Knockoff procedure improves causal gene identifications in conditional transcriptome-wide association studies

Abstract

Transcriptome-wide association studies (TWASs) have been developed to nominate candidate genes associated with complex traits by integrating genome-wide association studies (GWASs) with expression quantitative trait loci (eQTL) data. However, most existing TWAS methods evaluate the marginal association between a single gene and the trait of interest without accounting for other genes within the same genomic region or the same gene from different tissues. Additionally, false-positive gene-trait pairs can arise due to correlations with the direct effects of genetic variants. In this study, we introduce TWASKnockoff, a new knockoff-based framework for detecting causal gene-tissue pairs using GWAS summary statistics and eQTL data. Unlike marginal testing in traditional TWAS methods, TWASKnockoff examines the conditional independence for each gene-trait pair, considering both correlations in cis-predicted expression across genes and correlations between gene expression levels and genetic variants. TWASKnockoff estimates the theoretical correlation matrix for all genetic elements (cis-predicted expression across genes and genotypes for genetic variants) by averaging estimations from parametric boot-strap samples and then performs knockoff-based inference to detect causal gene-trait pairs while controlling the false discovery rate (FDR). Through empirical simulations and an application to type 2 diabetes (T2D) data, we demonstrate that TWASKnockoff achieves superior FDR control and improves the average power in detecting causal gene-trait pairs at a fixed FDR level.