Prediction accuracies (adjusted R2 on the liability scale (Lee et al., 2012), assuming 8% prevalence (Stahl et al., 2012)) and optimal weights for the 5 main methods (EUR, LAT, LAT+ANC, EUR+LAT, EUR+LAT+ANC) are reported in Table 3 (other prediction metrics are reported in S7 Table). In each case, the best prediction accuracy was obtained using LD-pruning threshold RLD2=0.8 (results using different LD-pruning thresholds are reported in S8 Table); the value of the optimal P-value threshold PT was equal to 0.05 for EUR and 0.2 for LAT. EUR performed only 33% better than LAT despite the much larger training sample size, again reflecting a tradeoff between sample size and target-matched LD patterns. EUR+LAT attained 75%-133% relative improvements vs. EUR and LAT respectively (and used a slightly larger weight for EUR than for LAT), again highlighting the advantages of incorporating multiple sources of training data. We also evaluated a meta-analysis PRS (EUR-LAT-meta) and determined that EUR+LAT attained a 19% relative improvement vs. EUR-LAT-meta (Table 3; also see S2 Fig), highlighting the advantages of optimizing mixing weights distinct from meta-analysis weights. Although adding
