ResaPred: A Deep Residual Network With Self-Attention to Predict Protein Flexibility

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning

Background:

• Protein flexibility is crucial for understanding biological mechanisms.
• Predicting protein flexibility aids in comprehending molecular mechanisms.
• Existing methods for protein flexibility prediction have limitations.

Purpose of the Study:

• To propose a novel deep learning method, ResaPred, for accurate protein flexibility prediction.
• To extract diverse and deep key features related to protein flexibility from sequences.
• To validate the method's effectiveness against existing approaches.

Main Methods:

• ResaPred utilizes a novel deep network architecture incorporating a modified 1D residual module and a self-attention mechanism.
• The modified 1D residual module includes three convolution layers with batchnorm and relu activations to ensure stable training.
• Self-attention mechanisms are employed to capture long-range dependencies within protein sequences.

Main Results:

• ResaPred achieves state-of-the-art results in protein flexibility prediction across non-strict and strict experimental cases.
• The method effectively extracts deep key features from protein sequences, including secondary structure, torsion angle, and solvent accessibility.
• Analysis revealed correlations between protein secondary structure, solvent accessibility, and flexibility.

Conclusions:

• ResaPred demonstrates superior performance in predicting protein flexibility compared to existing methods.
• The deep learning approach effectively captures complex features relevant to protein flexibility.
• Case studies on viral proteins confirm the method's practical applicability and effectiveness.