Question: Paired analysis (tumor vs normal) at three different conditions with EdgeR
0
5 months ago by
takumima10
takumima10 wrote:

Hi all,

I have 18 samples, 3 tumor and 3 paired normal tissue samples, belonging to the same individual, at three different conditions, each in three replicates.

 $df sample group tissue sample1-T-cond1 cond1 T sample1-T-cond1 cond1 T sample1-T-cond1 cond1 T sample1-T-cond2 cond2 T sample1-T-cond2 cond2 T sample1-T-cond2 cond2 T sample1-T-cond3 cond3 T sample1-T-cond3 cond3 T sample1-T-cond3 cond3 T sample1-N-cond1 cond1 N sample1-N-cond1 cond1 N sample1-N-cond1 cond1 N sample1-N-cond2 cond2 N sample1-N-cond2 cond2 N sample1-N-cond2 cond2 N sample1-N-cond3 cond3 N sample1-N-cond3 cond3 N sample1-N-cond3 cond3 N  My goal is to compare T versus N (tissue), in each condition (group). In other words, I would like to get the list of DE genes in tumor samples, and I'd like to know if this list is the same in the three different conditions. I can't figure out how to manage the replicates and how to set the DGEList object and design matrix:  geneList = DGEList(counts=round(geneData), genes=rownames(geneData), group = df$group)
design = model.matrix(~df$group+df$tissue)


Is this solution right?

Thanks

edger design de matrix • 176 views
modified 5 months ago by Aaron Lun24k • written 5 months ago by takumima10
Answer: Paired analysis (tumor vs normal) at three different conditions with EdgeR
1
5 months ago by
Aaron Lun24k
Cambridge, United Kingdom
Aaron Lun24k wrote:

Paste them together and use a one-way layout.

group <- paste0(df$group, ".", df$tissue)
design <- model.matrix(~0 + group)


And if you want to compare tissues in each condition, you can just do:

con <- makeContrasts(
groupcond1.T - groupcond1.N, # T vs N in cond1
groupcond2.T - groupcond2.N, # ... in cond2
groupcond3.T - groupcond3.N, # ... in cond3
levels=design)


Take the desired column and use it in contrasts= in glmQLFTest() to test individual comparisons. Or you can just supply contrasts=con to do an ANODEV, i.e., to identify any differences between T and N in any condition.

The formulation of design above assumes that there's no batch structure that requires additional blocking factors. If there is, then you'll need to include the necessary terms as additive factors.

It might be worth mentioning that the design matrix shown in the question also is correct but it's a bit harder to think about which coefficient to look at for which comparison of interest.

2

Yes, the original design matrix is valid but assumes that the effects of group and tissue are additive. My design sacrifices some residual d.f. to eliminate this assumption, which is important if the tumor/normal difference is condition-specific (i.e., there is a non-zero interaction between group and tissue). In contrast, the original design assumes that the tumor-normal difference is the same for all conditions.

Thank you @Aaron Lun.

I made two more comparisons, namely conditions 2 and 3 versus the cond1:

cond1vscond2=(groupcond1.T-groupcond1.N)-(groupcond2.T-groupcond2.N)
cond1vscond3=(groupcond1.T-groupcond1.N)-(groupcond3.T-groupcond3.N)


So, when I use geneQLFT = glmQLFTest(geneQLF, contrast=con), I get the following

> summary(decideTestsDGE(geneQLFT, adjust.method="BH"))
LR test on 3 degrees of freedom
NotSig                           1773
Sig                              13657


Accordingly, I can assess that 13657 genes are DE across all paired-samples. This means that different conditions impact differently the results, but what is the best one?