I am currently working with a dataset from a large microarray experiment where one of the aims is to look for evidence of the predicted doubling of variance in expression on the X chromosome in hemizygous males relative to females due to dosage compensation.
Speaking broadly, the experiment uses a controlled breeding scheme in Drosophila to generate individuals that are heterozygous or hemizygous for different known X chromosomes (the autosomal genotype is also controlled). I then want to use a genewise linear modelling approach to estimate additive genetic variance for gene expression for each gene in males and females. The arrays involved are Affymetrix GeneST arrays.
My initial inclination was to preprocess the array data using RMA, in my case from the oligo package in R. A fundamental part of RMA preprocessing implementations is a log2 transformation, and of course this has desirable properties for most gene expression analyses. However, it is apparent that the variance-stabilizing properties of the transformation have some undesirable consequences when looking for the sex differences in expression variance across samples that I am interested in. A toy example showed that variance in log2 expression is the same whether or not males show dosage compensation, obscuring the very signal I am interested in.
I would like to retain the benefits of the preprocessing steps (e.g., background correction), but I wonder if anyone can weigh in on the best way to avoid the log2 transformation’s effect on estimates of variance. Is it appropriate to simply take the antilog of the expression matrix after RMA preprocessing or is there another approach/package I should look into instead?