I mean prior steps, like Fluorescence-activated cell sorting (FACS), but it doesn't have to be FACS. Machine learning-wise, I mean the procedure that was used for label generation (thus, prior cell population classification).

Many thanks for your helpful suggestion. I am sorry for the late answer. I checked their datasets: unfortunately, the only complete transcriptome dataset has a high mt content according to metadata: filtering for <10% mt genes over a dataset of 13k cells yields 34 cells, 15% 778 and 20% 5.7k. However, I think could use this data to test the effect of higher % mitochondrial genes (since the other datasets I am using are filtered using a 5% cutoff).