---
license: cc-by-4.0
pretty_name: UniRef90
size_categories:
  - 100M<n<1B
task_categories:
  - other
language:
  - en
tags:
  - biology
  - proteins
  - sequences
  - fasta
  - uniref
  - clustering
---

# UniRef90

Normalized FASTA shards of UniRef90 cluster representative sequences (90% sequence identity threshold).

Processed and uploaded by the [MegaData](https://github.com/) post-download pipeline
(internal repo). Original source: <https://www.uniprot.org/help/uniref>.

## Statistics

| | |
|---|---|
| Source files | 1 |
| Shards | 189 |
| Compressed shard bytes | 38.65 GiB (41,498,595,315) |
| Records (per-source manifest sum) | 188,848,220 |
| Residues (per-source manifest sum) | 66,359,825,357 |
| Aggregate manifest `total_records` | 188848220 |

## Layout

```
.
├── _MANIFEST.json                      # aggregate manifest written by the pipeline
├── manifests/<source_slug>.json        # per-source manifest (records, residues, shards)
├── metadata/<source_slug>.records.jsonl  # per-record provenance
└── sequences/<source_slug>/shard-NNNNNN.fasta.zst
```

`<source_slug>` corresponds 1:1 with an upstream source archive; e.g.
`sequence_uniprotkb_uniprot_sprot.fasta.gz`.

## Loading

Stream every shard of one source (replace `<source_slug>` with the directory of
interest under `sequences/`):

```bash
hf download LiteFold/UniRef90 --repo-type dataset \
  --include 'sequences/<source_slug>/shard-*.fasta.zst' \
  --local-dir ./uniref90
zstd -dc ./uniref90/sequences/<source_slug>/shard-*.fasta.zst | head
```

Programmatic streaming with [`zstandard`](https://pypi.org/project/zstandard/):

```python
from huggingface_hub import snapshot_download
from pathlib import Path
import zstandard as zstd

local = snapshot_download(
    repo_id="LiteFold/UniRef90",
    repo_type="dataset",
    allow_patterns=["sequences/*/shard-*.fasta.zst"],
)

dctx = zstd.ZstdDecompressor()
for shard in sorted(Path(local).rglob("shard-*.fasta.zst")):
    with shard.open("rb") as f, dctx.stream_reader(f) as reader:
        buf = b""
        while chunk := reader.read(1 << 20):
            buf += chunk
            *lines, buf = buf.split(b"\n")
            for line in lines:
                ...  # naive splitter; swap in your FASTA parser
```

## License

CC BY 4.0 (UniProt Consortium).

## Citation

> Suzek BE, Wang Y, Huang H, McGarvey PB, Wu CH; UniProt Consortium. UniRef clusters: a comprehensive and scalable alternative for improving sequence similarity searches. Bioinformatics, 31(6):926-32, 2015.

## Provenance

Built from the local manifest entry `uniref90` of `manifests/atlas_download_plan.json`.
Pipeline source: `megadata-post normalize --dataset uniref90`.