How to deal with differently skewed biological data? I have a single-cell data set with around 40 variables per cell (protein expression, all variables are measured simultaneously). The expression distributions for the single channels look quite different. After log-transformation, some look gaussian distributed, others are skewed or even multimodal distributed (see histograms). How should I deal with this kind of data? 
What I want to do is basically to detect significant expression for the different channels for every cell. For the gaussian distributed expression, I can easily scale the data and use probability distribution function to infer p-values. However, this is not possible for channels with skewed data or data that follow a multimodal distribution.
Any suggestions on modeling approaches, data normalization/transformation that might work with such data?

 A: If I had seen protein expression data with multimodality in the log-transform, I would be very, very careful. Log-scaling is not just for convenience, it has a lot in common with the concept of concentrations. A multimodal pattern would make me suspect there's a bunch of heterogenous subpopulations in the sample. 
For example, you could be seeing two types of cells with different baseline expression patterns, or different propensity towards some treatment (e.g. drug uptake through the membrane), or different response to some treatment (e.g. the number of nuclear receptors for some drug => more translation as a result of the treatment).
You say your dataset is single-cell data, but the same principle holds. Might be different organelles, I suppose? Also, you're using marker measurements, something might be dimming the signal from the marker in some fraction of the protein...
Basically, some hidden variable is separating the two peaks you see on the histogram, otherwise they'd be a nice log-Gaussian. Whatever it is, this is a huge potential landmine. You could explain it away (probably using some hierarchical approach) or ignore it (and just take the mean of the two), but both impose a personal interpretation on what you see in the data.
One particular tricky case would be that you already have a high-expression subpopulation, and any treatment would just shift the measurements towards the right peak, and if you had simply averaged them together, you'd only see a slight increase in the skew.
