Abstract. Microarrays allow simultaneous measurement of the expression levels of thousands of genes in cells under different physiological or disease states. Because the number of genes exceeds the number of samples, class prediction on microarray expression data leads to an extreme “curse of dimensionality” problem. A principal goal of these studies is to identify a subset of informative genes for class prediction to reduce the curse of dimensionality. We propose a novel genetic approach that selects a subset of predictive genes for classification on the basis of gene expression data. Our genetic algorithm maximizes correlation between genes and classes and minimizes intercorrelation among genes. We tested the genetic algorithm on leukemia data sets and obtained improved results over previous results.