Merge pull request #22 from YosefLab/sparse_distances

Sparse distances

docs/source/hotspot.rst

@@ -1,8 +1,5 @@
 Function Reference
 ******************
 
-
-Where does this stuff get included?
-
 .. autoclass:: hotspot.Hotspot
     :members:

hotspot/hotspot.py

@@ -90,12 +90,6 @@ def __init__(
                 "Both `distances_obsp_key` and `tree` provided - only one of these should be provided."
             )
 
-        if distances is not None:
-            assert not issparse(distances)
-            distances = pd.DataFrame(
-                distances, index=adata.obs_names, columns=adata.obs_names
-            )
-
         if latent is not None:
             latent = pd.DataFrame(latent, index=adata.obs_names)
 
@@ -260,7 +254,7 @@ def legacy_init(
         input_adata.obs[tc_key] = umi_counts.values
         dkey = "distances"
         if distances is not None:
-            input_adata.obsp[dkey] = np.asarray(distances)
+            input_adata.obsp[dkey] = distances
             dist_input = True
         else:
             dist_input = False

hotspot/knn.py

@@ -5,6 +5,7 @@
 from numba import jit
 from tqdm import tqdm
 from pynndescent import NNDescent
+import warnings
 
 
 def neighbors_and_weights(data, n_neighbors=30, neighborhood_factor=3, approx_neighbors=True):
@@ -26,8 +27,10 @@ def neighbors_and_weights(data, n_neighbors=30, neighborhood_factor=3, approx_ne
     coords = data.values
 
     if approx_neighbors:
-        index = NNDescent(coords, n_neighbors=n_neighbors)
+        # pynndescent first neighbor is self, unlike sklearn
+        index = NNDescent(coords, n_neighbors=n_neighbors + 1)
         ind, dist = index.neighbor_graph
+        ind, dist = ind[:, 1:], dist[:, 1:]
     else:
         nbrs = NearestNeighbors(n_neighbors=n_neighbors,
                         algorithm="ball_tree").fit(coords)
@@ -45,7 +48,7 @@ def neighbors_and_weights(data, n_neighbors=30, neighborhood_factor=3, approx_ne
 
 
 def neighbors_and_weights_from_distances(
-    distances, n_neighbors=30, neighborhood_factor=3
+    distances, cell_index, n_neighbors=30, neighborhood_factor=3
 ):
     """
     Computes nearest neighbors and associated weights using
@@ -61,15 +64,23 @@ def neighbors_and_weights_from_distances(
     weights:  pandas.Dataframe num_cells x n_neighbors
 
     """
+    if isinstance(distances, pd.DataFrame):
+        distances = distances.values
 
     nbrs = NearestNeighbors(
         n_neighbors=n_neighbors, algorithm="brute", metric="precomputed"
-    ).fit(distances.values)
-    dist, ind = nbrs.kneighbors()
+    ).fit(distances)
+    try:
+        dist, ind = nbrs.kneighbors()
+    # already is a neighbors graph
+    except ValueError:
+        nn = np.asarray((distances[0] > 0).sum())
+        warnings.warn(f"Provided cell-cell distance graph is likely a {nn}-neighbors graph. Using {nn} precomputed neighbors.")
+        dist, ind = nbrs.kneighbors(n_neighbors=nn-1)
 
     weights = compute_weights(dist, neighborhood_factor=neighborhood_factor)
 
-    ind = pd.DataFrame(ind, index=distances.index)
+    ind = pd.DataFrame(ind, index=cell_index)
     neighbors = ind
     weights = pd.DataFrame(
         weights, index=neighbors.index, columns=neighbors.columns

pyproject.toml

@@ -29,7 +29,7 @@ packages = [
   {include = "hotspot"},
 ]
 readme = "README.md"
-version = "1.1"
+version = "1.1.1"
 
 [tool.poetry.dependencies]
 pytest = {version = ">=5.0", optional = true}
@@ -50,10 +50,11 @@ sphinx-book-theme = {version= ">=0.2.0", optional = true}
 nbsphinx = {version = "*", optional = true}
 sphinx = {version = ">=4.1", optional = true}
 ipython = {version = "*", optional = true}
+scanpy = {version = "*", optional = true}
 
 
 [tool.poetry.extras]
-test = ["pytest"]
+test = ["pytest", "scanpy"]
 docs=["sphinx-book-theme", "nbsphinx", "sphinx>", "ipython"]
 
 [tool.poetry.dev-dependencies]

tests/test_validations.py

@@ -4,6 +4,7 @@
 from hotspot import Hotspot
 import anndata
 import pytest
+import scanpy as sc
 
 from scipy.sparse import csc_matrix
 
@@ -73,6 +74,18 @@ def test_models():
             assert isinstance(hs.module_scores, pd.DataFrame)
             assert (hs.module_scores.index == gene_exp.columns).all()
 
+    # test precomputed distance matrix
+    sc.pp.neighbors(adata, use_rep="latent", n_neighbors=30)
+    hs = Hotspot(
+        adata,
+        model="normal",
+        umi_counts_obs_key="umi_counts",
+        layer_key="sparse",
+        distances_obsp_key="distances"
+    )
+    hs.create_knn_graph(False, n_neighbors=30)
+    hs.compute_autocorrelations()
+
 
 def test_filter_genes():
     """