We conclude with several limitations of our approach. First, disease-impacting variants that are independent of cis-expression – in general or in the training data – will not be identified. Second, as with any prediction, the number of genes that can be accurately imputed is still limited by the training cohort size and the quality of the training data. In particular, we found that prediction accuracy did not correspond to theoretical expectations and is likely driven by data quality. The impact of these weaknesses could be better quantified as expression from larger sample sizes and a more diverse set of tissues becomes available. Although in this work we utilized both microarray and RNA-seq as measure of gene expression thus showcasing the applicability of our approach to diverse data sets, the accuracy of our method intrinsically depends on the quality of the expression measurements. For the associated genes, it remains possible that the effect is actually mediated by phenotype (i.e. SNP → phenotype → cis-expression, Figure 2F). We attempted to quantify this in the YFS data by conditioning the heritability analyses on
