RomeroLab-Duke/prism2-antibody-dataset

PRISM2 Antibody Dataset

Private dataset backing the PRISM2 family of antibody language models trained at the Romero Lab (Duke University). Bundles the OAS immune-repertoire and SAbDab structural corpora used for the v6 / v7 unified VDJ runs, plus the V/D/J gene vocabulary files.

All sequence data has been pre-filtered for the production germline-distance threshold used in PRISM2 training, so what you load here is what the model trained on — no extra filtering required.

Layout

sabdab/
  04a_cased.parquet         IMGT-numbered, case-annotated antibodies
  04c_clustered.parquet     30%-identity clustered Phase 3 origin set (6,158 rows)
  prism2_sabdab.parquet     Final processed SAbDab for PRISM2

gene_vocab/
  gene_vocabulary.json      Combined V/D/J vocabulary (legacy)
  gene_vocab_v.json         V-gene allele → integer
  gene_vocab_d.json         D-gene allele → integer
  gene_vocab_j.json         J-gene allele → integer

paired/                     OAS paired heavy+light, filtered (heavy ngl≥3, light ngl≥2)
  train.parquet
  val.parquet
  test.parquet

unpaired/                   OAS unpaired, filtered (heavy ngl≥3 OR light ngl≥2)
  train/
    heavy/                  chunked, 7,334 files
    light/                  chunked, 1,291 files
  val/
    heavy/                  chunked, 1,058 files
  test/
    heavy/                  chunked
    light/                  chunked

Filter rationale

The PRISM2 training objective relies on somatic-hypermutation signal; reads indistinguishable from naive (germline-identical) sequences contribute trivial targets to the SHM loss. We therefore drop them upstream:

• Heavy chains: keep only sequences with germline-distance num_ngl_muts ≥ 3.
• Light chains: keep only sequences with num_ngl_muts ≥ 2.

The paired/ split applies both thresholds (a row is kept iff its heavy chain has ngl≥3 AND its light chain has ngl≥2). The unpaired/ split is per-chain — each row is a single chain and is kept if it passes its own threshold.

Chunking convention

Unpaired sets are stored as *_rg{NNNNN}.parquet chunks of roughly 100 K rows each, SNAPPY compressed. The data loader expands the glob and shuffles at the chunk level, which keeps memory bounded while preserving cross-chunk randomness.

Split	Chain	Files
train	heavy	7,334
train	light	1,291
val	heavy	1,058
val	light	— (not generated upstream)
test	heavy	1,312
test	light	361

The unpaired test chunks were re-created from the upstream concat using the same filter so they match train/val layout. val/light was absent from the upstream pipeline and is not present here either.

Loading

from datasets import load_dataset

# SAbDab structural antibodies
sabdab = load_dataset("RomeroLab-Duke/prism2-antibody-dataset", "sabdab", token=True)

# OAS paired (filtered)
paired = load_dataset("RomeroLab-Duke/prism2-antibody-dataset", "paired", token=True)

# OAS unpaired heavy (chunked — streaming recommended)
heavy = load_dataset(
    "RomeroLab-Duke/prism2-antibody-dataset",
    "unpaired_heavy",
    streaming=True,
    token=True,
)

Source attribution

• OAS — Olsen, T. H., Boyles, F., & Deane, C. M. (2022). Observed Antibody Space: A diverse database of cleaned, annotated, and translated unpaired and paired antibody sequences. Protein Science. Downloaded via the official bulk-CSV index and re-annotated with ANARCI (IMGT scheme) before deduplication, clustering, and the germline filter above.
• SAbDab — Dunbar, J., Krawczyk, K., Leem, J., Baker, T., Fuchs, A., Georges, G., Shi, J., & Deane, C. M. (2014). SAbDab: the structural antibody database. Nucleic Acids Research. Filtered to non-redundant structurally resolved antibodies with paired heavy/light annotations.

License

CC BY-NC 4.0 (research use). OAS itself is CC BY 4.0; SAbDab inherits the underlying PDB licensing. Commercial use must clear both upstream sources.

Citation

If you use this dataset, please cite the PRISM2 work and the upstream OAS / SAbDab references above.