Should batch correction be performed before or after feature selection in scRNA-seq

Question

[Lucy]

Hi,

I am analysing some Smart-seq2 data. The data comes from multiple plates, so I want to remove this batch effect. I was planning to use ComBat for batch effect correction and M3Drop for feature selection. However after ComBat analysis, the matrix of counts contains negative values and zero values are lost, which are required for M3Drop.

Should batch effect correction be performed prior to or after feature selection? If so, can ComBat be combined with M3Drop?

Many thanks,
Lucy