[kroy3]

When filter_rank_genes_groups recomputes fold changes or fractions (for example with a different use_raw, layer or groupby than rank_genes_groups), gene positions come from var_names.get_indexer, which returns -1 for names that are not present. x[:, -1] then silently scores those genes with the values of the last gene in adata, so genes that were removed can still pass the filter.

The original KeyError from #1487 no longer occurs on main, but the same situation now gives wrong results without any error.

Reproduction on current main

import scanpy as sc

adata = sc.datasets.pbmc68k_reduced().raw.to_adata()
sc.tl.rank_genes_groups(adata, "bulk_labels", method="wilcoxon", n_genes=20)
removed = list(adata.uns["rank_genes_groups"]["names"]["CD14+ Monocyte"][:3])  # FTL, AIF1, PSAP
sub = adata[:, ~adata.var_names.isin(removed)].copy()
order = [g for g in sub.var_names if g != "FCGR3A"] + ["FCGR3A"]  # strong marker last
sub = sub[:, order].copy()
sc.tl.filter_rank_genes_groups(
    sub, min_in_group_fraction=0.9, min_fold_change=1, max_out_group_fraction=0.3
)
sub.uns["rank_genes_groups_filtered"]["names"]["CD14+ Monocyte"][:3]
# -> FTL, AIF1, PSAP are kept although they are not in `sub`

Change

• Gene positions are looked up through a small helper, _get_var_idx, which raises a KeyError naming the missing genes, the uns key and the group, and suggests rerunning rank_genes_groups or passing the same use_raw / use.
• NaN names are skipped, so an already filtered result (key="rank_genes_groups_filtered") can still be filtered again.
• Only the recompute path is affected. When pts and logfoldchanges are reused from the stored results, behavior is unchanged.

I chose to raise rather than treat missing genes as filtered, since that would hide a likely mistake (for example ranking on .raw and filtering an HVG subset). Happy to switch to a warning plus NaN if you prefer.

Tests

• test_filter_rank_genes_groups_missing_genes: fails on main, passes with this change.
• test_filter_rank_genes_groups_refilter: filtering a filtered result still works.
• tests/test_rank_genes_groups.py, tests/test_filter_rank_genes_groups.py, tests/test_queries.py and the filter_rank_genes_groups plotting test pass locally; ruff check and ruff format are clean.

LLM usage

• LLM or coding agent used: Claude Opus 5.5 via Claude Code

• If one was used, I, a human, have reviewed, understood and tested every changed line

• Closes f"Values {list(not_found)}, from {list(indexer)}, " "are not valid obs/ var names or indices." #1487

• Tests included or not required because:

• Release notes not necessary because:

[read-the-docs-business]

Documentation build overview

📚 scanpy | 🛠️ Build #4692695 | 📁 Comparing a3d0edb against latest (117682e)

  🔍 Preview build  

3 files changed

± release-notes/index.html
± api/generated/scanpy.pp.calculate_qc_metrics.html
± api/generated/scanpy.settings.html

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 85.14%. Comparing base (117682e) to head (a3d0edb).
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #4497   +/-   ##
=======================================
  Coverage   85.13%   85.14%           
=======================================
  Files         138      138           
  Lines       12247    12253    +6     
=======================================
+ Hits        10427    10433    +6     
  Misses       1820     1820           

Flag	Coverage Δ
hatch-test.low-vers	78.67% <100.00%> (+0.01%)	⬆️
hatch-test.pre	84.93% <100.00%> (+<0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
src/scanpy/tools/_rank_genes_groups.py	96.12% <100.00%> (+0.07%)	⬆️

[sylvesterkaczmarek]

Could this refilter test also cover pts=True? With cached pts/pts_rest, the first filter leaves NaNs in names, and a second filter_rank_genes_groups(key='rank_genes_groups_filtered') calls .loc[var_names] with those NaNs, raising KeyError. The new helper skips NaNs for matrix indexing, but the cached-fraction path still needs the same handling.