Dissecting the Predictive Accuracy of Polygenic Indexes for Behavioral Phenotypes Across Genetic Ancestries

TL;DR

Polygenic indexes lose accuracy when applied to non-European ancestries, especially for behavioral traits, with different factors causing this loss depending on ancestry.

Contribution

The study systematically analyzes PGI portability across ancestries for behavioral and health-related traits and compares standard and family-based GWAS-based PGIs.

Findings

PGI predictive power drops significantly for non-European ancestries, with African ancestry showing the largest loss.

Biologically proximal traits show better portability than behavioral and social traits.

Family-based GWAS PGIs modestly improve portability for some traits like BMI in African ancestry.

Abstract

Polygenic indexes (PGIs) trained on samples of European genetic ancestries often lose substantial predictive power when applied to non-European ancestries. While this portability problem is well recognized, its manifestation in behavioral and social traits remains understudied, and the factors driving this accuracy loss warrant more comprehensive analysis. Using data from the UK Biobank and Health and Retirement Study, we conduct a systematic analysis of PGI portability for 52 health-related, behavioral, and social phenotypes. We advance prior literature by using genome-wide PGIs, assessing cross-ancestry heritability differences, and comparing the performance of PGIs based on standard versus family-based GWAS. Our findings confirm systematic reductions in PGI predictive power for non-European ancestries—lowest in African (24%), followed by East Asian (37%) and South Asian (51%) genetic ancestries—with biologically proximal traits exhibiting greater portability than behavioral and social traits. We show that the relative importance of factors underlying reduced portability varies across traits and ancestries: in African ancestries, linkage disequilibrium and allele frequency differences explain most of the loss (82%), compared with smaller contributions in East (34%) and South Asian (25%) ancestries. Finally, we find that family-based GWAS PGIs can modestly improve portability for select traits, such as BMI in African ancestry, suggesting that part of the portability gap may reflect population-specific confounds in standard PGIs.

Figures

Equations

Full text

Introduction

1

Genome-wide association studies (GWASs) have primarily been conducted on individuals of European (EUR) genetic ancestries, resulting in a lack of genetic diversity in GWAS datasets^1–3^. Recent estimates indicate that less than 14 percent of all GWAS participants are from non-EUR genetic ancestries, a disparity with significant implications for the generalizability of genetic research^4^. Although recent efforts have increased the representation of Asian ancestry groups in genetic studies, African, Latin American, and other populations remain largely underrepresented, contributing to persistent data imbalances^5–7^. These imbalances impact the predictive accuracy of polygenic indexes (PGIs, also called polygenic scores). Specifically, the predictive accuracy of a PGI has been shown to be substantially smaller when applied to samples with a different genetic ancestry than the GWAS discovery sample^2,8–10^. Martin et al.^2^ demonstrated that PGIs trained on EUR-genetic-ancestry discovery samples show a predictive accuracy reduction of approximately 37%, 50%, and 78% for individuals of South Asian (SAS), East Asian (EAS), and African (AFR) genetic ancestries, respectively, across 17 biologically-proximal traits. Privé et al.^11^ extended these findings by analyzing 245 traits, showing that PGI predictive accuracy not only diminishes across ancestries but also varies within continental ancestries as a function of genetic distance from the training population, a finding confirmed by Ding et al.^12^ for over 80 phenotypes in the genetically diverse Los Angeles Biobank (ATLAS).

The limited portability of PGIs across genetic ancestries poses a significant challenge, potentially exacerbating existing disparities as the benefits of genetic research may not be equitably distributed among populations^2,3,13–16^. As an initial step toward addressing this challenge, it is essential to understand the factors contributing to the reduced predictive accuracy of PGIs across different ancestries. Previous studies have suggested that factors such as cross-population differences in minor allele frequency (MAF), linkage disequilibrium (LD) between causal and tagging SNPs^8^, SNP heritability, and variability in causal SNP effect sizes (due to gene-environment interactions or population-specific causal variants)^17–20^ may contribute to the limited portability of PGIs. To quantify the contribution of LD and MAF differences to the cross-ancestry relative accuracy (RA) of PGIs, Wang et al.^8^ developed a theoretical framework which they then applied to 8 biologically proximal phenotypes in UK Biobank. In this study, we extend Wang et al.’s findings by investigating the predictive accuracy of PGIs across ancestries for 54 health-related, behavioral, social, and cognitive phenotypes in two cohorts. Since social and behavioral phenotypes are likely to be more strongly influenced by complex gene-environment interplay^21–24^, we anticipate that PGIs for these traits may exhibit even weaker portability than those for more biological proximal traits. To test this, we compare the cross-ancestry relative accuracy of PGIs across different phenotype categories, assessing whether the decline in predictive accuracy is more pronounced for social and behavioral traits. We further investigate whether the contribution of LD and MAF differences to loss of predictive accuracy differs across phenotype categories.

In addition to including a wide range of phenotypes, we extend Wang et al.’s analyses in several ways. First, we use PGIs based on weights for ~ 2.9 million SNPs adjusted for LD using the SBayesR methodology^25^ as opposed to unadjusted weights for only genome-wide significant SNPs. Second, our weights come from the Social Science Genetic Association Consortium’s Polygenic Index Repository^26^ and are based on the largest available GWAS samples for most traits. This allows us to analyze traits for which the UK Biobank subsample used by Wang et al. would not be a sufficiently large discovery sample. Thirdly, we extend the predictive framework by investigating how variability in SNP heritability across ancestries influences relative accuracy, rather than assuming constant heritability as in previous studies. These refinements allow for a more comprehensive understanding of the factors affecting PGI portability for complex traits.

A fourth extension considers how potential confounds in PGIs based on between-family GWAS SNP weights, hereafter standard PGIs, affect their cross-ancestry portability. Between-family GWAS SNP weights can be biased due to passive gene-environment correlation, including population stratification and indirect genetic effects from relatives, and assortative mating^27–30^. These biases can influence PGI predictive accuracy^31–34^ and if gene-environment correlations and assortative mating patterns differ across ancestries, they can also affect relative accuracy^33^. To assess this possibility, we also analyze PGIs based on family-based GWAS SNP weights, –fGWAS PGIs– which mitigate these biases by leveraging the random segregation of alleles within families. Comparing the cross-ancestry relative accuracy of standard and fGWAS PGIs allows us to test whether such confounds differ by ancestry and, in turn, to better understand the factors limiting PGI portability.

Finally, we extend our analyses from the UK Biobank (UKB) to the Health and Retirement Study (HRS). This cross-cohort comparison allows us to clarify whether conclusions regarding PGI portability across diverse ancestries hold under different demographic and environmental contexts. The two cohorts differ substantially in their design and composition. UKB recruited a large sample of middle-aged adults (40–69 years) who were, on average, healthier and more affluent than the general UK population, a common feature of volunteer-based cohorts^35,36^. In contrast, HRS was designed as a nationally representative sample of US adults over the age of 50, capturing a wider range of health and socioeconomic circumstances that more closely reflect the general population approaching retirement^37^.

In what follows, we first present the relative accuracy of PGIs in non-European genetic ancestries, then examine several factors that may underlie their reduced predictive power: (i) cross-ancestry differences in LD and MAF, (ii) differences in heritability, and (iii) variation in gene–environment correlation and assortative mating. We conclude with a discussion of the implications of our findings.

Results

2

Cross-Ancestry Predictive Accuracy of PGIs

2.1

We computed standard PGIs for 54 phenotypes across UKB and HRS (34 phenotypes are present in both cohorts, 16 are unique to UKB, and 4 are unique to HRS), while fGWAS PGIs were available for a subset of 24 of these UKB phenotypes (Supplementary Tables 1, 2 and 9). To assign individuals to one of four genetic ancestries—European (EUR), South Asian (SAS), East Asian (EAS), or African (AFR)—we first estimated principal component (PC) loadings from the 1000 Genomes Phase 3 reference panel^38^. We then projected study participants onto this PC space and assigned ancestry by comparing each participant’s first 10 PCs to the mean values for each 1000 Genomes ancestry (Methods). In UKB, sample sizes were sufficiently large for all four ancestries (162,963 EUR, 11,413 SAS, 2,216 EAS, 9,494 AFR). In HRS, however, the SAS and EAS subsamples were too small to permit reliable analyses, so we limited our analyses to the EUR (12,774) and AFR (3,593) groups (Table 1). Following Wang et al.^8^, we omit the AMR ancestry from our analyses because of their complex admixture patterns. For the standard PGIs, we use between-family GWAS weights for ~ 2.9 million SNPs from the second release of Polygenic Index Repository^26^. These weights are adjusted for LD using the SBayesR methodology^25^. The fGWAS PGIs are based on family-based GWAS weights from Tan et al.^39^ for HapMap3^40^ SNPs, adjusted for LD using PRS-CS^41^. All weights are based on EUR-genetic-ancestry GWAS that excluded the target samples. Further details on PGI computation are provided in the Methods. (Section 4.7).

We measure predictive accuracy using “incremental– $[eqn]$ ”: the increase in the coefficient of determination ( $[eqn]$ ) when the PGI is added to a regression of the phenotype on the first 20 PCs of the genomic relatedness matrix (GRM) and also batch dummies in UKB. Prior to these regressions, we residualize the phenotypes on a third-degree polynomial in birth year, sex, and their interactions (Methods). To assess the loss in predictive accuracy when analyzing PGIs based on EUR-genetic-ancestry GWAS in non-EUR-genetic-ancestry samples, we compute the observed relative accuracy ( $[eqn]$ ). Specifically, $[eqn]$ *-*to-EUR relative accuracy ( $[eqn]$ ) is defined as the ratio of the incremental $[eqn]$ in $[eqn]$ -ancestry to that in EUR-ancestry samples within the target cohorts (UKB and HRS):

$$[eqn]$$

where, $[eqn]$ and $[eqn]$ denote the incremental $[eqn]$ of the PGI for EUR and $[eqn]$ -genetic-ancestry populations, respectively, with $[eqn]$ . To obtain the 95% confidence intervals (CIs) for incremental $[eqn]$ and $[eqn]$ , we bootstrap with 1000 replications and take the 2.5th and 97.5th percentiles as the lower and upper bounds, respectively.

Estimates of the predictive accuracy of standard PGIs across genetic ancestries in our target samples are presented in Supplementary Tables 1–2. Figures 1 and S1 and Supplementary Tables 4–5 show the relative accuracies. Across all 47 phenotypes in UKB, the average standard PGI relative predictive accuracy ( $[eqn]$ ) is lowest in the African genetic ancestry group (28.03%, S.E. = 5.5), followed by East Asian (31.48%, S.E. = 3.4) and South Asian genetic ancestries (54.18%, S.E. = 4.7). In HRS, the average $[eqn]$ is 23.31% (S.E. =6.8) across 33 phenotypes. Supplementary Figures S2 and S3 provide a detailed comparison of the incremental $[eqn]$ estimates between EUR and non-EUR genetic ancestries for each phenotype in the UKB and HRS cohorts, respectively. Overall, our findings are consistent with previous studies by Martin et al.^2^ and Wang et al.^8^, which assessed the relative predictive accuracy of standard PGIs for a smaller set of biologically proximal traits (17 and 8 biologically proximal traits, respectively) in the UKB cohort. Expanding the analysis to 46 phenotypes, we find a smaller average relative predictive accuracy for standard PGIs in East Asian and South Asian ancestries compared to Martin et al., while the results for African ancestry are comparable.

To better understand differences in standard PGI relative accuracy across traits, we disaggregate the results by phenotype categories. In the UKB, the relative accuracy ( $[eqn]$ ) is consistently highest in South Asian (SAS) ancestry, followed by East Asian (EAS) and African (AFR) ancestries (Figure 2). The highest mean $[eqn]$ for SAS (0.92; 95% CI: 0.12, 1.72) and AFR (0.88; 95% CI: −0.16, 1.92) ancestries is observed for psychiatric traits in UKB, driven mainly by the schizophrenia PGI being more predictive in these ancestries which have a higher case prevalence compared to EUR ancestry. In EAS, the highest mean $[eqn]$ is observed for blood biomarkers (0.62; 95% CI: 0.58, 0.66), which also rank second for SAS (0.65; 95% CI: 0.59, 0.71) and AFR (0.36; 95% CI: 0.26, 0.46). By contrast, the lowest $[eqn]$ in all ancestries is found for cognition and education—AFR (0.12; 95% CI: 0.10, 0.14), SAS (0.43; 95% CI: 0.33, 0.53), and EAS (0.15; 95% CI: 0.11, 0.19).

In HRS, the highest mean $[eqn]$ is observed for substance use traits (0.57; 95%CI : −0.12, 1.26), largely driven by a greater predictive power of the PGI for alcohol misuse in AFR relative to EUR (Supplementary Table 5). This is followed by blood biomarkers (0.27; 95%CI : 0.19, 0.385, while the lowest predictive accuracy is observed for psychiatric conditions (0.08; 95%CI : −0.04, 0.20) and for fertility and sexual development (0.11; 95%CI : 0.00, 1.57)..

To formally summarize the extent of these differences in UKB, we applied Welch’s ANOVA to test whether mean $[eqn]$ differs across ancestry groups within each category. While we expect some level of cross-ancestry variation for all categories, we wanted to see if these differences are sufficiently large relative to within-category variability . We observed nominally significant differences in blood biomarkers, health, and fertility and sexual development. After Benjamini–Hochberg correction^42^ (FDR = 5%), blood biomarkers and health remained significant (adjusted $[eqn]$ and 0.045), whereas fertility and sexual development did not (Supplementary Table 6).

Our results are broadly consistent with prior studies, including Wang et al.^8^ and Martin et al.^2^, in that the average $[eqn]$ declines with increasing genetic distance from the reference group (EUR): highest in SAS, followed by EAS, and lowest in AFR ancestries. However, for some phenotypes, our estimates of $[eqn]$ differ from those reported by previous studies. For example, we find considerably higher $[eqn]$ for LDL-cholesterol in AFR and EAS but lower in SAS relative to Wang et al.^8^. For height, we observe consistently higher $[eqn]$ across all non-EUR ancestries, while for BMI, our estimates are lower across the board. In the case of HDL-cholesterol, our estimate exceeds that of Wang et al.^8^ only in AFR. For asthma, we observe uniformly lower $[eqn]$ across all ancestries. These divergences may be due to methodological differences, including the use of genome-wide SNPs (as opposed to genome-wide significant SNPs) and SNP weights derived from larger discovery samples in our study. For phenotypes whose GWS SNPs exhibit larger cross-ancestry MAF and LD differences compared to more weakly associated SNPs, all else being equal, we would expect our $[eqn]$ estimates to be larger. To provide a glimpse into the genetic architecture of the traits analyzed, Supplementary Figures S8–S17 display the effect sizes and ancestry-specific minor allele frequencies (MAF) of top associated SNPs, disaggregated by phenotype category.

Comparisons with Martin et al.^2^ are more limited due to smaller phenotype overlap, but our estimates are lower for BMI and higher for both height and educational attainment in AFR genetic ancestry. Here, it is important to note that these cross-study differences should only be interpreted descriptively, as they are not based on formal statistical tests. Such tests would require joint re-estimation of both sets of $[eqn]$ values on the same individual-level data to properly account for sampling covariance. Collectively, however, these contrasts underscore how PGI construction choices and differences in discovery samples can influence cross-ancestry PGI portability, even when general patterns remain aligned across studies.

To better understand how potential confounds in standard PGIs influence cross-ancestry prediction in UKB, we compared their performance to that of fGWAS PGIs, which are constructed using weights that are not biased by indirect genetic effects from relatives, population stratification, and assortative mating. While these biases can inflate predictive accuracy in the discovery population, they may not translate well across divergent genetic backgrounds—potentially reducing cross-ancestry portability. We present the incremental $[eqn]$ and $[eqn]$ for the fGWAS PGIs in Supplementary Tables 9 and 11, respectively.

Because the fGWAS have much smaller effective sample sizes, the incremental- $[eqn]$ values from fGWAS PGIs are generally smaller. In many cases, the non-EUR incremental- $[eqn]$ is statistically indistinguishable from zero (19/24 phenotypes in AFR, 8/24 in SAS, and 17/24 in EAS). Despite this, several traits exhibit comparable or higher $[eqn]$ in fGWAS PGIs relative to their standard GWAS counterparts. Most notably, the relative accuracy of the fGWAS PGI for BMI in AFR is significantly higher than that of the standard PGI, with $[eqn]$ (95% CI: 0.23, 0.46) for fGWAS versus 0.05 (95% CI: 0.03, 0.09) for the standard PGI; this difference remains statistically significant after multiple-testing correction ( $[eqn]$ ). We also observe nominally significant improvements in $[eqn]$ for BMI in SAS ( $[eqn]$ ), systolic blood pressure in AFR ( $[eqn]$ ), and ever smoking in EAS ( $[eqn]$ ), though these do not survive multiple testing correction (Figure 3). For the remaining traits, $[eqn]$ differences are not statistically distinguishable from zero, but the point estimates from fGWAS PGIs tend to be higher across most traits and ancestries (Supplementary Table 11). Details of the paired nonparametric bootstrap test for differences in $[eqn]$ between the two approaches (and the multiple-testing correction) are provided in Supplementary Note, Section 5.2.1.

Analysis of Factors Influencing PGI Predictive Accuracy

2.2

Cross-ancestry LD and MAF differences

2.2.1

To evaluate the contribution of different factors to the loss of cross-ancestry PGI predictive accuracy, we have generalized the approach proposed by Wang et al.^8^. Similar to Wang et al., we use a model that expresses the predicted PGI relative accuracy as a function of cross-population differences in LD, MAF, SNP-based heritability, as well as the cross-population correlation of causal SNP effect sizes. Under this model, Wang et al., derived the expected relative accuracy of PGIs ( $[eqn]$ ) as:

$$[eqn]$$

where $[eqn]$ is the correlation of causal SNP effect sizes between populations 1 and 2, $[eqn]$ is the SNP-based heritability in population $[eqn]$ , $[eqn]$ is the minor allele frequency (MAF) of PGI-SNP $[eqn]$ in population $[eqn]$ ; $[eqn]$ is the mean squared correlation of allele counts between PGI-SNP $[eqn]$ and all “candidate causal SNPs” in LD with it in population $[eqn]$ , $[eqn]$ is the mean product of the correlation of allele counts between PGI-SNP $[eqn]$ and all “candidate causal SNPs” in LD with it in populations 1 and 2, $[eqn]$ is the effect size of PGI-SNP $[eqn]$ as estimated in the discovery GWAS, and $[eqn]$ is the number of SNPs in the PGI.

We deviate from Wang et al. in the set of PGI-SNPs and “candidate causal SNPs” used in Equation 2. Wang et al. focus on PGIs constructed using only independent genome-wide significant (GWS) SNPs, based on two key considerations: first, genome-wide significant SNPs are more likely to pinpoint causal variants compared to weaker associations; second, prior studies have shown that including sub-significant SNPs reduces the cross-ancestry relative accuracy of PGIs^2,9,43,44^. Relying on results from a prior simulation study^45^, they define “candidate causal SNPs” as SNPs in LD ( $[eqn]$ ) with GWS-SNPs within a 100 kb window. In contrast, we construct PGIs using Bayesian methodologies that incorporate a much larger set of common SNPs. For the main analyses, we use weights from the second release of PGI Repository, obtained using the SBayesR methodology with ~ 2.9 million SNPs^25^. For the fGWAS PGIs, we use the weights for ~ 1.2 million HapMap3 SNPs^40^ generated by Tan et al.^39^ using PRS-CS^41^. The reasons behind this are two-fold, concerning feasibility and practical implications. First, we consider a wider set of behavioral and health-related phenotypes in this study. These phenotypes are highly polygenic, with each SNP having a very small effect size, and current largest GWAS available for many of these phenotypes are only able to identify a handful of GWS SNPs. Consequently, relying solely on GWS-SNPs to calculate the predicted relative accuracy may yield imprecise results regarding the relevant cross-ancestry differences that drive the loss. Second, most PGI studies construct PGIs with a focus on maximizing predictive power for the phenotype. Although, as Wang et al. state, the accuracy of GWS-based PGIs will get closer to that of genome-wide PGI methodologies as GWAS sample sizes become larger, we are not there yet. Therefore, we wanted to generalize and assess Wang et al.’s model under practically more relevant conditions.

Including more than a million SNPs in the PGIs irrespective of their association p-values leaves us with the challenge of defining “candidate causal SNPs”. We need a set of SNPs that would approximate the genome-wide level expected predictive accuracy of PGIs. Because the PGI-SNPs are not selected based on their p-values and are not pairwise independent, we cannot assume that causal SNPs are located within a 100kb window and are in LD with the PGI-SNPs. Therefore, we decided to use Wang et al.’s approach of defining candidate causal SNPs in relation to the most significant independent SNPs, but relaxing the p-value cutoff to consider more than only GWS SNPs. The challenge then becomes including enough SNPs to correspond to the predictive power of genome-wide PGIs while keeping the computational burden at a manageable level. In order to gauge how the expected relative accuracy changes in relation to the number of candidate causal SNPs included in the model, we generated three different candidate causal SNP sets for each phenotype based on the top 100, 1,000, and 10,000 independent SNPs by p-value (Section 4.8). We then expanded each set to include any SNPs in LD ( $[eqn]$ ) within 100 kb of these top SNPs. Using Equation 2, we computed the expected relative accuracy of the PGIs based on each candidate causal SNP set, fixing $[eqn]$ at unity. For the fGWAS PGIs, we repeated the procedure selecting top SNPs based on fGWAS p-values (Methods).

Our evaluations reveal that the expected relative accuracy ( $[eqn]$ ), modeled based on cross-ancestry differences in MAF and LD, is generally stable after the top 1,000 SNPs for most phenotypes, whether computed with standard GWAS or fGWAS candidate SNP sets (Figure 4, Supplementary Figures S4-S6). Using the top 10,000 SNPs provides the most precise estimates for both approaches; thus, we use this set as a proxy for genome-wide expectations in subsequent analyses. For the standard GWAS-based candidate causal SNPs, the average $[eqn]$ across 54 phenotypes was 36% for AFR, 79% for EAS, and 87% for SAS genetic ancestry. The fGWAS-based SNPs yielded similar averages for SAS (87%) and EAS (77%), with a modest increase for AFR (39%). Ancestry-specific $[eqn]$ values estimated based on the top 10,000 candidate causal standard GWAS-based SNPs for individual phenotypes are provided in Supplementary Table 3, while those based on fGWAS SNPs are shown in Supplementary Table 10.

After estimating $[eqn]$ and $[eqn]$ , we calculate the percentage loss in PGI predictive accuracy attributable to cross-ancestry differences in LD and MAF, denoted as $[eqn]$ . We compute this measure only for phenotypes that exhibit a $[eqn]$ statistically smaller than one, indicating meaningful shrinkage in the PGI’s predictive accuracy. For these phenotypes, $[eqn]$ is derived as the ratio of the expected loss to the observed loss in PGI accuracy, as formalized in Equation (3). A value of 100% indicates that differences in LD and MAF fully account for the reduction in predictive accuracy. Values below 100%—where $[eqn]$ —suggest that the decline in predictive performance is larger than what LD and MAF differences alone would predict, pointing to additional contributing factors such as imperfect genetic correlations between the target and reference group ( $[eqn]$ ) or lower heritability in the target population ( $[eqn]$ ). Conversely, values above 100%—where $[eqn]$ —indicate that the observed loss is smaller than expected based on LD and MAF, implying that other factors may be partially compensating for the predicted reduction. One plausible explanation is that the heritability of the trait is higher in the target population than in the discovery cohort ( $[eqn]$ ), which would enhance the observed predictive accuracy beyond what is predicted from LD and MAF differences alone.

$$[eqn]$$

Our analysis of standard GWAS-based PGIs in the UKB cohort reveals substantial differences in the mean loss of predictive accuracy due to LD and MAF across non-EUR genetic ancestries and phenotype categories (Figure 5, Supplementary Table 4). The mean $[eqn]$ is highest in AFR, averaging 83.04% (S.E. = 2.9%) across 41 phenotypes, followed by EAS at 32.50% (S.E. = 2.3%) over 32 phenotypes, and SAS at 24.90% (S.E. = 1.9%) across 36 phenotypes. Breaking down by phenotype category, $[eqn]$ is most pronounced for blood biomarkers across all non-EUR ancestries: 106.05% (S.E. = 8.8%) in AFR, 47.08% (S.E. = 6.3%) in EAS, and 35.59% (S.E. = 6.8%) in SAS. In contrast, the lowest mean $[eqn]$ values are observed for fertility and sexual development traits in AFR (69.70%; S.E. = 1.1%) and EAS (26.25%; S.E. = 0.5%), and for substance use traits in SAS (17.94%; S.E. = 1.5%). Beyond these patterns, differences across the remaining categories appear modest, with anthropometric traits showing a slightly elevated mean $[eqn]$ in EAS and SAS (Figure S7).

Zooming in on where $[eqn]$ is largest - blood biomarkers - we see pronounced heterogeneity within the category (Figure S7). In EAS, blood pressure phenotypes (systolic: 70.89%, S.E. = 6.1%; diastolic: 61.3%, S.E. = 7.3%; pulse: 56.41%, S.E. = 7.9), and HDL cholesterol (57.46%, S.E. = 7.7%) show outstandingly high $[eqn]$ values. In SAS, the largest $[eqn]$ is observed for triglycerides (71.69%, S.E. = 12.7%), followed by HDL cholesterol (47.91% (S.E. = 7.3%). In contrast, LDL and non-HDL cholesterol exhibit some of the lowest $[eqn]$ values in both EAS (LDL: 20.81%, S.E.=15.3; non-HDL: 24.80%, S.E.=17.3) and SAS (LDL: 15.80%, S.E. = 8.5; non-HDL: 15.42%, S.E. = 7.3) ancestries, which is to be expected given that LDL and non-HDL cholesterol are known to be more strongly influenced by lifestyle factors compared to HDL cholesterol^46^. Interestingly, this pattern does not hold in AFR ancestry, where LDL (133.61%, S.E.=6.05%) and non-HDL (126.41%, S.E.=5.75%) cholesterol exhibit the highest $[eqn]$ values, with HDL cholesterol following a few phenotypes behind (102.70%, S.E.=3.4).

The lowest $[eqn]$ is observed for COPD, the leading cause of which is cigarette smoking, in both EAS (9.67%, S.E.=6.8) and SAS (9.92%, S.E.=5.0) ancestries. COPD is also one of the lowest for AFR (68.66%, S.E.=4.8). Prostate cancer has the lowest $[eqn]$ in AFR genetic ancestry (62.96%, S.E.=2.5), ranks second lowest in SAS (12.06%, S.E.=2.4) and fourth lowest in EAS (62.96%, S.E.=2.5) (Supplementary Table 4 and Figure S7).

For a small subset of traits, we observe $[eqn]$ values exceeding 100% only in AFR. The phenotypes with $[eqn]$ significantly above 100% (95% CI entirely >100%) are: Life Satisfaction—Family (111.75%; 95% CI: 105.15, 118.35), total cholesterol (126.41%; 95% CI: 115.14, 137.69), non-HDL cholesterol (133.61%; 95% CI: 121.75, 145.47), and LDL cholesterol (142.42%; 95% CI: 130.50, 154.34). Similarly, a few other phenotypes had point estimates above 100% but were not statistically greater than 100%—namely, smoking cessation (101.06%; 95% CI: 96.46, 105.65), HDL cholesterol (102.70%; 95% CI: 95.96, 109.43), and type-II diabetes (105.43%; 95% CI: 99.67, 111.18). Consistent with this pattern, Wang et al. also report >100% values for certain traits in AFR (e.g., LDL cholesterol: 124.90% with S.E.=10.5%; asthma: 107.3% with S.E.=27.0%). $[eqn]$ higher than >100%) indicates that factors beyond LD and MAF may be positively influencing the observed predictive accuracy of PGIs for these traits. In subsequent sections, we explore whether accounting for additional factors such as heritability differences or using fGWAS-based approaches to compute PGIs alters this scenario.

Overall, our findings are in line with Wang et al.. Although there are substantial differences in point estimates for some phenotypes such as asthma in AFR and BMI in EAS where our estimates are lower and type-II diabetes in AFR and height in EAS where our estimates are higher, these estimates are contained within the substantially wider 95% confidence intervals reported by Wang et al.. Exceptions to this are LDL cholesterol in EAS and SAS, and height in AFR and SAS ancestries. Our $[eqn]$ values for LDL cholesterol in EAS and SAS reported above are much lower compared to Wang et al.’s, who found 97.6% (S.E. = 23.8%) for EAS and 42.1% (S.E. = 2.7%) for SAS ancestries. For height, we find larger $[eqn]$ values for AFR (82.54%, S.E. = 1.8% vs. 71.50%, S.E. = 1.8%) and SAS (30.32%, S.E. = 2.4% vs. 23.6%, S.E. = 1.8%).

As with the differences in relative accuracy estimates, the differences in $[eqn]$ estimates between our study and Wang et al. likely stem from methodological distinctions, particularly in SNP selection and PGI methodology. Lupi et al.^47^ demonstrate that $[eqn]$ and $[eqn]$ vary substantially across the genome, even within a given ancestry. They show that certain genomic regions—referred to as “high portability” segments—exhibit consistently strong cross-ancestry predictive accuracy, including in populations such as AFR where genome-wide $[eqn]$ is typically low. Lupi et al. also report that the $[eqn]$ estimates are markedly lower in these high portability regions across traits and ancestries. However, it is important to note that for certain phenotypes (height, LDL- and HDL-cholesterol), we observe substantially higher $[eqn]$ in our study relative to Wang et al., while the $[eqn]$ based on MAF and LD differences is similar across studies. In fact, we find that for most phenotypes, $[eqn]$ due to MAF and LD differences alone is relatively stable after the top 1,000 SNPs, and inclusion of more high portability areas in the model would increase both $[eqn]$ and $[eqn]$ . Several other explanations are possible. The simplest is that adjusting the SNP weights for LD improves the cross-ancestry portability of PGIs by getting closer to the causal effect sizes. Other explanations could include causal effects being more heterogeneous across ancestries for the top SNPs, or top SNPs not being representative of the whole genome in terms of the contribution of gene-environment correlation. , e.g.

We next estimated standard PGI $[eqn]$ values for AFR ancestry in HRS. The highest $[eqn]$ in HRS is observed for COPD (104.39%, S.E. = 7.4%), followed by smoking cessation (97.17%, S.E. = 2.4%), systolic blood pressure (88.99%, S.E. = 1.6%), and height (88.48%, S.E. = 1.9%). The lowest $[eqn]$ estimates are for coronary artery disease (65.62%, S.E. = 1.7%) followed by a suit of behavioral phenotypes: subjective well-being (67.70%, S.E. = 1.4%), family satisfaction (71.26%, S.E. = 2.0%) and depressive symptoms (71.92%, S.E. = 1.4%). For a direct comparison with UKB, we plotted the $[eqn]$ estimates for AFR in both cohorts, displaying point estimates alongside their 95% confidence intervals (Figure 6). In UKB, $[eqn]$ estimates for migraine, subjective well-being, and depression are significantly higher than those in HRS after correcting for multiple testing using the Benjamini–Hochberg procedure^42^ (FDR = 5%). Conversely, HRS shows significantly higher $[eqn]$ for phenotypes including BMI, educational attainment, drinks per week, height, cigarettes per day, and neuroticism. These differences suggest that the contribution of MAF and LD to PGI predictive accuracy may vary across cohorts, potentially reflecting differences in environmental exposures or sample characteristics. In some cases, substantial differences in phenotype definitions may also contribute to the observed discrepancies.

Cross-ancestry heritability differences

2.2.2

Next, we account for cross-ancestry differences in SNP-based heritability to further elucidate the combined effect of MAF, LD, and heritability ( $[eqn]$ ) in explaining the shrinkage of standard PGI predictive accuracy. To obtain the $[eqn]$ , we adjust Equation (3) by multiplying the $[eqn]$ by the ratio of the SNP-based heritability estimates in the target non-EUR and EUR genetic-ancestry samples which we estimate using BOLT-REML (Methods). Given the relatively small non-EUR sample sizes in both UKB and HRS cohorts, heritability estimates for most traits were imprecise, yielding large standard errors (Supplementary Tables 1 and 2). This imprecision resulted in $[eqn]$ estimates that were statistically indistinguishable from zero for many traits, particularly for SAS and EAS genetic ancestries in UKB.

Within the AFR ancestry of UKB, the $[eqn]$ estimates for BMI, height, and HDL cholesterol were calculated with reasonable precision and were all higher than the corresponding $[eqn]$ values. For height, adding heritability to the LD+MAF model raised the share of variance explained in the loss of accuracy to nearly 100%, indicating that LD, MAF, and $[eqn]$ together account for almost the entire reduction in PGI predictive power relative to EUR (Figure 7). In contrast, HDL and LDL cholesterol remained above 100% after accounting for heritability. This pattern suggests that factors beyond LD, MAF, and heritability contributing to the predictive accuracy of standard-GWAS PGIs such as passive gene–environment correlations or assortative mating may have inflated $[eqn]$ in AFR ancestry, producing $[eqn]$ and, consequently, $[eqn]$ .

In the HRS cohort, we could estimate $[eqn]$ with acceptable precision only for height and educational attainment; for height, the pattern mirrored UKB, while educational attainment exhibited a similar but non-significant rise.

Overall, for phenotypes for which we were able to estimate AFR-genetic-ancestry SNP-based heritability relatively precisely, our findings suggest that accounting for cross-ancestry differences in SNP-based heritability largely addresses the remaining unexplained shrinkage in PGI predictive accuracy (Figure 7). However, the persistence of $[eqn]$ estimates exceeding 100% for traits such as HDL and LDL cholesterol motivates our subsequent analysis, where we assess whether PGIs derived from family-based GWAS which are not affected by passive gene–environment correlations and assortative mating can help resolve this discrepancy.

LoA(LD+MAF) for fGWAS based PGIs

2.2.3

Finally, we employ fGWAS-based PGIs to re-compute $[eqn]$ . To compute $[eqn]$ , we start by computing $[eqn]$ using fGWAS. For AFR ancestry, $[eqn]$ decreases by more than 5 percentage points for BMI, height, HDL-cholesterol, extraversion, subjective well-being, and cannabis use. For most other phenotypes in AFR, values either decrease or remain stable, with the exceptions of educational attainment and drinks per week, which increase. For SAS, substantial decreases (¿5 percentage points) are observed for height, non-HDL cholesterol, migraine, and subjective well-being. In contrast, increases of similar magnitude occur for age at first birth, nearsightedness, and drinks per week. For EAS, most phenotypes show higher $[eqn]$ values compared to $[eqn]$ estimated using standard-GWAS. Increases of more than 5 percentage points are observed for age at first birth, nearsightedness, morning person, and depressive symptoms, while only height and migraine show substantial decreases. Across all ancestries, only three phenotypes—educational attainment, nearsightedness, and drinks per week—consistently exhibit increases in $[eqn]$ . Among these, drinks per week shows particularly pronounced gains.

We estimate the $[eqn]$ only for phenotypes where $[eqn]$ in a non-EUR ancestry was significantly lower than 1 ( $[eqn]$ ), indicating reduced predictive accuracy relative to the EUR reference group. In AFR, we find $[eqn]$ values ranging from 142.19% (S.E. = 16.1%) for non-HDL cholesterol to 61.49% (S.E. = 6.3%) for cigarettes per day; in SAS, from 132.11% (S.E. = 21.5%) for cognitive performance to 16.61% (S.E. = 4.5%) for age at first menses; and in EAS, from 46.74% (S.E. = 12.8%) for age at first birth to 16.59% (S.E. = 5.7%) for hayfever. Estimates with 95% confidence intervals for the full list of phenotypes are reported in Supplementary Table 11.

Relative to standard PGIs, fGWAS-based estimates of $[eqn]$ show both upward and downward shifts depending on the ancestry–phenotype combination. In AFR ancestry, $[eqn]$ goes down for the majority of phenotypes. In SAS, there are changes in both directions, and in EAS, most $[eqn]$ values go up. These contrasts indicate that fGWAS-based PGIs can meaningfully alter the relative contribution of LD and MAF to cross-ancestry prediction loss, with the direction and magnitude of these changes varying by phenotype and ancestry (Figure 8).

A particularly salient observation is that the $[eqn]$ estimate for non-HDL cholesterol in AFR consistently exceeds 100% with both standard-GWAS and fGWAS-based PGIs—including when heritability is added to the standard-GWAS model—suggesting that the observed predictive accuracy in this subgroup is higher than what LD, MAF, and $[eqn]$ differences alone would predict (Figure 8). This persistent observation of $[eqn]$ exceeding 100% may reflect inaccuracies in one or more parameters used to compute $[eqn]$ . For instance, if LD patterns are more similar than assumed, the LD similarity term $[eqn]$ would be underestimated thereby lowering $[eqn]$ . Overestimation of MAF differences would similarly depress $[eqn]$ by inflating the denominator of $[eqn]$ . Inaccurate assumptions about effect-size variances—expressed as $[eqn]$ —could further bias $[eqn]$ downward if causal effects explain more variance in AFR than predicted. These factors collectively may lead to $[eqn]$ exceeding 100%.

Discussion

3

In this study, we provide a comprehensive analysis of the factors driving the loss of predictive accuracy of PGIs when PGIs trained on European genetic ancestry samples are applied to other genetic ancestries. Analyzing an extensive set of 54 phenotypes, we expand beyond prior work by comparing the portability of PGIs for biologically proximal traits with more distal behavioral and social traits. Building on the prior framework by Wang et al.^8^, we introduce several methodological advances - including the use of genome-wide SNPs, accounting for ancestry-specific heritability estimates, and a comparison of standard versus family-based GWAS PGIs - to provide a more detailed account of PGI portability. Our findings confirm substantial reductions in the predictive accuracy of PGIs for non-European ancestries, with the lowest observed in African, followed by East Asian and South Asian genetic ancestries, consistent with prior studies^2,8,9,12^. Furthermore, we show that this loss is not uniform across trait categories. Specifically, we find that PGI portability is substantially lower for behavioral and social traits compared to more biologically proximal phenotypes. This pattern likely reflects their greater sensitivity to environmental, cultural, and socio-economic influences, which may interact with genetic effects in population-specific ways.

Our core findings are broadly consistent with Wang et al.^8^, showing that differences in LD and MAF account for a substantial portion of accuracy loss, particularly in African genetic ancestry (83%) compared to much lower contributions in East Asian (34%) and South Asian (25%) ancestries. However, we also observe notable differences in the estimated contribution of these factors for certain traits, which likely stem from methodological distinctions. Our use of genome-wide PGIs, rather than PGIs based only on GWS SNPs, enables a more comprehensive evaluation of polygenic contributions and provides a better benchmark for PGI applications, the majority of which are based on genome-wide PGIs. It also allows us to analyze PGIs for traits that would have too little predictive power if constructed solely from GWS SNPs, enhancing our ability to capture ancestry-specific patterns of predictive accuracy loss for behavioral traits.

Our results for African genetic ancestry align with recent work by Hu et al.^49^, who show widespread conservation of causal effect sizes and conclude that factors like LD and MAF are likely the primary drivers of PGI performance differences. However, our study provides the critical additional insight that this conclusion may not generalize across all non-European populations. The substantially smaller role of LD and MAF in East and South Asian ancestries suggests that other factors—such as cross-ancestry differences in SNP-based heritability or imperfect genetic correlations—are likely the dominant contributors to PGI accuracy loss in these groups.

Another key contribution of our study is the use of fGWAS-based PGIs to assess how PGI portability is affected by the standard-GWAS SNP weights being confounded by passive gene-environment correlation and assortative mating. We find that fGWAS-based PGIs can improve portability for few traits, most notably for BMI in individuals of African genetic ancestry, suggesting that some of the portability gap may be attributable to population-specific biases present in standard PGIs. This finding is in contrast with recent findings by Zhang and Conley^50^, who reported no improvement in cross-ancestry prediction accuracy in HRS and Add Health when using fGWAS-based PGIs for a number of traits including BMI. However, for the majority of traits analyzed, our findings align with Zhang and Conley^50^ showing that the relative accuracy between the standard and fGWAS-based approaches is broadly consistent.

Collectively, our findings present an interesting puzzle. While the overall loss of PGI accuracy follows a simple gradient corresponding to genetic distance from the reference group (EUR sample), our analysis suggests that portability of the PGI will vary considerably within a certain genetic ancestry depending on the phenotype. Behavioral phenotypes appear to be less portable across ancestries, and for these phenotypes, the relative contribution of cross-ancestry LD and MAF differences to the loss in predictive accuracy is smaller. Still, cross-ancestry LD and MAF differences appear to explain the majority of accuracy loss in AFR genetic ancestry, where the loss of accuracy is the greatest.

While our study provides valuable insights into the key factors influencing PGI portability, several aspects warrant further investigation. First, our analyses were limited to common SNPs (MAF ¿ 0.01 in each genetic ancestry), which do not fully capture the contribution of rare variants. Rare variants are often population-specific and poorly imputed in reference panels that lack adequate representation of non-European genetic ancestries. Incorporating rare variants and improving imputation strategies for diverse populations could alter the cross-ancestry portability of PGIs. Second, the use of genotype reference panels like the 1000 Genomes which are biased toward European ancestry, may introduce inaccuracies in estimating population parameters such as MAF and LD structure, particularly for non-European ancestries. These inaccuracies could lead to over- or under-estimation of the contribution of various factors, such as LD and MAF, to the loss of PGI predictive accuracy. Increasing the diversity and size of reference panels is crucial for accurate characterization of the genetic architecture of complex traits across genetic ancestries, enabling more robust analyses. Finally, while fGWAS-based PGIs showed promise for improving portability for certain traits, their smaller effective discovery sample sizes and stringent QC filters limit power and genome-wide SNP coverage. Expanding family-based GWAS datasets in scale and coverage could enhance the robustness of fGWAS-based approaches, further advancing our understanding of genetic architecture.

Methods

4

Study Cohorts

4.1

Our analyses were conducted using two prospective longitudinal studies with genomic data: the UK Biobank (UKB) and the Health and Retirement Study (HRS). Table 1 provides the number of genotyped samples by study cohort and genetic ancestry.

Discovery Samples:

For this study, we utilized GWAS summary statistics derived from two sources for the standard and family-based GWAS (fGWAS) approaches.

The GWAS summary statistics and PGI weights for the standard GWAS approach were obtained from the PGI-Repository version 2^26^, maintained by the Social Science Genetic Association Consortium (SSGAC). For UKB, we used the GWAS summary statistics that were generated for the third partition (UKB3) as defined by the Repository, which consists of one third of the sample after excluding individuals of non-European genetic ancestries and includes only individuals with no third-degree or closer relatives. The summary statistics are based on meta-analyses of GWAS from up to three sources, all of which include only individuals of European genetic ancestries: 23andMe, UKB, and published GWA studies. To avoid sample overlap, discovery meta-analyses for the UKB target cohort excludes the UKB3 partition (leveraging UKB1-UKB2, 23andMe, and published GWAS as applicable), and discovery meta-analyses for the HRS target cohort excludes HRS; Supplementary Table 8 details, by phenotype and cohort, the contributing GWAS sources and the total sample sizes.

The summary statistics for the fGWAS approach were obtained from the largest meta-analyzed fGWAS dataset recently released by Tan et al.^39^. The discovery sample includes up to 16 cohorts and is restricted to individuals of European genetic ancestries. Because neither of our prediction cohorts, UKB and HRS, are family samples, the fGWAS discovery samples do not include them.

Prediction Samples:

For the standard GWAS-based approach, we used both UKB and HRS as prediction cohorts. For the fGWAS-based approach, only UKB was used as the prediction sample.

In the UKB, the target sample comprised individuals of European genetic ancestries from the third partition (UKB3) and individuals of non-European ancestries. This resulted in a composition of 162,963 individuals of European (EUR) ancestries, 11,413 of South Asian (SAS) ancestries, 2,216 of East Asian (EAS) ancestries, and 9,494 of African (AFR) ancestries. In the Health and Retirement Study (HRS), the target sample included 12,774 individuals of EUR and 3,593 of AFR genetic ancestries. South Asian (SAS, N=87) and East Asian (EAS, N=162) ancestries were excluded from the HRS target sample due to their small sample sizes. Descriptive statistics for all phenotypes in these prediction (target) cohorts are reported in Supplementary Table 12.

Genotyping and Imputation

4.2

The details of genotyping and imputation for UKB and HRS can be found in references^51^ and^37^, respectively.

Identification of Genetic Ancestries

4.3

We follow a PCA-based approach to identify genetic ancestries. To estimate the PCs, in each cohort, we first restricted the genotypes to HapMap3 SNPs^40^ and converted the dosages to hard calls. Then, we merged these genotypes with the full 1000 Genomes Phase 3 reference sample^38^, keeping only SNPs that had a call rate > 99% and minor allele frequency > 1% after the merge. We estimated the loadings for the first 10 PCs in the 1000 Genomes subsample and then projected the remaining samples onto this PC space. We assigned an individual to a genetic ancestry as defined by the 1000 Genomes Project if each of their 10 PCs fell within four standard deviations of the average for that ancestry in the 1000 Genomes sample. Following Wang et al.^8^, we excluded individuals of American genetic ancestry (AMR) due to their complex patterns of genetic admixture.

Construction of PC Controls

4.4

We generated ancestry-specific principal components to control for population stratification when estimating the explanatory power (incremental R2) of PGIs. These PCs were obtained using a procedure different than the PCs used for ancestry identification. Prior to generating the PCs, in each ancestry, we removed SNPs meeting any of the following criteria: (1) call rate < 99; (2) MAF < 0.01; (3) HWE p-value < 10^−5^; (4) imputation accuracy < 0.7; (5) SNPs in long-range LD blocks in EUR ancestry (chr5:44mb–51.5 mb, chr6:25mb–33.5 mb, chr8:8mb–12mb and chr11:45mb–57mb). We pruned the remaining SNPs using a 1 Mb rolling window incremented in steps of 5 variants using a cutoff $[eqn]$ . Using these approximately independent variants, we constructed a genomic relatedness matrix in PLINK 1.9^52^ to identify pairs of individuals with a relatedness coefficient above 0.05. We excluded one individual from each such pair, estimated the PC loadings for the first 20 PCs in the sample of unrelated individuals that was obtained, and then projected the remaining individuals onto this PC space.

Phenotypes

4.5

We started with a set of 61 phenotypes available in the second release of PGI Repository, aggregated into seven categories: biomarkers, anthropometric traits, cognition and education, personality and well-being, health-related traits, fertility and sexual development, psychiatric conditions, and substance use. We analyzed a PGI if the phenotype was available in the target cohort and for binary phenotypes, the case proportion within each genetic ancestry was greater than 1%. 47 phenotypes satisfied these criteria in UKB and 33 in HRS. A full list of phenotypes and relevant inclusion criteria details are provided in Supplementary Tables S1 and S2.

Prior to analysis, we residualized all phenotypes on a set of covariates. If multiple measurements across time were available, we first obtained the standardized residuals from a regression in each wave of the phenotype on sex (unless the phenotype was sex-specific), a second-degree polynomial in age at the time of measurement, and their interactions, and then take the average of these residuals. This averaged phenotype was residualized a second time on the third-degree polynomial in birth year, sex and their interactions. If only a single measurement was available or the phenotype was defined using the maximum recorded value, the phenotype was residualized on sex, a third-degree polynomial in birth year, and their interactions. A more detailed description of phenotype definitions, pre-processing, and handling of repeated measures is provided in Supplementary Table 7.

Estimation of SNP-based heritability

4.6

We estimated SNP-based heritability ( $[eqn]$ ) for each phenotype within each ancestry group in UKB and HRS cohorts using the REML algorithm implemented in BOLT-LMM software (v2.3.4)^53^. To reduce computational burden, we randomly sampled 50,000 unrelated individuals from the European ancestry group in our UKB estimation sample for this analysis. For all other genetic ancestries in UKB and HRS, we used the full set of unrelated individuals identified through kinship filtering ( $[eqn]$ ) using PLINK v1.9^54,55^. We restricted the set of SNPs to those present in the HapMap 3 reference panel^40^ and filtered for MAF ≥ 1%, genotype missingness ≤ 15%, and individual-level missingness ≤ 15%. Phenotypes were residualized prior to analysis on covariates as described in the previous section, except that we additionally inlcuded the first 20 principal components of the genetic relatedness matrix (GRM) and, for UKB only, genotyping batch effects.

Computation of Polygenic Indexes (PGIs)

4.7

For the standard GWAS-based approach, we used weights from the second release of PGI Repository to construct the PGIs^26^, following the same methodology. These weights were obtained for ~ 2.9 million pruned common variants from the full UKB European-genetic-ancestry ( $[eqn]$ ) data set from Lloyd-Jones et al.^25^ by adjusting the estimated effects for LD using the SBayesR methodology implemented in GCTB software^25,56^. PGIs were computed using PLINK2^57^ using genotype dosages. Details on the input GWAS included in the discovery meta-analyses and their respective sample sizes for each phenotype are provided in Supplementary Table 8.

For the fGWAS-based approach, SNP weights were obtained from the Tan et al. family-based GWAS^39^. These weights were adjusted for LD using PRS-CS^41^ and variants were restricted to HapMap3 SNPs^40^. The remaining steps were the same as the construction of standard-GWAS based PGIs.

SNP Selection for LD Correlation Analysis

4.8

For each phenotype, we started by LD-clumping the GWAS that was used to obtain the weights for the PGI after restricting the set of SNPs to those included in the PGI. The algorithm, implemented in PLINK2^57^ starts by selecting the SNP with the lowest association p-value. SNPs within a 2000kb window that are correlated ( $[eqn]$ ) with the index SNP and had p-values below 0.5 are clumped together with the index SNP. The process iteratively continues by selecting the SNP with the lowest p-value among those that are not yet assigned to a clump and repeating the clumping steps until no SNPs with p-value < 0.5 remain. From this list of approximately independent SNPs, we extracted three subsets with the lowest p-values: the top 100, 1,000, and 10,000 SNPs.

Next, we identified SNPs common across the four genetic ancestries (African [AFR], South Asian [SAS], East Asian [EAS], and European [EUR]) from the 1000 Genomes Project Phase 3 reference panel^38^ after applying the following filters within each genetic ancestry: call rate > 95%, MAF > 1%, Hardy-Weinberg equilibrium (HWE) p-value > 10^−10^, and subject-level missingness < 1%. 4,576,403 SNPs were available in all four genetic ancestries after the filters. We restricted the 1000 Genomes data for each ancestry to this set of SNPs. Then we computed a genetic relatedness matrix using GCTA^58^ for each ancestry and excluded one individual from each pair that had a relatedness coefficient greater than 0.05.

Then, for each top SNP set (top 100, 1,000, and 10,000 SNPs), we defined candidate causal SNPs as SNPs available in the QC’d 1000 Genomes data^38^, which are within 100kb of a SNP in the top SNP set and that have ( $[eqn]$ ) with it.

To compute the LD-related summary statistics required for inferring the LD-correlation parameters in the $[eqn]$ formula (Equation 2), we used ldcorpair, a C++ program developed by Wang et al.^8^ and available at their GitHub repository^1^. This program calculates two key LD-based parameters. The first one, $[eqn]$ , is the average product of the LD between the $[eqn]$ -th PGI-SNP (i.e. k’th SNP available in the PGI weights) and candidate causal SNPs within a 100kb window of it. The second parameter, $[eqn]$ , is the mean squared correlation of allele counts between the $[eqn]$ -th PGI-SNP and the candidate causal SNPs within a 100kb window. We estimated these two parameters for each ancestry using the QC’d 1000 Genomes Project data described above. We repeated the process for each set of 100, 1,000, and 10,000 top SNPs that we identified.

Supplementary Material

Supplement 1

Supplementary Files

This is a list of supplementary files associated with this preprint. Click to download.

• SuppTablesAlemuetal.xlsx