On prs for complex polygenic trait prediction

Abstract

Polygenic risk score (PRS) is the state-of-art prediction method for complex traits using summary level data from discovery genomewide association studies (GWAS). The PRS, as its name suggests, is designed for polygenic traits by aggregating small genetic effects from a large number of causal SNPs and thus is viewed as a powerful method for predicting complex polygenic traits by the genetics community. However, one concern is that the prediction accuracy of PRS in practice remains low with little clinical utility, even for highly heritable traits. Another practical concern is whether genome-wide SNPs should be used in constructing PRS or not. To address the two concerns, we investigate PRS both empirically and theoretically. We show how the performance of PRS is influenced by the triplet (n, p, m), where n, p, m are the sample size, the number of SNPs studied, and the number of true causal SNPs, respectively. For a given heritability, we find that i) when PRS is constructed with all p SNPs (referred as GWAS-PRS), its prediction accuracy is controlled by the p/n ratio; while ii) when PRS is built with a set of top-ranked SNPs that pass a pre-specified threshold (referred as threshold-PRS), its accuracy varies depending on how sparse the true genetic signals are. Only when m is magnitude smaller than n, or genetic signals are sparse, can threshold-PRS perform well and outperform GWAS-PRS. Our results demystify the low performance of PRS in predicting highly polygenic traits, which will greatly increase researchers' awareness of the power and limitations of PRS, and clear up some confusion on the clinical application of PRS.

Can education be personalised using pupils’ genetic data?

“…The theoretical maximum bounds placed on the predictive ability of polygenic scores have been discussed in detail elsewhere (see [14][15][16] ). Briefly, polygenic scores are more predictive when genetic factors play a larger role in a phenotype (as measured by heritability) and in the case of binary phenotypes where prevalence in the outcome is higher [14][15][16] . It has been suggested that for complex genetic phenotypes, accurate prediction at the individual level may require a polygenic score that explains around 75% of the total genetic variance of the phenotype 15 .…”

Cross-trait prediction accuracy of high-dimensional ridge-type estimators in genome-wide association studies

“…This study is motivated by the increasing applications of high-dimensional GWAS marginal estimator, especially in out-of-sample polygenic risk prediction. A few studies (such as Chatterjee et al [2013], Daetwyler et al [2008], Dudbridge [2013], and Zhao and Zou [2019]) have explored the prediction accuracy of GWAS summary statistics in the special case Σ = I p . To the best of our knowledge, there is no study on the behavior of marginal estimator in dense high-dimensional settings with general Σ.…”

“…One of the state-of-art methods for genetic risk prediction of human complex traits is genome-wide polygenic risk score (PRS) [Purcell et al, 2009], which is a weighted sum of millions of SNPs where each SNP is weighted by their estimated effect size from discovery GWAS. As no need to access the personal DNA information of subjects in the training set, PRS is computationally efficient and has widespread applications with more than 3, 000 related publications in 2018 [Zhao and Zou, 2019]. Recent efforts have begun to explore the clinical utility of PRS on human diseases, such as heart disease and breast cancer [Khera et al, 2018, Mavaddat et al, 2019.…”