CoChat
Sign Up
Cited by 1,078Protein Structure and DynamicsMachine Learning in BioinformaticsEnzyme Catalysis and Immobilization

Large language models generate functional protein sequences across diverse families

Open access
Madani, Ali, Krause, Ben, Greene, Eric R., Subramanian, Subu, Mohr, Benjamin P., Holton, James M., Olmos, Jose Luis, Xiong, Caiming +4 more·2023·Nature Biotechnology
DOI 10.1038/s41587-022-01618-2PMC PMC10400306
Open this paper in CoChat — free accountOpen PDF

Abstract

Deep-learning language models have shown promise in various biotechnological applications, including protein design and engineering. Here we describe ProGen, a language model that can generate protein sequences with a predictable function across large protein families, akin to generating grammatically and semantically correct natural language sentences on diverse topics. The model was trained on 280 million protein sequences from >19,000 families and is augmented with control tags specifying protein properties. ProGen can be further fine-tuned to curated sequences and tags to improve controllable generation performance of proteins from families with sufficient homologous samples. Artificial proteins fine-tuned to five distinct lysozyme families showed similar catalytic efficiencies as natural lysozymes, with sequence identity to natural proteins as low as 31.4%. ProGen is readily adapted to diverse protein families, as we demonstrate with chorismate mutase and malate dehydrogenase.

Integrity signals

Registry

Crossref · 10.1038/s41587-022-01618-2

Cites

0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 23 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    ProGen was trained only to predict the next amino acid in over 280 million sequences from thousands of families, without explicit biophysical modeling.

    “Our protein language model is trained by simply learning to predict the next amino acid for over 280 million protein sequences from thousands of protein families, without explicit biophysical modeling.”
    Quote in the paperSubstantial claim
  • Findingreasoning

    After training, ProGen can generate full-length protein sequences from scratch for any specified protein family, with varying similarity to natural proteins.

    “After training, ProGen can be prompted to generate full-length protein sequences for any protein family from scratch, with a varying degree of similarity to natural proteins.”
    Quote in the paperSubstantial claim
  • Findingown data

    ProGen-generated artificial lysozymes cover the sequence landscape of natural lysozymes across five families with diverse folds, active sites, and mechanisms.

    “Our artificial lysozymes span the sequence landscape of natural lysozymes, Fig. 2a, across five families which contain diverse protein folds, active site architectures, and enzymatic mechanisms49,50.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [49] Baase, W. A., Liu, L., Tronrud, D. E. & Matthews, B. W. Lessons from the lysozyme of phage T4. Protein Sci. 19, 631–641 (2010). source

      Not checked in the free run.

    • cites [50] Kuroki, R., Weaver, L. H. & Matthews, B. W. A covalent enzyme–substrate intermediate with saccharide distortion in a mutant T4 lysozyme. Science 262, 2030–2033 (1993). source

      Not checked in the free run.

    • Rests on an assumption: The five lysozyme families chosen because they lack disulfide bonds can be expressed in high-throughput cell-free systems.
  • Findingown data

    Within each family, artificial and natural lysozyme sequences have similar residue-position entropies when aligned separately.

    “they demonstrate similar residue position entropies when forming separate multiple sequence alignments of natural and artificial proteins within each family, Fig. 2c.”
    Quote in the paperSubstantial claim
  • Findingown data

    In the natural positive-control set of 100 proteins, 72% expressed well by chromatography peaks and band visualization.

    “In the positive control set of 100 natural proteins, 72% were well-expressed based on chromatography peaks and band visualization.”
    Quote in the paperSubstantial claim
  • May combine two claims. 29
  • Classified as probably not a substantial claim. 16

  • May combine two claims.

    Findingnone given

    ProGen is a transformer language model with 1.2 billion parameters trained on a public dataset of 280 million protein sequences.

    “ProGen is a 1.2 billion-parameter neural network trained using a publicly available dataset of 280 million protein sequences.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The purified apparent lower-molecular-weight species of L056 and L070 were highly active and had Michaelis-Menten parameters comparable to HEWL.

    “we could readily measure the KM values for the purified apparent lower molecular weight species, where both L056 and L070 were highly active and had comparable Michaelis-Menten parameters to HEWL (Fig. 3e).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Of 90 artificial proteins tested, 66 (73%) were functional and showed high functionality across families.

    “Among our artificial proteins, 73% (66/90) were functional and exhibited high levels of functionality across families, Fig. 3c.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    A few natural and artificial proteins had activity within tenfold of HEWL, which was much more active than all negative controls.

    “Notably a small number of proteins, both within the natural and artificial proteins, were within an order of magnitude of HEWL, which was substantially more active than all negative controls.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    L056 and L070 expressed well and were bactericidal to E. coli BL21(DE3) during overnight induction at 16C.

    “Two proteins, L056 (max ID 69.6%) and L070 (max ID 89.2%) expressed well and incurred bactericidal activities towards the E. coli BL21(DE3) strain used during overnight induction at 16C.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    ProGen-generated proteins expressed as well as natural controls, with 72 of 100 expressing across maxID bins from 40% to 90%.

    “The ProGen-generated proteins express equally well (72/100 total) across all bins of sequence identity to any known natural protein (maxID 40–90%), Fig 2e.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    L070 and L056 share only 17.9% sequence identity despite comparable enzymatic activities.

    “Despite having comparable enzymatic activities, L070 and L056 only share 17.9% sequence ID.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Mutations separating L056 and L070 from their closest natural homologs are uniformly distributed, with no bias by location or structural element.

    “We also found that there is no bias to location or structural element to the mutations that diverge L056 and L070 from their respective nearest sequence homolog in nature. Differing residues are instead uniformly distributed.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The monomers of L056 and L070 were active in a pseudo-first-order assay, but their KM values were too weak to measure with the heterogeneous fluorescein-labeled Micrococcus lysodeikticus cell wall substrate.

    “The KM values of both monomers were too weak to be measured using a heterogeneous, fluorescein-labeled Micrococcus lysodeikticus cell wall substrate (Molecular Probes EnzChek Lysozyme kit); however, both monomers were active using a pseudo-first order kinetic assay (Fig. S7).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Some L056 mutations lie in the active site cleft and in regions that affect conformational state.

    “Some mutations are even found within the active site cleft and within regions that influence conformational state (for L056).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The 2.5 Å crystal structure of L056 has a global fold similar to predictions, with Cα RMSD 2.9 Å from trRosetta and 2.3 Å from WT T4 lysozyme.

    “We determined a 2.5 Å resolution crystal of L056 (see Methods, Figure 3e). The global fold was similar to predictions, with a Cα RMSD of 2.9 Å from the backbone structure predicted by trRosetta and 2.3 Å RMSD from a WT T4 Lysozyme structure59,60.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [59] Norn, C. et al. Protein sequence design by conformational landscape optimization. Proc. Natl Acad. Sci. USA 118, e2017228118 (2021). source

      Not checked in the free run.

    • cites [60] Anand, N. et al. Protein sequence design with a learned potential. Nat. Commun. 13, 746 (2022). source

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    bmDCA could not fit three of the five lysozyme families and gave detectable expression for 30 of 50 proteins in the remaining two.

    “bmDCA was unable to fit three out of the five lysozyme families, and exhibited 60% detectable expression (30/50 proteins) for the remaining two protein families.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [7] Russ, W. P. et al. An evolution-based model for designing chorismate mutase enzymes. Science 369, 440–445 (2020). source

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    The L056 active site cleft is well formed, with catalytic residue Glu15 and substrate-binding residue Thr30 correctly positioned.

    “Alignment with a structure featuring a covalently trapped substrate (PDB 148L) reveals that the active site cleft is well formed with the key catalytic residue Glu15 (Glu 11 in T4L) and key substrate binding residue Thr30 (Thr26 in T4L) correctly positioned.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    ProGen trained only on the universal dataset had a significant AUC drop of 0.18 for CM and 0.08 for MDH compared with fine-tuned ProGen.

    “These likelihoods showed a significant drop in AUC of 0.18 for CM (p<0.0001, two-tailed test, n=1617) and 0.08 for MDH (p<0.1, two-tailed test, n=56), as compared to fine-tuned ProGen when predicting if an artificial sequence should function.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    On MDH function data, ProGen log-likelihoods had AUC 0.94, better than ProteinGAN discriminator scores at 0.87 (p<0.1, two-tailed, n=56).

    “On MDH function data, ProGen log-likelihoods had an AUC of 0.94 (Fig. 4f), which was better than ProteinGAN discriminator scores, with an AUC of 0.87 (p<0.1, two-tailed test, n=56).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [66] Vig, J. et al. BERTology meets biology: Interpreting attention in protein language models. In International Conference on Learning Representations (ICLR, 2020). source

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    On CM function data, ProGen log-likelihoods had AUC 0.85, significantly better than bmDCA's 0.78 (p<0.0001, two-tailed, n=1617).

    “On CM function data, ProGen log-likelihoods had an area under the curve (AUC) of 0.85, significantly better (p<0.0001, two-tailed test, n=1617) than bmDCA, which had an AUC of 0.78 (Fig. 4c).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [7] Russ, W. P. et al. An evolution-based model for designing chorismate mutase enzymes. Science 369, 440–445 (2020). source

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    Six purified highly expressed low-identity artificial proteins were all active but had significantly lower Michaelis-Menten activities than HEWL or L056 and L070.

    “We purified 6 highly expressed proteins and found that they were all active, but with significantly lower Michaelis-Menten activities than HEWL or the previously generated artificial proteins L056 and L070 (Fig 3f).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Among 95 new low-identity artificial sequences for two lysozyme families, 78 of 89 (88%) expressed well and 24 of 78 (31%) were soluble.

    “Of the selected sequences, 78 out of the 89 (88%) expressed well and 24 out of the 78 (31%) were soluble (Supplemental Figure 8).”
    Quote in the paperSubstantial claim

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

    Findingown data

    Variants L013 and L038 expressed at high levels but accumulated in inclusion bodies.

    “L013 and L038 expressed highly to inclusion bodies.”
    Quote in the paperNot a substantial claim

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

    Findingown data

    The lowID lysozyme variant A5 had lower catalytic activity, requiring a modified assay protocol.

    “For lowID lysozyme A5 the above protocol was altered slightly to accommodate lower catalytic activity of these variants: reaction volumes were increased to 20 μL, plate was covered with an optically transparent seal (Microseal ® ‘B’ seal; BioRad) to mitigate sample evaporation, fluorescent reads were taken every 5 mins for 16 hours with 5 s of linear plate shaking prior to each measurement to min”
    Quote in the paperNot a substantial claim

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

    Findingnone given

    The MDH functional labels used in the analysis are binary.

    “The MDH functional labels are binary, so no thresholding was needed to compute AUCs.”
    Quote in the paperNot a substantial claim

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

    Findingown data

    The kinetic data for L056 and L070 were not fit by a single exponential model but were fit by a double exponential model.

    “The resultant data were not described by a single exponential model but were described by a double exponential model (Eq. 6), likely due to the heterogeneity of the substrate, and all data were fit in IgorPro7.”
    Quote in the paperNot a substantial claim

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

    Findingown data

    In induced cultures of L056 and L070, substantial turbidity appeared in the spent medium after cells were removed by centrifugation.

    “For induced cultures of L056 and L070, significant turbidity was observed in the spent media after cells were pelleted at 3,500 r.c.f. for 30 mins at 4°C.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    For each of L056 and L070, size-exclusion chromatography resolved two peaks, and both peaks had lysozyme activity against the fluorescein-labeled M. lysodeikticus cell wall substrate.

    “Two peaks were resolved for each variant and that harbored lysozyme activity against the fluorescein-labeled M. Lysodeikticus cell wall substrate (EnzChek kit; ThermoFisher).”
    • Also stated in: § Purification of L056 and L070 from spent media · plain
      “Two peaks were resolved for each variant and that harbored lysozyme activity against the fluorescein-labeled M. Lysodeikticus cell wall substrate (EnzChek kit; ThermoFisher) that corresponded to peaks observed in the first pool purification.”
    Quote in the paperNot a substantial claim

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

    Findingown data

    In the crystallographic analysis of L056, significant translational non-crystallographic symmetry and differences from the search model made the initial electron density maps difficult to interpret.

    “Significant translational non-crystallographic symmetry and differences with the search model resulted in maps that were initially hard to interpret.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    ProGen generates sequences that express well and retain enzymatic function across diverse sequence landscapes and families.

    “These results indicate that ProGen generates protein sequences that not only can express well but also maintain enzymatic function for diverse sequence landscapes across protein families.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Well-expression and solubility in the cell-free system indicate that the artificial proteins are structurally well-folded.

  • May combine two claims.

    Interpretationown data

    L056 and L070 have catalytic and bactericidal activity comparable to HEWL while differing from their closest natural sequences by 53 and 18 amino acids, respectively.

    “Taken together, L056 and L070 harbor potent catalytic activity and bactericidal capabilities that are comparable to HEWL, while diverging from their nearest known natural sequence by 53 and 18 amino acids respectively.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: HEWL is a suitable reference for native lysozyme catalytic activity.

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

    Study design and conductnone given

    Held-out test perplexity was monitored and never exceeded training-set perplexity during training.

    “The perplexity on a held-out test set was monitored and did not exceed training set perplexity throughout model training.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    ProGen can significantly expand the protein sequence space beyond evolution's samples, though it is not expected to generate proteins from a completely different distribution or domain.

    “While we do not expect our language model to generate proteins that belong to a completely different distribution or domain (for example creating a new fold that catalyzes an unnatural reaction), it can significantly expand the space of protein sequences from those sampled by evolution.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The 280 million-sequence training set is diverse enough to learn a universal protein representation transferable across families.

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductown data

    The dataset was split into 280 million training sequences and held-out out-of-distribution (100k from 20 families) and in-domain (1 million) test sets.

    “The aggregated dataset was split into a training set of size 280 million and two test sets, an out-of-distribution test set (OOD-test) of size 100k from 20 protein families and a randomly sampled in-domain test set (ID-test) of size 1 million, that were held-out for training and used for evaluation.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    ProGen's model likelihoods align better with experimental assay data for CM and MDH than the original family-specific generation methods.

    “In sum, ProGen’s model likelihoods are better aligned with experimentally-measured assay data on two diverse protein datasets—chorismate mutase and malate dehydrogenase—than the sequence generation methods from original studies specifically tailored for these families.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [7] Russ, W. P. et al. An evolution-based model for designing chorismate mutase enzymes. Science 369, 440–445 (2020). source

      Not checked in the free run.

    • cites [66] Vig, J. et al. BERTology meets biology: Interpreting attention in protein language models. In International Conference on Learning Representations (ICLR, 2020). source

      Not checked in the free run.

  • May combine two claims.

    Interpretationown data

    Both universal-dataset pretraining and family-specific fine-tuning contribute significantly to final model performance.

    “These results indicate that both components of our training strategy—initial training on the universal sequence dataset and fine-tuning on the protein family of interest—contribute significantly to final model performance.”
    Quote in the paperSubstantial claim

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

    Study design and conductown data

    Three adversarial discriminators were trained with different random seeds for robustness.

    “For robustness, we trained three discriminators using different random seeds.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    A protein was called functional if its fluorescence exceeded the maximum negative-control fluorescence by one standard deviation.

    “Proteins that generated fluorescence one standard deviation above the maximum fluorescence of any negative control were considered functional.”
    Quote in the paperSubstantial claim

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

    Interpretationown data

    The double exponential kinetics observed for L056 and L070 are likely caused by heterogeneity of the fluorescein-labeled M. lysodeikticus cell wall substrate.

    “The resultant data were not described by a single exponential model but were described by a double exponential model (Eq. 6), likely due to the heterogeneity of the substrate, and all data were fit in IgorPro7.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The average molecular weight of the fluorescein-labeled M. lysodeikticus cell wall substrate is unknown and the substrate is heterogeneous.

  • Classified as probably not a substantial claim.

    Study design and conductown data

    No two selected generated sequences shared more than 80% identity with each other.

    “We also ensured that no two generated samples share over 80% identity with each other.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductown data

    Bootstrapping was used to compute the statistical significance of the difference in AUC between fine-tuned ProGen and the reference method for CM and MDH datasets.

    “After measuring the AUC of each model for each dataset, we used bootstrapping to compute the statistical significance of the difference in AUC of fine-tuned ProGen vs. the reference method (bmDCA and ProGen ablations for CM, ProteinGAN and ProGen ablations for MDH).”
    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

    The lysozyme fine-tuning set was clustered at 80% identity, and 10% of clusters were held out for hyperparameter optimization and stopping.

    “The fine-tuning dataset was clustered at 80% sequence identity and 10% of the clusters were held-out as a validation set for hyper-parameter optimization and stopping criteria.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    A threshold of 0.42 norm relative enrichment approximately marks the cut-off between two modes in CM function data.

    “We also applied thresholding at 0.42 norm relative enrichment (r.e.) to obtain binary labels for CM function, which roughly corresponds to the cut-off point between two modes that exist in CM function data, to be used for ROC curves, following the original study7.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [7] Russ, W. P. et al. An evolution-based model for designing chorismate mutase enzymes. Science 369, 440–445 (2020). source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    AlphaFold2 structure prediction was run without an MSA input to avoid biasing predictions toward a known natural mode.

    “We performed structure prediction without an MSA as input so as to not heavily bias the structure prediction toward a known natural mode.”
    Quote in the paperSubstantial claim
40 of 40

  • The paper relies on this without support.

    The five lysozyme families chosen because they lack disulfide bonds can be expressed in high-throughput cell-free systems.

    “Within the lysozyme clan, we chose to work with five families that lack disulfide bonds as to be amenable to high-throughput cell-free expression: Phage lysozyme, Pesticin, Glucosaminidase, Glycoside hydrolase, and Transglycosylase.”

  • The paper relies on this without support.

    Well-expression and solubility in the cell-free system indicate that the artificial proteins are structurally well-folded.

    “These results indicate that ProGen can generate artificial proteins that are structurally well-folded for proper expression as compared to a batch of natural proteins”

  • The paper relies on this without support.

    The 280 million-sequence training set is diverse enough to learn a universal protein representation transferable across families.

    “Trained in this unsupervised manner from a large, varied protein sequence database, ProGen learns a universal, domain-independent representation of proteins that subsumes local and global structure motifs”

  • The paper relies on this without support.

    HEWL is a suitable reference for native lysozyme catalytic activity.

    “Hen egg white lysozyme (HEWL), a naturally-evolved exemplar protein, was measured as positive control”

  • The paper relies on this without support.

    Model likelihoods for GANs are intractable.

    “Since model likelihoods for GANs are intractable, we used discriminator scores corresponding to the probability at which the ProteinGAN discriminator predicted each sample was real as a ProteinGAN model score for each MDH sequence.”

  • The paper relies on this without support.

    The average molecular weight of the fluorescein-labeled M. lysodeikticus cell wall substrate is unknown and the substrate is heterogeneous.

    “(the average molecular weight of the fluorescein-labeled M. Lysodeikticus cell wall substrate was unknown and likely heterogeneous).”
6 of 6

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?

  • Synthetic proteins optimized for specific families of lysozymes using ProGen have exhibited catalytic efficiencies comparable to their naturally occurring counterparts with low sequence identity.

    “synthetic proteins optimized for specific families of lysozymes using ProGen have exhibited catalytic efficiencies comparable to their naturally occurring counterparts with low sequence identity.”

    From Large language models (LLMs): survey, technical frameworks, and future challenges (2024) · cites this paper as Madani et al. 2023

    The paper’s comparison of L056 and L070 (section labeled Abstract; page unavailable) reports catalytic activity comparable to a natural lysozyme and says they differ from their nearest known natural sequences by 53 and 18 amino acids, respectively. But its separate test of proteins with less than 40% identity to any known natural protein found significantly lower activity. The claim does not distinguish those low-identity proteins from the two with comparable activity. (LLM, conf=0.87)

    Experimental — accuracy not yet measured

  • AI models have already been used to design new CRISPR enzymes, functional proteins, and entire prokaryotic genomes.

    “AI models have already been used to design new CRISPR enzymes,53 functional proteins,128 and even entire prokaryotic genomes.51”

    From How to build the virtual cell with artificial intelligence: Priorities and opportunities (2024) · cites this paper as [128]

  • Success of LLMs has revolutionized AI learning in natural language and image processing and has been extended into learning of biological sequence datasets.

    “The success of LLMs has revolutionized AI learning in natural language and image processing [75] and has been extended into the learning of biological sequence datasets [80].”

    From Challenges in AI-driven Biomedical Multimodal Data Fusion and Analysis (2025) · cites this paper as [80]

  • ProGen language model employs deep learning techniques to produce protein sequences with predefined functionalities.

    “the ProGen (Madani et al. 2023) language model employs deep learning techniques to produce protein sequences with predefined functionalities.”

    From Large language models (LLMs): survey, technical frameworks, and future challenges (2024) · cites this paper as Madani et al. 2023

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.

  • Training-data claim conflicts with English-language initialization

    Minor

    The conclusion says the model was trained only with evolutionary sequence data, but the Methods say it was initialized with weights pretrained on English text.

1 of 1

Cites

Flagged first, then by citations

Research works printed but not linked: 14

  1. Vaswani, A. et al. Attention is all you need. In 31st Conference on Neural Information Processing Systems (NIPS, 2017).
  2. Devlin, J., Chang, M.-W., Lee, K. & Toutanova, K. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT, 2019).
  3. Brown, T. B. et al. Language models are few-shot learners. In 34th Conference on Neural Information Processing Systems (NeurIPS, 2020).
  4. Zellers, R. et al. Defending against neural fake news. In 33rd Conference on Neural Information Processing Systems (NeurIPS, 2019).
  5. Radford, A., Narasimhan, K., Salimans, T. & Sutskever, I. Improving language understanding by generative pre-training. Preprint at https://s3-us-west-2.amazonaws.com/openai-assets/research-covers/language-unsupervised/language_understanding_paper.pdf (2018).
  6. Dathathri, S. et al. Plug and play language models: a simple approach to controlled text generation. In 8th International Conference on Learning Representations (ICLR, 2020).
  7. Pettit, L. D. The IUPAC stability constants database. Chem. Int. 28, 14–15 (2006).
  8. Bengio, Y., Ducharme, R., Vincent, P. & Janvin, C. A neural probabilistic language model. J. Mach. Learn. Res. 3, 1137–1155 (2003).
  9. Goyal, K., Dyer, C. & Berg-Kirkpatrick, T. Exposing the implicit energy networks behind masked language models via metropolis–hastings. In 10th International Conference on Learning Representations (ICLR, 2022).
  10. Alley, E., Khimulya, G., Biswas, S., AlQuraishi, M. & Church, G. M. Unified rational protein engineering with sequence-based deep representation learning. Nat. Methods 16, 1315–1322 (2019). · DOI · Paper page
  11. Pascanu, R., Mikolov, T. & Bengio, Y. On the difficulty of training recurrent neural networks. In Proc. 30th International Conference on Machine Learning (eds. Dasgupta, S. & McAllester, D.) 1310–1318 (PMLR, 2013).
  12. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I. & Salakhutdinov, R. Dropout: a simple way to prevent neural networks from overfitting. J. Mach. Learn. Res. 15, 1929–1958 (2014).
  13. Holtzman, A., Buys, J., Du, L., Forbes, M. & Choi, Y. The curious case of neural text degeneration. In 8th International Conference on Learning Representations (ICLR, 2020).
  14. Goodfellow, I. J. et al. Generative adversarial networks. In 28th Conference on Neural Information Processing Systems (NIPS, 2014).

Cited by

Sorted by citations

Authors
Madani, A., Krause, B., Greene, E. R., Subramanian, S., Mohr, B. P., Holton, J. M., Olmos, J. L., Xiong, C., Sun, Z. Z., Socher, R., Fraser, J. S., & Naik, N.
Affiliations
  • Salesforce (United States) (US)
  • University of California, San Francisco (US)
  • Howard Hughes Medical Institute (US)
  • Lawrence Berkeley National Laboratory (US)
Corresponding author
  • Madani, Ali
  • Naik, Nikhil
ORCID
Venue
Nature Biotechnology
Publisher
Springer Science and Business Media LLC
Published
2023-01-26
Volume
41
Issue
8
Pages
1099-1106
ISSN
  • 0733-222X (ISSN-L)
  • 1087-0156
  • 1546-1696
DOI
10.1038/s41587-022-01618-2
PMC
PMC10400306
OpenAlex
W4318071656
Cited by
1,078
References
80
Citations by year
2026: 314 · 2025: 320 · 2024: 302 · 2023: 141 · 2022: 1
FWCI
95.97
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Biotechnology (publishedVersion)
  • PubMed (publishedVersion)
  • PubMed Central (acceptedVersion)
  • eScholarship (California Digital Library) (submittedVersion)
  • OSTI OAI (U.S. Department of Energy Office of Scientific and Technical Information) (submittedVersion)
Funders
U.S. Department of Health & Human Services | National Institutes of Health (GM123159)
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 › Molecular Biology › Enzyme Catalysis and Immobilization
Keywords
large language models, protein sequence generation, protein design, protein families, lysozyme, chorismate mutase, malate dehydrogenase, catalytic efficiency, controllable generation, protein engineering
MeSH terms
Amino Acid Sequence, Chorismate Mutase, Estrogens, Conjugated (USP), Language, 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.

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