Genetic determinants of polygenic prediction accuracy within a population

Abstract

Genomic risk prediction is on the emerging path towards personalized medicine. However, the accuracy of polygenic prediction varies strongly in different individuals. In this study, based on up to 352,277 White British participants in the UK Biobank, we constructed polygenic risk scores for 15 physiological and biochemical quantitative traits after performing genome-wide association studies (GWASs). We identified 185 polygenic prediction variability quantitative trait loci (pvQTLs) for 11 traits by Levene’s test among 254,376 unrelated individuals. We validated the effects of pvQTLs using an independent test set of 58,927 individuals. A score aggregating 51 pvQTL SNPs for triglycerides had the strongest Spearman correlation of 0.185 (p-value < 1.0x10−300) with the squared prediction errors. We found a strong enrichment of complex genetic effects conferred by pvQTLs compared to risk loci identified in GWASs, including 89 pvQTLs exhibiting dominance effects. Incorporation of dominance effects into polygenic risk scores significantly improved polygenic prediction for triglycerides, low-density lipoprotein cholesterol, vitamin D, and platelet. After including 87 dominance effects for triglycerides, the adjusted R2 for the polygenic risk score had an 8.1% increase on the test set. In addition, 108 pvQTLs had significant interaction effects with measured environmental or lifestyle exposures. In conclusion, we have discovered and validated genetic determinants of polygenic prediction variability for 11 quantitative biomarkers, and partially profiled the underlying complex genetic effects. These findings may assist interpretation of genomic risk prediction in various contexts, and encourage novel approaches for constructing polygenic risk scores with complex genetic effects.

Considering strategies for SNP selection in genetic and polygenic risk scores

“…We followed Forgetta et al (2022) when using data from the UK Biobank (ukbiobank.org (Bycroft et al, 2018)). We extracted 11 Ei phenotypes of interest as well as age, sex, and genetic principal components for 502,616 participants recruited into the UK Biobank dataset.…”

“…Then for each phenotype, we excluded those with missing covariates or phenotypes. Due to incomplete access to individual level hypothyroidism in the UK Biobank, our analysis incorporates only 11 of the 12 traits reported in (Forgetta et al, 2022).…”

“…These can be grouped into several classes: methods based expression Quantitative Trait Loci (eQTL) such as eCAVIAR (Hormozdiari et al, 2016) restrict attention to genes whose expression is influenced by associated SNPs; methods such as DEPICT (Pers et al, 2015) assume that functional annotation of genes will prioritize effectively; approaches such as MAGENTA (Segrè et al, 2010) leverage information in biological annotations; and methods that implement detailed fine-mapping (e.g., PAINTOR (Kichaev et al, 2014) use a combination of statistical arguments and functional annotations. In this latter category of methods, we recently published the Effector Index (Ei) for predicting which is the most likely causal gene at a locus showing multiple GWAS signals (Forgetta et al, 2022). For Ei, the predictions of the most likely causal gene are based on many types of information including features of the associated SNPs at a locus, and include the magnitudes of the β coefficients, linkage disequilibrium patterns, fine mapping results, and DNase hypersensitivity sites.…”

“…A compelling rationale for differential score performance can be made for males versus females ( Roberts et al, 2020 ; Manikpurage et al, 2021 ), since sex-specific disease incidence and regulation is the norm rather than the exception ( Ober et al, 2008 ). Furthermore, not only are there loci whose effects vary across subgroups, but loci exist that are associated with increased prediction variability in phenotypes ( Wang et al, 2019 ; Lu et al, 2022 ). Without a good understanding of the contexts where genetic or polygenic risk scores perform well, it is more challenging to argue for their clinical use.…”

“…Among the methods proposed, some attempt to increase the likelihood that the selected SNPs are truly causal. For example, fine mapping after a Genome-Wide Association Study (GWAS) predicts the most likely causal SNP(s) at a locus (e.g., (Kichaev et al, 2014;Chen et al, 2015;Benner et al, 2016;Wang et al, 2020;Zhang et al, 2021;Forgetta et al, 2022)). Even if the SNPs retained after such fine mapping algorithms are not truly causal, they are likely to be strongly correlated with the causal SNPs.…”