Impact of SNP Selection on Genomic Prediction for Different Reference Population Sizes

Methods for SNP selection can improve prediction accuracy over genomic BLUP, but in practice, the improvement is trait and population specific. This study investigates the importance of SNP selection in populations with 2000 to 25,000 genotyped animals. Populations were simulated with effective population sizes (Ne) of 20 or 100, and assuming that 10, 50, or 500 QTL were affecting a trait with heritability of 0.3. Pedigree information was available for 6 generations; phenotypes were recorded for the 4 middle generations. Animals from the last 3 generations were genotyped for 45,000 SNP. Single-step genomic BLUP (ssGBLUP) and weighted ssGBLUP (WssGBLUP) were used to estimate genomic EBV (GEBV). For WssGBLUP, 2 iterations of weights were calculated and were used to derive SNP variances and to construct a weighted genomic relationship matrix (G). Improved prediction accuracies are expected in WssGBLUP because more weight is placed on important SNP. Prediction accuracies were calculated for 1000 genotyped animals in the last generation. Reference populations included 2000, 5000 and 25,000 genotyped animals. The latter genotyped set was used to assess the dimensionality of genomic information (number of effective SNP or effective chromosome segments - Me). This was calculated as the number of the largest eigenvalues explaining 98% of the variation in the genomic relationship matrix with and without the weights. For the data sets with Ne=20 and 10 QTL, the accuracy gain from WssGBLUP was 12, 9, and 4 points for 2000, 5000, and 25,000 genotyped animals, respectively. With Ne=100, this gain was 8, 10, and 7 points, respectively. For both Ne of 20 and 100, the gain assuming 50 QTL was halved, and no gain was observed assuming 500 QTL. The number of effective SNP was about 4-fold less in weighted G (~1512) than in unweighted G (~5790), explaining the greater gain in accuracy with fewer genotyped animals. The impact of SNP selection decreases with increasing size of the reference population and number of QTL. In large populations, the detection of chromosome segments is more difficult, requiring more genotyped animals.