WGCNA with "interlinked" experimental conditions as traits

Question

[DBrauer]

Hey all,

I've got LC-MS protein data for a radiation experiment with the following conditions/set-up:

• 2 treatment types (X-ray, neutrino)
• 4 intensities (0, 0.25, 0.5, 1 Gy)
• 3 time points after irradiation (0.5, 48, 168 h)
• 3 replicates per condition
• (CFU and DNA DSB data)

I'm using WGCNA with the data and while I'm getting results, I'm wondering whether my approach is correct. At the moment my trait file consists of a time column, an intensity column, and a treatment type column (binary) (the CFU and DNA DSB data are not relevant to the question).

At first I used to have a big conceptual problem with whether WGCNA would correctly correlate the data with the experimental conditions or rather whether the results would be meaningful, as the intensity trait completely ignores how much time has passed after the irradiation and the time trait completely ignores the intensity of irradiation (or if there was any irradiation). In short I doubted that you could treat the experimental conditions as independent traits. I then simply performed WGCNA on the whole data set and also on subsets (only 0 Gy, only 0.25 Gy, .., only 0.5 h, etc.) with the above-mentioned trait file. The results were not convincing and some subsets did not "work" properly (e.g. problems with the scale-free topology fit index, few samples in each subset, ...), rendering at least the subset approach unusable.

Therefore I'm back to wondering whether it's legitimate to treat "interlinked" experimental conditions seperately. Should I rather make binary traits for every condition (e.g. 0.5 Gy & 0.5 h & X-ray - yes/1 or no/0)? Are there any other options? Is WGCNA even the right approach?

Any help/insight is greatly appreciated, thanks in advance!

Answer 1

[Lluís R.]

You seem to have very low number of samples per conditions which makes this type of analysis unsuitable for you (if there are high differences by condition), It is recommended to use at least 20 samples for analysis which would mean mixing different conditions, thus making the groups of genes correlated not really related to the condition.

If you want to compare what happens in each condition I think you might be better doing a differential experiment (comparing the replicates between conditions).

[DBrauer]

It was my understanding that the whole sample-pool should be >30, and I've got 72 samples (2x4x3x3). >20 Samples per condition seems a bit over the top for me as it would require >480 samples (2x4x3x20).

Nonetheless your recommendation of simply doing differential analysis between replicates sounds plausible; maybe I try too hard to make it work with WGCNA.

[Lluís R.]

In some context it seems that it should be enough to find some pattern that is common between conditions. If you want to find soft relationships between genes you need to have highly similar samples and that depends on the effect size of the changes of each condition, but with so many different conditions it is normal to fail to find these relationships.

[DBrauer]

Can you recommend an R package for this use case? Would msmsTests suit my needs? I'm relatively new to the field and kind of overwhelmed by the range of available packages and the respective feedback towards their viability.

[Lluís R.]

I haven't work with LC-MS data or tools so I can't recommend a package for this kind of data, but that one you linked seems a good one. Search the literature and papers that cite it to see if it fits your purpose (after you have read the methodology of the tool).

However it is different to test differences between conditions (what msmsTest do) than to test what is common in a condition (like WGCNA does). You can do both but be aware of the differences between these two kind of analysis