PUMAS: fine-tuning polygenic risk scores with GWAS summary statistics

Abstract

Polygenic risk scores (PRSs) have wide applications in human genetics research, but often include tuning parameters which are difficult to optimize in practice due to limited access to individual-level data. Here, we introduce PUMAS, a novel method to fine-tune PRS models using summary statistics from genome-wide association studies (GWASs). Through extensive simulations, external validations, and analysis of 65 traits, we demonstrate that PUMAS can perform various model-tuning procedures using GWAS summary statistics and effectively benchmark and optimize PRS models under diverse genetic architecture. Furthermore, we show that fine-tuned PRSs will significantly improve statistical power in downstream association analysis.

Quantifying portable genetic effects and improving cross-ancestry genetic prediction with GWAS summary statistics

“…Finally, we introduce an innovative strategy to linearly combine multiple PRS trained in different populations using summary association data alone. We employ a summary statistics-based repeated learning approach motivated from our recent work 45 to estimate the regression weights for combining multiple PRS. The entire X-Wing procedure only requires GWAS summary data and LD references as input, which is a major advance compared to existing approaches.…”

“…Typically, repeated learning (or cross-validation) requires individual-level genotype and phenotype data since it involves sample splitting. Generalizing the technique in our recent work 45 , we introduce a summary statistics-based repeated learning strategy, which does not need individual-level GWAS data. This approach has three main steps which we describe below.…”

“…We implemented 4-fold repeated learning to estimate the PRS combination weights using GWAS summary statistics and our equally divided 1000G reference panel 45,74 . In each fold, we first subsampled East Asian (or admixed American) summary statistics for 75% BBJ (or PAGE study) samples as the training and the remaining 25% as the validation set.…”

A new method for multi-ancestry polygenic prediction improves performance across diverse populations

“…The method requires three independent datasets: (1) GWAS summary statistics from training datasets across EUR and non-EUR populations; (2) a tuning dataset for the target population to find optimal model parameters; and (3) a validation dataset for the target population to report the final prediction performance. While this report assumes that individual-level data are available for model tuning and validation, summary-statistics-based methods 38,39 could also be used in these steps.…”

Computational Genomics in the Era of Precision Medicine: Applications to Variant Analysis and Gene Therapy

“…Methodological challenges in computing PRS reside in estimating the highly polygenic yet typically weak SNP effects for most complex traits and accounting for extensive LD in the human genome. Recently, penalized regression models re-estimate SNP effects from GWAS summary statistics while explicitly modeling LD have been shown to effectively improve the predictive performance of PRS [101][102][103], and novel resampling approaches now allow model fine-tuning without individual-level genotype and phenotype data [104]. Additionally, Khera et al convincingly demonstrated that individuals with very high PRS show substantially elevated coronary artery disease risk that is comparable to having monogenic mutations with large effects [105].…”