Dataset Card for flair-bio/uniref

Dataset Summary

This dataset is a cleaned, deduplication-clustered, and quality-scored version of UniRef100 (UniProt Reference Clusters), a comprehensive, non-redundant-by-design database of protein sequences derived from UniProtKB and select UniParc records. It has been reprocessed by the FLAIR modules/data pipeline into a single training-ready Parquet dataset (sharded), with additional per-sequence redundancy reduction (MMseqs2 cascaded clustering) and quality scores (RED — Residue Embedding Diversity) added, for use in pretraining protein language models (pLMs).

• Curated by: FLAIR (Mila / flair-bio)
• Source data: UniRef100 (UniProt)
• Language(s): Not applicable (protein sequences; FASTA header metadata is English free-text)
• License: CC BY 4.0, consistent with UniProt's own licensing terms (UniProt license).

Dataset Details

Source Data & Provenance

UniRef100 is distributed as a single large gzipped FASTA file:

https://ftp.uniprot.org/pub/databases/uniprot/uniref/uniref100/uniref100.fasta.gz

The file is downloaded with aria2c (16-way split, single concurrent download given there is only one source file). It was fetched in August 2026, reflecting the most recent UniProt release available at that time; UniRef100 is updated periodically upstream, so later pulls of this URL may reflect a newer release.

Processing Pipeline

Processed by the shared FLAIR data pipeline (modules/data):

Download → Preprocess → Cluster → Score → Assemble → Upload

1. Preprocess — the source FASTA is decompressed and linearized, split into 1,000,000-sequence chunks, and converted to Parquet shards in parallel. Each sequence is assigned a unique, collision-free ID of the form UNIREF_000000000000 (a dataset prefix followed by a zero-padded 12-digit counter), with the original FASTA header (UniRef cluster ID
   • description) preserved as description.
2. Cluster — cascaded MMseqs2 (easy-linclust) redundancy reduction at identity thresholds [0.9, 0.8, 0.7, 0.6, 0.5, 0.4, 0.3] (coverage threshold 0.8), producing one cluster- representative column per threshold.
3. Score — each sequence is scored with the RED (Residue Embedding Diversity) metric using flair-bio/amplify-350m as the embedding model — a measure of how distinct/informative a sequence's contextual residue embeddings are.
4. Assemble — base sequences, cluster assignments, and RED scores are joined on the shared sequence_id into the final sharded Parquet dataset published here.

Dataset Structure

One train split, stored as multiple Parquet shards (train/*.parquet).

Column	Type	Description
sequence_id	string	Unique FLAIR-assigned ID (UNIREF_{12-digit counter})
original_id	string	Original UniRef100 cluster/sequence identifier
description	string	Remaining original FASTA header text
sequence	string	Protein sequence (amino acid string)
cluster_rep_at_90 … cluster_rep_at_30	string	Cluster representative ID at each identity threshold (90%, 80%, 70%, 60%, 50%, 40%, 30%)
dataset_id	string	Source dataset identifier stamp (uniref)
sequence_length	uint32	Sequence length in residues
red_score	float16	RED (Residue Embedding Diversity) quality score

Data Statistics

Computed over the full assembled dataset:

Statistic	Value
Number of sequences	220,919,788
Sequence length — min / mean / median / max	2 / 416.8 / 325.0 / 46,734
Sequence length — std	417.46
Total residues (sum of lengths)	1,894,698,506
RED score — mean / median / std	0.0509 / 0.0377 / 0.0537
RED score — min / max	6.97e-05 / 0.938
RED score — p1 / p5 / p25 / p75 / p95 / p99	0.0031 / 0.0051 / 0.0222 / 0.0620 / 0.1353 / 0.2705
Ambiguous residues (count)	7,792,429
Sequences containing ambiguous residues	1,100,225 (0.4%)
Cluster count @ cluster_rep_at_30 / _at_60 / _at_90	26,576,306 / 43,191,925 / 111,744,815
Source files	221

Uses

Intended for self-supervised pretraining (e.g., masked language modeling) and representation learning of protein language models. The cluster_rep_at_* columns allow downstream users to subsample at a desired redundancy level (e.g., dedup at 30% identity for maximum diversity, or 90% for near-duplicate-only removal), and red_score can be used to filter or weight sequences by estimated informational diversity.

Bias, Risks, and Limitations

• UniRef100 aggregates sequences of highly varying provenance and experimental validation, from fully characterized reviewed (Swiss-Prot) entries to unreviewed, automatically annotated (TrEMBL/UniParc) records; annotation confidence is not distinguished in this release.
• RED scores are a heuristic computed from a specific pLM's embeddings (flair-bio/amplify-350m) and should not be interpreted as a ground-truth quality or functional label.
• Clustering provides an approximate indication of protein relatedness. These datasets use MMseqs2 Linclust, and resulting clusters can vary depending on the clustering algorithm and its parameters.
• Clustering thresholds reduce but do not eliminate redundancy; sequences sharing a cluster representative at coarse thresholds (e.g., 30%) may still be highly similar within.

Citation

If you use this dataset, please cite the original UniProt/UniRef source, MMseqs2, RED, and FLAIR:

@article{steinegger2017mmseqs2,
  title   = {MMseqs2 enables sensitive protein sequence searching for the analysis of massive sequence datasets},
  author  = {Steinegger, Martin and S{\"o}ding, Johannes},
  journal = {Nature Biotechnology},
  volume  = {35},
  pages   = {1026--1028},
  year    = {2017},
  doi     = {10.1038/nbt.3988}
}

@article{lebreton2026plm,
  title   = {pLM representations unlock metagenomic space beyond homology},
  author  = {Le Breton, Lola and Heurtel-Depeiges, David and Millar, Douglas C. and Zetzsche, Lara E. and Vernon, Robert M. and Langmead, Christopher James and Chandar, Sarath and Fournier, Quentin},
  journal = {bioRxiv},
  year    = {2026},
  doi     = {10.64898/2026.07.28.739874}
}

@article{uniref,
  title   = {UniRef clusters: a comprehensive and scalable alternative for improving sequence similarity searches},
  author  = {Suzek, Baris E. and Wang, Yuqi and Huang, Hongzhan and McGarvey, Peter B. and Wu, Cathy H. and {the UniProt Consortium}},
  journal = {Bioinformatics},
  year    = {2015},
  doi     = {10.1093/bioinformatics/btu739}
}

@misc{flair-plm,
  title        = {FLAIR: Protein Language Model Pretraining Data Pipeline},
  author       = {{Applied Machine Learning Research Team (AMLRT) collaborators at Mila and FLAIR-bio research students}},
  howpublished = {\url{https://github.com/flair-bio/amplify/tree/main/modules/data}},
}