Accurate Conformation Sampling via Protein Structural Diffusion.

Area of Science:

• Computational Biology
• Structural Biology
• Drug Discovery

Background:

• Accurate sampling of protein conformations is crucial for biological and medical advancements.
• Deep learning has improved protein structure prediction, but predicting diverse stable conformations remains challenging.
• Existing methods often exhibit conformational bias from training data.

Purpose of the Study:

• To develop a novel method, UFConf, for robustly sampling diverse protein conformations from amino acid sequences.
• To enhance protein structure prediction accuracy and structural validity.
• To demonstrate UFConf's utility in drug discovery applications, specifically protein-ligand docking.

Main Methods:

• Transformed AlphaFold2 into a diffusion model by implementing a conformation-based diffusion process.
• Adapted the model architecture to process diffused inputs effectively.
• Developed a hierarchical reweighting protocol using structural clustering to mitigate conformational bias.

Main Results:

• UFConf demonstrated superior performance in successful sampling and structural validity compared to existing methods.
• UFConf successfully overcame energy barriers in molecular dynamics simulations, enabling more efficient sampling.
• UFConf accurately predicted a novel protein-ligand complex in a blind test for drug discovery.

Conclusions:

• UFConf offers a powerful new approach for sampling diverse protein conformations.
• The method shows significant potential to accelerate drug discovery and advance biological research.
• UFConf provides versatile sampling capabilities, including partial sampling and structural interpolation.