||
最佳线性无偏预测(best linear unbiased prediction, 简称BLUP),又音译为“布拉普”[1],是统计学上用于线性混合模型对随机效应进行预测的一种方法。最佳线性无偏预测由C.R. Henderson提出。随机效应的最佳线性无偏预测(BLUP)等同于固定效应的最佳线性无偏估计(best linear unbiased estimates, BLUE)(参见高斯-马尔可夫定理)。因为对固定效应使用估计一词,而对随机效应使用预测,这两个术语基本是等同的。BLUP被大量使用于动物育种。——wiki
BLUP值,相当于是对混合线性模型中随机因子的预测;
BLUE值,相当于是对混合线性模型中固定因子的估算
predict means:预测均值,固定因子和随机因子都可以预测均值,它的尺度和表型值尺度一致
setwd("D:\\02 ASReml\\blue VS blup")library(asreml)library(tidyverse) dat <- read.csv("MaizeRILs.csv",head=T)for (i in 1:4) dat[,i] <- as.factor(dat[,i])
as1 <- asreml(height ~ location/rep + location*RIL,data=dat)
ASReml: Tue May 08 11:07:55 2018 LogLik S2 DF wall cpu -723.8797 64.8862 244 11:07:55 0.1 -723.8797 64.8862 244 11:07:55 0.0 Finished on: Tue May 08 11:07:55 2018 LogLikelihood Converged
ablue <- coef(as1)$fixed blue1 <- ablue[grep("^RIL_RIL*",rownames(ablue)),] %>% as.data.frame() head(blue1)
pv1 <- predict(as1,"RIL")$predictions$pvals
ASReml: Tue May 08 11:13:33 2018 LogLik S2 DF wall cpu -723.8797 64.8862 244 11:13:33 0.0 -723.8797 64.8862 244 11:13:33 0.0 Finished on: Tue May 08 11:13:33 2018 LogLikelihood Converged
head(pv1) # 类似SAS中的lsmeans
as2 <- asreml(height ~ 1,random = ~location/rep + location*RIL,data=dat)
ASReml: Tue May 08 11:13:34 2018 LogLik S2 DF wall cpu -1646.5302 233.5135 495 11:13:34 0.0 -1569.0397 137.9186 495 11:13:34 0.0 -1507.3257 94.6888 495 11:13:34 0.0 -1471.3354 74.5149 495 11:13:34 0.0 -1462.9209 67.6142 495 11:13:34 0.0 -1461.7649 65.3553 495 11:13:34 0.0 -1461.7228 64.9069 495 11:13:34 0.0 -1461.7228 64.8863 495 11:13:34 0.0 -1461.7228 64.8862 495 11:13:34 0.0 Finished on: Tue May 08 11:13:34 2018 LogLikelihood Converged
blup <- coef(as2)$random blup2 <- blup[grep("^RIL_RIL-*",rownames(blup)),] %>% as.data.frame() head(blup2)
pv2 <- predict(as2,"RIL")$predictions$pvals
ASReml: Tue May 08 11:13:34 2018 LogLik S2 DF wall cpu -1461.7228 64.8862 495 11:13:34 0.0 -1461.7228 64.8862 495 11:13:34 0.0 -1461.7228 64.8862 495 11:13:34 0.0 -1461.7228 64.8862 495 11:13:34 0.0 Finished on: Tue May 08 11:13:34 2018 LogLikelihood Converged
head(pv2)
summary(as2)$varcomp
str(dat)
'data.frame': 496 obs. of 9 variables: $ location: Factor w/ 4 levels "ARC","CLY","PPAC",..: 1 1 2 2 3 3 4 4 1 1 ... $ rep : Factor w/ 2 levels "1","2": 1 2 1 2 1 2 1 2 1 2 ... $ block : Factor w/ 8 levels "1","2","3","4",..: 4 6 5 4 8 5 1 4 1 2 ... $ plot : Factor w/ 122 levels "1","2","3","4",..: 28 47 36 92 64 40 7 27 6 9 ... $ RIL : Factor w/ 62 levels "RIL-1","RIL-11",..: 1 1 1 1 1 1 1 1 2 2 ... $ pollen : int 73 74 71 73 97 95 72 72 69 69 ... $ silking : int 77 79 74 77 101 100 78 78 71 72 ... $ ASI : int 4 5 3 4 4 5 6 6 2 3 ... $ height : num 182 169 213 203 156 ...
VSNR::pin(as2,h2 ~ V3/(V3 + V4/4 + V5/(2*4)))
library(openxlsx) write.xlsx(blue1,"blue.xlsx") write.xlsx(blup2,"blup.xlsx") write.xlsx(pv1,"pm1.xlsx") write.xlsx(pv2,"pm2.xlsx")
RIL是基因型
pm2-random是RIL作为随机因子的预测均值
pm1-fixed是RIL作为固定因子时的预测均值
blue是RIL作为固定因子的BLUE值
blup是RIL作为随机因子的BLUP值
pm2-blup 是随机因子的预测均值 减去 随机因子的BLUP值,可以看到得到的是一个常数(均值)
pm1-mu-random 是固定因子的预测均值 减去 固定依着你的BLUE值, 可以看到不是一个常数
blup/blue_effect=heritibility 是BLUP值 除以 BLUE效应值,得到的是遗传力常数
备注:blue_effect是用固定因子的预测均值 减去 整体均值
1,如果想用固定因子的效应值,可以使用固定因子的predict means(预测均值)-整体均值,因为固定因子的BLUE第一个强制为0,可以用预测均值代替
2,随机因子的预测均值和BLUP值,在数据平衡时,他们之差为常数(整体均值)
3,尽量使用BLUP值(或者随机因子的预测均值)作为排名依据,因为它相对于固定因子的效应值,考虑了遗传力的影响
Archiver|手机版|科学网 ( 京ICP备07017567号-12 )
GMT+8, 2024-11-23 13:33
Powered by ScienceNet.cn
Copyright © 2007- 中国科学报社