Dynamics-PLI

Official PyTorch-based implementation of the paper:
"Molecular dynamics-powered hierarchical geometric deep learning framework for protein-ligand interaction"
[IEEE Xplore] | [DOI]

In this work, we introduce Dynamics-PLI, a SO(3)-equivariant hierarchical graph neural network (EHGNN) designed to capture the intrinsic hierarchy of biomolecular structures. This framework leverages molecular dynamics simulations to enhance the prediction of protein-ligand interactions (PLIs), providing a powerful tool for computational drug discovery.

📦 Installation

git clone https://github.com/yourusername/Dynamics-PLI.git
cd Dynamics-PLI
conda env create -f environment.yml
conda activate Dynamics-PLI

🔮 Dataprocess

download

The following datasets and tools were utilized in the development and evaluation of our model:

• MISATO
  A multi-scale integrative tool for protein structure analysis, particularly useful for studying conformational changes and interactions.
  🔗 Zenodo Record

• Atom3D
  A collection of benchmark datasets for machine learning on 3D molecular structures, designed to facilitate research in structural biology.
  🔗 Official Website

MD_data_process

python ./datasets/pyg/MISATO.py

downstream_data_process

python ./datasets/pyg/pdbbind.py
python ./datasets/pyg/lep.py

🚀 Quick Start

python lba_main_count.py --dim 128 --dropout 0.15  --cross_cutoff 20 --local_cutoff 10 --depth_local 6 --depth_cross 2  --seed 4 --epochs 10 --atoms_bacth 4000 --lr 0.00005

📁 Project Structure

Dynamics-PLI/
├── baseline/            # baseline methods
├── datasets/            # Dataset loaders and preprocessors
├── Model/              # Model
├── util_mine.py/        # Training utilities, metrics, logging
├── lba_main_count.py    # lba task scripts
├── lep_main_count.py    # lep task scripts
├── pretrain.py          # pretrain scripts
└── README.md            # Project documentation

📊 Results

↑: Higher is better, ↓: Lower is better.

Pre-trained Model	LBA (30%) RMSE ↓	Pearson ↑	Spearman ↑	LEP AUROC ↑	AUPRC ↑
EGNN + Pre [[Satorras et al., 2021]]	1.341 ± 0.027	0.617 ± 0.016	0.613 ± 0.013	0.759 ± 0.010	0.752 ± 0.023
Dynamics-PLI	1.312 ± 0.005	0.636 ± 0.005	0.633 ± 0.008	0.779 ± 0.007	0.766 ± 0.031
Δ (%)	↓ 4.0%	↑ 3.5%	↑ 4.6%	↑ 3.7%	↑ 4.2%

Full benchmarking details are available in the paper.

📚 Reference

If you use Dynamics-PLI in scholarly publications, presentations, or to communicate with your satellite, please cite the following work:

@ARTICLE{10955744,
  author={Liu, Mingquan and Jin, Shuting and Lai, Houtim and Wang, Longyue and Wang, Jianmin and Cheng, Zhixiang and Zeng, Xiangxiang},
  journal={IEEE Transactions on Computational Biology and Bioinformatics}, 
  title={Molecular dynamics-powered hierarchical geometric deep learning framework for protein-ligand interaction}, 
  year={2025},
  volume={},
  number={},
  pages={1-12},
  keywords={Proteins;Atoms;Deep learning;Three-dimensional displays;Graph neural networks;Feature extraction;Drugs;Computational modeling;Training;Representation learning;Protein-ligand interactions;Molecular dynamics;Geometric deep learning;Pre-training},
  doi={10.1109/TCBBIO.2025.3558959}
}

🤝 Acknowledgements

Acknowledgements

We gratefully acknowledge the following works, which provided essential datasets and tools for this research:

• MISATO: Machine Learning Dataset of Protein–Ligand Complexes for Structure-Based Drug Discovery
  This work offers a valuable dataset for training and evaluating models in protein–ligand interaction prediction.
  🔗 GitHub Repository

• ATOM3D: Tasks On Molecules in 3 Dimensions
  This project provides a suite of benchmark tasks for machine learning on 3D molecular structures, advancing research in structural biology and drug discovery.
  🔗 GitHub Repository