DeepFrag-k

DeepFrag-k enhances protein fold recognition by identifying fold-discriminative structural fragments from protein sequences and classifying proteins into folds using a two-stage deep learning pipeline.

Key Features:

• Two-stage architecture: A sequential pipeline with a fragment-prediction stage followed by a fragment-classification stage implemented with deep learning models.
• Multi-modal Deep Belief Network (DBN): Predicts potential structural fragments from input protein sequences and produces fragment-level representations.
• Fragment vector representation: Encodes predicted structural fragments into fragment vectors that capture local structural information.
• Deep convolutional neural network (CNN): Classifies fragment vectors into corresponding protein folds by learning complex patterns across fragments.
• Fold-discriminative fragment identification: Detects fragments that are discriminative for major protein folds, serving as structural "keywords".
• Performance on fragment prediction: Reports 92.98% accuracy for predicting the top-100 most popular fragments.

Scientific Applications:

• Protein fold recognition and classification: Improves the assignment of proteins to fold categories based on fragment-level features.
• Protein structure–function analysis: Provides fragment-based features that aid interpretation of structure–function relationships.
• Drug discovery: Supplies fold-discriminative fragment information that can inform target structure characterization and compound design.
• Molecular modeling: Enhances modeling efforts by providing fragment-level constraints and features for structural prediction.

Methodology:

Stage 1 uses a multi-modal Deep Belief Network (DBN) to predict structural fragments from protein sequences and generate fragment vectors; Stage 2 applies a deep convolutional neural network (CNN) to classify those fragment vectors into protein folds, with identification of fold-discriminative fragments informing the feature vectors.