Synthyra/SwissProt-Annotation-Vocabulary

Swiss-Prot Annotation Vocabulary 2026_02

This release converts a pinned Swiss-Prot snapshot into a versioned protein annotation vocabulary. Stable, namespaced term identifiers are the biological identity. Integer tokens are specific to this vocabulary and grammar version.

Release summary

Load the data

Each table is a named Hugging Face configuration:

from datasets import load_dataset

clusters = load_dataset("Synthyra/SwissProt-Annotation-Vocabulary", "clusters")
excluded = load_dataset("Synthyra/SwissProt-Annotation-Vocabulary", "excluded")
migration = load_dataset("Synthyra/SwissProt-Annotation-Vocabulary", "migration")
minimal = load_dataset("Synthyra/SwissProt-Annotation-Vocabulary", "minimal")
rich = load_dataset("Synthyra/SwissProt-Annotation-Vocabulary", "rich")
terms = load_dataset("Synthyra/SwissProt-Annotation-Vocabulary", "terms")

Pin a Hub commit revision for reproducible downstream work.

Dataset configurations

Tables and columns

The rich protein table contains the full release record. Its main fields are:

All global_* arrays are parallel at the outer level; each inner provenance list contains the sorted unique values exposed by the source adapters for that term. All local_* arrays are parallel and have equal length within a row. The table also retains local source labels, descriptions, feature identifiers, coordinate modifiers, ligand identifiers, relations, and source references.

Taxonomy is rich protein metadata, not an annotation-vocabulary aspect. taxon_id, lineage_taxon_ids, and lineage_names are retained from the pinned NCBI taxonomy enrichment for stratification and confound control. Taxonomy does not receive biological tokens and is absent from global_term_ids, global_tokens, annotations, the term tables, and migration targets.

InterPro remains in the target vocabulary. Gene3D removal is a measured recipe decision, not a missing parser path: in the completed V1 coverage audit, only 468 of 574,627 Swiss-Prot proteins would lose their sole domain-level annotation when Gene3D was removed and InterPro retained. Raw Gene3D observations are used only for that separately recorded review. Gene3D is absent from global term and token arrays, minimal annotations, retained and excluded term tables, released per-aspect statistics, and migration targets.

The minimal table contains only sequence and annotations. The terms table maps every retained integer token to stable identity, label, description, locality, support, deterministic grammar-key material, observed evidence codes, relations, source references, hierarchy or replacement identifiers when supplied, source-pin fields, vocabulary version, and the primary source-specific release. sort_aspect_rank is the configured aspect-grammar rank. sort_identifier_key is compact JSON containing the remainder of the canonical term sort key: naturalized accession components, case-folded source database, and canonical term ID. Evidence codes, relations, and source references are sorted unique unions over all non-negated observations of the term. source_provenance names the pinned input fields that define or observe the term, while source_release is the primary source-specific release for that aspect.

parent_ids contains direct, canonical, source-prefixed parent term IDs exposed by the selected ontology adapter; a parent need not pass this vocabulary's support threshold. replacement_ids contains only canonical, source-prefixed automatic replacement IDs backed by an explicit source assertion such as Gene Ontology replaced_by. Gene Ontology consider suggestions are excluded from automatic replacements and remain migration manual-review candidates. Either array is empty when no selected source adapter exposes the assertion. The separate migration table is authoritative for old-token translation because multiple historical tokens can resolve to one current term.

Clustering and splits

Sequences were clustered with MMseqs2 eec9c354be4276d2373996af2e50808b1390d527 at 30% sequence identity and 50% alignment coverage. Coverage mode 0 requires the query and target to each meet the threshold. Equivalently, the minimum of query coverage and target coverage must be at least 50%.

Whole clusters were assigned with seed 42 toward 90% train, 5% validation, and 5% test by protein count. A cluster occurs in exactly one split.

Vocabulary identity and grammar

Canonical identifiers use av::aspect::source_database::stable_accession. Terms are ordered first by the explicit aspect grammar and then by natural stable-accession order. Biological tokens start at zero. Special tokens are appended after all biological tokens, so they cannot collide with biological token zero.

Terms observed in fewer than 10 distinct Swiss-Prot proteins are excluded from the integer vocabulary. Their measured counts remain in the release artifacts so another threshold can be applied without treating absence as nonexistence.

Local annotation coordinates

Local starts and ends are 1-based and inclusive, matching the normalized release convention. For a Python string, use sequence[start - 1:end]. Point features have start == end. Paired features, such as disulfide bonds, use local_pair_members to preserve endpoint membership.

Fuzzy, unknown, or out-of-range source coordinates are not converted into exact spans. Their exclusions and counts belong in the build manifest. A local term reports a source annotation at that position; it does not establish that every residue inside a broad region has the same function.

Migration from older raw tokens

Migration follows four explicit steps: old raw token to the old token dictionary, old label to a stable biological identifier, official replacement or obsolete-term mapping when available, and stable identifier to the new token dictionary. It never assumes that an integer has the same meaning across vocabulary versions.

Each migration row reports one of the API statuses: exact, replaced, current_but_filtered, current_not_observed, ambiguous, obsolete, removed/unmapped, unknown_old_token, or invalid_legacy_label. current_but_filtered means the term is current and observed in this Swiss-Prot release, but measured protein support is below the vocabulary threshold; its canonical candidate is retained with a null token and no automatic token is assigned. current_not_observed means a source-backed current or canonical candidate is absent from this pinned Swiss-Prot release; the candidate is retained with a null token and no automatic token is assigned. ambiguous includes every Gene Ontology consider suggestion, even a single suggestion, and records manual_review_candidates separately. Outcomes other than exact or replaced require review. Legacy token zero is also reported explicitly because the historical Translator vocabulary used zero for both a biological term and padding in different contexts. Callers must state whether input tokens are biological annotations or model-sequence tokens.

Legacy _threed labels retain their parsed Gene3D identity for audit but always report removed/unmapped with exclusion_reason=aspect_excluded_from_target_recipe. They have no candidate, new term, or new token, even if a caller supplies a replacement bridge.

Sources and licenses

This card uses license: other because the release combines records derived from multiple upstream resources. No new license grant is asserted here. Users must follow the terms named by each source and verify them for their intended use.

Artifact integrity

The release manifest fixes artifact paths, byte sizes, row counts where applicable, and SHA-256 digests. Verify the digest after downloading an artifact and pin the exact Hub revision used by an experiment.

Caveats

• —Swiss-Prot annotations are curated but incomplete and concentrated in well-studied proteins and taxa.
• —The support filter removes rare terms from integer labels, not from biological reality.
• —Sequence length and annotation count are correlated in Swiss-Prot. Naive pooled association or prediction can exploit annotation density and the number of opportunities for a model feature to activate. Use cluster-disjoint evaluation and control both quantities.
• —Taxonomy can act as a shortcut in predictive models. Evaluate results with taxonomy and sequence length controlled when they can confound the question.
• —MMseqs2 cluster-disjoint splits limit leakage under the stated sequence thresholds. They do not exclude every remote evolutionary, structural, or functional relationship.
• —Database annotations, migration matches, and model associations are evidence-bearing records. They are not experimental confirmation of a protein function or residue mechanism.
• —Rich provenance is lossless for values exposed by the selected source adapters, not necessarily for every raw source line. Biopython's Swiss-Prot parser normalizes keyword fields and does not expose raw keyword ECO blocks. A full scan found no evidence-bearing keyword lines in the pinned 2026_02 DAT; future source releases must repeat this audit.
• —Exact build exclusions, parser decisions, source checksums, and validation results are defined by the pinned manifests, not by mutable upstream URLs.

Citation

Cite the exact Hugging Face commit revision loaded by the experiment and the upstream resources listed in this card.

@dataset{annotation_vocabulary_2026_02,
  author = {Annotation Vocabulary contributors},
  title = {Swiss-Prot Annotation Vocabulary 2026_02},
  year = {2026},
  version = {2026_02-support10-v1},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/Synthyra/SwissProt-Annotation-Vocabulary},
  note = {Pin the immutable Hub commit revision used by the experiment}
}