Using k-Means? Consider ArrayMiner

With the (near-)availability of complete sequences of human and other genomes, research effort is turning from obtaining the sequence itself towards determining the biological function of the genes within the sequence. One approach, made practically feasible with the advent of DNA microarray technology, consists of clustering of genes into groups of coexpressed genes, i.e. genes exhibiting similar behavior in some circumstances. Among the various methods used for identification of groups of coexpressed genes, k-Means is one of the most popular. However, we show here that k-Means is a highly unreliable method of clustering, yielding high-quality solutions with low probability. We then present the ArrayMiner software based on Genetic Algorithms, which addresses the drawback, supplying high-quality solutions in short time with very high reliability. 1. Clustering of Expression Profiles With the (near-)availability of complete sequences of human and other genomes such as Drosophila and Arabidopsis, genomics has produced a significant wealth of sequence data, and the stage has been set for the next task, namely identifying the biological function of the genes within those sequences. Indeed, only this latter knowledge will enable researchers to establish correspondences between diseases and the genome, paving the route to new medication. A major difficulty lies with the fact that the detailed phenomena taking place within organisms are not fully understood and many probably remain to be discovered. Identifying the phenomena and the genes involved, together with the ways in which the genes interact, constitutes the major challenge of “post-sequencing” genetics. One approach to achieve this consists in identification of groups of coexpressed genes, i.e. groups of genes that behave similarly in some conditions. The rationale behind this approach is that similarly behaving genes probably participate together in some phenomenon or have similar functions – identifying such groups may thus help in identifying the phenomenon, discovering previously unknown phenomena, or assigning previously unknown functions to genes. As an additional benefit, clustering gene expression data into groups reduces the unmanageable volume of data into data sets that can be more easily handled by biologists. In this approach, each gene (or ORF) is represented by an expression profile, i.e. a series of numerical values of its activity. The profile of a gene may correspond to a set of readings of the gene’s activity under various conditions, or be a time-series of the gene’s activity after some event. The aim of clustering of the expression profiles being to decide which genes exhibit similar behaviors (i.e., are coexpressed), a measure of similarity between profiles must be defined. Several can be found in the literature, including the Euclidean distance, the correlation, and the Pearson coefficient, this last measure being widely used for the time-series profiles, as it measures the similarity of the shapes, rather than the absolute values of the measurements of the two profiles. 2. The k-Means Technique