Overestimated Polygenic Prediction due to Overlapping Subjects in Genetic Datasets

Recently, polygenic risk score (PRS) has gained significant attention in studies involving complex genetic diseases and traits. PRS is often derived from summary statistics, from which the independence between discovery and replication sets cannot be monitored. Prior studies, in which the independence is strictly observed, report a relatively low gain from PRS in predictive models of binary traits. We hypothesize that the independence assumption may be compromised when using the summary statistics, and suspect an overestimation bias in the predictive accuracy. To demonstrate the overestimation bias in the replication dataset, prediction performances of PRS models are compared when overlapping subjects are either present or removed. We consider the task of Alzheimers disease (AD) prediction across genetics datasets, including the International Genomics of Alzheimers Project (IGAP), AD Sequencing Project (ADSP), and Accelerating Medicine Partnership - Alzheimers Disease (AMP-AD). PRS is computed from either sequencing studies for ADSP and AMP-AD (denoted as rPRS) or the summary statistics for IGAP (sPRS). Two variables with the high heritability in UK Biobank, hypertension, and height, are used to derive an exemplary scale effect of PRS. Based on the scale effect, the expected performance of sPRS is computed for AD prediction. Using ADSP as a discovery set for rPRS on AMP-AD, {Delta}AUC and {Delta}R2 (performance gains in AUC and R2 by PRS) record 0.069 and 0.11, respectively. Both drop to 0.0017 and 0.0041 once overlapping subjects are removed from AMP-AD. sPRS is derived from IGAP, which records {Delta}AUC and {Delta}R2 of 0.051{+/-}0.013 and 0.063{+/-}0.015 for ADSP and 0.060 and 0.086 for AMP-AD, respectively. On UK Biobank, rPRS performances for hypertension assuming a similar size of discovery and replication sets are 0.0036{+/-}0.0027 ({Delta}AUC) and 0.0032{+/-}0.0028 ({Delta}R2). For height, {Delta}R2 is 0.029{+/-}0.0037. Considering the high heritability of hypertension and height of UK Biobank, we conclude that sPRS results from AD databases are inflated. The higher performances relative to the size of the discovery set were observed in PRS studies of several diseases. PRS performances for binary traits, such as AD and hypertension, turned out unexpectedly low. This may, along with the difference in linkage disequilibrium, explain the high variability of PRS performances in cross-nation or cross-ethnicity applications, i.e., when there are no overlapping subjects. Hence, for sPRS, potential duplications should be carefully considered within the same ethnic group.