DeepFrag-k: a fragment-based deep learning approach for protein fold recognition.

Area of Science:

• Structural bioinformatics
• Computational biology
• Deep learning applications in biology

Background:

• Protein fold recognition is a critical challenge in structural bioinformatics.
• Existing methods struggle to accurately identify features that distinguish protein folds.
• Identifying these features is essential for understanding protein structure and function.

Purpose of the Study:

• To develop a novel deep learning architecture, DeepFrag-k, for improved protein fold recognition.
• To identify fold-discriminative features at the fragment level.
• To enhance the accuracy of classifying protein structures based on sequence information.

Main Methods:

• A two-stage deep learning approach is proposed: DeepFrag-k.
• Stage 1: A multi-modal Deep Belief Network (DBN) predicts structural fragments from a sequence, generating a fragment vector.
• Stage 2: A deep Convolutional Neural Network (CNN) classifies the fragment vector to determine the corresponding protein fold.

Main Results:

• DeepFrag-k achieved 92.98% accuracy in predicting the top-100 most popular fragments.
• These predicted fragments serve as discriminative feature vectors.
• The fragment feature vectors significantly improve protein fold recognition accuracy.

Conclusions:

• A specific set of protein fragments act as structural 'keywords' for differentiating major protein folds.
• The DeepFrag-k deep learning architecture effectively identifies these key fragments.
• This approach offers a significant advancement in protein fold recognition accuracy.