Exploratory cluster analysis of genomic high-throughput data sets and their dependencies

Abstract

This thesis studies exploratory cluster analysis of genomic high-throughput data sets and their interdependencies. In modern biology, new high-throughput measurements generate numerical data simultaneously from thousands of molecules in the cell. This enables a new perspective to biology, which is called systems biology. The discipline developing methods for the analysis of the systems biology data is called bioinformatics. The work in this thesis contributes mainly to bioinformatics, but the approaches presented are general purpose machine learning methods and can be applied in many problem areas.

A main problem in analyzing genomic high-throughput data is that the potentially useful new findings are hidden in a huge data mass. They need to be extracted and visualized to the analyst as overviews.

This thesis introduces new exploratory cluster analysis methods for extracting and visualizing findings of high-throughput data. Three kinds of methods are presented to solve progressively better-focused problems. First, visualizations and clusterings using the self-organizing map are applied to genomic data sets. Second, the recently developed methods for improving the visualization and clustering of a data set with auxiliary data are applied. Third, new methods for exploring the dependency between data sets are developed and applied. The new methods are based on maximizing the Bayes factor between the model of independence and the model of dependence for finite data.

The methods outperform their alternatives in numerical comparisons. In applications they proved capable of confirming known biological findings, which validates the methods, and also generated new hypotheses. The applications included exploration of yeast gene expression data, yeast gene expression data in a new metric learned with auxiliary data, the regulation of yeast gene expression by transcription factors, and the dependencies between human and mouse gene expression.