We are soon going to be analysing the gut microbial communities of 80 pregnant women (Subjects). Each of those mothers gave 3 samples, one in each of the 3 trimesters of the pregnancy. ~30 of those mothers have T1D, while the others are healthy. I have been reading about performing time series differential abundance analysis using DESeq2. I would like to answer 2 main questions:

1.- Are there differentially abundant taxa (OTUs) in the different trimesters (time: T) in general (not taking into account if the mother has T1D)? Also specifically between T1 and T2 and between T2 and T3?

2.-  Are there differentially abundant taxa between mothers with and without T1D over the three trimesters?

FACTORS:

Subjects = Different mothers (IDs are 1 to 80)
T1D = Has T1D or not (T1D and non-T1D)
Time: The different trimesters (T1, T2 and T3)

QUESTIONS:

1.- For number 1 I think I should do a simple design in which I only take into account the factor time, which is the factor that I want to test for changes. —>  ~ Time

    Question: Should I control for the factor T1D (diseased?), with a model like —> ~ T1D + Time ? Using the default parameter test= “WaldTest” ?

    For answering if there are differences between any pair of trimesters (e.g. T1 vs T2) I could actually use contrasts, right?

2.- Here I think it would be more complicated and I think that I could actually apply something similar to what you (DESeq2 team) wrote in: http://www.bioconductor.org/help/workflows/rnaseqGene/#time-course-experiments
    
    Question: Should the design be something like —> ~ Time + T1D +Time:T1D with a reduced design: ~ Time + T1D ? Using the parameter test=“LTR” ?

     I could actually also just do a simple model in which I don’t consider the Time and just test if there are differentially abundant taxa in mothers with T1D compared with non-T1D:

    Question: If I do this, should I control for the differences between Subjects? —> ~ Subjects + T1D (test=“WaldTest”)  I guess that in this case what would potentially be wrong is that I would be ignoring differences that are caused by the factor Time, right?

Thanks in advance and hopefully you have the time to answer this questions

Cheers

Alex

For microbiome data, I can't say if DESeq2 is the best software. The methods for standard RNA-seq DE should work to find differences in mean, but I don't keep up in this literature and don't analyze microbiome data myself. So that's a caveat.

There is a problem which comes up often with users, in that you want to control for different mothers and also compare across groups of mothers. You cannot make such a comparison using a "fixed effects model" because the terms for the individual mothers are linearly dependent with the term for the difference across T1D. You can read about this in the DESeq2 vignette in the section on "Model matrix not full rank". There is a case where you *can* make comparisons, if you are interested in contrasting, for example, the difference between times across T1D and not T1D. This is discussed in the vignette in that section.

For your first goal, I would use a design that takes into account subject and time, that is ~subject + time. You can then use test="LRT" when you call DESeq() with a reduced design of ~subject to find genes which are different at any timepoint, and you can test individual effects by specifying a contrast when you call results(), e.g. contrast=c("time","T3","T1") and test="Wald".

For comparing across T1D and not, you cannot compare in general, but you could look for differences in the difference across time point. Check the vignette section describing how to do this, and then you can add a followup post if you have more questions.