To assess patterns of sharing beyond tissue pairs, we applied two Bayesian methods that assess, for each SNP-gene pair, the evidence for each of the 512 (29) possible null/alternative eQTL configurations. The first method (24) is “gene-based” and assumes a single causal eQTL per gene. We extended the model to (i) support the fact that not all tissues were sampled across all GTEx donors; (ii) calculate a gene-level FDR without requiring permutations (27); and (iii) include a fine mapping approach across multiple tissues (28). We also used a “SNP-based” approach (25), which assesses association of each SNP-gene pair separately, working directly with the z-statistics for each SNP-gene pair within each tissue [see also (29)]. We summarized the estimates of tissue specificity using marginal posterior probabilities for the number of tissues in which a randomly selected gene (gene-based model) or SNP-gene pair (SNP-based model) is active (Fig. 2C). Both approaches show a U-shaped pattern, with high tissue specificity (activity in a single tissue) or tissue ubiquity (activity in all nine tissues) more common than profiles involving only a few tissues, despite
