Fig. 2

Sensitivity for trans-eQTL discovery on simulated data. We compared the performance of Tejaas reverse regression, forward regression (FR) (similar to CPMA) and MatrixEQTL, by computing the partial area under the ROC curve (pAUC) up to a false positive rate (FPR) of 0.1. A perfect method has pAUC = 0.1 and a random one 0.005. pAUCs are averaged over 20 simulations. a pAUC for different confounder correction methods: no correction (none), correction using linear regression of known confounders (CCLM) on inverse normal transformed gene expression, and our k-nearest neighbors correction with K = 30 (KNN). The gray dotted line corresponds to the random expectation (pAUC = 0.005). b pAUC for different numbers of target genes for the cis transcription factor (TF) mediating the trans-eQTL (from top to bottom) and different mean effect sizes of the TF on the target genes (from left to right). c pAUC for predicting correct SNP-gene pairs (target gene discovery) by the different methods. SNP-gene pairs were ranked by their association p-values. For Tejaas and FR, trans-eQTLs were preselected using a cutoff on their trans-eQTL p-value (legend), while MatrixEQTL does not provide any such preselection. Gray dotted line as in a. One simulation setting with mean TF effect size of 0.4 and 100 target genes (see Figure S11 for other simulation settings)