Inclusion of variants discovered from diverse populations improves polygenic risk score transferability.

Accuracy of European-derived PRSs by proportion of total ancestryAccuracy of PRSs, with variants and weights from a European GWAS, decreases linearly with increasing proportion of African ancestry. Variants and weights were extracted from a GWAS of 10,000 European cases and 10,000 European controls. PRS accuracy was computed as the Pearson’s correlation between the true genetic risk and GWAS estimated risk score across 50 simulations in independent test populations of 5,000 Europeans, 5,000 Africans, and 5,000 admixed individuals. Admixed individuals were grouped based on their proportion of genome-wide European ancestry. Simulations assume 1,000 causal variants and a heritability of 0.5 to compute the true genetic risk. A p value of 0.01 and LD r2 cutoff of 0.2 was used to select variants for the estimated risk score.

PRS construction approaches and performance in admixed individualsUsing significant variants from an African ancestry GWAS with population-specific weights results in less disparity in PRS accuracy across populations. PRSs were constructed using variants and weights selected from either a European or African population (10,000 cases, 10,000 controls each) or a fixed-effects meta-analysis of both. An additional local ancestry-specific method was used for PRS weighting. Performance, measured as the Pearson’s correlation between the true and GWAS estimated risk score, is shown across 50 simulations. Simulations assume 1,000 causal variants and a heritability of 0.5 to compute the true genetic risk. A p value of 0.01 and LD r2 cutoff of 0.2 was used to select variants for the estimated risk scores.

Impact of African sample size on PRS accuracy and generalizationPRS accuracy in diverse populations can be improved by including data from an African ancestry GWAS with smaller sample sizes than in a European GWAS. The number of African samples used in the GWAS and subsequent PRS construction was decreased to reflect availability of diverse samples in real data. Analysis was conducted assuming 1%, 5%, 10%, 50%, and 100% (matched size of European dataset) of the total African ancestry cases. Average accuracy and the 95% CI were reported across the 50 simulations for different variant selection and weighting approaches. Simulations assume 1,000 causal variants and a heritability of 0.5 to compute the true genetic risk. A p value of 0.01 and LD r2 cutoff of 0.2 was used to select variants for the estimated risk score. A linear mixture of single-population PRSs (α1EUR+α2AFR), with variants and weights selected from that population, was also tested in the admixed population. The mixture coefficients (α1 and α2) were estimated in an independent African ancestry testing population.

Allele frequency distribution of GWAS selected variants and LD tagging of causal variantsGWAS significant variants are more common in the study population from which they were discovered; however, African Ancestry GWAS variants may result in better LD tagging across populations. Variants were selected from a European or African ancestry GWAS or a fixed-effects meta-analysis of both populations.(A) GWAS variants were binned by their MAF estimated from the European, African, and admixed populations. The error bar represents the 95% CI across simulations.(B) LD scores were calculated for every causal variant by adding up the LD r2 for each GWAS tag variant within ±1,000 kb of the causal variant. LD scores calculated in a Europeans and Africans were compared by Pearson’s correlation. The results were summarized across simulations as the average and 95% CI.(C) Raw LD scores for each causal variant (m = 1,000) calculated in a European or African population for one simulation. Each panel shows the approach used for variant selection. Causal variants directly discovered through the GWAS are colored in gray.