It is worth noting important limitations of PRS analyses. First, polygenic prediction is influenced by the ancestry of the population studied. For example, PRS for AUD generated in an African American (AA) cohort explained more of the variance in AUD than PRS derived from a much larger cohort of European Americans (12). This illustrates that the prediction from one population to another does not perform well (e.g. PRS based on European Americans but used to predict in AA) (57). Second, the method of ascertainment may bias the results. As an example, PRS for DSM-IV AD derived from a population based sample predicted increased risk for AD in other population samples but did not associate with AUD symptoms in a clinically ascertained sample (54). Third, the variance explained by PRS is still low, and hence PRS have limited clinical application. For example, in the largest study of alcohol consumption (21), the alcohol consumption PRS accounted for only ~2.5% of the variance in alcohol use in two independent datasets. Recent work suggested that predictions may improve by incorporating functional genomic information. For
