diff --git a/pertpy/tools/_perturbation_space/_perturbation_space.py b/pertpy/tools/_perturbation_space/_perturbation_space.py
index e3d1a0cf..ea25dae5 100644
--- a/pertpy/tools/_perturbation_space/_perturbation_space.py
+++ b/pertpy/tools/_perturbation_space/_perturbation_space.py
@@ -186,11 +186,6 @@ def _combine(
 
         if ensure_consistency:
             adata = self.compute_control_diff(adata, copy=True, all_data=True, target_col=target_col)
-            rename_back = {
-                key: key.removesuffix("_control_diff")
-                for key in [*adata.layers.keys(), *adata.obsm.keys()]
-                if key.endswith("_control_diff")
-            }
         else:
             warnings.warn(
                 "Combining perturbations without `ensure_consistency=True` is only well-defined "
@@ -198,7 +193,21 @@ def _combine(
                 "(otherwise perturbation - perturbation != control).",
                 stacklevel=3,
             )
-            rename_back = {}
+
+        # sc.get.aggregate can leave a `None`-keyed layer behind (the pre-aggregation .X);
+        # PseudobulkSpace.compute strips it but defensively re-strip here so callers passing
+        # a hand-built AnnData don't crash inside the string ops below.
+        layer_keys = [k for k in adata.layers if isinstance(k, str)]
+        obsm_keys = [k for k in adata.obsm if isinstance(k, str)]
+        rename_back = (
+            {
+                key: key.removesuffix("_control_diff")
+                for key in [*layer_keys, *obsm_keys]
+                if key.endswith("_control_diff")
+            }
+            if ensure_consistency
+            else {}
+        )
 
         def _running(values: np.ndarray) -> np.ndarray:
             result = values[adata.obs_names.get_loc(reference_key)].astype(float, copy=True)
@@ -207,13 +216,15 @@ def _running(values: np.ndarray) -> np.ndarray:
             return result
 
         new_layers: dict[str, np.ndarray] = {}
-        for layer_key, mat in adata.layers.items():
+        for layer_key in layer_keys:
+            mat = adata.layers[layer_key]
             new_layers[rename_back.get(layer_key, layer_key)] = np.concatenate(
                 (np.asarray(mat), _running(np.asarray(mat))[None, :]), axis=0
             )
 
         new_obsm: dict[str, np.ndarray] = {}
-        for embedding_key, mat in adata.obsm.items():
+        for embedding_key in obsm_keys:
+            mat = adata.obsm[embedding_key]
             new_obsm[rename_back.get(embedding_key, embedding_key)] = np.concatenate(
                 (np.asarray(mat), _running(np.asarray(mat))[None, :]), axis=0
             )