The Three Types Of Quantitative-Genetic Information In Genomic Prediction

Genomic prediction methods such as Genomic-BLUP and Bayesian methods with t-distributed priors need to capture relationships between training individuals and selection candidates at quantitative trait loci to be useful for artificial selection and disease risk prediction. Deterministic formulas and simulations were used to prove that those methods exploit linkage disequilibrium (LD), co-segregation (CS), and additive-genetic relationships (RS) to capture those useful relationships. Contributions of these information types to prediction accuracy, persistence over generations without retraining, and effect on the correlation of predictions within families were evaluated by simulations depending on SNP density, training size, and extent of LD. Accuracy due to RS and CS declines with training size and extent of LD depending on relatedness between training individuals and selection candidates, whereas accuracy due to LD increases with training size. Modeling polygenic effects via pedigree relationships jointly with markers prevents the decay of accuracy due to RS, and modeling LD and CS explicitly may prevent the decay of accuracy due to CS. CS contributes little to accuracy in current dairy cattle breeding schemes, but considerably to accuracy within families in corn breeding. Persistence of accuracy due to CS across generations is lower than that of LD. The correlation between predictions within families, an important parameter in breeding schemes, depends largely on RS information, which is determined by the effective number of SNPs and training size. Genomic-BLUP is not suitable to capture LD information under short-range LD as in humans; thus, to improve prediction of disease risk, Bayesian methods with t-distributed priors are recommended.