On the interpretation of transcriptome-wide association studies

Transcriptome-wide association studies (TWAS) aim to detect relationships between gene expression and a phenotype, and are commonly used for secondary analysis of genome-wide association study (GWAS) results. Results from TWAS analyses are often interpreted as indicating a genetic relationship between gene expression and a phenotype, but this interpretation is not consistent with the null hypothesis that is evaluated in the traditional TWAS framework. In this study we provide a mathematical outline of this TWAS framework, and elucidate what interpretations are warranted given the null hypothesis it actually tests. We then use both simulations and real data analysis to assess the implications of misinterpreting TWAS results as indicative of a genetic relationship between gene expression and the phenotype. Our simulation results show considerably inflated type 1 error rates for TWAS when interpreted this way, with 41% of significant TWAS associations detected in the real data analysis found to have insufficient statistical evidence to infer such a relationship. This demonstrates that in current implementations, TWAS cannot reliably be used to investigate genetic relationships between gene expression and a phenotype, but that local genetic correlation analysis can serve as a potential alternative.Author summary: The aim of transcriptome-wide association studies (TWAS) is to find genetic relationships between the expressions of genes and specific human traits of interest, and they do so using large-scale genetic association results from existing studies of those traits. However, the statistical methods traditionally used to perform these TWAS studies do not directly test such genetic relationships, and in this study we examine the implications of this limitation. Using both simulations as well as analyses of real data, we demonstrate that when traditional TWAS methods are interpreted as testing a genetic relationship between gene expression and a trait, this results in a considerable inflation in the false positive rates for these analysis. This suggests that these TWAS methods cannot reliably be used to study such genetic relationships, with their statistically valid interpretion being much more limited. We also show that methods such as local genetic correlation analysis can serve as potential alternative.