Best approach? (DE- RNAseq)

[SylvainL]

Dear all,

I have RNAseq to analyze (DE between different conditions). Mapping was done using STAR and I was planning to use edgeR for DE.

My concern is about how to proceed due to the particular design of the experiment.

I have two sets of experiments (each condition is done in triplicates, real biological triplicates).

Experiment #1:
3 conditions: Control - Treatment#1 - Treatment#2

Experiment #2:
3 conditions as well: Control - Treatment#1 - Treatment#3 (not #2 !!!)

If I analyze separately experiments 1&2, and first look at the DE genes between Treatment#1 and Control, I see some differences between both experiments. So, I did another analysis considering all the samples for Control and Treatment#1 (since then, it is like having 6 biological replicates for this Treatment).

My question is now, how do I process for Treatment#2 and #3 ? In this case, I won't have 6 replicates for each treatment (even though I will have 6 replicates for the Control and Treatment#1)? It has importance because Treatments 2 or 3 came after Treatment1 and the idea is to see the behavior of the genes affected by Treatment#1 after Treatments 2 and 3...

What is the best approach?



Thanks in advance,

s.

[dpryan]

I assume that experiment 1 and 2 were done at different times or that there's some other batch effect confounding just throwing everything together into the same experiment. In that case, I would load all of the samples at once and add an "experiment" variable to the design. You can then use contrasts for you comparisons of interest. The imbalance in the number of samples between the different conditions shouldn't be a problem. BTW, this will allow comparing treatment 2 and 3, in case that's at all biologically interesting.

[SylvainL]

Thanks dpryan,

These experiments are done with primary cells so I was focusing on the fact that differences may be due to the different donors (human primary cells). But true, both experiments were done at different times... Actually, each donor (so each biological replicate for both experiments) was obtained a different week, so I thought that considering the donor/patient as a variable to the design was enough... To summarize, the design considers the treatments and the donors.

My concern is that when I take the 6 biological replicates for Treatment#1, I detect some genes as differentially expressed (FC_threshold of 3 and pvalue_threshold of 0.001) which are not if I take only Experiment#1 (for example). So, do I exclude these guys for the rest of the analysis for Experiment#1?

My idea was to analyse independently both Experiments, analyse them together and take the overlaps between the 3 analyses... Questions: is it relevant? Would it be acceptable by referees? I'm aware I will reject some genes by taking only the overlaps (about 10% of the initial DE genes)...

s.

[dpryan]

It's not just a matter of treatment and donors, but also when the samples were actually prepared for sequencing (and RNA extracted). I've seen batch effects due to that sort of thing often enough to always include it in models when relevant.

Regarding the differences in results when using everything at once vs. just 3 samples at a time, that's not surprising. Statistical power comes form the number of samples you have, so particularly if you have one sample that's a bit of an outlier, you'll get a big bump in power with the additional samples.

I'm generally not a fan of taking overlaps of p-value delineated groups. There's just too much noise, particularly around p=0.05 or 0.1, for that to have much utility.

[SylvainL]

Ok, I will add the parameter Experiment in the design to see the results.

I am also not really concerned by the differences, particularly because it comes from human primary cells (so quite different donors, in addition to all possible batch effect, etc...).

About taking overlaps of p-value delineated groups, I am not a fan neither but it is the request of the groupleader. That's why I set the threshold to 0.001 (not 0.05).

s.

[dpryan]

...I am not a fan neither but it is the request of the groupleader.

Can't argue with that

[SylvainL]

Unfortunately me neither.... (even if sometimes I have to bite my tongue)