Datasets:

LiteFold
/

NCBI

pretty_name: NCBI RefSeq Protein Shard Index
license: other
tags:
  - biology
  - proteins
  - sequences
  - fasta
  - ncbi
  - refseq
  - parquet
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet
      - split: test
        path: data/test-*.parquet

NCBI RefSeq Protein Shard Index

The RefSeq protein database is the protein subset of NCBI's broader Reference Sequence collection, a curated non-redundant set of genomic, transcript, and protein sequences spanning the tree of life. Protein records are produced through a combination of expert curation, the NCBI Eukaryotic and Prokaryotic Genome Annotation pipelines, and propagation from collaborating resources, and are cross-linked to their source genome and transcript records. RefSeq protein accessions use prefixed identifiers (NP_ for curated, XP_ for predicted, WP_ for non-redundant prokaryotic proteins, YP_ for organelle and viral proteins) so that downstream tools can distinguish curation status at a glance.

Splits

The split is deterministic by file ID: sha256(file_id) % 10. Bucket 0 is test; buckets 1 through 9 are train.

Split	Rows
train	4,676
test	502
total	5,178

Source Statistics

Field	Value
Source FASTA files	1,725
RefSeq protein records	459,415,871
Residues	179,203,453,293
Sequence shards	1,725
Compressed sequence shard bytes	78,108,688,857
Metadata JSONL bytes	158,533,041,909

Usage

pip install datasets

Load the shard index:

from datasets import load_dataset

ds = load_dataset("LiteFold/NCBI")
print(ds)
print(ds["train"][0])

Load one split:

from datasets import load_dataset

train = load_dataset("LiteFold/NCBI", split="train")
test = load_dataset("LiteFold/NCBI", split="test")

List sequence shards:

from datasets import load_dataset

index = load_dataset("LiteFold/NCBI", split="train")
shards = index.filter(lambda row: row["is_sequence_shard"])
print(shards[0]["path"])

Find a source FASTA and its files:

from datasets import load_dataset

index = load_dataset("LiteFold/NCBI", split="train")
rows = index.filter(lambda row: row["source_file"] == "sequence/ncbi_refseq/release_complete/complete.1486.protein.faa.gz")
for row in rows:
    print(row["role"], row["path"], row["size_bytes"])

Download all sequence shards:

hf download LiteFold/NCBI --repo-type dataset \
  --include 'sequences/*/shard-*.fasta.zst' \
  --local-dir ./ncbi_refseq_protein

Download one source shard:

hf download LiteFold/NCBI --repo-type dataset \
  --include 'sequences/sequence_ncbi_refseq_release_complete_complete.1486.protein.faa.gz/shard-*.fasta.zst' \
  --local-dir ./ncbi_refseq_protein

Stream a downloaded shard with Python:

from pathlib import Path
import zstandard as zstd

shard = next(Path("./ncbi_refseq_protein").rglob("shard-*.fasta.zst"))
dctx = zstd.ZstdDecompressor()
with shard.open("rb") as f, dctx.stream_reader(f) as reader:
    print(reader.read(1024).decode("utf-8", errors="replace"))

Columns

Column	Description
`file_id`	Stable row ID, equal to the repository path.
`repo_id`	Hugging Face dataset repository.
`source_sha`	Source repository commit used to build the index.
`dataset_id`	Source dataset identifier from the manifest.
`source_slug`	Source slug from the original pipeline manifest.
`source_file`	Original source FASTA file path.
`path`	File path in the repository.
`role`	File role, such as `sequence_shard`, `metadata_records`, or `source_manifest`.
`shard_index`	Numeric shard index for sequence shards.
`size_bytes`	File size in bytes.
`compression`	Compression format, when applicable.
`records_in_source`	Protein record count for the source FASTA file.
`residues_in_source`	Residue count for the source FASTA file.
`shards_in_source`	Shard count for the source FASTA file.
`records_total`	Total protein record count from the aggregate manifest.
`residues_total`	Total residue count from the aggregate manifest.
`total_shards`	Total sequence shard count.
`is_sequence_shard`	Whether the row points to a FASTA shard.
`is_metadata_records`	Whether the row points to a per-record metadata JSONL.
`download_pattern`	Recommended path or glob for downloading.
`access_note`	Note describing the index scope.
`split_bucket`	Deterministic split bucket from `sha256(file_id) % 10`.

Preparation

The normalization script used to create the Parquet files is included at scripts/prepare_ncbi_dataset.py.

Citation

@article{goldfarb2025refseq,
  title     = {{NCBI RefSeq}: reference sequence standards through 25 years of curation and annotation},
  author    = {Goldfarb, Tamara and Kodali, Vamsi K. and Pujar, Shashikant and Brover, Vyacheslav and Robbertse, Barbara and Farrell, Catherine M. and Oh, Dong-Ha and Astashyn, Alexander and Ermolaeva, Olga and Haddad, Diana and Hlavina, Wratko and Hoffman, Jinna and Jackson, John D. and Joardar, Vinita S. and Kristensen, David and Masterson, Patrick and McGarvey, Kelly M. and McVeigh, Richard and Mozes, Eyal and Murphy, Michael R. and Schafer, Susan S. and Souvorov, Alexander and Spurrier, Brett and Strope, Pooja K. and Sun, Hanzhen and Vatsan, Anjana R. and Wallin, Craig and Webb, David and Brister, J. Rodney and Hatcher, Eneida and Kimchi, Avi and Klimke, William and Marchler-Bauer, Aron and Pruitt, Kim D. and Thibaud-Nissen, Fran{\c{c}}oise and Murphy, Terence D.},
  journal   = {Nucleic Acids Research},
  volume    = {53},
  number    = {D1},
  pages     = {D243--D257},
  year      = {2025},
  publisher = {Oxford University Press},
  doi       = {10.1093/nar/gkae1038}
}