Hi everyone! I’m running the nf-core/scrnaseq pipeline (version 4.1.0) and I've noticed an weird behavior regarding the CellBender outputs. I performed a comparison between the matrices and found that: - The .h5ad files in the mtx_conversions folder (specifically _cellbender_filter_matrix.h5ad) are identical to the raw_matrix.h5ad when subsetted for the same barcodes. It seems they contain the original raw counts. - However, the _filtered.h5 files located in the sample-specific CellBender output folder (cellbender_removebackground/) show different values (as expected, since they should be denoised). Has anyone else encountered this? Is there a specific parameter to ensure the mtx_conversions step uses the corrected denoised counts from CellBender rather than just the filtered barcode list with raw counts?

Hey, when we implemented this, the goal was only to replace the emptydrops empty droplet detection with cellbender. So that's what you get here - only empty droplet detection, no denoising. But I agree it would be low-handing fruit to include the denoised counts also in the output, maybe as an additional layer. This is a bit related to this issue in #C077PUX8DBP, which I was planning to work on soon anyway, so I think I can fix both in one go

Ok, now I see. I assumed both results from CellBender were already included in the .X of the _filter___matrix.h5ad. I agree, it would be great to have the denoised counts as a separate layer In the meantime, can you confirm that the _filtered.h5 matrix within the sample folder is the final output that I can use as denoised counts?