Milk infrared spectra are routinely used for phenotyping traits of interest through links developed between the traits and spectra. Predicted individual traits are then used in genetic analyses for estimated breeding value (EBV) or for phenotypic predictions using a single-trait mixed model; this approach is referred to as indirect prediction (IP). An alternative approach [direct prediction (DP)] is a direct genetic analysis of (a reduced dimension of) the spectra using a multitrait model to predict multivariate EBV of the spectral components and, ultimately, also to predict the univariate EBV or phenotype for the traits of interest. We simulated 3 traits under different genetic (low: 0.10 to high: 0.90) and residual (zero to high: ±0.90) correlation scenarios between the 3 traits and assumed the first trait is a linear combination of the other 2 traits. The aim was to compare the IP and DP approaches for predictions of EBV and phenotypes under the different correlation scenarios. We also evaluated relationships between performances of the 2 approaches and the accuracy of calibration equations. Moreover, the effect of using different regression coefficients estimated from simulated phenotypes (β_p), true breeding values (β_g), and residuals (β_r) on performance of the 2 approaches were evaluated. The simulated data contained 2,100 parents (100 sires and 2,000 cows) and 8,000 offspring (4 offspring per cow). Of the 8,000 observations, 2,000 were randomly selected and used to develop links between the first and the other 2 traits using partial least square (PLS) regression analysis. The different PLS regression coefficients, such as β_p, β_g, and β_r, were used in subsequent predictions following the IP and DP approaches. We used BLUP analyses for the remaining 6,000 observations using the true (co)variance components that had been used for the simulation. Accuracy of prediction (of EBV and phenotype) was calculated as a correlation between predicted and true values from the simulations. The results showed that accuracies of EBV prediction were higher in the DP than in the IP approach. The reverse was true for accuracy of phenotypic prediction when using β_p but not when using β_g and β_r, where accuracy of phenotypic prediction in the DP was slightly higher than in the IP approach. Within the DP approach, accuracies of EBV when using β_g were higher than when using β_p only at the low genetic correlation scenario. However, we found no differences in EBV prediction accuracy between the β_p and β_g in the IP approach. Accuracy of the calibration models increased with an increase in genetic and residual correlations between the traits. Performance of both approaches increased with an increase in accuracy of the calibration models. In conclusion, the DP approach is a good strategy for EBV prediction but not for phenotypic prediction, where the classical PLS regression-based equations or the IP approach provided better results.

牛奶红外光谱通常通过性状和光谱之间建立的联系用于感兴趣的性状的表型分析。然后将预测的单个性状用于遗传分析，以估算育种价值（EBV）或使用单性状混合模型进行表型预测。这种方法称为间接预测（IP）。一种替代方法[直接预测（DP）]是使用多特征模型对光谱（的缩减维数）进行直接遗传分析，以预测光谱成分的多变量EBV，并最终还预测该变量的单变量EBV或表型。感兴趣的特征。我们模拟了在不同遗传（低：0.10至高：0.90）和残留（零至高：±0）下的3个性状。90）3个特征之间的相关情景，并假设第一个特征是其他2个特征的线性组合。目的是比较IP和DP方法在不同相关情景下对EBV和表型的预测。我们还评估了这两种方法的性能与校准方程式之间的关系。此外，使用根据模拟表型（β_p），真育种值（β_克）和残差（β _{- [R}上2种方法中的性能）进行了评价。模拟数据包含2,100个父母（100头公母和2,000头母牛）和8,000个后代（每头母牛4个后代）。在8,000个观察值中，随机选择了2,000个，并使用偏最小二乘（PLS）回归分析将其用于建立第一个和其他两个特征之间的联系。不同的PLS回归系数，如β _p，β_克，和β _{- [R}，用于IP和DP方法的后续预测中。我们使用了用于模拟的真实（协）方差分量对剩余的6,000个观测值进行了BLUP分析。计算预测（EBV和表型）的准确性，作为预测值与真实值之间的相关性。结果表明，DP中的EBV预测准确性高于IP方法。使用β当反向是用于表型预测的准确性真_p但使用β时不_克和β _{- [R} ，其中在DP表型预测的准确率略高于在IP的方法。内的DP方法中，当使用EBV的精度β_克用β时高于_p仅在低遗传相关的情况下。然而，我们发现了β之间EBV预测准确率无显着差异_p和β _{g ^}在IP的方法。校准模型的准确性随着性状之间遗传和残留相关性的增加而增加。两种方法的性能都随着校准模型精度的提高而提高。总之，DP方法对于EBV预测是一种很好的策略，但对于表型预测却不是，因为经典的基于PLS回归的方程式或IP方法可以提供更好的结果。