henhaobio
Sign in

Nature BiotechnologyBioinformaticsPeer-reviewed journal article

Foldseek: fast and accurate protein structure search

Michel van KempeniD, Stephanie S. KimiD, Charlotte TumescheitiD, Milot MirditaiD, Jeong-Jae LeeiD, Cameron L. M. GilchristiD, Johannes SödingiD, Martin SteineggeriD

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

Abstract

As structure prediction methods are generating millions of publicly available protein structures, searching these databases is becoming a bottleneck. Foldseek aligns the structure of a query protein against a database by describing the amino acid backbone of proteins as sequences over a structural alphabet. Foldseek decreases computation times by four to five orders of magnitude with 86%, 88% and 133% of the sensitivities of DALI, TM-align and CE, respectively.

In-depth summary

Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text of version 5 (28 Mar 2023) · CC BY 4.0

Foldseek's 3Di tertiary-interaction alphabet enables structure search four to five orders of magnitude faster than Dali and TM-align at 86% and 88% of their respective sensitivities.

Structure comparison finds remote homologs that sequence search misses, but TM-align, Dali and CE lack prefilters and optimize non-local scores iteratively, so one TM-align query against 100 M structures takes about a month on one core, far too slow for databases holding over 214 million AlphaFold2 and 617 million ESMFold models. Foldseek recasts structures as 3Di sequences so MMseqs2-style prefiltering and alignment make structure-based homology search tractable.

Key results

  1. ShownOn all-versus-all SCOPe40 single-domain searches, Foldseek gives up some homology-detection sensitivity relative to Dali and TM-align in exchange for runtimes thousands of times shorter, while outperforming CE and earlier structural-alphabet tools.

    Sensitivity is 86%, 88% and 133% of that of Dali, TM-align and CE; by AUC up to the first false positive, Foldseek ranks below Dali and TM-align, above CE, and far above 3D-BLAST and CLE-SW. On 11 211 SCOPe40 domains (64-core AMD EPYC 7702P) it runs over 4,000-fold faster than Dali and TM-align and over 21,000-fold faster than CE. · Abstract; Abstract; Runtime evaluations on SCOPe and AlphaFoldDB

  2. ShownSpeed gains grow with database size: on AlphaFoldDB (v1) Foldseek runs about 184,600 and 23,000 times faster than Dali and TM-align, and an example webserver query returned known homologs as top hits within seconds.

    SARS-CoV-2 RNA-dependent RNA polymerase (RdRp; PDB 6M71_A, 942 residues) vs AlphaFoldDB (v1): Dali 10 days, TM-align 33 h, Foldseek webserver 6 s (reported as 180,000- and 23,000-fold faster); each of the 10 best-ranked hits was a known RdRp homolog. AlphaFoldDB runtimes were measured on the 100 benchmark queries using 64 threads or 64 parallel chunks. · Abstract; Runtime evaluations on SCOPe and AlphaFoldDB

  3. ShownOn full-length, multi-domain AlphaFold2 models, Foldseek's per-residue search sensitivity is on par with Dali, CE and TM-align and its alignments are competitive, though it trails slightly on the mostly single-domain HOMSTRAD reference alignments.

    100 AlphaFoldDB (v1) queries vs other representatives of 34,270 clusters (TP LDDT ≥ 0.6, FP < 0.25): sensitivity near Dali, CE, TM-align; far above CLE-SW, MMseqs2. Top-5 alignments: higher accuracy and sensitivity than MMseqs2, CLE-SW, TM-align; Dali-like accuracy; vs CE 13% less precise, 15% more sensitive. HOMSTRAD (1032 pairs): slightly below CE, Dali, TM-align. · Abstract; Reference-free multi-domain benchmarks; HOMSTRAD alignment benchmark

  4. ShownUsing Foldseek's prefilter to shortlist candidates for a vectorized TM-align reimplementation (Foldseek-TM) gives the best precision-recall performance on SCOPe40, ahead of TM-align alone, which the authors attribute to the prefilter suppressing high-scoring false positives.

    Area under the precision-recall curve: Foldseek-TM first and default Foldseek third at family, superfamily and fold levels; Foldseek-TM and TM-align both rank hits by mean query- and target-normalized TM-score. The TM-align reimplementation (parasail's vectorized Needleman-Wunsch with affine gaps) is 1.7 times faster and cuts memory from 17 to 4 bytes per residue pair. · Abstract; Pairwise global structural alignments using…; Evaluation SCOPe benchmark; Figure legends

  5. SuggestedInspecting high-scoring, low-TM-score Foldseek hits among confident AlphaFoldDB models revealed queries whose structured segments were misoriented relative to each other but individually aligned correctly, suggesting Foldseek, like Dali, can find homologs that global superposition misses.

    Self-search of unfragmented AlphaFoldDB (v1) models with mean pLDDT ≥ 80: 1,675 (of 133 813) high-scoring false positives (TM-score < 0.5 despite a per-column score ≥ 1.0) were inspected; the authors infer that global 3D aligners may overlook non-superposable homologs, while Foldseek, like Dali, does not depend on relative domain orientation. · Abstract; All-vs-all search of AlphaFoldDB with Foldseek

Methods, data and reproducibility

Each residue gets one of 20 3Di states describing its tertiary contact with its nearest neighbor, learned by a vector-quantized variational autoencoder (VQ-VAE) from 10 geometric features; MMseqs2's double-diagonal k-mer and ungapped prefilters then shortlist targets for vectorized Smith-Waterman alignment combining 3Di and amino-acid scores (or TM-align).

  • SCOPe40: SCOPe 2.01 single-domain structures clustered at 40% sequence identity (mean length 174 residues), compared all-versus-all · n = 11 211 domains
  • 3Di training data: residue pairs from TM-align alignments of SCOPe40 domains (TM-score ≥ 0.6, Cα distance ≤ 5 Å), split 80%/20% by fold; four-fold fold-level cross-validation
  • AlphaFoldDB (v1) protein models searched all-versus-all with Foldseek; unfragmented models with mean pLDDT ≥ 80 used to inspect high-scoring false positives · n = 365,198 models
  • Reference-free multi-domain benchmark: AlphaFoldDB (v1) clustered with BLAST and SPICi; random cluster representatives (longest member) queried against the remaining structures · n = 34,270 clusters; 100 queries
  • E-value network training set: structures sampled from AlphaFoldDB (v1) so that biased, disordered and poorly modeled regions are represented · n = 100 000 structures
  • HOMSTRAD expert-curated structural alignments; first-versus-last member pair of each family used as reference (median length 182 residues) · n = 1032 families (one pair each)
  • Case-study query: SARS-CoV-2 RNA-dependent RNA polymerase (RdRp) chain searched against AlphaFoldDB (v1) with the Foldseek webserver, TM-align and Dali · SARS-CoV-2 (virus) · n = 1 chain (942 residues) · PDB: 6M71_A
  • Prebuilt search databases: AlphaFoldDB (v4) UniProt, UniProt50, Proteome and Swiss-Prot sets; 100%-identity-clustered PDB; ESMatlas High-Quality Clu30 · n = 214M, 54M, 564k and 542k structures (AlphaFoldDB v4 sets)
Methods & tools
Foldseek (commit aeb5e) in default 3Di/amino-acid mode and Foldseek-TM mode, MMseqs2 (release 13-45111): prefilter framework and sequence-search baseline, TM-align (TMalign.cpp 2019/08/22; default and -fast), DaliLite.v5, CE (BioJava 5.4.0 implementation), CLE-Smith-Waterman (PDB Tool v4.80 with SSW commit ad452e), 3D-BLAST beta102 with BLAST+ 2.2.26, Geometricus via Caretta-shape (1e3adb0), VQ-VAE trained in PyTorch 1.9.0, integrated via Kerasify, BLAST 2.5.0+ and SPICi for AlphaFoldDB clustering, parasail vectorized Needleman-Wunsch and fast_protein_cluster TM-score code, PULCHRA, gemmi, Foldcomp, HMMER3 Gumbel fitting, mixtools gammamixEM
Software & environment
Foldseek commit aeb5e (GPLv3; SIMDe builds for x86_64, arm64, ppc64le on Linux and macOS); MMseqs2 13-45111; TMalign.cpp 2019/08/22; DaliLite.v5; BioJava 5.4.0; 3D-BLAST beta102 with BLAST+ 2.2.26; BLAST 2.5.0+; PyTorch 1.9.0 (VQ-VAE).
Assay conditions & compute
Runtimes measured on one AMD EPYC 7702P 64-core CPU, 1024 GB RAM (64 threads or 64 parallel query chunks; Geometricus on a single core); webserver keeps databases in RAM, e.g. 191GB for AlphaFoldDB/UniProt50 (v4) and 110GB for ESMatlas High-Quality Clu30.
Reproducibility gaps
SCOPe40 runtimes for TM-align, Dali and CE were estimated from a 10% query subset; the 3Di substitution matrix, alternative LDDT thresholds and cross-validation results sit in supplementary material absent from this text.
Limitations (6)
  • The SCOPe sensitivity benchmark has only single-domain queries and targets, so it cannot probe local similarity such as homologous domains shared by multi-domain proteins; the HOMSTRAD alignment benchmark is similarly limited (median length 182 residues). · stated by the authors
  • The reference-free multi-domain benchmark requires choosing LDDT cut-offs for true (≥ 0.6) and false (< 0.25) positives, set from median SCOPe40 intra- and inter-group LDDT scores; matches in between are ignored. · stated by the authors
  • Baseline gaps: Dali, CE and CLE-Smith-Waterman gave no alignment for 35, 1 and 1 of 1032 HOMSTRAD pairs (scored as zero sensitivity), 3D-BLAST was omitted from the multi-domain benchmark over occasional >10^7 scores for single-residue alignments, and one CE query (>16 days) was excluded. · stated by the authors
  • Geometricus was run on a single core because parallelizing it would need substantial engineering; the authors note an efficient multi-core version might match MMseqs2 speed on 64 cores, so its runtime comparison is not like-for-like. · stated by the authors
  • Speed comes at some accuracy cost: sensitivity stays below Dali and TM-align (86% and 88% of theirs) and HOMSTRAD alignment quality is slightly below CE, Dali and TM-align, so exhaustive aligners may still suit small, accuracy-critical comparisons. · noted in this summary
  • Headline SCOPe40 results use a 3Di alphabet, substitution matrix and score weights all fitted on SCOPe40 (overfitting checked only by a supplementary fold-split cross-validation), and deep-learning structure-search tools were excluded rather than benchmarked. · noted in this summary

Read this if you need homology search across millions of predicted protein structures, or want to see how a learned 3Di alphabet plus MMseqs2-style prefiltering trades modest sensitivity for orders-of-magnitude speedups.

Citations
2,594
FWCI
225.97
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
THISEdgar '26Zerefa '25Edgar '24Wang '26Lin '25Slanináková '26
82% semantic matchBioinformatics
R. C. Edgar·bioRxiv·Jul 30, 2026