UniClust30 (UniRef30)

Complete UniClust30 / UniRef30 dataset (release 2023_02) from the Söding Lab, converted from HH-suite A3M format to sharded Parquet. UniClust30 clusters UniProt sequences at 30% identity and includes precomputed multiple sequence alignments (MSAs) — widely used as input for protein structure prediction (AlphaFold, ColabFold) and protein language model pretraining.

Part of the ConvergeBio Protein Database Collection — see also UniRef50, UniRef90, and UniRef100.

Dataset Summary

Clusters	36,293,491
Shards	629
Release	2023_02
Includes	Precomputed A3M multiple sequence alignments per cluster

Schema

Each row represents one UniClust30 cluster with its representative sequence, MSA, and membership information.

Column	Type	Description
cluster_id	string	Cluster identifier (UniRef30 accession)
representative_id	string	UniProt accession of the representative sequence
sequence	large_string	Representative protein sequence (uppercase amino acid alphabet)
sequence_length	int32	Length of the representative sequence in residues
sequence_xxh128	string	xxHash-128 of the sequence (hex, computed at build time)
num_aligned	int32	Number of sequences in the A3M multiple sequence alignment
a3m	large_string	Full A3M-formatted MSA for the cluster
member_count	int32	Number of cluster members (from mapping file)
member_ids	list<string>	All member UniProt accessions

Usage

from datasets import load_dataset

# Stream without downloading everything
ds = load_dataset("ConvergeBio/uniclust30", streaming=True)
for row in ds["train"]:
    print(row["cluster_id"], row["sequence_length"], row["num_aligned"])
    break

# Or load fully
ds = load_dataset("ConvergeBio/uniclust30")

Data Processing

• Source: HH-suite ffindex/ffdata A3M database and uniref_mapping.tsv.gz from uniclust.mmseqs.com
• Parsing: Direct ffindex/ffdata binary reads; membership from mapping TSV
• Integrity: xxHash-128 computed per sequence; A3M representative sequence verified against extracted sequence
• Validation: Passed all tiers — schema conformance, zero null/empty sequences, xxHash roundtrip, A3M format checks, A3M–sequence consistency, member ID mapping verification, and field-by-field comparison against source ffindex/ffdata
• Format: Sharded Parquet with zstd compression

Source & Citation

UniClust30 is produced by the Söding Lab:

Mirdita M, von den Driesch L, Galiez C, Martin MJ, Söding J, Steinegger M. "Uniclust databases of clustered and deeply annotated protein sequences and alignments." Nucleic Acids Res. 45(D1):D170–D176 (2017). doi:10.1093/nar/gkw1081

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

UniClust30 data is available under CC BY-SA 4.0.