---
license: cc-by-4.0
pretty_name: PRISM Antibody Training Data
language:
  - en
tags:
  - biology
  - protein
  - antibody
  - immunology
  - OAS
size_categories:
  - 10M<n<100M
---

# PRISM Antibody Training Data

Training data for **PRISM** ([RomeroLab-Duke/prism-antibody](https://huggingface.co/RomeroLab-Duke/prism-antibody)),
a germline/non-germline-aware antibody language model. This repository contains the curated
unpaired and paired antibody sequence datasets and the exact OAS download manifest used to
build them.

## Files

| File | Rows | Description |
|------|------|-------------|
| `unpaired_anarci_relabeled.parquet` | 66,435,576 | Unpaired heavy/light chains (pretraining set) |
| `paired_anarci_relabeled.parquet` | 763,989 | Paired VH+VL antibodies (fine-tuning set) |
| `bulk_download.sh` | 14,433 URLs | Exact OAS download manifest used (see caveat below) |

Splits are encoded in the `split` column (`train` / `valid` / `test`).

## Provenance & pipeline

- **Source**: [Observed Antibody Space (OAS)](https://opig.stats.ox.ac.uk/webapps/oas/), human studies only
  (70 studies, 2021–2022 data freeze).
- **Pipeline**: download → QC filter (missing conserved cysteine / heavy fragmentation /
  non-canonical AA) → exact sequence deduplication → **NGL mutation filter** → CDR3 + whole-sequence
  identity clustering/dedup → ANARCI re-numbering.

## ⚠️ Important caveats (read before reproducing)

1. **This is NOT the full OAS human set.** Most of the largest study, **Briney_2019 heavy**
   (~1,100 files), was excluded from the unpaired set for download/disk size reasons
   (these lines are commented out in `bulk_download.sh`, and the loader skips `# wget` lines).
   This is why the unpaired heavy count is ~58M rather than the full OAS human heavy total.
2. **NGL (somatic) mutation filter**: sequences are kept only if
   `num_ngl_muts_hc > 3` (heavy) and `num_ngl_muts_lc > 2` (light). Applied to both sets.
3. **CDR3 clustering caveat**: the CDR3-identity clustering step used a fixed ANARCI CSV
   column window that can be shifted by IMGT insertion codes (affects heavy more than light).
   Treat `*_cdr3_cluster_id` / dedup counts with this in mind.

## Column dictionary

`HEAVY_CHAIN_AA_SEQUENCE`, `LIGHT_CHAIN_AA_SEQUENCE` (AA), `*_NT_SEQUENCE` (nucleotide),
`*_AA_GERMLINE_ALIGNMENT` (germline-aligned AA), `BType` (B-cell type), `Source_File` (OAS origin),
`hc_mut_codes` / `lc_mut_codes` (somatic mutation codes), `num_ngl_muts_hc` / `num_ngl_muts_lc` /
`total_num_ngl_muts` (somatic mutation counts), `*_cdr3_cluster_id` / `*_whole_seq_cluster_id`
(clustering IDs), `v_gene_*` / `j_gene_*` (V/J gene calls — decode with `gene_vocabulary.json`
in the model repo), `region_mask_*` (IMGT region masks), `split` (train/valid/test),
`oas_*_mut_codes` (original OAS mutation codes before ANARCI relabel).

## License & citation

Derived from OAS (CC-BY 4.0). Please cite OAS (Olsen et al. 2022, *Protein Science*) and PRISM.

Model: https://huggingface.co/RomeroLab-Duke/prism-antibody