Hierarchical learning architecture with automatic feature selection for multiclass protein fold classification

Area of Science:

• Bioinformatics
• Computational Biology
• Structural Biology

Background:

• Protein structure classification is crucial for predicting new protein structures.
• Effective classification relies on appropriate tools and features, with features often undervalued.
• Existing methods may not fully exploit the relationships and characteristics of known proteins.

Purpose of the Study:

• To develop novel methods for multiclass protein fold classification.
• To improve the accuracy and efficiency of protein structure prediction.
• To gain insights into protein folding processes through feature selection.

Main Methods:

• Utilized a gating neural network for online feature selection, opening gates for important features and closing them for less relevant ones.
• Implemented a hierarchical learning architecture (HLA) with two levels: classifying into four major classes and then into 27 folds.
• Introduced indirect coding features from amino-acid composition sequences using the N-gram concept for more discriminative local features.

Main Results:

• The proposed hierarchical learning architecture with new indirect coding features increased protein fold classification accuracy by approximately 12%.
• The gating neural network significantly reduced the number of features, achieving comparable accuracy with half the original features.
• The gating mechanism provided insights into feature importance for protein folding and reduced computation time.

Conclusions:

• The novel approach combining hierarchical learning architecture and indirect coding features offers a significant improvement in multiclass protein fold classification.
• Gating neural networks are effective for feature reduction and enhancing classification performance in bioinformatics.
• This research contributes to a better understanding of protein structure and folding mechanisms.