CoChat
Sign Up
Cited by 2,090Protein Structure and DynamicsMachine Learning in BioinformaticsComputational Drug Discovery Methods

Robust deep learning–based protein sequence design using ProteinMPNN

Open access
Dauparas, J., Anishchenko, I., Bennett, N., Bai, H., Ragotte, R. J., Milles, L. F., Wicky, B. I. M., Courbet, A. +14 more·2022·Science
DOI 10.1126/science.add2187PMC PMC9997061

Published version of a preprint

Open this paper in CoChat — free accountOpen PDF

Abstract

Although deep learning has revolutionized protein structure prediction, almost all experimentally characterized de novo protein designs have been generated using physically based approaches such as Rosetta. Here, we describe a deep learning–based protein sequence design method, ProteinMPNN, that has outstanding performance in both in silico and experimental tests. On native protein backbones, ProteinMPNN has a sequence recovery of 52.4% compared with 32.9% for Rosetta. The amino acid sequence at different positions can be coupled between single or multiple chains, enabling application to a wide range of current protein design challenges. We demonstrate the broad utility and high accuracy of ProteinMPNN using x-ray crystallography, cryo–electron microscopy, and functional studies by rescuing previously failed designs, which were made using Rosetta or AlphaFold, of protein monomers, cyclic homo-oligomers, tetrahedral nanoparticles, and target-binding proteins.

AI snapshot

Written from the paper’s full text.

TL;DR

ProteinMPNN is a deep learning–based protein sequence design method that uses a message-passing neural network with order-agnostic decoding and symmetry-aware coupling to design sequences for monomers, oligomers, nanoparticles, and interfaces. It achieves higher native sequence recovery than Rosetta and rescues previously failed experimental designs across multiple protein design challenges.

Key findings

  • On native protein backbones, ProteinMPNN achieved 52.4% sequence recovery versus 32.9% for Rosetta, while requiring only 1.2 seconds versus 4.3 minutes for a 100-residue protein.
  • Median sequence recovery was 52% for monomers, 55% for homomers, and 51% for heteromers; recovery ranged from 90–95% in the deep core to 35% on the surface.
  • Training with backbone noise improved robustness: a model trained with 0.3 Å noise generated 2–3 times more sequences with AlphaFold predictions within lDDT-Ca of 95.0 and 90.0 of the true structures than unnoised or slightly noised models.
  • Experimental rescue of AlphaFold-generated designs: of 96 designs produced in E. coli, 73 were soluble (median soluble yield 247 mg/L vs. 9 mg/L for AF sequences) and 50 had the target monomeric or oligomeric state; for C5/C6 cyclic oligomers, ProteinMPNN designs were 88% soluble with 27.7% correct oligomeric state versus 40% soluble and none correct for Rosetta designs.

Methods

  • ProteinMPNN is based on a message-passing neural network with 3 encoder and 3 decoder layers and 128 hidden dimensions, using backbone features including Ca-Ca distances, relative frame orientations, dihedral angles, and added N, Ca, C, O, and virtual Cb distances; edge updates and order-agnostic autoregressive decoding were introduced.
  • Training used protein assemblies from the PDB (as of Aug 02, 2021) determined by X-ray crystallography or cryoEM to better than 3.5 Å resolution and with fewer than 10,000 residues, clustered at 30% sequence identity with mmseqs2, yielding 25,361 clusters; monomer training/validation used 19.7k high-resolution single-chain structures split by CATH.
  • Evaluation included in silico test sets of 402 monomer backbones for Rosetta comparison and 690 monomers, 732 homomers (<2000 residues), and 98 heteromers, plus experimental characterization by E. coli expression, SEC, SEC-MALS, X-ray crystallography, cryoEM, and biolayer interferometry.

Limitations

  • The paper states that in silico metrics such as native sequence recovery are sensitive to crystallographic resolution and may not correlate with proper folding; even a single residue substitution can block folding.
  • The paper notes that deep learning methods lack the physical transparency of methods like Rosetta.
  • The text does not include a dedicated limitations section; evident from the text, experimental validation covers selected design challenges and backbones, and not all designed sequences were structurally characterized.

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

Integrity signals

Registry

Crossref · 10.1126/science.add2187

Cites

38 of 38 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

1 of 8 cited sources checked · 0 supported · 0 flagged · accuracy not yet measured

  • Findingown data

    AF generated sequences were mostly insoluble (median soluble yield: 9 mg per liter of culture equivalent).

    “we found that the AF generated sequences were mostly insoluble (median soluble yield: 9 mg per liter of culture equivalent Figure 3A).”
    Quote in the paperSubstantial claim
  • Findingown data

    Adding distances between N, Ca, C, O and a virtual Cb as additional input features increased sequence recovery from 41.2% (baseline model) to 49.0% (experiment 1).

    “This resulted in a sequence recovery increase from 41.2% (baseline model) to 49.0% (experiment 1), see Table 1 below”
    Quote in the paperSubstantial claim
  • Findingown data

    Combining additional input features and edge updates leads to a sequence recovery of 50.5% (experiment 3).

    “Combining additional input features and edge updates leads to a sequence recovery of 50.5% (experiment 3).”
    Quote in the paperSubstantial claim
  • Findingown data

    ProteinMPNN requires only a small fraction of the compute time of Rosetta (1.2 seconds versus 4.3 minutes for 100 residues).

    “Although requiring only a small fraction of the compute time (1.2 seconds versus 4.3 minutes for 100 residues)”
    • Also stated in: § Conclusion · plain
      “ProteinMPNN solves sequence design problems in a small fraction of the time (1.2 sec vs 258.8 sec on a single CPU for a 100 residue protein) required for physically based approaches such as Rosetta”
    Quote in the paperSubstantial claim
  • Findingown data

    Performance saturated at 32-48 nearest Ca neighbor neural networks.

    “found that performance saturated at 32–48 neighbors.”
    Quote in the paperSubstantial claim
  • May combine two claims. 19
  • Classified as probably not a substantial claim. 3

  • May combine two claims.

    Premisecitation

    Machine learning sequence design approaches have been developed previously, but have focused on the monomer design problem, achieve lower native sequence recoveries, and with the exception of a TIM barrel design study have not been extensively validated using crystallography and cryoEM.

    “Machine learning sequence design approaches have been developed previously (1–6, 6a), notably the previously described message passing method on which ProteinMPNN is based, but have focused on the monomer design problem, achieve lower native sequence recoveries, and with the exception of a TIM barrel design study (6) have not been extensively validated using crystallography and cryoEM to evaluate”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites 1–6, 6a

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [6] Anand N, Eguchi R, Mathews II, Perez CP, Derry A, Altman RB, & Huang PS (2022). Protein sequence design with a learned potential. Nature communications, 13(1), 1– 11. PMC8826426 35136054 source

      Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

      Experimental — accuracy not yet measured

  • May combine two claims.

    Findingown data

    On de novo designed scaffolds with ligand binding pockets, only a small fraction of original Rosetta designed sequences were predicted to fold to the design target structures, whereas following ProteinMPNN redesign the majority were confidently predicted to fold to close to the design target structures.

    “Whereas only a small fraction of the original Rosetta designed sequences were predicted to fold to the design target structures, following ProteinMPNN redesign the majority were confidently predicted to fold to close to the design target structures (Figure 2F).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    For homomers, averaging logits between symmetry related positions gave the best results: unconstrained design without symmetry, averaged probabilities, and averaged logits resulted in 52%, 53%, and 55% median sequence recoveries respectively.

    “unconstrained design without symmetry, averaged probabilities, and averaged logits resulted in 52%, 53%, and 55% median sequence recoveries respectively (Figure S1C).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    On a test set of 690 monomers, 732 homomers, and 98 heteromers, median overall sequence recoveries were 52% for monomers, 55% for homomers, and 51% for heteromers.

    “The median overall sequences recoveries were 52% for monomers, 55% for homomers, and 51% for heteromers (Figure 2B).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Many of the proteins were highly thermo-stable, with secondary structure being maintained up to 95 °C.

    “Many of the proteins were highly thermo-stable, with secondary structure being maintained up to 95 °C (Figure 3B).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The backbones of previously failed designs are correct and designable targets, so failure was due to the sequence.

  • May combine two claims.

    Findingown data

    The alpha-beta protein structure contains 5 beta strands and 4 alpha helices, and is close to the design target backbone (2.35 Å over 130 residues).

    “The alpha-beta protein structure contains 5 beta strands and 4 alpha helices, and is close to the design target backbone (2.35 Å over 130 residues)”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The backbones of previously failed designs are correct and designable targets, so failure was due to the sequence.

  • May combine two claims.

    Findingown data

    Of 96 ProteinMPNN designs produced in E. coli, 73 were expressed solubly (median soluble yield: 247 mg per liter of culture equivalent) and 50 had the target monomeric or oligomeric state as assessed by SEC.

    “of 96 designs produced in E. coli, 73 were expressed solubly (median soluble yield: 247 mg per liter of culture equivalent, Figure 3A) and 50 had the target monomeric or oligomeric state as assessed by SEC (Figure 3A,C).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The backbones of previously failed designs are correct and designable targets, so failure was due to the sequence.

  • May combine two claims.

    Findingown data

    Training models on backbones with Gaussian noise (std=0.02Å) improved sequence recovery on confident AlphaFold models (average pLDDT>80.0) from UniRef50, while sequence recovery on unperturbed PDB structures decreased.

    “training models on backbones to which Gaussian noise (std=0.02Å) had been added improved sequence recovery on confident protein structure models generated by AlphaFold (average pLDDT>80.0) from UniRef50, while the sequence recovery on unperturbed PDB structures decreased as expected (Table 1).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Biolayer interferometry showed strong binding to Grb2 SH3 domain for ProteinMPNN sequences, with considerably higher signal than free proline rich peptide; point mutations predicted to disrupt design completely eliminated binding signal.

    “Biolayer interferometry experiments showed strong binding to the Grb2 SH3 domain (Figure 4B), with considerably higher signal than the free proline rich peptide; point mutations predicted to disrupt the design completely eliminated the binding signal.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The backbones of previously failed designs are correct and designable targets, so failure was due to the sequence.

  • May combine two claims.

    Findingcitation

    ProteinMPNN designs have a much higher propensity to crystallize, greatly facilitating structure determination of designed proteins.

    “ProteinMPNN designs also have a much higher propensity to crystallize, greatly facilitating structure determination of designed proteins.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Wicky et al.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Findingown data

    For C5/C6 cyclic oligomers, the Rosetta designed set had 40% soluble and none with correct oligomeric state confirmed by SEC-MALS; the ProteinMPNN designed set had 88% soluble and 27.7% with correct oligomeric state.

    “For the Rosetta designed set, 40% were soluble and none had the correct oligomeric state confirmed by SEC-MALS. For the ProteinMPNN designed set, 88% were soluble and 27.7% had the correct oligomeric state.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The backbones of previously failed designs are correct and designable targets, so failure was due to the sequence.

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationreasoning

    Deep learning methods are trained directly to find the most probable amino acid for a protein backbone given all the examples in the PDB, and hence such ambiguities do not arise, making sequence design more robust and less dependent on the judgement of a human expert.

    “they are trained directly to find the most probable amino acid for a protein backbone given all the examples in the PDB, and hence such ambiguities do not arise, making sequence design more robust and less dependent on the judgement of a human expert.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    In traditional physically based approaches, sequence design maps to the problem of identifying an amino acid sequence whose lowest energy state is the desired structure.

    “In traditional physically based approaches, sequence design maps to the problem of identifying an amino acid sequence whose lowest energy state is the desired structure.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisenone given

    The methods described thus far are limited in their applicability to the wide range of current protein design challenges, and have not been extensively validated experimentally.

    “However, the methods described thus far are limited in their applicability to the wide range of current protein design challenges, and have not been extensively validated experimentally.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites 1–6, 6a

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Classified as probably not a substantial claim.

    Premisecitation

    Deep learning approaches have shown considerable promise in rapidly generating plausible amino acid sequences given monomeric protein backbones without need for compute intensive explicit consideration of sidechain rotameric states.

    “Recently deep learning approaches have shown considerable promise in rapidly generating plausible amino acid sequences given monomeric protein backbones without need for compute intensive explicit consideration of sidechain rotameric states (1–6, 6a).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites 1–6, 6a

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Premisecitation

    AlphaFold and RoseTTAfold produce remarkably good structure predictions for native proteins given multiple sequence alignments, but generally produce much poorer structures when provided only with a single sequence.

    “AlphaFold (9) and RoseTTAfold (13) produce remarkably good structure predictions for native proteins given multiple sequence alignments which can contain substantial co-evolutionary and other information reflecting aspects of the 3D structure, but generally produce much poorer structures when provided only with a single sequence.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [9] Uria B, Murray I, & Larochelle H (2014, January). A deep and tractable density estimator. In International Conference on Machine Learning (pp. 467– 475). PMLR. source
    • cites [13] Leman JK, Weitzner BD, Lewis SM, Adolf-Bryfogle J, Alam N, Alford RF, … & Bonneau R (2020). Macromolecular modeling and design in Rosetta: recent methods and frameworks. Nature methods, 17(7), 665– 680. PMC7603796 32483333 source

  • May combine two claims.

    Interpretationown data

    Locally connected graph neural networks can be used to model the structure to sequence mapping problem because protein backbones provide a notion of local neighborhoods which primarily determine sequence identities.

    “locally connected graph neural networks can be used to model the structure to sequence mapping problem because protein backbones provide a notion of local neighborhoods which primarily determine sequence identities.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisereasoning

    This is computationally intractable as it requires computing energies over all possible structures, including unwanted oligomeric and aggregated states.

    “This is, however, computationally intractable as it requires computing energies over all possible structures, including unwanted oligomeric and aggregated states”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    ProteinMPNN requires no expert customization for specific design challenges, and it should thus make protein design more broadly accessible.

    “ProteinMPNN requires no expert customization for specific design challenges, and it should thus make protein design more broadly accessible.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    In silico metrics such as native sequence recovery are very sensitive to crystallographic resolution and may not correlate with proper folding.

    “In silico metrics such as native sequence recovery are very sensitive to crystallographic resolution (Figure S3 B, C) and may not correlate with proper folding”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisereasoning

    Rosetta and other approaches as a proxy carry out a search for the lowest energy sequence for a given backbone structure, and structure prediction calculations are required in a second step to confirm that there are no other structures in which the sequence has still lower energy.

    “instead Rosetta and other approaches as a proxy carry out a search for the lowest energy sequence for a given backbone structure, and structure prediction calculations are required in a second step to confirm that there are no other structures in which the sequence has still lower energy.”
    Quote in the paperSubstantial claim
21 of 21

  • The paper relies on this without support.

    Evolution in most cases does not optimize for stability.

    “as evolution in most cases does not optimize for stability”

  • The paper relies on this without support.

    The backbones of previously failed designs are correct and designable targets, so failure was due to the sequence.

    “we kept the backbones of the original designs fixed but discarded the original sequences and generated new ones using ProteinMPNN.”
2 of 2

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?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • CATH citation points to the wrong reference

    Minor

    The text cites reference 7 for CATH, but the CATH paper is listed as reference 8. Reference 7 is a different paper.

1 of 1

Cites

Flagged first, then by citations

Cited by

Sorted by citations

Related

Authors
Dauparas, J., Anishchenko, I., Bennett, N., Bai, H., Ragotte, R. J., Milles, L. F., Wicky, B. I. M., Courbet, A., de Haas, R. J., Bethel, N., Leung, P. J. Y., Huddy, T. F., Pellock, S., Tischer, D., Chan, F., Koepnick, B., Nguyen, H., Kang, A., Sankaran, B., Bera, A. K., King, N. P., & Baker, D.
Affiliations
  • University of Washington (US)
  • Howard Hughes Medical Institute (US)
  • Department of Physical Chemistry and Soft Matter, Wageningen University and Research, Wageningen, Netherlands.
  • Lawrence Berkeley National Laboratory (US)
Corresponding author
Baker, D.
ORCID
Venue
Science
Publisher
American Association for the Advancement of Science (AAAS)
Published
2022-10-07
Volume
378
Issue
6615
Pages
49-56
ISSN
  • 0036-8075 (ISSN-L)
  • 1095-9203
DOI
10.1126/science.add2187
PMC
PMC9997061
OpenAlex
W4296032638
Semantic Scholar
98926d43356e87c22c82efc132dcaaac1ff40ebe
Cited by
2,070
References
35
Citations by year
2026: 732 · 2025: 669 · 2024: 419 · 2023: 216 · 2022: 33
FWCI
92.42
Type
journal article
Language
en
Open access
Licence
cc-by-nc-nd
Locations
  • Science (publishedVersion)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • Socio-Environmental Systems Modeling (acceptedVersion)
  • OSTI OAI (U.S. Department of Energy Office of Scientific and Technical Information) (submittedVersion)
  • PubMed Central (submittedVersion, cc-by)
  • Data Archiving and Networked Services (DANS) (submittedVersion)
Funders
  • National Science Foundation (DEAC02-05CH11231, DGE-214-0004, 1937533, P30 GM124169, 2140004)
  • Howard Hughes Medical Institute (P30 GM124169-01, DE-AC02-05CH11231, GT11817)
  • U.S. Department of Energy (-AC02-05CH11231, 05CH11231, DE AC02 05CH11231, GM124169, DE-AC02-)
  • Alfred P. Sloan Foundation (g-2021-16899, DE-AC02-05CH11231)
  • Washington Research Foundation
  • Open Philanthropy Project
  • National Institutes of Health (DE-AC02–05CH11231, 5p30gm124169-08, P30 GM124169, P30-GM124169-01, GM124169)
  • Office of Science (DEAC02–05CH11231, AC02-05CH11231, DE-AC02)
  • National Institute of General Medical Sciences (P30 GM124169., P30 GM124169‐01, DE-AC02-05CH11231, GM124169-01, GM124169)
  • Basic Energy Sciences (DE-AC02, DE-AC02-05CH11231., AC02–05CH11231)
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
  • Physical Sciences › Computer Science › Computational Theory and Mathematics › Computational Drug Discovery Methods
Keywords
protein sequence design, de novo protein design, ProteinMPNN, deep learning, protein structure prediction, sequence recovery, tetrahedral nanoparticles
MeSH terms
Deep Learning, Amino Acid Sequence, Protein Conformation, Proteins, Protein Engineering, Crystallography, X-Ray, Cryoelectron Microscopy

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.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-10.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-10.

Cited sources are checked by 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