A systematic comparison of genome-scale clustering algorithms.

Area of Science:

• Bioinformatics
• Computational Biology
• Systems Biology

Background:

• Gene co-expression analysis is crucial for understanding biological pathways.
• Numerous clustering algorithms exist, including parametric (e.g., k-means) and graph-based methods (e.g., WGCNA).
• Prior comparative studies have largely excluded graph theoretical approaches.

Purpose of the Study:

• To compare the effectiveness of various parametric and graph theoretical clustering algorithms.
• To evaluate clustering performance on genome-scale transcriptomic data from Saccharomyces cerevisiae.
• To guide algorithm selection and development in gene co-expression studies.

Main Methods:

• Tested multiple parametric and graph theoretical clustering algorithms.
• Utilized Jaccard similarity to assess cluster agreement with GO and KEGG annotations.
• Calculated the Best Average Top 5 (BAT5) score for each method.

Main Results:

• Clustering methods were ranked based on their positive matches to known biological pathways.
• The ability of methods to identify consistent clusters with other approaches was evaluated.
• Graph-based methods demonstrated superior performance in identifying biologically relevant gene clusters.

Conclusions:

• Graph-based clustering techniques significantly outperform conventional methods for this dataset.
• The findings support further development and application of combinatorial and graph-based strategies.
• This research provides valuable insights for selecting optimal clustering algorithms in transcriptomic analysis.