Draft PR: Add modularity and modularity_adata functions to scanpy.metrics

[amalia-k510]

This adds two functions to compute modularity scores from a given graph and a clustering like Leiden or Louvain. The goal is to make it easier to compare different community detection methods using an external metric. This follows up on issue #2908. To my knowledge, there is no built-in way to compare clustering results nor ways to calculate modularity score.

Functions:

• modularity(): accepts a connectivity matrix and label array, generates an igraph, and returns the modularity score
• modularity_adata(): AnnData wrapper that pulls graph and clustering labels

[flying-sheep]

Hi! Apart from the issue with igraph being an optional dependency, this looks good!

I think we have get_igraph_from_adjacency which might be useful, but maybe not.

Could you please add tests? We have many examples on how, best would probably be

1. for the direct variant, manually create very small graphs to run this on so you can be sure the results are correct
2. for the anndata version, use neighbors to create the connectivity matrix.

Please add @needs.igraph so the test only runs when igraph is installed

if you’re unsure about anything, please search the code for examples or ask me!

If you end up implementing a non-igraph flavor for this, please test using parametrization, e.g.: @pytest.mark.parametrize("directed", [True, False], ids=["directed", "undirected"])

[flying-sheep]

Hi, this is shaping up! The tests are looking particularly great!

Please address these previous comments:

1. Please follow the doc style as described here: Draft PR: Add modularity and modularity_adata functions to scanpy.metrics #3613 (comment)

Also I mentioned in there “But here a is_directed: bool parameter would be better anyway, there will never be more than two options.”

What do you think?

2. Don’t densify (in modularity_adata or anywhere): Draft PR: Add modularity and modularity_adata functions to scanpy.metrics #3613 (comment)
3. Why the np.asarray? Draft PR: Add modularity and modularity_adata functions to scanpy.metrics #3613 (comment)

[flying-sheep]

Oh, I just saw that we have get_igraph_from_adjacency, can you use that?

[flying-sheep]

Looks good! A few points:

• please add a release note (see the other PR, please adapt the PR number in the command I mention there when running it for this PR)
• please move the correct change to unique_labels (to_numpy) from the other PR into this one:

  scanpy/src/scanpy/metrics/_metrics.py

  Line 63 in 040b8b7

  unique_labels = pd.unique(np.concatenate((orig.to_numpy(), new.to_numpy())))

• please change mode to is_directed: bool (no default value) for modularity

Once your other PR (the neighbors one) is merged, let’s finish this one up:

• we should make neighbors store an is_directed param in .uns[key_added or "neighbors"]["params"]
• then we change the modularity_adata parameter obsp: str parameter to key: str = "neighbors"
• then we use adata.uns[key]["connectivities_key"] and adata.uns[key]["params"]["is_directed"] in modularity_adata to call modularity.