Hi All,

I am analyzing RNASeq data from 12 samples. I did the alignment and the count using Rsubread package. For differential expression genes I am using DESeq2, but it return 41% low count genes and only 71 significant.

>info<data.frame(condition=c(rep("TOX",6),rep("NONTOX",6)),DOX=c(rep("untreated",3),rep("treated",3),rep("untreated",3),rep("treated",3)))

>rownames(info) <- colnames(counts)

>info

                         condition       DOX
RARG_1_0uM       TOX untreated
RARG_2_0uM       TOX untreated
RARG_3_0uM       TOX untreated
RARG_1_1uM       TOX   treated
RARG_2_1uM       TOX   treated
RARG_3_1uM       TOX   treated
WT_1_0uM      NONTOX untreated
WT_2_0uM      NONTOX untreated
WT_3_0uM      NONTOX untreated
WT_1_1uM      NONTOX   treated
WT_2_1uM      NONTOX   treated
WT_3_1uM      NONTOX   treated

>dds <- DESeqDataSetFromMatrix(countData=counts, colData=info,design=~DOX+condition)

> levels(dds$DOX)
[1] "untreated" "treated"

> levels(dds$condition)
[1] "NONTOX" "TOX"   

>dds <-DESeq(dds)

> summary(res)

out of 40911 with nonzero total read count
adjusted p-value < 0.1
LFC > 0 (up)     : 48, 0.12% 
LFC < 0 (down)   : 23, 0.056% 
outliers [1]     : 550, 1.3% 
low counts [2]   : 19079, 47% 
(mean count < 5)
[1] see 'cooksCutoff' argument of ?results
[2] see 'independentFiltering' argument of ?results

Is this high % of low counts normal?

Note the mean count value that the automatic independent filtering found to be optimal:

low counts [2]   : 19079, 47% 
(mean count < 5)

This makes sense, because genes with average counts less than 5 are typically not powered enough to rise out of the sampling noise. You would need many more samples in order to find differences at such a low count.

So yes it is normal and expected to discard these low counts / low power genes before performing multiple test correction.