Accurate Prediction of Protein Structural Flexibility by Deep Learning Integrating Intricate Atomic Structures and...

Area of Science:

• Computational Biology
• Structural Biology
• Biophysics

Background:

• Protein dynamics are essential for biological function.
• Predicting protein dynamics computationally is a significant challenge.
• Existing methods lack speed and accuracy.

Purpose of the Study:

• To develop a novel neural network model, RMSF-net, for accurate and rapid prediction of protein dynamics.
• To improve upon existing computational methods for protein dynamic analysis.
• To leverage integrated structural and cryo-electron microscopy data for enhanced prediction efficacy.

Main Methods:

• Development of a neural network architecture (RMSF-net).
• Integration of experimental protein structure data and cryo-electron microscopy (cryo-EM) maps.
• Training the model to identify bidirectional constraints between cryo-EM maps and PDB models.
• Rigorous 5-fold cross-validation for performance assessment.

Main Results:

• RMSF-net achieved high test correlation coefficients: 0.746 ± 0.127 (voxel level) and 0.765 ± 0.109 (residue level).
• The model demonstrates prediction accuracy comparable to molecular dynamics simulations.
• Real-time dynamic inference is achieved with minimal storage requirements (megabytes).

Conclusions:

• RMSF-net significantly outperforms previous methods in predicting protein dynamics.
• The model offers a fast, accurate, and accessible tool for studying protein dynamics.
• RMSF-net is expected to be valuable for various research applications in structural biology and beyond.