Genotype Imputation Methods and Their Effects on Genomic Predictions in Cattle

Abstract

In this study, we reviewed six imputation methods (Impute 2, FImpute 2.2, Beagle 4.1, Beagle 3.3.2, MaCH, and Bimbam) and evaluated the accuracy of imputation from simulated 6K bovine SNPs to 50K SNPs with 1800 beef cattle from two purebred and four crossbred populations and the impact of imputed genotypes on performance of genomic predictions for residual feed intake (RFI) in beef cattle. Accuracy of imputation was reported in both concordance rate (CR) and allelic $$r^{2}$$ and assessed via fivefold cross-validations. Running times of different methods were compared. Impute 2, FImpute and Beagle 4.1 yielded the most accurate imputation results (with CR > 91%). FImpute was the fastest and had advantages over all other methods in imputing rare variants. Minor allele frequency (MAF) and genetic relatedness between individuals in reference and validation populations can affect accuracy of imputation. For all methods, imputation accuracy for genotypes carrying the minor allele increases as the MAF increases. Impute 2 outperformed all other methods on MAF > 5% and onwards. FImpute and Impute 2 that adopted the nearest neighbour scheme coped better with individuals of distant relativeness. Bimbam yielded the poorest CR (76%) due to admixed reference panels. Imputed genotypes and actual 50K/6K genotypes were employed to predict genomic breeding values (GEBVs) of RFI using a Bayesian method and GBLUP. Accuracies of GEBV were similar using actual 50K genotypes or imputed genotypes, except those from Bimbam, and the imputation errors had minimal impact on the genomic predictions.