edgeR中三种标准化方法TMM\UQ\RLE的比较

关于RNA-seq中的reads count标准化处理的方法汇总，请先看看这篇：
当我们在说RNA-seq reads count标准化时，其实在说什么？

本文集中讨论常用的edgeR包中三种标准化方法TMM\UQ\RLE的比较

英文原贴Normalisation methods implemented in edgeR

1.首先创建一个数据集

包含四个样品c1,c2是正常组，p1,p2是病人。共有50个转录本，每个样品内转录本counts的总数都是500个，前25个转录本在四个样品里都有表达，其中病人转录本的数目(20)是对照组(10)的两倍, 后25个转录本只在正常组中检测到。

#prepare example
control_1 <- rep(10, 50)
control_2 <- rep(10, 50)
patient_1 <- c(rep(20, 25),rep(0,25))
patient_2 <- c(rep(20, 25),rep(0,25))
 
df <- data.frame(c1=control_1,
                 c2=control_2,
                 p1=patient_1,
                 p2=patient_2)
 
head(df)
  c1 c2 p1 p2
1 10 10 20 20
2 10 10 20 20
3 10 10 20 20
4 10 10 20 20
5 10 10 20 20
6 10 10 20 20
 
tail(df)
   c1 c2 p1 p2
45 10 10  0  0
46 10 10  0  0
47 10 10  0  0
48 10 10  0  0
49 10 10  0  0
50 10 10  0  0
 
#equal depth
colSums(df)
 c1  c2  p1  p2 
500 500 500 500

数据集信息详见Robinson and Oshlack http://genomebiology.com/2010/11/3/R25

2.如果不做标准化处理

#load library
library(edgeR)
 
#create group vector
group <- c('control','control','patient','patient')
 
#create DGEList object
d <- DGEList(counts=df, group=group)
 
#check out the DGEList object
d
An object of class "DGEList"
$counts
  c1 c2 p1 p2
1 10 10 20 20
2 10 10 20 20
3 10 10 20 20
4 10 10 20 20
5 10 10 20 20
45 more rows ...
 
$samples
     group lib.size norm.factors
c1 control      500            1
c2 control      500            1
p1 patient      500            1
p2 patient      500            1
 
d <- DGEList(counts=df, group=group)
d <- estimateCommonDisp(d)
 
#perform the DE test
de <- exactTest(d)
 
#how many differentially expressed transcripts?
table(p.adjust(de$table$PValue, method="BH")<0.05)
 
TRUE
  50

可以看到：检测出共50个转录本有差异，即每个转录本都是差异表达的，假阳性很高。

3.TMM normalisation

TMM <- calcNormFactors(d, method="TMM")
TMM
An object of class "DGEList"
$counts
  c1 c2 p1 p2
1 10 10 20 20
2 10 10 20 20
3 10 10 20 20
4 10 10 20 20
5 10 10 20 20
45 more rows ...
 
$samples
     group lib.size norm.factors
c1 control      500    0.7071068
c2 control      500    0.7071068
p1 patient      500    1.4142136
p2 patient      500    1.4142136

我们看到对前25个转录本而言，正常组和病人之间没有差异 (10/0.7071068 (~14.14) 等于 20/1.4142136 (~14.14))。因此检测出有25个转录本存在差异（后25个转录本）

TMM <- estimateCommonDisp(TMM)
TMM <- exactTest(TMM)
table(p.adjust(TMM$table$PValue, method="BH")<0.05)
 
FALSE  TRUE
   25    25

4.RLE normalisation

RLE
An object of class "DGEList"
$counts
  c1 c2 p1 p2
1 10 10 20 20
2 10 10 20 20
3 10 10 20 20
4 10 10 20 20
5 10 10 20 20
45 more rows ...
 
$samples
     group lib.size norm.factors
c1 control      500    0.7071068
c2 control      500    0.7071068
p1 patient      500    1.4142136
p2 patient      500    1.4142136
RLE <- estimateCommonDisp(RLE)
RLE <- exactTest(RLE)
table(p.adjust(RLE$table$PValue, method="BH")<0.05)
 
FALSE  TRUE
   25    25

5.UQ normalisation

uq
An object of class "DGEList"
$counts
  c1 c2 p1 p2
1 10 10 20 20
2 10 10 20 20
3 10 10 20 20
4 10 10 20 20
5 10 10 20 20
45 more rows ...
 
$samples
     group lib.size norm.factors
c1 control      500    0.7071068
c2 control      500    0.7071068
p1 patient      500    1.4142136
p2 patient      500    1.4142136
 
uq <- estimateCommonDisp(uq)
uq <- exactTest(uq)
table(p.adjust(uq$table$PValue, method="BH")<0.05)
 
FALSE  TRUE
   25    25

因为数据比较简单，这里三种标准化方法得到的结果一致，那么真实测序数据的情况又如何呢？

6.测试一套真实数据

my_url <-"[https://davetang.org/file/pnas_expression.txt](https://davetang.org/file/pnas_expression.txt)"

data <-read.table(my_url, header=TRUE, sep="\t")

dim(data)

[1] 37435     9

ensembl_ID lane1 lane2 lane3 lane4 lane5 lane6 lane8  len

1 ENSG00000215696     0     0     0     0     0     0     0  330

2 ENSG00000215700     0     0     0     0     0     0     0 2370

3 ENSG00000215699     0     0     0     0     0     0     0 1842

4 ENSG00000215784     0     0     0     0     0     0     0 2393

5 ENSG00000212914     0     0     0     0     0     0     0  384

6 ENSG00000212042     0     0     0     0     0     0     0   92

准备DGEList

rownames(d) <- data[,1]
group <- c(rep("Control",4),rep("DHT",3))
d <- DGEList(counts = d, group=group)
 
An object of class "DGEList"
$counts
                lane1 lane2 lane3 lane4 lane5 lane6 lane8
ENSG00000215696     0     0     0     0     0     0     0
ENSG00000215700     0     0     0     0     0     0     0
ENSG00000215699     0     0     0     0     0     0     0
ENSG00000215784     0     0     0     0     0     0     0
ENSG00000212914     0     0     0     0     0     0     0
37430 more rows ...
 
$samples
        group lib.size norm.factors
lane1 Control   978576            1
lane2 Control  1156844            1
lane3 Control  1442169            1
lane4 Control  1485604            1
lane5     DHT  1823460            1
lane6     DHT  1834335            1
lane8     DHT   681743            1

还是先不做标准化处理

no_norm <- exactTest(no_norm)
table(p.adjust(no_norm$table$PValue, method="BH")<0.05)
 
FALSE  TRUE
33404  4031

TMM normalisation

TMM <- calcNormFactors(d, method="TMM")
TMM
An object of class "DGEList"
$counts
                lane1 lane2 lane3 lane4 lane5 lane6 lane8
ENSG00000215696     0     0     0     0     0     0     0
ENSG00000215700     0     0     0     0     0     0     0
ENSG00000215699     0     0     0     0     0     0     0
ENSG00000215784     0     0     0     0     0     0     0
ENSG00000212914     0     0     0     0     0     0     0
37430 more rows ...
 
$samples
        group lib.size norm.factors
lane1 Control   978576    1.0350786
lane2 Control  1156844    1.0379515
lane3 Control  1442169    1.0287815
lane4 Control  1485604    1.0222095
lane5     DHT  1823460    0.9446243
lane6     DHT  1834335    0.9412769
lane8     DHT   681743    0.9954283
 
TMM <- estimateCommonDisp(TMM)
TMM <- exactTest(TMM)
table(p.adjust(TMM$table$PValue, method="BH")<0.05)
 
FALSE  TRUE
33519  3916

RLE

RLE <- calcNormFactors(d, method="RLE")
RLE
An object of class "DGEList"
$counts
                lane1 lane2 lane3 lane4 lane5 lane6 lane8
ENSG00000215696     0     0     0     0     0     0     0
ENSG00000215700     0     0     0     0     0     0     0
ENSG00000215699     0     0     0     0     0     0     0
ENSG00000215784     0     0     0     0     0     0     0
ENSG00000212914     0     0     0     0     0     0     0
37430 more rows ...
 
$samples
        group lib.size norm.factors
lane1 Control   978576    1.0150010
lane2 Control  1156844    1.0236675
lane3 Control  1442169    1.0345426
lane4 Control  1485604    1.0399724
lane5     DHT  1823460    0.9706692
lane6     DHT  1834335    0.9734955
lane8     DHT   681743    0.9466713
 
RLE <- estimateCommonDisp(RLE)
RLE <- exactTest(RLE)
table(p.adjust(RLE$table$PValue, method="BH")<0.05)
 
FALSE  TRUE
33465  3970

the upper quartile method

uq <- calcNormFactors(d, method="upperquartile")
uq
An object of class "DGEList"
$counts
                lane1 lane2 lane3 lane4 lane5 lane6 lane8
ENSG00000215696     0     0     0     0     0     0     0
ENSG00000215700     0     0     0     0     0     0     0
ENSG00000215699     0     0     0     0     0     0     0
ENSG00000215784     0     0     0     0     0     0     0
ENSG00000212914     0     0     0     0     0     0     0
37430 more rows ...
 
$samples
        group lib.size norm.factors
lane1 Control   978576    1.0272514
lane2 Control  1156844    1.0222982
lane3 Control  1442169    1.0250528
lane4 Control  1485604    1.0348864
lane5     DHT  1823460    0.9728534
lane6     DHT  1834335    0.9670858
lane8     DHT   681743    0.9541011
 
uq <- estimateCommonDisp(uq)
uq <- exactTest(uq)
table(p.adjust(uq$table$PValue, method="BH")<0.05)
 
FALSE  TRUE
33466  3969

以上四种处理方法找到的差异基因取交集，可以看出不做标准化处理会得到405个假阳性和342个假阴性的转录本

library(gplots)
 
get_de <- function(x, pvalue){
  my_i <- p.adjust(x$PValue, method="BH") < pvalue
  row.names(x)[my_i]
}
 
my_de_no_norm <- get_de(no_norm$table, 0.05)
my_de_tmm <- get_de(TMM$table, 0.05)
my_de_rle <- get_de(RLE$table, 0.05)
my_de_uq <- get_de(uq$table, 0.05)
 
gplots::venn(list(no_norm = my_de_no_norm, TMM = my_de_tmm, RLE = my_de_rle, UQ = my_de_uq))

$edgeR中三种标准化方法TMM\UQ\RLE的比较_第1张图片$

不做标准化会得到405个假阳性和342个假阴性的转录本

三种标准化方法找到的差异基因大部分是一致的

gplots::venn(list(TMM = my_de_tmm, RLE = my_de_rle, UQ = my_de_uq))

$edgeR中三种标准化方法TMM\UQ\RLE的比较_第2张图片$

1.png

小结

三种标准化方法效果类似，处理结果都比不做标准化要好
The normalisation factors were quite similar between all normalisation methods, which is why the results of the differential expression were quite concordant. Most methods down sized the DHT samples with a normalisation factor of less than one to account for the larger library sizes of these samples.