Clustering of gene expression data: performance and similarity analysis

Area of Science:

• Bioinformatics
• Computational Biology
• Molecular Biology

Background:

• DNA Microarray technology generates extensive gene expression data.
• Selecting appropriate clustering algorithms for this data is challenging.
• Evaluating clustering algorithm performance is crucial in bioinformatics.

Purpose of the Study:

• To experimentally compare three major clustering algorithms: Hierarchical Clustering (HC), Self-Organizing Map (SOM), and Self Organizing Tree Algorithm (SOTA).
• To introduce and utilize a new data mining tool, Cluster Diff, for analyzing cluster similarity.
• To provide guidance on choosing effective clustering methods for gene expression data.

Main Methods:

• Experimental comparison of HC, SOM, and SOTA using Yeast Saccharomyces cerevisiae gene expression data.
• Development and application of the Cluster Diff tool for similarity analysis of algorithm-generated clusters.
• Performance evaluation based on efficiency and robustness.

Main Results:

• The Self Organizing Tree Algorithm (SOTA) is more efficient than Self-Organizing Map (SOM); Hierarchical Clustering (HC) is the least efficient.
• Cluster Diff effectively identifies the closest matching cluster for a given target cluster.
• SOTA offers a balance of visual representation, noise robustness, and flexibility.

Conclusions:

• Hierarchical Clustering (HC) is visually convenient but lacks robustness and efficiency.
• Self-Organizing Map (SOM) is robust to noise but requires pre-defined cluster numbers.
• Self Organizing Tree Algorithm (SOTA) integrates the strengths of HC and SOM, providing visual structure, noise insensitivity, and flexibility, enhanced by Cluster Diff for comparative analysis.