fix dataframe dtype conversion for empty df with non-default index

closes #145

Author: ilia-kats

CHANGELOG.md

@@ -8,12 +8,23 @@ and this project adheres to [Semantic Versioning][].
 [keep a changelog]: https://keepachangelog.com/en/1.1.0/
 [semantic versioning]: https://semver.org/spec/v2.0.0.html
 
-## [0.3.8] (Unreleased)
+## [0.4.0] (unreleased)
+
+### Changed
+
+- `update()` no longer automatically pulls obs/var columns from individual modalities by default. Set `mudata.set_options(pull_on_update=true)`
+  to restore the old behavior. Use `pull_obs/pull_var` and `push_obs/push_var` for more flexibility.
+- The settings API has changed. Use e.g. `mudata.settings.pull_on_update = True` instead of `mudata.set_options(pull_on_update=True)` and use
+  `mudata.settings.override` as context manager for local settings overrides.
+
+## [0.3.8]
 
 ### Fixed
 
 - `obs_vector` and `var_vector` once again return Pandas Series objects when appropriate. This fixes
   applying `sc.pl.umap` on a MuData object.
+- `update` with `pull_on_update != None` works again when run on a MuData object where all modalities have empty `.obs` / `.var` but the global
+  `.obs` or `.var` is not empty.
 
 ## [0.3.7]
 
@@ -169,6 +180,7 @@ To copy the annotations explicitly, you will need to use `pull_obs()` and/or `pu
 
 Initial `mudata` release with `MuData`, previously a part of the `muon` framework.
 
+[0.4.0]: https://github.com/scverse/mudata/releases/tag/v0.4.0
 [0.3.8]: https://github.com/scverse/mudata/releases/tag/v0.3.8
 [0.3.7]: https://github.com/scverse/mudata/releases/tag/v0.3.7
 [0.3.6]: https://github.com/scverse/mudata/releases/tag/v0.3.6

src/mudata/_core/utils.py

@@ -144,4 +144,4 @@ def try_convert_dataframe_to_numpy_dtypes(df: pd.DataFrame | Mapping[str, pd.Ser
     new_cols = {}
     for colname, col in df.items():
         new_cols[colname] = try_convert_series_to_numpy_dtype(col)
-    return pd.DataFrame(new_cols)
+    return pd.DataFrame(new_cols, index=df.index)

tests/test_pandas_utils.py

@@ -33,7 +33,8 @@ def test_try_convert_series_to_numpy_dtype(numpy_dtype, nullable_dtype):
 
 
 def test_try_convert_dataframe_to_numpy_dtypes():
-    series = pd.Series(np.arange(10))
+    idx = pd.Index(np.arange(10)[::-1].astype(str))
+    series = pd.Series(np.arange(10), index=idx)
     df = pd.DataFrame({"numpy": series})
 
     series = series.astype(pd.Int64Dtype())
@@ -45,3 +46,11 @@ def test_try_convert_dataframe_to_numpy_dtypes():
     df = try_convert_dataframe_to_numpy_dtypes(df)
     assert df["numpy"].dtype == df["nullable"].dtype == np.int64
     assert df["nullable_noconvert"].dtype == pd.Int64Dtype()
+    assert (df.index == idx).all()
+
+
+def test_try_convert_empty_dataframe_to_numpy_dtypes():
+    idx = pd.Index(np.arange(10)[::-1].astype(str))
+    df = pd.DataFrame(index=idx)
+    df = try_convert_dataframe_to_numpy_dtypes(df)
+    assert (df.index == idx).all()

tests/test_update.py

@@ -176,6 +176,23 @@ def test_update_simple(self, mdata: MuData, axis: Axis):
             getattr(mdata, f"{attr}_names")[: mdata["mod1"].shape[axis]] == getattr(mdata["mod1"], f"{attr}_names")
         ).all()
 
+    def test_update_simple_empty_modalities(self, modalities: Mapping[str, AnnData], axis: Axis):
+        for mod in modalities.values():
+            mod.obs = pd.DataFrame(index=mod.obs_names)
+            mod.var = pd.DataFrame(index=mod.var_names)
+        mdata = MuData(modalities)
+
+        old_obsnames = mdata.obs_names
+        old_varnames = mdata.var_names
+
+        mdata.obs["test"] = True
+        del mdata._obshash
+        del mdata._varhash
+        mdata.update()  # https://github.com/scverse/mudata/issues/145
+
+        assert (old_obsnames == mdata.obs_names).all()
+        assert (old_varnames == mdata.var_names).all()
+
     def test_update_add_modality(self, rng: np.random.Generator, modalities: Mapping[str, AnnData], axis: Axis):
         modnames = list(modalities.keys())
         mdata = add_mdata_global_columns(