Large proportion of "NM" cells

[Laolga]

Hi! Thanks for your tool.
What is a typical number of not annotated cells that should be produced by Nico? I observe almost 30% (187615 NM cells) after integration with a cell atlas. Sounds like quite a lot or rather normal for not matched single cell data?

[Laolga]

And a follow up suggestion, I have noticed that some NMs are produced because the tool can not confidently pick between two cell types, but sometimes it's a cluster that is really far away from anything else. I would be comfortable to use resolved_tie_issue_with_weighted_nearest_neighbor='Yes' in the first case, but not in the second. Hope that makes sense.

[ankitbioinfo]

Thanks, @Laolga, for your question.
In my experience so far, I’ve primarily worked with homeostatic tissues, where the cell types identified by scRNA-seq closely match those found in spatial transcriptomics data. In these cases, the proportion of NM (non-matching) cells has typically remained under 12%.
However, in non-homeostatic tissues (e.g., inflamed or diseased conditions), it’s quite possible that not all cell states are captured in the scRNA-seq reference, while they are present in the spatial data — which could lead to a higher percentage of NM cells. As you mentioned, when clusters are clearly far away, this is often a sign that new or missing cell states may be involved. In this case spatial cells anchors could not be mapped to reference data.
If you're confident that your reference dataset captures all major cell states and wants less confident spatial cells also to annotate, I recommend setting "Yes" to resolve the remaining NM population. But if you suspect that some cell states are missing in the reference and want to retain only high-confidence cell type assignments, then "No" is the safer option.
Hope this helps clarify!