Why can deep convolutional neural networks improve protein fold recognition? A visual explanation by interpretation

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• Protein fold recognition is crucial for predicting protein structure and function.
• Current machine learning methods for fold recognition are limited by handcrafted features.
• Deep convolutional neural networks (DCNNs) show promise for automatic feature extraction but lack interpretability.

Purpose of the Study:

• To explore the internal mechanisms of DCNNs in protein fold recognition.
• To explain why DCNNs are effective for identifying protein folds.
• To develop a visual explanation technique for DCNN-based fold recognition.

Main Methods:

• Trained a VGGNet-based DCNN (VGGNet-FE) to extract fold-specific features from protein contact maps.
• Developed a contact-assisted predictor (VGGfold) using the trained VGGNet-FE.
• Employed deconvolution techniques to visualize features extracted by convolutional layers and identify fold-discriminative regions.

Main Results:

• VGGNet-FE successfully extracted fold-specific features from protein contact maps.
• Visualizations revealed distinct fold-discriminative regions identified by VGGNet-FE for different protein folds.
• The identified regions explain the enhanced performance of the VGGfold predictor.

Conclusions:

• DCNNs effectively extract salient, fold-discriminative regions from protein contact maps.
• The proposed visualization method enhances understanding of DCNNs in protein fold recognition.
• This approach is applicable to other DCNN-based problems in bioinformatics and computational biology.