A systematic comparison of GWAS pathway analysis methods

Abstract

Though rooted in genomic expression studies, pathway analysis for genome-wide association studies (GWAS) has gained increasing popularity, since it provides a way of discovering hidden disease causal mechanisms by combining statistical methods with biological knowledge. Algorithms or programs currently available can be categorized by different types of input data, null hypothesis or number of analysis stages. Due to complexity caused by SNP, gene and pathway relationships, re-sampling strategies like permutation are always utilized to derive empirical distributions for test statistics, and then evaluate the significance of candidate pathways. However, thorough performance evaluation of these algorithms on real GWAS datasets and real biological pathway databases needs to be done before these methods should become common practice in GWAS. Seven algorithms were selected to conduct pathway analysis on SNP genotypes together with simulated and real phenotypes from the WTCCC Crohn’s disease study. All 7 methods control type I error rate (at 0.05) well, and are mostly slightly conservative. However, the methods varied greatly in terms of power and running time. In real data analysis, raw data-based algorithms turn out to be best, provided sufficient computation capacity is available. Given the variability in performance, in general, particularly when underlying disease causal mechanism is ambiguous, it is worthwhile to apply two or more pathway analysis algorithms on the same GWAS dataset.