Knowledge-assisted recognition of cluster boundaries in gene expression data

DNA microarray technology has made it possible to determine the expression levels of thousands of genes in parallel under multiple experimental conditions. Genome-wide analyses using DNA microarrays make a great contribution to the exploration of the dynamic state of genetic networks, and further lead to the development of new disease diagnosis technologies. An important step in the analysis of gene expression data is to classify genes with similar expression patterns into the same groups. To this end, hierarchical clustering algorithms have been widely used. Major advantages of hierarchical clustering algorithms are that investigators do not need to specify the number of clusters in advance and results are presented visually in the form of a dendrogram. However, since traditional hierarchical clustering methods simply provide results on the statistical characteristics of expression data, biological interpretations of the resulting clusters are not easy, and it requires laborious tasks to unveil hidden biological processes regulated by members in the clusters. Therefore, it has been a very difficult routine for experts.

Here, we propose a novel algorithm in which cluster boundaries are determined by referring to functional annotations stored in genome databases.

The algorithm first performs hierarchical clustering of gene expression profiles. Then, the cluster boundaries are determined by the Variance Inflation Factor among the Gene Function Vectors, which represents distributions of gene functions in each cluster. Our algorithm automatically specifies a cutoff that leads to functionally independent agglomerations of genes on the dendrogram derived from similarities among gene expression patterns. Finally, each cluster is annotated according to dominant gene functions within the respective cluster.

In this paper, we apply our algorithm to two gene expression datasets related to cell cycle and cold stress response in budding yeast Saccharomyces cerevisiae. As a result, we show that the algorithm enables us to recognize cluster boundaries characterizing fundamental biological processes such as the Early G1, Late G1, S, G2 and M phases in cell cycles, and also provides novel annotation information that has not been obtained by traditional hierarchical clustering methods. In addition, using formal cluster validity indices, high validity of our algorithm is verified by the comparison through other popular clustering algorithms, K-means, self-organizing map and AutoClass.