Negative selection on complex traits limits phenotype prediction accuracy between populations

Abstract

Accurate genetic risk prediction is a key goal for medical genetics and great progress has been made toward identifying individuals with extreme risk across several traits and diseases (Collins and Varmus, 2015). However, many of these studies are done in predominantly European populations (Bustamante et al., 2011; Popejoy and Fullerton, 2016). Although GWAS effect sizes correlate across ancestries (Wojcik et al., 2019), risk scores show substantial reductions in accuracy when applied to non-European populations (Kim et al., 2018; Martin et al., 2019; Scutari et al. , 2016). We use simulations to show that human demographic history and negative selection on complex traits result in population specific genetic architectures. For traits under moderate negative selection, ~50% of the heritability can be accounted for by variants in Europe that are absent from Africa. We show that this directly leads to poor performance in risk prediction when using variants discovered in Europe to predict risk in African populations, especially in the tails of the risk distribution. To evaluate the impact of this effect in genomic data, we built a Bayesian model to stratify heritability between European-specific and shared variants and applied it to 43 traits and diseases in the UK Biobank. Across these phenotypes, we find ~50% of the heritability comes from European-specific variants, setting an upper bound on the accuracy of genetic risk prediction in non-European populations using effect sizes discovered in European populations. We conclude that genetic association studies need to include more diverse populations to enable to utility of genetic risk prediction in all populations. The past decade of genome wide association studies (GWAS) has uncovered a plethora of trait associated loci scattered across the genome (The Wellcome Trust Case Control Consortium, 2007;The UK10K Consortium, 2015;Visscher et al., 2017;Bycroft et al., 2018). Geneticists have devoted many resources to turning these associations into risk prediction models that can be used to guide healthcare decisions for a variety of traits and diseases (Vilhjálmsson et al., 2015). Recent work has suggested these risk scores may be ready for clinical use (Khera et al., 2018(Khera et al., , 2019. While individuals with high genetic risk for diseases have been found using these scores (for example atherosclerosis (Natarajan et al., 2017) and breast cancer (Maas et al., 2016)), .

Quantitative Human Paleogenetics: What can Ancient DNA Tell us About Complex Trait Evolution?

“…For large sample sizes, low-frequency variants (MAF ≤ 0.05) make a significant contribution to the heritability of many complex traits ( Mancuso et al, 2016 ; Hartman et al, 2019 ), but the role of rare variants is less well established. Both empirical and simulation studies have shown that for traits under either negative or stabilising selection, there is an inverse correlation between effect size and MAF ( Simons et al, 2018 ; Schoech et al, 2019 ; Durvasula and Lohmueller, 2021 ). For the many traits thought to be under negative selection ( O’Connor et al, 2019 ), large effect variants that are rare in present-day populations may have had higher allele frequencies in ancient populations due to selection.…”

“…The consequence of this missing heritability may be particularly acute for trait prediction in ancient samples, as large-effect rare variants which contributed to variability in the past may no longer be segregating in present-day populations. Indeed, simulations suggest that the genetic architecture of complex traits is highly specific to each population, and that negative selection enriches for private variants, which contribute to a substantial component of the heritability of each trait ( Durvasula and Lohmueller, 2021 ). Empirical studies have also identified that functionally important regions, including conserved and regulatory regions, are enriched for population-specific effect sizes, and that this pattern may have been driven by directional selection ( Shi et al, 2021 ).…”

Embryo Screening for Polygenic Disease Risk: Recent Advances and Ethical Considerations

“…There are now many validations of polygenic prediction in the scientific literature, which were conducted using groups of people born on different continents and in different decades with respect to the original populations used in training [10,13,14]. However, it is important to note that predictors work best when applied to ancestry groups that are similar to the original training population, and performance falls off with genetic distance [11,15]. It has also been shown that predictors can differentiate between siblingsfor example, determining which one of them will experience a heart attack-despite similarity in childhood environments and genotype.…”

Enrichment analyses identify shared associations for 25 quantitative traits in over 600,000 individuals from seven diverse ancestries

“…For a given phenotype, SNP-level GWA summary statistics may differ across ancestries due to a variety of factors, including: (i ) ascertainment bias in genotyping 2,5 , (ii ) varying linkage disequilibrium (LD) patterns 12,13 , (iii ) variation in allele frequencies due to different selective pressures from unique population histories [13][14][15][16][17] , and (iv ) the effect of environmental factors on phenotypic variation [18][19][20][21] . These confounders and the observed low transferability of GWA results across ancestries 2,22 have generated an important call for increasing GWA efforts focused on diverse, non-European ancestry samples. However, we note that non-European ancestry GWA studies have --and will continue to have -smaller sample sizes than existing and emerging European-ancestry GWA cohorts, reducing the precision of summary statistic estimation in these studies.…”

“…Replication of SNP-level GWA results across ancestries is the exception, not the rule One question that has not been interrogated fully is the extent to which SNP-level associations for a given trait replicate across ancestries (however, also see Wojcik et al 6 , Durvasula and Lohmueller 22 , Carlson et al 38 , Liu et al 39 , Eyre-Walker 40 , Shi et al 41 , 42 ). To test this in our analyses, we first examined the number of genome-wide significant SNP-level associations that replicated exactly based on chromosomal position in multiple ancestries (see Supplementary Figure 3a and Supplementary Figure 3c, with Bonferronicorrected thresholds provided in Supplementary Table 10).…”