---
license: cc-by-4.0
task_categories:
  - fill-mask
  - feature-extraction
tags:
  - protein
  - proteins
  - biology
  - bioinformatics
  - protein-language-model
  - metagenomics
  - fasta
  - sequence-clustering
pretty_name: BFD (Big Fantastic Database) — FLAIR Preprocessed
size_categories:
  - 1B<n<10B
configs:
  - config_name: default
    data_files:
      - split: train
        path: "train/*.parquet"
---

# Dataset Card for `flair-bio/bfd`

## Dataset Summary

This dataset is a cleaned, deduplication-clustered, and quality-scored version of the
**Big Fantastic Database (BFD)** — a large metagenomic protein sequence database originally
assembled from metaclust and used as an MSA source in AlphaFold. It has been reprocessed by the
[FLAIR](https://github.com/flair-bio/amplify) `modules/data` pipeline into a single
training-ready Parquet dataset (sharded), with 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:** [BFD — `bfd_metaclust_clu_complete_id30_c90_final_seq`](https://bfd.mmseqs.com/)
- **Language(s):** Not applicable (protein sequences; FASTA header metadata is English free-text)
- **License:** [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) for the FLAIR-processed
  release.

## Dataset Details

### Source Data & Provenance

BFD is distributed as a single large tarball of pre-clustered sequences:

```
https://bfd.mmseqs.com/bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt.tar.gz
```

The tarball is downloaded with `aria2c` (16-way split, single concurrent download given there is
only one source file). It was fetched in **August 2026**, the most recent version available at
that time; BFD is distributed as a single unversioned static archive (no dated release identifier
is published upstream), so reproducibility relies on the file remaining unchanged at that URL.

### Processing Pipeline

Processed by the shared FLAIR data pipeline
([`modules/data`](https://github.com/flair-bio/amplify/tree/main/modules/data)):

```
Download → Preprocess → Cluster → Score → Assemble → Upload
```

1. **Preprocess** — BFD's tarball is streamed and only the `*_a3m.ffdata` member is extracted
   (the `cs219`/`hhm` profile members are skipped, as they are not raw sequence data). Alignment
   gaps, consensus rows, and database-match rows are stripped from the A3M records. The remaining
   sequences are linearized into a monolithic FASTA, 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 `BFD_000000000000` (a dataset prefix followed by a zero-padded 12-digit counter),
   with the original FASTA header 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 (`BFD_{12-digit counter}`) |
| `original_id` | string | Original BFD/A3M 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 (`bfd`) |
| `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 | 2,032,226,987 |
| Sequence length — min / mean / median / max | 49 / 178.0 / 142.0 / 25,830 |
| Sequence length — std | 107.96 |
| Total residues (sum of lengths) | 902,945,803 |
| RED score — mean / median / std | 0.0184 / 0.0153 / 0.0145 |
| RED score — min / max | 5.98e-05 / 0.924 |
| RED score — p1 / p5 / p25 / p75 / p95 / p99 | 0.0032 / 0.0042 / 0.0068 / 0.0259 / 0.0446 / 0.0642 |
| Ambiguous residues (count) | 811,636,585 |
| Sequences containing ambiguous residues | 708,960,173 (89.9%) |
| Cluster count @ `cluster_rep_at_30` / `_at_60` / `_at_90` | 294,546,612 / 446,346,719 / 1,439,636,576 |
| Source files | 2,439 |

## 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

- BFD is a metagenomic database; sequence provenance, quality, and annotation completeness vary
  widely and largely lack experimental validation.
- 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 BFD source, MMseqs2, RED, and FLAIR:

```bibtex
@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}
}
```

```bibtex
@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}
}
```

```bibtex
@misc{bfd,
  title        = {Big Fantastic Database (BFD)},
  howpublished = {\url{https://bfd.mmseqs.com/}},
}
```

```bibtex
@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}},
}
```