Selection of reliable genes from a huge gene expression data containing high intergene correlation is essential to carry out a diagnostic test and successful treatment. In this regard, a rough set based gene selection algorithm is reported, which selects a set of genes by maximizing the relevance and significance of the selected genes. A gene ontology-based similarity measure is proposed to analyze the functional diversity of the selected genes. It also helps to analyze the effectiveness of different gene selection methods. The performance of the rough set based gene selection algorithm, along with a comparison with other gene selection methods, is studied using the predictive accuracy of K-nearest neighbor rule and support vector machine on two cancer and one arthritis microarray data sets. An important finding is that the rough set based gene selection algorithm selects more functionally diverse set of genes than the existing algorithms. © 2011 Springer-Verlag.