(feat): igraph leiden implementation now included as an option in sc.tl.leiden

[ilan-gold]

• Closes Leiden now included in python-igraph #1053
• Tests included or not required because:

• Release notes not necessary because:

[ilan-gold]

TODOs:

1. Figure out why some tests are passing when they shouldn't (hence why I pushed the branch, curious about CI). UPDATE: tol for matplotlib.testing.compare.compare_images is too high for a sparse-ish plot like rank_genes_groups. This is somewhat worrying so will need to be amended. Other than that, changed plotting outputs make sense so this should be resolved.
2. Check with scanpy tutorials to see what needs to be changed there as well, if anything (if needed, the two PRs should be merged in tandem). The following use leiden in some capacity:
   a. https://scanpy-tutorials.readthedocs.io/en/latest/pbmc3k.html
   b. https://scanpy-tutorials.readthedocs.io/en/latest/plotting/core.html
   c. https://scanpy-tutorials.readthedocs.io/en/latest/spatial/basic-analysis.html
   d. https://scanpy-tutorials.readthedocs.io/en/latest/spatial/integration-scanorama.html
3. Do a large dataset test - check NMI for accuracy of the new default against the old one, check speed to confirm what we're doing makes sense (although this was covered, it seems, in Leiden now included in python-igraph #1053), and scalability

[ilan-gold]

Failing violin_2 does not fail locally, so that's bizarre.

[codecov]

Codecov Report

Attention: 4 lines in your changes are missing coverage. Please review.

Comparison is base (1ac74a7) 74.57% compared to head (2549f61) 74.63%.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #2815      +/-   ##
==========================================
+ Coverage   74.57%   74.63%   +0.05%     
==========================================
  Files         115      115              
  Lines       12716    12756      +40     
==========================================
+ Hits         9483     9520      +37     
- Misses       3233     3236       +3     

Files	Coverage Δ
scanpy/_utils/__init__.py	68.20% <90.90%> (+1.00%)	⬆️
scanpy/tools/_leiden.py	88.40% <94.44%> (+3.29%)	⬆️

[ilan-gold]

I think both of these should be added to uns as outputs, weights in the params part of the dict and use_igraph at the top levels

adata['leiden']
# { 'params': { ... 'use_weights': True }, 'use_igraph': False }

[ivirshup]

There are a lot of branches here like if flavor == "leidenalg", I think it could be cleaner to just have one higher level branch for the two cases.

---

I think we can change the default for directed to be None then dynamically set the value based on which flavor is passed.

We could do the same with n_iterations, though I'm a little on the fence about this.

---

A number of plots are being changed that don't seem to be very different. Mostly things pre-clustering, like pca and qc metrics. I'm going to check if these are actually different enough to trigger test failures.

[ilan-gold]

@ivirshup You are right about some of the pre-processing plots. I should have created a separate branch for those. I simply think hte outputs have changed and we never caught it but would be curious to see what you get. It could be my M1

[ilan-gold]

@ivirshup I simplified the conditionals a bit and there are only two sets now. One to check for various {Value/Import}Errors and another to do the clustering_kwargs building. I think this is cleaner and faster since no code will run that doesn't have to. I didn't really see a way to do it with only one set of conditionals without code duplication. There's some code that's just common to both, but that shouldn't be run in the case of one of the {Value/Import}Errors .

[ivirshup]

About the preprocessing plots:

scanpy/tests/notebooks/_images_pbmc3k/filter_genes_dispersion/expected.png

• Text and some points are shifted slightly. I'm not totally sure whether any points are actually in a different place

scanpy/tests/notebooks/_images_pbmc3k/highest_expr_genes/expected.png
scanpy/tests/notebooks/_images_pbmc3k/pca/expected.png
scanpy/tests/notebooks/_images_pbmc3k/pca_variance_ratio/expected.png
scanpy/tests/notebooks/_images_pbmc3k/scatter_2/expected.png

• Axis text shifted slightly
• Can probably be reverted if the tests still pass

scanpy/tests/notebooks/_images_pbmc3k/scatter_1/expected.png

• y axis moved

[ilan-gold]

@ivirshup I will revert those, and hopefully tests pass

[ivirshup]

Looks good! But I don't think that you addressed these:

I think we can change the default for directed to be None then dynamically set the value based on which flavor is passed.

We could do the same with n_iterations, though I'm a little on the fence about this.

I think I would definitely change directed, and add a note in the docs that setting n_iterations=2 will be faster and is the default for the underlying packages.

[ilan-gold]

Ah, you're right. Will change.

[ivirshup]

Great, looks good!