Single nucleotide polymorphisms (SNPs) are the most abundant form of genetic variations amongst species. With the genome-wide SNP discovery, many genome-wide association studies are likely to identify multiple genetic variants that are associated with complex diseases. However, genotyping all existing SNPs for a large number of samples is still challenging even though SNP arrays have been developed to facilitate the task. Therefore, it is essential to select only informative SNPs representing the original SNP distributions in the genome (tag SNP selection) for genome-wide association studies. These SNPs are usually chosen from haplotypes and called haplotype tag SNPs (htSNPs). Accordingly, the scale and cost of genotyping are expected to be largely reduced. We introduce binary particle swarm optimization (BPSO) with local search capability to improve the prediction accuracy of STAMPA. The proposed method does not rely on block partitioning of the genomic region, and consistently identified tag SNPs with higher prediction accuracy than either STAMPA or SVM/STSA. We compared the prediction accuracy and time complexity of BPSO to STAMPA and an SVM-based (SVM/STSA) method using publicly available data sets. For STAMPA and SVM/STSA, BPSO effective improved prediction accuracy for smaller and larger scale data sets. These results demonstrate that the BPSO method selects tag SNP with higher accuracy no matter the scale of data sets is used. (C) 2009 American Institute of Chemical Engineers Biotechnol. Prog., 26: 580-588, 2010