I'm doing a WCGNA analysis (signed network) on microbiome 16S data. I have transformed counts to centeres log-ratio transformed data (CLR) to address the compositional characteristics of the data and have obtained a pretty decent cluster (around 7 modules, 20 taxa each).

I correlated the different modules eigengenes to my features of interest using both bicor and pearson correlation measures. Overall, the results I'm getting make sense but I'm having some questions regarding the method: The % of variance explained of each eigengen is around 30%. Is this value normal? Is it ok that I'm using only the PC1 of the module as a summary of the module itself when variances explained of the PC1 are so low?


  • $\begingroup$ Which bacterial groups, or perhaps rather what was the original sample for the meta genomics study? $\endgroup$
    – M__
    Sep 21, 2022 at 14:05

2 Answers 2


Part 1 Just to clarify on the specific issue of using PC1 alone given the reported result ...

Without knowing the underlying ecology, which is important, 30% is too low to any draw conclusions in PCA-based analysis, particularly because this could be a really complex niche, ecotope, ecoregion, etc ...

Thus this specific result from the unsupervised learning shouldn't be used as the basis for any downstream analysis. This is because the remaining 70% of the variance could overturn PC1, where PC1 could be separating a minor feature of the community ecology.

Thus, its informative because it states the first component is insufficient to proceed - however I stress its difficult to give a clear answer without knowing the expectedly complexity of the meta-genomic sample.

I suspect its a nice result, but in broad general terms its not a good idea.

Part 2

of each eigengen is around 30%.

If PC1-PC3 is 90%, then thats good and maybe its not so complex. Personally, I'd be delighted with that result. However, the inference is that the clustering is not clear. I would simply suggest further data mining is required.


Proportion of variance explained of about 30% is indeed on the low side, but nothing out of the ordinary. I have no experience with 16S data; in gene expression data, between 30 and 50% is actually considered good. PVE byt the eigengene depends strongly on the number of samples; the more samples, the lower the PVE will be.

You can try to calculate additional singular vectors and their PVE. Usually this is much lower than the eigengene (1st singular vector). So I would focus more on the biological significance - do the modules make biological sense?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.