A Principal Component Informed Approach to Address Polygenic Risk Score Transferability Across European Cohorts

Abstract

One important confounder in genome-wide association studies (GWASs) is population genetic structure, which may generate spurious associations if not properly accounted for. This may ultimately result in a biased polygenic risk score (PRS) prediction, especially when applied to another population. To explore this matter, we focused on principal component analysis (PCA) and asked whether a population genetics informed strategy focused on PCs derived from an external reference population helps in mitigating this PRS transferability issue. Throughout the study, we used two complex model traits, height and body mass index, and samples from UK and Estonian Biobanks. We aimed to investigate 1) whether using a reference population (1000G) for computation of the PCs adjusted for in the discovery cohort improves the resulting PRS performance in a target set from another population and 2) whether adjusting the validation model for PCs is required at all. Our results showed that any other set of PCs performed worse than the one computed on samples from the same population as the discovery dataset. Furthermore, we show that PC correction in GWAS cannot prevent residual population structure information in the PRS, also for non-structured traits. Therefore, we confirm the utility of PC correction in the validation model when the investigated trait shows an actual correlation with population genetic structure, to account for the residual confounding effect when evaluating the predictive value of PRS.

Ancestral genetic components are consistently associated with the complex trait landscape in European biobanks

“…In other words: although we took measures to control for confounders by including multiple geographic and socio-economic variables into our models, only by enriching for biologically-relevant regions, and by controlling for genome-wide covA stratification, we could expose robust and consistent signals. This finding mirrors the known difficulty in discerning functional gene-trait associations from spurious correlations mediated by genetic structure 19–21,25,27,28 , which are indeed independent between different populations and different biobanks with different recruitment strategies. In addition, while combining GWAS-derived effect sizes at genome scale is sensitive to subtle correction-surviving biases 26 , our approach does not rely on such methodology.…”

“…A plethora of methods [11][12][13][14] have been developed to correct for these unwanted sources of variance that might bias GWAS discovery. Indeed even the finer cases of population structure present in national Biobanks [15][16][17] have been demonstrated to affect GWAS [18][19][20][21][22][23] and, if not carefully addressed, hamper analyses following up on these results, such as Polygenic Risk Scoring 19,21,22,[24][25][26] and polygenic selection testing 27,28 .…”

Ensembled best subset selection using summary statistics for polygenic risk prediction

“…The current model is designed for single ancestry analysis. While a large portion of GWAS data currently comes from individuals of European ancestry, it is well-known that polygenic risk scores do not transfer well between individuals of different ancestries, which can impact their utility for patients of non-European ancestry [64][65][66][67][68][69] . Many methods that utilize summary data from multiple populations have already been proposed and demonstrate improved prediction in under-represented populations 25,29,30,[70][71][72][73][74] .…”