Cluster busters.

Bioinformatics experts are always working to design better statistical algorithms to comprehend the expression patterns of tens of thousands of genes. Different algorithms may better serve diverse scientific goals, such as screening for potential tumor markers or obtaining a comprehensive window into the state of a cell as it reacts to an environmental toxicant. Now, in a new use for a preexisting methodology, Raj Acharya and Jyotsna Kasturi, two computer scientists at The Pennsylvania State University, have applied a mathematical approach called Kullback-Leibler (KL) clustering to the identification of patterns in microarray data.

Microarrays shed light the effects of environmental toxicants on genes by measuring the expression of thousands of messenger RNAs simultaneously. Microarray experiments generate vast amounts of data, which bioinformatics experts examine using statistical algorithms designed to detect patterns. Similar genes are sorted into groups, or clusters, that provide insights into gene interactions and thus help to explain underlying biological processes.

Traditionally, mathematicians and engineers have used KL methods to explore theoretical concepts. But KL clustering has proven to be a powerful method for looking at gene expression over time in response to drugs or environmental toxicants, says Murali Ramanathan, an associate professor of pharmaceutical sciences at the University at Buffalo--The State University of New York, who collaborated with the Penn State scientists in proof-of-concept studies on the approach.

In general, clustering procedures find similarities among data set items that form the basis for sorting them into a series of groups. "Clustering is like sorting different-color balls into bins, each containing one color," explains Acharya, who is director of the Penn State Advanced Laboratory for Information Systems and Analysis. Each ball is sorted by how closely it matches the color of the other balls already assigned to bins. The algorithm calculates a similarity score for all pairs of genes and assigns them to a cluster. Genes within clusters carry out similar tasks, such as cholesterol synthesis or wound healing. Any genes of unknown function are "guilty by association"--they are suspected to have a function similar to those of known genes in the same cluster.

The main aim of cluster analysis of microarray data is to expose regulatory networks and assign function to sequences of no known function. So researchers desire small numbers of clusters, each densely packed with interrelated genes that reflect common pathways or biological functions. In contrast, techniques that generate many clusters containing just one or a few genes are undesirable.

In a proof-of-concept experiment, the researchers applied KL clustering to genetic data from the Onto-Express database of the Wayne State University Intelligent Systems and Bioinformatics Laboratory. Graduate student Kasturi had written a computer program to test KL clustering, which was applied to 517 genes from human fibroblasts treated with serum, representing 12 time points. In addition, a larger set of 4,579 yeast genes containing 18 time points in the cell cycle was analyzed. This test run, reported in the March 2003 issue of Bioinformatics, showed that KL clustering performed better at sorting microarray data than the standard method of hierarchical clustering, which uses a different algorithm to measure the similarity of genes.

For the fibroblast data, KL clustering produced about half as many clusters as did hierarchical clustering, and KL clusters were densely packed with similar genes. In contrast, many of the hierarchical clusters contained just one or two genes. A similar pattern emerged for the yeast data. "Using a small and large data set shows that the program is scalable to large sets of genes," says Kasturi.

Some KL clusters were compared with known genes from Onto-Express, which confirmed that clusters shared similar gene functions. For example, one cluster held genes related to call-cell communication, whereas another cluster controlled cellular development. The Penn State team's KL clustering program is available by request by contacting Kasturi at [e-mail].

Toxicologists are always searching for better ways to extract data from high-throughput screens, and new approaches to mining large data sets appear regularly in the bioinformatics literature. But Christopher Bradfield, a professor of oncology at the University of Wisconsin, Madison, and CEO of the Madison-based toxicogenomics service company Functional Biosciences, points out that new algorithms are only as good as their translation into clearer biologic understanding. "This may be a better mouse trap," he says, "but-the real proof will be in how many mice it catches."