Constrained classification issue

[Alea4jacta6est]

Hi, I try to get predictions for my atlas using your notebook:

target_celltypes = ["B cell", "CD4 T cell", "DN T cell", "MAIT cell", "Myeloid cell", "NK cell", "Progenitor cell", "CD8 T cell", "gd T cell"]
ca.safelist_celltypes(target_celltypes)

And this annotation works for more than 48 hours:
adams = ca.annotate_dataset(adams, skip_preprocessing=True)

Is it normal? May the issue be related to the way I write down the cell types? Do you have an approach to estimate the time of running for the constrained classification?

[tony-kuo]

Safelisting will be slower due to circumventing the indexing of the hnswlib kNN implementation. With too few valid cell types, it will perform basically an exhaustive search. It will perform faster if you expand the list of celltypes, or even without restrictions first to see the unconstrained results before any filtering.

We are looking at alternative kNN implementations that can perhaps perfom better in this scenario.

[Alea4jacta6est]

Thanks for the answer! Can the problem also be only with naming in the list? I actually made a subset of cell names from constrained predictions and it worked much faster (2 hours instead of 2 days).

[DarioS]

I have only cancer patient-derived fibroblast cell cultures. Fast constrained classification would be desirable for one cell type.