Word correlation matrices for protein sequence analysis and remote homology detection

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Protein sequence classification is crucial in computational biology.
• Kernel-based methods offer high accuracy but lack interpretability and are computationally expensive.
• Interpretable models are needed for analyzing discriminative sequence features.

Purpose of the Study:

• To develop a novel, interpretable, and computationally efficient kernel for protein sequence classification.
• To enable fast classification of new protein sequences and analysis of discriminative features.
• To improve protein remote homology detection.

Main Methods:

• A novel kernel based on average word similarity between protein sequences was developed.
• The kernel creates a feature space allowing for analysis of discriminative features.
• The method was evaluated on a benchmark dataset for protein remote homology detection.

Main Results:

• The proposed word correlation approach achieves highly competitive performance compared to state-of-the-art methods.
• The method provides an interpretable model in terms of biologically meaningful features.
• Analysis of discriminative words identifies characteristic regions in biological sequences.

Conclusions:

• The novel word correlation kernel offers a powerful tool for protein remote homology detection.
• Its interpretability facilitates the identification of key sequence features.
• High computational efficiency allows application to large-scale database searches for potential homologs.