# DrugRNA Dataset

This repository contains the RNA-small molecule interaction prediction dataset used in the paper "LLM Agents Enable RNA–Small Molecule Interaction Prediction at Performance Comparable to Human-Designed Models".

## Dataset Description

The dataset is derived from the RNAInter repository and contains experimentally validated RNA-compound interactions. It includes 45,049 RNA–compound pairs spanning 3,258 unique RNAs and 345 unique compounds, with a 1:4 positive/negative ratio.

## Data Splits

The dataset is organized into three evaluation scenarios:

### 1. In-Distribution Split (`in_distribution/`)
- **Purpose**: Standard 80/10/10 train/validation/test split for head-to-head performance comparison
- **Characteristics**: Complete entity overlap, testing interpolation within known molecular space
- **Files**:
  - `train_text.csv` (36,040 samples, 20.0% positive)
  - `val_text.csv` (4,504 samples, 19.8% positive)
  - `test_text.csv` (4,505 samples, 20.2% positive)
- **Total**: 45,049 samples with 3,258 unique RNAs and 345 unique compounds

### 2. Full Out-of-Domain Split (`full_ood/`)
- **Purpose**: Cold-start prediction with entirely novel entities
- **Characteristics**: Zero overlap for both RNAs and compounds between training and test sets
- **Files**:
  - `train_text.csv` (21,794 samples, 18.4% positive)
  - `val_text.csv` (1,081 samples, 22.1% positive)
  - `test_text.csv` (1,020 samples, 24.2% positive)
  - `DATASET_REPORT.md` (detailed statistics and construction methodology)
- **Total**: 23,895 samples with 3,142 unique RNAs and 363 unique compounds

### 3. Compound Out-of-Domain Split (`compound_ood/`)
- **Purpose**: Virtual screening of novel compounds against known RNA targets
- **Characteristics**: Zero compound overlap, complete RNA reuse (86% of training RNAs appear in test)
- **Files**:
  - `train_text.csv` (31,576 samples, 21.0% positive)
  - `val_text.csv` (6,694 samples, 20.1% positive)
  - `test_text.csv` (6,779 samples, 15.2% positive)
  - `DATASET_REPORT.md` (detailed statistics and construction methodology)
- **Total**: 45,049 samples with 3,258 unique RNAs and 345 unique compounds

## Data Format

Each CSV file contains the following columns:
- `rna_id`: RNA identifier
- `compound_id`: Compound identifier (PubChem CID)
- `rna_sequence`: RNA nucleotide sequence (A, U, G, C)
- `compound_smiles`: SMILES string representation of the compound
- `label`: Binary interaction label (1 = interaction, 0 = no interaction)

## Scripts

The `scripts/` folder contains:
- `create_disjoint_splits.py`: Script used to generate the out-of-domain splits with strict entity disjointness

## Data Sources

- **RNAs**: Mapped from NCBI, Ensembl, miRBase, and circBase
- **Compounds**: Mapped from PubChem
- **Interactions**: Curated from RNAInter repository with experimental evidence

## Citation

If you use this dataset in your research, please cite:

```bibtex
@article{drugrna2025,
  title={LLM Agents Enable RNA–Small Molecule Interaction Prediction at Performance Comparable to Human-Designed Models},
  author={...},
  journal={...},
  year={2025}
}
```

## License

This dataset is provided for research purposes. Please refer to the original data sources for specific licensing information.

## Contact

For questions or issues, please open an issue on the GitHub repository or contact the authors.