Fig. 4

An overview of TWAS pipeline (PrediXcan [116]). The general method of TWAS is composed of three steps. First, using individual-level genotype and matching gene expression data from a reference eQTL dataset, predictive models are trained to estimate the expression level of each gene based on local genotype. Second, the models are used to predict (or “impute”) the expression level of genes (normally not captured in GWAS) for each individual-level genotype in a GWAS dataset. Third, an association test is conducted for each predicted expression with the trait to elucidate gene–trait associations. The first step has been improved by subsequent methods, which allow summary-level GWAS data as input (e.g., FUSION [117], S-PrediXcan [118], MOSTWAS [119], and UTMOST [120]). For example, UTMOST takes summary-level data and simultaneously trains models across multiple tissues to increase power [120]