{
 "page": "tools/foldseek.html",
 "title": "Foldseek card",
 "source": {
  "repo": "seanome/2024-kmerseek-analysis",
  "commit": "4a4b11714ddb740f8a7e324058b26bee8fb35af4",
  "commit_date": "2026-10-15T13:12:56Z",
  "path": "results/leaderboard",
  "copied_at": "2026-10-15T13:44:05Z",
  "sha256": {
   "results.parquet": "0c0cf94b5fc40587cab2ffbf7aaf84df9b5475037cc0d818ad2f452a9c73a837",
   "compute.parquet": "8a47ba2123c60cbb41b4833c9f96661934369137b72c5d81ef0808c98247e435",
   "coverage.md": "c10b26730d8dc44b16120bd2f6ee87eb8ed0db3efa40398b339ddac0b0cb7781",
   "provenance.json": "22e54f206354e42208f8a78c4f4827f980ac1db2ce13153f2b077274552e4141",
   "inventory.md": "8dc7040a38cf06af126bd08211edf1f6055c69fe9ac2b4088c6ba8f7f90e6c71",
   "pileup/regions.parquet": "e97701d2fcb601babe25c7d5c6c82760ea19e460e9bc1d555fe5f8f69cb4d6d8",
   "pileup/calls.parquet": "41bc1f6db7bf54f00370b11db2ffed61031d90a42895aa1697be26fc449ea958",
   "pileup/provenance.json": "5ed6e70dd9e82b5200f089b9da3afc62d0a10cc6a51da86715f5f26d3aac9b9e"
  },
  "pileup_commit": "39d81ec90815cacbf3744199d1fa8e99ab28c6f4"
 },
 "text": [
  {
   "tag": "h1",
   "text": "Foldseek"
  },
  {
   "tag": "p",
   "text": "Predicted 3-D shapes."
  },
  {
   "tag": "h2",
   "text": "Will it work for you?"
  },
  {
   "tag": "li",
   "text": "Is the closest same-family domain in the relative you will search at least 40% identical, for most of your proteins? If yes, Fmax here is 40–60% 0.164 (n = 563), 60–100% 0.241 (n = 1,289). If no, 20–30% 0.268 (n = 93), 30–40% 0.167 (n = 165). (Pfam truth set, raw calls, zebrafish (430 Mya) as the relative searched, setting shape letters + amino acids; subsets with at least 30 true instances.)"
  },
  {
   "tag": "li",
   "text": "Do you have a 3D structure prediction for every protein you want to annotate? If not, Foldseek cannot search the proteins without one."
  },
  {
   "tag": "li",
   "text": "Can you spend about 0.129 core-hours per 1,000 proteins for each relative you search? That is the measured mean per relative; see \"Cost per 1,000 queries\" below."
  },
  {
   "tag": "h2",
   "text": "Best Fmax and cost"
  },
  {
   "tag": "p",
   "text": "Under the Pfam truth set, raw calls, zebrafish (430 Mya) as the relative searched, its best setting is shape letters + amino acids (3di_aa): Fmax 0.250 over 2,435 true instances. Searching costs about $0.00574 per 1,000 queries (0.129 core-hours), for each relative searched."
  },
  {
   "tag": "h2",
   "text": "Where it is strong and weak"
  },
  {
   "tag": "p",
   "text": "Each line compares this setting with the best setting of any other tool on the same subset of the proteins, so the other tool's setting can change from line to line. Only subsets with at least 30 true instances are compared."
  },
  {
   "tag": "p",
   "text": "Where it is closest to, or furthest ahead of, the best other tool:"
  },
  {
   "tag": "li",
   "text": "no same-family domain in the relative: Fmax 0.027 (n = 321); best other: HH-suite (hhsearch), single-protein profiles, 0.032."
  },
  {
   "tag": "li",
   "text": "identity 30–40% to the closest same-family domain: Fmax 0.167 (n = 165); best other: MMseqs2, three rounds, sensitivity 7, 3 rounds, 0.209."
  },
  {
   "tag": "p",
   "text": "Where it is furthest behind the best other tool:"
  },
  {
   "tag": "li",
   "text": "regions 16–30 aa long: Fmax 0.004 (n = 271); best other: kmerseek, wwmj5, k = 11, 0.330."
  },
  {
   "tag": "li",
   "text": "regions 31–60 aa long: Fmax 0.034 (n = 420); best other: kmerseek, gbmr4, k = 21, 0.232."
  },
  {
   "tag": "p",
   "text": "Swiss-Prot features, raw calls, zebrafish, setting shape letters + amino acids: best feature type active site, Fmax 0.285 (n = 176); worst membrane-spanning, Fmax 0.000 (n = 1,124)."
  },
  {
   "tag": "figcaption",
   "text": "Figure 2. Where Foldseek is ahead and behind: its Fmax on each subset of the proteins, against the best other tool on that subset (Pfam truth set, raw calls, zebrafish as the relative; its best setting, shape letters + amino acids). Left out, under 30 true instances: 0–20%."
  },
  {
   "tag": "h2",
   "text": "What it compares"
  },
  {
   "tag": "p",
   "text": "Predicted 3-D shapes. Each residue of an AlphaFold model gets one of 20 shape letters (3Di) from its local fold. The shape letters are aligned together with the amino acids."
  },
  {
   "tag": "h2",
   "text": "What it needs"
  },
  {
   "tag": "li",
   "text": "sequence only: no"
  },
  {
   "tag": "li",
   "text": "3D structure prediction: yes"
  },
  {
   "tag": "li",
   "text": "GPU: no"
  },
  {
   "tag": "h2",
   "text": "Reference searched"
  },
  {
   "tag": "p",
   "text": "AlphaFold models of the relative's proteome (Quest for Orthologs reference proteomes, release 2020_04 accessions), as a Foldseek database."
  },
  {
   "tag": "h2",
   "text": "How to read its score"
  },
  {
   "tag": "p",
   "text": "Bit score: higher ranks first. The E-value cutoff was 10."
  },
  {
   "tag": "h2",
   "text": "Known failure modes"
  },
  {
   "tag": "li",
   "text": "A protein with no AlphaFold model cannot be searched. Five relatives have no Foldseek rows at all (chimpanzee, cow, dog, opossum, western clawed frog): not run."
  },
  {
   "tag": "li",
   "text": "AlphaFold models of long proteins come in overlapping pieces (F1, F2, ...). The task script moves piece n's coordinates 200 x (n - 1) residues along the protein; if a piece started elsewhere, its calls would land in the wrong place. Checked against the real piece starts: not measured."
  },
  {
   "tag": "h2",
   "text": "Cost per 1,000 queries"
  },
  {
   "tag": "p",
   "text": "Setting 3di_aa, searching one relative (mean over the 10 relatives it searched): 0.129 core-hours reserved per 1,000 queries (16 cores for the task's wall time), measured on Sherlock. Peak memory 1.5 GB. That is about $0.00574 at $0.044625 per CPU-hour. Building its databases took 0.267 core-hours over 2 databases. A database is built once and reused, so this is not in the per-query cost. This covers the 2 database builds the task records hold; the search used 10 relatives."
  },
  {
   "tag": "p",
   "text": "The per-query cost counts the search only. The run recorded cost per tool and setting, summed over all relatives, so the cost for one relative is the mean."
  },
  {
   "tag": "h2",
   "text": "Licence"
  },
  {
   "tag": "li",
   "text": "Licence: GPL-3.0 (source)"
  },
  {
   "tag": "li",
   "text": "Commercial use allowed: Yes. If you distribute a changed version, you must release its source under GPL-3.0 too."
  },
  {
   "tag": "h2",
   "text": "Settings in the table"
  },
  {
   "tag": "li",
   "text": "shape letters + amino acids (3di_aa): 3Di shape letters and amino acids together"
  },
  {
   "tag": "h2",
   "text": "Version, container, command"
  },
  {
   "tag": "li",
   "text": "Version: Foldseek 10.941cd33"
  },
  {
   "tag": "li",
   "text": "Container: quay.io/biocontainers/foldseek:10.941cd33--h5021889_1 (sha256:1156a052f31b2afb85257c02e83a962f559c9752273fe1064ab735f90ac29d1a)"
  },
  {
   "tag": "pre",
   "text": "foldseek search \\ human-38fdd87d_fsdb/db zebrafish_fsdb/db result foldseek_tmp \\ --threads 16 \\ -e 10.0 \\ --max-seqs 1000"
  },
  {
   "tag": "p",
   "text": "Followed by foldseek convertalis to write the table of hits."
  },
  {
   "tag": "p",
   "text": "Tool id foldseek. Numbers from seanome/2024-kmerseek-analysis commit 4a4b1171."
  },
  {
   "tag": "p",
   "text": "All tool cards · this card as Markdown · as JSON"
  }
 ],
 "data": {
  "tool": "foldseek",
  "name": "Foldseek",
  "kind": "compares 3-D shape letters (3Di)",
  "compares": "Predicted 3-D shapes. Each residue of an AlphaFold model gets one of 20 shape letters (3Di) from its local fold. The shape letters are aligned together with the amino acids.",
  "needs": {
   "sequence only": false,
   "3D structure prediction": true,
   "GPU": false
  },
  "reference": "AlphaFold models of the relative's proteome (Quest for Orthologs reference proteomes, release 2020_04 accessions), as a Foldseek database.",
  "score": "Bit score: higher ranks first. The E-value cutoff was 10.",
  "failure_modes": [
   "A protein with no AlphaFold model cannot be searched. Five relatives have no Foldseek rows at all (chimpanzee, cow, dog, opossum, western clawed frog): not run.",
   "AlphaFold models of long proteins come in overlapping pieces (F1, F2, ...). The task script moves piece n's coordinates 200 x (n - 1) residues along the protein; if a piece started elsewhere, its calls would land in the wrong place. Checked against the real piece starts: not measured."
  ],
  "version": "Foldseek 10.941cd33",
  "licence": "GPL-3.0",
  "commercial_use": "Yes. If you distribute a changed version, you must release its source under GPL-3.0 too.",
  "licence_url": "https://github.com/steineggerlab/foldseek",
  "container": "quay.io/biocontainers/foldseek:10.941cd33--h5021889_1 (sha256:1156a052f31b2afb85257c02e83a962f559c9752273fe1064ab735f90ac29d1a)",
  "command": "foldseek search \\\n    human-38fdd87d_fsdb/db zebrafish_fsdb/db result foldseek_tmp \\\n    --threads 16 \\\n    -e 10.0 \\\n    --max-seqs 1000",
  "numbers_from": {
   "repo": "seanome/2024-kmerseek-analysis",
   "commit": "4a4b11714ddb740f8a7e324058b26bee8fb35af4",
   "filters": "pfam, raw, zebrafish"
  },
  "best_setting": "3di_aa",
  "strong": [
   {
    "axis": "identity",
    "stratum": "no_homolog",
    "fmax": 0.027082308589475693,
    "n": 321,
    "other_tool": "hhblits",
    "other_setting": "single_seq",
    "other_fmax": 0.03236195478661682,
    "margin": -0.005279646197141125
   },
   {
    "axis": "identity",
    "stratum": "30-40%",
    "fmax": 0.16701190421314652,
    "n": 165,
    "other_tool": "mmseqs2_iterative",
    "other_setting": "s7",
    "other_fmax": 0.20886402027992682,
    "margin": -0.0418521160667803
   }
  ],
  "weak": [
   {
    "axis": "feature_length_bin",
    "stratum": "16-30",
    "fmax": 0.004474321206119519,
    "n": 271,
    "other_tool": "kmerseek",
    "other_setting": "wwmj5_k11_lcFalse",
    "other_fmax": 0.3297506857792661,
    "margin": -0.32527636457314657
   },
   {
    "axis": "feature_length_bin",
    "stratum": "31-60",
    "fmax": 0.034179736822508223,
    "n": 420,
    "other_tool": "kmerseek",
    "other_setting": "gbmr4_k21_lcFalse",
    "other_fmax": 0.23209537155652643,
    "margin": -0.19791563473401821
   }
  ],
  "cost_per_1000_queries": {
   "core_h_reserved_per_relative": 0.1287018481407259,
   "n_relatives": 10,
   "database_build_core_h": 0.26666666666666666,
   "database_build_gpu_h": null,
   "peak_rss_gb": 1.5,
   "hardware": "Stanford Sherlock, SLURM partition hns"
  },
  "questions": [
   "Is the closest same-family domain in the relative you will search at least 40% identical, for most of your proteins? If yes, Fmax here is 40–60% 0.164 (n = 563), 60–100% 0.241 (n = 1,289). If no, 20–30% 0.268 (n = 93), 30–40% 0.167 (n = 165). (Pfam truth set, raw calls, zebrafish (430 Mya) as the relative searched, setting shape letters + amino acids; subsets with at least 30 true instances.)",
   "Do you have a 3D structure prediction for every protein you want to annotate? If not, Foldseek cannot search the proteins without one.",
   "Can you spend about 0.129 core-hours per 1,000 proteins for each relative you search? That is the measured mean per relative; see \"Cost per 1,000 queries\" below."
  ],
  "markdown": "foldseek.md"
 }
}