Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 85 of the 85 references it cites were found in a registry or an index; 13 of the 36 sources its claims cite were checked for support.

Accurate prediction of protein structures and interactions using a 3-track network

Open access

Abstract

DeepMind presented remarkably accurate protein structure predictions at the CASP14 conference. We explored network architectures incorporating related ideas and obtained the best performance with a 3-track network in which information at the 1D sequence level, the 2D distance map level, and the 3D coordinate level is successively transformed and integrated. The 3-track network produces structure predictions with accuracies approaching those of DeepMind in CASP14, enables rapid solution of challenging X-ray crystallography and cryo-EM structure modeling problems, and provides insights into the functions of proteins of currently unknown structure. The network also enables rapid generation of accurate models of protein-protein complexes from sequence information alone, short circuiting traditional approaches which require modeling of individual subunits followed by docking. We make the method available to the scientific community to speed biological research. One-Sentence Summary: Accurate protein structure modeling enables rapid solution of structure determination problems and provides insights into biological function.

Overview

CoChat Summary

TL;DR

RoseTTAFold is a 3-track neural network that integrates 1D sequence, 2D distance-map, and 3D coordinate information to predict protein structures and interactions. It approaches AlphaFold2 accuracy in CASP14, solves challenging crystallography and cryo-EM modeling problems, and generates protein-complex models directly from sequence.

Key findings

  • In CASP14, the 3-track RoseTTAFold outperformed the top 2 server groups (Zhang-server and BAKER-ROSETTASERVER), the BAKER human group, and the authors’ 2-track models, though its performance was still not as good as AlphaFold2.
  • On CAMEO, over 69 medium and hard targets released May 15–June 19, 2021, RoseTTAFold outperformed all other evaluated servers, including Robetta, IntFold6-TS, BestSingleTemplate, and SWISS-MODEL.
  • RoseTTAFold models enabled molecular-replacement solutions for four previously unsolved crystallographic datasets; for the cryo-EM p101 Gβγ binding domain, the Cα-RMSD between predicted and refined structure was 3.0 Å over the beta-sheets.
  • Among 693 domains from human disease-related proteins, over one-third had predicted lDDT > 0.8, corresponding to an average Cα-RMSD of 2.6 Å on CASP14 targets; many protein-complex models had TM-score > 0.8, with complex generation taking ~30 min on a 24G TITAN RTX GPU.

Methods

  • The 3-track network transforms and integrates information at the 1D sequence/MSA level, the 2D distance/orientation-map level, and the 3D backbone-coordinate level, with information flowing between tracks; training used discontinuous sequence crops spanning 260 residues.
  • Final 3D structures were generated either by feeding predicted distance/orientation distributions into pyRosetta for all-atom models or by an end-to-end SE(3)-equivariant layer producing backbone coordinates directly.
  • Evaluation included CASP14 targets, CAMEO blind server assessment, four X-ray molecular-replacement datasets, a cryo-EM PI3Kγ complex, GPCRs, 693 human disease-related domains, and paired-sequence complexes of known structure, using accuracy comparisons, Cα-RMSD, lDDT, TM-score, and MR/cryo-EM fit.

Limitations

  • Stated: hardware memory limitations constrained model size and training; performance remained below AlphaFold2; the end-to-end version was less accurate than the pyRosetta version due to incomplete optimization and neglect of side-chain information; the accuracy predictor did not consistently identify models better than the rapid single-pass method; the network was trained on monomeric proteins, not complexes; and current RoseTTAFold uses only the top 1000 sequences due to memory limits.
  • Stated: the Perceiver-based extension for using all available sequences gave promising initial results, but more training is required for rigorous comparison.
  • Evident: CASP14 and CAMEO evaluations cover limited target sets and time windows, and the end-to-end model does not produce side chains.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1101/2021.06.14.448402

Cites

85 of 85 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Network architecture and performance. (A) RoseTTAFold architecture with 1D, 2D, and 3D attention tracks. Multiple connections between tracks allow the network to simultaneously learn relationships within and between sequences, distances, and coordinates (see Methods and fig. S1 for details). (B) Average TM-score of prediction methods on the CASP14 targets. Zhang-server and BAKER-ROSETTASERVER were the top 2 server groups while AlphaFold2 and BAKER were the top 2 human groups in CASP14; BAKER-ROSETTASERVER and BAKER predictions were based on trRosetta. Predictions with the 2-track model and RoseTTAFold (both end-to-end and pyRosetta version) were completely automated. (C) Blind benchmark results on CAMEO medium and hard targets; model accuracies are TM-score values from the CAMEO website (https://cameo3d.org/).
Fig. 1 Network architecture and performance. (A) RoseTTAFold architecture with 1D, 2D, and 3D attention tracks. Multiple connections between tracks allow the network to simultaneously learn relationships within and between sequences, distances, and coordinates (see Methods and fig. S1 for details). (B) Average TM-score of prediction methods on the CASP14 targets. Zhang-server and BAKER-ROSETTASERVER were the top 2 server groups while AlphaFold2 and BAKER were the top 2 human groups in CASP14; BAKER-ROSETTASERVER and BAKER predictions were based on trRosetta. Predictions with the 2-track model and RoseTTAFold (both end-to-end and pyRosetta version) were completely automated. (C) Blind benchmark results on CAMEO medium and hard targets; model accuracies are TM-score values from the CAMEO website (https://cameo3d.org/).CC BY · PubMed Central
Fig. 2 Enabling experimental structure determination with RoseTTAFold. (A-B) Successful molecular replacement with RoseTTAFold models. (A) SLP. (top) C-terminal domain: comparison of final refined structure (gray) to RoseTTAFold model (blue); there are no homologs with known structure. (bottom) N-terminal domain: refined structure is in gray, and RoseTTAFold model is colored by the estimated RMS error (ranging from blue for 0.67 Å to red for 2 Å or greater). 95 Cα atoms of the RoseTTAFold model can be superimposed within 3 Å of Cα atoms in the final structure, yielding a Cα-RMSD of 0.98 Å. In contrast, only 54 Cα atoms of the closest template (4l3a, brown) can be superimposed (with a Cα-RMSD of 1.69 Å). (B) Refined structure of Lrbp (gray) with the closest RoseTTAFold model (blue) superimposed; residues having estimated RMS error greater than 1.3 Å are omitted (full model is in fig. S5C). (C) Cryo-EM structure determination of p101 Gβγ binding domain (GBD) in a heterodimeric PI3Kγ complex using RoseTTAFold. (top) RoseTTAFold models colored in a rainbow from the N-terminus (blue) to the C-terminus (red) have a consistent all-beta topology with a clear correspondence to the density map. (bottom) Comparison of the final refined structure to the RoseTTAFold model colored by predicted RMS error ranging from blue for 1.5 Å or less to red 3 Å or greater. The actual Cα-RMSD between the predicted structure and final refined structure is 3.0 Å over the beta-sheets. Figure prepared with ChimeraX (35).
Fig. 2 Enabling experimental structure determination with RoseTTAFold. (A-B) Successful molecular replacement with RoseTTAFold models. (A) SLP. (top) C-terminal domain: comparison of final refined structure (gray) to RoseTTAFold model (blue); there are no homologs with known structure. (bottom) N-terminal domain: refined structure is in gray, and RoseTTAFold model is colored by the estimated RMS error (ranging from blue for 0.67 Å to red for 2 Å or greater). 95 Cα atoms of the RoseTTAFold model can be superimposed within 3 Å of Cα atoms in the final structure, yielding a Cα-RMSD of 0.98 Å. In contrast, only 54 Cα atoms of the closest template (4l3a, brown) can be superimposed (with a Cα-RMSD of 1.69 Å). (B) Refined structure of Lrbp (gray) with the closest RoseTTAFold model (blue) superimposed; residues having estimated RMS error greater than 1.3 Å are omitted (full model is in fig. S5C). (C) Cryo-EM structure determination of p101 Gβγ binding domain (GBD) in a heterodimeric PI3Kγ complex using RoseTTAFold. (top) RoseTTAFold models colored in a rainbow from the N-terminus (blue) to the C-terminus (red) have a consistent all-beta topology with a clear correspondence to the density map. (bottom) Comparison of the final refined structure to the RoseTTAFold model colored by predicted RMS error ranging from blue for 1.5 Å or less to red 3 Å or greater. The actual Cα-RMSD between the predicted structure and final refined structure is 3.0 Å over the beta-sheets. Figure prepared with ChimeraX (35).CC BY · PubMed Central
Fig. 3 RoseTTAFold models provide insights into function. (A) TANGO2 model, colored in a rainbow from the N-terminus (blue) to the C-terminus (red), adopts an Ntn hydrolase fold. Pathogenic mutation sites are in magenta spheres. (B) Predicted TANGO2 active site colored by ortholog conservation in rainbow scale from variable (blue) to conserved (red) with conserved residues in stick and labeled. Pathogenic mutations (spheres with wild-type side chains in the sticks) are labeled in magenta; select neighboring residues are depicted in the sticks. (C) ADAM33 prodomain adopts a lipocalin-like barrel shown in a rainbow from N-terminus (blue) to C-terminus (red). (D) ADAM33 model surface rendering colored by ortholog conservation from blue (variable) to red (conserved), highlighting a conserved surface patch. (E) CERS1 transmembrane structure prediction is colored from N-terminus (blue) to C-terminus (red), with a pathogenic mutation in TMH2 near a central cavity in magenta. (F) Zoom of CERS1 active site with residues colored by ortholog conservation from variable (blue) to conserved (red). Residues that contribute to catalysis (H182 and D213) or are conserved (W298 and D213) line the cavity. The conserved pathogenic mutation is adjacent to the active site.
Fig. 3 RoseTTAFold models provide insights into function. (A) TANGO2 model, colored in a rainbow from the N-terminus (blue) to the C-terminus (red), adopts an Ntn hydrolase fold. Pathogenic mutation sites are in magenta spheres. (B) Predicted TANGO2 active site colored by ortholog conservation in rainbow scale from variable (blue) to conserved (red) with conserved residues in stick and labeled. Pathogenic mutations (spheres with wild-type side chains in the sticks) are labeled in magenta; select neighboring residues are depicted in the sticks. (C) ADAM33 prodomain adopts a lipocalin-like barrel shown in a rainbow from N-terminus (blue) to C-terminus (red). (D) ADAM33 model surface rendering colored by ortholog conservation from blue (variable) to red (conserved), highlighting a conserved surface patch. (E) CERS1 transmembrane structure prediction is colored from N-terminus (blue) to C-terminus (red), with a pathogenic mutation in TMH2 near a central cavity in magenta. (F) Zoom of CERS1 active site with residues colored by ortholog conservation from variable (blue) to conserved (red). Residues that contribute to catalysis (H182 and D213) or are conserved (W298 and D213) line the cavity. The conserved pathogenic mutation is adjacent to the active site.CC BY · PubMed Central
Fig. 4 Complex structure prediction using RoseTTAFold. (A, B) Prediction of structures of E.coli protein complexes from sequence information. Experimentally determined structures are on the left, RoseTTAFold models, on the right; the TMscores below indicate the extent of structural similarity. (A) Two chain complexes. The first subunit is colored in gray, and the second subunit is colored in a rainbow from blue (N-terminal) to red (C-terminal). (B) Three chain complexes. Subunits are colored in gray, cyan, and magenta. (C) IL-12R/IL-12 complex structure generated by RoseTTAFold fits the previously published cryo-EM density (EMD-21645).
Fig. 4 Complex structure prediction using RoseTTAFold. (A, B) Prediction of structures of E.coli protein complexes from sequence information. Experimentally determined structures are on the left, RoseTTAFold models, on the right; the TMscores below indicate the extent of structural similarity. (A) Two chain complexes. The first subunit is colored in gray, and the second subunit is colored in a rainbow from blue (N-terminal) to red (C-terminal). (B) Three chain complexes. Subunits are colored in gray, cyan, and magenta. (C) IL-12R/IL-12 complex structure generated by RoseTTAFold fits the previously published cryo-EM density (EMD-21645).CC BY · PubMed Central

Verification

Claims the paper makes

3 Valid · 8 Not valid · 14 Not checked · 44 Own results and interpretations

33 more claims

Create a free account to see all 69 claims and to verify the 2 references still open.

See all claims — free account
  • It may state two separate claims as one. 24
  • Probably not a substantial claim. 13
  • The quote only partly states this claim. 2

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

17 more claims about this paper

Create a free account to see every claim other papers make about this one.

See all claims — free account

Citations

Cites

Show all 85 references

Research works printed but not linked: 2

  1. 2011, phenix. ensembler: a tool for multiple superposition, Comput Crystallogr Newsl, 2, 8
  2. 2015, in ACTA CRYSTALLOGRAPHICA A-FOUNDATION AND ADVANCES (INT UNION CRYSTALLOGRAPHY 2 ABBEY SQ, CHESTER, CH1 2HU, ENGLAND, 71, S19

Cited by

See all 5,595

Metadata

Authors
Baek, M., DiMaio, F., Anishchenko, I., Dauparas, J., Ovchinnikov, S., Lee, G. R., Wang, J., Cong, Q., Kinch, L. N., Schaeffer, R. D., Millán, C., Park, H., Adams, C., Glassman, C. R., DeGiovanni, A., Pereira, J. H., Rodrigues, A. V., van Dijk, A. A., Ebrecht, A. C., Opperman, D. J., Sagmeister, T., Buhlheller, C., Pavkov-Keller, T., Rathinaswamy, M. K., Dalwadi, U., Yip, C. K., Burke, J. E., Garcia, K. C., Grishin, N. V., Adams, P. D., Read, R. J., & Baker, D.
Affiliations
  • University of Washington (US)
  • Harvard University (US)
  • The University of Texas Southwestern Medical Center (US)
  • Howard Hughes Medical Institute (US)
  • University of Cambridge (GB)
  • Stanford Medicine (US)
  • Lawrence Berkeley National Laboratory (US)
  • North-West University (ZA)
  • University of the Free State (ZA)
  • University of Graz (AT)
  • University of Victoria (CA)
  • University of British Columbia (CA)
Corresponding author
Baker, David
Venue
Science
Publisher
openRxiv
Published
2021-06-15
Volume
373
Issue
6557
Pages
871-876
ISSN
0036-8075 (ISSN-L)
OpenAlex
W3186179742
Cited by
5,549
References
85
Citations by year
2026: 912 · 2025: 1108 · 2024: 1165 · 2023: 1144 · 2022: 982 · 2021: 227
FWCI
247.22
Type
journal article
Language
en
Open access
Licence
cc-by
Locations
  • Science (publishedVersion, cc-by)
  • bioRxiv (Cold Spring Harbor Laboratory) (submittedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • PubMed Central (acceptedVersion, cc-by)
  • eScholarship (California Digital Library) (submittedVersion, cc-by)
  • OSTI OAI (U.S. Department of Energy Office of Scientific and Technical Information) (submittedVersion, cc-by)
  • DESY Publication Database (PUBDB) (Deutsches Elektronen-Synchrotron) (submittedVersion)
  • Apollo (University of Cambridge) (submittedVersion, cc-by)
  • Apollo (cc-by)
Funders
  • National Science Foundation (GM127390, 1937533)
  • Welch Foundation (I‐1505)
  • Washington Research Foundation
  • Microsoft
  • TU Graz, Internationale Beziehungen und Mobilitätsprogramme
  • Wellcome Trust (Z/17/Z, 209407/Z/17/Z, 209407)
  • G. Harold and Leila Y. Mathers Foundation
  • Diamond Light Source
  • Open Philanthropy Project
  • Global Challenges Research Fund (ST/R002754/1)
  • Schmidt Family Foundation
  • European Synchrotron Radiation Facility
  • Austrian Science Fund (P 29432, DOC 50)
  • National Institutes of Health (DP5OD026389, RO1-AI51321)
  • Canadian Institutes of Health Research (168998, 168907)
  • Science and Technology Facilities Council (ST/R002754/1)
  • Deutsches Elektronen-Synchrotron
  • National Institute of General Medical Sciences (P01GM063210, R35GM127390, R01GM123089)
  • National Institute of Allergy and Infectious Diseases (R01AI051321)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Protein Structure and Dynamics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Machine Learning in Bioinformatics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Structural Biology › Enzyme Structure and Function
Keywords
protein structure prediction, protein-protein complex prediction, distance maps, sequence-based structure prediction, CASP14, protein function prediction
MeSH terms
Deep Learning, Amino Acid Sequence, Computer Simulation, Membrane Proteins, Models, Molecular, Protein Conformation, Proteins, Neural Networks, Computer, Protein Folding, Crystallography, X-Ray, Cryoelectron Microscopy, Protein Subunits, Databases, Protein, Receptors, G-Protein-Coupled, Multiprotein Complexes, Sphingosine N-Acyltransferase, ADAM Proteins

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Accurate prediction of protein structures and interactions using a 3-track network. https://app.cochat.ai/papers/10.1101/2021.06.14.448402
The paper
Baek, M., DiMaio, F., Anishchenko, I., Dauparas, J., Ovchinnikov, S., Lee, G. R., Wang, J., Cong, Q., Kinch, L. N., Schaeffer, R. D., Millán, C., Park, H., Adams, C., Glassman, C. R., DeGiovanni, A., Pereira, J. H., Rodrigues, A. V., van Dijk, A. A., Ebrecht, A. C., . . . Baker, D. (2021). Accurate prediction of protein structures and interactions using a 3-track network. Science. https://doi.org/10.1101/2021.06.14.448402

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.