Evaluation and application of methodology for omic imputation into genome-wide association studies of complex human traits to infer potential causal mechanisms

Abstract

To date, genome-wide association studies (GWAS) have been successful at identifying associations between common genetic variants and complex traits. However, little is known about the mechanisms by which trait-associated variants identified through GWAS affect the traits. One method developed to address this problem is the transcriptome-wide association study (TWAS), in which known relationships between genotypes and gene expression are leveraged to impute gene expression levels into GWAS samples. These imputed gene expression levels are then tested for association with traits to identify potentially causal trait-associated genes. Here, I investigated a number of ways of improving TWAS to enable the detection of more trait-associated genes, before extending the TWAS approach to investigate other omics measurements. First, to identify the best software for conducting TWAS, a range of packages were compared through application to data from the Geuvadis and Wellcome Trust Case Control Consortium projects. Overall, the investigated packages predicted gene expression with similar accuracy and detected similar expression-trait associations, although some tested a broader set of genes, so were preferable. Following this, the accuracy with which gene expression could be predicted from genotype data was investigated by comparing different statistical modelling approaches using data from the Geuvadis project. Overall, the expression of most genes could not be predicted accurately using any approach, but the best estimates were achieved when using approaches that assumed sparsity. Furthermore, prediction accuracy was improved by increasing sample size and by carefully matching training and testing data in terms of ancestry and tissue. Finally, the TWAS approach was extended to investigate the prediction of other omics measurements from genotype data. By generating prediction models for these omics measurements and applying these models to publicly available GWAS data, many associations between omics measurements and complex traits were detected, improving understanding of the mechanisms underlying GWAS risk loci