Dynamic PDB: A New Dataset and a SE(3) Model Extension by Integrating Dynamic Behaviors and Physical Properties in Protein Structures

Ce Liu^1* Jun Wang^1* Zhiqiang Cai^1* Yingxu Wang^1,3 Huizhen Kuang² Kaihui Cheng² Liwei Zhang¹

Qingkun Su¹ Yining Tang² Fenglei Cao¹ Limei Han² Siyu Zhu^2† Yuan Qi^2†

¹Shanghai Academy of Artificial Intelligence for Science ²Fudan University
³Mohamed bin Zayed University of Artificial Intelligence

Overview

Dynamic PDB is a large-scale dataset that enhances existing prestigious static 3D protein structural databases, such as the Protein Data Bank (PDB), by integrating dynamic data and additional physical properties. It contains approximately 12.6k filtered proteins, each subjected to all-atom molecular dynamics (MD) simulations to capture conformational changes.

Compared with previously existing protein MD datasets, dynamic PDB provides three key advancements:

Extended simulation durations: Up to 1 microsecond per protein, facilitating a more comprehensive understanding of significant conformational changes.
Finer-grained sampling intervals: 1 picosecond intervals, allowing for the capture of more detailed allosteric pathways.
Enriched array of physical properties: Captured during the MD process, including atomic velocities and forces, potential/kinetic energies, and the temperature of the simulation environment, etc.

What dynamic PDB contains?

The attributes contained in dynamic PDB are listed as follows:

File Name	Attribute	Data Type	Unit
`{protein_id}_T.pkl`	Trajectory coordinates	float array	Å
`{protein_id}_V.pkl`	Atomic velocities	float array	Å/ps
`{protein_id}_F.pkl`	Atomic forces	float array	kcal/mol·Å
`{protein_id}_npt_sim.dat`	Potential energy Kinetic energy Total energy Temperature Box volume System density	float float float float float float	kJ/mole kJ/mole kJ/mole K nm³ g/mL

In addition, the following data are stored during the MD simulation:

File Name	Description
`{protein_id}_minimized.pdb`	PDB structure after minimization
`{protein_id}_nvt_equi.dat`	Information in NVT equilibration
`{protein_id}_npt_equi.dat`	Information in NPT equilibration
`{protein_id}_T.dcd`	DCD format for trajectory coordinates
`{protein_id}_state_npt1000000.0.xml`	Status file for MD prolongation

Download dataset

You can easily get dynamic PDB dataset from our ModelScope repo.

Clone the dataset into ${DATA_ROOT}/dynamicPDB directory by cmd below:

git lfs install
git clone git clone https://www.modelscope.cn/datasets/fudan-generative-vision/dynamicPDB.git dynamicPDB

Finally, the dataset should be organized as follows:

./dynamicPDB/
|-- 1ab1_A_npt1000000.0_ts0.001
|   |-- 1ab1_A_npt_sim_data
|   |   |-- 1ab1_A_npt_sim_0.dat
|   |   `-- ...
|   |-- 1ab1_A_dcd
|   |   |-- 1ab1_A_dcd_0.dcd
|   |   `-- ...
|   |-- 1ab1_A_T
|   |   |-- 1ab1_A_T_0.pkl
|   |   `-- ...
|   |-- 1ab1_A_F
|   |   |-- 1ab1_A_F_0.pkl
|   |   `-- ...
|   |-- 1ab1_A_V
|   |   |-- 1ab1_A_V_0.pkl
|   |   `-- ...
|   |-- 1ab1_A.pdb
|   |-- 1ab1_A_minimized.pdb
|   |-- 1ab1_A_nvt_equi.dat
|   |-- 1ab1_A_npt_equi.dat
|   |-- 1ab1_A_T.dcd
|   |-- 1ab1_A_T.pkl
|   |-- 1ab1_A_F.pkl
|   |-- 1ab1_A_V.pkl
|   `-- 1ab1_A_state_npt1000000.0.xml
|-- 1uoy_A_npt1000000.0_ts0.001
|   |-- ...
|   `-- ...
`-- ...

Applications

Trajectory Prediction

We extend the SE(3) diffusion model to incorporate sequence features and physical properties for the task of trajectory prediction.
Specifically, given an initial 3D structure of the protein, the task is to predict 3D structure at the next time step.

Showcase

We present the predicted 3D structures by our method and SE(3)-Trans.

SE(3) Trans	Ours	Ground Truth

Framework

We present the network architecture, where the predicted 3D structures are conditioned on the amino acid sequence and physical properties.

Installation

  pip install -r requirements.txt
  pip install .

Data Preparation

./DATA/
|-- 16pk_A
|   |-- 16pk_A.pdb
|   |-- 16pk_A.npz
|   |-- 16pk_A_new_w_pp.npz
|   |-- 16pk_A_F_Ca.pkl
|   `-- 16pk_A_V_ca.pkl
|-- 1b2s_F
|   |-- ...
|   `-- ...
`-- ...

For each protein xxxx_x,
the xxxx_x.pdb is the pdb file for protein;
the xxxx_x.npz is the node features and edge features from OmegaFold; produced by ./data_preprocess/extract_embedding.py
the xxxx_x_new_w_pp.npz is the trajectory of the protein; produced by first ./data_preprocess/post_process.py and then ./data_preprocess/post_process.py;prep_atlas_with_forces.py;
the xxxx_x_F_Ca.pkl is the force of C alpha atoms of the protein; produced by ./data_preprocess/atom_select.py;
the xxxx_x_V_ca.pkl is the velocity of C alpha atoms of the protein; produced by ./data_preprocess/atom_select.py;

Prepare a list of proteins for training in train_proteins.csv as below:

name	seqres	release_date	msa_id	atlas_npz	embed_path	seq_len	force_path	vel_path	pdb_path
16pk_A	EKKSIN...	1998/11/25	16pk_A	./DATA/16pk_A/16pk_A_new_w_pp.npz	./DATA/16pk_A/16pk_A.npz	415	./DATA/16pk_A/16pk_F_Ca.pkl	./DATA/16pk_A/16pk_V_ca.pkl	./DATA/16pk_A/16pk.pdb
...

Similarly, prepare for the test_proteins.csv

Training

sh run_train.sh

Key arguments in run_train.sh:
data.keep_first: we use frames in [0, data.keepfirst) in each trajectory for training
csv_path: the path for train_proteins.csv

Inference

sh run_eval.sh

Key arguments in run_eval.sh:
model_path: path of pretrained models
start_idx: the time index in the trajectory for evaluation
data.test_csv_path: path for test_proteins.csv

Acknowledgements

We would like to thank the contributors to the OpenFold, OmegaFold.
If we missed any open-source projects or related articles, we would like to complement the acknowledgement of this specific work immediately.

coder-drinker / dynamicPDB