What's the best log-transformation for compositional data (microbiology) ?

Question

[pablo]

I have a matrix with gene abundances through different samples. I know I work with compositional data, and I need to normalize my dataset. I also know the clr transformation is used with compositional data in microbiology.

On this matrix, I will compute the correlations between my genes in order to contrust clusters. I would like to know if the alr or ilr transformation could be more suitable for my analysis of correlations? Or if I should stay with the clr transformation?

Bests

Answer 1

[Martin Modrák]

My best guess is that you will get quite similar results with either. A good practice would be to try both and any features that are not replicated are likely problematic anyway.

Answer 2

[Asaf]

Normalize and transform are two different things. For normalization I usually use the set of genes used by MUSiCC, I do the normalization on the read count level using DESeq2. As for transformation, I agree with @Martin, I don't think it will have much effect, you'll probably end-up with some parametric test so why bother. Just make sure you're not inflating or deflating zeros.