[Bug] >99% cells annotated as NaN

[pakiessling]

Hi Quentin,

I just tried Novae out on my heart cohort. 46 samples ~ 2 million cells.
I used novae-human-0 with both zero_shot and fine_tune .

I used counts and let Novae handle the preprocessing

Both times almost all cells were annotated as NaN:

adata.obs["novae_domains_7"].value_counts(normalize=False,dropna=False)
novae_domains_7
NaN      1997204
D1015      12403
D1016       1755
D1012        827
D966         266
D1008        164
D993         102
D282           2

Do you know what could cause this?

Novae 0.1.0

[quentinblampey]

Hi @pakiessling, sorry for the delay, I just found out that I don't receive notifications from new issues on Novae, by bad...

NaN values are given for cells that are not connected to any cells in the Delaunay graph (or too few, e.g., one). So, my guess is that your Delaunay graph was not computed properly. You can refer to this function to understand how we built the graph (read the "notes"), maybe this can help you.

Did you get a warning saying that your cells have very few neighbors? Novae should detect if something is going wrong with the graph and warn you about it.

Did you already have the key spatial_distances in obsp before using Novae? If yes, to what do these distances correspond? Otherwise, are you spatial coordinates in microns?

[quentinblampey]

Hello @pakiessling,

I released a new version (novae==0.2.0) which changes a few things, I let you refer to the new tutorial.
The main difference concerns the computation of the graph of cells. I now force the user to run novae.utils.spatial_neighbors itself (it was automatic before), so that you can check if the graph looks good before running Novae. An example of a graph is shown here.

Let me know if that helps!