Spectral clustering of protein sequences

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Automatic clustering of homologous proteins using only sequence data is a significant challenge in genomics.
• Current local clustering methods, based on sequence distance thresholds, have performance limitations due to their localized approach.

Purpose of the Study:

• To address the limitations of local clustering methods for homologous proteins.
• To introduce and validate a novel global spectral clustering method for improved protein clustering.

Main Methods:

• Analysis of protein sequence distance distributions to demonstrate the limitations of local methods.
• Development and theoretical justification of a global spectral clustering algorithm.
• Extensive performance comparison against established local and global clustering methods using SCOP database subsets.

Main Results:

• The proposed global spectral clustering method consistently produced cluster numbers close to known protein superfamilies.
• The method reduced the number of singleton clusters and improved the grouping of remote homologs.
• Quantified cluster quality showed significant improvements: 84% over hierarchical clustering, 34% over Connected Component Analysis (CCA), and 72% over TribeMCL.

Conclusions:

• Global spectral clustering offers a superior approach for automatically clustering homologous proteins based on sequence information.
• The new method demonstrates enhanced accuracy and efficiency compared to existing local and some global techniques.
• This advancement has strong implications for large-scale genomic and proteomic data analysis.