RFdiffusion
Last reviewed
May 21, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 · 3,931 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 21, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 · 3,931 words
Add missing citations, update stale details, or suggest a clearer explanation.
RFdiffusion (short for RoseTTAFold diffusion) is a deep-learning system for de novo protein design developed at the University of Washington Institute for Protein Design (IPD) in David Baker's lab. It is built by fine-tuning the RoseTTAFold structure prediction network as a denoising diffusion probabilistic model that operates on residue rigid-body frames in SE(3)-equivariant space, generating new protein backbones conditioned on user-supplied design targets such as fold topology, symmetry, binding hotspots, or functional motifs.[^1][^2] Reported in Nature in 2023 by Joseph L. Watson, David Juergens, Nathaniel R. Bennett, Brian L. Trippe, Jason Yim and colleagues, the model achieved roughly an order-of-magnitude improvement in experimental success rates over previous physics-based and hallucination-based protein design methods.[^1][^3] RFdiffusion has been released under an open-source BSD license through the RosettaCommons GitHub organization and has spawned a family of follow-up tools, including RFdiffusion All-Atom and RFantibody.[^4][^5][^6] The system is widely regarded as a landmark in computational biology and a defining example of how generative diffusion can be repurposed beyond image generation to a high-stakes scientific domain.[^23][^28]
| Property | Value |
|---|---|
| Type | Generative diffusion model for protein backbone design |
| Developer | Baker Lab, UW Institute for Protein Design, with collaborators at Columbia and MIT |
| Initial preprint | 10 December 2022 (bioRxiv 2022.12.09.519842)[^7] |
| Code release | March 2023 (open source, BSD)[^4] |
| Peer-reviewed publication | Nature vol. 620, pp. 1089-1100, 31 August 2023[^1] |
| Backbone architecture | Fine-tuned RoseTTAFold three-track network with SE(3) equivariant attention[^1][^2] |
| Sequence design partner | ProteinMPNN, with AlphaFold2 used for in silico filtering[^8] |
| License | BSD (free for non-profit and commercial use)[^5] |
| Repository | github.com/RosettaCommons/RFdiffusion[^5] |
De novo protein design, the task of computationally generating an amino-acid sequence that folds to a target structure and performs a desired function, was historically dominated by physics-based search inside the Rosetta modeling suite. The traditional Rosetta pipeline used Monte Carlo sampling against a hand-crafted energy function, then exhaustive experimental screening of tens of thousands of candidates per project, and even then yielded only modest success rates on hard problems such as protein binder design.[^3][^9]
The launch of AlphaFold in 2020 and RoseTTAFold in 2021 demonstrated that deep neural networks could predict native protein structures with near-experimental accuracy, suggesting that the same networks could be inverted to drive design.[^10][^11] Two early "structure-prediction-as-design" approaches emerged in the Baker lab: hallucination, which uses gradient descent through the network to optimize a sequence whose predicted structure matches a target, and inpainting (RFjoint), which conditions RoseTTAFold on partial structural inputs and fills in the rest.[^3] Both approaches scaled poorly: hallucination became unstable beyond about 100 residues, and inpainting could not perform unconstrained generation of long, novel structures.[^12]
In parallel, diffusion models led by DDPM (Ho, Jain and Abbeel, 2020) had become the dominant generative paradigm for images, powering systems such as Stable Diffusion, DALL E, and Midjourney.[^13] The Baker-lab team explicitly drew the analogy: just as a text-to-image diffusion model can be conditioned by a prompt to denoise a random pixel grid into a coherent picture, a structure diffusion model could be conditioned on a protein-design specification to denoise random residue frames into a coherent protein backbone.[^3][^15] Several groups, including Trippe et al. on bioRxiv and Yim et al. on SE(3) diffusion, showed that the same denoising framework could be ported to protein backbones expressed as sequences of rigid frames.[^14] RFdiffusion combined these threads. By fine-tuning RoseTTAFold's pretrained weights as a denoiser, the IPD team reused all of the structural inductive biases the network had already learned for prediction, while adapting it to the very different task of producing realistic protein geometries from pure noise under user-specified conditioning.[^1][^7] The development team has described RFdiffusion's design as the moment generative diffusion fully arrived in structural biology, and the broader research literature treats it as the canonical example of repurposing a structure-prediction backbone as a generative model.[^2][^23]
| Date | Event |
|---|---|
| 1 December 2022 | IPD posts a blog announcement, "A diffusion model for protein design"[^15] |
| 10 December 2022 | Preprint posted on bioRxiv as "Broadly applicable and accurate protein design by integrating structure prediction networks and diffusion generative models"[^7] |
| March 2023 | Code released under open-source BSD license on GitHub and via ColabFold[^4] |
| 11 July 2023 | Advance Online Publication of the Watson et al. Nature paper[^1] |
| 31 August 2023 | Issue date of the Nature paper, vol. 620, pp. 1089-1100[^1] |
| 9 October 2023 | RoseTTAFold All-Atom (RFAA) and RFdiffusion All-Atom (RFdiffusionAA) posted on bioRxiv[^16] |
| 19 December 2023 | Vázquez Torres et al. Nature report picomolar peptide binders via partial diffusion[^17] |
| 7 March 2024 | RoseTTAFold All-Atom published in Science (Krishna et al.)[^6] |
RoseTTAFold is a three-track architecture that exchanges information between a one-dimensional sequence track, a two-dimensional pairwise residue track, and a three-dimensional structural track. During structure prediction, the structure track outputs the rigid frame of each residue, defined by a backbone Cα coordinate and an N to Cα to C orientation matrix that fixes the rotation of the residue in space.[^1][^11] RFdiffusion keeps this exact frame representation but re-purposes the network as the denoising function of a diffusion probabilistic model.[^1][^2]
In the forward (noising) process, a true protein structure from the Protein Data Bank is gradually corrupted by independent perturbations of each residue's translation and rotation: 3D Gaussian noise on the Cα coordinate, and Brownian motion on the manifold of rotation matrices SO(3) for the orientation.[^1][^2] After roughly 200 timesteps the structure is reduced to an effectively random cloud of frames.[^1] In the reverse (denoising) process, RFdiffusion is asked, at each timestep, to predict the corresponding noiseless structure given the current noisy frames and any conditioning information. The predicted structure is then mixed back into the noise schedule to produce the input to the next step. Iterating from random noise to t equals 0 yields a fully formed protein backbone.[^1][^2]
Because rotating or translating a protein leaves its identity unchanged, the denoiser is built to be SE(3)-equivariant: applying a rigid transformation to the input frames produces exactly the same transformation in the output. RFdiffusion inherits this property from RoseTTAFold's invariant-point-attention (IPA) modules and from NVIDIA's SE(3)-Transformer implementation, which it uses inside its equivariant graph operations.[^2][^5] Translations are perturbed and predicted in Cα coordinates with mean-squared-error loss, while rotations are handled on SO(3).[^1][^2]
A central trick is self-conditioning, inspired by AlphaFold2's "recycling": at each denoising step the network is given not only the noisy frames but also the structure it predicted at the previous step. This dramatically stabilizes long trajectories and is critical for achieving high in silico designability on long monomers.[^1][^2] The model is fine-tuned from pretrained RoseTTAFold weights rather than trained from scratch, which both shortens training and lets the network exploit structural priors learned from the entire PDB.[^1] Training uses a mean-squared-error loss on Cα coordinates and on the rotation prediction in SO(3), rather than the frame-aligned point error (FAPE) used by AlphaFold2 for structure prediction; the authors found that this simpler objective worked better for the denoising setting and that the resulting model recovered the local stereochemistry of natural proteins (Ramachandran statistics, sensible secondary-structure proportions, well-packed cores) despite never being trained explicitly on those properties.[^1][^2] Training is performed on roughly 100,000 high-resolution Protein Data Bank chains between 60 and 512 residues, with various data-augmentation schemes including random cropping and motif-conditioning examples.[^2]
Unconditional RFdiffusion just denoises random noise into any plausible monomer. The real power of the system is that the same network supports a wide range of conditioning signals, all expressed by selectively freezing or partially noising features of the input frames or pair representation:[^1][^2]
The pipeline is two-stage. RFdiffusion outputs a backbone; ProteinMPNN, a separate graph-based sequence design network from the Baker lab, then proposes amino-acid sequences for that backbone; AlphaFold2 or RoseTTAFold predicts the structure of each candidate sequence and filters those whose prediction does not match the design model.[^1][^8] This sequence design and prediction loop has become the de facto standard "RFdiffusion plus ProteinMPNN plus AF2" workflow used across the field.[^8][^9]
The Nature paper accompanies its computational claims with a substantial wet-lab validation campaign, in which hundreds of designs across multiple problem classes were expressed and characterized.[^1][^3]
RFdiffusion produces diverse monomeric structures from roughly 50 to 600 residues, with AlphaFold2 self-consistency RMSDs of about 0.5 to 1.7 Å up to 400 residues, after which AlphaFold2 agreement begins to deteriorate.[^1][^18] Generation is fast: a 100-residue protein can be generated in about 11 seconds on an NVIDIA RTX A4000 GPU, compared to about 8.5 minutes for RoseTTAFold hallucination at the same length.[^18]
On a published benchmark of 25 motif-scaffolding problems, RFdiffusion solved 23, including challenging tasks such as scaffolding the catalytic site of a retroaldolase and the binding loops of complex epitopes.[^1] Designed proteins built around immunogenic motifs were experimentally shown to recapitulate the structure of the motif within near-atomic accuracy.[^1]
RFdiffusion produced cyclic (Cn), dihedral (Dn), tetrahedral, octahedral, and icosahedral assemblies that matched their designed symmetry by negative-stain and cryo-electron microscopy in dozens of independent cases.[^1][^3] Symmetric metal-binding cages with C4 and other symmetries were designed to coordinate ions such as Ni2+ with low-nanomolar dissociation constants.[^1]
The most discussed result is protein binder design. RFdiffusion was used to generate de novo binders against multiple unrelated targets, including influenza hemagglutinin (HA), interleukin-7 receptor alpha (IL-7Rα), insulin receptor (InsR), PD-L1, and tropomyosin receptor kinase A (TrkA).[^1] Across these targets, experimental hit rates from a single design pool reached approximately 19 percent for some campaigns, with measured affinities ranging from nanomolar to picomolar, and roughly a tenfold improvement over the prior Cao et al. Rosetta-based binder design pipeline.[^1][^3][^9] A cryo-EM structure of an RFdiffusion-designed HA binder bound to influenza hemagglutinin matched the design model within 0.63 Å backbone RMSD, confirming atomic-level accuracy.[^1][^3]
In binder mode, the user provides the target structure, optional hotspot residues, and a desired length for the binder. RFdiffusion then grows a binder backbone that contacts the chosen hotspots. The IPD team highlighted the practical implication: with prior Rosetta-based binder design, researchers typically had to express and test tens of thousands of designs per campaign to discover a single confirmed binder, whereas the RFdiffusion pipeline could deliver experimentally validated binders from pools as small as 96 to a few hundred designs.[^3] This compression of the design-build-test loop is one of the central reasons RFdiffusion is described in popular coverage as transforming protein engineering rather than merely improving it.[^23][^26]
The paper demonstrates that RFdiffusion can scaffold catalytic constellations of residues for multiple enzyme classes, holding the spatial geometry of the active site fixed while designing a stable surrounding fold.[^1] This established the foundation for follow-up work on de novo enzyme design with subsequent RFdiffusion variants, and dovetails with parallel efforts on enzyme design that leveraged the underlying RoseTTAFold network for inverse-folding tasks.[^19]
The authors report that RFdiffusion designs are highly designable (a high fraction of generated backbones admit at least one sequence whose predicted structure agrees with the design) and at the same time produce backbones that are dissimilar to anything in the PDB by TM-score, indicating genuine de novo generation rather than retrieval or paraphrase of known structures.[^1][^18] On the unconditional generation benchmark, designability at length 100 reaches scTM scores around 0.97 and remains above 0.9 even at length 500, with novelty (lowest TM-score to PDB) decreasing monotonically as length increases, consistent with the intuition that longer designs explore more novel topologies.[^2][^18]
RFdiffusion is best understood as the founding member of a family of related Baker-lab generative models.
A companion paper by Nathaniel R. Bennett, Brian Coventry, Inna Goreshnik and colleagues, "Improving de novo protein binder design with deep learning" (Nature Communications, 6 May 2023), showed that adding AlphaFold2 or RoseTTAFold post-hoc filters and ProteinMPNN sequence design on top of an otherwise Rosetta-style binder pipeline already raises experimental success rates roughly tenfold.[^9] This work supplied much of the validation methodology used in the RFdiffusion Nature paper.
Vázquez Torres, Leung, Venkatesh and collaborators reported, also in Nature in December 2023, the use of RFdiffusion (with partial-diffusion refinement) to generate de novo binders against intrinsically disordered bioactive peptides such as parathyroid hormone and glucagon, reaching picomolar affinities.[^17]
Limitations of the original RFdiffusion (it sees only the protein backbone, not ligands, ions, nucleic acids, or post-translational modifications) motivated an "all-atom" successor. Krishna, Wang, Ahern and colleagues, "Generalized biomolecular modeling and design with RoseTTAFold All-Atom" (Science vol. 384, 7 March 2024), introduced RoseTTAFold All-Atom (RFAA), which models entire biological assemblies including small molecules, metals, DNA, RNA and covalent modifications, and RFdiffusion All-Atom (RFdiffusionAA), which extends the RFdiffusion design framework to grow new protein binders around small-molecule targets.[^6][^16] The team experimentally validated proteins that bind the cardiac drug digoxigenin, the cofactor heme, and bilin chromophores relevant to engineered photosynthesis.[^6][^16]
Subsequent versions, RFdiffusion2 (inference code released September 2025) and RFdiffusion3 (announced December 2025), generalize the all-atom design framework further. They report improved enzyme design from sequence-level prompts, as well as designed DNA-binding proteins.[^20][^21]
RFantibody, a Baker-lab system that fine-tunes the RFdiffusion machinery on antibody Fv complexes, was published in Nature in late 2025 and represents the first reported atomically accurate de novo design of antibodies entirely from scratch.[^22]
RFdiffusion has rapidly become a default tool inside academic and industrial protein engineering. By the time of the paper's publication in mid-2023, multiple Baker-lab spinouts (including Xaira Therapeutics, Vilya, A-Alpha Bio, and others affiliated with the IPD) had begun integrating RFdiffusion or its derivatives into their internal therapeutic-discovery pipelines, and use spread quickly into academic groups outside Seattle. As of 2024, the GitHub repository was among the most-starred biology-oriented machine-learning projects on the platform, and NVIDIA, Microsoft, and other cloud vendors began offering hosted versions through their managed scientific-AI services.[^4][^24][^23]
RFdiffusion is powerful but not magic, and several limitations are well documented in both the paper and follow-up work.
| Method | Year | Approach | Notes |
|---|---|---|---|
| Rosetta-based design | 1990s onward | Physics-based Monte Carlo over fragments | Workhorse before deep learning; very large screening libraries[^3] |
| RoseTTAFold hallucination | 2021-2022 | Gradient descent through prediction net | Effective up to about 100 residues[^3] |
| RFjoint inpainting | 2022 | Conditional fill-in of structure with RoseTTAFold | Strong on partial-structure tasks but not unconstrained generation[^3] |
| Chroma (Generate Biomedicines) | 2023 | Programmable diffusion model for proteins | Concurrent with RFdiffusion[^23] |
| FrameDiff / SE(3) diffusion (Yim et al.) | 2023 | SE(3) diffusion trained from scratch | Inspired RFdiffusion's noise schedule[^14] |
| RFdiffusion | 2023 | RoseTTAFold fine-tuned as SE(3) denoiser | Roughly 10x experimental improvement over prior Baker-lab pipelines[^1][^3] |
| RFdiffusion All-Atom | 2024 | All-atom extension with small molecules and cofactors | Protein-ligand binder design[^6] |
| RFantibody | 2025 | Antibody-specialized RFdiffusion | First atomically accurate de novo antibody design[^22] |
The Nature paper attracted broad media attention. The news article in the same issue, "AI tools are designing entirely new proteins that could transform medicine" (12 July 2023), described diffusion-based protein design as representing an "explosion in capabilities" and surveyed work from the Baker lab and competing groups, quoting Gevorg Grigoryan of Generate Biomedicines and Mohammed AlQuraishi at Columbia among others.[^26] Quanta Magazine's feature "How AI Revolutionized Protein Science, but Didn't End It" (Yasemin Saplakoglu, 26 June 2024) likewise placed RFdiffusion at the center of the post-AlphaFold wave of generative design, quoting David Baker and others on the shift from prediction to programmable design.[^23] Trade publications such as Chemical & Engineering News covered the broader rise of generative protein design and the role of RFdiffusion within it.[^27] David Baker himself was a co-recipient of the 2024 Nobel Prize in Chemistry, with the prize citation explicitly referencing the lab's work on computational protein design, including the line of work that produced RFdiffusion.[^28]
The community impact is also visible in the academic literature. RFdiffusion accumulated thousands of citations within the first two years of publication and became a baseline against which essentially every subsequent generative protein-design method is compared, including flow-matching variants, sparse-denoising models, antibody-specific systems, and combined sequence-structure co-design frameworks.[^2] Critics have noted that the reliance on a single hand-engineered pipeline (RFdiffusion plus ProteinMPNN plus AlphaFold2 filtering) can mask the limits of each component, and that head-to-head benchmarks on functional binder design have at times favored newer AlphaFold-driven design methods such as BindCraft for certain target classes.[^25] Despite this, RFdiffusion remains the canonical reference point for diffusion-based protein generation and the entry point for most newcomers to deep-learning-based protein design.[^23]