Experiment Design - power and sample size

[kindergarten]

Hi
We are going o colonize germ free (GF) mice with two microbiota - A and B. These microbiota differ in composition. We aim to find taxa exclusively present in A, but absent in B. We planned to use 5 mice for each group and perform V1-V3 sequencing on fecal samples (taxa we are looking for should be present in fecal samples). Analysis will be either heat map (showing presence absence based on absolute counts in each sample) and/or LefSe.
IACUC reviewer wants us to calculate the power and determine if n=5 is appropriate animal number.
What would be the appropriate sample (n) size for such study?
Is there software or online tool that I can use?

[jwdebelius (Justine Debelius)]

Hi @kindergarten,

I assume you're doing some other phenotypic characterization and you've already done a power calculation for that? Otherwise, you might consider whether you need the mice at all to look at replicates in a community.

From a microbiome perspective, you might check out the evident plugin. I dont know that it will work in your case since I think it addresses something else.

But, let's also look at the analyses you're proposing form a super basic perspective.

I want to start with a basic issue. First, if you're doing a kruskal wallis test, like LefSe, the minimum sample size you need to even make the test work is about 5 samples per group. If you test for a difference in 10 samples over 100 features (a very conservative number) with either kruskal wallis, checkerboard, or Poisson test, do you think you'll be able to get p-values small enough to overcome your FDR?

Second, how are you planning to handle well known issues with confounding? Rodents are pretty well known to be coprophagic, which means that your animals cluster by cage. If you have 5 animals per group, are you single housing them (rolling the cage effect into the animal effect but making your mice very sad)? Will you have a single cage (if so, how do you know A vs B is actually form the community and not just a cage effect)? Will you have multiple cages (can your modeling tolerate multiple cases and cage-related variation? How will that change your minimum p-values?)

I also want to mention Murphy's law of sequencing, which says that the harder samples are to get, the more likely they are to fail to amplify in the wetlab when you sequence them. I've been doing this for years and I have very few projects where we didn't lose one or two samples at some stage in processing. If you only collect 10 samples and 1 fails, what does that do to your modeling?

Best,
Justine

[kindergarten]

Thanks for taking time to respond.

Sorry, I did not explain the experiment well.

We are going to colonize germ free (GF) mice with microbiota A and B. These microbiota differ in composition. After 3 week colonization, mice will be gavaged with pathogen S. After 14 days (post pathogen exposure) we will enumerate (by plating) the pathogen in mice feces.
Based on preliminary data (omitting some details here, sorry), we believe pathogen will be absent (counts below detection) in mice A feces, but present in mice B feces. (Outcomes are theoretical. If outcome is same for both, we have a dead-end. However, all mice in the group will either have pathogen S or not have. It is all or none. No proportions. Our experiment in not measuring effect of microbiota on mice, rather which taxa in the microbiota are causing the effect). Mice are crucial to assess the function of microbiota.
We aim to find taxa exclusively present in A, but absent in B. We hypothesize that few taxa (1-5) in microbiota A eliminated (or prevented) the pathogen S. We aim to identify these taxa.
We will house one mouse per cage.
We planned to use 5 mice for each group and perform V1-V3 sequencing on fecal samples (taxa we are looking for should be present in fecal samples). Analysis will be either heat map (showing presence absence based on absolute counts in each sample) and/or LefSe.
I read somewhere that atleast 5 samples are required for LefSe analysis. That is why selected n=5.

[jwdebelius (Justine Debelius)]

Hi @kindergarten,

Thanks for the additional information.

I really don't think 5 animals per group will work. I want you to go back to this question:

@jwdebelius

If you test for a difference in 10 samples over 100 features (a very conservative number) with either kruskal wallis, checkerboard, or Poisson test, do you think you'll be able to get p-values small enough to overcome your FDR?

I don't think you will, based on results I've seen lots of places. But, if you wanted to test it, you could simulate it.

I see broader issues with the experimental design, and recommend you explore these more with a biostatistian or experimental design expert. If S is a bacteria and therefore detectable by 16S, you're already introducing a bias into your sequencing results, since microbiome data is compositional etc etc etc. FWIW, I would not be willing to analyze or supervise this project because I dont think your methodology will provide valid conclusions to answer the questions you're interested in.

Best,
Justine