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

Ce Liu, Jun Wang+11 more|arXiv (Cornell University)|Aug 22, 2024

Protein Structure and Dynamics4 citations

TL;DR

This paper introduces Dynamic PDB, a large-scale dataset of ~12.6K proteins with 1-µs all-atom molecular dynamics simulations, capturing conformational changes and rich physical properties (e.g., velocities, forces, energies). It extends an SE(3) diffusion model by integrating these physical properties, achieving improved trajectory prediction accuracy, with MAE reduced to 0.277 and RMSD to 0.528 on benchmark proteins.

ABSTRACT

Despite significant progress in static protein structure collection and prediction, the dynamic behavior of proteins, one of their most vital characteristics, has been largely overlooked in prior research. This oversight can be attributed to the limited availability, diversity, and heterogeneity of dynamic protein datasets. To address this gap, we propose to enhance existing prestigious static 3D protein structural databases, such as the Protein Data Bank (PDB), by integrating dynamic data and additional physical properties. Specifically, we introduce a large-scale dataset, Dynamic PDB, encompassing approximately 12.6K proteins, each subjected to all-atom molecular dynamics (MD) simulations lasting 1 microsecond to capture conformational changes. Furthermore, we provide a comprehensive suite of physical properties, including atomic velocities and forces, potential and kinetic energies of proteins, and the temperature of the simulation environment, recorded at 1 picosecond intervals throughout the simulations. For benchmarking purposes, we evaluate state-of-the-art methods on the proposed dataset for the task of trajectory prediction. To demonstrate the value of integrating richer physical properties in the study of protein dynamics and related model design, we base our approach on the SE(3) diffusion model and incorporate these physical properties into the trajectory prediction process. Preliminary results indicate that this straightforward extension of the SE(3) model yields improved accuracy, as measured by MAE and RMSD, when the proposed physical properties are taken into consideration. https://fudan-generative-vision.github.io/dynamicPDB/ .

Motivation & Objective

• To address the lack of comprehensive, diverse, and long-duration dynamic protein datasets in existing static structural databases like the PDB.
• To capture detailed conformational dynamics and physical properties (e.g., velocities, forces, energies) at 1 ps resolution over 1 µs simulations.
• To improve protein trajectory prediction by extending the SE(3) diffusion model with physical property integration.
• To evaluate the impact of physical properties on prediction accuracy and model generalization across unseen proteins.

Proposed method

• Generated a large-scale dataset, Dynamic PDB, by performing all-atom molecular dynamics simulations on ~12.6K proteins for 1 microsecond each at 1 ps time intervals.
• Collected comprehensive physical properties including atomic velocities, forces, potential and kinetic energies, and simulation temperature at each time step.
• Extended the SE(3) diffusion model to incorporate physical properties such as velocity and force as additional input features during trajectory prediction.
• Trained the model on multiple proteins simultaneously to enable zero-shot generalization to unseen proteins.
• Conducted ablation studies to isolate the contribution of physical properties (velocity, force) to prediction performance.
• Evaluated model performance using standard metrics: mean absolute error (MAE) and root mean square deviation (RMSD).

Experimental results

Research questions

• RQ1 How does integrating physical properties such as velocity and force improve the accuracy of protein conformational trajectory prediction?
• RQ2 What is the impact of longer simulation durations (1 µs vs. 100 ns vs. 50 ns) on the model’s ability to learn dynamic patterns?
• RQ3 To what extent can a diffusion model trained on diverse protein trajectories generalize to predict trajectories of previously unseen proteins?
• RQ4 How do the physical properties captured in Dynamic PDB contribute to modeling allosteric transitions and conformational changes?

Key findings

• Incorporating velocity into the SE(3) model reduced MAE from 0.284 to 0.279 and RMSD from 0.546 to 0.533 on protein 2ERL_A.
• Adding force information further reduced MAE to 0.277 and RMSD to 0.528, demonstrating a cumulative benefit of physical properties.
• The model trained on 214 proteins achieved a MAE of 0.299 and RMSD of 0.584 on unseen proteins, outperforming training on only 10 proteins (MAE: 0.303, RMSD: 0.596).
• Longer simulation durations (1 µs) improved the performance of the flow-based model (FM), while SE(3)-Trans, EGNN, and SEGNN showed comparable accuracy across simulation lengths.
• Qualitative results show that predictions incorporating physical properties align more closely with ground truth, particularly in preserving secondary structure features like alpha helices and beta sheets.
• The dataset enables detailed analysis of conformational changes, including RMSD and radius of gyration trends, contact map frequencies, and Ramachandran plot distributions.