Statistical colocalization of genetic risk variants for related autoimmune diseases in the context of common controls

Abstract

Determining whether potential causal variants for related diseases are shared can identify overlapping etiologies of multifactorial disorders. Colocalization methods disentangle shared and distinct causal variants. However, existing approaches require independent data sets. Here we extend two colocalization methods to allow for the shared-control design commonly used in comparison of genome-wide association study results across diseases. Our analysis of four autoimmune diseases--type 1 diabetes (T1D), rheumatoid arthritis, celiac disease and multiple sclerosis--identified 90 regions that were associated with at least one disease, 33 (37%) of which were associated with 2 or more disorders. Nevertheless, for 14 of these 33 shared regions, there was evidence that the causal variants differed. We identified new disease associations in 11 regions previously associated with one or more of the other 3 disorders. Four of eight T1D-specific regions contained known type 2 diabetes (T2D) candidate genes (COBL, GLIS3, RNLS and BCAR1), suggesting a shared cellular etiology.

Figure 1

A Venn diagram showing summary of disease assignments to 90 regions which showed association to at least one disease, based upon the results of the Bayesian analysis. In cases where assignment was uncertain, the assignment most supported by the posterior probabilities was used. The numbers in brackets correspond to how many of these regions show evidence of distinct causal variants. Thirty–six regions analyzed did not demonstrate association to any disease within our available data, and so are not included in this figure.

Figure 2

The distribution of $\widehat{\mathrm{\eta }}$, the estimated proportionality coefficient, together with its 95% confidence interval. In the case of colocalization, η is the ratio of the effects the region exert upon the two traits. ∣η∣ > 1 corresponds to a stronger effect in Trait 2 than Trait 1. We estimate η by $\widehat{\mathrm{\eta }}$. Labels on the x–axis give the traits and regions analyzed; D for T1D, R for RA, C for CEL and M for MS. Note that in some regions, the conditional analysis supports the existence of multiple associated variants: if none of these overlap, then we consider the region to have separate SNP effects. (a) Regions with novel evidence of disease association, in which we believe there to be colocalization present between the novel association and at least one of the existing associations. Regions have been ordered such that $\widehat{\mathrm{\eta }}$ estimates the effect size for the novel trait divided by the effect size for the known association. Labels give the novel association being given first. It can be seen that the effect size tends to be smaller in the new disease. (b) Regions with strong evidence of colocalization $\left(\mathbb{P}\right({\mathbb{H}}_4)>0.9)$. As we would expect, $\widehat{\mathrm{\eta }}$ is distributed about 1, which corresponds to the regions having equal effects on each trait. Note that 6q25.3, containing the candidate causal gene TAGAP, has $\widehat{\mathrm{\eta }}<0$, indicating opposite effects on the two diseases. Trait 1 is listed first, and trait 2 second.

Figure 3

The 2q33.1 region containing the candidate gene CTLA4. Three potential causal variants are partially shared between T1D, RA and CEL. (a) A Manhattan plot of the region. The blue signal corresponds to the tag rs231775, the green to rs1980422 and the magenta to rs3087243. All other SNPs are colored according to their linkage disequilibrium with these three SNPs. SNPs rs231775 and rs3087243 have r² = 0.50; all other pairwise r² < 0.1. (b) Each possible model involving these three SNPs was tested; the four models with highest posterior probabilities, which together encompass over 90% of the total posterior probability, are shown. (c) Effect size estimates (including 95% confidence intervals) of each SNP for each disease for the most likely model. (d) Effect size estimates (including 95% confidence intervals) of each SNP for each disease for the second most likely model.

Figure 4

The 6q23.3 region containing candidate causal gene TNFAIP3. Our results show that T1D, RA and CEL all colocalize, suggesting a single shared causal variant affecting the three diseases; rs6933404 being the most likely SNP. There is also evidence of MS association, driven by a distinct causal variant. Note that this region was associated with MS at genome–wide significant levels in the analysis of the international MS dataset.