Efficient Generation of Protein and Protein-Protein Complex Dynamics via SE(3)-Parameterized Diffusion Models

Area of Science:

• Computational biology
• Structural biology
• Deep learning

Background:

• Protein conformations are crucial for biological functions but challenging to simulate using traditional molecular dynamics (MD).
• Enhanced sampling methods improve efficiency but struggle with vast conformational spaces.
• Generative deep learning offers novel approaches for protein conformational sampling.

Purpose of the Study:

• To develop a novel deep learning framework, protein trajectory diffusion (PTraj-Diff), for generating protein and protein-protein complex trajectories.
• To enable efficient exploration of protein conformational landscapes.
• To integrate with existing protein structure prediction tools like AlphaFold3.

Main Methods:

• PTraj-Diff utilizes a geometric diffusion framework, simulating protein dynamics via a denoising process.
• It employs residue-level SE(3) transformations to capture geometric constraints and structural relationships.
• Tensor product attention and a power Bert Encoder are integrated to reduce computational cost and capture long-range temporal dependencies.

Main Results:

• PTraj-Diff efficiently explores conformational trajectories for both protein monomers and complexes.
• The framework demonstrates compatibility with AlphaFold3-generated conformations, enabling high-quality trajectory prediction.
• The model effectively captures geometric constraints and long-range temporal dependencies in protein dynamics.

Conclusions:

• PTraj-Diff presents a powerful new tool for investigating protein conformational dynamics.
• This deep generative modeling approach enhances molecular dynamics simulations.
• The framework facilitates a deeper understanding of biological functions through accurate protein dynamics exploration.