README.md

---
license: cc-by-4.0
pretty_name: UniProtKB Processed
size_categories:
  - 100M<n<1B
task_categories:
  - feature-extraction
language:
  - en
tags:
  - biology
  - proteins
  - uniprot
  - uniprotkb
  - swiss-prot
  - trembl
  - protein-sequences
  - bioinformatics
  - train-validation-test-split
configs:
  - config_name: sprot
    data_files:
      - split: train
        path:
          - tables/source_set=sprot/split=train/*.jsonl.gz
      - split: validation
        path:
          - tables/source_set=sprot/split=validation/*.jsonl.gz
      - split: test
        path:
          - tables/source_set=sprot/split=test/*.jsonl.gz
  - config_name: sprot_varsplic
    data_files:
      - split: train
        path:
          - tables/source_set=sprot_varsplic/split=train/*.jsonl.gz
      - split: validation
        path:
          - tables/source_set=sprot_varsplic/split=validation/*.jsonl.gz
      - split: test
        path:
          - tables/source_set=sprot_varsplic/split=test/*.jsonl.gz
  - config_name: trembl
    data_files:
      - split: train
        path:
          - tables/source_set=trembl/split=train/*.jsonl.gz
      - split: validation
        path:
          - tables/source_set=trembl/split=validation/*.jsonl.gz
      - split: test
        path:
          - tables/source_set=trembl/split=test/*.jsonl.gz
---

# UniProtKB Processed

This dataset is a table-friendly, split-aware version of LiteFold/UniProtKB.
It converts UniProtKB FASTA headers into normalized records that can be browsed
in the Hugging Face Dataset Viewer and loaded with the `datasets` library.

The table contains parsed protein metadata and deterministic train/validation/test
split assignments. To keep the dataset practical for table viewing, full amino
acid sequences are not duplicated here; each record includes `sequence_sha256`
for sequence-level identity and deduplication.

## Dataset Summary

| Source set | Description | Records |
|---|---|---:|
| `sprot` | Swiss-Prot reviewed canonical proteins | 574,627 |
| `sprot_varsplic` | Swiss-Prot alternative isoform sequences | 41,333 |
| `trembl` | TrEMBL unreviewed proteins | 202,556,314 |
| **Total** |  | **203,172,274** |

## Splits

Splits are deterministic and assigned from the exact amino-acid sequence hash.
Exact duplicate sequences are kept in the same split.

| Split | Records |
|---|---:|
| `train` | 162,548,965 |
| `validation` | 20,308,533 |
| `test` | 20,314,776 |

These are exact-sequence splits, not homology-cluster splits. For strict
homology-aware model evaluation, split again by UniRef/MMseqs cluster ids.

## Loading

Load Swiss-Prot reviewed entries:

```python
from datasets import load_dataset

sprot = load_dataset("LiteFold/UniProtKB", "sprot")
train = sprot["train"]
valid = sprot["validation"]
test = sprot["test"]
```

Load Swiss-Prot alternative isoforms:

```python
from datasets import load_dataset

isoforms = load_dataset("LiteFold/UniProtKB", "sprot_varsplic")
```

Load only TrEMBL:

```python
from datasets import load_dataset

trembl = load_dataset("LiteFold/UniProtKB", "trembl", streaming=True)
for row in trembl["train"]:
    print(row["accession"], row["protein_name"])
    break
```

Filter examples:

```python
from datasets import load_dataset

ds = load_dataset("LiteFold/UniProtKB", "sprot", split="train")
human = ds.filter(lambda x: x["taxon_id"] == 9606)
reviewed_kinases = human.filter(lambda x: "kinase" in x["protein_name"].lower())
```

For large-scale TrEMBL use, prefer streaming:

```python
from datasets import load_dataset

rows = load_dataset("LiteFold/UniProtKB", "trembl", split="train", streaming=True)
for row in rows:
    ...
```

Because the source sets are very different in size, this dataset does not define
a single mixed `default` config. Load one of `sprot`, `sprot_varsplic`, or
`trembl` explicitly.

## Columns

| Column | Type | Description |
|---|---|---|
| `record_id` | string | Full UniProt FASTA record id, e.g. `sp|Q6GZX4|001R_FRG3G` |
| `accession` | string | UniProt accession |
| `base_accession` | string | Accession without isoform suffix |
| `entry_name` | string | UniProt entry name |
| `reviewed` | bool | `true` for Swiss-Prot reviewed entries |
| `source_set` | string | `sprot`, `sprot_varsplic`, or `trembl` |
| `is_isoform` | bool | Whether the row is an isoform entry |
| `isoform_number` | int | Isoform suffix number, or `-1` if not applicable |
| `protein_name` | string | Parsed protein name from the FASTA header |
| `organism_name` | string | Parsed organism name |
| `taxon_id` | int | NCBI taxonomy id |
| `gene_name` | string | Parsed gene name, or `""` if unavailable |
| `protein_existence` | int | UniProt PE value, or `-1` if unavailable |
| `sequence_version` | int | UniProt sequence version, or `-1` if unavailable |
| `length` | int | Amino-acid sequence length |
| `sequence_sha256` | string | SHA-256 hash of the amino-acid sequence |
| `split` | string | `train`, `validation`, or `test` |

## Missing Values

For stable Hugging Face Dataset Viewer schema inference, table files do not use
JSON `null` values.

- Missing optional integer values are encoded as `-1`.
- Missing optional string values are encoded as `""`.

## Processing Notes

The public tables intentionally omit internal provenance columns such as raw
FASTA header, source shard, source file, record index, and split internals.
The postprocessing manifest is available in `_POSTPROCESS_MANIFEST.json`.

## Source

Derived from LiteFold/UniProtKB, which was built from UniProtKB FASTA releases.

Original source: UniProtKB.

## License

CC BY 4.0.

## Citation

If you use this dataset, cite UniProt:

The UniProt Consortium. UniProt: the Universal Protein Knowledgebase in 2023.
Nucleic Acids Research, 51(D1):D523-D531, 2023.