A simulation study to compare gene set analysis methods

Abstract

Genome-wide association studies (GWA studies) identify alleles that are associated with a disease. These allele variations are called single nucleotide polymorphisms (SNPs). However, GWA studies do not account for interaction between SNPs. Gene set analysis (GSA) is used in GWA studies to account for interaction. GSA methods map SNPs to gene sets and identify gene sets that are associated with a disease. Comprehensive reviews of GSA exist in the literature. However, these reviews do not compare specific methods or implement them on data. In this thesis, we compare six GSA methods. We use seven factors highlighted by the reviews as important in GSA to compare these methods. For example, we analyse how each method accounts for parameters that could affect the analysis. These parameters include gene size and SNP interaction. We consider the null hypothesis tested by each method. We also analyse the sensitivity of methods to individual SNPs with small p-values. In contrast, the marginal effect of many SNPs that cause diseases is often small. The p-values of such SNPs need not be small. We conduct a simulation study to compare four GSA methods. We investigate the sensitivity of these methods to SNPs with very small p-values. We use Manhattan plots to display gene sets that were assigned disparate p-values by different methods. We also use receiver operating characteristic curves to compare the performance of each method. Finally, we recommend a method that gave excellent performance.