Deseq2 positive results for genes highly variable between replicates

Question

[guillaume.rbt]

Hi all,

I'm using DESeq2 to find differentially expressed genes between two conditions from RNAseq data, with lots of replicates (46 in condition "1", 20 in condition "2").

I get results with significative adjusted p-values, but for most of them the gene expression values are highly variable between replicates.

For example for the gene with the lowest adjusted p-value, I've got all samples from both conditions with low normalized counts (around 10), and just one sample in one condition with >200000 normalized counts, which drives the differential expression toward this condition.

See log2(normalized counts + 1) boxplot below ( the adjusted p-value is 8.05e-12, and the log2FC is -5.87 between condition "1" and "2" for this gene)

Here is the code I used :

dds <- DESeqDataSetFromTximport(tx_import_data, coldata, ~condition)
keep <- rowSums(counts(dds)) >= 10
dds <- dds[keep,]
dds$condition <- relevel(dds$condition, ref = "R")
dds <- DESeq(dds)
res05 <- results(dds, alpha=0.05)

I'm wondering if this is "normal" that DESeq2 keeps those kinds of results and I that should filter it if I find it irrelevant, of if I made some mistake during the process and that DEseq2 should only keep genes without such expression dispersion between replicates?

Thank for your help

[ATpoint]

With only words but no plots illustrating your question it is difficult to make any statements. Please provide e.g. some boxplots of normalized counts or tables.

[guillaume.rbt]

Ok I've just put a link with a boxplot illustrating my example.

[ATpoint]

log2 scale please ;-) and see How to add images to a Biostars post. You have to paste the link with the full suffix like https...foo.png to the image box.

[guillaume.rbt]

done ;) sorry I never uploaded a plot before

[ATpoint]

I would check if these outliers samples also show outlier-like behaviour in a PCA maybe indicating a batch effect and if so, think about removing them.

[guillaume.rbt]

Ok thanks, I've checked that and unfortunately they don't seem to be different from the other ones on the PCA.

[Martombo]

In my experience, this kind of result typically stems from the presence of a very high variability in samples of the same group (compared to between groups). You may want to correct for possible co-variates in your data (see svaseq) or simply filter out results with high dispersion.