Adding "networkit" flavour to scanpy.tl.leiden for parallel community detection

src/scanpy/_settings/presets.py

@@ -35,7 +35,7 @@
 
 type DETest = Literal["logreg", "t-test", "wilcoxon", "t-test_overestim_var"]
 type HVGFlavor = Literal["seurat", "cell_ranger", "seurat_v3", "seurat_v3_paper"]
-type LeidenFlavor = Literal["leidenalg", "igraph"]
+type LeidenFlavor = Literal["leidenalg", "igraph", "networkit"]
 
 
 @dataclass

src/scanpy/_utils/__init__.py

@@ -72,6 +72,7 @@
     "ensure_igraph",
     "get_igraph_from_adjacency",
     "get_literal_vals",
+    "get_networkit_from_adjacency",
     "indent",
     "is_backed_type",
     "is_backed_type",
@@ -98,6 +99,17 @@ def ensure_igraph() -> None:
     raise ImportError(msg)
 
 
+def ensure_network() -> None:
+    if importlib.util.find_spec("netowrkit"):
+        return
+    msg = (
+        "Please install the networkit package: "
+        "`conda install conda-forge::networkit` or"
+        "`pip install networkit`"
+    )
+    raise ImportError(msg)
+
+
 def _getdoc(c_or_f: Callable | type) -> str | None:
     if getattr(c_or_f, "__doc__", None) is None:
         return None
@@ -303,6 +315,25 @@ def get_igraph_from_adjacency(adjacency: CSBase, *, directed: bool) -> Graph:
     return g
 
 
+def get_networkit_from_adjacency(adjacency: CSBase, *, weighted: bool = True):
+    """Get a NetworKit graph from an adjacency matrix."""
+    import networkit as nk
+    import scipy.sparse as sps
+
+    shape = adjacency.shape
+    assert shape is not None
+    upper = sps.triu(adjacency, k=1).tocoo()
+    return nk.graph.GraphFromCoo(
+        (
+            upper.data.astype(np.float64),
+            (upper.row.astype(np.uint64), upper.col.astype(np.uint64)),
+        ),
+        n=adjacency.shape[0],
+        weighted=weighted,
+        directed=False,
+    )
+
+
 # --------------------------------------------------------------------------------
 # Group stuff
 # --------------------------------------------------------------------------------

src/scanpy/tools/_leiden.py

@@ -185,6 +185,33 @@ def leiden(  # noqa: PLR0913
             "MutableVertexPartition",
             leidenalg.find_partition(g, partition_type, seed=seed, **clustering_args),
         )
+    elif flavor == "networkit":
+        from types import SimpleNamespace
+
+        import networkit
+
+        seed = int(rng.integers(np.iinfo(np.int64).max))
+        networkit.setSeed(seed, useThreadId=True)
+        # only undirected for Parallel Leiden
+        g = _utils.get_networkit_from_adjacency(adjacency, weighted=use_weights)
+        iterations = n_iterations if n_iterations > 0 else 3
+        gamma = 1.0 if resolution is None else resolution
+        # randomization was removed as an option, so it is randomize = True
+        algorithm = networkit.community.ParallelLeiden(
+            g, iterations=iterations, gamma=gamma
+        )
+        # applying algorithm to the graph
+        algorithm.run()
+        nk_part = algorithm.getPartition()
+        # NetworKit's Partition exposes getVector() for the labels and a
+        # separate Modularity measure, rather than .membership / .modularity.
+
+        part = SimpleNamespace(
+            # get the actual vector representing the partition data structure
+            membership=np.asarray(nk_part.getVector()),
+            modularity=networkit.community.Modularity().getQuality(nk_part, g),
+        )
+
     else:
         g = _utils.get_igraph_from_adjacency(adjacency, directed=False)
         if use_weights:
@@ -243,6 +270,13 @@ def _validate_flavor(
             if partition_type is not None:
                 msg = "Do not pass in partition_type argument when using igraph."
                 raise ValueError(msg)
+        case "networkit":
+            if directed:
+                msg = "Cannot use NetworKit's leiden implementation with a directed graph."
+                raise ValueError(msg)
+            if partition_type is not None:
+                msg = "Do not pass in partition_type argument when using networkit."
+                raise ValueError(msg)
         case "leidenalg":
             msg = (
                 "The `igraph` implementation of leiden clustering is *orders of magnitude faster*. "