[dpryan]

The ~patient+condition vs. ~patient will handle all of that.

[SylvainL]

Ok, thanks a lot...

Just to give a comparison, with the same pvalue_threshold (0.01), when I take into consideration the paptients, I have 565 DE genes, when I do not consider the patient parameter, I have 2563!!!
I believe you when you say this is due to the noise in human RNAseq (and of course of the low number of replicates) but I have to confess, I am quite surprised. My god feeling was that the number of DE genes would be higher taking into consideration the patients than not considering this parameter...

Can I stricly use the same pvalue_threshold in both analyses?

[dpryan]

I assume these are the adjusted p-values. Yes, you can use the same threshold regardless of the comparison. 0.01 is extremely restrictive. A common threshold is 0.1 for adjusted p-values (though depending on what you want to do with things, perhaps 0.01 makes more sense for you).

[SylvainL]

Yes, I am speaking about the adjusted pvalues...

I prefer to start the analysis being restrictive in order to see which common points the regulated genes are sharing... The idea is to make predictions for the researcher who can go back to bench and then test some hypostheses (based on my analysis)...

The difference in the number of concerned genes really does not seem to shock you .

[dpryan]

I've done enough analyses that I'm not easily shocked

[SylvainL]

Hi,

so I had a closer look at the data and I still have a problem...

For example, for gene a (one of the genes of interest, already validated by qPCR by the researcher), I have these values (here in rpkm to make easier the comparison but I was obviously using the raw counts for the DE analyses):
Patient A: no stimulus: 0.22; stimulus 1: 130 so fold-induction: 590
Patient B: no stimulus: 0.43; stimulus 1: 332 so fold_induction: 770
Patient C: no stimulus: 0.50; stimulus 1: 600 so fold_induction: 600

This gene is considered as significantly induced when I do not take into consideration the patient parameter; but not when I compare ~patient+condition to ~patient...

I really do not understand why. The only way to have it in both analyses is to increase like hell the pvalue_threshold...

s.

[dpryan]

My guess is that this is due to the limited N. It's real in any case but I suspect that the low sample number is really hindering things.

[SylvainL]

Yes, I agree the N is quite low but in this particular case, the fold-changes are really high... I guess we are at the limit of the understanding between Biologists and Statisticians here. Every Biologist will obviously not understand why this gene is not considered significant...

I am using the "BH" method for the GLM analysis, I was wondering if changing the model would allow the detection of such genes. I will try next week...

I also did my own script for DE, which consider the fold-changes and look wheither they are similar or not between patients but the problem of this approach is I have to pass by a normalization step for all the patients at the "no stimulus" condition, so for this condition, I arrive with the same normalized values for the 3 patients. Then, I detect genes I know they are regulated but I wonder how to justify it in the mat&meth (of course the finality of this is publication...).

I will also try edgeR to see the results...

[mbblack]

Looking at your RPKM values, and as mentioned, given your lack of replication, I'd not be surprised either. It looks to me like your controls barely detected these genes in the first place, so I'm not at all sure how much (if any) faith I would have in their actual abundance estimate to begin with. Those sorts of massive fold change are really only believable to me if you were able to capture the background signal (i.e. the controls) with reasonably high confidence. Without high numbers of replicates, and much higher read count than you appear to have had, I simple don't believe such high estimates of fold change given how very low the control sample signal appears to be. And your ability to distinguish such changes with statistical confidence will be poor at best.