Xiaowei0402/uniprot_subcellular_localization

UniProt Subcellular Localization (Vertebrates) — ProVADA

A curated collection of vertebrate UniProt/Swiss‑Prot protein domains labeled for cytosolic and extracellular localization. We remove signal peptides, restrict domain lengths, and provide both the full set and a 30% identity‑clustered representative set with train/test/validation splits (70/20/10). This dataset underpins the subcellular localization oracle in ProVADA (preprint).

---

See the Files section for exact filenames and formats.

Supported Tasks

• Sequence classification: Predicting whether a protein domain is viable in the cytosol/extracellular space.
• Protein representation learning: Pretraining or evaluating embeddings for subcellular localization viability.

Languages / Alphabets

• Amino‑acid sequences using the standard 20‑letter alphabet (uppercase single‑letter codes).

---

Data Structure

Files

• full_dataset.csv — Full curated set (10,652 domains).
• representative_data_train.csv, representative_data_val.csv, representative_data_test.csv — 30%‑identity clustered representative set with predefined splits for model development.

Column schema is consistent across files.

Fields

• sequence (str) — Amino‑acid sequence of the domain.
• converted_id (str) — UniProt accession plus the domain region in brackets (e.g., P12345[35-210]).
• cytosolic_label (bool) — True if the domain localizes to the intracellular/cytosolic region.
• extracellular_label (bool) — True if the domain localizes to the extracellular space (secreted or extracellular domain of a TM protein).

Notes:

• The two labels are not mutually exclusive; a small number of domains have both annotations.

• For TM proteins we include only the annotated extracellular domain within the length range.

---

Processing & Quality Control

1. Curation from Swiss‑Prot. We include entries with experimental evidence and exclude organelles such as lysosomes, endosomes, peroxisomes, and mitochondria.
2. Domain extraction. We keep mature regions only (signal sequences removed). Where multiple domains are annotated, we retain shorter, individual domains between 50–1000 aa. For transmembrane proteins, we extract annotated extracellular domains within this length.
3. Redundancy control. We cluster all sequences using MMseqs2 at 30% identity and keep one representative/cluster.
4. Splitting. We create stratified train/test/validation splits on the representative subset to balance labels and protein families.

---

Suggested evaluation splits

If you only need a single‑split for quick evaluation, prefer the representative set (representative_data_*.csv) to reduce family‑level redundancy.

---

Licensing

• Upstream data: Derived from UniProt/Swiss‑Prot (released under CC BY 4.0).
• This dataset release: CC BY 4.0. Please ensure your downstream use complies with UniProt’s terms and institutional policies.

---

Citation

If you use this dataset, please cite the ProVADA preprint where the curation procedure and usage are described:

@article{Lu2025,
  title = {ProVADA: Generation of Subcellular Protein Variants via Ensemble-Guided Test-Time Steering},
  url = {http://dx.doi.org/10.1101/2025.07.11.664238},
  DOI = {10.1101/2025.07.11.664238},
  publisher = {Cold Spring Harbor Laboratory},
  author = {Lu, Wenhui Sophia and Zhang, Xiaowei and Mille-Fragoso, Luis S. and Dai, Haoyu and Gao, Xiaojing J. and Wong, Wing Hung},
  year = {2025},
  month = jul
}