CoChat
Sign Up
PreprintCited by 629Protein Structure and DynamicsMachine Learning in BioinformaticsComputational Drug Discovery Methods

De novo protein design by deep network hallucination

Open access
Anishchenko, Ivan, Chidyausiku, Tamuka M., Ovchinnikov, Sergey, Pellock, Samuel J., Baker, David·2020·Nature
DOI 10.1101/2020.07.22.211482PMC PMC9293396

Preprint — published as 10.1038/s41586-021-04184-w · Paper page

Open this paper in CoChat — free accountOpen PDF

Abstract

There has been considerable recent progress in protein structure prediction using deep neural networks to infer distance constraints from amino acid residue co-evolution 1–3 . We investigated whether the information captured by such networks is sufficiently rich to generate new folded proteins with sequences unrelated to those of the naturally occuring proteins used in training the models. We generated random amino acid sequences, and input them into the trRosetta structure prediction network to predict starting distance maps, which as expected are quite featureless. We then carried out Monte Carlo sampling in amino acid sequence space, optimizing the contrast (KL-divergence) between the distance distributions predicted by the network and the background distribution. Optimization from different random starting points resulted in a wide range of proteins with diverse sequences and all alpha, all beta sheet, and mixed alpha-beta structures. We obtained synthetic genes encoding 129 of these network hallucinated sequences, expressed and purified the proteins in E coli, and found that 27 folded to monomeric stable structures with circular dichroism spectra consistent with the hallucinated structures. Thus deep networks trained to predict native protein structures from their sequences can be inverted to design new proteins, and such networks and methods should contribute, alongside traditional physically based models, to the de novo design of proteins with new functions.

AI snapshot

Written from the paper’s full text.

TL;DR

The authors invert the trRosetta structure-prediction network by Monte Carlo sampling in amino acid sequence space to “hallucinate” new 100-residue proteins with sequences unrelated to native proteins. Of 129 selected designs expressed in E. coli, 27 folded to monodisperse species with CD spectra matching the hallucinated structures, and three experimentally solved structures closely matched the models.

Key findings

  • Hallucinated sequences had very low similarity to native proteins, with best BLAST E-values to UniProt of ~0.1, while their predicted structures had TM-align scores of 0.6–0.9 to native PDB structures.
  • Of 129 hallucinated proteins expressed and purified from E. coli, 27 yielded SEC peaks corresponding to monomeric or small oligomeric species and CD spectra consistent with the target structures; 21 were highly thermostable with apparent melting temperatures above 70 °C, and alpha-beta designs showed no unfolding transitions up to 95 °C.
  • In orthogonal Rosetta folding simulations, 82 of 129 designs had lowest-energy structures close to the hallucinated structures with Cα-RMSDs <3.0 Å, and for all 129 the design model was lower in energy than any structure obtained from an extended chain.
  • Experimental structures matched hallucinated models: NMR structure of design 0515 had ~1.84 Å Cα RMSD to the model; X-ray structure of 0217 at 2.9 Å resolution had 2.53 Å backbone RMSD over all 100 residues; X-ray structure of 0738_mod at 2.4 Å resolution had 3.68 Å Cα RMSD over 96 residues, with N- and C-terminal halves aligning at 1.32 Å over 57 residues and 2.17 Å over 43 residues.

Methods

  • The trRosetta network predicts inter-residue distance and orientation distributions from sequence; a background network with similar architecture but no amino acid sequence identity information provided a reference distribution. The objective combined KL-divergence of trRosetta predictions from the background with negative KL-divergence of amino acid composition from PDB frequencies, optimized by simulated annealing in sequence space with 40,000 attempted moves, starting temperature 0.1 halved every 5,000 steps, cysteines excluded, and length L = 100.
  • Two thousand designs were generated, clustered by TM-score into 95 clusters at a 0.75 threshold; 129 sequences from 27 structural clusters were selected for experimental testing, expressed in E. coli BL21(DE3), purified by IMAC and size-exclusion chromatography, and characterized by circular dichroism, NMR, and X-ray crystallography.
  • Structural comparisons used TM-align, Theseus superposition, and Rosetta ab initio folding simulations; crystallographic data were processed with HKL2000/Scalepack and structures solved by molecular replacement.

Limitations

  • The paper states that many hallucinated proteins form oligomers in solution, likely because the network incorporated sequence features from native protein-protein interfaces; surface hydrophobic substitutions were needed to produce monomeric species in some cases.
  • The paper states that design 0738_mod showed multiple monomeric conformations in slow conformational exchange, and suggests that incorporating an explicit sidechain representation in the hallucination method could reduce such structural heterogeneity.
  • Evident from the text: only 27 of 129 expressed designs folded to monodisperse species, only three structures were determined experimentally, one of those required surface modification, and the study did not test designed functions.

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

Integrity signals

Registry

Crossref · 10.1101/2020.07.22.211482

Cites

68 of 68 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 8 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    The overall distributions of hallucinated sequences and structures are very different from those of naturally occurring proteins of the same 100-residue length used during trRosetta training.

    “The overall distributions of hallucinated sequences and structures are very different from those of naturally occurring proteins of the same (100 residue) length which were used during trRosetta training (Extended Data Fig. 1a–e).”
    Quote in the paperSubstantial claim
  • Findingown data

    For all 129 hallucinated proteins, the lowest energy structure sampled starting from the design model was lower in energy than any other structure obtained starting from an extended chain.

    “for all 129 the lowest energy structure sampled starting from the design model was lower in energy than any other structure obtained starting from an extended chain.”
    Quote in the paperSubstantial claim
  • Findingown data

    Hallucinated proteins resemble one another more closely than they resemble natural proteins.

    “The hallucinated proteins are more similar to each other (blue lines) than they are to natural proteins (grey lines).”
    Quote in the paperSubstantial claim
  • Findingown data

    The hallucinated designs are more ideal than natural proteins in having fewer loops but longer secondary structure elements.

    “Hallucinations are more ideal than natural proteins in having less loops but longer secondary structure elements.”
    Quote in the paperNot a substantial claim
  • Findingcitation

    The trRosetta deep network, trained with multiple sequence information, can consistently predict structure accurately for de novo designed proteins from a single sequence, without co-evolution information.

    “we found that the trRosetta deep neural network trained using multiple sequence information could consistently predict structure quite accurately for de novo designed proteins from just a single sequence -- i.e., in the complete absence of co-evolution information3.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [3] Yang J Improved protein structure prediction using predicted interresidue orientations. Proc. Natl. Acad. Sci. U. S. A 117, 1496– 1503 (2020). 31896580 PMC6983395 source

      Not checked in the free run.

  • May combine two claims. 30
  • The quote only partly states this claim. 2
  • Classified as probably not a substantial claim. 5

  • May combine two claims.

    Findingown data

    For the 27 proteins, circular dichroism spectra were consistent with the target structures, with characteristic profiles of all-alpha helical proteins for all-alpha designs and of alpha-beta proteins for alpha-beta designs.

    “In all cases, the CD spectra were consistent with the target structures (Figs. 3e, 4e and Extended Data Fig. 3e,k), with the characteristic profiles of all alpha helical proteins for the all alpha helical designs (Fig. 3e and Extended Data Fig. 3e), and of alpha-beta proteins for the alpha-beta designs”
    Quote in the paperSubstantial claim

  • The quote only partly states this claim.

    Findingown data

    All experimentally validated protein sequences are predicted by Rosetta large-scale energy calculations to have funneled landscapes leading into the target structure.

    “all of the sequences are predicted by Rosetta large scale energy calculations to have funneled landscapes leading into the target structure”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The surface-modified hallucination 0738_mod yielded crystals, and its 2.4 Å resolution structure was determined.

    “One of these surface-modified hallucinations, 0738_mod, yielded crystals and the 2.4 Å resolution structure was determined.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Both de novo designed proteins and the hallucinated proteins have regular alpha helices and beta sheets and lack long loops and other idiosyncrasies of native protein structures.

    “Both de novo designed proteins and the hallucinated proteins generated here have regular alpha helices and beta sheets, and lack the long loops and other idiosyncrasies of native protein structures”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The 2.9 Å resolution crystal structure of design 0217 revealed a 3-helix bundle with an overall fold similar to the hallucinated model, with a backbone RMSD of 2.53 Å over all 100 residues.

    “We also succeeded in determining a 2.9 Å resolution crystal structure of design 0217, which revealed a 3-helix bundle with an overall fold similar to the hallucinated model; the backbone RMSD between model and crystal structure is 2.53 Å over all 100 residues”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The homodimer 0217 has a lower average heteronuclear NOE than 0515 and 0738_mod, and it shows fewer peaks than expected because of exchange broadening.

    “The average hetNOE for the homodimer 0217 is lower than for 0515 and 0738_mod, and it has fewer peaks than expected due to exchange broadening.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Despite register shifts, the N- and C-terminal halves of the 0738_mod crystal structure align remarkably well to the corresponding regions in the hallucinated model, with backbone RMSDs of 1.32 Å over 57 residues and 2.17 Å over 43 residues, respectively.

    “the N- and C-terminal halves of the crystal structure align remarkably well to the corresponding regions in the hallucinated model with backbone RMSDs of 1.32 Å over 57 residues, and 2.17 Å over 43 residues for the N-terminal and C-terminal half, respectively”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    NMR data on design 0417 are consistent with the alpha/beta fold of the hallucinated model, and SEC and NMR relaxation measurements indicate it is primarily monomeric with some transient self-association in solution.

    “NMR data on design 0417 are consistent with the alpha/beta fold of the hallucinated model, and both SEC and NMR relaxation measurements (15N T1 ~730 ms, 15N T2 ~77 ms, and τc ~10.4 ns at 25 °C, Extended Data Fig. 6) indicate that it is primarily monomeric but with some transient self-association in solution.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Design 0217 forms a dimer in the crystal structure, consistent with SEC-MALS analysis and NMR rotational correlation time measurements.

    “For example, 0217 forms a dimer in the crystal structure (Extended Data Fig. 5), consistent with SEC-MALS analysis and NMR rotational correlation time measurements”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The predicted structures resemble but are not identical to native structures in the PDB, with TM-align scores of 0.6-0.9.

    “the predicted structures resemble but are not identical to native structures in the PDB, with TM-align scores of 0.6–0.9 (Fig. 1g).”
    Quote in the paperSubstantial claim

  • The quote only partly states this claim.

    Findingown data

    Monte Carlo optimization of KL divergence between trRosetta-predicted distance distributions and background distributions yields novel proteins spanning a wide range of sequences and predicted structures.

    “Optimization from different random starting points results in novel proteins spanning a very wide range of sequences and predicted structures.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The 0217 model displays amphipathic sequence patterning across the 3-helix topology, with numerous solvent-exposed hydrophobic residues that mediate the dimer contacts in the crystal structure.

    “the 0217 model displays clear amphipathic sequence patterning across the 3-helix topology, with numerous solvent-exposed hydrophobic residues (Suppl. Fig. 2) that mediate the dimer contacts in the crystal structure.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Solution data for designs 0515 and 0217 suggest well-ordered structures in solution, with internal dynamics typical of small natural proteins.

    “The solution data for designs 0515 and 0217 (dimer) suggest well-ordered structures in solution, with internal dynamics typical of small natural proteins.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    A single molecular replacement solution for 0217 was obtained in space group I 2 3 using one ab initio model; it had two copies per asymmetric unit and a TFZ score of 13.3, while no other model scored above 8.

    “A single solution was found in I 2 3 from one of the ab initio models with two copies in the asymmetric unit and a TFZ score of 13.3 (no other model yielded a TFZ score >8).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    SEC data demonstrate that the proteins are monodisperse in solution, with 0515 and 0738_mod predominantly monomeric and 0217 predominantly dimeric.

    “SEC data demonstrating monodispersity of these proteins in solution, with predominantly monomer for 0515 and 0738_mod and predominantly dimer for 0217.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The accuracy of the 0738_mod structure does not reflect PDB memorization; the closest BLAST hits in the PDB for the N- and C-terminal halves have E-values of 0.29 and 0.63, respectively.

    “The accuracy does not reflect PDB memorization; the closest BLAST hits in the PDB for the N and C terminal halves have E-values of 0.29 and 0.63, respectively.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Two thousand proteins were generated using the hallucination procedure and compared structurally using TM-score; average-linkage hierarchical clustering yielded 95 clusters with an average inter-cluster similarity of TM-score 0.75.

    “Two thousand proteins were generated using the hallucination procedure described above, and structurally compared to each other using the template modeling score (TM-score)29. Average-linkage hierarchical clustering yielded 95 clusters with an average inter-cluster similarity of TM-score = 0.75.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [29] Zhang Y & Skolnick J TM-align: a protein structure alignment algorithm based on the TM-score. Nucleic Acids Res. 33, 2302– 2309 (2005). 15849316 PMC1084323 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    One hundred twenty-nine hallucinated sequences from the remaining 27 structural clusters were selected for experimental testing.

    “One-hundred twenty-nine hallucinated sequences from the remaining 27 structural clusters (no more than 10 designs per cluster) were selected for experimental testing.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The hallucinated structures span all-alpha, all-beta, and mixed alpha-beta fold classes, with 95 different sub-folds at a TM-score clustering threshold of 0.75.

    “The structures span all alpha, all beta and mixed alpha-beta fold classes, with 95 different sub-folds at a TM-score clustering threshold of 0.75.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The solution NMR structure of design 0515 is a monomeric antiparallel four-helix bundle, and the ensemble of 20 structures has a Cα RMSD to the hallucinated model of about 1.84 Å.

    “We determined the solution NMR structure for design 0515 to be a monomeric antiparallel four-helix bundle (1D estimated 15N T1 ~ 780 ms, 15N T2 ~ 77 ms, and τc ~ 9.6 ns at 25 °C); the ensemble of 20 structures had a Cɑ root mean square deviation (RMSD) to the hallucinated model of ~1.84 Å”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    For 0217, molecular replacement with trRosetta and classical Rosetta ab initio models both produced converged ensembles on one topology, but the ab initio set showed substantial internal diversity.

    “While both sets of predictions yielded converged ensembles on a single topology, the classical ab initio models had significant diversity within that ensemble.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    0738_mod has a heteronuclear NOE distribution similar to that of monomeric 0515, but it shows over twice the expected number of peaks, which suggests at least two folded conformations in solution undergoing slow exchange on the NMR timescale.

    “Although 0738_mod has a similar hetNOE distribution as monomeric 0515, it has more than double the expected number of peaks, indicating at least two folded conformations (for all or parts of the protein) in solution that are in slow conformational exchange on the NMR time-scale.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Study design and conductown data

    SDS-PAGE data show that each of the three designs is more than 95% homogeneous, which together with MALDI-TOF mass spectrometry indicates that the observed spectral heterogeneity is not due to chemical heterogeneity.

    “SDS-PAGE data (not shown) show that each is >95% homogeneous, which together with MALDI-TOF mass spectrometry indicate that the spectral heterogeneity observed is not due to chemical heterogeneity.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Random sequences give diffuse trRosetta predictions, while existing de novo designs produce peaked distributions with low variance.

    “Random sequences give diffuse predictions, while existing de novo designs produce peaked distributions with low variance3.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [3] Yang J Improved protein structure prediction using predicted interresidue orientations. Proc. Natl. Acad. Sci. U. S. A 117, 1496– 1503 (2020). 31896580 PMC6983395 source

      Not checked in the free run.

  • May combine two claims.

    Study design and conductreasoning

    The Rosetta folding test is completely orthogonal because the network was trained exclusively on native protein structures and has no access to the Rosetta energy function.

    “This is a completely orthogonal test as the network was trained exclusively on native protein structures, and has no access to the Rosetta energy function.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductown data

    For the 0738_mod crystal structure, the final Ramachandran favored and outlier percentages were 96% and 0%, respectively.

    “For 0738_mod refinement, final Ramachandran favored and outliers were 96% and 0%, respectively.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductown data

    For the 0217 crystal structure, the final Ramachandran favored and outlier percentages were 99% and 0%, respectively.

    “For 0217 final Ramachandran favored and outliers were 99% and 0%, respectively.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Many hallucinated proteins are monomeric, stable, have the expected secondary structure, and are strongly predicted by Rosetta to fold to the target structure in orthogonal calculations.

    “Many of the hallucinated proteins are monomeric, stable, have the expected secondary structure, and are strongly predicted to fold to the target structure by Rosetta in completely orthogonal calculations”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The network appears to incorporate sequence features associated with protein-protein interfaces of native oligomeric proteins in the PDB training set, likely explaining why many hallucinated proteins self-associate.

    “The network appears to incorporate sequence features associated with the protein-protein interfaces of the native oligomeric proteins included in the PDB training set, likely explaining why many of the network hallucinated proteins self associate.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Even under optimized conditions, 0417 shows evidence of exchange broadening, temperature- and glycerol-dependent resonances, and doubled resonances, all indicating transient self-association.

    “Even under optimized conditions, there is still evidence of exchange broadening (e.g. Trp side chain NεHs are weak), resonances that appear only at high temperature and high glycerol concentrations, and some resonances that are doubled; all indications of transient self-association.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Study design and conductown data

    The 0515 sample used for structure determination was over 98% pure by SDS-PAGE and over 95% labeled with isotopes according to MALDI-TOF mass spectrometry.

    “This sample was >98% homogeneous by SDS-PAGE analysis, and >95% isotope enriched based on MALDI-TOF mass spectrometry.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    In benchmark tests, the trRosetta network outperformed other methods.

    “in benchmark tests this network outperformed other methods3.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [3] Yang J Improved protein structure prediction using predicted interresidue orientations. Proc. Natl. Acad. Sci. U. S. A 117, 1496– 1503 (2020). 31896580 PMC6983395 source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    The trRosetta network is a 2D residual-convolutional neural network that takes 1- and 2-site features from a multiple sequence alignment or a single sequence and outputs probabilistic distance and orientation distributions for all residue pairs.

    “The trRosetta protein structure prediction network, described in detail elsewhere3, is a 2D residual-convolutional neural network which takes 1- and 2-site features derived from a multiple sequence alignment or a single sequence as an input and produces a 2D output describing distances and orientations for all residue pairs in a protein in a probabilistic manner”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [3] Yang J Improved protein structure prediction using predicted interresidue orientations. Proc. Natl. Acad. Sci. U. S. A 117, 1496– 1503 (2020). 31896580 PMC6983395 source

      Not checked in the free run.

  • May combine two claims.

    Study design and conductown data

    For 0738_mod, the two-conformation interpretation was supported by new peaks at 288 K, altered peaks at 308 and 318 K, and by 15N ZZ-exchange cross peaks detected at 318 K using 600 and 750 ms mixing times.

    “This was further validated by the appearance of new peaks in spectra at lower temperature (288K), and different peaks at higher temperatures (308 and 318K), and confirmed by detection of 15N ZZ-exchange cross peaks at 318K with 600 and 750 ms mixing times (Bruker pulse sequence hsqcetexf3gp, data not shown)58.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [58] Montelione GT & Wagner G 2D Chemical exchange NMR spectroscopy by proton-detected heteronuclear correlation. Journal of the American Chemical Society vol. 111 3096– 3098 (1989).

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    The NMR samples were each about 98% or more homogeneous based on SDS-PAGE.

    “These samples were each >~98% homogeneous, based on SDS-PAGE.”
    Quote in the paperNot a substantial claim
35 of 35

  • The paper relies on this without support.

    Random amino acid sequences have a vanishingly small probability of folding to a defined structure.

    “As expected for random sequences, which have a vanishingly small probability of folding to a defined structure”

  • The authors state this as an assumption.

    The background network's predictions can be viewed as representing a generic molten globule state.

    “this can loosely be viewed as representing a generic “molten globule” state”
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?

  • Deep learning methods such as AlphaFold and trRosetta, which extract information from the PDB, outperform traditional approaches that explicitly model the folding process.

    “The bi-annual Critical Assessment of Structure (CASP) meetings have demonstrated that deep learning methods such as AlphaFold (1, 2) and trRosetta (3), that extract information from the large database of known protein structures in the PDB, outperform more traditional approaches that explicitly model the folding process.”

    From Accurate prediction of protein structures and interactions using a 3-track network (2021) · cites this paper as [2]

    The Introduction (page not provided) says deep networks improved protein-structure prediction accuracy, notes an AlphaFold observation at CASP13, and reports that trRosetta outperformed other methods in benchmark tests. The Conclusion says the network learned from native protein structures and contrasts it with physically based approaches. The paper does not establish the claimed CASP comparison in which both AlphaFold and trRosetta outperform approaches that explicitly model folding. (LLM, conf=0.82)

    Experimental — accuracy not yet measured

  • Crystal and NMR structures confirm that hallucinated sequences fold to the hallucinated structures.

    “crystal and NMR structures confirm that the hallucinated sequences fold to the hallucinated structures (12).”

    From Scaffolding protein functional sites using deep learning (2022) · cites this paper as [12]

  • The trRosetta structure-prediction neural network can be used to generate new proteins by maximizing the trRosetta output probability that a sequence folds to some unspecified three-dimensional structure during Monte Carlo sampling in sequence space.

    “We previously demonstrated that the trRosetta structure-prediction neural network (11) can be used to generate new proteins by maximizing the trRosetta output probability that a sequence folds to some (unspecified) three dimensional structure during Monte Carlo sampling in sequence space (12).”

    From Scaffolding protein functional sites using deep learning (2022) · cites this paper as [12]

  • There has been recent success in applying the hallucination technique to general protein design.

    “There has also been recent success in applying the hallucination technique to general protein design [72].”

    From Computational and artificial intelligence-based methods for antibody development (2023) · cites this paper as [72]

  • The RoseTTAFold approach has allowed for de novo design through network hallucination by training a similar model specially for inpainting portions of proteins given context.

    “The competing RoseTTAFold approach has allowed for de novo design through network hallucination by training a similar model specially for inpainting portions of proteins given context [72,83].”

    From Computational and artificial intelligence-based methods for antibody development (2023) · cites this paper as [72]

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.

  • The reported resolution for 0217 differs from its table entry

    Minor

    The text reports a 2.9 Å crystal structure for 0217, but the crystallographic table gives its data-collection resolution as 3.0 Å.

1 of 1

Cites

Flagged first, then by citations

Cited by

Sorted by citations

Related

Authors
Anishchenko, I., Chidyausiku, T. M., Ovchinnikov, S., Pellock, S. J., & Baker, D.
Affiliations
  • University of Washington (US)
  • Howard Hughes Medical Institute (US)
Corresponding author
Baker, David
ORCID
Venue
Nature
Publisher
openRxiv
Published
2020-07-23
Volume
600
Issue
7889
Pages
547-552
ISSN
0028-0836 (ISSN-L)
DOI
10.1101/2020.07.22.211482
PMC
PMC9293396
OpenAlex
W3216341763
Cited by
629
References
61
Citations by year
2026: 102 · 2025: 120 · 2024: 139 · 2023: 122 · 2022: 110 · 2021: 29 · 2020: 6
FWCI
31.36
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature (publishedVersion)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • PubMed Central (acceptedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
Funders
  • Howard Hughes Medical Institute
  • Novo Nordisk Fonden (NNF18OC0030446)
  • National Institutes of Health (1s10od021527-01, DP5OD026389)
  • National Institute of General Medical Sciences (R01GM120574, R35GM141818, P30GM124165)
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
de novo protein design, deep neural network, hallucinations, protein structure prediction, Kullback-Leibler divergence, synthetic proteins, protein folding, circular dichroism, X-ray crystallography, nuclear magnetic resonance
MeSH terms
Amino Acid Sequence, Hallucinations, Humans, Protein Conformation, Proteins, Neural Networks, Computer, Crystallography, X-Ray

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 2026-10-05.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-05.

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