foldseek

name	foldseek
description	Foldseek — ultrafast protein structure search and clustering via the 3Di structural alphabet (Steinegger / Söding labs). Use when: (1) Searching MAG or predicted structures against AFDB/PDB/ESM Atlas, (2) Foldseek-Multimer complex search for distant systems (e.g. CRISPR–Cas), (3) Structure clustering at AFDB scale, (4) Sequence→structure search with ProstT5. Upstream: https://github.com/steineggerlab/foldseek. Web: https://search.foldseek.com. Route via homology-search / metagenomics-llm; pairs with mmseqs2 (sequence) and folddisco (motifs).
license	MIT
category	analysis-tools
tags	["Foldseek","3Di","structure-search","AFDB","multimer","ProstT5","homology"]
upstream	https://github.com/steineggerlab/foldseek
stage	function

Foldseek

Upstream: steineggerlab/foldseek · Web: https://search.foldseek.com

Citation

van Kempen, M. et al. Fast and accurate protein structure search with Foldseek. Nat. Biotechnol. 42, 243–246 (2024). https://doi.org/10.1038/s41587-023-01773-0

Kim, W. et al. Rapid and sensitive protein complex alignment with Foldseek-Multimer. Nat. Methods 22, 469–472 (2025). https://doi.org/10.1038/s41592-025-02593-7

Barrio-Hernandez, I. et al. Clustering predicted structures at the scale of the known protein universe. Nature 622, 637–645 (2023). https://doi.org/10.1038/s41586-023-06510-w

See also docs/references.md.

Analytical thinking

Sequence search (mmseqs2 / diamond) fails for many remote homologs. Foldseek turns tertiary geometry into a 3Di letter sequence and reuses MMseqs2-style prefilters — roughly 10⁴–10⁵× faster than TM-align/Dali/CE with comparable sensitivity. Metagenomics use cases: AFDB/ESMAtlas annotation of dark MAG proteins, structural NR clustering, and complex-level mining (easy-multimersearch).

Martin Steinegger “AI and evolution” stack (conservation level → tool):

Level	Prefer
Sequence / single-chain structure	mmseqs2 · foldseek
Multimer / interface	Foldseek-Multimer (easy-multimersearch)
Local structural motif	folddisco
Conserved gene neighborhoods	spacedust (uses Foldseek)

vs	Prefer Foldseek when
mmseqs2 / diamond	Structures or ProstT5 path available; twilight-zone homology
plmsearch / deepblast	Real/predicted structures exist (or AFDB targets)
folddisco	Need whole-chain / complex fold match, not a short motif
TM-align / US-align	Database-scale search (not gold-standard pairwise polish)

Hard rules:

1. Structural hits ≠ experimental function — pin AFDB/PDB release + Foldseek version
2. Low-confidence AF models (pLDDT) inflate false positives — filter inputs
3. Multimer claims need chain pairing / complex TM reporting, not monomer E-values alone
4. Do not merge Foldseek tables with BLAST/MMseqs unlabeled

How to run

# databases (examples)
foldseek databases Alphafold/Proteome afdb tmp
foldseek databases PDB pdb tmp

# monomer structure search
foldseek easy-search query.pdb afdb aln.m8 tmp

# FASTA without structures → ProstT5 3Di
foldseek databases ProstT5 weights tmp
foldseek easy-search QUERY.faa pdb res.m8 tmp --prostt5-model weights

# structural clustering
foldseek easy-cluster structures/ clu tmp -c 0.9

# complexes (Foldseek-Multimer)
foldseek easy-multimersearch query_complex.pdb.gz pdb mm_result tmp

Webserver for interactive queries: https://search.foldseek.com

Key parameters

Setting	Notes
Input	PDB/mmCIF structures, or FASTA + --prostt5-model
Target DBs	AFDB proteomes, PDB100, ESMAtlas clusters — pin release
--alignment-type	3Di/AA vs TM-align-style rescoring when needed
--cluster-search	Clustered DBs (AFDB50): reps-only vs expand members
--gpu	Optional GPU prefilter (Ampere+ preferred)
Multimer report	Use _report for chain pairing, TM, transforms

Decision tree

Homology / structure need?
├─ Close sequence homologs → mmseqs2 / diamond
├─ Dark proteins, have/ predict structures → foldseek easy-search
├─ Sequence-only, no GPU structures → foldseek+ProstT5 or plmsearch / deepblast
├─ Multi-chain complexes / interfaces → foldseek easy-multimersearch
├─ Short active-site / motif geometry → folddisco
├─ Conserved neighborhoods → spacedust
└─ Structure NR catalogue → foldseek easy-cluster

Related skills

homology-search · protein-language-model · prostt5 · esm · foldcomp · folddisco · foldmason · unicore · mmseqs2 · hh-suite · petasearch · spacedust · plmsearch · deepblast · eggnog-mapper · microbial-mining · tool-selection