Normalization and log transformation of log count data. Pseudocounts, library size adjustment, Centered Log Ratios (CLR), Variance Stabilizing Transformation, and all that. There are many variations on ways to achieve this, but here's something I haven't seen done:
Transform counts to log2(count/scale+1), with a scale chosen per sample such that each sample adds to the same total.
I've put an implementation in my old varistran package. There is a numerical optimization per sample, but I can apply Newton's method so it's fast. I have more testing to do to really understand how well this performs, but wanted to have implementation I could point people too.