Polygenic scoring accuracy varies across the genetic ancestry continuum

Abstract

Polygenic scores (PGSs) have limited portability across different groupings of individuals (for example, by genetic ancestries and/or social determinants of health), preventing their equitable use^1,2,3. PGS portability has typically been assessed using a single aggregate population-level statistic (for example, R²)⁴, ignoring inter-individual variation within the population. Here, using a large and diverse Los Angeles biobank⁵ (ATLAS, n = 36,778) along with the UK Biobank⁶ (UKBB, n = 487,409), we show that PGS accuracy decreases individual-to-individual along the continuum of genetic ancestries⁷ in all considered populations, even within traditionally labelled ‘homogeneous’ genetic ancestries. The decreasing trend is well captured by a continuous measure of genetic distance (GD) from the PGS training data: Pearson correlation of −0.95 between GD and PGS accuracy averaged across 84 traits. When applying PGS models trained on individuals labelled as white British in the UKBB to individuals with European ancestries in ATLAS, individuals in the furthest GD decile have 14% lower accuracy relative to the closest decile; notably, the closest GD decile of individuals with Hispanic Latino American ancestries show similar PGS performance to the furthest GD decile of individuals with European ancestries. GD is significantly correlated with PGS estimates themselves for 82 of 84 traits, further emphasizing the importance of incorporating the continuum of genetic ancestries in PGS interpretation. Our results highlight the need to move away from discrete genetic ancestry clusters towards the continuum of genetic ancestries when considering PGSs.

Main

PGSs—estimates of an individual’s genetic predisposition for complex traits and diseases (that is, genetic liability; also referred to as genetic value)—have garnered tremendous attention recently across a wide range of fields, from personalized genomic medicine^4,8,9,10 to disease risk prediction and prevention^11,12,13,14 to socio-genomics^3,15. However, the variation in PGS performance across different genetic ancestries and/or socio-demographic features (for example, sex, age and social determinants of health)² poses a critical equity barrier that has prevented widespread adoption of PGSs. Similar portability issues have also been reported for non-genetic clinical models^16,17,18. The interpretation and application of PGSs are further complicated by the conflation of genetic ancestries with social constructs such as nationality, race and/or ethnicity. Here we investigate PGS performance across genetically inferred ancestry (GIA), which describes the genetic similarity of an individual to a reference dataset (for example, 1000 Genomes¹⁹) as inferred by methods such as principal component analysis (PCA); GIAs do not represent the full genetic diversity of human populations.

Genetic prediction and its accuracy (or reliability) have been extensively studied in agricultural settings with a focus on breeding programmes^20,21,22,23. At the population level, PGS accuracy can be expressed as a function of heritability, training sample size and the number of markers used in the predictor in single^24,25,26 or multi-population settings with or without effect size heterogeneity²⁷. At the individual level, accuracy of genetic prediction from pedigree data^28,29,30 can be derived as a function of the inverse of the coefficient matrix of mixed-models equations, whereas accuracy of genetic prediction using whole-genome genetic data can be derived similarly, with the pedigree matrix replaced with the genomic relationships matrix^{21,22,23,27,31,32} among training and testing individuals. Simulations guided by dairy breeding programmes showcase that genomic prediction accuracy varies with genetic relatedness of the testing individual to the training data^33,34 as well as across generations, owing to the decay of genetic relationships³⁵.

In humans, PGS performance evaluation has traditionally relied on population-level accuracy metrics (for example, R²)^2,4. PGS accuracy decays as the target populations become more dissimilar from the training data using either relatedness^36,37 or continental or subcontinental ancestry groupings^1,38,39,40; the decay may be explained by differences in linkage disequilibrium, minor allele frequencies and/or heterogeneity in genetic effects due to gene–gene and gene–environment interactions⁴¹. However, population-level metrics of accuracy provide only an aggregate (average) metric for all individuals in the population, thus implicitly assuming some level of homogeneity across individuals^2,4,42. Homogeneous populations are an idealized concept that only roughly approximate human data; human diversity exists along a genetic ancestry continuum without clearly defined clusters and with various correlations between genetic and socio-environmental factors^{7,42,43,44,45,46}. Grouping individuals into discrete GIA clusters obscures the impact of individual variation on PGS accuracy. This is evident among individuals with recently admixed genomes for which genetic ancestries vary individual-to-individual and locus-to-locus in the genome. For example, a single population-level PGS accuracy estimated across all African Americans overestimates PGS accuracy for African Americans with large proportions of African GIA⁴⁰; likewise, coronary artery disease PGS performs poorly in Hispanic individuals with high proportions of African GIA⁴⁷. The genetic ancestry continuum affects PGS accuracy even in traditionally labelled ‘homogeneous’ or ‘non-admixed’ populations. For example, PGS accuracy decays across a gradient of subcontinental ancestries within Europe as the target cohorts become more genetically dissimilar from the PGS training data^39,45. Assessing PGS accuracy using population-level metrics is further complicated by technical issues in assigning individuals to discrete clusters of GIA. Different algorithms and/or reference panels may assign the same individual to different clusters^39,42,48, leading to different PGS accuracies. Moreover, many individuals are not assigned to any cluster owing to limited reference panels used for genetic ancestry inference^5,39, leaving such individuals outside PGS characterization. This poses equity concerns as it limits PGS applications only to individuals within well-defined GIAs.

Here we leverage classical theory^28,29,30 and methods that characterize PGS performance at the level of a single target individual⁴⁹ to evaluate the impact of the genetic ancestry continuum on PGS accuracy. We use simulations and real-data analyses to show that PGS accuracy decays continuously individual-to-individual across the genetic continuum as a function of GD from the PGS training data; GD is defined as a PCA projection of the target individual on the training data used to estimate the PGS weights. We leverage a large and diverse Los Angeles biobank at the University of California, Los Angeles⁵ (ATLAS, n = 36,778) along with the UK Biobank⁶ (UKBB, n = 487,409) to investigate the interplay between genetic ancestries and PGS for 84 complex traits and diseases. The accuracy of PGS models trained on individuals labelled as white British (WB; see Methods for naming convention used in this work) in the UKBB (n = 371,018) is negatively correlated with GD for all considered traits (average Pearson R = −0.95 across 84 traits), demonstrating pervasive individual variation in PGS accuracy. The negative correlation remains significant even when restricted to traditionally defined GIA clusters (ranging from R = −0.43 for East Asian GIA to R = −0.85 for the African American GIA in ATLAS). On average across the 84 traits, when rank-ordering individuals according to distance from training data, PGS accuracy decreases by 14% in the furthest versus closest decile in the European GIA. Notably, the furthest decile of individuals of European ancestries showed similar accuracy to the closest decile of Hispanic Latino individuals. Characterizing PGS accuracy across the continuum allows the inclusion of individuals unassigned to any GIA (6% of all ATLAS), thus allowing more individuals to be included in PGS applications. Finally, we explore the relationship between GD and PGS estimates themselves. Of 84 PGSs, 82 show significant correlation between GD and PGS with 30 showing opposite correlation (GD, trait) versus (GD, PGS); we exemplify the importance of incorporating GD in interpretation of PGSs using height and neutrophils in the ATLAS data. Our results demonstrate the need to incorporate the genetic ancestry continuum in assessing PGS performance and/or bias.

Overview of the study

PGS accuracy has conventionally been assessed at the level of discrete GIA clusters using population-level metrics of accuracy. Individuals from diverse genetic backgrounds are routinely grouped into discrete GIA clusters using computational inference methods such as PCA⁵⁰ and/or admixture analysis⁵¹ (Fig. 1a). Population-level metrics of PGS accuracy are then estimated for each GIA cluster and generalized to everyone in the cluster (Fig. 1b). This approach has three major limitations: the inter-individual variability within each cluster is ignored; the GIA cluster boundary is sensitive to algorithms and reference panels used for clustering; and a substantial proportion of individuals may not be assigned to any GIA owing to a lack of reference panels for genetic ancestry inference (for example, individuals of uncommon or admixed ancestries)….