Evaluation of polygenic prediction methodology within a reference-standardized framework

• DOI: 10.1371/journal.pgen.1009021
• Publication Date: 2021-05-05T16:12:36Z
• Authors: Oliver Pain; Kylie P. Glanville; Saskia P. Hagenaars; Saskia Selzam; Anna E. Fürtjes; Héléna A. Gaspar; Jonathan R. I. Co...; Kaili Rimfeld; Gerome Breen; Robert Plomin; Lasse Folkersen; Cathryn M. Lewis
• Keywords: 0301 basic medicine; Multifactorial Inheritance; Genotype; Models, Genetic; Twins; Datasets as Topic; Reproducibility of Results; QH426-470; Polymorphism, Single Nucleotide; United Kingdom; 3. Good health; 03 medical and health sciences; Genetics; Humans; Twin Studies as Topic; Computer Simulation; Precision Medicine; Research Article; Genome-Wide Association Study

The predictive utility of polygenic scores is increasing, and many scoring methods are available, but it unclear which method performs best. This study evaluates the within a reference-standardized framework, uses common set variants reference-based estimates linkage disequilibrium allele frequencies to construct scores. Eight score were tested: p-value thresholding clumping (pT+clump), SBLUP, lassosum, LDpred1, LDpred2, PRScs, DBSLMM SBayesR, evaluating their performance predict outcomes in UK Biobank Twins Early Development Study (TEDS). Strategies identify optimal thresholds shrinkage parameters compared, including 10-fold cross validation, pseudovalidation infinitesimal models (with no validation sample), multi-polygenic elastic net models. lassosum PRScs performed strongly using cross-validation most threshold or parameter, giving relative improvement 16-18% over pT+clump correlation between observed predicted outcome values. Using pseudovalidation, best SBayesR. was only 3% worse than identified by validation. Elastic containing based on range consistently improved prediction any single score. Within achieved modeling multiple derived parameters. will help researchers performing studies select powerful analysis methods.