Evaluation of polygenic prediction methodology within a reference-standardized framework

The predictive utility of polygenic scores is increasing, and many polygenic scoring methods are available, but it is unclear which method performs best. This study evaluates the predictive utility of polygenic scoring methods within a reference-standardized framework, which uses a common set of variants and reference-based estimates of linkage disequilibrium and allele frequencies to construct scores. Eight polygenic score methods were tested: p-value thresholding and clumping (pT+clump), SBLUP, lassosum, LDpred1, LDpred2, PRScs, DBSLMM and SBayesR, evaluating their performance to predict outcomes in UK Biobank and the Twins Early Development Study (TEDS). Strategies to identify optimal p-value thresholds and shrinkage parameters were compared, including 10-fold cross validation, pseudovalidation and infinitesimal models (with no validation sample), and multi-polygenic score elastic net models. LDpred2, lassosum and PRScs performed strongly using 10-fold cross-validation to identify the most predictive p-value threshold or shrinkage parameter, giving a relative improvement of 16–18% over pT+clump in the correlation between observed and predicted outcome values. Using pseudovalidation, the best methods were PRScs, DBSLMM and SBayesR. PRScs pseudovalidation was only 3% worse than the best polygenic score identified by 10-fold cross validation. Elastic net models containing polygenic scores based on a range of parameters consistently improved prediction over any single polygenic score. Within a reference-standardized framework, the best polygenic prediction was achieved using LDpred2, lassosum and PRScs, modeling multiple polygenic scores derived using multiple parameters. This study will help researchers performing polygenic score studies to select the most powerful and predictive analysis methods.Author summary: An individual’s genetic predisposition to a given outcome can be summarized using polygenic scores. Polygenic scores are widely used in research and could also be used in a clinical setting to enhance personalized medicine. A range of methods have been developed for calculating polygenic scores, but it is unclear which methods are the best. Several methods provide multiple polygenic scores for each individual which must then be tested in an independent tuning sample to identify which polygenic score is most accurate. Other methods provide a single polygenic score and therefore do not require a tuning sample. Our study compares the prediction accuracy of eight leading polygenic scoring methods in a range of contexts. For methods that calculate multiple polygenic scores, we find that LDpred2, lassosum, and PRScs methods perform best on average. For methods that provide a single polygenic score, not requiring a tuning sample, we find PRScs performs best, and the faster DBSLMM and SBayesR methods also perform well. Our study has provided a comprehensive comparison of polygenic scoring methods that will guide future implementation of polygenic scores in both research and clinical settings.