Bump version: 1.1.0 → 1.1.1; counterfactual edges donor -> seed only

.bumpversion.cfg

@@ -1,5 +1,5 @@
 [bumpversion]
-current_version = 1.0.0
+current_version = 1.1.1
 commit = True
 tag = True
 

CHANGELOG.md

@@ -7,6 +7,7 @@ All notable changes to this project will be documented in this file.
 
 ## [Unreleased]
 
+## [1.1.1] — 2026-09-05
 ### Removed
 - **Breaking**: `condition_on_intrinsic` parameter removed from `Cellina`,
   `CellinaModule`, `CellinaGCN`, and `CellinaGCNModule`; behaviour is now always
@@ -27,6 +28,11 @@ All notable changes to this project will be documented in this file.
 - Explicitly stored zeros in the spatial connectivity matrix (e.g. from
   in-place weight thresholding) no longer become message-passing edges; they
   are dropped via `eliminate_zeros()` on a copy, leaving `adata.obsp` untouched.
+- `CellinaGCN` counterfactual loaders now add donor -> seed edges only instead of
+  bidirectional seed <-> donor edges. With multi-hop sampling, the reverse edges let
+  donors aggregate over the (control) seeds and pulled counterfactuals back toward
+  the control state. Donor draws use the same RNG stream, so `seed` still reproduces
+  the same donor sets; predictions change.
 
 ## [1.0.0] — 2026-06-04 - Release
 

pyproject.toml

@@ -5,7 +5,7 @@ requires = ["hatchling"]
 
 [project]
 name = "cellina"
-version = "1.1.0"
+version = "1.1.1"
 description = "Cellina: Dual-encoder VAE for spatial transcriptomics with adversarial domain forgetting"
 readme = "README.md"
 requires-python = ">=3.10"

src/cellina/_cellina_gcn_model.py

@@ -215,18 +215,15 @@ def _make_counterfactual_loader(
                 f"len(neighbour_indices) ({len(neighbour_indices)})"
             )
 
-        cf_src_parts, cf_dst_parts = [], []
-        for s in indices:
-            chosen = rng.choice(neighbour_indices, size=n_neighbors_per_seed, replace=False)
-            cf_src_parts.append(np.full(n_neighbors_per_seed, s))
-            cf_dst_parts.append(chosen)
-        cf_src = np.concatenate(cf_src_parts)
-        cf_dst = np.concatenate(cf_dst_parts)
-
-        cf_edges = np.stack([
-            np.concatenate([cf_src, cf_dst]),
-            np.concatenate([cf_dst, cf_src]),
-        ], axis=0)
+        # Same RNG stream / draw order as before, so seeds reproduce the same donor sets.
+        donors = [rng.choice(neighbour_indices, size=n_neighbors_per_seed, replace=False)
+                  for _ in indices]
+        cf_src = np.repeat(indices, n_neighbors_per_seed)
+        cf_dst = np.concatenate(donors)
+
+        # Donor -> seed only. Bidirectional edges let donors aggregate over the (control)
+        # seeds under multi-hop sampling, pulling the counterfactual back toward control.
+        cf_edges = np.stack([cf_dst, cf_src], axis=0)
 
         new_edge_index = np.concatenate([filtered_edges, cf_edges], axis=1)
         new_edge_index = torch.tensor(new_edge_index, dtype=torch.long)
@@ -715,7 +712,8 @@ def _resolve_num_neighbors(num_neighbors: Optional[List[int]], n_layers: int) ->
     - any other length    -> ``UserWarning``
     """
     if num_neighbors is None:
-        return [-1] * n_layers
+        # NOTE: this was the number of neighbors that did best in our benchmarks
+        return [-20] * n_layers
     num_neighbors = list(num_neighbors)
     if len(num_neighbors) != n_layers:
         warnings.warn(

tests/test_cellina_gcn.py

@@ -81,7 +81,7 @@ def test_num_neighbors_resolution(adata_with_spatial):
     with warnings.catch_warnings():
         warnings.simplefilter("error", UserWarning)
         model_default = CellinaGCN(adata_with_spatial, n_latent=5, n_layers=3)
-    assert model_default._num_neighbors == [-1, -1, -1]
+    assert model_default._num_neighbors == [-20, -20, -20]
 
     # Length 1 != n_layers -> warns, used as-is (no broadcast).
     with pytest.warns(UserWarning):
@@ -1036,3 +1036,21 @@ def test_counterfactual_rejects_oversized_neighbourhood(trained_model):
     donors = np.arange(60, 70)
     with pytest.raises(ValueError, match="must be less than"):
         model.get_counterfactual_latents(np.arange(10), donors, n_neighbors_per_seed=len(donors))
+
+
+def test_counterfactual_edges_are_donor_to_seed_only(trained_model):
+    """Counterfactual rewiring adds donor -> seed edges only; no edge may point at a donor from a seed."""
+    model, adata = trained_model
+    indices = np.arange(10)
+    donors = np.arange(60, adata.n_obs)
+    k = 5
+    loader = model._make_counterfactual_loader(indices, donors, n_neighbors_per_seed=k, batch_size=10, seed=0)
+    src, dst = loader.node_loader.data.edge_index.numpy()
+    seed_src, seed_dst = np.isin(src, indices), np.isin(dst, indices)
+    assert not seed_src.any(), "seeds must not be message sources (all seed edges were dropped)"
+    assert seed_dst.sum() == len(indices) * k
+    assert np.isin(src[seed_dst], donors).all()
+    # seed sets are reproducible and drawn from the same stream as before the fix
+    rng = np.random.default_rng(0)
+    expected = np.concatenate([rng.choice(donors, size=k, replace=False) for _ in indices])
+    np.testing.assert_array_equal(np.sort(src[seed_dst]), np.sort(expected))