---
pretty_name: Pfam Current Release
license: cc0-1.0
tags:
- biology
- protein
- protein-family
- protein-domain
- domain-annotation
- pfam
- jsonl
configs:
- config_name: clans
  data_files:
  - split: train
    path: tables/annotation_pfam_current_release_Pfam-A.clans.tsv.gz.jsonl
- config_name: regions
  data_files:
  - split: train
    path: tables/annotation_pfam_current_release_Pfam-A.regions.tsv.gz.jsonl.parts/part-*.jsonl
---

# Pfam

Pfam is a widely used protein family and domain annotation resource built from curated multiple sequence alignments and profile hidden Markov models. This Hugging Face dataset mirrors tabular exports from the Pfam current release in a normalized JSONL format with row-level provenance. It is intended for large-scale domain annotation lookup, protein family analysis, model training over Pfam-A region assignments, and integration with protein sequence or structure datasets.

This mirror contains two Pfam-A tables: a small clan mapping table and a very large regions table. The clan table maps Pfam-A families to clans and family names, while the regions table records where Pfam-A domains occur on protein sequences. Each JSONL record preserves the upstream row under `row` and adds provenance fields so downstream users can trace the source file and row index.

## Splits

- `clans/train`: 27,480 rows
- `regions/train`: 128,192,943 rows

Total rows: 128,220,423.

The large regions table is split into two JSONL parts:

- `tables/annotation_pfam_current_release_Pfam-A.regions.tsv.gz.jsonl.parts/part-00000.jsonl`
- `tables/annotation_pfam_current_release_Pfam-A.regions.tsv.gz.jsonl.parts/part-00001.jsonl`

## Columns

Every row is a JSON object with the same outer provenance fields:

- `dataset_id`: dataset identifier, always `pfam`
- `row`: raw upstream Pfam row represented as a nested JSON object
- `row_index`: zero-based row index within the upstream source table
- `source_file`: original Pfam source path

The `clans` table contains Pfam-A family and clan metadata. Its upstream fields correspond to family name, Pfam family identifier, clan identifier, clan short name, and Pfam accession.

The `regions` table contains Pfam-A sequence-region assignments. Its nested `row` fields include:

- `pfamseq_acc`: protein sequence accession
- `seq_version`: sequence version
- `crc64`: CRC64 checksum for the protein sequence
- `md5`: MD5 checksum for the protein sequence
- `pfamA_acc`: Pfam-A family accession
- `seq_start`, `seq_end`: domain coordinates on the full sequence
- `ali_start`, `ali_end`: alignment coordinates for the Pfam match

## Usage

Load the smaller clans table:

```python
from datasets import load_dataset

clans = load_dataset("LiteFold/Pfam", "clans", split="train")
print(clans[0])
```

Stream the large regions table:

```python
from datasets import load_dataset

regions = load_dataset("LiteFold/Pfam", "regions", split="train", streaming=True)
row = next(iter(regions))

print(row["dataset_id"])
print(row["row"]["pfamseq_acc"])
print(row["row"]["pfamA_acc"])
print(row["row"]["seq_start"], row["row"]["seq_end"])
```

Load directly from JSONL paths if you do not want to use named configs:

```python
from datasets import load_dataset

regions = load_dataset(
    "json",
    data_files={
        "train": "hf://datasets/LiteFold/Pfam/tables/annotation_pfam_current_release_Pfam-A.regions.tsv.gz.jsonl.parts/part-*.jsonl"
    },
    split="train",
    streaming=True,
)
```

Download the full repository snapshot:

```bash
hf download LiteFold/Pfam --repo-type dataset --local-dir ./pfam
```

Iterate over the raw JSONL files locally:

```python
import json
from pathlib import Path

root = Path("./pfam")
for jsonl in sorted(root.glob("tables/**/*.jsonl")):
    with jsonl.open() as handle:
        for line in handle:
            record = json.loads(line)
            upstream_row = record["row"]
            break
```

## Data Notes

This dataset is a normalized table mirror, not a full mirror of every Pfam artifact. It includes Pfam-A clan mappings and Pfam-A region assignments, but not necessarily the full profile HMM files, seed/full alignments, or upstream sequence databases. Use `row_index` and `source_file` when you need reproducible provenance back to the Pfam release files.

Because the two tables have different nested `row` schemas, load them separately using the `clans` and `regions` configs or explicit `data_files`. Loading every JSONL file in the repository as a single dataset may fail schema casting in the Hugging Face dataset viewer.

## License

This mirror follows the upstream Pfam release license metadata provided with the dataset: `CC0 1.0`.

# Citation

```bibtex
@article{mistry2021pfam,
  title   = {{Pfam}: The protein families database in 2021},
  author  = {Mistry, Jaina and Chuguransky, Sara and Williams, Lowri and Qureshi, Matloob and Salazar, Gustavo A. and Sonnhammer, Erik L. L. and Tosatto, Silvio C. E. and Paladin, Lisanna and Raj, Shriya and Richardson, Lorna J. and Finn, Robert D. and Bateman, Alex},
  journal = {Nucleic Acids Research},
  volume  = {49},
  number  = {D1},
  pages   = {D412--D419},
  year    = {2021},
  doi     = {10.1093/nar/gkaa913},
  url     = {https://doi.org/10.1093/nar/gkaa913}
}
```