Adjustment for the uncertainty of predicted expression in transcriptome-wide association study: a fusion of measurement error theory and bootstrapping

Abstract

Transcriptome-wide association study (TWAS) is a promising approach to identifying novel genes associated with complex traits. The method commonly used in a standard TWAS approach combines two subsequent regressions. A significant criticism of the method is that the uncertainty of the predicted expression is ignored in the second-step regression, which can lead the final inference on the gene trait association subject to an uncontrolled rate of false positives, reducing the reliability of the detected association signals. We propose a novel approach to adjust for the uncertainty of predicted expression in TWAS. We improvise techniques from measurement error theory to derive the adjustment factor that needs to be incorporated in the test statistic obtained from standard TWAS. Next, we apply bootstrapping techniques for penalized regression to estimate the adjustment factor. We use simulations to show that the traditional TWAS approach inflates the type 1 error rate, whereas our adjusted TWAS approach adequately controls it. Due to an inflated false positive rate, the standard TWAS produces a higher power than the adjusted TWAS approach. For a multi-trait TWAS, ASSET has an excessive type 1 error rate based on the traditional TWAS statistics, contrary to the adjusted TWAS statistics. Overall, the proposed TWAS approach is a valid test for the gene-trait association while adequately controlling the false positive rate.