Building multiclass classifiers for remote homology detection and fold recognition.

Area of Science:

• Computational biology
• Bioinformatics
• Structural bioinformatics

Background:

• Protein remote homology detection and fold recognition are critical in computational biology.
• Support vector machines (SVMs) are effective for binary classification but less explored for multiclass problems.
• Extending SVMs to multiclass scenarios is essential for comprehensive protein classification.

Purpose of the Study:

• To evaluate various methods for building SVM-based multiclass classification schemes.
• To assess the effectiveness of these schemes for protein remote homology prediction and fold recognition.
• To compare different approaches for multiclass SVM model construction.

Main Methods:

• Comprehensive evaluation of SVM-based multiclass classification schemes.
• Direct multiclass SVM model building.
• Second-level learning to combine binary SVM classifier predictions.
• Building and combining binary classifiers for hierarchical SCOP levels.

Main Results:

• Most proposed multiclass SVM approaches are effective for remote homology prediction and fold recognition.
• Schemes using predictions from ancestral SCOP categories reduce error rates.
• These hierarchical schemes minimize misclassifications between superfamilies and folds.

Conclusions:

• Multiclass SVM approaches, particularly those leveraging hierarchical SCOP data, show strong performance.
• Hierarchical schemes improve accuracy and reduce fold-level misclassifications.
• Model complexity should be moderate due to training data limitations.