Prediction accuracies (adjusted R2) and optimal weights for the 5 main methods (EUR, LAT, LAT+ANC, EUR+LAT, EUR+LAT+ANC) are reported in Table 2A. In each case, the best prediction accuracy was attained using LD-pruning threshold RLD2=0.8 (results using different LD-pruning thresholds are reported in S1 Table); the median value of the optimal P-value threshold PT was equal to 0.01 for EUR and 0.05 for LAT. On average, the EUR method performed only 23% better than the LAT method, despite having twice as much training data. This reflects a tradeoff between the larger training sample size for EUR and the target-matched LD patterns for LAT. EUR+LAT attained 64%-101% relative improvements vs. EUR and LAT respectively (and used a slightly larger weight for EUR than for LAT), highlighting the advantages of incorporating multiple sources of training data. When including an ancestry predictor, EUR+LAT+ANC attained a 10% relative improvement vs. EUR+LAT (≥80% relative improvement vs. EUR or LAT), reflecting small genetic effects of ancestry on phenotype that can arise from random genetic drift between populations at causal markers (which is better-captured by ancestry components than by SNPs used in a PRS).
