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

Area of Science:

• Structural Bioinformatics
• Computational Biology
• Deep Learning

Background:

• Protein conformational flexibility is crucial for biological functions.
• Experimentally determining protein dynamics is resource-intensive and slow.
• Existing methods often require computationally expensive molecular dynamics (MD) simulations.

Purpose of the Study:

• To create a deep learning framework for predicting protein flexibility from static structural descriptors.
• To bypass the need for MD simulations in assessing protein dynamics.
• To enable large-scale, computationally efficient screening of protein flexibility.

Main Methods:

• Utilized the ATLAS database to encode 1,374 protein chains into 30-dimensional Gaussian integral (GI) vectors.
• Employed principal component analysis (PCA) to identify structural clusters based on GI profiles.
• Trained an attention-based 1D convolutional neural network (1D-CNN) for classification and a recurrent neural network (RNN) for regression.

Main Results:

• The 1D-CNN classifier achieved an AUC of 0.772 for predicting flexible vs. non-flexible proteins.
• The RNN regression model attained an R² of 0.537, with a tendency to underestimate high flexibility.
• Identified specific GI components as highly predictive of protein flexibility.
• Found that coil-rich and beta-sheet-rich proteins were more predictable than alpha-helical proteins.

Conclusions:

• Compact GI descriptors capture sufficient information for predicting MD-derived flexibility trends.
• The deep learning framework offers a computationally efficient alternative to MD simulations.
• This approach facilitates large-scale analysis of protein dynamics from structural data, aiding drug design and functional annotation.