A unified hierarchical Bayesian approach to transcriptome-wide association study

Abstract

The transcriptome-wide association study (TWAS) has identified novel gene-trait associations, providing essential biological insights. TWAS combines reference panel transcriptome and genome-wide association study (GWAS) data. Traditional TWAS methods construct a prediction model for gene expression based on the transcriptome data, which is then employed to impute gene expression in the GWAS data. The complex trait in GWAS is regressed on the predicted expression to identify gene-trait associations. Such a two-step approach ignores the uncertainty of the imputed expression and can lead to reduced inference accuracy. We develop a unified Bayesian approach for TWAS that avoids the need for a two-step approach by modeling the two datasets simultaneously. We consider the horseshoe prior to model the relationship between gene expression and local SNPs, and the spike and slab prior to test for an association between the genetic component of expression and the trait to build an integrated Bayesian framework. We extend the method to conducting a multi-ancestry TWAS, focusing on discovering genes that affect the trait in either or both ancestries. Using extensive simulations, we demonstrate that the new approach performs better than existing methods in terms of correctly classifying non-null genes and the accuracy of effect size estimation. To demonstrate our approach, we perform a single and multi-ancestry TWAS for intraocular pressure (IOP), integrating the Geuvadis transcriptome and UK Biobank GWAS data. We find that the identified set of IOP-associated genes is enriched in relevant pathways.