The RNA-seq reads were aligned to the human reference genome (GRCh37) using the GEM aligner19 and alignments were filtered for properly paired and uniquely mapping reads (mapping quality greater than or equal to 150). Genotypes originated from 1000 Genomes phase 1 data, which is based on 1,092 individuals with 5× whole-genome sequencing data, 80× exome sequencing data, and high-quality genotyping. The genotype data were filtered for variants with MAF < 5% and HWE p < 1 × 10−6 for each population separately and were corrected for population stratification using the first three and two eigenvectors for Europeans and Africans, respectively.7 The Altrans link counts were normalized using the first 15 principal components calculated from these link counts. We first looked at all pairwise links between exon groups considering the union of all exons in the exon group as one entity and filter so that we keep only pairs of exon groups that have 15 links in 80% of the samples. Then we count the links between exons of the initial exon group and exons of the terminal exon group and keep only links where the exon in the initial exon group made at least ten links with any of the exons in the terminal exon group in at least 30% of the samples. Cufflinks quantifications were run using the annotation with the –GTF option. In the case of Cufflinks, the transcript quantifications were converted to link quantifications and we assessed links originating from the same genes where there were Altrans quantifications. The cis-window for asQTL discovery was 1 Mb flanking the transcription start site of each gene. The associations were run with the FastQTL package.20 The observed nominal p values were calculated by correlating the genotype and link quantifications, which were Gaussian transformed. Subsequently, we ran permutations for each link separately to assign empirical p values to each link. The permutation scheme involved permuting all links of a given gene together 1,000 times and in each permutation iteration, we record the most significant p value from an association between any variant in the cis- window and any link of a given gene, thereby accounting for the dependencies among the link quantifications of a gene, allowing us to find significant asQTL genes. From this distribution of null p values we use an approximation using the beta distribution to estimate the extremes of the null p value distribution, and using this we calculate an adjusted p value. These adjusted p values are then corrected for multiple testing using the qvalue R package.21