Allele frequency differences of causal variants have a major impact on low cross-ancestry portability of PRS

Abstract

Genome-wide association studies (GWAS) are overwhelmingly biased toward European ancestries. Nearly all existing studies agree that transferring genetic predictions from European ancestries to other populations results in a substantial loss of accuracy. This is commonly referred to as low portability of polygenic risk scores (PRS) and is one of the most important barriers to the ethical clinical deployment of PRS. Yet, it remains unclear how much various genetic factors, such as linkage disequilibrium (LD) differences, allele frequency differences or causal effect differences, contribute to low PRS portability. In this study, we used gene expression levels in lymphoblastoid cell lines (LCLs) as a simplified model of complex traits with minimal environmental variation, in order to understand how much each genetic factor contributes to PRS portability from European to African populations. We found thatcis-genetic effects on gene expression are highly similar between European and African individuals (). This stands in stark contrast to the very low estimates ofcis-genetic correlation between Europeans and Africans in previous studies, which we demonstrate are artifacts of statistical bias. We showed that portability decreases with increasing LD differences in thecis-regions. We also found that allele frequency differences of causal variants have a striking impact on PRS portability. For example, PRS portability is reduced by more than 32% when the causalcis-variant is common (minor allele frequency, MAF > 5%) in European samples (training population) but is rarer (MAF < 5%) in African samples (prediction population). While large allele frequency differences can decrease PRS portability through increasing LD differences, we also show that causal allele frequency can significantly impact portability independently of LD. This observation suggests that improving statistical fine-mapping alone does not overcome the loss of portability caused by causal allele frequency differences. Lastly, we also found that causal allele frequency is the main genetic factor underlying differential gene expression levels across ancestries. We conclude that causal genetic effects are highly similar in Europeans and Africans, and low PRS portability is primarily due to allele frequency differences.

Sources of gene expression variation in a globally diverse human cohort

“…This suggests that effect size heterogeneity of eQTLs between populations is rare, and apparent heterogeneity may instead reflect the failure to control for the additive effects of multiple independent causal signals. This conclusion is consistent with previous studies using orthogonal approaches for evaluating effect size heterogeneity based on analysis of admixed individuals 45,53 . For the small number of lead eQTLs that do have a significant interaction effect after controlling for multiple causal signals (9 eQTLs; 0.07% of all eQTLs that passed filtering), this effect could be driven by non-additive epistatic interactions between variants, additional untested causal variants that did not meet nominal MAF thresholds, or population-specific LD patterns with a untyped causal variants.…”

“…The scale and diversity of the data set enabled the discovery of numerous potentially novel genetic associations, while also offering high resolution for identifying putative causal variants and elucidating their mechanisms of action. Our study also demonstrates that conditional on the correct identification and presence of causal variants, the effects of such variants tend to be additive and highly consistent across populations—a point of recent debate within the field 9,44,45,53,54 . This observation in turn suggests that ancestry-dependent epistatic effects tend to be weak and/or rare in human genomes, in contrast to some observations from other model systems 57 .…”

“…Given the debate regarding the prevalence of GWAS/eQTL hits exhibiting effect-size heterogeneity between ancestries/populations 9,44,45,53,54 , we sought to test for he-eQTLs in MAGE. Because the genotypes are derived from high-coverage whole genome sequencing in 1KGP, MAGE should be robust to effect size heterogeneity resulting from population-specific LD patterns with an untyped casual variant (as commonly affects microarray data), barring large structural variation that may escape detection with short-read sequencing.…”

Improved multi-ancestry fine-mapping identifiescis-regulatory variants underlying molecular traits and disease risk

“…Altogether, these results further affirm previous results 20,21,23,[68][69][70][71][72][73][74] demonstrating primarily shared genetic architectures for molecular traits across ancestries.…”

Highly parameterized polygenic scores tend to overfit to population stratification via random effects

“…The poor cross-ancestry portability of PGSs tailored for specific ancestries has been attributed to inter-population differences in allele frequencies and linkage disequilibrium (LD) patterns (11; 12; 13; 14), closely related population-specific factors including genetic drift (15), and differences in environmental and social factors including social determinants of health (5). In contrast, variable cross-group portability of PGSs within the same ancestry has been linked to environmental stratification driven by demographic history (16) and confounding by non-genetic factors or assortative mating (17).…”