Lee J, Kachman S D, Spangler M L
J Anim Sci. 2017 Aug;95(8):3406-3414. doi: 10.2527/jas.2017.1604.
Genomic selection (GS) has become an integral part of genetic evaluation methodology and has been applied to all major livestock species, including beef and dairy cattle, pigs, and chickens. Significant contributions in increased accuracy of selection decisions have been clearly illustrated in dairy cattle after practical application of GS. In the majority of U.S. beef cattle breeds, similar efforts have also been made to increase the accuracy of genetic merit estimates through the inclusion of genomic information into routine genetic evaluations using a variety of methods. However, prediction accuracies can vary relative to panel density, the number of folds used for folds cross-validation, and the choice of dependent variables (e.g., EBV, deregressed EBV, adjusted phenotypes). The aim of this study was to evaluate the accuracy of genomic predictors for Red Angus beef cattle with different strategies used in training and evaluation. The reference population consisted of 9,776 Red Angus animals whose genotypes were imputed to 2 medium-density panels consisting of over 50,000 (50K) and approximately 80,000 (80K) SNP. Using the imputed panels, we determined the influence of marker density, exclusion (deregressed EPD adjusting for parental information [DEPD-PA]) or inclusion (deregressed EPD without adjusting for parental information [DEPD]) of parental information in the deregressed EPD used as the dependent variable, and the number of clusters used to partition training animals (3, 5, or 10). A BayesC model with π set to 0.99 was used to predict molecular breeding values (MBV) for 13 traits for which EPD existed. The prediction accuracies were measured as genetic correlations between MBV and weighted deregressed EPD. The average accuracies across all traits were 0.540 and 0.552 when using the 50K and 80K SNP panels, respectively, and 0.538, 0.541, and 0.561 when using 3, 5, and 10 folds, respectively, for cross-validation. Using DEPD-PA as the response variable resulted in higher accuracies of MBV than those obtained by DEPD for growth and carcass traits. When DEPD were used as the response variable, accuracies were greater for threshold traits and those that are sex limited, likely due to the fact that these traits suffer from a lack of information content and excluding animals in training with only parental information substantially decreases the training population size. It is recommended that the contribution of parental average to deregressed EPD should be removed in the construction of genomic prediction equations. The difference in terms of prediction accuracies between the 2 SNP panels or the number of folds compared herein was negligible.
基因组选择（GS）已成为遗传评估方法中不可或缺的一部分，并已应用于所有主要家畜品种，包括肉牛和奶牛、猪以及鸡。在奶牛实际应用GS后，已清楚地表明其在提高选择决策准确性方面做出了重大贡献。在美国的大多数肉牛品种中，也已做出类似努力，通过使用各种方法将基因组信息纳入常规遗传评估，以提高遗传价值估计的准确性。然而，预测准确性可能会因基因分型密度、用于交叉验证的折数以及因变量的选择（例如，估计育种值[EBV]、去回归估计育种值、调整后的表型）而有所不同。本研究的目的是评估在训练和评估中使用不同策略时，红安格斯肉牛基因组预测器的准确性。参考群体由9776头红安格斯动物组成，其基因型被推算到两个中等密度的基因分型面板，分别包含超过50000个（50K）和大约80000个（80K）单核苷酸多态性（SNP）。使用推算后的面板，我们确定了标记密度、在用作因变量的去回归估计育种值中排除（根据亲本信息调整的去回归估计育种值[DEPD-PA]）或纳入（不根据亲本信息调整的去回归估计育种值[DEPD]）亲本信息以及用于划分训练动物的聚类数（3、5或10）的影响。使用π设置为0.99的贝叶斯C模型来预测13个有估计育种值的性状的分子育种值（MBV）。预测准确性通过MBV与加权去回归估计育种值之间的遗传相关性来衡量。使用50K和80K SNP面板时，所有性状的平均准确性分别为0.540和0.552，交叉验证分别使用3、5和10折时，平均准确性分别为0.538、0.541和0.561。将DEPD-PA用作响应变量时，生长和胴体性状的MBV准确性高于使用DEPD时。当使用DEPD作为响应变量时，阈值性状和性别受限性状的准确性更高，这可能是因为这些性状缺乏信息含量，并且在训练中仅排除有亲本信息的动物会大幅减少训练群体规模。建议在构建基因组预测方程时去除亲本平均值对去回归估计育种值的贡献。本文比较的两个SNP面板或折数之间的预测准确性差异可忽略不计。
