Powerful gene set analysis in GWAS with the Generalized Berk-Jones statistic

A common complementary strategy in Genome-Wide Association Studies (GWAS) is to perform Gene Set Analysis (GSA), which tests for the association between one phenotype of interest and an entire set of Single Nucleotide Polymorphisms (SNPs) residing in selected genes. While there exist many tools for performing GSA, popular methods often include a number of ad-hoc steps that are difficult to justify statistically, provide complicated interpretations based on permutation inference, and demonstrate poor operating characteristics. Additionally, the lack of gold standard gene set lists can produce misleading results and create difficulties in comparing analyses even across the same phenotype. We introduce the Generalized Berk-Jones (GBJ) statistic for GSA, a permutation-free parametric framework that offers asymptotic power guarantees in certain set-based testing settings. To adjust for confounding introduced by different gene set lists, we further develop a GBJ step-down inference technique that can discriminate between gene sets driven to significance by single genes and those demonstrating group-level effects. We compare GBJ to popular alternatives through simulation and re-analysis of summary statistics from a large breast cancer GWAS, and we show how GBJ can increase power by incorporating information from multiple signals in the same gene. In addition, we illustrate how breast cancer pathway analysis can be confounded by the frequency of FGFR2 in pathway lists. Our approach is further validated on two other datasets of summary statistics generated from GWAS of height and schizophrenia.Author summary: Researchers are frequently interested in the association between a biologically related set of genes—for example, a particular immune response pathway—and a complex phenotype. Such associations are often explored by applying various gene set analysis methods to genotype data from genome-wide association studies. However, many common methods are ad-hoc in nature and possess unknown statistical operating characteristics; reviews of existing procedures often show poor Type I error and power. We propose conducting gene set analysis with a class of tests that possesses both rigorous statistical motivation and excellent performance in application. Comparisons with popular alternatives including GSEA and MAGMA show a substantial increase in power. In addition, we introduce a novel step-down inference procedure that mitigates the confounding introduced by different gene set databases. For example, this procedure identifies that a seemingly strong association between breast cancer and Ear Morphogenesis is actually an association between breast cancer and just one single gene in the Ear Morphogenesis pathway. Use of the step-down procedure can improve reproducibility and result in much more interpretable findings when performing gene set analysis.