Fast prediction of protein flexibility

Area of Science:

• Computational Biology
• Structural Biology
• Bioinformatics

Background:

• Molecular dynamics (MD) simulations are crucial for understanding protein flexibility but are computationally intensive.
• Existing MD resources provide valuable data, enabling the development of predictive models for protein dynamics.
• Estimating protein flexibility accurately, especially for large systems and long timescales, remains a significant computational challenge.

Purpose of the Study:

• To introduce a novel, efficient, and accurate computational model for predicting protein flexibility.
• To develop a method that directly estimates protein flexibility from atomic coordinates, reducing computational burden.
• To create a generalizable model applicable across various protein structures and sizes.

Main Methods:

• Introduction of the Graphlet Degree Vector (GDV) as a 15-dimensional feature vector.
• GDV captures local atomic packing and spatial connectivity.
• Model training and validation using the ATLAS database and independent NMR/cryo-EM datasets.

Main Results:

• The GDV model achieves a high Spearman correlation of 0.828 with MD data for protein flexibility prediction.
• The model demonstrates robustness and generalizability across independent datasets (NMR, cryo-EM).
• Near real-time prediction (seconds) is achievable for large proteins (20,000 atoms) on standard hardware.

Conclusions:

• The lightweight, coordinate-based GDV model provides accurate and generalizable predictions of protein flexibility.
• This approach significantly reduces the computational cost associated with estimating protein flexibility.
• The GDV model offers a fast and accessible tool for analyzing protein dynamics across diverse biological structures.