Problems in conducting batch effect correction

Question

I am now trying to conduct batch effect correction for my data matrix, whose samples (human) have different gender and come from different sequencing platforms. I have tried Combat, Combat-seq, and limma. However, the outcomes are bad. It seems that Combat-seq is based on Negative binomial regression models, Combat is based on Gaussian distribution, and limma is based on the linear model. I wonder if my data violates their assumption, so I check my data and get this outcome. It seems that my data follows Gamma distribution. I am wondering what I could do in this situation.

---

The data I am working on is DNAse-seq data, which is a matrix with 16 samples. As the samples come from different platforms, I would like to correct the batch effect. I am also considering gender information as there are male and female samples. To sum up, my questions are:

1. Do we need to consider the distribution of the data when we apply the batch correction (limma, combat, and combat-seq)?
2. If we need to consider the distribution, which method do I need to correct the data which follows gamma distribution?
3. Are there methods designed for read counts? (I know there are many methods designed for surat objects, but do we have others that are more general?)

Thank you very much for all kinds of help.

Answer 1

The ComBat-seq method requires raw counts which can be modelled as a negative binomial. The other methods are typically used with normalized counts transformed to log2 scale. Obviously for any method that tries to regress a factor out of a dataset you have to make sure that none of the experimental variables are confounded with the batch variables that you want to regress. More details can be given if you edit the post to say what the data are, how the experimental setup is and what the factors to regress are.