More recent studies have leveraged predicted expression to improve the power of TWAS. Under this paradigm, transcriptome reference data is used to predict gene expression in the GWAS data set (using cis SNPs, e.g. within 1Mb of the transcription start site), followed by a test for association between predicted expression and trait. As an alternative to TWAS using individual-level data32, TWAS using predicted expression can also be performed using only summary association statistics and summary LD information33–35. These studies respectively employed expression predictors that do not account for LD33, account for LD and allow for sparsity in eQTL effect sizes34, or utilize the top eQTL at the locus35. The key intuition is that the correlation between a weighted linear combination of SNPs (i.e. predicted gene expression) and trait is equivalent to a weighted linear combination of correlations between SNPs and trait (i.e. summary association statistics from GWAS) (see Figure 2). Since TWAS using predicted expression is conceptually similar to a test for non-zero genetic covariance between gene expression and trait34, it can also be performed via a two-sample Mendelian randomization
