Contrasts with DESeq2

[mistrm]

Hi;

I am working with an interesting design and have setup the contrasts with DESeq2 and unsure if it setup correct as I get quite alot significant genes.

The setup is as follows:
6 neurontypes (each with three replicates) which can be subdivided into 2 groups (LTMR and Nociceptors). I want to compare each neurontype against all neurontypes of the opposite group. So for example given that neurontype1 is in LTMR - what genes are DE between neurontype1 and all Nociceptor samples? The code I used is pasted below:

# Create DESeq2 dataset
se <-SummarizedExperiment(assays=SimpleList(counts=exprs(sub.eset)), colData=DataFrame(pData(sub.eset)))
dds <- DESeqDataSet(se=se, design = ~ neurontype)
dds <- DESeq(dds)

# Create list object for the results
resultsAll <- vector("list", 6)

# adding the effects across all neurontypes within a group to create background contrast
bg.ltmr <- resultsNames(dds)[grep("LTMR", resultsNames(dds))]
bg.nociceptor <- resultsNames(dds)[grep("Nociceptor", resultsNames(dds))]

# Extract data of specified contrasts with appropriate background
for (n in 2:length(resultsNames(dds))){
r <- resultsNames(dds)[n]
if (r %in% bg.ltmr) contrast <- list(r, c(bg.nociceptor))
if (r %in% bg.nociceptor) contrast <- list(r, c(bg.ltmr))
res <- results(dds, contrast=contrast)
resultsAll[n-1] <- list(res)
}
names(resultsAll) <- resultsNames(dds)[-1]

I get about ~3000 genes significant for each of the neurontypes which seems a bit high to me. Any thoughts on if there is a better way to do this?

Thanks!

[Richard Finney]

Others have noted the large number of genes identified in deseq2 : http://seqanswers.com/forums/showthread.php?t=30907

I notice it too.

If you really, really need a smaller number of genes, then change the statistical significance cut off.

[Michael Love]

I'd recommend you look at a PCA plot (see vignette for example). This just sounds like you have small within group variance compared to the between group differences, which you will see by looking at the top 500 genes with plotPCA. You can either reduce the FDR cutoff as Richard recommends, or increase the lfcThreshold if you are interested in larger fold changes (as Simon just responded there http://seqanswers.com/forums/showpos...3&postcount=21).

[mistrm]

That makes sense, the within group variance for the single neurontype is going to be much smaller since the samples are just replicates. Thanks, I'll probably just increase lfcThreshold and take genes from there