swhitfield/uniprot-GO_cc

GO_cc

Sourced from the UniProt REST API (https://rest.uniprot.org/uniprotkb/stream) and prepared for Hugging Face datasets usage.

Intended use

Protein-level Gene Ontology annotation prediction for the cellular_component aspect. Labels measure recovery of UniProt annotations, not experimental causality.

Source

• Selected GO aspect: cellular_component
• Card generated (UTC): 2026-10-05
• Organism Taxonomy ID: 9606 (all = all organisms).
• Review status: reviewed (reviewed = Swiss-Prot, unreviewed = TrEMBL, all = both). UniProtKB TSV fields requested: accession, reviewed, id, protein_name, gene_names, organism_name, length, sequence, go_f, go_p, go_c.

Query:

(organism_id:9606) AND (reviewed:true)

Preprocessing

• Read the selected aspect's Gene Ontology TSV field and extract distinct GO identifiers from each protein's annotation cell.
• Consider GO identifiers found in at least 5 query-result proteins. After splitting, keep terms with at least 20 positives in training, 3 in validation, and 3 in test. Sort the retained identifiers lexicographically; this order defines the target-vector positions in label_vocabulary.json. Human-readable term names are recorded there in label_names, keyed by GO identifier.
• Keep rows with a non-empty sequence and at least one retained GO identifier.
• Apply the optional sequence-length filter before splitting these rows by sequence clusters. Sequences are whitespace-trimmed. No maximum sequence-length filter was applied.
• Output columns are id, sequence, targets, and split. IDs are generated as <dataset>_<split>_<row-index> after split assignment.

Labels

targets is a multi-hot vector over 224 GO ids (aspect: cellular_component). Each term meets the configured source and canonical split support cutoffs. The ordered vocabulary is in label_vocabulary.json.

Splitting and MMseqs2

MMseqs2 clusters the eligible sequences in one easy-linclust pass. The effective clustering settings are:

• Minimum sequence identity (--min-seq-id): 0.3
• Minimum coverage (-c): 0.8
• Coverage mode (--cov-mode): 1
• Threads (--threads): 1
• Linclust algorithm version (--linclust-version): 2
• Other MMseqs2 options are omitted and use MMseqs2 defaults.

Whole clusters are assigned to exactly one split. Target row counts are calculated from fractions {'train': 0.8, 'validation': 0.1, 'test': 0.1} (rounded to the nearest integer, with rounding drift assigned to the largest split). Clusters are considered largest-first and greedily assigned to the split with the greatest relative size deficit; the configured seed 1957723 breaks ties between equal-sized clusters. No upstream split assignment is available.

When create_split_subsets is enabled, pooled random, stratified, and hold-cluster-out subsets are also included. Hold-cluster-out subsets require enough MMseqs clusters to populate all three roles.

Split row counts after preprocessing:

• test: 1890 rows
• test_cluster: 1890 rows
• test_random: 1890 rows
• test_stratified: 1890 rows
• train: 15107 rows
• train_cluster: 15107 rows
• train_random: 15107 rows
• train_stratified: 15107 rows
• validation: 1888 rows
• validation_cluster: 1888 rows
• validation_random: 1888 rows
• validation_stratified: 1888 rows

Dataset statistics

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