RoseTTAFold
Last reviewed
May 21, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 ยท 4,634 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 ยท 4,634 words
Add missing citations, update stale details, or suggest a clearer explanation.
RoseTTAFold is a deep-learning system for predicting the three-dimensional structures of proteins from amino acid sequence, developed at the Institute for Protein Design (IPD) at the University of Washington and released publicly on 15 July 2021.[1][2] Led by postdoctoral researcher Minkyung Baek in the laboratory of David Baker, the project introduced a "three-track" neural network that simultaneously processes one-dimensional sequence information, two-dimensional residue-pair geometries, and three-dimensional atomic coordinates, with information passing between all three representations during inference.[3] It appeared in the journal Science shortly after DeepMind's AlphaFold 2 and reproduced much of that system's accuracy while running in roughly ten minutes on a single consumer GPU, making accurate structure prediction broadly accessible.[1][4] RoseTTAFold has since spawned a family of follow-up systems including RoseTTAFold2, RoseTTAFoldNA, RoseTTAFold All-Atom, and the RFdiffusion generative design framework, and it played a central role in the work for which David Baker was awarded one half of the 2024 Nobel Prize in Chemistry.[5][6]
| Field | Value |
|---|---|
| Type | Protein structure prediction neural network |
| Developer | Baker Lab / Institute for Protein Design, University of Washington |
| Lead author | Minkyung Baek |
| Initial release | 15 July 2021 (Science publication and GitHub) |
| Original paper | Baek et al., Science, 2021 (DOI 10.1126/science.abj8754) |
| Architecture | Three-track network (1D sequence, 2D pair, 3D coordinates) with SE(3) equivariant transformer |
| Code license | MIT (code); separate non-commercial license for weights |
| Repository | github.com/RosettaCommons/RoseTTAFold |
| Web server | Robetta (robetta.bakerlab.org) |
| Follow-ups | RoseTTAFoldNA, RoseTTAFold2, RoseTTAFold All-Atom, RFdiffusion, RFdiffusionAA |
RoseTTAFold sits at the convergence of two long-running research programs at the University of Washington. The first is the Rosetta software suite, which originated in the Baker laboratory in the late 1990s as a Monte Carlo fragment assembly framework for predicting and designing protein structure.[7] The second is the use of deep learning to predict interresidue contacts and geometries from multiple sequence alignments, an approach that matured rapidly after 2018. Together these two lines led to a sequence of named systems in which each generation enlarged the role of neural networks at the expense of traditional energy-based search.
The immediate predecessor of RoseTTAFold was trRosetta, introduced by Jianyi Yang, Ivan Anishchenko, Hahnbeom Park, Sergey Ovchinnikov, and David Baker. The trRosetta paper "Improved protein structure prediction using predicted interresidue orientations" appeared in the Proceedings of the National Academy of Sciences on 21 January 2020 (volume 117, pages 1496 to 1503).[8] trRosetta used a residual convolutional network on coevolutionary features to predict not only interresidue distances but also the relative orientations of backbone frames, then fed those predictions as restraints into Rosetta minimization. At the 14th Critical Assessment of Structure Prediction (CASP14) competition in late 2020 the BAKER-trRosetta server ranked second among automated server groups, with only DeepMind's AlphaFold 2 doing meaningfully better.[8]
The CASP14 results made clear that DeepMind's system represented a step-change in accuracy. While the broader community waited for AlphaFold 2's full architecture to be disclosed, the Baker group set out to match its performance using publicly available descriptions and reasonable architectural guesses. Minkyung Baek led that effort. The team converged on what they called a "three-track" architecture, which extended AlphaFold 2's two-track (sequence and pair) design by adding an explicit third track that operated on 3D atomic coordinates and was connected to the other two by an SE(3) equivariant transformer.[9][3]
The team released the RoseTTAFold preprint on bioRxiv on 14 June 2021 (manuscript 2021.06.14.448402), about a month before the AlphaFold 2 paper was published in Nature.[10] The Science version appeared on 15 July 2021 alongside the open-source code drop on GitHub. The simultaneity of the publications was not coincidental: DeepMind's AlphaFold 2 paper in Nature also appeared in mid-July 2021, and the Baek paper explicitly acknowledged that the AlphaFold 2 design at CASP14 inspired RoseTTAFold's architectural choices.[1][4]
The defining technical idea of RoseTTAFold is its three-track structure. The network jointly maintains and refines three representations of a protein, with attention modules and explicit transforms exchanging information between them at every block.[3][9]
The first track is the 1D track. Its inputs are derived from a multiple sequence alignment (MSA) for the target protein, including per-position features such as conservation, gap statistics, and embedded queries used in row and column attention. This track is conceptually similar to the MSA representation in AlphaFold 2 and captures evolutionary signal that is otherwise difficult to recover from a single sequence.[3]
The second track is the 2D track. It carries a pairwise representation indexed by ordered pairs of residues. Features include predicted distances, dihedral and planar angles, and continuous embeddings of the relative geometry. Updates to the 2D track use triangle-style attention informed by the MSA track. This track is the descendant of the contact and orientation maps that earlier methods such as trRosetta produced as their final output.[3][8]
The third track is the 3D track, and it is the novel ingredient. RoseTTAFold maintains an explicit set of backbone frames and atomic coordinates as a first-class network state, not merely as a postprocessing step. The 3D track is updated using an SE(3) equivariant transformer originally introduced by Fabian Fuchs, Daniel Worrall, Volker Fischer, and Max Welling. Equivariance ensures that rotating or translating the input coordinates produces a correspondingly rotated or translated output, which is a desirable inductive bias for molecules in free space.[9]
In each block of the network, information flows in both directions between the three tracks. The 1D and 2D tracks bias the 3D track by providing predicted geometries; the 3D track in turn updates the 1D and 2D tracks by feeding back features computed from current coordinates. The Baek paper describes this design as allowing the network "to collectively reason about the relationships within and between sequences, distances, and coordinates".[3] The end-to-end version produces backbone coordinates directly. A separate "pyRosetta" mode uses RoseTTAFold's 2D predictions as restraints in a traditional energy minimization run, similar in spirit to trRosetta but with much stronger neural priors.
The original RoseTTAFold network was trained on protein chains from the Protein Data Bank, with template pairs and sequence alignments constructed from public sequence databases. Hardware constraints at the time of the original work meant that training operated on cropped subsequences of length 260 residues; for longer proteins, the system tiled inference and combined results.[1][3] The published model required roughly ten minutes to fold a typical small protein on a single NVIDIA RTX 2080 consumer GPU, a striking contrast with AlphaFold 2 in its initial form which was reported to need substantially more compute per prediction at CASP14.[4][1]
RoseTTAFold was released publicly on the same day as the Science paper. The reference implementation lives at github.com/RosettaCommons/RoseTTAFold, licensed under the MIT license, while the trained network weights are distributed under a separate non-commercial license that has been broadly described as the "Rosetta-DL Software" license.[11][1] The repository contains the network definition in PyTorch, scripts for MSA construction using HHblits and HHsearch over UniRef30 and the Big Fantastic Database (BFD), template search against PDB100, and end-to-end and pyRosetta-mode inference pipelines.[11]
Alongside the code, the Baker group made the network accessible through the existing Robetta web service at robetta.bakerlab.org. Robetta had hosted Baker-Lab tools since the early 2000s and added a RoseTTAFold option in 2021 that allowed researchers without local GPU resources to submit sequences for prediction. The IPD reported that within a month of release "over 4,500 proteins" had been submitted through the new server, including hundreds of structures of pharmaceutical or biological interest such as targets in lipid metabolism, inflammation, and oncology.[2]
The combination of a permissive code license, a free public web server, and consumer-grade GPU requirements was the proximate cause of RoseTTAFold's rapid uptake. By the time DeepMind released the AlphaFold 2 code and the AlphaFold Protein Structure Database in late July and 2021, RoseTTAFold was already in widespread use as a complementary or alternative tool, and several downstream open-source efforts drew on its codebase.[4][1]
In its CASP14 retrospective and in benchmarks reported in the Science paper, RoseTTAFold approached but did not equal AlphaFold 2's accuracy on the hardest "free modeling" targets. The Baek paper reports that on the CASP14 free modeling targets, RoseTTAFold attained accuracy comparable to the top-ranked groups other than DeepMind's, while remaining below AlphaFold 2's accuracy.[3] The paper attributes part of the gap to hardware-imposed crops on training input and to other choices that DeepMind had been free to optimize across a larger compute budget.[3]
On more standard CAMEO targets and a curated set of recent PDB chains, RoseTTAFold consistently produced models with backbone accuracy in the range that supports downstream uses such as molecular replacement in X-ray crystallography, function annotation, and homology-based design. The Baek paper also demonstrated that the network's 2D outputs were sufficient to identify many previously unknown protein-protein interactions across the yeast proteome, suggesting that the embeddings carried physically meaningful information about interaction interfaces.[3]
The practical headline metric, beyond raw accuracy, was speed. By computing on a single consumer GPU rather than a cluster, RoseTTAFold reduced the per-prediction time and cost by roughly an order of magnitude.[1][4] In contemporary coverage, Google DeepMind and the Baker group converged on a tacit division of labor: AlphaFold 2 served the role of an authoritative reference for the human proteome and other large databases, while RoseTTAFold provided a flexible, modifiable, and faster open source baseline that other research groups could fork.[4]
The picture changed somewhat with AlphaFold 2.3 and AlphaFold-Multimer, and again more dramatically with AlphaFold 3 in 2024, which generalized to nucleic acids and small molecules. RoseTTAFold All-Atom, discussed below, was the Baker Lab's response to that generalization. Independent benchmarks published in 2024 and 2025 generally place AlphaFold 3 above RoseTTAFold All-Atom on heterogeneous biomolecular targets, with RoseTTAFold 2 and ESMFold trailing on most subsets of the benchmarks, although the open-source status of RoseTTAFold makes it the natural starting point for groups that need to retrain or modify a structure predictor.[12]
Following the 2021 release, the Baker Lab and collaborators developed a sequence of derivative systems that share the three-track core but extend it to new modalities or new tasks.
RoseTTAFoldNA, posted to bioRxiv in September 2022 and published in Nature Methods in 2024, extended the three-track architecture to handle nucleic acids and protein-nucleic acid complexes.[13] The system represents DNA and RNA residues alongside amino acids in the 1D and 2D tracks, and updates atomic coordinates in the 3D track for both polymer types. Its authors reported that on protein-nucleic acid benchmarks, RoseTTAFoldNA's confident predictions were considerably more accurate than prior methods that handled nucleic acids only indirectly, such as docking on top of separately predicted protein structures.[13]
RoseTTAFold2, posted to bioRxiv on 24 May 2023 by Baek, Anishchenko, Humphreys, Cong, Baker, and DiMaio, was a substantial reengineering of the original system.[14] The team explicitly set out to combine features of RoseTTAFold and AlphaFold 2 into a single network that retained the three-track design but borrowed architectural choices from AlphaFold 2 where those had proved superior. The result was a network that closed much of the accuracy gap with AlphaFold 2 on monomer prediction and that scaled efficiently to large protein complexes. RoseTTAFold2 became the structural backbone for the all-atom and diffusion generalizations that followed.[14]
RoseTTAFold All-Atom, abbreviated RFAA, was published in Science on 19 April 2024 by Krishna and colleagues with David Baker as senior author (DOI 10.1126/science.adl2528).[5] RFAA combines a residue-based representation of amino acids and DNA or RNA bases with an explicit atom-and-bond graph representation of arbitrary small molecules, metals, and covalent modifications. The network can therefore predict the structure of an assembly that contains, for example, an enzyme, its peptide substrate, a metal ion cofactor, and a small molecule inhibitor in one forward pass, given their sequences and chemical structures.[5]
To accommodate small molecules, the authors fed chemical element identities to the 1D track, chemical bonds to the 2D track, and chirality and other 3D information to the 3D track of an RoseTTAFold-2-style network. They reported that RFAA can jointly model the relative positions of multiple ligands and a protein in a single inference step, and that it generalizes to ligand types not present in its training set.[5]
In the same paper Krishna et al. introduced RFdiffusionAA, a denoising diffusion variant of RFAA that designs protein backbones around target small molecules. They reported that proteins designed to bind the cardiac glycoside digoxigenin, the cofactor heme, and the light-harvesting pigment bilin folded as designed and showed the expected binding behavior, with crystal structures confirming the design models in several cases.[5]
RFdiffusion is a generative protein design model built by fine-tuning the RoseTTAFold structure prediction network on a denoising task. The work was first posted as a bioRxiv preprint titled "Broadly applicable and accurate protein design by integrating structure prediction networks and diffusion generative models" on 9 December 2022 by Joseph Watson, David Juergens, Nathaniel Bennett, Brian Trippe, Jason Yim, and coauthors.[15] A revised version was published in Nature in 2023 with the title "De novo design of protein structure and function with RFdiffusion".[16]
RFdiffusion adapts the diffusion model framework to backbone geometry. The forward (noising) process gradually corrupts an existing backbone toward an isotropic prior over backbone frames; the reverse (denoising) process is parameterized by RoseTTAFold and is trained to invert that corruption while satisfying user-specified constraints such as a target motif, a binding site, or a desired symmetry. Watson and coauthors reported strong performance across unconditional monomer design, topology-constrained design, protein binder design, symmetric oligomer design, enzyme active-site scaffolding, and metal-binding scaffolding, with several designs validated experimentally.[16]
A widely cited result from the Nature paper is the de novo design of a picomolar-affinity binder generated entirely by computation rather than by directed evolution, along with a series of symmetric assemblies confirmed by cryo-electron microscopy.[16] Equally striking is the reported efficiency of the design pipeline: where earlier de novo binder workflows often required testing tens of thousands of candidate sequences to find one with the desired affinity, RFdiffusion paired with sequence design tools such as ProteinMPNN reduced that ratio dramatically, in some cases to a single candidate per target.[16]
The RFdiffusion code was released as open source under a permissive BSD license in March 2023 from the RosettaCommons GitHub organization, with weights also distributed openly.[17] This permissive licensing, combined with documentation and example notebooks, made RFdiffusion a foundation for a generation of follow-on protein design work in academia and industry.
Other named extensions exist within the family. RoseTTAFold2NA combines the RoseTTAFold 2 backbone with nucleic acid handling, distributed as a separate GitHub repository at uw-ipd/RoseTTAFold2NA. Specialized variants have been released for antibody design, multi-state design, and sequence-space diffusion. LightRoseTTA, published by Wang and coauthors in 2025, is a distilled graph network designed to reproduce RoseTTAFold-class accuracy at much smaller parameter and inference cost.[18]
A consistent thread through the RoseTTAFold lineage is the deliberate choice to release code, weights, and a web server at the time of publication. This stands in some contrast to DeepMind's initial release pattern with AlphaFold 2, which delivered code and the AlphaFold Protein Structure Database shortly after the Nature paper but withheld it during the CASP14 cycle, and again with AlphaFold 3, whose code and weights were initially provided only through the Isomorphic AlphaFold Server with strong usage restrictions before later open-source releases.[4][12]
The open licensing of RoseTTAFold has had at least three measurable effects. First, it enabled a broad community of academic and industrial labs to fork and modify the network for specialized purposes such as antibody modeling, peptide design, and structure-based virtual screening, without negotiating commercial licenses. Second, it provided the substrate for the entire RFdiffusion family of protein design tools, which would not have been possible without an editable structure predictor. Third, it served as an open benchmark and reference implementation for neural network approaches to molecular structure, against which closed systems can be compared.
Among notable peers in this open ecosystem, OpenFold is an open-source reimplementation of AlphaFold 2 maintained by an academic and industry consortium, ESMFold (from Meta's Evolutionary Scale Modeling team) uses a Transformer protein language model to predict structures from a single sequence rather than from an MSA, and the Boltz family of models, released in 2024 and 2025, aimed at open reproductions of AlphaFold 3 capabilities. RoseTTAFold predates all of these and continues to be cited alongside them, both as a methodological inspiration and as a working tool.[12]
RoseTTAFold and its descendants have been applied across structural biology, biochemistry, and drug discovery. Common application areas include the following.
Structural genomics and target characterization. After release, the IPD used RoseTTAFold to compute structural models for hundreds of human proteins of unknown structure, many of them implicated in disease. These models were deposited in the public record and have been used to seed mutagenesis experiments and structure-based design.[2]
Modeling of protein-protein complexes. Because the 2D track is symmetric in its handling of pairs, RoseTTAFold can be run on concatenated sequences to predict heterodimer or homodimer structures. The Baek 2021 paper reported large-scale screening of candidate yeast protein-protein interactions using this approach.[3]
Protein design. RFdiffusion, RFdiffusionAA, and related fine-tuned variants form the design backbone for binders to therapeutic targets, novel enzymes, symmetric assemblies, and metal-coordinating proteins. Several reported designs have been validated by cryo-EM, X-ray crystallography, and biochemical assays.[16][5]
Nucleic acid and ligand-aware modeling. RoseTTAFoldNA, RoseTTAFold All-Atom, and their variants are used to model RNA folds, protein-DNA recognition, enzyme-substrate complexes, and protein-ligand co-folding in AI Drug Discovery workflows.[13][5]
Mutation effect prediction and protein engineering. Several derivative networks built on top of RoseTTAFold weights have been used to score mutations in candidate enzymes and to rank designed variants prior to experimental synthesis.[19]
Several limitations of RoseTTAFold and its lineage are widely acknowledged. The original 2021 network, while striking for its hardware footprint, was meaningfully less accurate than AlphaFold 2 on the most difficult CASP14 free modeling targets and on long multidomain assemblies.[3] Hardware constraints on training crops, and a smaller compute budget than DeepMind, account for some but not all of that gap.[3]
Second, like other structure prediction networks of its era, RoseTTAFold's accuracy degrades on proteins for which deep MSAs are unavailable, including many orphan sequences and engineered proteins. Methods that operate on a single sequence, including ESMFold, can outperform RoseTTAFold in that regime, while still falling short of the MSA-conditioned methods elsewhere.[12]
Third, the licensing terms for trained weights are not fully open in the same sense that the code is. The code base of RoseTTAFold uses the MIT license, but trained weights have historically been distributed under a separate license that permits non-commercial use only.[11] RFdiffusion's weights were released under BSD and are free for commercial use, but a researcher mixing weights across the family must inspect the individual licenses for each component.[17]
Fourth, the RFdiffusion family of generative models inherits the well-known risks of generative biological design. Although the Baker Lab and the broader biosecurity community have written extensively about responsible use, the same architectures that produce binders for therapeutic targets could in principle be turned to harmful ends, and the rapid proliferation of design tools has been a topic of policy discussion.[16]
Finally, RoseTTAFold 2 and All-Atom are reported in benchmarks to trail AlphaFold 3 and certain closed systems on heterogeneous biomolecular targets, although they typically lead on speed and on the breadth of available source code and weights.[12]
On 9 October 2024, the Royal Swedish Academy of Sciences announced that the 2024 Nobel Prize in Chemistry would be divided, with one half awarded to David Baker "for computational protein design" and the other half jointly to Demis Hassabis and John Jumper of Google DeepMind "for protein structure prediction". The total prize amount was 11 million Swedish kronor, with Baker receiving one half and Hassabis and Jumper sharing the other half between them.[6][20]
The Academy's scientific background explicitly recognized two intertwined achievements: the deep-learning revolution in protein structure prediction, exemplified by AlphaFold 2, and the design of entirely new proteins with desired functions, exemplified by the work of David Baker and his collaborators. Although RoseTTAFold was not the sole subject of the citation, it was part of the body of work that the Academy considered. More directly, the protein design framework RFdiffusion, which is built on RoseTTAFold, exemplifies the type of computational protein design for which Baker was honored.[21]
David Baker, born in Seattle in 1962, is a professor of biochemistry at the University of Washington, director of the Institute for Protein Design, and a Howard Hughes Medical Institute investigator.[20] In 2003 his group reported Top7, a 93-residue protein with a fold not observed in nature, designed using the Rosetta software (Kuhlman et al., Science, 21 November 2003). Top7 was a proof of principle that arbitrary protein folds could be designed from scratch with atomic-level accuracy and is widely treated as the starting point of de novo protein design.[22]
The other half of the 2024 chemistry prize went to Demis Hassabis and John Jumper for the development of AlphaFold 2. The combination of recipients in a single prize reflects an unusual situation in which two complementary lines of work, structure prediction and structure design, matured on roughly the same timescale and converged on shared neural network ideas. RoseTTAFold sits at exactly that intersection: a structure predictor used as a generative model of geometry, developed in the laboratory that the Nobel Committee honored for protein design.[6]
The protein modeling landscape in 2026 contains several systems closely related to RoseTTAFold, each with distinct lineage and licensing.
| System | Developer | First release | Modalities | Open source |
|---|---|---|---|---|
| AlphaFold (2) | DeepMind (Hassabis, Jumper) | 2020 to 2021 | Proteins (monomers and complexes) | Code and weights released |
| AlphaFold 3 | DeepMind and Isomorphic Labs | 2024 | Proteins, nucleic acids, ligands | Initial release server-only; code later released |
| RoseTTAFold (1) | Baker Lab / IPD | 2021 | Proteins | MIT code; non-commercial weights |
| RoseTTAFoldNA | Baker Lab | 2022 (bioRxiv) | Protein + nucleic acid | Open source |
| RoseTTAFold 2 | Baker Lab | 2023 (bioRxiv) | Proteins (improved) | Open source |
| RoseTTAFold All-Atom | Baker Lab | 2024 (Science) | Proteins, nucleic acids, small molecules, metals | Open source |
| RFdiffusion | Baker Lab | 2023 (Nature) | Protein backbone design | BSD, open |
| ESMFold | Meta FAIR ESM team | 2022 | Proteins (single sequence) | Open source |
| OpenFold | Academic consortium | 2022 | Proteins (AlphaFold 2 reimplementation) | Open source |
| Boltz | MIT and collaborators | 2024 | Open AlphaFold 3-class biomolecular modeling | Open source |
| ESM3 | EvolutionaryScale | 2024 | Multimodal protein modeling and design | Partial open source |
This table is intentionally selective and shows the systems most often discussed alongside RoseTTAFold in benchmarking and policy contexts. Among the open-source systems for biomolecular structure prediction, RoseTTAFold occupies a special position as the oldest established lineage and as the one most tightly integrated with a major protein design program.
In retrospect, RoseTTAFold's significance lies less in any single accuracy number than in three structural facts about its release and design.
First, RoseTTAFold demonstrated that a small academic group with consumer hardware could reproduce most of the capability of a much larger industrial system within a few months, given a clear architectural goal. This shaped expectations in machine learning for biology more broadly, and contributed to the open-source acceleration of structure prediction across 2022 to 2025.
Second, it provided a usable, editable, GPU-friendly baseline for the field at exactly the moment when AI drug discovery pipelines and academic protein design groups needed one. Open licensing and a public web server amplified that effect.[2][4]
Third, by exposing the 3D coordinate track as a first-class object inside the network, RoseTTAFold made it natural to think of structure predictors as priors over geometry that could be inverted for design. The leap to RFdiffusion was conceptually short once that framing was in place, and it produced one of the most impactful generative model deployments outside of generative model research in vision and language.[16][5]