CoChat
Sign Up
Cited by 362Protein Structure and DynamicsEnzyme Structure and FunctionGenomics and Phylogenetic Studies

Clustering predicted structures at the scale of the known protein universe

Open access
Barrio-Hernandez, Inigo, Yeo, Jingi, Jänes, Jürgen, Mirdita, Milot, Gilchrist, Cameron L. M., Wein, Tanita, Varadi, Mihaly, Velankar, Sameer +2 more·2023·Nature
DOI 10.1038/s41586-023-06510-wPMC PMC10584675

Published version of a preprint

Open this paper in CoChat — free accountOpen PDF

Abstract

Proteins are key to all cellular processes and their structure is important in understanding their function and evolution. Sequence-based predictions of protein structures have increased in accuracy 1 , and over 214 million predicted structures are available in the AlphaFold database 2 . However, studying protein structures at this scale requires highly efficient methods. Here, we developed a structural-alignment-based clustering algorithm—Foldseek cluster—that can cluster hundreds of millions of structures. Using this method, we have clustered all of the structures in the AlphaFold database, identifying 2.30 million non-singleton structural clusters, of which 31% lack annotations representing probable previously undescribed structures. Clusters without annotation tend to have few representatives covering only 4% of all proteins in the AlphaFold database. Evolutionary analysis suggests that most clusters are ancient in origin but 4% seem to be species specific, representing lower-quality predictions or examples of de novo gene birth. We also show how structural comparisons can be used to predict domain families and their relationships, identifying examples of remote structural similarity. On the basis of these analyses, we identify several examples of human immune-related proteins with putative remote homology in prokaryotic species, illustrating the value of this resource for studying protein function and evolution across the tree of life.

AI snapshot

Written from the paper’s full text.

TL;DR

The authors developed Foldseek cluster, a scalable structural-alignment-based clustering method, and applied it to all 214 million predicted structures in the AlphaFold Protein Structure Database. They identified 2.30 million non-singleton structural clusters, including 31% lacking annotations, and used the resource to explore evolutionary relationships, dark clusters, domain families, and remote homology.

Key findings

  • Clustering 214,684,311 AFDB structures in two steps yielded 52,327,413 sequence clusters, then 18,661,407 Foldseek structural clusters, and finally 2,302,908 non-singleton AFDB clusters after removing fragments; these had on average 13.05 proteins per cluster and average pLDDT 71.64, with clustering taking 129 h on 64 cores.
  • 711,705 clusters (30.9% of AFDB clusters) were “dark” with no known structural or domain-family annotation, but they represented only 4.06% of AFDB proteins; 49% of clusters were at least partially similar to PDB structures, and 38% remained dark after Pfam search.
  • Cluster quality was high: median LDDT 0.77, median TM score 0.71, 66.4% of Pfam-annotated clusters had 100% Pfam consistency, and 97.4% of pairwise cluster members were conserved at the ECOD H-group (homology) level.
  • Taxonomic analysis mapped 23% of non-singleton clusters to the cellular organism level, 16.1% to bacteria, 13.5% to Eukaryota, and 0.5% to Archaea; 3.91% were species-specific, and only 13 human-specific clusters were found, 9 of which had pLDDT <70.

Methods

  • Two-step clustering: MMseqs2 clustered 214 million UniProtKB sequences at 50% sequence identity and 90% overlap, selecting the highest-pLDDT representative per cluster; Foldseek cluster then used the 3Di structural alphabet with Linclust/MMseqs2-style linear-time clustering, E-value <0.01 and 90% structural alignment overlap, followed by removal of UniProt “fragment” sequences.
  • Cluster purity was assessed by structural metrics (LDDT, TM score) and Pfam/EC consistency; homology versus analogy was benchmarked with ECOD and MALISAM; dark clusters were defined by Foldseek search against PDB, MMseqs2 search against Pfam, and UniProt/TrEMBL/SwissProt Pfam or TIGRFAM annotations; lowest common ancestors were assigned with MMseqs2 lca.
  • Domain families were predicted from local Foldseek hits by clustering start/stop positions, linking structurally similar regions, and graph-based clustering; function and pockets were predicted with AutoSite and DeepFRI, and results were made available through a web server.

Limitations

  • The paper states that the 90% alignment overlap requirement may exclude similar structures with significant insertions or unique repeat arrangements, and the strict E-value threshold of 0.01 may miss similarities.
  • The paper states that AFDB does not contain the full extent of metagenomic or viral protein sequences, limiting detection of retroviral proteins, and that domain prediction used only Foldseek cluster representatives and did not exhaustively search other sequence-based domain annotations.
  • The paper cautions that structural similarity should not always be interpreted as evolutionary homology; evident from the text, the main analyses focus on non-singleton clusters after removing fragments, so singleton and fragment structures are not represented in the reported cluster-level findings.

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

Integrity signals

Registry

Crossref · 10.1038/s41586-023-06510-w

Cites

63 of 63 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 20 cited sources checked · 0 supported · 0 flagged

  • Findingnone given

    Foldseek cluster can cluster hundreds of millions of structures.

    “Here, we developed a structural-alignment-based clustering algorithm—Foldseek cluster—that can cluster hundreds of millions of structures.”
    Quote in the paperSubstantial claim
  • Findingown data

    From 711,705 dark clusters, 33,842 clusters with average pLDDT >90 were selected.

    “From the 711,705 clusters without annotations (dark clusters), we selected 33,842 clusters with the highest average AlphaFold2 prediction confidence (that is, average pLDDT >90).”
    Quote in the paperSubstantial claim
  • Findingown data

    4% of clusters seem to be species specific.

    “but 4% seem to be species specific”
    • Also stated in: § Taxonomic analysis of the clusters · plain
      “we found a small fraction (3.91%) of species-specific structural clusters”
    Quote in the paperSubstantial claim
  • Findingown data

    The analyses identify examples of human immune-related proteins with putative remote homology in prokaryotic species.

    “we identify several examples of human immune-related proteins with putative remote homology in prokaryotic species”
    Quote in the paperNot a substantial claim
  • Findingown data

    883,788 (38% of AFDB clusters) dark clusters were identified after annotating remaining representatives to Pfam by MMseqs2.

    “resulting in 883,788 (38% of AFDB clusters) dark clusters (Methods).”
    Quote in the paperSubstantial claim
  • May combine two claims. 32
  • Classified as probably not a substantial claim. 27
  • The quote only partly states this claim. 1

  • May combine two claims.

    Findingown data

    After removing fragments, 2,302,908 non-singleton clusters remained, with on average 13.05 proteins per cluster and average pLDDT 71.64.

    “This identified 2,302,908 non-singleton clusters that have on average 13.05 proteins per cluster with an average pLDDT of 71.64.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    With increasing LDDT from 0.5 to 0.9, the percentage of perfectly consistent clusters increases from 29.2% to 93.8% in Pfam and from 40.3% to 81.3% for EC (level 4).

    “With increasing LDDT from 0.5 to 0.9, there is an increase in the percentage of perfectly consistent clusters—from 29.2% to 93.8% in Pfam, and from 40.3% to 81.3% for EC (level 4), respectively.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The AFDB clusters have a median LDDT of 0.77 and a median TM score of 0.71 across all clusters.

    “Our clusters have a median LDDT of 0.77 and a median TM score of 0.71 across all clusters”
    • Also stated in: § Cluster purity analysis · plain
      “median LDDT of 0.77 and a median template modelling (TM) score of 0.71”
    Quote in the paperSubstantial claim

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

    Findingown data

    The structural clustering took 129 h on 64 cores.

    “It took 129 h on 64 cores to finish the clustering.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The remaining four human-specific clusters contained a herpes virus U54 unit; annexin with limited human homologues in UniRef50; a U2 snRNP-specific A′ protein that appeared to be a fragment but is not labelled as one; and VPS53, a single long coil structure not clustered by Foldseek due to high random chances.

    “The remaining four clusters contained a herpes virus U54 (UniProt: A0A126LB04) unit; annexin (UniProt: A0A4D5RA95) with limited human homologues in UniRef50; a U2 snRNP-specific A′ protein (UniProt: Q9UEN1) that appeared to be a fragment but is not labelled as one; and VPS53 (UniProt: A0A7P0T9Z7), a single long coil structure that was not clustered by Foldseek due to high random chances of observ”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    We found 5,388 non-redundant Pfam annotations for predicted domain families with at least 5 representatives, corresponding to around 29% of the 19,000 known Pfam families.

    “We found 5,388 non-redundant Pfam annotations for predicted domain families with at least 5 representatives, corresponding to around 29% of the 19,000 known Pfam families.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Pfam annotations used as ground truth correctly reflect domain families.

  • May combine two claims.

    Findingown data

    In total, 1,770 pockets in 1,707 structures were identified and 5,324 functional assignments were made.

    “In total, we identified 1,770 pockets in 1,707 structures and made 5,324 functional assignments within these proteins with predicted pockets.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Species-specific clusters tend to have fewer members (twice more likely to have just two members), are more likely to be dark (56% no annotation), and are composed of smaller proteins (median length around 40 amino acids fewer).

    “Compared with other clusters, the species-specific clusters tend to have fewer members (that is, twice more likely to have just two members); they are more likely to be dark, with 56% having no annotation; and composed of smaller proteins, with a median length of around 40 amino acid fewer).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The other non-annotated group (cluster 3) showed a large degree of diversity and exhibited features of both mammalian and bacterial gasdermin.

    “The other non-annotated group (cluster 3) showed a large degree of diversity and exhibited features of both mammalian and bacterial gasdermin.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The AFDB predicted structures are sufficiently accurate for structural clustering and domain family prediction.

  • May combine two claims.

    Findingown data

    The predicted gasdermin structures from all three groups exhibited conservation of a twisted central antiparallel β-sheet and shared placement of connecting helices and strands of gasdermin.

    “The predicted gasdermin structures from all three groups exhibited the structural characteristic conservation of a twisted central antiparallel β-sheet and the shared placement of connecting helices and strands of gasdermin.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The AFDB predicted structures are sufficiently accurate for structural clustering and domain family prediction.

  • May combine two claims.

    Findingown data

    A homologous protein 64% identical to UniProt A0A1C5UEQ5 from a cultured Lachnospiraceae bacterium was predicted with ColabFold and confirmed to have a similar DNA-binding domain structure (TM score 0.97 and 0.56 in relation to UniProt A0A1C5UEQ5 and human AIM, respectively).

    “We predicted the structure of one homologous protein that is 64% identical to UniProt A0A1C5UEQ5 (Extended Data Fig. 8)—which originates from a cultured Lachnospiraceae bacterium that is part of the Culturable Genome Reference30 of the human gut—using ColabFold31 and confirmed that it has a similar structure DNA-binding domain structure (TM score of 0.97 and 0.56 in relation to UniProt A0A1C5UEQ5”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [30] Zou, Y. et al. 1,520 reference genomes from cultivated human gut bacteria enable functional microbiome analyses. Nat. Biotechnol. 37, 179–185 (2019). source

      Not checked in the free run.

    • cites [31] Mirdita, M. et al. ColabFold: making protein folding accessible to all. Nat. Methods 19, 679–682 (2022). source

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    Bacterial structures related to human CD4 like protein B4E1T0 were found, although these can also be annotated by sequence matching to the immunoglobulin-like domain family in InterPro (IPR013783).

    “We also found bacterial structures that are related to the human CD4 like protein B4E1T0 (Extended Data Fig. 7a), although these can also be annotated by sequence matching to the immunoglobulin-like domain family in InterPro (IPR013783).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Out of 13 human-specific clusters, 9 are predicted non-confident with pLDDT <70 and did not contain structural proteins.

    “Out of the 13 human-specific clusters identified, 9 are predicted non-confident with a pLDDT score of less than 70 and did not contain structural proteins.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    UniProt A0A849ZK06 is predicted to be a ribonucleotide-binding protein with an overall structure resembling a protein kinase fold.

    “For example, UniProt A0A849ZK06 (Fig. 2c) is predicted to be a ribonucleotide-binding protein with an overall structure having an organization that resembles a protein kinase fold.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The structure of S0EUL8 resembles members of the structural maintenance of chromosomes family but is missing several characteristic elements.

    “The structure resembles members of the structural maintenance of chromosomes family but it is missing several characteristic elements.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Non-singleton structural clusters were mapped as conserved at cellular organism (23%), bacterial (16.1%), Eukaryota (13.5%) and Archaea (0.5%) levels.

    “we mapped non-singleton structural clusters that appear to be conserved at the cellular organism (23%) (that is, universal to all life), bacterial (16.1%), Eukaryota (13.5%) and Archaea (0.5%) levels.”
    Quote in the paperSubstantial claim

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

    Findingown data

    The distribution of the number of structures found for each predicted domain family annotated with a known Pfam (n = 3,875) or DUF domain (n = 1,513).

    “The distribution of the number of structures found for each predicted domain family annotated with a known Pfam (n = 3,875) or DUF domain (n = 1,513).”
    Quote in the paperNot a substantial claim

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

    Findingown data

    Non-redundant count of Pfam and DUF domain families found in the structure-based predicted families.

    “Non-redundant count of Pfam and DUF domain families found in the structure-based predicted families.”
    Quote in the paperNot a substantial claim

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

    Findingown data

    The distribution of protein region length in predicted domain families stratified by annotations: Pfam domain (n = 1,048,276), DUF (n = 72,798) and not annotated (n = 1,904,498).

    “The distribution of protein region length in the predicted domain families, stratified by their annotations: Pfam domain (n = 1,048,276), DUF (n = 72,798) and not annotated (n = 1,904,498).”
    Quote in the paperNot a substantial claim

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

    Findingnone given

    Terms such as nucleus, chromatin organization and microtubule motor activity are mostly restricted to eukarya.

    “Mostly restricted to eukarya, terms such as nucleus, chromatin organization and microtubule motor activity are included.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    From 1,770 pockets, 579 (32.7%) encompass more than 40% of the total protein sequence, indicating that the predicted structure is not compact.

    “From 1,770 pockets, 579 (32.7%) encompass more than 40% of the total protein sequence, indicating that the predicted structure is not compact.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The high level of consistency in our clustering is mainly attributed to the stringent E-value of 10−2; when raising it to 10, the consistencies decrease to 69.7%, 55.7%, 53.3%, 51.9% and 36.6%, respectively.

    “The high level of consistency in our clustering is mainly attributed to the stringent E-value of 10−2; when raising it to 10, the consistencies decrease to 69.7%, 55.7%, 53.3%, 51.9% and 36.6%, respectively.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The AFDB predicted structures are sufficiently accurate for structural clustering and domain family prediction.

  • May combine two claims.

    Interpretationnone given

    Species-specific clusters represent lower-quality predictions or examples of de novo gene birth.

    “representing lower-quality predictions or examples of de novo gene birth”
    • Also stated in: § Main · hedged
      “could contain examples of de novo gene birth events”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Premisereasoning

    Studying protein structures at the scale of the AlphaFold database requires highly efficient methods.

    “However, studying protein structures at this scale requires highly efficient methods.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Foldseek increases the speed of structural comparisons by four to five orders of magnitude relative to previous approaches while maintaining sensitivity.

    “Foldseek can increase the speed of comparisons of structures by four to five orders of magnitudes relative to previous approaches while maintaining sensitivity15”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [15] van Kempen, M. et al. Fast and accurate protein structure search with Foldseek. Nat. Biotechnol. https://doi.org/10.1038/s41587-023-01773-0 (2023). source

      Not checked in the free run.

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

    Premisenone given

    Comparing all 214 million structures against each other using current methods would take approximately 10 years on a 64-core machine.

    “However, comparing all 214 million structures against each other using current methods would take approximately 10 years on a 64-core machine.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Structural clusters can provide hypotheses as to the evolutionary origin of specific biological processes and illustrate cross-kingdom similarities in immune systems.

    “These results exemplify how the structural clusters can provide hypotheses as to the evolutionary origin of specific biological processes and further illustrate the cross-kingdom similarities in immune systems.”
    Quote in the paperSubstantial claim

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

    Premisecitation

    Gasdermin is activated by proteolytic cleavage, which results in N-terminal activation through lengthening of strands β3, β5, β7 and β8, and oligomerization.

    “Gasdermin is activated by proteolytic cleavage, which results in N-terminal activation through the lengthening of strands β3, β5, β7 and β8, and oligomerization36.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [36] Ruan, J., Xia, S., Liu, X., Lieberman, J. & Wu, H. Cryo-EM structure of the gasdermin A3 membrane pore. Nature 557, 62–67 (2018). source

      Not checked in the free run.

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

    Interpretationown data

    It is possible that the bacterial versions may also have a role in pathogen DNA sensing and response.

    “It is possible that the bacterial versions may also have a role in pathogen DNA sensing and response.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S1 (asserted), S3 (load-bearing).

    Interpretationown data

    Several of these high-confidence structures are probably incorrect predictions.

    “We hypothesize that several of these are probably incorrect predictions.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    The preceding gene in the genome encodes a RecN homologue, giving additional evidence for a role of UniProt S0EUL8 in chromosome maintenance.

    “The preceding gene in the genome encodes a RecN homologue (a member of the structural maintenance of chromosomes family), giving additional evidence for a role of UniProt S0EUL8 in chromosome maintenance.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Residues contributing most to the DeepFRI prediction for A0A849ZK06 are in contact with the top scoring pocket, suggesting a potential nucleotide-binding function for this pocket.

    “The residues contributing the most to the DeepFRI prediction are directly in contact with the top scoring pocket (Fig. 2c), suggesting a potential nucleotide-binding function for this pocket.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: DeepFRI predictions are accurate enough to infer functions of dark clusters.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Study design and conductnone given

    The strict E-value threshold of 0.01 may result in missed similarities.

    “Moreover, our strict E-value threshold of 0.01 may result in missed similarities.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    We defined potential domain boundary positions for each protein sequence by clustering start-stop positions with hierarchical clustering (height parameter 250).

    “We defined potential domain boundary positions for each protein sequence by clustering start–stop positions (hierarchical clustering, height parameter of 250 to establish clusters).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The AFDB predicted structures are sufficiently accurate for structural clustering and domain family prediction.

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    Each predicted domain inside selected communities was annotated using Pfam-A regions mapped to UniProt identifiers (v.35.0), requiring more than 75% overlap.

    “Each predicted domain inside the selected communities was annotated using Pfam-A regions mapped to UniProt identifiers (v.35.0), more than 75% of the Pfam domain has to overlap with the predicted domain.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The AFDB predicted structures are sufficiently accurate for structural clustering and domain family prediction.
    • Rests on an assumption: Pfam annotations used as ground truth correctly reflect domain families.

  • May combine two claims.

    Interpretationown data

    The clustering effectively clusters homologous proteins, demonstrating a nearly exclusive distinction between homologues and analogues.

    “This indicates an effective clustering of homologous proteins, demonstrating a nearly exclusive distinction between homologues and analogues.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    We connected the remaining communities on the basis of structure similarities, allowing connections with a P < 10−3.

    “We connected the remaining communities on the basis of the structure similarities, allowing connections with a P < 10−3.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The AFDB predicted structures are sufficiently accurate for structural clustering and domain family prediction.

  • May combine two claims.

    Interpretationreasoning

    Additional domains are expected to be detected if each structure was searched against a larger set of structures, because domain prediction requires multiple observations on the same structural region.

    “As the domain prediction requires multiple observations on the same structural region, additional domains are expected to be detected if each structure was searched against a larger set of structures.”
    Quote in the paperNot a substantial claim

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

    Study design and conductcitation

    We used DeepFRI to predict enriched GO/EC terms and residue-level saliency weights.

    “To assign putative function and predict catalytic residues, we used DeepFRI49 to predict enriched GO/EC terms and residue-level saliency weights across available GO/EC categories (BP, CC, EC, MF).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [49] Gligorijević, V. et al. Structure-based protein function prediction using graph convolutional networks. Nat. Commun. 12, 3168 (2021). source

      Not checked in the free run.

    • Rests on an assumption: The AFDB predicted structures are sufficiently accurate for structural clustering and domain family prediction.

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    We considered only the representatives of Foldseek clusters when performing the domain prediction.

    “Note that we considered only the representatives of Foldseek clusters when performing the domain prediction.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    We applied graph-based clustering (walktrap, 6 steps), keeping communities with at least 5 members.

    “We applied graph-based clustering (walktrap, 6 steps), keeping communities with at least 5 members.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The AFDB predicted structures are sufficiently accurate for structural clustering and domain family prediction.

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    To eliminate clusters similar to previously known experimental structures, we conducted a search using Foldseek against the PDB with an E-value threshold of 0.1.

    “To eliminate clusters similar to previously known experimental structures, we conducted a search using Foldseek against the PDB (v.2022-10-14) for each cluster representative, with an E-value threshold of 0.1.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The AFDB predicted structures are sufficiently accurate for structural clustering and domain family prediction.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Study design and conductnone given

    The current AFDB does not contain the full extent of protein sequences from metagenomics studies or viral proteins, limiting the potential to detect retroviral proteins.

    “Another limitation is that the current AFDB does not contain the full extent of protein sequences from metagenomics studies or viral proteins, limiting the potential to detect retroviral proteins.”
    Quote in the paperNot a substantial claim

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

    Study design and conductnone given

    We used AutoSite to predict pockets and selected pockets with an AutoSite empirical composite score >60 and mean pocket residue pLDDT >90 for additional analyses.

    “We used AutoSite to predict pockets48, and selected pockets with an AutoSite empirical composite score of >60 and mean pocket residue pLDDT of >90 for additional analyses.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [48] Ravindranath, P. A. & Sanner, M. F. AutoSite: an automated approach for pseudo-ligands prediction-from ligand-binding sites identification to predicting key ligand atoms. Bioinformatics 32, 3142–3149 (2016). source

      Not checked in the free run.

    • Rests on an assumption: The AFDB predicted structures are sufficiently accurate for structural clustering and domain family prediction.

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

    Study design and conductnone given

    The resulting network was trimmed excluding connections with E-value higher than 10−5, predicted domains with more than 350 amino acids, and connected components with less than 5 nodes.

    “The resulting network was then trimmed excluding connections with E-value higher than 10−5, predicted domains with more than 350 amino acids and connected components with less than 5 nodes.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The AFDB predicted structures are sufficiently accurate for structural clustering and domain family prediction.

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    For domain prediction, we filtered out low-scoring Foldseek hits using an E-value of 10−3 as the threshold.

    “First, we filtered out low-scoring Foldseek hits using an E-value of 10−3 as the threshold.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The AFDB predicted structures are sufficiently accurate for structural clustering and domain family prediction.

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

    Study design and conductnone given

    We did not perform exhaustive searches with other sequence-based domain family annotations that could annotate additional clusters with previous knowledge.

    “We did not perform exhaustive searches with other sequence-based domain family annotations that could annotate additional clusters with previous knowledge.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Gasdermin was recently found in bacteria and archaea, in which it is similarly activated by dedicated proteases and defends against phages by pore-mediated cell death.

    “Gasdermin was also recently found in bacteria and archaea, in which it is similarly activated by dedicated proteases and defends against phages by pore-mediated cell death37.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [37] Johnson, A. G. et al. Bacterial gasdermins reveal an ancient mechanism of cell death. Science 375, 221–225 (2022). source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    We removed clusters with members annotated with Pfam or TIGRFAM annotations in UniProt/TrEMBL and SwissProt.

    “Finally, we removed clusters with members annotated with Pfam or TIGRFAM20 annotations in the UniProt/TrEMBL and SwissProt database.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The AFDB predicted structures are sufficiently accurate for structural clustering and domain family prediction.
    • Rests on an assumption: Pfam annotations used as ground truth correctly reflect domain families.

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

    Interpretationnone given

    Some ancient systems have been co-opted for use in the mammalian immune response system.

    “In particular, we focused on several examples relating human immunity to bacterial structures, emphasizing how some ancient systems have been co-opted for use in the mammalian immune response system.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    In the inactive structure of mammalian gasdermin (A, B, D and E), the N terminus forms interfaces with the repressor C-terminal domain mediating autoinhibition, one of these is the primary interface at the α1 helix.

    “In the inactive structure of mammalian gasdermin (A, B, D and E), the N terminus forms interfaces with the repressor C-terminal domain mediating autoinhibition, one of these is the primary interface at the α1 helix35.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [35] Liu, Z. et al. Crystal structures of the full-length murine and human gasdermin D reveal mechanisms of autoinhibition, lipid binding, and oligomerization. Immunity 51, 43–49 (2019). source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Study design and conductown data

    We excluded clusters annotated with Pfam domains by searching cluster representatives against the Pfam database.

    “We then excluded clusters annotated with Pfam domains by searching the cluster representatives using MMseqs2 with parameters -s 7.5 --max-seqs 100000 -e 0.001 against the Pfam database.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The AFDB predicted structures are sufficiently accurate for structural clustering and domain family prediction.
    • Rests on an assumption: Pfam annotations used as ground truth correctly reflect domain families.

  • Probably not a substantial claim: may not pass S1 (asserted), S4 (based). May combine two claims. The quote only partly states this claim.

    Study design and conductown data

    Cluster purity was assessed using average LDDT and TM score per cluster, Pfam consistency, and EC number consistency.

    “To assess cluster purity, we followed a two-step approach. First, we calculated the average LDDT and TM score per cluster to assess the structural similarity.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The AFDB predicted structures are sufficiently accurate for structural clustering and domain family prediction.
53 of 53

  • The paper relies on this without support.

    The AlphaFold Protein Structure Database is representative of the known protein universe.

    “over 214 million predicted structures are available in the AlphaFold database2”
    “The AlphaFold Protein Structure Database (AFDB) is a publicly available data repository of protein structures and their confidence metrics, predicted using the AlphaFold2 AI system1,2.”

  • The paper relies on this without support.

    Lack of annotation in PDB, Pfam or TIGRFAM indicates that a cluster is previously undescribed or novel.

    “of which 31% lack annotations representing probable previously undescribed structures”
    “This resulted in the identification of 711,705 (30.9% of AFDB clusters) dark clusters, probably enriched for novel structures.”

  • The paper relies on this without support.

    DeepFRI predictions are accurate enough to infer functions of dark clusters.

    “To predict potential novel enzymes, we searched each structure for pockets and predicted Gene Ontology (GO) and EC number using DeepFRI, a structure-based function prediction method (Methods).”

  • The paper relies on this without support.

    The human BPI protein has a strong affinity for negatively charged lipopolysaccharides found in Gram-negative bacteria.

    “The human bactericidal permeability-increasing (BPI) protein (B4DKH6) is a key component of the innate immune system and is known to have a strong affinity for negatively charged lipopolysaccharides found in Gram-negative bacteria.”

  • The paper relies on this without support.

    The E. coli YceB protein is a tubular putative lipid-binding protein without a well-characterized function.

    “The E. coli YceB protein is a tubular putative lipid-binding protein without a well-characterized function.”

  • The paper relies on this without support.

    The AFDB predicted structures are sufficiently accurate for structural clustering and domain family prediction.

    “We used Pfam annotations to assess the quality of these predictions (Fig. 4b–g).”

  • The paper relies on this without support.

    Pfam annotations used as ground truth correctly reflect domain families.

    “We used Pfam annotations to assess the quality of these predictions (Fig. 4b–g).”

  • The paper relies on this without support.

    Protein structure is conserved for longer periods of evolutionary time than protein sequences.

    “As protein structure is conserved for longer periods of evolutionary time than protein sequences, we expect that AFDB will empower the identification of remote homology.”

  • The paper relies on this without support.

    The 3Di alphabet representation preserves structural similarity sufficiently for clustering.

    “Foldseek’s 3Di alphabet (Extended Data Fig. 1) was used to represent the structures as one-dimensional sequences.”
9 of 9

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?

  • The clustering process first used MMseqs2 to cluster 214 million UniProtKB protein sequences from AlphaFold DB, trimming to 52 million clusters based on defined sequence criteria, with the protein with peak pLDDT elected as each cluster's representative.

    “The clustering process was two-pronged: the MMseqs2 tool first clustered 214 million UniProtKB protein sequences from AlphaFold DB, trimming it down to 52 million clusters based on defined sequence criteria. A protein with the peak pLDDT score was elected as each cluster's representative.”

    From AlphaFold Protein Structure Database in 2024: providing structure coverage for over 214 million protein sequences (2024) · cites this paper as [22]

  • The Foldseek Cluster method was applied to the AlphaFold DB archive, making derived structure similarity clusters available to the research community.

    “Tackling this challenge, Steinegger et al. recently unveiled the Foldseek Cluster, a state-of-the-art structural-alignment-based clustering algorithm designed for vast datasets (22). They applied this novel method to the AlphaFold DB archive, making the derived structure similarity clusters available to the research community.”

    From AlphaFold Protein Structure Database in 2024: providing structure coverage for over 214 million protein sequences (2024) · cites this paper as [22]

  • A secondary clustering via Foldseek using specific structural delineations resulted in 18.8 million clusters; after dismissing fragment sequences, 2.30 million robust clusters each housing at least a pair of structures were finalised.

    “This set then underwent a secondary clustering via Foldseek, using specific structural delineations, resulting in 18.8 million clusters. After dismissing sequences recognised as fragments, we finalised 2.30 million robust clusters, each housing at least a pair of structures.”

    From AlphaFold Protein Structure Database in 2024: providing structure coverage for over 214 million protein sequences (2024) · cites this paper as [22]

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.

  • Pfam consistency is reported as both 66.5% and 66.4%

    Minor

    The figure caption and accompanying text give different percentages for clusters with complete Pfam consistency, both to one decimal place.

  • Structural-cluster count differs between figure caption and text

    Minor

    The Fig. 1 caption reports 18.8 million Foldseek clusters, but the detailed result is 18,661,407, which rounds to 18.7 million at one decimal place.

  • Frag1-enriched family count differs between figure and text

    Minor

    The Fig. 5 caption says three clusters were enriched for the Frag1 annotation, whereas the text describes one Frag1-enriched family linked to two additional families, one DUF998-enriched and one unannotated.

3 of 3

Cites

Flagged first, then by citations

Cited by

Sorted by citations

Related

Authors
Barrio-Hernandez, I., Yeo, J., Jänes, J., Mirdita, M., Gilchrist, C. L. M., Wein, T., Varadi, M., Velankar, S., Beltrao, P., & Steinegger, M.
Affiliations
  • Seoul National University (KR)
  • ETH Zurich (CH)
  • Weizmann Institute of Science (IL)
  • European Bioinformatics Institute (GB)
Corresponding author
Steinegger, Martin
ORCID
Venue
Nature
Publisher
Springer Science and Business Media LLC
Published
2023-09-13
Volume
622
Issue
7983
Pages
637-645
ISSN
  • 0028-0836 (ISSN-L)
  • 1476-4687
DOI
10.1038/s41586-023-06510-w
PMC
PMC10584675
OpenAlex
W4386708656
Cited by
362
References
63
Citations by year
2026: 97 · 2025: 139 · 2024: 103 · 2023: 23
FWCI
32.26
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature (publishedVersion, cc-by)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by)
  • Research Square (acceptedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
  • Seoul National University Open Repository (Seoul National University) (submittedVersion)
  • Repository for Publications and Research Data (ETH Zurich) (submittedVersion, cc-by)
  • ETH Zürich Research Collection
Funders
  • National Research Foundation (2021-R1C1-C102065)
  • Seoul National University
  • Eidgenössische Technische Hochschule Zürich
  • National Research Foundation of Korea (2021-R1C1-C102065, 2021-M3A9-I4021220, 2020M3‐A9G7‐103933, RS-2023-00250470)
  • ETH Zürich Foundation
  • Helmut Horten Stiftung
  • Samsung
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Protein Structure and Dynamics
  • Physical Sciences › Materials Science › Materials Chemistry › Enzyme Structure and Function
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Genomics and Phylogenetic Studies
Keywords
protein structure clustering, structural alignment, protein evolution, domain families, de novo gene birth, immune-related proteins, prokaryotic proteins, protein function
MeSH terms
Algorithms, Humans, Phylogeny, Prokaryotic Cells, Proteins, Species Specificity, Cluster Analysis, Sequence Alignment, Evolution, Molecular, Databases, Protein, Structural Homology, Protein, Molecular Sequence Annotation

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.

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