edgeR: Correct pipeline for DE analysis with multiple conditions and batches
1
2
Entering edit mode
7.1 years ago
MT ▴ 40

I'm using edgeR to test for DE genes in an expression matrix. For each sample if have a condition and a batch. I want to use the glm-functionality in edgeR to test for DE between conditions, while taking into account batches.

Some example data to show my problem. Say if have a count matrix, EM,  of 10 sample with the following labels:

EM = EM # Imagine matrix of counts here...

conditions = c("con1", "con1", "con1", "con2", "con2", "con2", "con3", "con3", "con3", "con3")

batches = c("batch1", "batch2", "batch3", "batch1", "batch2", "batch3", "batch1", "con2", "con3", "con3")

The pipeline could look like this:

dge = DGEList(counts=EM) # Create object

dge = calcNormFactors(dge, method='TMM') # Normalize library sizes using TMM

design = model.matrix(~0+conditions+batches) # Create design matrix for glm

colnames(design) = c(levels(conditions), levels(batches)[2:length(levels(batches))]) # Set prettier column names

dge = estimateGLMCommonDisp(dge, design) # Estimate common dispersion

dge = estimateGLMTagwiseDisp(dge, design) # Estimate tagwise dispersion

fit = glmFit(dge,design) # Fit glm

pair_vector = sprintf("%s-%s", "con1", "con3") # Samples to be compared
pair_contrast = makeContrasts(contrasts=pair_vector, levels=design) # Make contrast

lrt = glmLRT(fit, contrast=pair_contrast) # Likelihood ratio test

 

My questions:

1) The design matrix: There is no baseline conditions, so I remove the intersect with the 0+. Is this necessary to do for batches as well, even though they are not directly used as contrasts?

2) Does the glmFit take into account norm-factors for library sizes?

R edgeR RNA-Seq DE • 16k views
ADD COMMENT
0
Entering edit mode

Do you want to just compare "con3" and "con2" vs. "con1" or do all of the pairwise comparisons? Your setup is more targeted toward comparing things versus con1 and including an intercept rather than using contrasts.

ADD REPLY
0
Entering edit mode

I want to test all pairwise comparisons between conditions (no condition can be considered baseline'). I omitted the for-loop for clarity.

ADD REPLY
7
Entering edit mode
7.1 years ago

There are two ways to go about this. Firstly, either allow an intercept and then just use a contrast for the cond3 to cond2 comparison, or don't allow an intercept and group things different (e.g., group=sprint("%s.%s",batch,condition)).

If we just allow an intercept then:

design <- model.matrix(~conditions+batches)
dge = estimateGLMCommonDisp(dge, design)
dge = estimateGLMTagwiseDisp(dge, design)
fit = glmFit(dge,design)
lrt2vs1 <- glmLRT(fit, coef=2)
lrt3vs1 <- glmLRT(fit, coef=3)
lrt3vs2 <- glmLRT(fit,contrast=c(0,-1,1,0,0))

BTW, have a read through the edgeR user's guide, particularly section 3.4.2, which has an almost identical example (it's a really well done user's guide).

ADD COMMENT
0
Entering edit mode

Thanks for the response! I would like to directly use the names of conditions when doing the testing, rather than using an Intercept column and coef - how would you go about that?

ADD REPLY
1
Entering edit mode

You can use names for the coef= argument, but I believe that contrast only takes numeric vectors and matrices. You should be able to use my example to determine how to do one without an intercept.

ADD REPLY
0
Entering edit mode

Hi Devon

I found a interesting but naive staff.

When I use

design1 <- model.matrix(~conditions) # with Intercept column
d <- estimateDisp(d, design, robust=TRUE)
fit <- glmFit(d, design)
lrt <- glmLRT(fit)

and

design2 <- model.matrix(~0+conditions) # without Intercept

e.g. design2 like

  conditionN conditionT
1          1          0
2          1          0
3          0          1
4          0          1

The design2 result is much more liberal than design1.

Why?

Which do you recommend? (I don't have muti-factor, the comparison is based on same timepoint).Thank you!

ADD REPLY
1
Entering edit mode

They end up being identical, though you MUST explicitly specify the contrast with the second method.

ADD REPLY
0
Entering edit mode

I got. design1 is

lrt <- glmLRT(fit)

coef = last column(default),

design2 is something like

lrt <- glmLRT(fit,contrast=(-1,1))

Am I right?

ADD REPLY
1
Entering edit mode

That looks correct at least

ADD REPLY
0
Entering edit mode

I made a quick script for this.

pdacDGE <- function(group,ref.groups,comp.groups) {
  contrast <- c(0,0,0,0,0,0,0,0,0,0,0,0,0)
  for (i in 1:length(ref.groups)) {
    contrast[which(levels(group) == ref.groups[i], arr.ind = TRUE)] <- -(1/length(ref.groups))
  }
  for (j in 1:length(comp.groups)) {
    contrast[which(levels(group) == comp.groups[j], arr.ind = TRUE)] <- 1/length(comp.groups)
  }
  print(contrast)
  lrt<-glmLRT(glm.fit,contrast = contrast)
  tt<-topTags(lrt, n=500000, p.value = .05, sort.by = "logFC")
  print(contrast)
  print(length(rownames(tt)))
  return(tt)
}
ADD REPLY
0
Entering edit mode

How can i run edgeR for multiple counts and sample info file using a for-loop? Thanks

ADD REPLY

Login before adding your answer.

Traffic: 1224 users visited in the last hour
Help About
FAQ
Access RSS
API
Stats

Use of this site constitutes acceptance of our User Agreement and Privacy Policy.

Powered by the version 2.3.6