Protein Sequences · LiteFold
UniRef50
UniProt reference cluster representatives at 50% sequence identity for protein machine learning and sequence searches.
What UniRef50 is
UniRef50 is the most condensed tier of the UniProt Reference Clusters, which also includes UniRef100 and UniRef90. It provides a non-redundant catalogue of protein sequences derived from the UniProt Knowledgebase (UniProtKB) and selected UniParc records, clustered to a maximum of 50% sequence identity.
Raw protein databases are heavily skewed towards frequently studied organisms and protein families. By reducing sequence redundancy, UniRef50 produces a substantially flatter representation of sequence space. This makes it a standard corpus for pretraining protein language models, including ESM and ProtTrans, and for sequence searching in structural prediction systems such as AlphaFold2.
Packaged by LiteFold, this dataset release contains 60,315,044 cluster representative sequences comprising 17,282,055,793 amino acid residues, divided into compressed sequence shards accompanied by per-record metadata.
How it was made
UniRef50 is constructed by clustering UniRef90 cluster representatives using the MMseqs2 algorithm. Clusters are formed using a threshold of 50% sequence identity and an 80% sequence length overlap.
The dataset distribution comprises 61 Zstandard-compressed FASTA sequence shards and corresponding per-record JSONL metadata files. The primary repository index is organised into Parquet tables that define a deterministic split across files using sha256(file_id) % 10, assigning bucket 0 to testing and buckets 1 to 9 to training.
Key facts
- Publisher
- LiteFold
- Total records
- 60,315,044
- Total residues
- 17,282,055,793
- Clustering algorithm
- MMseqs2
- Clustering thresholds
- 50% sequence identity, 80% length overlap
- Source data
- UniRef90 representatives (UniProtKB and selected UniParc records)
- Sequence shards
- 61
- File formats
- Parquet (index), Zstandard-compressed FASTA (.fasta.zst), JSONL (metadata)
What you can use it for
- Protein Language Model Pretraining
- Provides a diverse, de-duplicated sequence corpus used to train large-scale protein language models such as ESM and ProtTrans.
- Protein Structure Prediction Pipelines
- Serves as an input database for sequence search and multiple sequence alignment generation in systems like AlphaFold2.
- Broad Sequence Similarity Searches
- Enables rapid sequence homology searches across the protein universe with significantly reduced search spaces compared to full UniProtKB.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from LiteFold/UniRef50 at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 60 rows of default/train
access_note
string_label- File/shard index for UniRef50; download sequence shards for FASTA records.100%
compression
string_label- zstd92%
8.3% empty
download_pattern
string_label- sequences/sequence_uniref50_uniref50.fasta.gz/shard-*.fasta.zst92%
- metadata/sequence_uniref50_uniref50.fasta.gz.records.jsonl2%
- manifests/sequence_uniref50_uniref50.fasta.gz.json2%
file_id
string_text- min
- 9
- median
- 68
- mean
- 64.75
- max
- 68
is_metadata_records
bool- False98%
- True2%
is_sequence_shard
bool- True92%
- False8%
path
string_text- min
- 9
- median
- 68
- mean
- 64.75
- max
- 68
records_total
int- min
- 60M
- median
- 60M
- mean
- 60M
- max
- 60M
repo_id
string_label- LiteFold/UniRef50100%
residues_total
int- min
- 17.3B
- median
- 17.3B
- mean
- 17.3B
- max
- 17.3B
role
string_label- sequence_shard92%
- readme2%
- aggregate_manifest2%
shard_index
float- min
- 0
- median
- 30
- mean
- 29.91
- max
- 60
8.3% empty
size_bytes
int- min
- 462
- median
- 131M
- mean
- 520M
- max
- 21.0B
source_sha
string_label- 45f6d1f6eb68c847f89b719d7f1da3e7fa9e6611100%
source_slug
string_label- sequence_uniref50_uniref50.fasta.gz100%
split_bucket
int- min
- 1
- median
- 5
- mean
- 5.27
- max
- 9
total_shards
int- min
- 61
- median
- 61
- mean
- 61
- max
- 61
Strengths and limits
Good at
- Substantial redundancy reduction yields a balanced sample of protein sequence space.
- Provides a deterministic train and test split based on cryptographic file hashing.
- Structured index allows selective downloading and streaming of compressed FASTA shards and metadata.
Watch out for
- Aggressive clustering at 50% sequence identity removes fine-grained sequence variations and close homologs.
- The main Parquet index contains file-level metadata rather than individual sequence rows.
Questions
Can I look inside UniRef50 without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is UniRef50 under?
Its authors declare cc-by-4.0. Read the terms on the original page before reusing the data.
How is UniRef50 generated?
It is created by taking UniRef90 representative sequences and clustering them with MMseqs2 at a 50% sequence identity threshold and an 80% length overlap.
What machine learning models make use of UniRef50?
It is widely used as a training corpus and search resource for structural prediction networks like AlphaFold2 and protein language models like ESM and ProtTrans.
How are the training and test splits partitioned?
The split is computed deterministically from the file path using sha256(file_id) % 10, placing bucket 0 into the test set (6 index rows) and buckets 1 to 9 into the train set (60 index rows).
How are the sequence files stored and compressed?
Sequences are stored across 61 FASTA shards compressed using Zstandard (.fasta.zst), complemented by per-record JSONL metadata.
How to cite it
Suzek, B. E., Wang, Y., Huang, H., McGarvey, P. B., Wu, C. H., & UniProt Consortium (2015). UniRef clusters: a comprehensive and scalable alternative for improving sequence similarity searches. Bioinformatics, 31(6), 926–932. https://doi.org/10.1093/bioinformatics/btu739
Written from UniRef50's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.