Designing flexible protein structures and sampling protein conformations with a unified model using vector...

Area of Science:

• Computational biology
• Structural biology
• Deep learning for protein science

Background:

• Protein conformational dynamics are crucial for biological functions.
• Predicting and designing protein structures with dynamic capabilities is a key challenge in molecular biology.
• Deep learning approaches show promise for understanding and engineering protein structures.

Purpose of the Study:

• To introduce Protein Vector Quantization and Diffusion (PVQD), a novel deep learning framework.
• To enable accurate prediction of protein conformational distributions and design of proteins with desired dynamics.
• To capture sequence-dependent effects on protein conformational dynamics.

Main Methods:

• Utilized a vector-quantized auto-encoder to learn latent representations of protein backbones.
• Employed latent-space diffusion models for protein backbone generation and conformation sampling.
• Conditioned conformation sampling on native protein sequences.

Main Results:

• PVQD generates protein backbones with natural distributions of secondary structures, loop lengths, and domain sizes.
• PVQD demonstrates superior performance in reproducing experimental structural variations for benchmark proteins compared to existing methods.
• PVQD accurately captures sequence-specific influences on functional conformational dynamics in proteins like K-Ras and KaiB.

Conclusions:

• The PVQD framework offers a unified approach for protein structure prediction and design.
• Latent space diffusion is a powerful tool for modeling and generating protein conformational dynamics.
• PVQD advances the field of protein engineering by enabling the design of proteins with controlled dynamic properties.