Generating Dynamic Structures Through Physics-Based Sampling of Predicted Inter-Residue Geometries.

Area of Science:

• Structural Biology
• Computational Biology
• Biophysics

Background:

• Deep learning methods like AlphaFold2 excel at predicting static protein structures.
• Predicting dynamic protein structures and alternative conformations remains a significant challenge in structural biology.

Purpose of the Study:

• To introduce trRosettaX2-Dynamics (trX2-D), a novel method for predicting protein alternative conformations and dynamic structures.
• To address the limitations of current methods in capturing protein flexibility and conformational heterogeneity.

Main Methods:

• trX2-D utilizes a Transformer-based neural network to predict inter-residue geometric constraints.
• Physics-based iterative sampling is employed on these constraints to generate dynamic structures, bypassing the need for known native states.
• The model was pre-trained on X-ray structures and fine-tuned on Nuclear Magnetic Resonance (NMR) dynamic structures.

Main Results:

• Benchmarking on datasets for alternative conformations and dynamics demonstrated trX2-D's capability.
• The method shows promise in predicting diverse protein conformations and accurately capturing structural dynamics.
• trX2-D successfully generated dynamic structures without prior knowledge of specific structural states.

Conclusions:

• Integrating deep learning predictions with physics-based sampling offers a powerful approach for protein dynamic structure prediction.
• trX2-D represents a significant advancement in modeling protein flexibility and conformational ensembles.
• This work opens new avenues for understanding protein function through dynamic structural insights.