UniRef50

Complete UniRef50 dataset from UniProt, converted from XML to sharded Parquet. UniRef50 clusters sequences at 50% identity, providing the most aggressively deduplicated UniRef tier — ideal for training protein language models and building diverse, non-redundant sequence sets.

Part of the ConvergeBio Protein Database Collection — see also UniRef90, UniRef100, and UniClust30.

Dataset Summary

Clusters	60,315,044
Shards	130
Compressed size	~17 GB (zstd)
Sequence lengths	11 – 49,499 aa (median 189, mean 287)
Members per cluster	1 – 321,476 (median 1, mean 8.7)
GO annotation coverage	MF 18.6% · BP 12.0% · CC 12.4%
Updated range	2006-10-31 to 2026-01-28

Schema

Each row represents one UniRef50 cluster with its representative sequence and metadata.

Column	Type	Description
id	string	Cluster identifier (e.g. UniRef50_P12345)
name	string	Cluster name from UniProt
updated	string	Last update date (YYYY-MM-DD)
member_count	int32	Number of sequences in the cluster
common_taxon	string	Lowest common taxon across members
common_taxon_id	int32	NCBI Taxonomy ID of common taxon
seed_id	string	ID of the seed sequence
go_mf	list<string>	GO Molecular Function terms (GO:XXXXXXX)
go_bp	list<string>	GO Biological Process terms
go_cc	list<string>	GO Cellular Component terms
member_ids	list<string>	All member sequence IDs
rep_member_id	string	Representative member ID
rep_member_id_type	string	ID type (e.g. UniProtKB ID, UniParc ID)
rep_organism	string	Source organism of representative
rep_organism_tax_id	int32	NCBI Taxonomy ID of representative organism
rep_protein_name	string	Protein name of representative
rep_accessions	list<string>	UniProtKB accessions of representative
rep_uniparc_id	string	UniParc ID of representative
rep_uniref90_id	string	Child UniRef90 cluster ID
rep_uniref100_id	string	Child UniRef100 cluster ID
rep_is_seed	bool	Whether the representative is the seed sequence
sequence	large_string	Representative protein sequence (uppercase amino acid alphabet)
sequence_length	int32	Length of the sequence in residues
sequence_crc64	string	CRC64 checksum from UniProt (hex)
sequence_xxh128	string	xxHash-128 of the sequence (hex, computed at build time)

Usage

from datasets import load_dataset

# Stream without downloading everything
ds = load_dataset("ConvergeBio/uniref50", streaming=True)
for row in ds["train"]:
    print(row["id"], row["sequence_length"])
    break

# Or load fully
ds = load_dataset("ConvergeBio/uniref50")

Data Processing

• Source: uniref50.xml.gz from the UniProt FTP
• Parsing: Streaming XML parse with lxml.etree.iterparse, multi-process for throughput
• Integrity: xxHash-128 computed per sequence; CRC64 preserved from source XML
• Validation: Passed all tiers — schema conformance, zero null/empty sequences, xxHash roundtrip, CRC64 format, GO term format, member ID consistency, and field-by-field comparison against source XML
• Format: Sharded Parquet with zstd compression

Source & Citation

UniRef is produced by the UniProt Consortium:

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):926–932 (2015). doi:10.1093/bioinformatics/btu739

About

Built by Converge Bio — accelerating drug discovery with generative AI. Converge Bio develops foundation models for protein engineering, antibody design, and gene expression optimization, powering its computational lab products ConvergeAB, ConvergeGEO, and ConvergeCELL.

License

UniProt data is available under CC BY 4.0.