Microarray technology has been widely used for the study of gene expression in cancer diagnosis. This paper presents a novel method for application of genetic algorithms and constructive neural networks for the analysis of microarray cancer data. Our aim is to establish a gene selection method without assumption restriction to reduce the dimension of the data set. We have used the modified genetic algorithm to evolve gene subsets. It is found that this algorithm can reduce the dimension of the data set and all test samples can be classified correctly. The modified genetic algorithm strategy combines mutual information and classification models to predict cancer outcome. Furthermore, a constructive neural network model, C-Mantec, is applied providing reduced network architectures with competitive results in comparison to other classifiers. Six free-public cancer databases have been used to test our approach. This method increases the selection accuracy of genes close to 100% after more than 120 runs as against the selection accuracy of genes of SFS which is around 80% after more than 1000 runs. The evolution time in MGA is faster than SFS by close to 20 minutes. © Research India Publications.