q-value is not as robust. Here is why.. or am I wrong?

Question

[mfahim]

I have four subjects at two different conditions (a total of 8 samples).. cuffdiff tells me there is no significant difference based on adjusted p-value (aka q-value).

This is why I guess q-value is just overrated.. and is not applicable to every RNAseq analysis.

What do you say?

Answer 1

[Sam]

There can be a lot of reason for there to be no significant differences. One simple explanation is that there is indeed no significant difference between your two conditions. You can try and plot a QQ plot of your un-adjusted p-value and see if they are normal. If all of them follow the null, then there is no significance.

[mfahim]

Sam, I tried posting picture with it.. but could not find attachment thingy here.. here is the same post on seqanswers.. There is a huge difference between WT and mutants.. check it out and kindly comment.. also let me know a bit more about qqplot and how it is plotted and where... I am a biologist and have no specific statistics or bioinformatics skills except for the basic usage..
http://seqanswers.com/forums/showthread.php?p=174983#post174983

[Sam]

One thing though, if I understand your graph correctly, you have 3 mutants and 1 control, and expose them to different treatment, so that if you draw a sample information table, you have something like

             Control  Mutant-1  Mutant-2  Mutant-3
Treatment-1  1        1         1         1
Treatment-2  1        1         1         1

And here, when you are talking about the difference in treatment response, you are grouping different samples together based only on their treatment, e.g.

               Treatment-1    Treatment-2
Sample Size    4              4

First thing that you can notice is that, although there seems to be difference between mutants, there isn't such a big difference between treatments. So I won't be surprised if you found no significance when comparing between the samples across treatment.

To draw the qq-plot, you can follow the instructions in wiki

[mfahim]

Spot on Sam..That's exactly the situation I am in. And this is affecting my volcano plots..

Can you please advise how should I plot my data or analyze it.

[Sam]

To be honest, with only 1 samples per condition, it is very difficult to get any sensible p-value out of an analysis. Even, when you merge the samples by their treatment group, the in-group variance will still be big considering how different your groups are. With this huge in-group variance and relatively small between group difference, it is most likely to not have any significance from your data. And what you've got from cufflink is actually correct: there is no significance

As Devon mentioned, it is definitely NOT the problem of the q-value, rather, it is the problem of your experimental design that is the problem.

To illustrate it, you are kind of comparing the effect of two different chemical on apple and orange and you tried to group the apple and orange together and test what is the effect of the chemical in general to fruits.

Answer 2

[Devon Ryan]

I already replied to this on your post on seqanswers.