fix: correct binary logistic marker score direction

docs/release-notes/4348.fix.md

@@ -0,0 +1 @@
+Correct the direction of binary logistic-regression marker scores in `tl.rank_genes_groups` when the reported group is the classifier’s first class {smaller}`D O’Toole`

src/scanpy/tools/_rank_genes_groups.py

@@ -597,11 +597,14 @@ def logreg(self, **kwds) -> Generator[_TestResult, None, None]:
         # not all codes necessarily appear in data
         existing_codes = np.unique(self.grouping.cat.codes)
         for igroup, cat in enumerate(self.groups_order):
+            cat_code: int = np.argmax(self.grouping.cat.categories == cat)
             if len(self.groups_order) <= 2:  # binary logistic regression
+                # Binary coefficients point toward classes_[1]; orient them
+                # toward the group under which the scores will be reported.
                 scores = scores_all[0]
+                if cat_code == clf.classes_[0]:
+                    scores = -scores
             else:
-                # cat code is index of cat value in .categories
-                cat_code: int = np.argmax(self.grouping.cat.categories == cat)
                 # index of scores row is index of cat code in array of existing codes
                 scores_idx: int = np.argmax(existing_codes == cat_code)
                 scores = scores_all[scores_idx]

tests/test_rank_genes_groups_logreg.py

@@ -62,3 +62,66 @@ def test_rank_genes_groups_with_unsorted_groups():
         bdata.uns["rank_genes_groups"]["scores"]["Three"]
     ).to_numpy()
     np.testing.assert_equal(array_ad, array_bd)
+
+
+@pytest.mark.parametrize("categories", [["A", "unused", "B"], ["B", "unused", "A"]])
+@pytest.mark.parametrize("target", ["A", "B"])
+@pytest.mark.parametrize("representation", ["dense", "sparse", "raw", "layer"])
+def test_binary_logreg_scores_point_toward_requested_group(
+    categories, target, representation
+):
+    from anndata import AnnData
+    from scipy.sparse import csr_matrix  # noqa: TID251
+
+    x = np.vstack([
+        np.tile([10.0, 0.0, 1.0], (30, 1)),
+        np.tile([0.0, 10.0, 1.0], (30, 1)),
+    ])
+    adata = AnnData(
+        x,
+        obs=pd.DataFrame(
+            {"group": pd.Categorical(["A"] * 30 + ["B"] * 30, categories=categories)},
+            index=[f"cell_{i}" for i in range(60)],
+        ),
+        var=pd.DataFrame(index=["A_marker", "B_marker", "shared"]),
+    )
+    kwargs = {"use_raw": False}
+    if representation == "sparse":
+        adata.X = csr_matrix(adata.X)
+    elif representation == "raw":
+        adata.raw = adata.copy()
+        adata.X = np.zeros_like(x)
+        kwargs = {"use_raw": True}
+    elif representation == "layer":
+        adata.layers["expression"] = adata.X.copy()
+        adata.X = np.zeros_like(x)
+        kwargs["layer"] = "expression"
+    reference = "B" if target == "A" else "A"
+    sc.tl.rank_genes_groups(
+        adata, "group", groups=[target], reference=reference, method="logreg", **kwargs
+    )
+    result = adata.uns["rank_genes_groups"]
+    names = result["names"][target]
+    scores = dict(zip(names, result["scores"][target], strict=True))
+    assert names[0] == f"{target}_marker"
+    assert scores[f"{target}_marker"] > 0
+    assert scores[f"{reference}_marker"] < 0
+
+
+@pytest.mark.parametrize("categories", [["A", "B"], ["B", "A"]])
+def test_binary_logreg_default_group_direction(categories):
+    from anndata import AnnData
+
+    adata = AnnData(
+        np.vstack([np.tile([10.0, 0.0], (20, 1)), np.tile([0.0, 10.0], (20, 1))]),
+        obs=pd.DataFrame(
+            {"group": pd.Categorical(["A"] * 20 + ["B"] * 20, categories=categories)},
+            index=[f"cell_{i}" for i in range(40)],
+        ),
+        var=pd.DataFrame(index=["A_marker", "B_marker"]),
+    )
+    sc.tl.rank_genes_groups(adata, "group", method="logreg", use_raw=False)
+    result = adata.uns["rank_genes_groups"]
+    target = categories[0]
+    assert result["names"][target][0] == f"{target}_marker"
+    assert result["scores"][target][0] > 0