CLR: add a flavor argument and add flavors stoeckius and standard (closes #144)

The Seurat implementation differs from both the standard definition
of CLR and the version of CLR applied in the original CITE-Seq paper.
Therefore, allow the user to choose the CLR flavor.

Author: ilia-kats

muon/_prot/preproc.py

@@ -1,10 +1,11 @@
-from typing import Optional, Iterable, Tuple, Union
+from typing import Optional, Iterable, Tuple, Union, Literal
 from numbers import Integral, Real
 from warnings import warn
 
 import numpy as np
 import pandas as pd
 from scipy.sparse import issparse, csc_matrix, csr_matrix
+from scipy.stats import gmean
 from sklearn.mixture import GaussianMixture
 from sklearn.decomposition import PCA
 from sklearn.linear_model import LinearRegression
@@ -198,7 +199,12 @@ def dsb(
     return toreturn
 
 
-def clr(adata: AnnData, inplace: bool = True, axis: int = 0) -> Union[None, AnnData]:
+def clr(
+    adata: AnnData,
+    inplace: bool = True,
+    axis: int = 0,
+    flavor: Literal["seurat", "stoeckius", "standard"] = "seurat",
+) -> AnnData | None:
     """
     Apply the centered log ratio (CLR) transformation
     to normalize counts in adata.X.
@@ -207,6 +213,19 @@ def clr(adata: AnnData, inplace: bool = True, axis: int = 0) -> Union[None, AnnD
         data: AnnData object with protein expression counts.
         inplace: Whether to update adata.X inplace.
         axis: Axis across which CLR is performed.
+        flavor: How to perform the CLR transformation.
+
+            - seurat: Uses log1p transformations throughout. This results in non-negative values and preserves
+                sparse matrices.
+            - stoeckius: Follows the original CITE-Seq paper by adding a pseudocount of 1 to the data before
+                performing any transformations and using the standard log transform. This adheres more closely
+                to the standard definition of the CLR transform, but can yield negative values and does not
+                preserve sparse matrices (the result is always a dense matrix.)
+            - standard: The standard CLR transform without any pseudocounts. Does not preserve sparse matrices
+                and may yield infinite values if the input contains zeros.
+
+    References:
+        Stoeckius et al, 2017 (`doi:10.1038/nmeth.4380 <https://dx.doi.org/10.1038/nmeth.4380>`_)
     """
 
     if axis not in [0, 1]:
@@ -215,25 +234,33 @@ def clr(adata: AnnData, inplace: bool = True, axis: int = 0) -> Union[None, AnnD
     if not inplace:
         adata = adata.copy()
 
-    if issparse(adata.X) and axis == 0 and not isinstance(adata.X, csc_matrix):
-        warn("adata.X is sparse but not in CSC format. Converting to CSC.")
-        x = csc_matrix(adata.X)
-    elif issparse(adata.X) and axis == 1 and not isinstance(adata.X, csr_matrix):
-        warn("adata.X is sparse but not in CSR format. Converting to CSR.")
-        x = csr_matrix(adata.X)
-    else:
-        x = adata.X
+    x = adata.X
 
-    if issparse(x):
-        x.data /= np.repeat(
-            np.exp(np.log1p(x).sum(axis=axis).A / x.shape[axis]), x.getnnz(axis=axis)
-        )
-        np.log1p(x.data, out=x.data)
+    if flavor == "seurat":
+        if issparse(x) and axis == 0 and not isinstance(x):
+            warn(
+                "adata.X is sparse but not in CSC format. CSC format required for `axis=0`. Converting to CSC."
+            )
+            x = csc_matrix(x)
+        elif issparse(x) and axis == 1 and not isinstance(x):
+            warn(
+                "adata.X is sparse but not in CSR format. CSR format required for `axis=1`. Converting to CSR."
+            )
+            x = csr_matrix(x)
+
+        if issparse(x):
+            x.data /= np.repeat(np.exp(np.log1p(x).mean(axis=axis).toarray()), x.getnnz(axis=axis))
+            np.log1p(x.data, out=x.data)
+        else:
+            np.log1p(x / np.exp(np.log1p(x).mean(axis=axis, keepdims=True)), out=x)
+    elif flavor in ("stoeckius", "standard"):
+        if issparse(x):
+            x = x.toarray()
+        if flavor == "stoeckius":
+            x += 1
+        np.log(x / gmean(x, axis=axis, keepdims=True), out=x)
     else:
-        np.log1p(
-            x / np.exp(np.log1p(x).sum(axis=axis, keepdims=True) / x.shape[axis]),
-            out=x,
-        )
+        raise ValueError(f"Unknown flavor `{flavor}`.")
 
     adata.X = x