README.md

---
license: cc-by-4.0
task_categories:
  - text-generation
  - other
task_ids:
  - text2text-generation
tags:
  - biology
  - protein
  - antibody
  - antigen
  - drug-discovery
  - therapeutic-design
  - bioinformatics
  - structural-biology
size_categories:
  - 1K<n<10K
language:
  - en
pretty_name: AbDes - Antibody Description Dataset
dataset_info:
  features:
    - name: heavy_chain
      dtype: string
    - name: light_chain
      dtype: string
    - name: antigen
      dtype: string
    - name: description
      dtype: string
    - name: drugbank_description
      dtype: string
    - name: delta_g
      dtype: float64
    - name: pdb_id
      dtype: string
  splits:
    - name: train
      num_examples: 7800
---

# AbDes: Antibody Description Dataset

## Dataset Description

**AbDes** (Antibody Description Dataset) is a comprehensive multi-modal dataset designed for text-based
antigen-conditioned antibody redesign. It unifies antibody-antigen structural information with rich textual
descriptions, enabling novel AI-driven approaches to therapeutic antibody discovery.

### Key Features

- **7,684 complete antibody-antigen pairs** with heavy chain, light chain, and antigen sequences
- **Rich textual descriptions** derived from PDB annotations including experimental details, source organisms,
  classification, and biological context
- **Extended Description from DrugBank**: Inclusion of 871 samples with additional detailed textual information about
  the corresponding antibodies, offering deeper context for advanced analysis.
- **Binding affinity measurements** (ΔG values) for 411 entries from SAbDab
- **Multi-modal integration** combining sequences, structures, and natural language descriptions

## Associated Paper

**TeBaAb: Text-Based Antigen-Conditioned Antibody Redesign via Directed Evolution**

*Project Page / Source Code*: https://github.com/HySonLab/TeBaAb

This dataset was created to support the TeBaAb framework, which combines:

- Conditional Variational Autoencoder (CVAE) jointly conditioned on antigen sequences and textual descriptions
- Two-stage binding affinity predictor
- Iterative enrichment loop inspired by directed evolution

**Key Results**: TeBaAb improves predicted binding affinity by an average of 10% compared to original antibodies while
preserving structural confidence (RMSPE < 1.0Å).

## Dataset Structure

### Data Fields

- **`heavy_chain`** (string): Complete amino acid sequence of the antibody heavy chain
- **`light_chain`** (string): Complete amino acid sequence of the antibody light chain
- **`antigen`** (string): Target antigen sequence(s). Multiple antigen fragments are separated by `/` following SAbDab
  conventions
- **`description`** (string): Free-text description of antibody properties including:
    - Protein classification
    - Source organism
    - Expression system
    - Experimental details
    - Structural characteristics
    - Symmetry information
- **`drugbank_description`** (string, optional): Extended, detailed textual description of the antibody collected from
  the DrugBank database. Available for a subset of 871 entries.
- **`delta_g`** (float, optional): Binding free energy in kJ/mol (lower values indicate stronger binding). Available for
  411 entries
- **`pdb_id`** (string): Protein Data Bank identifier for accessing structural data

### Data Splits

The dataset contains:

- **`train`**: 7,684 complete antibody-antigen pairs with sequences and descriptions

## Dataset Creation

### Source Data

AbDes was curated from multiple authoritative sources:

- **AbSet**: Foundation dataset providing antibody structures and molecular descriptors
- **Protein Data Bank (PDB)**: Structural annotations and experimental metadata
- **SAbDab** (Structured Antibody Database): Binding affinity measurements (ΔG values)
- **DrugBank**: Source for extended textual descriptions related to the therapeutic context of the antibodies.

### Curation Process

1. **Filtering**: Selected 7,684 complete entries from AbSet with paired heavy/light chains and antigen information
2. **Text annotation**: Systematically collected descriptive metadata from PDB entries using PDB IDs
3. **Text annotation (DrugBank)**: Integrated a set of 871 extended descriptions from DrugBank using PDB ID mapping and
   cleaned the text for direct use.
4. **Affinity integration**: Merged experimentally determined ΔG values from SAbDab for applicable entries
5. **Quality control**: Validated sequence completeness and description quality

## Usage

### Loading the Dataset

```python
from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("hysonlab/abdes")

```

### Basic Usage Example

```python
# Iterate through training examples
for example in dataset["train"]:
    heavy = example["heavy_chain"]
    light = example["light_chain"]
    antigen = example["antigen"]
    description = example["description"]

    # Process sequences and text
    print(f"Description: {description}")
    print(f"Binding affinity: {example.get('delta_g', 'N/A')}")
```

### Use Cases

1. **Text-conditioned antibody design**: Train generative models that incorporate natural language descriptions
2. **Binding affinity prediction**: Develop models to predict antibody-antigen binding strength
3. **Multi-modal learning**: Research at the intersection of protein sequences and natural language
4. **Therapeutic antibody discovery**: Accelerate drug development through AI-driven design
5. **Structural bioinformatics**: Study relationships between sequence, structure, and function

### Citation

If you use this dataset, please cite:

```bibtex
@article{tebaab2025,
  title={TeBaAb: Text-Based Antigen-Conditioned Antibody Redesign via Directed Evolution},
  author={[Your Name(s)]},
  journal={[Journal/Conference]},
  year={2025}
}
```

## Dataset Statistics

| Statistic                  | Value                                |
|----------------------------|--------------------------------------|
| Total entries              | 7,684                                |
| Entries with ΔG            | 411                                  |
| Entries with DrugBank Desc | 871                                  |
| Average heavy chain length | 121.22                               |
| Average light chain length | 107.95                               |
| Average description length | 294.60                               |
| Unique PDB entries         | 7,684                                |
| Source organisms           | Multiple (primarily mouse, human)    |
| Expression systems         | Multiple (E. coli, mammalian, yeast) |

## License

This dataset is released under the **CC-BY-4.0** license. Users are free to share and adapt the material with
appropriate credit to the original sources (PDB, SAbDab, AbSet, DrugBank).

## Acknowledgments

We acknowledge the following resources:

- **Protein Data Bank (PDB)** for structural and annotation data
- **SAbDab** for binding affinity measurements
- **AbSet** for the foundational antibody-antigen dataset
- **DrugBank** for providing extended drug/antibody description metadata.

## Contact

For questions, issues, or collaborations,
please [open an issue](https://huggingface.co/datasets/HySonLab/AbDes/discussions/new) or
contact [dodo5](s.cuongnm5@gmail.com).

## Version History
- **v1.1** (2025-11): Added drugbank_description feature with 871 extended entries.
- **v1.0** (2025-11): Initial release with 7,684 entries