DyneTrion: Spatiotemporally Coherent Generative Emulation of Protein Dynamics Across Timescales

Kaihui Cheng^1,2*  Zhiqiang Cai^2*  Peng Tu²  Yisong Yao² 

Limei Han^1,2  Libo Wu¹  Siyu Zhu^2†  Tzuhsiung Yang†²  Yuan Qi^2† 

¹Fudan University  ²Shanghai Academy of AI for Science 

^*Equal Contribution  ^†Corresponding Author

1ail_A	1ifg_A	2kxl_A	2rcs_H
5b3k	7ejg	7n0j_E	7vsx

📅️ Roadmap

Status	Milestone	ETA
✅	Inference code and pretrained checkpoints released	2025-12-22
🚀	Training code release	TBD
🚀	Accelerate inference performance	TBD

⭐️ Source Data

DynamicPDB (https://github.com/fudan-generative-vision/dynamicPDB): a large-scale dataset that augments existing static 3D protein structural databases (e.g., PDB) with dynamic information and additional physical properties. It contains approximately 12.6k filtered proteins, each subjected to all-atom molecular dynamics (MD) simulations to capture conformational changes.

🛠️ Installation

# Create virtual environment (Python 3.10.12 is recommended)
python -m venv .venv
source .venv/bin/activate

# Install PyTorch (CUDA 12.4)
pip install torch==2.4.0 --index-url https://download.pytorch.org/whl/cu124

# Install other dependencies
pip install -r requirements.txt

📥 Download Pretrained Models & Pre-Processed Data

Pretrained weights for DyneTrion are available on Hugging Face. The pre-processed test data can be found in the dynamicPDB dataset repository, DyneTrion-test-data.

▶️ Inference

Run inference using:

bash inference.sh

• Model checkpoint: step_400000.pth
• Input CSV: datasets/inference/inference_data.csv
• Frame number: n_frame = 16
• Motion number: n_motion = 2
• Frame sampling step: sample_step = 40
• Extrapolation time: extrapolation_time = 16
• Noise scale: noise_scale = 1.0

Inference results will be saved to save_root (default: ./test/inference/).