Remote homology detection using a kernel method that combines sequence and secondary-structure similarity scores

Area of Science:

• Bioinformatics
• Computational Biology
• Structural Bioinformatics

Background:

• Distant evolutionary relationships between proteins with low sequence similarity are difficult to detect computationally.
• Many protein sequences from large-scale projects remain unassigned to known families.
• Existing sequence-based methods often lack sensitivity for remote homology detection.

Purpose of the Study:

• To develop and evaluate a novel kernel-based method for remote homology detection.
• To integrate sequence and secondary structure information for improved protein classification.
• To assess the method's performance in predicting superfamily membership using the SCOP database.

Main Methods:

• Introduced a kernel-based method combining sequence and secondary-structure similarity scores.
• Employed a discriminative approach for remote homology detection.
• Utilized Support Vector Machines (SVM) classifiers for prediction.

Main Results:

• The combined sequence and predicted secondary-structure kernel method performed comparably to using true secondary structures.
• This approach significantly outperformed sequence-only based classifiers.
• The method achieved superior mean performance compared to recent published results on the same dataset.

Conclusions:

• Combining sequence and secondary structure similarity scores enhances the accuracy of remote homology detection.
• SVM classifiers benefit from joint sequence/secondary-structure similarity for predicting homology between distantly related proteins.
• The developed method offers a more sensitive approach for assigning uncharacterized protein sequences to evolutionary families.