remove patsy dependency

theislab · Oct 27, 2023 · d03a079 · d03a079
1 parent 9c23d03
commit d03a079
Show file tree

Hide file tree

Showing 4 changed files with 12 additions and 16 deletions.
diff --git a/ncem/tl/fit/backend/design_matrix.py b/ncem/tl/fit/backend/design_matrix.py
@@ -4,6 +4,7 @@
 import numpy as np
 import pandas as pd
 import patsy
+from formulaic import Formula
 
 from ncem.tl.fit.constants import PREFIX_INDEX, PREFIX_NEIGHBOR
 
@@ -248,11 +249,9 @@ def get_dmat_from_obs(obs: pd.DataFrame, obs_niche: pd.DataFrame, formula: str,
     # Merge sample annotation:
     obs_full = pd.concat([obs, obs_index_type, obs_niche], axis=1)
     columns_names = formula.replace('~0+','').split('+')
-
-    dmat = patsy.dmatrix(formula, obs_full)
-    # Simplify names, this is necessary for patsy to accept these as terms later.
-    column_names = [f"{x.split('[')[0]}{x.split('[')[1].split(']')[-1]}" for x in dmat.design_info.column_names]
-    dmat = pd.DataFrame(np.asarray(dmat), index=obs.index, columns=column_names)
+
+    dmat = Formula(formula).get_model_matrix(obs_full)
+    dmat = pd.DataFrame(np.asarray(dmat), index=obs.index, columns=dmat.columns)
     return dmat
 
 
@@ -298,13 +297,8 @@ def get_dmats_from_deconvoluted(
         obs_index_type_x.index = obs.index
         # Merge sample annotation:
         obs_full = pd.concat([obs, obs_index_type_x, obs_niche], axis=1)
-        dmats[x] = patsy.dmatrix(formulas[x], obs_full)
-        # ensure that column names start with index type name
-        dmat_columns = [
-            col if col.startswith(PREFIX_INDEX) else PREFIX_INDEX + x + col for col in dmats[x].design_info.column_names
-        ]
-        dmat_columns = [f"{x.split('[')[0]}{x.split('[')[1].split(']')[-1]}" for x in dmat_columns]
-        dmats[x] = pd.DataFrame(np.asarray(dmats[x]), index=obs.index, columns=dmat_columns)
+        dmats[x] = Formula(formulas[x]).get_model_matrix(obs_full)
+        dmats[x] = pd.DataFrame(np.asarray(dmats[x]), index=obs.index, columns=dmats[x].columns)
     return dmats
 
 
@@ -353,7 +347,8 @@ def get_dmat_global_from_deconvoluted(obs: pd.DataFrame, deconv: pd.DataFrame, f
         obs_index_type[x].index = obs.index
         # Merge sample annotation:
         obs_full = pd.concat([obs, obs_index_type_x, obs_niche], axis=1)
-        dmat_x = patsy.dmatrix(formula, obs_full)
-        dmats.append(pd.DataFrame(np.asarray(dmat_x), index=obs.index, columns=dmat_x.design_info.column_names))
+
+        dmat_x = Formula(formula).get_model_matrix(obs_full)
+        dmats.append(pd.DataFrame(np.asarray(dmat_x), index=obs.index, columns=dmat_x.columns))
     dmat = pd.concat(dmats, axis=0)
     return dmat
diff --git a/ncem/tl/fit/backend/linear_model.py b/ncem/tl/fit/backend/linear_model.py
@@ -274,7 +274,6 @@ def linear_ncem_deconvoluted(
     )
     dmats = get_dmats_from_deconvoluted(deconv=adata.obsm[key_deconvolution], formulas=formulas, obs=adata.obs)
     for k, v in dmats.items():
-        print(k)
         dmat_key = f"{OBSM_KEY_DMAT}_{k}"
         adata.obsm[dmat_key] = v
         params = ols_fit(x_=adata.obsm[dmat_key].values, y_=adata.layers[k])

diff --git a/ncem/tl/fit/backend/testing.py b/ncem/tl/fit/backend/testing.py
@@ -150,6 +150,7 @@ def test_deconvoluted(
             for y in coef_to_test:
                 if y in parameter_names:
                     idx = parameter_names.index(y)
+                    print(idx)
                     theta_mle = params.values[:, idx]
                     theta_sd = fisher_inv[:, idx, idx]
                     theta_sd = np.nextafter(0, np.inf, out=theta_sd, where=theta_sd < np.nextafter(0, np.inf))

diff --git a/pyproject.toml b/pyproject.toml
@@ -22,10 +22,11 @@ python = ">=3.8,<=3.10"
 click = "^8.0.0"
 rich = "^10.1.0"
 PyYAML = "^5.4.1"
-Jinja2 = ">=2.11.3,<4.0.0"
+Jinja2 = ">=2.11.3,<=4.0.0"
 scanpy = "^1.9.3"
 squidpy = "^1.2.3"
 patsy = "^0.5.1"
+formulaic = "=0.6.6"
 scipy = "=1.9.1"
 seaborn = "^0.12.2"
 matplotlib = "^3.7.1"