hvg flavors seurat and cellranger with batch: bug in subset

[eroell]

• Closes Unexplainable sc.pp.highly_variable_genes(subset = True) behavior #3027
• Tests included or not required because:

• Release notes not necessary because: they are, added

This PR fixes the bug reported in the linked issue.

A new test which spots the erroneous computations has been added.

I would use this chance to refactor the _highly_variable_genes.py, rather than using the 2-lines fix suggested in the first commit:
Doing the multi-batch hvg flagging differently for seurat_v3 and seurat/cell_ranger is what made this bug hard to spot in the first place I think.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 76.31%. Comparing base (896e249) to head (13b1f6c).
Report is 50 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #3042   +/-   ##
=======================================
  Coverage   76.31%   76.31%           
=======================================
  Files         109      109           
  Lines       12513    12515    +2     
=======================================
+ Hits         9549     9551    +2     
  Misses       2964     2964           

Files with missing lines	Coverage Δ
src/scanpy/preprocessing/_highly_variable_genes.py	95.23% <100.00%> (+0.03%)	⬆️

[flying-sheep]

Looks simple enough! Please deduplicate the tests though, they have too many identical lines.

[eroell]

Please deduplicate the tests though, they have too many identical lines.

To do so did across-setting tests with for loop on top of former test...

Would you prefer one separate test with that for loop for the across-settings check?

[flying-sheep]

Oh, sorry for missing this. I’m always for pytest.parametrize instead of for loops. That way it’s immediately visible which variants of the test fail.

[eroell]

To test in the future that this bug doesnt happen, the 4 combinations of inplace=True/False and subset=True/False need to be compared for their selected var_names - if these arguments were be in pytest.mark.parametrize, the different produced anndatas don't live at the same time and can't be compared...

Could

• extract this specific section into a new, dedicated smaller test

    # check that the results are consistent for subset True/False: inplace True
    adata_post_subset = adatas["subset_False_inplace_True"][
        :, adatas["subset_False_inplace_True"].var["highly_variable"]
    ]
    assert adata_post_subset.var_names.equals(
        adatas["subset_True_inplace_True"].var_names
    )

    # check that the results are consistent for subset True/False: inplace False
    df_post_subset = dfs["subset_False_inplace_False"][
        dfs["subset_False_inplace_False"]["highly_variable"]
    ]
    assert df_post_subset.index.equals(dfs["subset_True_inplace_False"].index)

    # check that the results are consistent for inplace True/False: subset True
    assert adatas["subset_True_inplace_True"].var_names.equals(
        dfs["subset_True_inplace_False"].index
    )

• hardcode the known-to-be-selected features and check against them

what's your preference here?

[flying-sheep]

Ah, silly me, this makes sense.

Since some entries of the dicts are never used, I just removed them and replaced the string with a bool (subset=True/False). This makes it all quite a bit more compact.

Also please remember itertools: If we’re not writing numba code, it’s always preferable to use it as opposed to nesting for loops.

[flying-sheep]

please add a relnote

[eroell]

Timeout and Scrublet failing in Python3.12?

[eroell]

Ah on main too I see

[eroell]

coverage decreased, I think not detected because some pytest.parametrize were removed?
I can also add the new specific test separately do avoid this, else it looks good to me

[flying-sheep]

coverage decreased, I think not detected because some pytest.parametrize were removed?

I think codecov just hadn’t updated its comment yet when you saw that.

What you say can’t be, it doesn’t matter how a line was hit: If a line is run, it’ll be reported as hit, if your changes would have caused it to no longer be it, it would have been reported as a miss.