DeepSF: deep convolutional neural network for mapping protein sequences to folds

Area of Science:

• Structural bioinformatics
• Computational biology
• Genomics

Background:

• Protein fold recognition is crucial for understanding protein function and structure.
• Traditional methods rely on sequence homology, limiting direct sequence-structure relationship insights.
• Existing methods for direct sequence-to-fold classification are limited in scope and practical utility.

Purpose of the Study:

• To develop a novel deep learning method for direct protein sequence-to-fold classification.
• To improve the accuracy and scope of protein fold recognition.
• To facilitate the study of the intrinsic relationship between protein sequence and its three-dimensional structure.

Main Methods:

• A deep 1D-convolutional neural network (DeepSF) was developed.
• The model automatically extracts fold-related features directly from protein sequences.
• The network classifies sequences into one of 1195 known folds.

Main Results:

• DeepSF achieved an average classification accuracy of 75.3% on SCOP1.75 and 73.0% on SCOP2.06.
• The method demonstrated superior performance compared to HHSearch on both template-free and hard template-based modeling targets in CASP9-12.
• Extracted features proved robust against sequence variations and applicable to other protein analysis tasks.

Conclusions:

• DeepSF offers a powerful, direct approach to protein fold recognition.
• The method enhances understanding of sequence-structure relationships.
• DeepSF provides a robust tool for various protein pattern recognition applications.