In metagenomics one typically collects data representing gene/species counts (or their proportion) in an individual. One then performs comparisons for every gene/species in groups of individuals affected by different conditions, with the aim to determine which of the genes/species are characteristic of a specific condition. This necessitates correction for multiple comparisons / false discovery rate.
One of the complicating issues here is that with increasing the depth of metagenomic sequencing one is bound to discover more and more species, increasing the number of comparisons, which renders the correction for multiple testing somewhat arbitrary. Alternatively, a common practice is to filter the species with low prevalence, thus reducing the number of comparisons.
I have recently found works that forgo the correction for multiple comparisons altogether - instead simply selecting features passing certain threshold, and then ranking them in terms of importance (using some kind of regression or machine learning approach.) This struck me as particularly odd due to the use of the unadjusted p-values for the initial feature selection, although I suppose this contradiction could be bypassed using Bayesian approach.
I would like to get a clearer perspective on this issue: both frequentist and Bayesian: when/whether/how to correct for multiple comparisons, when the number of features is variable or could be subject to prior selection/filtering.
(Worth mentioning that other complication is the compositionality of the data, due to the finite sequencing depth and/or working with relative abundances. Perhaps, independent tests are simply not applicable here... but they are likely to persist, since the number of features is often larger than the number of samples, and needs to be reduced.)
