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

Zijie Zhao,Xiaoyuan Zhong,Jie Song,Qiongshi Lu,Yanyao Yi,Timothy J Hohman,Yupei Lin,Jason Fletcher,Yuchang Wu

doi:10.1186/s13059-021-02479-9

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.

Highlights

Accurate prediction of complex traits with genetic data is a major goal in human genetics research and precision medicine [1]
We sample marginal association statistics for a subset of individuals based on the complete genome-wide association studies (GWASs) summary data (Eqs. (7) and (12), the “Methods” section)
We propose an approach to evaluate the predictive performance of polygenic risk scores (PRSs) using summary statistics in the validation set so that we can select the best model based on its superior performance (Eq (20), the “Methods” section)

Summary

Introduction

Accurate prediction of complex traits with genetic data is a major goal in human genetics research and precision medicine [1]. Advancements in genotyping and imputation techniques have greatly accelerated discoveries in genome-wide association studies (GWASs) for numerous complex diseases and traits [2]. These data have enabled statistical learning applications that leverage genome-wide data in genetic risk prediction [3,4,5,6,7,8]. Despite these advances, it remains challenging to access, store, and process individual-level genetic data at a large scale due to privacy concerns and high computational burden. Most PRS models have tuning parameters, including the p-value threshold in traditional PRS, the penalty strength in penalized

Methods

Results

Conclusion