lhallee's picture
protein_annotation_compendium_v1: the compendium of UniProtKB 2026_03
c0863b7 verified
|
Raw History Blame Contribute Delete
30.6 kB
metadata
pretty_name: protein annotation compendium v1
license: other
task_categories:
  - other
tags:
  - protein
  - biology
  - multilabel
  - synthyra
size_categories:
  - 1M<n<10M
configs:
  - config_name: whole_vocabulary
    default: true
    data_files:
      - split: train
        path: train.parquet
      - split: validation
        path: validation.parquet
      - split: test
        path: test.parquet
      - split: calibration
        path: calibration.parquet
  - config_name: go_bp
    data_files:
      - split: train
        path: go_bp/train.parquet
      - split: validation
        path: go_bp/validation.parquet
      - split: test
        path: go_bp/test.parquet
      - split: calibration
        path: go_bp/calibration.parquet
  - config_name: go_mf
    data_files:
      - split: train
        path: go_mf/train.parquet
      - split: validation
        path: go_mf/validation.parquet
      - split: test
        path: go_mf/test.parquet
      - split: calibration
        path: go_mf/calibration.parquet
  - config_name: go_cc
    data_files:
      - split: train
        path: go_cc/train.parquet
      - split: validation
        path: go_cc/validation.parquet
      - split: test
        path: go_cc/test.parquet
      - split: calibration
        path: go_cc/calibration.parquet
  - config_name: localization
    data_files:
      - split: train
        path: localization/train.parquet
      - split: validation
        path: localization/validation.parquet
      - split: test
        path: localization/test.parquet
      - split: calibration
        path: localization/calibration.parquet
  - config_name: metal_ions
    data_files:
      - split: train
        path: metal_ions/train.parquet
      - split: validation
        path: metal_ions/validation.parquet
      - split: test
        path: metal_ions/test.parquet
      - split: calibration
        path: metal_ions/calibration.parquet
  - config_name: modifications
    data_files:
      - split: train
        path: modifications/train.parquet
      - split: validation
        path: modifications/validation.parquet
      - split: test
        path: modifications/test.parquet
      - split: calibration
        path: modifications/calibration.parquet
  - config_name: ec
    data_files:
      - split: train
        path: ec/train.parquet
      - split: validation
        path: ec/validation.parquet
      - split: test
        path: ec/test.parquet
      - split: calibration
        path: ec/calibration.parquet
  - config_name: homodimer
    data_files:
      - split: train
        path: homodimer/train.parquet
      - split: validation
        path: homodimer/validation.parquet
      - split: test
        path: homodimer/test.parquet
      - split: calibration
        path: homodimer/calibration.parquet
  - config_name: measurements
    data_files:
      - split: train
        path: measurements/train.parquet
      - split: validation
        path: measurements/validation.parquet
      - split: test
        path: measurements/test.parquet
      - split: calibration
        path: measurements/calibration.parquet

protein annotation compendium v1

UniProtKB 2026_03's Swiss-Prot and Swiss-Prot-quality TrEMBL entries, with every annotation their sources state: GO, EC, Rhea, cofactors, locations, topology, keywords, located features, disordered regions, modifications, InterPro and Pfam families, Reactome pathways, lineage and organism traits, each with evidence code and tier. The default config's row is a distinct sequence with its tokens, terms and tiers; GO, localization, metal ion, modification, EC, homodimer, pH, melting temperature, kcat and KM views, residue-level modifications, disorder and features, lineage and every assertion ship beside it. Split 85/5/8/2 by cluster at 0.5 identity (1,423,992 train, 18,680 validation, 31,662 test, 6,830 calibration), none at 0.5 identity or more across splits; another split holds out whole species. The default config holds the 1,481,135 with a vocabulary term.

Task prediction of a protein's annotation terms from its sequence over one vocabulary in eight fundamental aspects (Oracle One, the translator), retrieval between a sequence and its annotation (CAMP), per-aspect views for GO, localization and EC, which metal ions a protein binds and which modifications it carries, which residues carry them and each other located feature, whether it is a homodimer, the optimum pH, the melting temperature, kcat and KM as regression views, and its lineage as chained choices and its organism's traits, with a split that holds whole species out
Rows 1,481,135 across four splits
Homology enforced at 0.5 identity, measured at 0.3
Acceptance 7 of 7 acceptance checks passed, among them 12 directed homology searches with none at 0.5 identity or more
Status private

Homology leakage

Enforced: 0.5 identity. No held-out sequence has a relative at 0.5 identity or more in another split: every cluster at that level, and so every finer one, sits in exactly one split. 12 directed searches between the splits found no pair above 0.499 identity. Measured, not enforced: 0.3 identity. 18,767 of the 57,172 held-out proteins (33%) have no relative at 0.3 identity or more in train; the rest have one between 0.3 and 0.5.

Clustered at 0.9, 0.7, 0.5, 0.3 identity with 0.8 coverage of both sequences (MMseqs2).

Identity Guarantee
0.9 implied: no cluster at this level spans two splits
0.7 implied: no cluster at this level spans two splits
0.5 enforced: no cluster at this level spans two splits
0.3 measured, not enforced: 18,767 of 57,172 held-out proteins are below this identity to train
Split Rows Clusters (0.9 identity)
train 1,423,964 798,403
validation 18,680 12,446
test 31,661 21,515
calibration 6,830 5,167
Why this level

CAMP, which reads this split, may enforce 0.5 at most, and no level up to it is well formed, so the survey of the clusters recommends 0.5, the finest such level: there 886 of the 6,066 labels spread over 20 or more units at 0.9 identity fall to fewer (14.6 percent, over the survey's 10), as their proteins sit in a few large families; 0.7 is the lowest well-formed level. Repair returned to train every held-out cluster with a relative in another split at 0.5 identity, since the clusters join 0.8 of both sequences and the audit counts half of one.

How it was built

  1. Read UniProtKB release 2026_03, every element with its own evidence codes: the 575,748 Swiss-Prot entries and 1,083,685 TrEMBL entries (1,033,251 of annotation score 4 or 5 or protein existence 1, streamed pinned to the release, and 50,434 that an experimental GOA line names), 42,063,735 statements; then GOA's 694,440,912 lines, of which the pool's accessions hold 28,164,146; InterPro's 904,908,849 match lines, which gave 39,403,909 protein-level and located InterPro and Pfam rows; Reactome's 223,594 pathway assignments; GlyGen's 98,092 glycosylation-site rows of GlyConnect, the wwPDB and its own literature curation and the O-GlcNAcAtlas's 46,517 unambiguous sites, of which 125,650 place a modification UniProt names on the residue their entry's sequence holds (3,402 on the serine or threonine at a position the Atlas states without its residue), 99,916 residue-level and protein-level assertions once each; DisProt's 13,983 regions, of which 8,324 of its IDPO terms with a paper and a mapped evidence code stand on the sequence they annotate (its GO terms reach the build through GOA); the accessions' lineages, 13,158,683 taxonomy terms; Madin et al.'s 164,832 trait rows of 26 sources (KEGG's 4,670 and ProTraits' 2,818 left out), which give 17,289 species their traits and the accessions 2,921,314 inherited trait terms; and the Meltome Atlas workbook's 40,209 rows, fetching from the release, for their sequence and organism alone, the 3,704 entries of its accessions that no read entry holds.
  2. Gave every statement the tier its evidence code earns: experimental (ECO:0000269 and its family, ECO:0007744, the GO experimental and high-throughput codes), signature (InterPro and Pfam matches), curated (ECO:0000250, 0000255 and 0000305; GO ISS, ISO, ISA, ISM, IC and TAS) or guard (every other code, kept but never a label or a negative); a keyword takes the tier of what implies it, an oligomeric state the evidence of the sentence of its subunit comment that states it (that sentence's inline citations, else what no sentence cites), and 1,799,118 keywords of the technical, coding-sequence-diversity, disease and developmental-stage categories were dropped. Then pooled distinct sequences: dropped 30,918 fragments and 4,946 sequences over 2,046 residues, merged 142,405 accessions into the entry of an identical sequence, and kept the statements of the pooled accessions, 1,481,164 sequences with 121,750,559 assertions (17,932,059 experimental, 2,911,749 of them inherited traits, 38,690,305 signature, 9,538,625 curated, 55,544,461 guard and 45,109 of codes that assert nothing).
  3. Clustered with MMseqs2 at 0.95, 0.9 and 0.8 identity (linclust) and 0.7, 0.6, 0.5, 0.4 and 0.3 (sensitivity 4, 5, 6, 7 and 7.5), 0.8 coverage of both sequences, each level clustering the representatives of the one before: 1,098,016 clusters at 0.95, 837,556 at 0.9, 590,815 at 0.8, 390,350 at 0.7, 302,677 at 0.6, 237,715 at 0.5, 190,909 at 0.4 and 158,689 at 0.3; then searched every sequence against every other once (MMseqs2 at sensitivity 5.7 and 300 candidates, e-value 1e-3 scaled to each split's residues, half of the query aligned) and split by cluster 85/5/8/2 (seed 42), holding out only clusters with an experimental assertion outside the organism aspect, balanced over the 11,982 experimental terms of 10 or more proteins, enforcing 0.5 identity; repair returned 162,452 clusters with a relative in another split to train, after which the twelve directed comparisons between the splits found no pair at 0.5 identity or more (the highest is 0.499), and the worst label-balance excess (a held-out split's distance from train's label shares over a random sample's of its size) is 2.11, on test, within the 2.5 the release states: the splits convention's allowance at any share, where test's own at 8 percent is 2.6.
  4. Kept in the vocabulary the 126,254 terms that 10 or more train proteins carry at a label tier, in eight fundamental aspects; wrote the views: whole_vocabulary (1,481,135 rows over 126,254 tokens), GO BP (126,251 rows over 12,789 terms), GO MF (106,856 rows over 3,891 terms), GO CC (149,474 rows over 1,929 terms), localization (343,874 rows over 2,309 terms; a set view's terms are the vocabulary's and, here 3,739, 603, 265 and 294, the ancestors 10 or more train proteins reach), metal ions (83,849 rows over 21 of GO's terms below metal ion binding, read at each statement's tier from the metal binding sites of 293,878 proteins, the metal cofactors of 363,674 and the metal GO terms of 316,293, a metal being an ion below ChEBI's metal cation; 153,591 statements name no metal, or one GO has no term for), modifications (104,707 rows over 44 of keywlist's 47 PTM keywords, each at the tier of the sites that imply it, a cross-link to ubiquitin or a ubiquitin-like protein implying Ubl conjugation), modification sites (tables/modification_sites.parquet, 80,016 rows over 176 site classes, each a modification ptmlist lets a residue carry anywhere on the chain, or a cross-link to a ubiquitin-like protein, whose train sites sit on one residue, the other residues of that kind in a protein with a site of it its negatives: on test 3,050 sites among 69,740 such residues, 4.4%), disordered residues (tables/disordered_regions.parquet, 3,115 proteins with a disordered residue at a scored tier, from DisProt's curated regions and UniProt's, MobiDB-lite's predictions masked and a paper's ordered residues marked, a protein's other residues its negatives as CAID's DisProt reference counts them: on test 14,092 disordered among 58,236 scored residues, 24.2%; its DisProt-PDB reference keeps as negatives only the 33,365 of them a PDB structure observes (SIFTS; 119 of the 138 test proteins have one), so 29.7% of its 47,457 scored residues are disordered), located features (tables/located_features.parquet, 337,790 proteins with a residue at a scored tier of one of 3,390 classes, the binding and active sites, transmembrane, signal and transit spans, disulfide bonds, motifs, repeats, domains and regions of UniProt's entries, each feature at its tier, an aspect's type term (Active site, Transmembrane) holding its described classes' residues too, and a protein's other residues the negatives of a class it holds: on test 2,098 pairs of a protein and a class, led by binding (675), signal peptide (271) and disulfide bond (227), 8.3% of whose 790,963 scored residues are positive), lineage (tables/lineage.parquet, every protein's taxon at the eight ranks where all its organisms agree, 1,450,150 proteins to the species, 30,898 to a rank above and 116 to none, with an organism split beside the global one: 85/5/8/2 of the 29,969 species, balanced over their lineage, the 136 of more than 0.1% of the proteins in train, so test holds 2,398 species with 16,059 proteins; 2,736 sequences whose species fall in different splits are in none), homodimer (67,564 proteins, 47,063 homodimers and 20,501 monomers by the state their subunit comments give at their best tier, 19 of them stating the other at a lesser one, and 374 stating both at their best left out), pH optimum (3,096 proteins with 3,098 observations: the one point optimum each of 5,416 Swiss-Prot pH-dependence texts may state from a cited experiment, read as ph_optimum_experimental_v2 reads it, and a protein's mean), melting temperature (23,974 proteins with 24,107 of the Meltome Atlas workbook's 40,209 rows, read as thermal_tm_experimental_v1 reads them, and a protein's mean, 356 of the points found through an accession UniProt has since merged into a read entry and 738 through another entry of a pooled sequence, fetched from the release; 8,210 rows state no melting point and 1 no UniProtKB accession, 1,107 name an isoform, 3,104 an accession the release no longer holds and 6 one it has split, 2,685 an entry whose sequence the pool lacks, 113 another species' entry, 481 a fragment and 395 a sequence over the pool's length limit), kcat and KM (2,727 proteins with 7,964 turnover numbers and 9,298 with 29,836 Michaelis constants: the distinct values, in inverse seconds and molar, that Swiss-Prot's kinetics statements give the entry's own sequence from a cited experiment, each with its substrate, pH and temperature as stated, and a protein's mean log10, kcat_experimental_v1's target; 22,315 of the 39,228 constants read have a substrate ChEBI names, by UniProt's name for an entity or else the entity's own; of the values stated 177 name a mutant, an isoform or another sequence, 8 are bounds, 147 are paired with their substrates by "respectively", 111 are a KM in another unit than molar and 46 do not read as a positive number with its unit, none of them read; the pH, melting temperature, kcat and KM views are the measurements config's columns) and EC (175,460 enzymes with one EC number, 11,965 with several left out, and 23,884 confident non-enzymes; 27,040 rows matched in lineage and length), a non-enzyme being a pooled protein with a reviewed entry that breaks none of ec_experimental_v3's twelve rules and no sign of catalysis on any accession at any tier, matched to the enzymes in each split cell by cell (the order and 8 length bins, one of a class from a 0.5-identity cluster in a cell, 2 non-enzymes in 5); and scored the train prior and identity transfer from the split's own table on every view's validation and test rows, keeping each held-out protein's 50 nearest train proteins.

How protein_annotation_compendium_v1 was built

Row counts at every step
Step Rows in Rows out
the nine line sources 1,657,926,551 1,657,926,551
line sources: the pool's accessions' rows (goa_other_database removes 34,180,904; goa_isoform_or_form removes 10,674; goa_accession_outside_pool removes 632,085,188; interpro_other_protein removes 885,846,109; reactome_isoform removes 1,720; reactome_other_protein removes 99,504) 1,657,926,551 105,702,452
InterPro: protein-level and Pfam rows 105,702,452 126,043,621
terms, hierarchy and keyword tiers 126,043,621 126,043,621
pooled accessions, kept keywords, GO terms (swissprot_accession_not_in_pool removes 418,791; swissprot_dropped_keyword_category removes 504,046; trembl_accession_not_in_pool removes 295,497; trembl_dropped_keyword_category removes 1,295,072; goa_accession_not_in_pool removes 752,080; go_unresolved_or_other_aspect removes 3,261; interpro_accession_not_in_pool removes 713,604; reactome_accession_not_in_pool removes 4,213; taxonomy_accession_not_in_pool removes 284,946; glygen_accession_not_in_pool removes 11,611; disprot_accession_not_in_pool removes 376; traits_accession_not_in_pool removes 9,565) 126,043,621 121,750,559
union of assertions per sequence 121,750,559 1,481,164
Meltome melting points, by accession 1,481,164 1,481,164
85/5/8/2 by cluster, cascade 0.95 to 0.3 1,481,164 see splits
vocabulary, views, audit and baselines 1,481,164 1,481,164

Every source and filter is also listed as text in pipeline.md.

What the compendium holds, by aspect, tier and split

What the compendium holds, by aspect, tier and split. A, each fundamental aspect's protein-term pairs by their best evidence tier, with the total: domain and family 27.4M, 82 percent signatures; organism 14.2M, all experimental, 11.8M lineage terms and 2.4M traits a species' proteins inherit; the other six 1.9M to 9.7M, of which 0.8 to 8.7 percent are experimental and most are guards. B, vocabulary terms per aspect (10 or more train proteins at a label tier), 126,254 in all, from 77,406 domain and family terms to 38 biophysical ones. C, proteins per split from Swiss-Prot and TrEMBL: train is 69 percent TrEMBL and the held-out splits 62 to 74 percent Swiss-Prot, since only clusters with an experimental assertion are held out. D, each held-out split's highest identity to a train protein in the split's table, cumulative: about a quarter have no hit, the median is 0.36, and none reaches the enforced 0.5 (dashed). E, the tier mix of each split's positive assertions, the organism's inherited terms among them: 14 percent experimental in train, 22 to 27 percent in the held-out splits. F, held-out AUROC of a logistic regression on metadata alone for whether a protein carries a label in each aspect: lineage and length (circles) from 0.59 to 0.87, the whole shortcut registry (squares) from 0.84 to 0.93, against chance at 0.5 and the gate's 0.55, which binds only a constructed class (the EC view's non-enzymes, audited on their own rows); organism, which every protein holds, is left out.

Columns

The first columns are the ones to read. The ids and cluster columns after them are for joining and audit.

whole_vocabulary

Column What it holds
name The protein's first Swiss-Prot accession, else its first TrEMBL one, as UniProtKB:P77619.
seqs The protein sequence: a distinct sequence of at most 2,046 residues, never a fragment.
tokens The vocabulary tokens of the terms the protein carries at a label tier, ascending; tables/tokens.parquet maps each to its term.
terms The term of each token, in the same order, such as av::go_bp::GO::GO:0006915.
tiers The best evidence tier of each term, in the same order: experimental, signature, or curated (in train and calibration only).
masked Terms the protein holds only at a tier that neither scores nor counts as a negative here: a guard anywhere, a curated term in validation and test.
observed_units The units the protein has a positive in, each an aspect within its fundamental aspect, such as go_mf/enzyme (GO MF's catalytic terms) or keyword/ptm_processing; the unit's other unmasked terms are its negatives, and a term of another unit is unknown for it, never a negative.
protein_id SHA-256 of the sequence. Protify's pinned format requires it; it is an id, not something to read.
cluster_id The protein's finest cluster (0.9 identity). Every member of a cluster sits in one split.

go_bp, go_mf, go_cc, localization, metal_ions and modifications

Also holds, with the same meaning as above: name, seqs, masked, protein_id, cluster_id.

Column What it holds
positives The view's terms the protein carries at a label tier, with every ancestor along is_a and part_of that is a view term: a vocabulary term, or an ancestor 10 or more train proteins reach (tables/view_terms.parquet).
negated View terms an experimental NOT assertion denies for the protein: explicit negatives.

ec

Also holds, with the same meaning as above: name, seqs, protein_id, cluster_id.

Column What it holds
level_1 The protein's one EC number to its level 1, the rest dashed, such as 3.-.-.-, or non_enzyme: a reviewed entry with a function from a cited experiment and no sign of catalysis on any accession of its sequence (ec_experimental_v3's definition).
level_2 The protein's one EC number to its level 2, the rest dashed, such as 3.4.-.-; null for a non-enzyme.
level_3 The protein's one EC number to its level 3, the rest dashed, such as 3.4.21.-; null for a non-enzyme.
level_4 The protein's one EC number to its level 4, the rest dashed, such as 3.4.21.4; null for a non-enzyme.
matched Whether the row is in the set matched in lineage and length, 3 enzymes to 2 non-enzymes, that asks whether a protein is an enzyme.

homodimer

Also holds, with the same meaning as above: name, seqs, protein_id, cluster_id.

Column What it holds
homodimer Whether the protein's subunit comments state it a homodimer (true) or a monomer (false) at its best tier, each state with its own sentence's evidence; a protein stating both at that tier is left out.
tier The best evidence tier of the state: experimental, or curated in train and calibration only.

measurements

Also holds, with the same meaning as above: name, seqs, protein_id, cluster_id.

Column What it holds
ph_optimum The mean of the protein's distinct observed optimum pH values; null when unmeasured.
ph_observations Each distinct point optimum a Swiss-Prot pH-dependence text states with its own experimental evidence and a cited paper: the pH, the entry, the text verbatim and the citations (ph_optimum_experimental_v2's reading).
melting_temperature The mean of the protein's Meltome Atlas melting points, in degrees Celsius; null when unmeasured.
tm_observations Each melting point thermal proteome profiling of a lysate measured: the value, the accession the workbook names, the lysate's species, its optimal growth temperature, the workbook row and how the accession found this sequence: its entry's own, a secondary accession of the entry, or another entry of the same sequence (thermal_tm_experimental_v1's reading).
log10_kcat The mean log10 of the protein's distinct turnover numbers in inverse seconds, over all substrates and conditions (kcat_experimental_v1's target); null when unmeasured.
kcat_observations Each distinct kcat a Swiss-Prot kinetics statement gives the entry's own sequence (no mutant, variant, isoform, allele, truncation or chimera named) with experimental evidence and a cited paper: the value in inverse seconds, the substrate as stated and the ChEBI entity UniProt's name for it, else the entity's own, names, the stated pH and temperature, the entry, the statement verbatim and its citations.
log10_km The mean log10 of the protein's distinct Michaelis constants in molar, over all substrates and conditions; null when unmeasured.
km_observations Each distinct KM, in molar, read as kcat_observations are, with the same fields.

Example rows

name seqs terms
UniProtKB:P81809 ASPYAFGL av::keyword::UniProtKB-KW...
UniProtKB:A7Z3X4 MSEVLATYLLTDTTDSEKRAEQIAA... av::interpro::InterPro::I...
UniProtKB:A9MS91 MIKERRTELVEGFRHSVPYINTHRG... av::interpro::InterPro::I...

What a model trained on it may say

It says how likely a protein is to carry each vocabulary term at the tier it was trained on. A low score does not show a term absent, since a protein unobserved in an aspect has no negatives there; a signature term means InterPro's models match, not that a function was measured; and an organism term names where the sequence was read, or a phenotype of its species as culture collections and papers record it, not a property of the protein. A metal ion score ranks which metals a metal-binding protein binds, among the proteins that name one, and does not say whether a protein binds a metal at all; a modification or site score ranks what UniProt's sources would record for a protein that records one, and a low one does not show a modification absent; a homodimer score weighs a homodimer against a monomer under the conditions such papers report, a pH is the optimum an assay of the literature would read, not a constant of the protein, a kcat or KM is the mean log10 over the substrates and conditions Swiss-Prot's statements give, not one substrate's constant at one condition, and a melting temperature is a lysate's thermal proteome profiling, not a purified protein's.

Sources and licence

  • UniProtKB/Swiss-Prot XML 2026_03: CC-BY-4.0
  • UniProtKB/TrEMBL 2026_03 entries: CC-BY-4.0
  • GOA UniProt, all species (2026-10-11): CC-BY-4.0
  • InterPro protein2ipr (2026-08-31): CC0-1.0
  • Reactome UniProt2Reactome (2026-06-19): CC0-1.0
  • NCBI Taxonomy lineages of the entries: NCBI, no restrictions on use
  • GlyGen v-2.12.1 and O-GlcNAcAtlas sites: CC-BY-4.0 and CC0-1.0
  • DisProt 2026_06 regions and SIFTS residues: CC-BY-4.0 (DisProt); EMBL-EBI terms (SIFTS)
  • Madin et al. organism traits v1.2.0: CC-BY-4.0
  • Ontologies and UniProt lists: CC-BY-4.0, CC0-1.0 or LGPL (MEROPS)
  • Meltome Atlas workbook (figshare 11829780): CC-BY-4.0
  • 8 more sources were considered and set aside; each is listed with its reason in pipeline.md.

Files

File What it is
{train,validation,test,calibration}.parquet the four splits of whole_vocabulary
<config>/{train,validation,test,calibration}.parquet the four splits of each other config in its folder: go_bp, go_mf, go_cc, localization, metal_ions, modifications, ec, homodimer and measurements
pipeline.md the pipeline as text, for agents
card_facts.json the numbers this card states, written by the exporter from the build's receipts
receipts/ the build's records: build.json (the sources by digest, every stage's counts, time and peak memory), split.json (clusters, strata, repair and the twelve directed searches), survey.json (the identity survey that chose the enforced level), audit.json (each split's tier mix and what metadata alone scores), baselines.json (each view's prior and identity transfer on validation and test), non_enzymes.json (the EC view's non-enzyme rules and their funnel) and acceptance.json (each acceptance check)
tables/ what the views are read from: proteins.parquet (every pooled protein: sequence, accessions, review state, evidence level, annotation score, lineage and study metadata), assertions.parquet (every pooled assertion: protein, term, aspect, evidence code, tier, source, reference and residue positions), measurement_texts.parquet (the entries' biophysicochemical statements, raw, with each evidence code's citations), kinetics.parquet (every KM and kcat those statements give, in molar and inverse seconds, with substrate, ChEBI entity, pH, temperature, evidence and citations, the kinetic views' input), vocabulary.parquet (every term with its name, definition, support per tier and split, and whether it is in the vocabulary: 10 train proteins at a label tier), view_terms.parquet (each set view's terms: its vocabulary terms and the ancestors 10 or more train proteins reach), modification_sites.parquet (the residue-level view, a protein a row with its split: each modification it has a scored site of, with its residue, positions and masked positions; the other residues of that kind are its negatives), disordered_regions.parquet (the residue-level disorder view, a protein a row with its split: its disordered residues, from DisProt's curated regions and UniProt's, its masked ones, MobiDB-lite's predictions among them, and the ones a paper shows ordered; its other residues are negatives, as CAID's DisProt reference counts them, and the residues a PDB structure observes, by SIFTS, are the only negatives of CAID's DisProt-PDB reference), located_features.parquet (the residue-level view of UniProt's other located features, a protein a row with its split: each binding or active site, transmembrane, signal or transit span, disulfide bond, motif, repeat, domain or region class it has a scored residue of, with those residues and its masked ones as runs; its other residues are that class's negatives), lineage.parquet (every protein's taxon at the eight ranks from domain to species, where all its organisms agree, for the taxon's chained choices, its split, and its organism split, 85/5/8/2 of the species with whole species held out, to train and score any view by organism), organism_traits.parquet (each bacterial and archaeal species' traits as Madin et al. 2020 compile them, KEGG's and ProTraits' rows left out: a categorical one, which the species' proteins inherit as terms, the value most of its sources state, and the optimum temperature and pH, growth temperature and doubling time as means, each with its rows and sources; join it to lineage.parquet by species), tokens.parquet (each vocabulary term's token), concepts.parquet and hierarchy.parquet (the ontology concepts the terms name and their typed edges: GO, EC, taxonomy, InterPro, subcellular location, keywords, ChEBI and PSI-MOD), clusters.parquet (each protein's cluster at 0.95, 0.9, 0.8, 0.7, 0.6, 0.5, 0.4 and 0.3 identity), split.parquet (each protein's split, whether it could be held out, and a held-out protein's stratum and identity to train) and neighbours.parquet (each held-out protein's 50 nearest train proteins in the split's own search, with identity and e-value: the identity baseline's input)
composition.png a figure of the build, shown above