Bayesian hierarchical clustering for microarray time series data with replicates and outlier measurements.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Post-genomic era yields vast biological data, necessitating advanced analysis tools.
• Time series experiments are common, requiring methods to capture temporal data structures.
• Existing methods often struggle with outlier measurements and underutilize replicate data.

Purpose of the Study:

• To develop a novel clustering algorithm for microarray time series data.
• To address challenges posed by outlier measurements and incorporate replicate information.
• To improve the quality and biological relevance of clustering results.

Main Methods:

• Generative model-based Bayesian hierarchical clustering (BHC).
• Gaussian process regression for data structure modeling.
• Mixture model likelihood to handle outlier measurements.
• Empirical Bayes approach using replicate observations for noise variance estimation.

Main Results:

• The algorithm automatically determines the optimal number of clusters.
• It effectively models outlier measurements and incorporates non-uniformly sampled time points.
• Demonstrated superior performance in generating higher quality, biologically meaningful clusters compared to state-of-the-art methods.
• Showcased the importance of modeling outliers and utilizing replicate data for improved gene expression profile discrimination.

Conclusions:

• The developed clustering algorithm offers improved handling of inherent noise in high-throughput genomic data.
• Incorporating outlier and replicate information leads to more robust time series analysis.
• The 'Timeseries BHC' algorithm is available as an R package for broader scientific use.