"Clustering" is a commonly used exploratory data analysis technique for pattern discovery in gene expression datasets. The goal is to find groups, or clusters, of genes that tend to vary together across experiments. The idea is that such genes may be functionally related, or co-regulated. While most previously used clustering algorithms were largely based on heuristics, clustering algorithms based on probability models offer a principled alternative, including an objective way to estimate the number of clusters. Our paper was possibly the first to apply model-based clustering methods to gene expression data, and demonstrated that clusters suggested by model-based methods are similar to those known to have biological significance.

“Now that the Human Genome Project has provided a list of genes, the focus has turned to discovering what genes do and how they work.„

Our work presented a novel application of model-based clustering algorithms to gene expression data. We showed that the quality of clusters from model-based clustering methods is relatively high compared to leading heuristic-based methods. In addition, the probability framework allows us to come up with a reasonable estimate of the optimal number of clusters. We hope that our results will motivate biomedical researchers to use probability-based methods to analyze their data, and that our method will facilitate the data mining process in biomedical sciences.

This paper is part of my dissertation project titled "Cluster Analysis of Gene Expression Data" in Computer Science at the University of Washington. Seeking more principled alternatives to the heuristic methods then prevalent in microarray cluster analysis, my dissertation advisor, Dr. Ruzzo, and I decided to study the performance of model-based clustering methods on gene expression datasets, and to collaborate with Drs. Raftery, Fraley, and Murua, who are experts in model-based clustering methods.

Now that the Human Genome Project has provided a list of genes, the focus has turned to discovering what genes do and how they work. The analysis of gene expression data from microarrays, which simultaneously measure the activity levels of many different genes, is one of the most prevalent approaches to this task. Genes that vary similarly across experimental conditions may be functionally related, and finding groups of such genes is one goal of gene expression data analysis. Clustering, which is the automatic assigment of data points to meaningful groups, is particularly difficult in this setting because of natural variability and measurement error in the data.

Before our paper, the clustering methods used were largely heuristic. Our contribution was to apply model-based clustering methods to this task. Based on established statistical principles, these methods also provide a way of estimating the number of groups in the data, which was previously lacking. We showed that model-based method produces clusters similar to those known to have biological significance, and subsequent research by others has confirmed the promise of this approach to gene expression data analysis.