uniprot_structures

Residue-level annotation dataset sourced from the UniProtKB REST API.

Intended use

Residue-level structural-motif prediction from protein sequence. Scores indicate recovery of curated UniProt features, not independent experimental evidence.

Source and labels

• Source API: https://rest.uniprot.org/uniprotkb/stream
• Card generated (UTC): 2026-10-07
• Organism Taxonomy ID: 9606 (all = all organisms).
• Review status: reviewed (reviewed = Swiss-Prot, unreviewed = TrEMBL, all = both).
• Query: (organism_id:9606) AND (reviewed:true)
• Requested UniProt fields: accession, sequence, length, ft_helix, ft_strand, ft_turn, ft_mod_res, ft_lipid, ft_disulfid, ft_carbohyd, ft_transmem, ft_intramem, ft_topo_dom, ft_propep, ft_signal, ft_transit, ft_act_site, ft_binding, ft_dna_bind, ft_domain, ft_region, ft_repeat, ft_coiled, ft_zn_fing
• UniProt feature headers: Repeat, Coiled coil, Zinc finger
• UniProt feature fields: ft_repeat, ft_coiled, ft_zn_fing
• Per-residue class source: UniProt feature codes.
• Feature blocks are labeled only when their own /evidence qualifier contains ECO:0000269 or ECO:0000305; secondary-structure features also accept ECO:0007829|PDB:<structure ID> (PDB-derived structural assignments). Missing, malformed, and other evidence is excluded; /note text is never used as evidence.
• Background residues use -. Conflicting annotations and positions covered by features with rejected evidence or uncertain coordinates use the ignored target -100. Other integer targets index the sorted label_vocabulary.json.
• Disulfide bonds label only their two bonded cysteines, not the span between them.
• Note-derived labels drop parenthetical text (e.g. N-linked (GlcNAc...) asparagine becomes N-linked asparagine); notes starting with a qualifier such as (Microbial infection) fall back to the feature code.
• The ordered label vocabulary contains 4 labels.
• Rows without a non-empty sequence are dropped; all target vectors have the same length as their sequence.

Preprocessing and splits

Identical sequences are collapsed to one entry before vocabulary construction, negative sampling, and splitting; their residue labels are merged, with disagreeing residues ignored (-100). Features with uncertain (?) positions are treated like evidence-rejected features. Proteins with at least one qualified annotated residue are retained. Genuinely feature-free proteins may be sampled as unannotated negatives to balance the positive pool; proteins with candidate features rejected only for missing, malformed, or non-allowed evidence or uncertain positions are excluded from that negative pool. Background residues on retained proteins use - as an unannotated background class, not experimentally verified absence. id is the UniProtKB accession of the entry kept for each unique sequence. Splits are assigned by whole MMseqs2 clusters targeting {'train': 0.8, 'validation': 0.1, 'test': 0.1}. No maximum sequence-length filter was applied. Pooled random, stratified, and hold-cluster-out subsets are also included; hold-cluster-out subsets require enough clusters.

Clustering settings:

• MMseqs workflow: easy-cluster
• Minimum sequence identity (--min-seq-id): 0.3
• Minimum coverage (-c): 0.8
• Coverage mode (--cov-mode): 1
• Threads: 1
• Other MMseqs2 options use MMseqs2 defaults.

Split sizes

• test: 42 rows
• test_cluster: 42 rows
• test_random: 42 rows
• test_stratified: 42 rows
• train: 338 rows
• train_cluster: 338 rows
• train_random: 338 rows
• train_stratified: 338 rows
• validation: 42 rows
• validation_cluster: 42 rows
• validation_random: 42 rows
• validation_stratified: 42 rows

Dataset statistics

stats.json at the dataset root contains row counts by split, columns, and SeqKit sequence-length metrics.