n_pcs are no longer ignored for representations other than "X_pca" (#…

…2179)

* n_pcs are no longer ignored for non (scanpy-internal) X_pca reductions

This change allows to only take the dimensions needed also for non X_pca representations. This can be useful when using harmony for example or storing different PCA embeddings in the same object. Then it is not needed to first store them as X_pca to be able to fully use the n_pcs parameter. It might make sense then to adapt the documentation as such as well.

* Added error if n_pcs > dimensions in representation

* Fixed typo in variable name

* Fixed styling

* using consistent quotation marks; all done from my side

* first idea for test; check in detail tomorrow

* Used wrong quotes before

* shortened lines

* Improved style

* improved style in new function

* trying to finish the test today after all

* Removing wrong toarray()

* Test changes as per pre-commit

* Function changes as per pre-commit

* Another try at implementing pre-commit changes

* Added release note

* Attempt to fix reference error

Co-authored-by: pmajev <[email protected]>

docs/release-notes/1.9.0.md

@@ -6,3 +6,4 @@
 - {func}`~scanpy.tl.filter_rank_genes_groups` now allows to filter with absolute values of log fold change {pr}`1649` {smaller}`S Rybakov`
 - {func}`~scanpy.pl.embedding_density` now allows more than 10 groups {pr}`1936` {smaller}`A Wolf`
 - {func}`~scanpy.logging.print_versions` now uses `session_info` {pr}`2089` {smaller}`P Angerer` {smaller}`I Virshup`
+- `_choose_representation` now subsets the provided representation to n_pcs, regardless of the name of the provided representation (should affect mostly {func}`~scanpy.pp.neighbors`)  {pr}`2179`  {smaller}`I Virshup` {smaller}`PG Majev`

scanpy/tests/test_pca.py

@@ -136,3 +136,20 @@ def test_pca_chunked(pbmc3k_normalized):
         np.abs(chunked.uns["pca"]["variance_ratio"]),
         np.abs(default.uns["pca"]["variance_ratio"]),
     )
+
+
+def test_pca_n_pcs(pbmc3k_normalized):
+    """
+    Tests that the n_pcs parameter also works for
+    representations not called "X_pca"
+    """
+    pbmc = pbmc3k_normalized
+    sc.pp.pca(pbmc, dtype=np.float64)
+    pbmc.obsm["X_pca_test"] = pbmc.obsm["X_pca"]
+    original = sc.pp.neighbors(pbmc, n_pcs=5, use_rep="X_pca", copy=True)
+    renamed = sc.pp.neighbors(pbmc, n_pcs=5, use_rep="X_pca_test", copy=True)
+
+    assert np.allclose(original.obsm["X_pca"], renamed.obsm["X_pca_test"])
+    assert np.allclose(
+        original.obsp["distances"].toarray(), renamed.obsp["distances"].toarray()
+    )

scanpy/tools/_utils.py

@@ -47,10 +47,16 @@ def _choose_representation(adata, use_rep=None, n_pcs=None, silent=False):
             logg.info('    using data matrix X directly')
             X = adata.X
     else:
-        if use_rep in adata.obsm.keys():
+        if use_rep in adata.obsm.keys() and n_pcs is not None:
+            if n_pcs > adata.obsm[use_rep].shape[1]:
+                raise ValueError(
+                    f'{use_rep} does not have enough Dimensions. Provide a '
+                    'Representation with equal or more dimensions than'
+                    '`n_pcs` or lower `n_pcs` '
+                )
+            X = adata.obsm[use_rep][:, :n_pcs]
+        elif use_rep in adata.obsm.keys() and n_pcs is None:
             X = adata.obsm[use_rep]
-            if use_rep == 'X_pca' and n_pcs is not None:
-                X = adata.obsm[use_rep][:, :n_pcs]
         elif use_rep == 'X':
             X = adata.X
         else: