Key message: Including additive and additive-by-additive epistasis in a NOIA parametrization did not yield orthogonal partitioning of genetic variances, nevertheless, it improved predictive ability in a leave-one-out cross-validation for wheat grain yield. Abstract: Additive-by-additive epistasis is the principal non-additive genetic effect in inbred wheat lines and is potentially useful for developing cultivars based on total genetic merit; nevertheless, its practical benefits have been highly debated. In this article, we aimed to (i) evaluate the performance of models including additive and additive-by-additive epistatic effects for variance components (VC) estimation of grain yield in a wheat-breeding population, and (ii) to investigate whether including additive-by-additive epistasis in genomic prediction enhance wheat grain yield predictive ability (PA). In total, 2060 sixth-generation (F
6) lines from Nordic Seed A/S breeding company were phenotyped in 21 year-location combinations in Denmark, and genotyped using a 15 K-Illumina-BeadChip. Three models were used to estimate VC and heritability at plot level: (i) “I-model” (baseline), (ii) “I + G
A-model”, extending I-model with an additive genomic effect, and (iii) “I + G
A + G
AA-model”, extending I + G
A-model with an additive-by-additive genomic effects. The I + G
A-model and I + G
A + G
AA-model were based on the Natural and Orthogonal Interactions Approach (NOIA) parametrization. The I + G
A + G
AA-model failed to achieve orthogonal partition of genetic variances, as revealed by a change in estimated additive variance of I + G
A-model when epistasis was included in the I + G
A + G
AA-model. The PA was studied using leave-one-line-out and leave-one-breeding-cycle-out cross-validations. The I + G
A + G
AA-model increased PA significantly (16.5%) compared to the I + G
A-model in leave-one-line-out cross-validation. However, the improvement due to including epistasis was not observed in leave-one-breeding-cycle-out cross-validation. We conclude that epistatic models can be useful to enhance predictions of total genetic merit. However, even though we used the NOIA parameterization, the variance partition into orthogonal genetic effects was not possible.
