Datasets:

flair-bio
/

oas

license: cc-by-4.0
task_categories:
  - fill-mask
  - feature-extraction
tags:
  - protein
  - proteins
  - antibody
  - immunology
  - biology
  - bioinformatics
  - protein-language-model
  - fasta
  - sequence-clustering
pretty_name: OAS (Observed Antibody Space, Paired) — FLAIR Preprocessed
size_categories:
  - 1M<n<10M
configs:
  - config_name: default
    data_files:
      - split: train
        path: train/*.parquet

Dataset Card for `flair-bio/oas`

Dataset Summary

This dataset is a cleaned and quality-scored version of the paired heavy/light chain subset of the Observed Antibody Space (OAS) database (Oxford Protein Informatics Group), a large repository of Next-Generation Sequencing (NGS) antibody repertoires. It has been reprocessed by the FLAIR modules/data pipeline into a single training-ready Parquet dataset (sharded — one shard per source study/run), with per-sequence redundancy reduction (MMseqs2 cascaded clustering) and quality scores (RED — Residue Embedding Diversity) added, for use in pretraining antibody/protein language models.

Curated by: FLAIR (Mila / flair-bio)
Source data: Observed Antibody Space (OAS), paired sequences (Oxford Protein Informatics Group)
Language(s): Not applicable (protein sequences; header/metadata text is English)
License: CC BY 4.0 for the FLAIR-processed release, consistent with the original OAS terms; please review OAS's terms of use for the underlying source data and constituent studies.

Dataset Details

Source Data & Provenance

Paired heavy/light antibody repertoire CSVs are downloaded directly from OPIG's OAS web service, covering multiple published studies (e.g. Alsoiussi 2020, Eccles 2020, Goldstein 2019, Jaffe 2022), one gzipped CSV per sequencing run:

https://opig.stats.ox.ac.uk/webapps/ngsdb/paired/<Study_Year>/csv/<Run_ID>_paired.csv.gz

Each file is downloaded with aria2c (16-way split per file, single concurrent download). OAS is a continuously growing database; the studies/runs included here were fetched in August 2026, reflecting the most recent OAS snapshot available at that time — later OAS pulls may include additional studies not present in this release.

Processing Pipeline

Processed by the shared FLAIR data pipeline (modules/data), with an OAS-specific preprocessing path:

Download → Preprocess → Cluster → Score → Assemble → Upload

Preprocess — unlike the FASTA-based datasets, OAS is read directly from the paired CSVs with pandas. Rows where ANARCI flags either chain as "Shorter" are dropped. Each retained pair is written as a FASTA record (heavy_seq + XXXXX linker + light_seq, used for statistics only) and, since the source already provides structured columns, the base Parquet shard is written directly from the CSV (one shard per source run), bypassing the generic FASTA→Parquet conversion. Each pair is assigned a unique, collision-free ID of the form OAS_000000000000 (a dataset prefix followed by a zero-padded 12-digit counter).
Cluster — cascaded MMseqs2 (easy-linclust) redundancy reduction at identity thresholds [0.99, 0.98, 0.97, 0.96, 0.95, 0.94, 0.93, 0.92, 0.91, 0.9, 0.85, 0.8] (coverage threshold 0.95) — notably finer-grained and higher-identity than the other FLAIR datasets, reflecting the naturally lower diversity/higher similarity of paired antibody sequences.
Score — each paired sequence is scored with the RED (Residue Embedding Diversity) metric using flair-bio/amplify-350m as the embedding model.
Assemble — base sequences, cluster assignments, and RED scores are joined on the shared sequence_id into the final sharded Parquet dataset published here (one output shard per source run).

Dataset Structure

One train split, stored as multiple Parquet shards (train/*.parquet, one shard per source sequencing run).

Column	Type	Description
`sequence_id`	string	Unique FLAIR-assigned ID (`OAS_{12-digit counter}`)
`original_id`	string	Original OAS pair identifier
`description`	string	Remaining original header text
`sequence`	string	Paired sequence: `heavy_seq` + `XXXXX` linker + `light_seq`
`cluster_rep_at_99` … `cluster_rep_at_80`	string	Cluster representative ID at each identity threshold (99%, 98%, 97%, 96%, 95%, 94%, 93%, 92%, 91%, 90%, 85%, 80%)
`dataset_id`	string	Source dataset identifier stamp (`oas`)
`sequence_length`	uint32	Combined paired sequence length in residues (including linker)
`red_score`	float16	RED (Residue Embedding Diversity) quality score

Data Statistics

Computed over the full assembled dataset:

Statistic	Value
Number of sequences (pairs)	1,887,487
Sequence length — min / mean / median / max	212 / 232.6 / 232.0 / 277
Sequence length — std	4.54
Total residues (sum of lengths)	439,101,472
RED score — mean / median / std	0.0858 / 0.0825 / 0.0287
RED score — min / max	0.0327 / 0.734
RED score — p1 / p5 / p25 / p75 / p95 / p99	0.0600 / 0.0666 / 0.0756 / 0.0901 / 0.1065 / 0.1704
Ambiguous residues (count)	0
Sequences containing ambiguous residues	0 (0.0%)
Cluster count @ `cluster_rep_at_80` / `_at_90` / `_at_99`	1,117,924 / 1,559,067 / 1,787,995

Uses

Intended for self-supervised pretraining (e.g., masked language modeling) and representation learning of antibody/protein language models, particularly modeling heavy/light chain co- occurrence. The cluster_rep_at_* columns allow downstream users to subsample at a desired redundancy level, and red_score can be used to filter or weight sequences by estimated informational diversity.

Bias, Risks, and Limitations

OAS aggregates repertoires from multiple studies/donors/species; sampling depth, immunization status, and health/disease state vary by source study and are not exhaustively normalized here.
The XXXXX linker is an artificial separator introduced for sequence-level modeling and stats computation; it is not a biological sequence.
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.
Antibody sequences are naturally low-diversity relative to general proteomes; even at the finest clustering threshold (99%), many near-identical sequences may remain.

Citation

If you use this dataset, please cite the original OAS 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{oas,
  title   = {Observed Antibody Space: A diverse database of cleaned, annotated, and translated unpaired and paired antibody sequences},
  author  = {Olsen, Tobias H. and Boyles, Fergus and Deane, Charlotte M.},
  journal = {Protein Science},
  year    = {2022},
  doi     = {10.1002/pro.4205}
}

@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}},
}

Dataset Card for flair-bio/oas