A promising method for optimizing PRS was recently reported by Khera et al. (2018) 28. In this study, three large independent patient samples were used for each of five common medical conditions (coronary artery disease, atrial fibrillation, type 2 diabetes, inflammatory bowel disease, and breast cancer). The first sample was the discovery sample. A GWAS was performed, and a series of PRS was produced by using multiple P value thresholds, “pruning” for non-independence, and optimal number of variants for the final dataset. The second sample was used as a validation dataset, in which each of the PRS was applied to find the one that best discriminated cases from controls for that disease. The third sample was the testing dataset, in which only the best validated PRS was used. Using this method, the authors were able to generate good predictive ability (with 80% accuracy for coronary artery disease, for instance). Notably, they were also able to identify subsets of the testing sample with odds ratios of 3, 4, or 5 for having illness. This is comparable to the predictive ability of
