differential gene expression - Illumina

[nicole_01]

Hi all,

my apologies if this question has already been asked (I couldn't find it anywhere in either of the forums) and also that this one is a question that probably could also fit into the RNAseq forum.

Last time I did differential gene expression work I used custom microarrays, but I've now been asked to switch to RNAseq for it and as I'm still fairly new to it, I'm a bit unsure how to tackle it.

What I want to do is: find differentially expressed genes between different treatments. The species I'm using is a novel one (molluscan), but I already have a reference transcript (I made it with strand-specific, normalised data of multiple individuals that were control and treated, so it should have a wide range of genes in it).

The questions I have are (and I think it links to some degree into the bioinformatics bit, so that's why I posted it here instead of the RNAseq forum):
- should I use single reads or paired end reads
- should I use strand-specific data
- would 3Gb of data be enough or too much per sample to find meaningful results (i.e. how many reads should I have per sample for differential gene expression)
- is there a good paper that one could recommend

Sorry for all the questions. I've tried to find info online, but there is so much information and it's starting to become really confusing and also difficult to weed out papers that might not be worth following.

Thank you so much for your help.

Nicole

[gringer]

I recommend doing a pilot study with as much information as possible, which will give you data that you can remove information from to simulate your other conditions:

* paired-end reads -- can simulate single-end reads by ignoring the linking and/or one read end
* strand-specific data -- can simulate unstranded sequencing by ignoring alignment direction
* rRNA depletion / sample enrichment [not mentioned in your post] -- would need to do at least two tests (one with, one without) to compare bias / rRNA contamination, because this tends to be organism/kit specific
* spike-in RNA controls [not mentioned in your post] -- can simulate data without the controls by ignoring spike-in read counts
* use only a few samples -- can simulate high-numbers of samples / multiplexing by randomly removing reads [to check e.g. if 3Gb is enough per sample]

edit: Most of the time you have a fixed cost, and need to choose between spending more money per sample to get more detail and analysing more samples to get more variance and statistical robustness