VGGfold

VGGfold applies deep convolutional neural networks to extract and visualize fold-discriminative features from predicted residue-residue contact maps for protein fold recognition.

Key Features:

• Deep Convolutional Neural Networks (DCNNs): VGGfold leverages DCNNs to learn discriminative patterns for fold recognition from contact maps.
• VGGNet-FE model: The method uses a VGGNet-FE architecture based on VGGNet to perform feature extraction.
• Predicted residue-residue contact maps: Predicted contact maps serve as the input representation for fold-specific feature learning.
• Automatic feature extraction: The approach automatically extracts discriminative features without relying on handcrafted descriptors.
• Contact-assisted predictor: VGGfold functions as a contact-assisted predictor that uses contact-map information to improve fold classification.
• Deconvolution visualization: Deconvolution techniques are used to visualize features extracted at each convolutional layer.
• Fold-discriminative regions: Visualization reveals high-level semantic fold-discriminative regions within predicted contact maps.
• Generalizable visualization method: The visualization approach is applicable to other DCNN-based questions in bioinformatics and computational biology and aids interpretation of contact map–structure relationships.

Scientific Applications:

• Protein fold recognition: Enhances recognition of protein folds to support protein structure and function prediction.
• Interpreting DCNNs in biology: Provides visual explanations of how DCNNs derive features for biological classification tasks.
• Contact map–structure analysis: Illuminates relationships between predicted residue-residue contact maps and protein tertiary structures.
• DCNN visualization for bioinformatics: Supplies a visualization technique that can be applied to other DCNN-based analyses in computational biology.

Methodology:

Train the VGGNet-FE model on predicted residue-residue contact maps to identify fold-specific features and apply deconvolution techniques to visualize convolutional-layer features and reveal fold-discriminative regions.