Should I use raw or log normalized expression values for the input?

[wykpenguin]

Hi,

It wasn't very clear if raw, normalized, log transformed or scaled inputs were required. Does it matter and which would be the best?

Many thanks

[martinjzhang]

use raw count input please and leave flag_raw_count to be on.

[wykpenguin]

Thanks for the quick reply. If I combine data from more than one sequencing run, would it be better to normalize within each data set to equalize individual differences before running scDRS on the combined data with --flag_raw_count set to False?

On the flip side, you mentioned that a large data set could be split into 3 and the results concatenated. Does that imply scDRS results on my cell types of interest only with say a whole brain data set is still valid?

[martinjzhang]

Sorry for having missed this thread.

Thanks for the quick reply. If I combine data from more than one sequencing run, would it be better to normalize within each data set to equalize individual differences before running scDRS on the combined data with --flag_raw_count set to False?

It is fine to just concatenate the data. scDRS results of a given cell depends on other cells in the dataset only through the mean and var of the genes.

On the flip side, you mentioned that a large data set could be split into 3 and the results concatenated. Does that imply scDRS results on my cell types of interest only with say a whole brain data set is still valid?

No. The mean and var of the genes would have changed between restricting to only one cell type and using the full data.