Basal gene expression threshold

[R.Vin]

Dear Colleagues

I am using the RNAseq pipeline: Tophat -> HTseq -> DEseq
I have RNAseq samples in triplicate (3x wt , 3x mutant)

I wanted to see which genes are switched ON (from no expression to expression) in the mutant.
Just by looking the amount of read counts on genes one can see the easy scenario

wt1 wt2 wt3 mutant1 mutant2 mutant3 DESeq log2FC
0 0 0 13 5 7 Inf
0 0 0 1240 1378 1200 Inf

my question arises when I see a small number of reads on the wt
for example

wt1 wt2 wt3 mutant1 mutant2 mutant3 DESeq log2FC
0 1 0 715 1024 920 11,3
1 1 1 2107 2997 2572 11,2
2 2 0 2660 3131 3472 11,1
0 1 0 747 801 642 11,1
2 2 2 3827 4187 4222 11
1 0 0 548 783 789 11
0 1 1 758 1813 1464 11
0 1 0 547 597 695 11
2 1 0 1618 2251 1709 11


Are these small number of read counts random? noise? background expression? bad aligned reads? or real expression?.
Which is the threshold to say that a gene is expressed ? (for example: 3 read counts in all 3 wt samples).

Thanks in advance for your insights.

[dpryan]

There is no accepted threshold for this. The best you can do is look at regions of similar composition and compare how many reads map to them and how many map to these genes (you'll need to length normalize).

Also, this a biologically questionable thing to ask anyway. Say you look at protein levels with a Western. Does no band mean no expression or just no detected expression? If there's one copy of the protein floating around, you're very likely to miss it with any method (likewise with measuring RNA), but it's still expressed (Or is this just transcriptional/translational noise? Is it even possible to determine this (generally it's not)).

[mbblack]

RNA-seq does not measure absolute expression, so I agree your very question is inappropriate for the data you have. To begin with, by definition, the only genes you could even detect with RNA-seq are ones that are "switched on" as otherwise, there would not be a cellular transcript to capture when you sampled. A raw count of zero is simply a failure to detect, not confirmation of the absence of a transcript. And your greatest measurement bias is inherently in the low count features, as those are the ones with the highest probability of failure to detect. And for those which are detected, they will have the least reliable estimate of abundance (since they are the rarer events).

Add in the bias in raw counts related to library artifacts and other experimental vagaries, and any comparison of raw counts between samples becomes highly questionable at best. You cannot infer the initiation of transcription just from relative expression data.

[danielr]

I find that expression in RNA-seq often follows a bimodal expression, with two bumps of genes and a valley with its lowest point at like 0.1-1 RPKM. That makes the lower bump background of some sort. It's sometimes there in single-cell RNA-seq as well, at well below single-molecule levels, so mismapping is my guess.
Plot the distribution of expression levels and see if there's a natural cutoff.