Determination of optimum weighting factors for single-step genetic evaluation via genetic variance partitioning

Abstract

<p>It is important in single-step genetic evaluations to use appropriate lambdas (λ) for calculating weighted average of NRM (numerator relationship matrix) and GRM (genomic relationship matrix) in joint relationship matrix. λ is usually estimated using a single-trait cross-validation procedure. However, it can be shown that a univariate single-step model applying a scalar λ is simply a condensed form of an extended model containing two genetic factors, factor <i>H~N</i>(0, <i>H</i>) and factor <i>A~N</i>(0, <i>A</i>), where the partitioning of the total genetic variance reflects λ. For multivariate single-step genetic evaluation, this model condensation implies that all involved genetic variances may yield the same λ, which is highly unlikely. Hence, it is required to estimate λ by accounting for its heterogeneity using the extended model for variance component estimation. This study used an extended single-step model to estimate variances and λs for calving difficulty (CD), gestation length (GL), and birth weight (BW) using Australian Angus data. A total of 129,851 animals with 45,575 genotypes were analysed. Initial variances obtained from a pedigree-only model were then used as starting values for the extended single-step model assigning 90% of the genetic variance to factor <i>A</i> and 10% to factor <i>H</i>. Since CD is a categorical trait with three categories, a threshold model-Gibbs sampling method was used to estimate variances. Heritability estimates for the extended single-step model were very similar to those from the pedigree only model implying that the single-step model was not explaining more variation in the data than the pedigree only model. For CD, GL, and BW, the total heritability estimates were 0.39 ± 0.04, 0.68 ± 0.02, and 0.44 ± 0.01, respectively. For the same traits, the total maternal heritability estimates were 0.17 ± 0.02, 0.11 ± 0.01, and 0.09 ± 0.01, respectively. In contrast, to the Gibbs sampling starting values, the genetic variance was partitioned between <i>A</i> and <i>H</i> such that direct genetic λ estimates for CD, GL, and BW were 0.36 ± 0.05, 0.62 ± 0.03, 0.75 ± 0.03, respectively. Maternal genetic λ estimates ranged from 0.01 ± 0.01 (for BW) to 0.05 ± 0.01 (for CD). The results imply that λ values are heterogeneous in multivariate single-step genomic evaluation. Further studies are needed to investigate the consequences of using heterogenous λ values for direct genetic and maternal genetic components in multivariate single-step evaluation in terms of model dimensions, solver convergence rate, and model forward predictive ability.</p>