Remove duplicated row names from the reference before marker detection.

These cause trouble as they confuse our name-based approach to matters.

src/singler/_utils.py

@@ -101,14 +101,17 @@ def _clean_matrix(x, features, assay_type, check_missing, num_threads):
 
 
 def _restrict_features(data, features, restrict_to):
-    if restrict_to is not None:
-        if not isinstance(restrict_to, set):
-            restrict_to = set(restrict_to)
-        keep = []
-        new_features = []
-        for i, x in enumerate(features):
-            if x in restrict_to:
-                keep.append(i)
-                new_features.append(x)
-        return delayedarray.DelayedArray(data)[keep, :], new_features
-    return data, features
+    if restrict_to is None:
+        return data, features
+
+    if not isinstance(restrict_to, set):
+        restrict_to = set(restrict_to)
+    keep = []
+    for i, x in enumerate(features):
+        if x in restrict_to:
+            keep.append(i)
+
+    return (
+        delayedarray.DelayedArray(data)[keep, :],
+        biocutils.subset_sequence(features, keep)
+    )

src/singler/train_single.py

@@ -102,8 +102,9 @@ def train_single(
     marker_method: Literal["classic"] = "classic",
     marker_args: dict = {},
     nn_parameters: Optional[knncolle.Parameters] = knncolle.AnnoyParameters(),
+    check_duplicated: bool = True,
     num_threads: int = 1,
-    ) -> TrainedSingleReference:
+) -> TrainedSingleReference:
     """Build a single reference dataset in preparation for classification.
 
     Args:
@@ -163,6 +164,10 @@ def train_single(
             Algorithm for constructing the neighbor search index, used to
             compute scores during classification.
 
+        check_duplicated:
+            Whether to remove rows with duplicate feature names. This can be
+            set to False if ``ref_features`` does not contain any duplicates.
+
         num_threads:
             Number of threads to use for reference building.
 
@@ -179,8 +184,19 @@ def train_single(
         num_threads=num_threads,
     )
 
+    if check_duplicated:
+        encountered = set()
+        keep = []
+        for i, rg in enumerate(ref_features):
+            if rg not in encountered:
+                encountered.add(rg)
+                keep.append(i)
+        if len(keep) != len(ref_features):
+            ref_features = biocutils.subset_sequence(ref_features, keep)
+            ref_data = delayedarray.DelayedArray(ref_data)[keep,:]
+
     unique_labels, label_idx = _factorize(ref_labels)
-    markers = identify_genes(ref_data, ref_features, ref_labels, unique_labels, markers, marker_method, test_features, restrict_to, marker_args, num_threads)
+    markers = _identify_genes(ref_data, ref_features, ref_labels, unique_labels, markers, marker_method, test_features, restrict_to, marker_args, num_threads)
     markers_idx = [None] * len(unique_labels)
     for outer_i, outer_k in enumerate(unique_labels):
         inner_instance = [None] * len(unique_labels)
@@ -217,7 +233,7 @@ def train_single(
     )
 
 
-def identify_genes(ref_data, ref_features, ref_labels, unique_labels, markers, marker_method, test_features, restrict_to, marker_args, num_threads):
+def _identify_genes(ref_data, ref_features, ref_labels, unique_labels, markers, marker_method, test_features, restrict_to, marker_args, num_threads):
     ref_data, ref_features = _restrict_features(ref_data, ref_features, test_features)
     ref_data, ref_features = _restrict_features(ref_data, ref_features, restrict_to)
 

tests/test_train_single.py

@@ -38,6 +38,16 @@ def test_train_single_markers():
     assert built.markers == mbuilt.markers
 
 
+def test_train_single_dedup():
+    ref = numpy.random.rand(10000, 10)
+    labels = ["A", "B", "C", "D", "E", "E", "D", "C", "B", "A"]
+    features = [str(i) for i in range(ref.shape[0])]
+    features[0] = "1"
+    built = singler.train_single(ref, labels, features)
+    assert built.features == features[1:] # duplicates are ignored
+    assert built._full_data.shape[0] == len(built.features)
+
+
 def test_train_single_restricted():
     ref = numpy.random.rand(10000, 10)
     labels = ["A", "B", "C", "D", "E", "E", "D", "C", "B", "A"]