I have recently acquired some 16S metagenomics data, and was wondering if anyone can speak of the potential limitations, challenges as well as advantages to conducting a network-based study on metagenomics, such as what is done w.r.t co-occurrence networks.

I thought it would be interesting to deviate from the "classic" study of taxonomic abundancies present in different timepoints / conditions. However, it seems as though for such studies they are usually dealing with a relatively large number of samples - please find paper below:


Could such analyses be done with much smaller sample sizes such as those generated via mouse models? What are the potential limitations of such an approach, and how could they be overcome?


  • $\begingroup$ Do you mean covariance, ie this is the measurement of co-occurance? $\endgroup$
    – M__
    Commented Feb 11, 2022 at 15:29
  • $\begingroup$ No. There are many possible ways to generate the networks. Based on my understanding, they are generated primarily using correlation metrics, in order to capture 'pairwise affinities', between different taxonomic classifications - i.e. bacteria i, and bacteria j. $\endgroup$
    – h3ab74
    Commented Feb 11, 2022 at 15:57
  • $\begingroup$ I do agree, its just 'co-occurance' isn't a statistical method. Pearson's or Spearman's correlations for example? Pairwise is a statistical method but not a metric and might even suggest distance method like P-distances :-/. Without knowing the metric you run the risk that the assumptions for one, may not be applicable. At present its not clear whether the analysis is applied directly to sequence data - which would likely be a mistake - or their taxonomic designation. $\endgroup$
    – M__
    Commented Feb 11, 2022 at 16:48
  • $\begingroup$ In one of the studies I am currently reading, they use a 'treshold-Kendall correlation'. My understanding is that this is to be applied to processed sequencing data, after taxonomic classification, rather than simply the sequences. $\endgroup$
    – h3ab74
    Commented Feb 11, 2022 at 19:01
  • $\begingroup$ Okay, that makes sense. Would be good to update your question. Its Spearman type stuff which is part of SciPy stats. I dunno enough about R, but there are some very capable R-coders here. Really this is population dyanamics of a microbial community, lactobaccilus (if its gut) and all that. Its definitely not my thing and I wouldn't approach the question in this way. $\endgroup$
    – M__
    Commented Feb 11, 2022 at 19:32

1 Answer 1


To answer your question more formally. I strongly believe you are treating the mice with something that will shift its microbial flora. Clearly the experiment is not described.

Thus what I simply feel is you essential have a discrete time series comprising before, during and after treatment. 'Treatment' could be anything drug, vaccine, antibiotic, diet who knows.

I think that is a very different time-series to a one where classic correlation statistics will be informative, I'm just guessing in the absence of a hypothesis, thus I could be wrong.

I also kinda suspect lactobacillus are strong players.

Anyway the stat you are looking at is a special case of the general correlation coefficient in much the same way Spearman's classic statistic is. What I can say is that the statistic is not a specialist approach for large sample sizes, which I believe is your question.

'New fangled cool', might be an approach of data mining and that might be okay if you've loads of data you don't know how to analyse, like it goes this way, that way. Ultimately, statistics are used to assess an exact hypothesis. If e.g. lacks power (very common) or the assumptions are violated then look deeper. I do agree that in population biology generating a hypothesis is often helped by data mining, but what I think is you already have an exact hypothesis and I'm far from sure this would nail it.


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.