Best practices for filtering cells based on prediction scores after Label Transfer (scATAC → scRNA)

[cmj-dresen]

Hi everyone,

I’m using the LabelTransfer function to annotate a scATAC-seq dataset based on a matching scRNA-seq reference. The overall certainty of the transferred labels is quite high, with over 92% of cells having a prediction score > 0.5.

I’ve noticed that some people further filter cells based on these prediction scores. However, when I apply such filtering, it disproportionately affects smaller cell populations (which I’m not particularly interested in) and samples with lower total cell counts.

I’ve already performed stringent doublet filtering. However, since I unintentionally overloaded the chip, some of the low-confidence cells may be still doublets, even though they don’t show high doublet enrichment scores.

My question:
Is there a generally accepted best practice for filtering based on prediction scores after LabelTransfer?

Thanks in advance for any advice or references!