If LD differences between African and European haplotypes drive the pattern seen in Figure 1, then a PRS constructed from SNPs in low recombination regions should be more transferable than a PRS constructed from SNPs in high recombination regions of the genome. Imputed data. 2019). In another approach, we tested for correlation between and LD scores (Bulik-Sullivan et al. 2009) with k = 3 and identified 7,285 individuals with self-reported “African American” ancestry with at most 0.8 of the first ADMIXTURE component (interpreted as reflecting European ancestry), and at most 0.05 of the second (reflecting Native American ancestry; Figure S2). We estimated weights separately for the SNPs present in each dataset using both the Gibbs sampler and the infinitesimal model and evaluated the partial-R2 as described above. Weighting the combination by the ancestry proportion of each individual produces a similar improvement: 3.9% for WHI_afr, 4% for JHS_afr, and 3.2% for HRS_afr (Figure 2). We first show that the predictive accuracy of height PRS increases linearly with European ancestry and is partially explained by European ancestry segments of the admixed genomes. 2015) (ρ = 0.0292, P = 0.0379) (Figure S12). Our results are broadly consistent with simulation studies showing that these two factors are expected to decrease variance explained when the test cohort has different ancestry from the GWAS cohort and, specifically, that together they explain up to 72% of the loss of accuracy in prediction between European and African ancestry (Wang et al. Datasets used in this study. Vertical bars show 95% bootstrap confidence intervals. GWAS results: We obtained UK Biobank summary statistics for height from the Neale Lab GWAS on 360,388 individuals of European ancestry (round 2; https://www.nealelab.is/uk-biobank, accessed April 2, 2019). Prediction of breast cancer risk based on profiling with common genetic variants. We repeated this analysis with imputed genotypes, unweighted PRS and sibling-estimated effect sizes. These results suggest that marginal effect sizes differ across ancestries and that this is one of the factors underlying the reduction in predictive power. 2003) (WHI), Jackson Heart Study (Taylor 2005) (JHS), and Health and Retirement Study (Sonnega et al. 2020). In , we weight in all individuals by a common factor α ranging from 0-1, and in , in addition to α, each individual’s is weighted by , the proportion of European ancestry for the individual. File S1 contains 14 supplementary figures. Finally, we constructed a combined PRS where Admixed African effect sizes are used for SNPs falling in African ancestry regions of the genome, and European effect sizes are used for SNPs falling in European ancestry regions. Each admixed dataset is split up into quantiles of European ancestry proportion. This polygenic architecture is a feature of most common diseases (Watanabe et al. We used recombination maps estimated in African Americans (AA_Map) (Hinch et al. Despite statistical methods to control for population stratification, it continues to be a confounding factor in the analysis of GWAS results (Berg et al. 2012) with a window size of 91. Effect of recombination rates on predictive power. Genotype and phenotype data were obtained from UK Biobank or dbGaP. In simulations, we can control and quantify these effects. 2017; Veturi et al. For example, 3,290 genome-wide significant loci explain approximately 25% of the phenotypic variation in height in European ancestry individuals (Yengo et al. However, this requires realistic simulations of complex traits in admixed populations. Dispersals and genetic adaptation of Bantu-speaking populations in Africa and North America. The orange lines represent the equation , for k={1,1.5,2}. The dashed line shows the regression with standard errors shaded in light gray. On the other hand, we note that, even for the quartile of lowest recombination, the reduction in partial-R2 for admixed individuals is substantial – 76% on average across datasets – compared to 84% for the fourth quantile (Figure 3A). Several explanations have been proposed to explain why the predictive power of PRS is lower in non-European ancestry populations. On the other hand, most of the ancestry of the populations that our cohorts are drawn from is West African, even other African ancestries are largely symmetrically related to Europe, and in these cohorts, the admixture proportion is the major component of variation (Zakharia et al. First, we describe the reduction in the predictive power of height PRS as a function of ancestry in populations of recently admixed African and European ancestry. 2017). For some analyses, we used between-sibling effect sizes estimated at a subset of 1,284,881 SNPs (Cox et al. We used the UKB_eur imputed genotypes as an LD reference panel and the UKB GWAS summary statistics for height. We considered SNPs below the p-value thresholds: 5×10−7, 5×10−6; 5×10−5 5×10−4, 5×10−3. The dashed line shows the regression with standard errors shaded in light gray. 2009; Tishkoff et al. The dashed line shows the regression with standard errors shaded in light gray. We next restricted the PRS SNPs to those found in segments of the genome inferred to have European ancestry (Figure 1B). Although many genome-wide significant GWAS hits do replicate in non-European ancestry cohorts (N’Diaye et al. Our members work to advance knowledge in the basic mechanisms of inheritance, from the molecular to the population level. Thus, we focused on strategies that are independent of LD and chose a set of SNPs using a p-value threshold of 0.0005 and a physical window of 100 Kb, which includes ∼5,600-7,100 SNPs (Table 1) and obtains partial-R2 values close to the LD clumping strategies while requiring about 10-fold fewer SNPs. 2019; Marnetto et al. Evaluation of genome wide association study associated type 2 diabetes susceptibility loci in sub Saharan Africans. While these absolute improvements are modest, this is likely due to GWAS sample size discrepancy (N = 8,700 Admixed African and N = 361,194 European). Reduced signal for polygenic adaptation of height in UK Biobank. 2018). In the admixed cohorts (WHI_afr, JHS_afr, HRS_afr, UKB_afr), partial-R2 was 3.1–4.1%, or between 3.8- to 5-fold lower than in HRS_eur, consistent with previous observations (Ware et al. We also used a strategy of clumping based on empirical LD structure. in addition to α, weights the African component based on individual African ancestry.