Welcome to AbAgym, a manually curated dataset containing 68 deep mutational scanning (DMS) datasets on antibody-antigen complexes, along with the 3D structure of each complex. The dataset includes around 324,000 non-redundant mutations, including 36,541 interface mutations, which can support the development and evaluation of computational methods for antibody design and immune escape prediction.
📄 For full details, please refer to our publication:
G. Cia, D. Li, S. Poblete, M. Rooman, F. Pucci "AbAgym: a well-curated dataset for the mutational analysis of antibody–antigen complexes. mAbs, 17(1)"
The table below describes the data and files contained in this repository. Additionaly, FoldX structures generated to perform the benchmark presented in the publication can be downloaded at https://zenodo.org/records/17328791 (FoldX licensing terms apply).
| File | Description |
|---|---|
AbAgym_data_full.csv.zip |
~572k redundant mutation data (i.e. all individual mutations, including those in identical chains of homo-oligomeric PDB structures) |
AbAgym_data_full_interface.csv |
Subset of 37,361 redundant mutations located at the antibody-antigen interface, which we used to benchmark computational prediction methods. Interface residues are determined using a 6 Å heavy atom distance threshold |
AbAgym_data_non-redundant.csv.zip |
~324k non-redundant mutation data (i.e. mutations in identical chains of homo-oligomeric chains combined into a single row) |
AbAgym_data_non-redundant_interface.csv |
Subset of 36,541 non-redundant mutations located at the antibody-antigen interface. Interface residues are determined using a 6 Å heavy atom distance threshold |
AbAgym_metadata.csv |
General information of each DMS dataset |
PDB_files.zip |
3D structures of the antibody-antigen complexes |
In the AbAgym data files, the score of each mutation is given by the DMS_score column. This score originates from deep mutational scanning (DMS) experiments and is therefore not a direct experimental measurement of binding affinity changes upon mutation. Although related, DMS scores and binding affinities are not equivalent, and DMS scores depend on the specific experimental assay used. As noted in our publication, different assay types produce scores with varying ranges. Some assays output scores between 0 and 1, where 0 indicates complete neutralization (no effect on binding) and 1 indicates full immune escape. Others produce scores ranging from negative to positive values: in these cases, 0 typically indicates no effect, positive scores indicate escape mutations, and negative scores suggest enhanced binding. If you plan to use this dataset to train a ML model, it could be necessary to normalize the scores across datasets to account for these differences; we provide two normalized DMS_score columns (MinMax and RankQuartile normalized). For details, please refer to our publication and Supplementary Material.
The dataset is freely accessible for non-commercial use only. For commercial applications (e.g., selling access to a prediction tool trained using AbAgym), please contact us (Fabrizio.Pucci@ulb.be).
Please cite our publication when using AbAgym in your research:
@article{cia2024abagym,
title = {AbAgym: a well-curated dataset for the mutational analysis of antibody-antigen complexes},
author = {Cia, G. and Li, D. and Poblete, S. and Rooman, M. and Pucci, F.},
journal = {submitted},
year = {2025}
}