Bug with get_all_available_embeddings

[huguesva]

Hello, thanks for the great repo!

I have this bug when trying to reproduce https://chanzuckerberg.github.io/cellxgene-census/notebooks/api_demo/census_embedding.html

Describe the bug

I have the bug :

No module named 'tiledb.vector_search'

To Reproduce

from cellxgene_census.experimental import get_all_available_embeddings

Environment

Name: cellxgene-census
Version: 1.16.1

Name: tiledb
Version: 0.32.0

[ebezzi]

Hey @huguesva ,

thanks for reporting the issue. This is a newly introduced dependency, so if you run pip install -U "cellxgene-census[experimental]" it will fix the issue.

That said, in a future version we will move the import so that the dependency can be optional.

[huguesva]

great ! thanks a lot.

Quick follow up: is there a way to select the scGPT embeddings in the cellxgene_census.get_obs function ?
get_anndata seems to have this option but not get_obs .

[ebezzi]

get_obs only returns the obs dataframe - are you looking for a way to retrieve the embeddings without loading the full anndata?

[huguesva]

yes exactly !

I am looking for the API to retrieve embeddings and labels (as numpy arrays or torch tensors for instance) given all the parameters that identify it : organism, measurement_name, tissue, emb_names ("scgpt" or "geneformer" or "scvi") ...

I am not sure what to use between get_obs, get_var, or get_anndata. get_anndata seems to have more parameters than the other two.

Thanks a lot for your help !

[ivirshup]

@huguesva, while get_anndata retrieves the whole anndata object, get_obs and get_var only retrieve the adata.obs and adata.var fields specifically.

[huguesva]

thanks for the answer.
To be more precise, I think what perturbs me from this tutorial : https://chanzuckerberg.github.io/cellxgene-census/notebooks/api_demo/census_access_maintained_embeddings.html
is that at the cell [1] of the notebook, get_anndata is used with a specified obs_embeddings = ["scvi", "geneformer"] whereas later when get_obs is used (cell 12) there is no obs_embeddingsspecified so I have the impression that we don't know if the data is from "scvi" or "geneformer" or "scgpt" when using get_obs.

Thanks for the help.