The architecture of gene regulatory variation across multiple human tissues: the MuTHER study

Abstract

While there have been studies exploring regulatory variation in one or more tissues, the complexity of tissue-specificity in multiple primary tissues is not yet well understood. We explore in depth the role of cis-regulatory variation in three human tissues: lymphoblastoid cell lines (LCL), skin, and fat. The samples (156 LCL, 160 skin, 166 fat) were derived simultaneously from a subset of well-phenotyped healthy female twins of the MuTHER resource. We discover an abundance of cis-eQTLs in each tissue similar to previous estimates (858 or 4.7% of genes). In addition, we apply factor analysis (FA) to remove effects of latent variables, thus more than doubling the number of our discoveries (1,822 eQTL genes). The unique study design (Matched Co-Twin Analysis--MCTA) permits immediate replication of eQTLs using co-twins (93%-98%) and validation of the considerable gain in eQTL discovery after FA correction. We highlight the challenges of comparing eQTLs between tissues. After verifying previous significance threshold-based estimates of tissue-specificity, we show their limitations given their dependency on statistical power. We propose that continuous estimates of the proportion of tissue-shared signals and direct comparison of the magnitude of effect on the fold change in expression are essential properties that jointly provide a biologically realistic view of tissue-specificity. Under this framework we demonstrate that 30% of eQTLs are shared among the three tissues studied, while another 29% appear exclusively tissue-specific. However, even among the shared eQTLs, a substantial proportion (10%-20%) have significant differences in the magnitude of fold change between genotypic classes across tissues. Our results underline the need to account for the complexity of eQTL tissue-specificity in an effort to assess consequences of such variants for complex traits.

Figure 1. P-value distribution of cis eQTLs (10⁻³ PT) gained with FA correction in the uncorrected data.

The significant overrepresentation of low p-values for the new eQTLs (π₁ = 0.99) shows that the signal existed in the uncorrected data but wasn't called significant due to low power. In each tissue, the exact SNP-gene combinations (eQTLs) tested are presented for both co-twin sets (Twin 1—first column, Twin 2—second column).

Figure 2. Proportion of tissue shared and tissue-specific eQTLs (10⁻³ PT) from the SRC analysis and SRC-FA respectively.

Both methods reveal similarly high extents of tissue-specificity. Skin specific eQTLs of smaller effects are harder to detect due to low power.

Figure 3. Fold change within twins and across tissues for LCL eQTLs (10⁻³ PT, SRC) discovered in Twin 1.

The plotted fold change on the X-and Y-axes was calculated as the difference in mean expression of the heterozygous and major homozygous genotypic classes. For each pairwise tissue comparison, the Pearson's correlation coefficient between fold changes is shown.