Learning Protein-Ligand Binding in Hyperbolic Space

⚙ Installization

Clone the repository:

git clone https://github.com/jianhuiwemi/HypSeek.git
cd HypSeek

Install OpenBabel:

apt-get install -y openbabel

Install Python packages:

pip install numpy scikit-bio==0.6.2 rdkit biopandas

Install Uni-Core:

git clone https://github.com/dptech-corp/Uni-Core.git
cd Uni-Core
python setup.py install
cd ..

Install ProDy:

git clone https://github.com/prody/ProDy.git
cd ProDy
python setup.py build_ext --inplace --force
pip install -Ue .
cd ..

We also provide a full Conda environment file (environment.yml) for users who encounter dependency issues.

🚀 Quick Start

HypSeek can be directly evaluated using the provided test.sh script.

⚡ Run Virtual Screening

Use the Screening checkpoint (checkpoint_avg_41-50_vs.pt) for:

DUD-E

bash test.sh DUDE three_hybrid_model /path/checkpoint_avg_41-50_vs.pt ./results

LIT-PCBA

bash test.sh PCBA three_hybrid_model /path/checkpoint_avg_41-50_vs.pt ./results

⚡ Run Affinity Ranking

Use the Ranking checkpoint (checkpoint_avg_41-50_rk.pt) for FEP:

bash test.sh FEP three_hybrid_model /path/checkpoint_avg_41-50_rk.pt ./results

🏋️ Training

We provide a unified training script (train.sh). You may train either the Virtual Screening (VS) model or the Affinity Ranking (RK) model depending on the validation set used.

🔥 Train Virtual Screening Model

Use the CASF validation set:

bash train.sh CASF

🔥 Train Affinity Ranking Model

Use the FEP validation set:

bash train.sh FEP

📚 Citation

If you find this work useful in your research, please cite:

@misc{wang2025learningproteinligandbindinghyperbolic,
      title={Learning Protein-Ligand Binding in Hyperbolic Space}, 
      author={Jianhui Wang and Wenyu Zhu and Bowen Gao and Xin Hong and Ya-Qin Zhang and Wei-Ying Ma and Yanyan Lan},
      year={2025},
      eprint={2508.15480},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2508.15480}, 
}

💐 Acknowledgments

This work builds upon Uni-Mol, Uni-Core, and LigUnity. We especially thank the LigUnity team for providing their data resources, and we thank all authors for their open-source contributions.

📬 Contact

For any questions or collaboration requests, please contact Jianhui: jianhuiwang309@gmail.com