An empirical study on the matrix-based protein representations and their combination with sequence-based approaches

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Automatic protein classification is crucial for many scientific domains.
• Existing feature extraction methods often lack generalizability across different classification problems.
• Developing robust protein representations is key to improving classification accuracy.

Purpose of the Study:

• To evaluate various feature extraction approaches for sequence-based protein classification.
• To identify protein representations that enhance classification performance.
• To investigate the effectiveness of combining different feature descriptors.

Main Methods:

• Evaluation of protein representations derived from Position Specific Scoring Matrices (PSSM) and amino-acid sequences.
• Development of a texture descriptor from PSSM representations.
• Comparison of Support Vector Machines (SVM) and ensemble SVMs trained on different descriptors.
• Fusion of classifiers trained on diverse protein descriptors.

Main Results:

• A texture descriptor extracted from PSSM significantly improves classification performance.
• No single descriptor consistently outperforms others across all datasets.
• Classifier fusion using multiple descriptors achieves robust and high performance on all tested datasets.
• Matlab code and datasets are available for reproducibility.

Conclusions:

• Feature extraction from PSSM, particularly using texture descriptors, enhances protein classification.
• Classifier fusion is a powerful strategy for achieving generalizable and reliable automatic protein classification.
• The proposed methods offer a significant advancement in the field of computational protein analysis.