evidence of disease association on their own—under the assumption that many genuine significant associations are potentially missed because of inadequate power in the original GWAS. This approach yields greater predictive power across many psychiatric disorders, maximizing the predictive power and accuracy for discrimination between cases and controls (as measured by predicted areas under the curve, or AUCs) of around 82% for schizophrenia, 68% for bipolar disorder, 58% for major depressive disorder, and 54% for anxiety 7 (compared with an AUC of 50% describing a test with prediction no better than chance), although discrimination estimates vary considerably across cohorts in “leave one out” analyses reflecting independent datasets 8. These discrimination estimates will likely improve with more powerful and diverse discovery GWAS 9, 10 and improved methods for quantifying polygenic risk. Ideally, in order to be useful, predictive capacity would need to be at least moderately (that is, over 70% to 80%) but preferably highly (that is, over 90%) predictive. Recently, PRS reflecting a selection of SNP sets based on biological processes (for example, signaling pathway membership, protein–protein interactions, or pharmacological treatment response) have been calculated and applied 11, 12. These pathway-specific PRS may predict phenotypic variation more robustly than risk scores
