Co-Expression Networks Of Rna-Seq

Question

[Diana]

Hi everyone!

I'm interested in making a co-expression network out of RNA-seq data that I have for 3 different samples (developmental stages) without replicates. Is there a minimum no. of samples needed to make informative co-expression networks? or is 3 good enough? Are there any good programs that anyone can recommend?

Thanks a lot!!!

[Istvan Albert]

What is your definition for co-expression network. What are the nodes in your network and what are the conditions for connecting two nodes?

[Diana]

The nodes would be genes and the condition for connecting the nodes would be if they have similar expression profiles across the 3 developmental stages. If they are similar they should form a module and it could be speculated that these are co-regulated.

Answer 1

[Devon Ryan]

It sounds like WGCNA is what you're looking for. I don't know of any good a priori minimum sample number, though I would worry that 3, particularly with no replicates, might be too few to be of use. I would think that the biological question you want to ask is, "Can the change in gene coexpression over the surveyed developmental timepoints be meaningfully modeled by recruitment/suppression of a limited number of gene modules, each presumably driven by different promoters or some other coordinated event?" You really need replicates to answer most of the more interesting questions in biology.

[Sudeep]

For WGCNA the minimum number of samples per phenotype is something like 12 see this post WGCNA Network Construction Issues & Best practices?.

[Adamc]

I don't think an absolute minimum number of samples is cited by the authors of the package as the number of samples you need likely depends on your particular dataset.

Answer 2

[vj]

In my opinion the more the samples the better the network. You could try getting the correlation coefficient between the genes based on their expression across conditions and construct a network based on a cut-off (say 0.9). Other methods for eg., information theoretical approaches seem to work well with larger sample size.

Answer 3

[Federico Giorgi]

For simple correlation based networks you need at least 8-10 samples to be able to build some sort of network.

So far the smallest networks accepted in literature using RNASeq data are those from Iancu (Mouse, 21 samples) and Giorgi (Arabidopsis, 65 samples)