Judging the quality of gene expression-based clustering methods using gene annotation

Area of Science:

• Bioinformatics
• Computational Biology
• Systems Biology

Background:

• Gene expression data analysis is crucial for understanding biological systems.
• Clustering algorithms are widely used to group genes with similar expression patterns.
• Evaluating the performance of these algorithms is essential for reliable biological interpretation.

Purpose of the Study:

• To compare the performance of various gene clustering algorithms.
• To assess the impact of different distance metrics on clustering results.
• To determine optimal cluster numbers for biological function enrichment.

Main Methods:

• Utilized a figure of merit based on mutual information between cluster membership and gene attributes.
• Compared Euclidean and Pearson distances for ratio-based and non-ratio-based measurements.
• Evaluated algorithms including self-organized maps and hierarchical clustering (single- and average-linkage).
• Analyzed multiple publicly available gene expression datasets.

Main Results:

• Biological function enrichment is generally highest at lower cluster numbers.
• Euclidean distance is optimal for ratio-based measurements, and Pearson distance for non-ratio-based measurements at optimal cluster numbers.
• Self-organized map (SOM) approach performs best for both measurement types at higher cluster numbers.
• Hierarchical clustering methods (single- and average-linkage) yielded worse-than-random results.

Conclusions:

• The choice of clustering algorithm and distance metric significantly impacts the biological interpretability of gene expression data.
• Lower cluster numbers often lead to more biologically meaningful gene groupings.
• Self-organized maps offer a robust approach for gene clustering, particularly at higher cluster counts.