I would like some advice on how to properly identify significantly abundant features within my groups.
I have this experimental design:
Group A: T0, T1, T2: Control diet
Group B: T0: Control diet; T1: Diet 1; and T2: Diet 2
I am not finding anything clear with volatility results (since diets are changing). I do not think I can use LefSe (since it performs unpaired KW analysis), I do not think I can use ANCOM because I cannot see how to add the longitudinal effect.
So, meanwhile, I just performed in R a pairwise.wilcox.test. I separated my data per time and performed three unpaired test (one per time). Moreover, I separated the two groups and assessed using a paired and in a pairwise manner the longitudinal differences within a group. Do you think that is a good approach? Has anyone another suggestion?
All of this is correct, and might I add you have a bit of a complex experimental design! So longitudinally-aware methods that make linear assumptions may be inappropriate.
The ANCOM test currently implemented in QIIME 2 performs a one-way test and is not longitudinally aware, though we’ve been planning for a while to upgrade this. You could run ANCOM directly in R, as the ANCOM 2.0 implementation supports the use of LME for longitudinal designs.
However, this (and the other differential abundance methods available in R and elsewhere) may still be hamstrung by your complex experimental design. I think we’ve already discussed this in regards to your use of LME vs. pairwise tests in q2-longitudinal, that the use of 3 different diets in Group B make for a complicated scenario. So for these same reasons, you may want to do the same for differential abundance testing: e.g., filter samples by timepoint and compare groups at those individual timepoints. E.g., you could use q2-composition (ANCOM) or q2-aldex2 or q2-sample-classifier to compare groups A vs. B at T0, T1, and T2 — hopefully you will see no or minor differences at T0 (you could see differences! e.g., if you have cage effects if these are animal subjects, or other population covariates if these are human subjects), and then you will see more substantial differences at T1 or T2. If you see major differences at T0 (due to covariates), or minor differences at T1 or T2 (due to high population heterogeneity in human subjects) then using a paired test would be appropriate, in which case I’d recommend running ANCOM directly in R, since it gives access to a number of different tests for hypothesis testing.
Alas, this is not an appropriate test. Wilcoxon tests assume that observations (e.g., individual taxa) are independent, which they are not in this case since you have compositional data (i.e., you are comparing the frequencies at which each ASV or taxon is observed, but the frequencies in this case are really relative frequencies since they do not relate to absolute abundance, and hence increasing abundance of one ASV/taxon displaces the observed frequency of all other ASVs/taxa, and hence these measurements are not independent). Note, Wilcoxon tests are fine for other microbiome measurements that are independent, e.g., for comparing alpha diversity and some measurements of beta diversity, but will be prone to a very high false-positive detection rate if you are using this test on ASV/taxon frequencies.