Pharmaco-Explainer Datasets

This repository contains datasets used in the Pharmaco-Explainer project. They are shared separately on Hugging Face and are used by the training and experimentation code hosted on GitHub:

👉 Training code and scripts:
https://github.com/AdamSulek/pharmaco-explainer/

---

Available Datasets

The following datasets are available:

• k3: 3-element pharmacophore
• k4_2ar: 4-element pharmacophore with two aromatic features
• k4: 4-element pharmacophore
• k5 5-element pharmacophore

Each dataset consists of three files:

• <dataset>.parquet – main feature data
• <dataset>_split.parquet – predefined train/val/test split
• <dataset>_labels.parquet – labels

Repository structure:

k3/
├── k3.parquet
├── k3_split.parquet
└── k3_labels.parquet
k4_2ar/
├── k4_2ar.parquet
├── k4_2ar_split.parquet
└── k4_2ar_labels.parquet
k4/
├── k4.parquet
├── k4_split.parquet
└── k4_labels.parquet
k5/
├── k5.parquet
├── k5_split.parquet
└── k5_labels.parquet

Downloading the Data

Datasets can be downloaded directly from Hugging Face using the Python script below.

Requirements

• Python ≥ 3.8
• requests
• Environment variable PHARM_PROJECT_ROOT pointing to the project root

import argparse
from pathlib import Path
import requests
import os

def project_path(*parts):
    root = os.environ.get("PHARM_PROJECT_ROOT")
    if root is None:
        raise RuntimeError(
            "Environment variable PHARM_PROJECT_ROOT is not set.\n"
            "Run:\n"
            "   export PHARM_PROJECT_ROOT=/path/to/project"
        )
    return os.path.join(root, *parts)

def download_file(url, dest_path: Path):
    dest_path.parent.mkdir(parents=True, exist_ok=True)
    if dest_path.exists():
        print(f"File already exists, skipping: {dest_path}")
        return
    print(f"Downloading {url} -> {dest_path}")
    r = requests.get(url, stream=True)
    r.raise_for_status()
    with open(dest_path, "wb") as f:
        for chunk in r.iter_content(chunk_size=1024 * 1024):
            if chunk:
                f.write(chunk)

def download_dataset(dataset_name):
    base_url = (
        f"https://huggingface.co/datasets/"
        f"klimczakjakubdev/pharmaco-explainer/resolve/main/{dataset_name}/"
    )
    files = [
        f"{dataset_name}.parquet",
        f"{dataset_name}_split.parquet",
        f"{dataset_name}_labels.parquet",
    ]

    dest_dir = Path(project_path("data", dataset_name))
    for file_name in files:
        url = base_url + file_name
        dest_path = dest_dir / file_name
        download_file(url, dest_path)

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--dataset",
        type=str,
        default="k3",
        choices=["k3", "k4", "k4_2ar", "k5"],
        help="Which dataset to download"
    )
    args = parser.parse_args()

    download_dataset(args.dataset)

Example Usage

python download_dataset.py --dataset k4

Files will be downloaded to:

$PHARM_PROJECT_ROOT/data/k4/

Related Repository

These datasets are used by the main Pharmaco-Explainer codebase:

🔗 https://github.com/AdamSulek/pharmaco-explainer/

That repository contains:

• model training code
• experiment configurations
• preprocessing scripts
• evaluation and explainability pipelines