奶牛产奶性状贝叶斯方法与机器学习全基因组选择研究

Study on genome-wide selection of dairy cow milk production traits by Bayesian method and machine learning

为了比较不同全基因组选择方法估计奶牛产奶性状育种值的性能,试验选用了3种贝叶斯方法(BB、BL和BRR方法)和3种机器学习方法(GB、RF和RKHS方法)分别采取3倍交叉验证(3FCV)、5FCV、10FCV和20FCV共4种分组方案对乳脂率、产奶量和体细胞评分3种性状进行10次独立运行育种值估计研究,同时比较不同方法所需运行时间。结果表明:20FCV验证方案所得育种值估计准确度最高;就每种性状而言,乳脂率的最高准确度是BB方法产生的0.881±0.005;产奶量的最高准确度是BL方法产生的0.804±0.008;体细胞评分的最高准确度是RKHS方法产生的0.773±0.007;育种值无偏性的优劣顺序与准确度的高低顺序一致。3种贝叶斯方法所需的运行时间较长,GB和RKHS方法所需的运行时间明显少于其他方法。说明BB、BL和RKHS方法分别在乳脂率、产奶量和体细胞评分的全基因组育种值估计准确度方面有较明显的优势。

To compare the performance in estimating the breeding value of milk production traits in dairy cows by different genome-wide selection methods, three Bayesian methods(BB, BL and BRR me thods) and three machine learning methods(GB, RF and RKHS methods) were selected to conduct 10 independent runs of breeding value estimation studies on three traits, namely milk fat rate, milk yield and somatic cell score which were determined by 3-fold cross validation(3FCV), 5FCV, 10FCV and 20FCV, respectively. The breeding values of these three traits were estimated for 10 independent runs, and the running time required by different methods were compared. The results showed that the estimated accuracies of breeding value obtained by 20FCV were the highest. For each trait, the highest accuracy of milk fat rate was 0.881±0.005 produced by BB method. The highest accuracy of milk yield was 0.804±0.008 produced by BL method. The highest accuracy of somatic cell score was 0.773±0.007 generated by RKHS. The order of unbiased breeding value unbiasedness was consistent with the order of accuracy. The running times required by the three Bayesian methods were longer, and the running times required by the GB and RKHS methods were significantly shorter than those required by the other methods. The results indicated that BB method in milk fat rate, BL method in milk yield and RKHS method in somatic cell score had obvious advantages for genome-wide breeding value evaluation accuracy.