Mining for gene-environment and gene-gene interactions: parametric and non-parametric tests for detecting variance quantitative trait loci

Introduction Detection of variance quantitative trait loci (vQTL) can facilitate the discovery of gene-environment (GxE) and gene-gene interactions (GxG). Identifying vQTLs before direct GxE and GxG analyses can considerably reduce the number of tests and the multiple-testing penalty. Methods Despite some methods proposed for vQTL detection, few studies have performed a head-to-head comparison simultaneously concerning false positive rates (FPRs), power, and computational time. This work compares three parametric and two non-parametric vQTL tests. Results Simulation studies show that the deviation regression model (DRM) and Kruskal-Wallis test (KW) are the most recommended parametric and non-parametric tests, respectively. The quantile integral linear model (QUAIL, non-parametric) appropriately preserves the FPR under normally or non-normally distributed traits. However, its power is never among the optimal choices, and its computational time is much longer than that of competitors. The Brown-Forsythe test (BF, parametric) can suffer from severe inflation in FPR when SNP’s minor allele frequencies <0.2. The double generalized linear model (DGLM, parametric) is not valid for non-normally distributed traits, although it is the most powerful method for normally distributed traits. Discussion Considering the robustness (to outliers) and computation time, I chose KW to analyze four lipid traits in the Taiwan Biobank. I further showed that GxE and GxG were enriched among 30 vQTLs identified from the four lipid traits.