WGCNA network (unsigned or signed or signed hybrid) and TOMtype (unsigned or signed)

Question

[liyong]

Hello All,

How to choose the right parameters for networkType and TOMType when using WGCNA's blockwiseModules ?

net = blockwiseModules(datExpr, power = 12, networkType = "signed hybrid", TOMType = "signed")

For networkType, it offers unsigned or signed or signed hybrid. For TOMType, it also has unsigned or signed. I am still very confused after reading through the manual, does anyone have an easier way to understand those?

Thanks a lot.

Answer 1

[LChart]

networkType controls the adjacency matrix with:

signed=[(1 + cor)/2]^pow
unsigned=|cor|^pow
signed hybrid = ifelse(cor < 0, 0, (1 + cor)/2)^pow

the TOMType determines whether the signs of the correlation are used to 'reinforce' or 'cancel out' edge projects of the topological overlap (see https://horvath.genetics.ucla.edu/html/CoexpressionNetwork/Rpackages/WGCNA/TechnicalReports/signedTOM.pdf). Note that for signed and signed-hybrid networks, the adjacency for negative correlations are 0 or close to 0, so the impact of using a signed topological overlap is negligible.

[liyong]

Thank you for the explanation.

So, in that case, the "signed hybrid" network will only keep the positively correlated ones, and neglect the negative ones?

The shared file is really informative. Thanks a lot.

[LChart]

Yes. But the signed network pretty much does this too; since cor=0 -> (1 + cor)/2 = 0.5

for a power of 10 the resulting adjacency is around 0.001 which is not all that much different from the hybrid value of 0.

[liyong]

Indeed. I have RNAseq data from 140 samples as well as some trait data (e.g., plant height, flowering time...). What I want to do is to identify which modules is highly related to which traits through WGCNA.

So practically speaking, which types of network and TOMtype should I choose will make more sense biologically?

Sorry to bug you with further questions.

[LChart]

For so many samples, I'd recommend using signed for both network and topological overlap. It's simple and will result in more clusters than unsigned.

[liyong]

That's very good to know.

As the WGCNA authors suggest remove "non-varying genes, which usually represent noise". I only pick the top 15,000 variable genes across the samples to run WGCNA, is it sufficient?

I am pretty new to this, thank you so much for your Q&As!

[LChart]

I only pick the top 15,000 variable genes across the samples to run WGCNA, is it sufficient?

Absolutely.

[liyong]

That's awesome. Thank you. Really appreciate it!