Oxford Protein Informatics Group
Published October 13, 2026 | Version 0.3.0
Dataset Open

SAbDab2 Machine Learning Dataset

Description

Clean, curated antibody and antibody–antigen-complex structures from SAbDab2, specifically post-processed for ML applications, alongside standardised train/test-splits. 
 
The current release contains 16415 structures from 8992 PDB IDs, corresponding to 5556 unique antibodies (SAbDab2 IDs), composed as follows:
 
antibody type with antigen no antigen total structures
paired-chain (FV-like)
8618 (52.5%)
4223 (25.7%)
12841 (78.2%)
single-domain heavy (VHH-like)
2156 (13.1%)
1271 (7.7%)
3427 (20.9%)
single-domain light (VL-like)
37 (0.2%)
26 (0.2%)
63 (0.4%)
VNAR
67 (0.4%)
17 (0.1%)
84 (0.5%)
total structures
10878 (66.3%)
5537 (33.7%)
16415 (100.0%)

Each structure file contains a single paired-chain or single-chain antibody, cropped to the variable region, alongside any antigen chains. Around 9.4% of structures bind antigens consisting of multiple polymer chains.

We provide two distinct train/test splits of these structures, both based on sequence similarity.
  • The ab-split (ab_split.csv) accounts for similarity between antibody sequences. To avoid data leakage via the antigen, use this only for antigen-agnostic tasks like prediction of antibody structure in solution.
  • The ab-ag-split (abag_split.csv) additionally considers similarity between any protein, peptide, DNA and RNA antigen sequences. Use this for antigen-aware tasks like antibody–antigen complex modelling.
These same splits are also available filtered down to just single-domain (VHH-like and VL-like) structures (ab_split_sd.csv and abag_split_sd.csv).
 
Unless otherwise noted, splits are backward-compatible, allowing models trained on previously published versions of this dataset to be benchmarked on the most recent test split. If you want to split the data by date, you can do so yourself using the relevant columns, though we strongly advise against it.
 
More details in the README.md accompanying this download.
 
How to Cite
 
If you use this data in your work, please cite the specific data release with its corresponding DOI (see right-hand "Versions" panel), alongside the SAbDab2 publication:

```bibtex
   author = {
       Capel, Henriette L and Vavourakis, Odysseas and Williams, Benjamin H and Taylor, Christopher R and Deane, Charlotte M
   },
   title = {SAbDab2: the structural antibody database in the age of machine learning},
   journal = {Nucleic Acids Research},
   pages = {gkag1091},
   year = {2026},
   month = {10},
   issn = {1362-4962},
   doi = {10.1093/nar/gkag1091},
   url = {https://doi.org/10.1093/nar/gkag1091},
   eprint = {https://academic.oup.com/nar/advance-article-pdf/doi/10.1093/nar/gkag1091/72639853/gkag1091.pdf}, }
```

Files

Files (958.7 MB)

Name Size
md5:75662fe08cb542177971fa3ded2cd800
958.7 MB Download

Additional details

Funding

Engineering and Physical Sciences Research Council
EPSRC CDT in Sustainable Approaches to Biomedical Science: Responsible and Reproducible Research - SABS:R^3 EP/S024093/1
AstraZeneca (United Kingdom)
Fusion Antibodies plc

Dates

Created
2026-05-08