The use of polygenic risk scores11,12 (PRS, also known as risk profile scoring, genetic scoring, and genetic risk scoring) has become widespread in biomedical and social science disciplines13–15. Businesses have commercialized this technology, including direct-to-consumer testing from 23andMe and other companies. Perhaps most importantly, there is hope that polygenic risk scores can improve health outcomes by accelerating diagnosis and matching patients to tailored treatments16. Polygenic scoring studies have demonstrated reliable, though modest, prediction using straightforward scoring methods11,12 for many complex genetic phenotypes (e.g., blood pressure13,17, height18, diabetes9,19, depression7,20, and schizophrenia14). Polygenic risk scores are calculated by summing risk alleles, which are weighted by effect sizes derived from GWAS results11,12,21. Commonly used methods account for ancestry using principal components (calculated on pruned genetic data). In the parlance of polygenic scoring studies, the training GWAS is referred to as the discovery sample, and the testing dataset is referred to as the target sample. No overlap between training and testing datasets is essential to maintain independence of predictions, and the removal of related individuals is also needed, as demonstrated by Wray et al.21. Methods of prediction that offer modest improvements on this basic framework are also available22–25.
