Improved polygenic prediction by Bayesian multiple regression on summary statistics

Abstract

The capacity to accurately predict an individual's phenotype from their DNA sequence is one of the great promises of genomics and precision medicine. Recently, Bayesian methods for generating polygenic predictors have been successfully applied in human genomics but require the individual level data, which are often limited in their access due to privacy or logistical concerns, and are computationally very intensive. This has motivated methodological frameworks that utilise publicly available genome-wide association studies (GWAS) summary data, which now for some traits include results from greater than a million individuals. In this study, we extend the established summary statistics methodological framework to include a class of point-normal mixture prior Bayesian regression models, which have been shown to generate optimal genetic predictions and can perform heritability estimation, variant mapping and estimate the distribution of the genetic effects. In a wide range of simulations and cross-validation using 10 real quantitative traits and 1.1 million variants on 350,000 individuals from the UK Biobank (UKB), we establish that our summary based method, SBayesR, performs similarly to methods that use the individual level data and outperforms other state-of-the-art summary statistics methods in terms of prediction accuracy and heritability estimation at a fraction of the computational resources. We generate polygenic predictors for body mass index and height in two independent data sets and show that by exploiting summary statistics on 1.1 million variants from the largest GWAS meta-analysis (n ≈ 700, 000) that the SBayesR prediction R 2 improved on average across traits by 6.8% relative to that estimated from an individual-level data BayesR analysis of data from the UKB (n ≈ 450, 000). Compared with commonly used state-of-the-art summarybased methods, SBayesR improved the prediction R 2 by 4.1% relative to LDpred and by 28.7% relative to clumping and p-value thresholding. SBayesR gave comparable prediction accuracy to the recent RSS method, which has a similar model, but at a computational time that is two orders of magnitude smaller. The methodology is implemented in a very efficient and user-friendly software tool titled GCTB. Introduction 1The capacity to accurately predict an individual's phenotype from their DNA sequence 2 is one of the great promises of genomics and precision medicine 1-5 , recognising that the 3 accuracy of a genetic risk predictor is dependent on the genetic contribution to variation 4 in the trait. It is anticipated that genetic risk prediction will be useful for informing early 5 disease intervention and aiding diagnosis by identifying individuals with an increased 6 genetic risk of disease 5-7 . Accurate genetic predictors for complex traits and disorders are 7 currently limited, due mainly to an incomplete understanding of complex genetic varia-8 tion, small training sample sizes and suboptimal modelling 4,8,9 . Through large consortia 9 and biobank initiatives, sample sizes for gen...

Polygenic Prediction via Bayesian Regression and Continuous Shrinkage Priors

“…Individual-level Bayesian regression models (1) with a prior on SNP effect sizes can often be approximated using an external LD reference panel and turned into summary statistics based methods 4,6,21,22 . Here we enable posterior inference of SNP effect sizes from GWAS summary statistics under continuous shrinkage priors using an efficient Gibbs sampler with multivariate block update of the effect sizes (see Methods).…”

“…Although this approach has advantages in terms of computational and conceptual simplicity, and has been used to predict genetic liability across a broad phenotypic spectrum, recent studies have shown that this conventional method for PRS construction discards information and limits prediction accuracy 4 . More sophisticated Bayesian polygenic prediction methods that rely on GWAS summary statistics, including LDpred 4 and the normal-mixture model recently developed 5,6 , can incorporate genome-wide markers and accommodate varying genetic architectures, and thus have enhanced performance and flexibility. However, the type of prior on SNP effect sizes used in these methods, known as discrete mixture priors, imposes daunting computational challenges and may result in inaccurate adjustment for local LD patterns.…”

The emerging landscape of health research based on biobanks linked to electronic health records: Existing resources, statistical challenges, and potential opportunities

“…124,[132][133][134][135][136][137][138][139][140][141][142][143] For a recent exploration of PRS construction, we refer the reader to Choi et al 144 Recently, statistical methods have been developed to leverage published GWAS and other omics summary statistics to improve the performance of prediction algorithms and perform analyses adjusting for many genetic loci simultaneously. [145][146][147][148][149] Researchers may also be interested in studying relationships between phenotypes or joint relationships between phenotypes and other patient-level factors such as treatments or genotypes. Existing statistical methods for dealing with correlated outcomes such as mixed modeling and generalized estimating equations (when the model coefficients are of primary interest) can often be applied.…”

Bayesian analysis of GWAS summary data reveals differential signatures of natural selection across human complex traits and functional genomic categories

“…where b is a point mass at zero, and S (the relationship between MAF and effect sizes), F% (the effect variance factor common to all SNPs) and π (the proportion of SNPs with nonzero effects, i.e., the polygenicity) are considered as unknown, with prior distributions of a standard normal, a scaled inverse chi-squared distribution (Supplementary Note), and a uniform distribution between zero and one, respectively. Specifying a different prior distribution to Z H gives a form of other summary-data-based Bayesian alphabet models 47 .…”