Predicting signatures of "synthetic associations" and "natural associations" from empirical patterns of human genetic variation

Abstract

Genome-wide association studies (GWAS) have in recent years discovered thousands of associated markers for hundreds of phenotypes. However, associated loci often only explain a relatively small fraction of heritability and the link between association and causality has yet to be uncovered for most loci. Rare causal variants have been suggested as one scenario that may partially explain these shortcomings. Specifically, Dickson et al. recently reported simulations of rare causal variants that lead to association signals of common, tag single nucleotide polymorphisms, dubbed "synthetic associations". However, an open question is what practical implications synthetic associations have for GWAS. Here, we explore the signatures exhibited by such "synthetic associations" and their implications based on patterns of genetic variation observed in human populations, thus accounting for human evolutionary history -a force disregarded in previous simulation studies. This is made possible by human population genetic data from HapMap 3 consisting of both resequencing and array-based genotyping data for the same set of individuals from multiple populations. We report that synthetic associations tend to be further away from the underlying risk alleles compared to "natural associations" (i.e. associations due to underlying common causal variants), but to a much lesser extent than previously predicted, with both the age and the effect size of the risk allele playing a part in this phenomenon. We find that while a synthetic association has a lower probability of capturing causal variants within its linkage disequilibrium block, sequencing around the associated variant need not extend substantially to have a high probability of capturing at least one causal variant. We also show that the minor allele frequency of synthetic associations is lower than of natural associations for most, but not all, loci that we explored. Finally, we find the variance in associated allele frequency to be a potential indicator of synthetic associations.

Figure 1. Distance of synthetic and natural associations from the causal variant it is in greatest LD with.

Box plot of the distance between any associated SNP and causal variant it is in highest LD with, measured in r², for (a) YRI and (b) CEU in four scenarios: 2 common causal variants with a GRR of 1.5 (dark blue), 2 common causal variants with an unrealistic GRR of 3 (light blue), 5 and 9 rare causal variants with a GRR of 3 (red and gold respectively). Distances vary greatly between the different disease loci (x-axis) as well as between populations, but in all regions the median (line within each box) is larger for rare causal variants than for common causal variants of lower effect size. Increasing the effect size can result in higher association distance as is observed most notably in region #5.

Figure 2. Distance of causal variant from “synthetic associations” partitioned by the age of the mutation.

Box plot similar to Figure 1, while separating rare variants in CEU and YRI into a more recent and an older class (Materials and Methods). Variants due to more recent mutations result in much increased distance between the associated SNP and the causal variant with highest LD in 3 regions in YRI and 2 regions in CEU. Results are presented for only 4 of the disease loci due to lack of relevant data in locus #1. Note that the risk allele frequency range for rare variants is narrower compared to Figure 2 (Materials and Methods) and that the y-axis scale is different between the two populations.

Figure 3. Resequence window size necessary to capture at least one causal variant.

The figure presents for a given window size, the fraction of tests combined over all regions with significant associations where at least one association is within the given distance from the causal variant it is in highest LD with. The colors correspond to the same scenarios as in Figure 1. Resequencing need not extend much further than in the common causal variant case, as a window of size of 0.1 cM has at least one association tagging a rare causal variant in >90% of the tests between both populations and all regions.

Figure 4. Minor allele frequency (MAF) of associated variants.

Box plot of the minor allele frequency for all associated variants in the different scenarios. Although synthetic associations have median MAF lower than that of natural associations, the range of MAF for synthetic associations varies across the different loci and populations. The median MAF is similar between the natural and synthetic associations for a few loci (disease locus #2 in CEU and #1 in YRI).