Leveraging local ancestry to detect gene-gene interactions in genome-wide data

Abstract

Background: Although genome-wide association studies have successfully identified thousands of variants associated to complex traits, these variants only explain a small amount of the entire heritability of the trait. Gene-gene interactions have been proposed as a source to explain a significant percentage of the missing heritability. However, detecting gene-gene interactions has proven to be very difficult due to computational and statistical challenges. The vast number of possible interactions that can be tested induces very stringent multiple hypotheses corrections that limit the power of detection. These issues have been mostly highlighted for the identification of pairwise effects and are even more challenging when addressing higher order interaction effects. In this work we explore the use of local ancestry in recently admixed individuals to find signals of gene-gene interaction on human traits and diseases.

Results: We introduce statistical methods that leverage the correlation between local ancestry and the hidden unknown causal variants to find distant gene-gene interactions. We show that the power of this test increases with the number of causal variants per locus and the degree of differentiation of these variants between the ancestral populations. Overall, our simulations confirm that local ancestry can be used to detect gene-gene interactions, solving the computational bottleneck. When compared to a single nucleotide polymorphism (SNP)-based interaction screening of the same sample size, the power of our test was lower on all settings we considered. However, accounting for the dramatic increase in sample size that can be achieve when genotyping only a set of ancestry informative markers instead of the whole genome, we observe substantial gain in power in several scenarios.

Conclusion: Local ancestry-based interaction tests offer a new path to the detection of gene-gene interaction effects. It would be particularly useful in scenarios where multiple differentiated variants at the interacting loci act in a synergistic manner.

Fig. 1

Simulation schemes. Main and interaction effects are simulated assuming either a single genetic variant per locus (a) or multiple genetic variants per locus (b). In the latter case, the main and interaction effects on the outcome Y are moderated through two latent variables Z₁ and Z₂ that directly depend on the causal variants. Example of local ancestry derived for the two haplotypes of three individuals (c)

Fig. 2

Power comparison for a single causal SNP per locus. Upper panels show the sample size required for 80 % power for the interaction test based on 1 M genotyped GWAS SNPs (S _G) (a), and the interaction test based on local ancestry segment (S _L) assuming a total of 1 K local ancestry segments (b) against the interaction test based on full sequencing data (S _S) assuming a total of 20 M genetic variants (blue curve). Sample size is plotted for increasing ρ _GC and ρ _GL (defined by the red gradient), the correlation between the true interaction term and the best tag from 1 M genotyped SNPs, and the best tag from local ancestries, respectively. The variance explained by the interaction effect is unrealistically large for illustration purposes and varied between 1 and 10 %. Lower panels show the observed distribution of ρ _GC (c) and ρ _GL (d) for a randomly selected region from the 1 M Illumina chip and local ancestry, respectively

Fig. 3

Tagging interaction effects in a multiple causal model. A latent variable Z is generated as a function of an increasing number of SNPs at a single locus, explaining altogether 50 % of its variance. The average value of Z across 20,000 replicates of 10,000 admixed samples is plotted for each three local ancestry classes. The effect of the SNPs is drawn from a normal (a) and left-truncated normal (b) distribution with a mean of 0 (upper panel). When the SNP effects are null on average, the average values of Z do not differ by local ancestry and ρ _ZL, the correlation between Z and local ancestry, is also null on average. Conversely, when the average effect of the SNPs is not null, ρ _ZL increases with an increasing number of causal variants (lower panel)

Fig. 4

Power comparison for multiple causal SNPs per locus. Power across 25,000 replicates using a Bonferroni correction resulting in p-value thresholds of 1 × 10⁻⁷ and 1 × 10⁻¹⁵ for the local ancestry-based interaction test (S _L) and the SNP-based interaction test (S _G), respectively. One to five common causals SNPs were selected per interacting locus while assuming either low (a), moderate (b) or high (c) differentiation of those SNPs between the two admixed populations. We considered three case scenarios for the additional increase in sample size that would be achieve when using local ancestry derived from AIMs, no increase (pink), a lower bound of six fold increase (light red) and an upper bound of 10 fold increase. We varied the baseline sample size (for S _G) across scenarios to emphasize the differences between the tests