CoBALT Protein Interaction Graphs

This dataset contains the precomputed antigen and antibody graph files used by CoBALT: Coupling Antigen–Antibody Binding-Site Prediction and Affinity Ranking through a Contact-Energy Bottleneck. The files are intended for reproducing the graph-based experiments described in the paper and for developing models for antigen–antibody binding-site prediction and binding-affinity ranking.

The dataset contains serialized Python/PyTorch graph objects. It does not contain the original protein structures, language-model checkpoints, or the training code. The corresponding model and evaluation code are available in the CoBALT GitHub repository.

Repository contents

File	Description	Approximate size
graph_antibody.pkl	Precomputed antibody graph collection used by the binding-affinity ranking task	13 GB
graph_antigen.pkl	Precomputed antigen graph collection used by the binding-affinity ranking task	26 GB
graph.pkl	Precomputed graph collection used by the binding-site prediction task	36 GB

The BAR affinity table, BAR five-fold split files, BAR chunk-bank manifest, and BSP five-fold split file are distributed separately with the CoBALT reproduction package. They are required to run the complete five-fold experiment.

Data format

The files use Python pickle serialization and are loaded by the CoBALT data loader. The stored objects contain graph data and the node, edge, geometric, and protein-language-model features required by the model. The exact graph schema is defined by the loader in the CoBALT source repository.

Because these files use pickle serialization, load them only from a trusted source. Do not unpickle arbitrary files in a security-sensitive environment.

Loading the files

The following example downloads a file from a Hugging Face dataset repository and loads it with PyTorch:

from pathlib import Path

import torch
from huggingface_hub import hf_hub_download


repo_id = "JosieFanXU/CoBALT"
local_file = hf_hub_download(
    repo_id=repo_id,
    repo_type="dataset",
    filename="BAR/graph_antigen.pkl",
)

graphs = torch.load(Path(local_file), map_location="cpu")
print(type(graphs))

Replace filename with the relevant file path when downloading another file. The files are large, so ensure that the local storage and memory available to the loading process are sufficient.

Intended use

The data are provided for:

• reproducing the CoBALT binding-site prediction and affinity-ranking experiments;
• developing graph-based methods for antigen–antibody interactions;
• evaluating alternative interaction encoders and task-specific decoders.

Limitations

This dataset is intended for research use only. It is not approved for clinical diagnosis, treatment selection, or other medical applications.

Citation

If you use these files, please cite the CoBALT paper and the associated software repository:

@article{cobalt,
  title   = {CoBALT: Coupling Antigen--Antibody Binding-Site Prediction and Affinity Ranking through a Contact-Energy Bottleneck},
  author  = {Fan Xu},
  journal = {Preprint},
  year    = {2026}
}

Please also cite the original sources of any structures, protein language models, and graph-construction tools used to generate the released files.