmedas of 2025-07-06

    14. [14]

      Boltz-2, described in a June 2025 preprint by MIT and Recursion researchers, predicts both complex structures and small molecule–protein binding affinity, and its authors say it is the first AI model to approach free-energy perturbation accuracy while being at least 1,000 times more computationally efficient. confirmedas of 2025-06-18

    15. [15]

      DOE's October 2026 Phase II Genesis Mission awards include a Commonwealth Fusion Systems digital twin of a fusion demonstration device, University of Washington enzyme design tools, a UC San Diego project to expand the RNA structure database five-fold and a Fermilab project to design rugged microchips for extreme environments. confirmedas of 2026-10-08

    16. [16]

      AlphaFold2 was trained on a copy of the Protein Data Bank downloaded in August 2019. confirmedas of 2021-07-15

    Revision history (2)
    1. Page created.
    2. Added the wider generative design field (Chai-2 antibodies, Boltz-2 affinity prediction), the Genesis Mission enzyme-design award and links to Anthropic and Twist Bioscience.

    Created Oct 10, 2026. Last reviewed by an editor on Oct 10, 2026. Next scheduled review: Jan 10, 2027.

    Cite this page

    "RFdiffusion." ContentLora, updated Oct 10, 2026. https://contentlora.com/wiki/rfdiffusion

    Spotted an error? Suggest a correction or emailcorrections@contentlora.com.

    Keep exploring

    Connections
    AlphaFold Fellow Nobel-recognised p…Anthropic Used in Claude-led design…Twist Bioscience Designs made and tested by
    ☰ Contents