Introduction
Gene set analysis (GSA) is useful in understanding the biological mechanisms underneath a phenotype by assessing the overall evidence of association of variations in an entire set of genes with a disease or a quantitative trait. GSA has recently been used to investigate many common diseases as an approach for the secondary analysis of a genome-wide association study (GWAS) result [1-3]. Currently, several software tools for GSA are available: GSA-SNP [4], i-GSEA4GWAS [5], GSEA-SNP [6], GeSBAP [7], and so on.
GSA-SNP is useful only when p-values of the single nucleotide polymorphism (SNP) markers are available. While GSA-SNP has implemented several options for estimating the significance of a gene set, its implementation of Z-statistics may be the most convenient. Other methods require permuted p-values that are obtained from sample permutation trials; this requires lengthy computation runs. The Z-statistics method accepts only one set of unpermuted original p-values and compares the score of a gene set against the background distribution made by all the genes; these p-values should be readily available for a typical GWAS. Similarly, i-GSEA4GWAS also uses only the original set of p-values and thus is as convenient as GSA-SNP. Instead of sample permutation, it estimates the significance of a gene set via SNP permutation [8]. One of its unique features is a scaling step that emphasizes the gene sets that are enriched with strongly associated genes.
Often, these approaches give different results in terms of the number of gene set hits. Hence, we compared these two methods using the same dataset while controlling the input parameters as much as possible. Here, we reanalyzed the type 2 diabetes mellitus (T2DM) GWAS results for the Korean population against the Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway databases. The GWAS has been done with both the original unimputed and imputed genotypes. A large discrepancy in the number of significant gene set hits was observed between the two programs as well as the gene set databases. We also observed that the results were strongly affected by imputation.