henhaobio
Sign in

NatureBioinformaticsPeer-reviewed journal article

Clustering predicted structures at the scale of the known protein universe

Inigo Barrio‐HernandeziD, Jingi YeoiD, Jürgen JänesiD, Milot MirditaiD, Cameron L. M. GilchristiD, Tanita WeiniD, Mihály VáradiiD, Sameer S VelankariD, Pedro BeltrãoiD, Martin SteineggeriD

PDFDOIPreprintCodeData
In-depth full-text digest available · Jump to digest ↓

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 with over 214 million predicted structures available in the AlphaFold database (AFDB). 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 structures in AFDB, identifying 2.27M non-singleton structural clusters, of which 31% lack annotations representing likely novel structures. Clusters without annotation tend to have few representatives covering only 4% of all proteins in the AFDB. Evolutionary analysis suggests that most clusters are ancient in origin but 4% seem species specific, representing lower quality predictions or examples of de-novo gene birth. Additionally, we show how structural comparisons can be used to predict domain families and their relationships, identifying examples of remote homology. Based on these analyses we identify several examples of human immune related proteins with remote homology in prokaryotic species which illustrates the value of this resource for studying protein function and evolution across the tree of life. AvailabilityMethods and data are available at cluster.foldseek.com

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (10 Mar 2023) · CC BY 4.0

A Foldseek-based linear-time structural clustering groups 52 million AlphaFold DB representatives into 2.27 million non-singleton clusters; 31% lack known structural or domain annotation.

All-versus-all comparison of the 214 million AlphaFold DB predictions with existing tools was estimated at roughly 10 years on a 64-core machine, and linear-time Linclust clustering had only been used for sequences. Porting Linclust to Foldseek's 3Di structural alphabet makes whole-database structural clustering feasible, enabling a census of unannotated folds, their taxonomic distribution, and structure-based domain family discovery.

Key results

  1. ShownA two-stage pipeline (MMseqs2 sequence pre-clustering, then Linclust-style clustering on Foldseek 3Di strings) structurally clusters the full AlphaFold DB in linear time, and the resulting clusters are largely coherent in structure and Pfam annotation.

    From 214,684,311 structures, MMseqs2 pre-clustering gave 52,327,413 groups; Foldseek clustering (E-value 0.01, 90% overlap) gave 18,661,407 clusters in 129 hours on 64 cores, leaving 2,278,854 non-singleton clusters (mean 13.2 members) after fragment removal. Median LDDT 0.77, median TM-score 0.71; 64% of clusters with at least two Pfam-annotated members fully consistent. · Structure-based clustering of the AlphaFold…; Structural clustering algorithm; Cluster purity analysis

  2. ShownAbout a third of non-singleton clusters lack PDB similarity and Pfam/TIGRFAM annotation, but these dark clusters are small and cover few proteins. DeepFRI on confident dark structures with pockets points to membrane and transporter functions; oversized pockets flag some non-compact, likely erroneous models.

    1,132,440 clusters (50%) resembled PDB entries; after Pfam search 882,608 (39%) stayed dark, and after member Pfam/TIGRFAM filtering 705,936 (31%), covering 3.85% of AFDB proteins. Among 46,997 dark representatives with pLDDT >90, 1,723 pockets were found in 1,660 structures; 559 (32%) spanned >40% of the protein. · Structurally and functionally unknown clusters…; Prediction of putative novel enzymes and small…; Identification of “dark” clusters and the lowest…; Pocket and functional activity predictions for…

  3. ShownLowest-common-ancestor mapping indicates that most structural clusters are ancient. Species-specific clusters are rare, small, shorter and more often dark, and the data do not support newly emerging human-specific structural clusters.

    Non-singleton clusters conserved at cellular-organism, Bacteria, Eukaryota and Archaea levels: 532,478 (23.4%), 365,561 (16%), 307,365 (13.5%), 11,151 (0.5%). 88,828 (3.8%) were species-specific; 81.65% of these have two members vs 35.59% for older clusters, and 56.62% are dark vs 20.31% for universal clusters. Of 12 human-specific clusters, 8 had pLDDT <70. · Evolutionary conservation of the structural…; Evolutionary conservation of human-related…; Identification of “dark” clusters and the lowest…

  4. ShownClustering the boundaries of local Foldseek hits among cluster representatives, then grouping the regions into walktrap communities, yields putative domain families that mostly agree with Pfam, along with many families that have no Pfam match.

    Families with ≥5 representatives: 7,599 match Pfam, 2,032 match Pfam DUFs and 10,722 have no Pfam match; 5,388 non-redundant Pfam annotations recovered, ~29% of the 19,000 known Pfam families; most families carry a homogeneous dominant Pfam label. · Prediction of domain families using structural…; Domain prediction from Foldseek local hits

  5. SuggestedStructure-based links point to cross-kingdom remote homology. Examples include bacterial relatives of the human innate immunity proteins BPI and AIM2, and two unannotated domain families resembling gasdermin, one of them close to bacterial gasdermin.

    Human BPI (B4DKH6) clusters with bacterial A0A2D5ZNG0 (TM-score 0.81, normalized to human length). AIM2 (O14862) clusters with three Lokiarchaeota structures and Clostridium sp. A0A1C5UEQ5 despite no bacterial NR BLAST hits; a ColabFold model of a 64%-identical Lachnospiraceae homolog scored TM 0.97 vs A0A1C5UEQ5 and 0.56 vs human AIM2. · Bacterial remote homology of human immunity…; Structural similarity across distantly related…; Identification of novel gasdermin like domains

Methods, data and reproducibility

Sequences were pre-clustered with MMseqs2 (50% identity, 90% bidirectional overlap). Highest-pLDDT representatives were converted to Foldseek 3Di strings and clustered by Linclust-style k-mer grouping, ungapped and Smith-Waterman structural alignment, and cascaded clustering. Clusters were then annotated against PDB, Pfam and taxonomy.

  • AlphaFold Protein Structure Database (UniProt v3) predictions spanning Swiss-Prot, most of TrEMBL, model organism and WHO priority proteomes · multiple species (across the tree of life) · n = 214,684,311 structures
  • AFDB50: MMseqs2 clusters at 50% identity and 90% overlap, with the highest-pLDDT member as representative · multiple species · n = 52,327,413 clusters
  • Non-singleton AFDB structural clusters retained after removing UniProt fragments · multiple species · n = 2,278,854 clusters
  • Protein Data Bank experimental structures (version 2022-10-14) used to flag known folds; bacterial gasdermin structure for comparison · PDB 7n51
  • Pfam database, Pfam-A regions mapped to UniProt (v35.0), and UniProtKB Pfam/TIGRFAM member annotations
  • High-confidence dark cluster representatives (pLDDT >90) used for pocket and function prediction · multiple species · n = 46,997 structures
  • Human protein-containing clusters analyzed for GO terms and lowest common ancestor · human · n = 12 human-specific clusters (total human clusters not stated)
  • Bacterial AIM2-like protein on a long contig and a ColabFold-modeled Lachnospiraceae homolog · Clostridium sp.; Lachnospiraceae bacterium (human gut) · A0A1C5UEQ5; GenBank FMFM01000010; RJW57900.1
Methods & tools
Foldseek (3Di structural alphabet; structurealign module), MMseqs2 (sequence clustering, lca module, Pfam search), Linclust (adapted to 3Di sequences), AlphaFold2 (source of AlphaFold DB predictions), AutoSite (pocket prediction), DeepFRI (GO/EC function prediction), ColabFold, NCBI BLAST against NR, Pavian (Sankey LCA visualization), Walktrap community detection (6 steps), Hierarchical clustering of hit start/stop positions, Pfam, TIGRFAM and InterPro annotations
Software & environment
Database versions stated: AlphaFold DB UniProt v3, PDB 2022-10-14, Pfam-A regions mapped to UniProt v35.0; parameters stated (k-mer length 10, m=300, E-value 0.01, walktrap 6 steps). Software versions and containers not reported.
Experimental conditions
Foldseek structural clustering of the 52 million representatives took 129 hours on 64 cores.
Reproducibility gaps
Licences are stated (data CC-BY, Foldseek GPLv3), but scripts for dark-cluster filtering, LCA and domain-family prediction are not separately linked; software versions and runtimes for downstream analyses are not given.
Limitations (6)
  • Requiring 90% alignment overlap for cluster membership may exclude true homologs that carry large insertions or distinctive repeat arrangements. · stated by the authors
  • The strict E-value cutoff of 0.01 for structural alignments may leave some homologs unclustered. · stated by the authors
  • AlphaFold DB omits much of metagenomic and viral protein space, which limits detection of, for example, retroviral proteins. · stated by the authors
  • Domains were predicted only from Foldseek cluster representatives, so more likely went undetected, and other sequence-based domain resources were not exhaustively searched, so some putative novel families may be known. Counting partial hits as annotated likely underestimates unexplored structural space. · stated by the authors
  • Functional and evolutionary inferences (DeepFRI functions, pockets, AIM2/BPI/gasdermin homologies) are computational and untested experimentally. Some pLDDT >90 dark structures had pockets covering >40% of the protein, so confidence scores can mask non-compact models. · noted in this summary
  • Summary statistics differ between the main text and the figure legends (18,661,407 vs 18.8 million Foldseek clusters; median LDDT 0.77 and TM-score 0.71 vs average 0.75 and 0.69). A 152,959-structure community was also excluded from the domain analysis. · noted in this summary

Read this if you mine AlphaFold DB or other large predicted-structure collections and need scalable structural clustering, a census of unannotated folds, or structure-based domain family and remote homology discovery.

Citations
362
FWCI
32.16
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISBarone '25Yeo '25Han '26Lau '24Slanináková '26Bayly-Jones '22
76% semantic matchBioinformatics
Federico Barone, Alessandro Laio, Marco Punta+3 more·PRX Life·May 27, 2025· 3 citations