Using the predicted cell-type proportions, we aimed to identify cell type-dependent eQTLs. To increase the robustness of our results, we excluded 50 samples with a cell-proportion z-score > 4 on one or more cell type and limited the analysis to eQTLs with <95% missingness per dataset, a joint MAF > 5% and a joint Hardy–Weinberg P < 0.0001. With the remaining 25,497 eQTLs and 2,633 samples, we used Decon-QTL19, which employs a non-negative least-squares model to identify cell-type interaction effects. For this analysis, we used the steps as described in the Decon-QTL manuscript19. For the pre-processing of the TMM expression counts, we corrected for dataset indicator variables, 20 RNA-seq alignment metrics and four genotype multidimensional scaling components using OLS. As an additional step, we forced the data to the normal distribution per gene to reduce outliers. Finally, we evaluated whether the multiple-testing correction applied by Decon-QTL properly reflects the null distribution by comparing a permutation-based method to the default BH-FDR multiple-testing correction. We found that the vast majority (87.76%) of FDR significant interactions were also significant using permutations (Supplementary Note and Supplementary Figs. 37,38).
