基于多层感知机的绵羊限性性状基因组选择模拟研究

Simulation Study on Genomic Selection of Sex-limited Traits Using Multilayer Perceptron in Sheep

旨在将多层感知机(multilayer perceptron,MLP)应用于绵羊限性性状基因组选择中,并在多种情况下与其他经典基因组选择方法进行比较分析。本研究利用Qmsim软件模拟2个绵羊群体Pop1和Pop2的表型数据和基因型数据。在MLP中使用人工神经网络(artificial neural network,ANN),线性模型中使用约束性最大似然法(residual maximum likelihood,REML)估计不同群体的遗传参数。利用Python语言自编MLP模型,利用DMU软件实现最佳线性无偏预测(best linear unbiased prediction,BLUP)、基因组最佳线性无偏预测(genomic BLUP)和一步法(single-step GBLUP,SSGBLUP)模型,评估不同情况下各方法遗传力(heritability,h^(2))和育种值估计方面的差异。各情况下,MLP和SSGBLUP均显著(P<0.05)优于GBLUP和BLUP。在3种情况下MLP的h^(2)估值与SSGBLUP差异不显著:h^(2)为0.05,标记数为10K且QTL数为100时的Pop2群体;h^(2)为0.2,QTL数为500的两个标记数下Pop1群体和QTL数为100且标记数为50K时Pop2群体;h^(2)为0.5且QTL数为100时,标记数10K下Pop1群体和标记数50K下Pop2群体;除上述情况之外,MLP的h^(2)估计结果均显著(P<0.05)优于SSGBLUP、GBLUP和BLUP。在不同h^(2)初值下,QTL数和标记数变化时,Pop1和Pop2群体中MLP的h^(2)估值与当代群体h^(2)的差值小于SSGBLUP、GBLUP和BLUP;SSGBLUP和GBLUP法在不同标记数下遗传参数估计结果差别较大,MLP差别较小。在各情况下,MLP基因组估计育种值(genomic estimated breeding value,GEBV)的准确性均为最高。h^(2)初值为0.05时,MLP在标记数为10K时GEBV准确性略高于SSGBLUP在标记数为50K时的预测准确性。在h^(2)、QTL数和标记数相同的情况下,Pop2群体中各方法的EBV预测准确性较Pop1群体均有提升。根据上述模拟结果表明,在绵羊限性性状基因组选择中,MLP优于其他经典基因组选择方法。

This study aimed to apply multilayer perceptron(MLP)to genome selection of sex-limi ted traits in sheep,and compare it with other classical genome selection methods in various situations.In this study,Qmsim software was used to simulate the phenotype data and genotype data of 2 sheep population(Popl and Pop2).Artificial neural network(ANN)was used in MLP,and residual maximum likelihood(REML)method was used in linear model to estimate genetic parameters of different populations.Using Python to write MLP model,using DMU software to achieve best linear unbiased prediction(BLUP),genomic BLUP(GBLUP)and single-step GBLUP(SSGBLUP).The differences in the estimation of heritability(h^(2))and breeding values under different conditions were evaluated.In all cases,MLP and SSGBLUP were significantly(P<0.05)better than GBLUP and BLUP.There is no significant difference between h^(2) estimation of MLP and SSGBLUP in the 3 conditions:When h^(2) was 0.05,the QTL number was 100 and the marker number was 10K in the Pop2 population;when h^(2) was 0.2,the QTL number was 500 under the two marker number in Pop1,and the QTL number was 100 under the marker number of 50K in Pop2;when h^(2) was 0.5 and QTL number was 100,the marker number of 10K in Pop1 and the marker number of 50K in Pop2.Except for the above,the h^(2) estimates of MLP were significantly better(P<0.05)than SSGBLUP,GBLUP,and BLUP.Under different prior values of h^(2),when QTL number,marker number changed,the difference between the estimated h^(2) of MLP and that of contemporary population was smaller than that of SSGBLUP,GBLUP,and BLUP.The h^(2) estimation results of SSGBLUP and GBLUP methods were very different under different marker number,and the MLP difference was small.In all cases,the prediction accuracy of GEBV by MLP was the highest.When the prior value of h^(2) was 0.05,the GEBV accuracy of MLP at 10K was slightly higher than that of SSGBLUP at 50K.Under the same h^(2),number of QTL and marker number,the EBV prediction accuracy of each method in the Pop2 population was improved compared with that in the Pop1 population.According to the above simulation results,MLP is superior to other classical genome selection methods in the genome selection of sex-limited traits in sheep.