Polygenic prediction across populations is influenced by ancestry, genetic architecture, and methodology

Abstract

Polygenic risk scores (PRSs) developed from multi-ancestry genome-wide association studies (GWASs), PRS_multi, hold promise for improving PRS accuracy and generalizability across populations. To establish best practices for leveraging the increasing diversity of genomic studies, we investigated how various factors affect the performance of PRS_multi compared with PRSs constructed from single-ancestry GWASs (PRS_single). Through extensive simulations and empirical analyses, we showed that PRS_multi overall outperformed PRS_single in understudied populations, except when the understudied population represented a small proportion of the multi-ancestry GWAS. Furthermore, integrating PRSs based on local ancestry-informed GWASs and large-scale, European-based PRSs improved predictive performance in understudied African populations, especially for less polygenic traits with large-effect ancestry-enriched variants. Our work highlights the importance of diversifying genomic studies to achieve equitable PRS performance across ancestral populations and provides guidance for developing PRSs from multiple studies.

Keywords: genetic architecture; genome-wide association studies; multi-ancestry; polygenic risk scores.

Graphical abstract

Figure 1

Study design in both simulations and empirical analyses (1) In the context of single-ancestry GWASs, we randomly split individuals in European (EUR) and other minority populations, including East Asian and African populations, into equally sized bins. Simulations involved a total of 52 bins per population, each containing 10,000 individuals. For empirical analysis, bin number was dependent on the sample size of the respective phenotype in that population (Table S3), with 5,000 individuals per bin. A GWAS was conducted within each bin for each individual population, followed by meta-analysis of GWASs from various numbers of bins within each population. To construct PRSs derived from single-ancestry GWASs (PRS_single) in the target population, we applied P + T for both simulations and empirical analyses, utilizing PRS-CS for the latter. Subsequently, we combined PRS_single developed from EUR GWAS (PRS_{EUR_GWAS}) and other minority population-based GWAS (PRS_{Minor_GWAS}) through a linear weighted strategy (denoted as PRS_weighted, highlighted in red box) for empirical analyses. Note that PRS_weighted was also developed using PRS-CSx, which utilizes GWAS summary statistics from multiple populations. (2) For meta-analyzed multi-ancestry GWASs (referred to as Meta), we ran meta-analyses on EUR GWASs and Minor GWASs with varying ancestry compositions. In simulations, we incrementally included 4 bins from EUR GWASs for the meta-analysis, while in empirical analyses, we increased the number to 8 bins. Simultaneously, we varied the number of bins in Minor GWASs from 1 to the total number. Following the meta-analysis, we constructed PRSs based on Meta (referred to as PRS_multi), using the P + T method for simulations, and employing both P + T and PRS-CS for empirical analyses.

Figure 2

Improvement of PRS accuracy through meta-analyzed multi-ancestry GWASs compared with large-scale EUR GWASs across 6 simulated genetic architectures The multi-ancestry GWASs included populations of EUR and East Asian (EAS) ancestry, with the EAS sample size varying as indicated on the x axis. For illustrative purposes, we present the results using 32 EUR bins, each consisting of 10,000 individuals, which were included in both EUR GWASs and multi-ancestry GWASs. The PRS was separately evaluated in African (AFR), EAS, and EUR populations. Full results are shown in Table S1. $M_c$ indicates the number of causal variants, and $h^2$ refers to SNP-based heritability. In each panel, the red vertical dashed line indicates the point where an equal number of bins from EUR and EAS populations was included in the multi-ancestry GWAS. The error bars represent the SEs of predictive accuracy differences between PRS_multi and PRS_{EUR_GWAS}.

Figure 3

Genetic architecture of 17 studied traits between the BioBank Japan and the UK Biobank The error bar is the standard deviation of the corresponding estimate. The vertical dashed line was the median estimate. Full results are shown in Table S4. The phenotypes were ranked according to their polygenicity estimates using GWASs from the UKBB, including: BMI (body mass index); height; DBP (diastolic blood pressure); SBP (systolic blood pressure); WBC (white blood cell count); lymphocyte (lymphocyte count); RBC (red blood cell count); neutrophil (neutrophil count); HB (hemoglobin concentration); HT (hematocrit percentage); eosinophil (eosinophil count); PLT (platelet count); monocyte (monocyte count); MCV (mean corpuscular volume); MCH (mean corpuscular hemoglobin); basophil (basophil count); and MCHC (mean corpuscular hemoglobin concentration).

Figure 4

Accuracy improvement of PRS in the UKBB-EAS population using multi-ancestry GWASs compared with using EUR GWASs for P + T and PRS-CS The multi-ancestry GWASs were obtained by meta-analyzing EUR GWASs and EAS GWASs, with the EAS sample size from the BBJ varying as indicated on the x axis. For illustrative purposes, we present the results using 64 EUR bins, each containing 5,000 individuals, which were included in both EUR GWASs and multi-ancestry GWASs. The y axis is the accuracy difference of PRSs when using multi-ancestry GWASs (PRS_multi) compared with using EUR GWASs (PRS_{EUR_GWAS}). The error bars indicate the SE of accuracy improvement. The red dashed line is y = 0. We showed the results for 7 traits with SNP-based heritability >0.1 in both the BBJ and the UKBB, and they were ranked by polygenicity estimates using the UKBB (Figure 3). Abbreviations are the same as in Figure 3. Full results are shown in Table S7.

Figure 5

Predictive accuracy using different PRS methods in the UKBB-EAS population We showed the results for 7 traits with SNP-based heritability >0.1 in both the BBJ and the UKBB. Traits were ranked by polygenicity estimates using the UKBB (Figure 3). Boxes represent the first and third quartiles, with the whiskers extending to 1.5-fold the interquartile range. Abbreviations are the same as in Figure 3. Full results are shown in Tables S8 and S9.

Figure 6

Accuracy of PRSs derived from local ancestry-informed GWASs vs. other discovery GWASs in the UKBB-AFR population AFR_Tractor denotes the AFR-specific GWAS performed using Tractor on the UKBB admixed AFR-EUR individuals. EUR_standard refers to standard GWASs performed on the EUR population in the UKBB. Meta_standard is the meta-analysis performed on AFR_Tractor and EUR_standard. Furthermore, we constructed a weighted PRS by combining PRSs generated from AFR_Tractor and EUR_standard through a linear weighted approach. The figure shows the results for traits with SNP-based heritability >0.1 in the UKBB-AFR. Full results are shown in Table S10.

Figure 7

General practices for developing PRSs using different discovery GWASs We summarized the general practice for developing PRSs (A) using single-ancestry GWASs (PRS_single) and (B) using GWASs from multiple ancestries (PRS_multi or PRS_weighted). $r_g$, cross-ancestry genetic correlation; $h_d^2$ and $h_t^2$, SNP-based heritability in discovery and target populations, respectively; $N_d$, discovery GWAS sample size; $M_d$, the number of genome-wide independent segments in the discovery population.