Ability to index by label along one axis into obsp, varp arrays

[wrosko]

So I've run into an issue when slicing my anndata object to get a vector.

My anndata is structured like:

• X: obs x var
• obs: iterable of observation names
• var: iterable of variable names
• obsp: connectivity matrices defined by my different obsp keys and having a shape of obs x obs
• varp: connectivity matrices defined by my different varp keys and having a shape of var x var

I initiate the object with (note that all X's are sparse matrices):

adata = anndata.AnnData(adata_X, obs=pd.DataFrame(index=obs_keys))

adata.obsp['obsp_1'] = obsp_1_X
adata.obsp['obsp_2'] = obsp_2_X

adata.varp['varp_1'] = varp_1_X
adata.varp['varp_2'] = varp_2_X

I would like to slice with: adata[key].obsp[obsp_1_X] but this returns a 1x1 matrix whereas I would like a 1 x n_obs

To get around this I am having to do: adata.obsp['obsp_1_X'][adata.obs.index.get_loc('key')]

I assume similar behavior with .varm and .obsm

[falexwolf]

Maybe empower obs_vector to allow accessing obsp? That could be easy, @ivirshup?

Or adding a .loc accessor to the obsp attribute.

[ivirshup]

Maybe empower obs_vector to allow accessing obsp?

That's the plan! One of the last remaining things to do in #342 (@flying-sheep). IIRC we were going to leave the ability to select which axis of the pairwise array is being subset and default to rows for now.

Right now I do what @wrosko is doing:

adata.obsp['obsp_1_X'][adata.obs.index.get_loc('key')]

I do think it would be nice to make adata.obs.index.get_loc('key') more concise.

Or adding a .loc accessor to the obsp attribute.

This could be interesting. Right now the idea of a subset of one of the aligned mappings is pretty tightly coupled to a view of an AnnData object, so this could take some doing.