This is a test version of Biostars. For the public version, visit https://www.biostars.org.
If I have a single very large module in my WGCNA (possibly due to noise from low sample size), are other modules still potentially useable?

Hey there!

I am having some difficulties with a dataset I am currently working on. There are only twelve samples, and I knew from some time in February that the low sample size could potentially pose a problem later, but I had read success stories with fewer than fifteen samples and a student who worked with the lab before me had been successful with fourteen. However, there is a module containing some 65% of the total genes, and adjusting the cut height does little to break it apart. I still have some strategies including using resampling to verify module stability, alternate network construction methods, and perhaps adjusting the deepSplit parameter. I am preparing myself for the possibility that this colossal module is stable and the data may not be useful.

So, I understand that I may have to eventually concede that the data is simply inherently noisy. However, I am wondering if, in the event some modules are associated with traits in a statistically significant manner (even after false discovery rate p-value correction), that information might still be usable or whether those associations are likely to also be spurious.

Basically, does it all need to be thrown out if I come to suspect noise might be to blame? How can I verify whether noise is to blame (at the very least so I can demonstrate it in a quantifiable way to my supervisor)? I was essentially planning on using the bootstrapping to verify the stability of this behemoth and attribute its existence to noise if it turns out to be stable, but that isn't quite as direct a demonstration of cause as I would like.

Any additional information, corrections or suggestions are most welcome.

wgcna
• 3,804 views • link
updated by andres.firrincieli andres.firrincieli • written by Jacob • Jacob

Small modules can still be informative, even if most of the genes in the dataset cluster in a single module. Small modules can still capture small set of genes that significantly change between two or more group of samples.

To understand what is going in each cluster why don't you plot the module heatmaps and check how gene expression levels change across all your samples/conditions. As a first approach this should be good enough to understand if the clustering make sense or not (noise)

Would you mind to post the heatmap of the largest module in your network?

• 0 views • link
written by andres.firrincieli andres.firrincieli

I have two heatmaps produced at the moment. The first is the standard one portraying the relationship between modules and traits, which in this case is the production of fatty acids of interest. The second is simply the relationship between eigengenes and each experimental treatment.

The large module is the black one. I should mention that I have not adjusted the p-values yet in the module-trait heatmap and I suppose that may be significant given the large number of traits we are interested in (though you would probably be hard pressed to even read them).

enter image description here

enter image description here

It is difficult for me to make good sense of the heatmaps alone at the moment as I have little knowledge of how much we would expect each of the fatty acids of interest to be expressed at different stages of plant development. It might make more sense talking through it with my lab supervisor, but if you can glean anything from the maps at a glance you'd be helping me immensely.

• 0 views • link
written by Jacob • Jacob

Use this chunck of code to print a gene expression heatmap for each module:

library(gplots)
col = colorpanel(300, 'purple','black','yellow')
colorsA1 = names(table(moduleColors))
pdf("Heatmap.pdf",height=9,width=9)
for (c in 1:length(colorsA1)){
      moduleColors == colorsA1[c]
      heatmap.2(t(datExpr[moduleColors==colorsA1[c]]), scale = "row", col=col, density.info ="none", trace="none", cexCol=0.5, cexRow=0.8, margin=c(19,11), main = colorsA1[c], Colv = FALSE)
      }; dev.off()

There is a clear pattern in the module-trait relationship heatmap. Many modules, including the black, shows significant correlations against the same species of FA, DAG and TAG. Probably most of the modules in the network are sub-cluster of the black module. On the other hand, for other moduels, like the green, the correlation pattern is the opposite.

Finally, by looking at the second heatmap seems clear that most of the gene in the black module (largest module in the network) are expressed at low level in T4 samples.

• 0 views • link
written by andres.firrincieli andres.firrincieli

I appreciate all this quite a bit, I've been a little stumped for the last few days.

Regarding your comment that most of the modules are likely sub-clusters of the black module, do you feel it would be sensible to continue to adjust parameters to try to break the black into smaller modules if this is the case? I do see your point in that there is a clear pattern of expression within the black module, but many of the genes in the pathway(s) of interest are also located in this big module, which makes it difficult to study their relationships in particular. Before the merging of similar modules, they were much smaller (though there were about forty of them), so I know it is at least possible. I suppose the question is whether it would make a meaningful difference to do so, or if there is an upper limit on the number of modules one should be willing to accept.

• 0 views • link
written by Jacob • Jacob

There is no need to break down the black into smaller modules.

many of the genes in the pathway(s) of interest are also located in this big module, which makes it difficult to study their relationships in particular.

The black includes 65% of the genes in the dataset, so the fact the many of the genes of interest are located in this module is not surprising. With such big modules is always difficult to study the direct relationship between genes of interest because everything is connected. You should check if the pathway of interest is still preserved in the hub (genes with the highest connectivity) of the black module.

I suppose the question is whether it would make a meaningful difference to do so, or if there is an upper limit on the number of modules one should be willing to accept.

There is no upper limit on the number of modules you should get from WGCNA.

• 0 views • link
written by andres.firrincieli andres.firrincieli

0 answers

No answers yet.

Log in to answer this question.

More posts like this