wgcna network interpretation

Question

[grant.hovhannisyan]

Apart from that it looks crap, how would you interpret this WGCNA network?

Cheers,

[andres.firrincieli]

Most of your genes/modules populate the upper part of the gene dendogram. This make me think that your conditions have no or little effect on gene clustering. Did you look at the expression profiles of each module?

[grant.hovhannisyan]

thanks! This is a meta-analysis of around 700 samples, and they come from many different conditions and experimental set-ups, so I guess the network is messed up due to this. On the other hand, we are not interested in correlating any modules with any conditions, just interested in the network properties of some transcripts (connectivity, hubs or not, etc).

Answer 1

[i.sudbery]

If I had a particular question in mind I'd answer that question, and not worry about a general interpretation. So if you are interested in the network properties of some transcripts, then I'd calculate the network properties of those transcripts and ignore everything else. One thing you can tell from this is that the clustering hasn't worked very well. Which probably means the network construction hasn't worked very well, so you should treat everything you find with a massive pinch of salt.

But then that should always be the case with these sorts of analysis - either you have a specific hypothesis which you are testing. If which case you should know what you are going to look for before you start, or you are hypothesis generating, in which case you'll have to design further tests of any hypothesis you generate anyway.

[grant.hovhannisyan]

good points, agree with you in general. Though I was wondering if there could be any technical issues that result in this kind of network (I bet there are since its a very heterogenous data). If it would be possible to minimize them, then I'd just go and continue with overall hypothesis-generating analyses, as you mention.

[i.sudbery]

Yes. If you are combining many heterogeneous datasets, then a lot of the correlation structure is likely to be driven by the difference between the different experiments.You could start by plotting a PCA of all the data and checking that the samples don't cluster by study.