A guide to performing Polygenic Risk Score analyses

Key results

1. ClaimedPRS analyses need QC beyond standard GWAS filters because errors add up across SNPs. Check file integrity and genome build, confirm the effect allele, drop strand-ambiguous and duplicate SNPs, check sex, and remove base/target sample overlap and close relatives.

   These are recommendations, not empirical tests: remove all A/T and C/G SNPs rather than aligning them by allele frequency; exclude 1st/2nd degree relatives across base and target; treat a base GWAS with hsnp^2 < 0.05 (from LD Score regression or LDAK) with caution. · 2. Quality Control of Base and Target data; Ambiguous SNPs; Relatedness; Heritability check

2. ClaimedPicking the most predictive P-value threshold (or any tuning parameter) trains the model on the target sample, so a true out-of-sample check needs a third dataset. Using one arbitrary threshold with no optimisation risks underfitting.

   No data are given. The suggested fallbacks are repeated partitioning of the target (e.g. 10-fold cross-validation) or empirical P-values from permutation. The authors argue the latter keeps type 1 error appropriate when the aim is hypothesis testing rather than prediction. · 4.4 Overfitting in PRS association testing

3. SuggestedIn simulation, the Lee et al liability-scale pseudo-R^2 corrects for case/control ratios that differ from disease prevalence, whereas Nagelkerke R^2 is biased by that ascertainment.

   2,000,000 individuals were simulated with a normal phenotype driven partly by a normal predictor; case status followed a liability threshold model; 5,000 cases and 5,000 controls were sampled. Both pseudo-R^2 values were compared with an empirical R^2 from linear regression. The text gives no error sizes. · 4.1 Association and goodness-of-fit metrics; Figure legends

4. SuggestedA real-data reference shows how PRS variance explained and association P-values change with validation sample size for UK Biobank traits of high, medium and low heritability.

   Traits: height (h^2 = 0.49), Forced Volume Capacity (h^2 = 0.23) and hand grip (h^2 = 0.11). Two random sets of 100,000 individuals served as base and target; validation sets ran from 50 to 3000 individuals, repeated 5 times. Results are shown only in Figure 4, not in the text. · 4.5 Power and accuracy of PRS: target sample…; Figure legends

5. ClaimedBecause PRS aggregate the estimated effects of many null SNPs, population stratification is potentially a greater problem for them than for GWAS. Correlated allele-frequency and environmental differences between base and target can create spurious associations even if the GWAS corrected for structure perfectly.

   This is a conceptual argument without data. It advises many principal components when base and target are genetically similar, extreme caution across populations, and checking PRS distributions, since mixed ancestry or correlated SNPs can yield non-normal scores. · 3.4 Population structure and global…; 4.3 PRS distribution