How to present a subset of gene expression values from RNAseq in a publication?

[sindrle]

Hi!
Im curious, my supervisor want to include 5 genes from the Cuffdiff results in a table for publication.

He just wants fold changes and standard deviation (and star if significant etc), but first of all I don't know how to calculate SD for FC.. Any suggestions?

If FC+SD is not the solution, what should I send him?


Thank you very much.

[sindrle]

My suggestion so far is:
1. Take FPKM values and test values

Here is an example:

q1_FPKM q1_conf_lo q1_conf_hi
1,20921 0,404779 2,01361

q2_FPKM q2_conf_lo q2_conf_hi
0,833905 0,24662 1,4278

Gives

FC loFC hiFC
0,689627939 0,609270738 0,709074746

I present this as: 0,69 +/-0.05*

*just marks if significant.. (based on FDR)

[dpryan]

Note that the confidence intervals aren't symemetric, so "0.69, 95% CI [0.61, 0.71]" would seem to make more sense.

Coming up with reasonable confidence intervals or SDs by hand when dispersion is being squeezed with an empirical Bayes method is going to be non-trivial. I remember you mentioning edgeR and such before, so you might ask on the bioconductor email list if there's a convenient function to do this built-in. The authors are pretty responsive on that list.

[sindrle]

Thank you, its my concerns as well.. But Im on a time limit and must present something very soon.

Ill try the mailing list!


Problem with edgeR is the calculation of FPKM and my problem reproducing fold changes etc from CPM/logCPM.

With Cuffdiff it is possible, FPKM calculations may be reproduced and it has a lot more easy accessible results in the output files.

[dpryan]

Ah, yeah, the fold change is also shrunken (I think, I'd have to double check), so again you won't normally get the same answer by hand. As an aside, your supervisor is being a bit unreasonable to request these metrics when you actually require a pretty high level of statistical understanding to know how they were created (and your supervisor lacks that understanding). Best of luck dealing with that!

[sindrle]

I know :S

But, that is why Im asking here, to get your expert advise how to best present a subfraction of Cuffdiff results in a publication, without doing lots of stupid **** (like Im doing now) :P