From co-expression to co-regulation: how many microarray experiments do we need?

Area of Science:

• Computational Biology
• Bioinformatics
• Systems Biology

Background:

• Cluster analysis is a common method to infer gene regulatory modules and functions based on gene expression patterns.
• However, the biological relevance of clustering results is not always guaranteed.

Purpose of the Study:

• To evaluate the effectiveness of various clustering algorithms in identifying co-regulated genes from microarray data.
• To assess the impact of dataset size and data sources on clustering accuracy.

Main Methods:

• Applied diverse clustering algorithms to microarray datasets of varying sizes.
• Evaluated clustering results by analyzing shared transcription factors between gene pairs within clusters.
• Utilized yeast transcription factor databases (SCPD, YPD) and chromatin immunoprecipitation (ChIP) data for validation.

Main Results:

• Clustering accuracy for identifying co-regulated genes is highly dependent on the number of microarray experiments, plateauing around 50-100 experiments for yeast data.
• The model-based clustering algorithm MCLUST consistently outperformed traditional methods in accurately assigning co-regulated genes.
• ChIP data analysis revealed a high false-negative rate (approx. 80%) with YPD using a p-value of 0.001.

Conclusions:

• While clustering can infer gene regulatory networks, even with extensive data, true-positive rates for known transcription factor interactions may be limited (e.g., 28%).
• False-positive rates can exceed true-positive rates, highlighting the need for careful interpretation of clustering results.
• Model-based clustering offers improved accuracy in identifying co-regulated genes compared to traditional approaches.