Investigation of prediction accuracy and the impact of sample size, ancestry, and tissue in transcriptome-wide association studies

In transcriptome-wide association studies (TWAS), gene expression values are predicted using genotype data and tested for association with a phenotype. The power of this approach to detect associations relies, at least in part, on the accuracy of the prediction. Here we compare the prediction accuracy of six different methods-LASSO, Ridge regression, Elastic net, Best Linear Unbiased Predictor, Bayesian Sparse Linear Mixed Model, and Random Forests-by performing cross-validation using data from the Geuvadis Project. We also examine prediction accuracy (a) at different sample sizes, (b) when ancestry of the prediction model training and testing populations is different, and (c) when the tissue used to train the model is different from the tissue to be predicted. We find that, for most genes, the expression cannot be accurately predicted, but in general sparse statistical models tend to outperform polygenic models at prediction. Average prediction accuracy is reduced when the model training set size is reduced or when predicting across ancestries and is marginally reduced when predicting across tissues. We conclude that using sparse statistical models and the development of large reference panels across multiple ethnicities and tissues will lead to better prediction of gene expression, and thus may improve TWAS power.