Integrative approaches for large-scale transcriptome-wide association studies.

Overview of methodsCartoon representation of TWAS approach. In the reference panel (top) estimate gene expression effect-sizes: directly (i.e. eQTL); modeling LD (BLUP); or modeling LD and effect-sizes (BSLMM). A: Predict expression directly into genotyped samples using effect-sizes from the reference panel and measure association between predicted expression and trait. B: Indirectly estimate association between predicted expression and trait as weighted linear combination of SNP-trait standardized effect sizes while accounting for LD among SNPs.

Modes of expression causalityDiagrams are shown for the possible modes of causality for the relationship between genetic markers (SNP, blue), gene expression (GE, green), and trait (red). A–D describes scenarios that would be considered null by the TWAS model; E–G describes scenarios that could be identified as significant.

Number of genes with significant cis-heritability observed at varying sample sizesThe number of genes with significant cis-heritability was estimated by down-sampling each cohort (YFS, METSIM, and NTR/Wright et al.) into quintiles.

Accuracy of direct expression imputation algorithmsAdjusted accuracy was estimated using cross-validation R^2 between prediction and true expression, and normalized by corresponding cis-h2g. Bars show mean estimate across three cohorts and three methods: eQTL – single best cis-eQTL in the locus; BLUP using all SNPs in the locus; BSLMM using all SNPs in the locus and non-infinitesimal priors.

Power of summary-based expression imputation algorithmsRealistic disease architectures were simulated and power to detect a genome-wide significant association evaluated across three methods (accounting for 15,000 eGWAS/TWAS tests, and 1,000,000 GWAS tests). Colors correspond number of causal variants simulated and methods used: GWAS where every SNP in the locus is tested; eGWAS where only the best cis-eQTL is tested; and TWAS computed using summary-statistics. Expression reference panel was fixed at 1,000 out-of-sample individuals and simulated GWAS sample size designated by x-axis. Power was computed as the fraction of 500 simulations where significant association was identified.