Quantifying genomic connectedness and prediction accuracy from additive and non-additive gene actions

Background: Genetic connectedness is classically used as an indication of the risk associated with breeding value comparisons across management units because genetic evaluations based on best linear unbiased prediction rely for their success on sufficient linkage among different units. In the whole-genome prediction era, the concept of genetic connectedness can be extended to measure a connectedness level between reference and validation sets. However, little is known regarding (1) the impact of non-additive gene action on genomic connectedness measures and (2) the relationship between the estimated level of connectedness and prediction accuracy in the presence of non-additive genetic variation. Results: We evaluated the extent to which non-additive kernel relationship matrices increase measures of connectedness and investigated its relationship with prediction accuracy in the cross-validation framework using best linear unbiased prediction and coefficients of determination. Simulated data assuming additive, dominance, and epistatic gene action scenarios and real swine data were analyzed. We found that the joint use of additive and non-additive genomic kernel relationship matrices or non-parametric relationship matrices led to increased capturing of connectedness, up to 25%, and improved prediction accuracies compared to those of baseline additive relationship counterparts in the presence of non-additive gene action. Conclusions: Our findings showed that connectedness metrics can be extended to incorporate non-additive genetic variation of complex traits. Use of kernel relationship matrices designed to capture non-additive gene action increased measures of connectedness and improved whole-genome prediction accuracy, further broadening the scope of genomic connectedness studies.