Inferring Dynamic Information from Protein Structures by Gaussian Integrals and Deep Learning.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning

Background:

• Protein dynamics are crucial for function but challenging to study experimentally.
• Current methods like molecular dynamics (MD) simulations are resource-intensive and low-throughput.
• There is a need for scalable, structure-based approaches to infer protein dynamics.

Purpose of the Study:

• To develop a deep learning framework for predicting protein dynamics from static structures.
• To enable rapid and cost-effective inference of protein flexibility and motion.

Main Methods:

• Utilized Gaussian integral (GI) descriptors of Cα backbone topology.
• Employed an attention-based 1D-convolutional neural network (CNN) for classification.
• Developed regression models to predict Root Mean Square Fluctuation (RMSF).

Main Results:

• GI descriptors stratified proteins into functionally relevant clusters.
• The 1D-CNN accurately classified protein flexibility (AUC=0.772) and dynamics modes (AUC=0.91).
• Regression models achieved high accuracy in predicting mean RMSF (r=0.72) and slow-mode RMSF (r=0.83).

Conclusions:

• The deep learning framework provides a scalable method for inferring protein dynamics.
• This approach accelerates the analysis of protein flexibility and collective motion.
• The method offers a valuable alternative to traditional experimental and simulation techniques.