Assessment of genomic prediction accuracy of discrete traits with imputation of missing genotypes

Genomic selection as a promising tool for discovering genetic variants influencing complex traits, along with genotype imputation has an important role in increasing economic efficiency as well as genetic gain by accelerating animal breeding programs and potentially improving the accuracy of breeding values. The objectives of present research were: (i) to quantify the accuracy of genotype imputation and to evaluate factors affecting it, and (ii) to assess the effects of genotype imputation and genomic architecture on the performance of the Random Forest (RF), GBLUP and threshold Bayes A (TBA) methods for genomic predictions of binary traits. According to disease incidence and genomic architecture (heritability (h(2)) = 0.25 or 0.05, QTL=81 or 810 and linkage disequilibrium (LD) = low or high), reference and validation sets were organised in different simulated scenarios for the 54K SNP panel. To evaluate imputation accuracy, we randomly masked (90 and 50 percent of markers) and subsequently imputed certain genotypes using the FImpute programme. The disease incidence slightly affected prediction accuracies. A negative effect of increased missing genotypes on accuracies of genomic prediction was observed when applying TBA and GBLUP rather than RF. The TBA method performed better than the RE and GBLUP methods for genomic prediction. Nonetheless, for a scenario affected by a high number of QTLs and a high level of heritability, RF was more precise with an extension of computational time. The results suggested that genotype imputation from sparse panels (5.4 K SNPs) with high LD to 50K panels could be a cost-effective approach for genomic selection.