Leveraging Multi-ethnic Evidence for Mapping Complex Traits in Minority Populations: An Empirical Bayes Approach

Abstract

Elucidating the genetic basis of complex traits and diseases in non-European populations is particularly challenging because US minority populations have been under-represented in genetic association studies. We developed an empirical Bayes approach named XPEB (cross-population empirical Bayes), designed to improve the power for mapping complex-trait-associated loci in a minority population by exploiting information from genome-wide association studies (GWASs) from another ethnic population. Taking as input summary statistics from two GWASs-a target GWAS from an ethnic minority population of primary interest and an auxiliary base GWAS (such as a larger GWAS in Europeans)-our XPEB approach reprioritizes SNPs in the target population to compute local false-discovery rates. We demonstrated, through simulations, that whenever the base GWAS harbors relevant information, XPEB gains efficiency. Moreover, XPEB has the ability to discard irrelevant auxiliary information, providing a safeguard against inflated false-discovery rates due to genetic heterogeneity between populations. Applied to a blood-lipids study in African Americans, XPEB more than quadrupled the discoveries from the conventional approach, which used a target GWAS alone, bringing the number of significant loci from 14 to 65. Thus, XPEB offers a flexible framework for mapping complex traits in minority populations.

Figure 1

Decision Boundary for XPEB versus That of the Conventional Single-Population Approach S (x axis) and S′ (y axis) represent the chi-square test statistics in the target and base GWASs, respectively. The dashed magenta vertical line corresponds to the conventional approach, which uses the target GWAS alone and a genome-wide significance of p = 5 × 10⁻⁸. Solid curves delineate the decision boundaries constructed by XPEB, in which the overlap in genetic architecture (κ₁) varies; the yellow line labeled κ₁ = κ₀ corresponds to the case where trait-associated loci are independently distributed in the base and target populations. All other model parameter values are taken from the estimates obtained from the LDL data.

Figure 2

Partial Receiver Operating Characteristic under Independent Simulations with Varying Degrees of Overlap Number of true-positive (y axis) and false-positive (x axis) discoveries according to four methods—XPEB, p value in the target GWAS alone (target only), p value from a fixed-effects meta-analysis of base and target GWASs (meta-FE), and p value from a random-effects meta-analysis of base and target GWASs (meta-RE produced by METASOFT)—for δ = 1 (A), 0.75 (B), 0.5 (C), and 0.001 (D). Each simulated dataset consisted of 10⁶ independent markers, of which 1,000 causal SNPs explained 70% of the phenotypic variance; the base GWAS had a sample size of 10⁵ individuals, and the target GWAS included 10⁴ individuals. Each curve is based on the average of 100 simulations. The inset in (A) indicates that meta-analysis (meta-FE and meta-RE) performed better than XPEB at very low false-positive counts (<10) when δ = 1; otherwise, XPEB performed as well as or better than other methods considered.

Figure 3

XPEB-Estimated Overlap Parameter Values versus the True Values (A) The XPEB-estimated overlap parameter (${\hat{\kappa }}_1$, y axis) corresponds well to the true overlap parameters used in the simulation (δ, x axis) in simulations of independent loci. The red bar indicates the median of ${\hat{\kappa }}_1$ based on 100 simulations; the simulation setting is identical to that described in the legend of Figure 2. (B) In the presence of LD, XPEB under-estimated the degree of overlap. The red bar is the median of ${\hat{\kappa }}_1$ based on 20 simulations. Each simulation included ∼727,000 markers, of which 1,000 SNPs were causal and together explained 70% of the phenotypic variance. Genotypes were simulated to represent ten cohorts in the base population and one cohort in the target population (each cohort consists of 10⁴ individuals). LD in the base and target populations was simulated to resemble that of HapMap CEU and YRI, respectively.