Association studies between SNP and complex disease using statistics and machine learning methods has been faced serious curse of dimensionality in a large-scale SNP set. Reducing a large-scale SNP set to a smaller one is the key and primary problem for the association research. To solve the problem, a novel method, called Multi-group Genetic Algorithm (MGA), is proposed for rough feature selection in SNPs. Mutual information (MI) as the fitness of genetic algorithm is used to measure the association relation of SNPs to disease. Optimal SNP subsets searched by MGA method are combined to form a feature SNP set. In contrast to Maximum Entropy (ME) method, this method can reduce the number of redundant SNPs ,which have nothing to do with disease, while disease-related SNPs are retained. Experimental results on simulated datasets of SNPs show that the MGA method provides the appropriate size of SNP data for the future research, and can be employed in a middle-scale or large-scale of SNPs set.
