Comparing algorithms for clustering of expression data: how to assess gene clusters

Area of Science:

• Bioinformatics
• Computational Biology
• Systems Biology

Background:

• Clustering gene expression data is vital for identifying co-regulated genes.
• Numerous clustering algorithms exist, yielding diverse results that require objective evaluation.
• Assessing the quality of clustering solutions is challenging without robust validation methods.

Purpose of the Study:

• To present a novel method for assessing gene expression data clustering algorithms.
• To combine internal and external validity indices for superior model selection.
• To compare different clustering algorithms based on biological relevance and accuracy.

Main Methods:

• Developed a hybrid validation approach using Minimum Description Length (MDL) principle for internal validity.
• Incorporated an external validity index measuring consistency with experimental data and functional gene links.
• Tested the method on various clustering algorithms, including those designed for noisy data.

Main Results:

• The combined internal and external validity indices effectively pinpoint the optimal clustering model.
• The proposed method successfully identifies clustering algorithms that maximize correlation with gene networks.
• Demonstrated the ability to minimize error rates and assess cluster significance in relation to biochemical pathways.

Conclusions:

• A robust method for evaluating gene expression clustering algorithms has been established.
• The combined validity index approach offers a superior strategy for selecting optimal clustering models.
• This work facilitates more accurate identification of gene clusters and their biological significance.