This might be a bit long, please bare with me. I'm conducting a differential expression analysis using limma - voom. My comparison is regarding response vs non-response to a cancer drug. However, I'm not getting any DE genes, absolute zeros. Someone here once recommended not to use contrast matrix for such a simple task, so I did that but then I get almost all 18 thousand genes as DE, so that's not good either. I don't know what is the problem here. The counts matrix and metadata matrix have the same samples with the exact same order. I filtered my counts matrix from noise genes earlier. I did everything according to this guide: https://bioconductor.org/packages/release/workflows/vignettes/RNAseq123/inst/doc/limmaWorkflow.html

Let me show you the code and the results I get:

Code:

d0 = DGEList(counts)
group = metadata$Response
d0$samples$group = group
geneid = rownames(d0)
d0$genes = geneid
d0 =  calcNormFactors(d0, method = 'TMM')
dim(d0)

## Design
design = model.matrix(~ 0 + Response, metadata) %>% as.data.frame()
colnames(design)[c(1,2)] = c('NoResponse','Response')

## Contrast
contrast = makeContrasts(NoResp_VS_Resp = NoResponse - Response, levels = colnames(design))

Voom <- voom(d0, design, plot = TRUE)
vfit <- lmFit(Voom, design = design) 
vfit  <- contrasts.fit(vfit , contrasts = contrast) 
efit <- eBayes(vfit) 
plotSA(efit, main = 'final model: Mean-Variance trend')

summary(decideTests(efit))

deg <- topTable(efit, 
                coef = 'NoResp_VS_Resp',
                 p.value = 0.05,
                 adjust.method = 'fdr',
                 number = Inf)

Now you might ask how does the Voom and the eBayes plots look like, here they are:

And here is the summary(decideTests(efit)) results when using contrast matrix:

And this is without contrast matrix: ( Almost 17 thousand significant genes, out of 18 thousand)

Can anyone help me figure out what is the problem ? If any additional info is needed I'll add it!

Failing to get DE genes is not an error. Indeed one of the key purposes of the limma software is to avoid giving DE genes when there isn't worthwhile evidence of differences. It is very common for drug response studies with human patients to show either no differences or weak differences. That isn't the fault of the software, just the reality of biological variation between human subjects and drug responses.

You asked the same question a few weeks ago and I replied:

"The limma documentation provides several ways to plot and trouble-shoot your data. At very least you should be making the plots shown in the limma workflow paper, especially the MDS plot."

It makes no difference to the results whether you use a contrast or not, provided you do it correctly. Steve Lianoglou showed you complete code last time to get the same results with and without a contrast. If you don't want to use a contrast then you have to reform the design matrix as Steve showed you.