A TWAS Method Calibrated for the Uncertainty of Predicted Expression

The transcriptome-wide association study (TWAS) is a powerful approach to identifying novel genes associated with complex phenotypes. Standard TWAS approaches build a prediction model for the genetic component of expression based on reference transcriptome data. Next, an outcome is regressed on the predicted expression in separate GWAS data. The traditional TWAS approach disregards the uncertainty of predicted expression, which can lead to unreliable inference on gene-phenotype associations. We propose a novel approach that adjusts for the uncertainty of predicted expression in TWAS. We adapt techniques from measurement error theory and implement bootstrapping algorithms for penalized regression to explicitly obtain an adjustment factor to be incorporated into the unadjusted TWAS. We base the framework on adaptive Lasso. Using extensive simulations, we show that our approach produces more accurate estimates of the gene's effect size than a traditional TWAS approach. Traditional TWAS marginally inflates the type I error rate, whereas the adjusted TWAS adequately controls it. At the expense of a modestly inflated false positive rate, an unadjusted TWAS offers a limited increase in power compared to the adjusted TWAS. Simulations show that some other unadjusted and unified TWAS approaches can also inflate the type I error rate, particularly when using summary-level data. We demonstrate the merits of the proposed adjusted approach by conducting TWAS for height and lipid phenotypes, LDL and HDL cholesterol, and triglycerides while integrating the Geuvadis transcriptome and UK Biobank GWAS data.