anndata tokenizer
#102
by
giovp
- opened
- geneformer/tokenizer.py +95 -12
geneformer/tokenizer.py
CHANGED
@@ -14,6 +14,9 @@ Usage:
|
|
14 |
tk.tokenize_data("loom_data_directory", "output_directory", "output_prefix")
|
15 |
"""
|
16 |
|
|
|
|
|
|
|
17 |
import pickle
|
18 |
from pathlib import Path
|
19 |
|
@@ -85,42 +88,119 @@ class TranscriptomeTokenizer:
|
|
85 |
# protein-coding and miRNA gene list dictionary for selecting .loom rows for tokenization
|
86 |
self.genelist_dict = dict(zip(self.gene_keys, [True] * len(self.gene_keys)))
|
87 |
|
88 |
-
def tokenize_data(
|
|
|
|
|
|
|
|
|
|
|
|
|
89 |
"""
|
90 |
Tokenize .loom files in loom_data_directory and save as tokenized .dataset in output_directory.
|
91 |
|
92 |
Parameters
|
93 |
----------
|
94 |
loom_data_directory : Path
|
95 |
-
Path to directory containing loom files
|
96 |
output_directory : Path
|
97 |
Path to directory where tokenized data will be saved as .dataset
|
98 |
output_prefix : str
|
99 |
Prefix for output .dataset
|
|
|
|
|
100 |
"""
|
101 |
-
tokenized_cells, cell_metadata = self.tokenize_files(
|
|
|
|
|
102 |
tokenized_dataset = self.create_dataset(tokenized_cells, cell_metadata)
|
103 |
|
104 |
output_path = (Path(output_directory) / output_prefix).with_suffix(".dataset")
|
105 |
tokenized_dataset.save_to_disk(output_path)
|
106 |
|
107 |
-
def tokenize_files(
|
|
|
|
|
108 |
tokenized_cells = []
|
109 |
loom_cell_attr = [attr_key for attr_key in self.custom_attr_name_dict.keys()]
|
110 |
-
cell_metadata = {
|
|
|
|
|
111 |
|
112 |
-
# loops through directories to tokenize .loom files
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
)
|
|
|
118 |
tokenized_cells += file_tokenized_cells
|
119 |
for k in loom_cell_attr:
|
120 |
cell_metadata[self.custom_attr_name_dict[k]] += file_cell_metadata[k]
|
121 |
|
122 |
return tokenized_cells, cell_metadata
|
123 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
124 |
def tokenize_file(self, loom_file_path):
|
125 |
file_cell_metadata = {
|
126 |
attr_key: [] for attr_key in self.custom_attr_name_dict.keys()
|
@@ -162,7 +242,9 @@ class TranscriptomeTokenizer:
|
|
162 |
|
163 |
# scan through .loom files and tokenize cells
|
164 |
tokenized_cells = []
|
165 |
-
for
|
|
|
|
|
166 |
# select subview with protein-coding and miRNA genes
|
167 |
subview = view.view[coding_miRNA_loc, :]
|
168 |
|
@@ -174,6 +256,7 @@ class TranscriptomeTokenizer:
|
|
174 |
* 10_000
|
175 |
/ norm_factor_vector[:, None]
|
176 |
)
|
|
|
177 |
# tokenize subview gene vectors
|
178 |
tokenized_cells += [
|
179 |
tokenize_cell(subview_norm_array[:, i], coding_miRNA_tokens)
|
|
|
14 |
tk.tokenize_data("loom_data_directory", "output_directory", "output_prefix")
|
15 |
"""
|
16 |
|
17 |
+
|
18 |
+
from __future__ import annotations
|
19 |
+
from typing import Literal
|
20 |
import pickle
|
21 |
from pathlib import Path
|
22 |
|
|
|
88 |
# protein-coding and miRNA gene list dictionary for selecting .loom rows for tokenization
|
89 |
self.genelist_dict = dict(zip(self.gene_keys, [True] * len(self.gene_keys)))
|
90 |
|
91 |
+
def tokenize_data(
|
92 |
+
self,
|
93 |
+
data_directory: Path | str,
|
94 |
+
output_directory: Path | str,
|
95 |
+
output_prefix: str,
|
96 |
+
file_format: Literal["loom", "h5ad"] = "loom",
|
97 |
+
):
|
98 |
"""
|
99 |
Tokenize .loom files in loom_data_directory and save as tokenized .dataset in output_directory.
|
100 |
|
101 |
Parameters
|
102 |
----------
|
103 |
loom_data_directory : Path
|
104 |
+
Path to directory containing loom files or anndata files
|
105 |
output_directory : Path
|
106 |
Path to directory where tokenized data will be saved as .dataset
|
107 |
output_prefix : str
|
108 |
Prefix for output .dataset
|
109 |
+
file_format : str
|
110 |
+
Format of input files. Can be "loom" or "h5ad".
|
111 |
"""
|
112 |
+
tokenized_cells, cell_metadata = self.tokenize_files(
|
113 |
+
Path(data_directory), file_format
|
114 |
+
)
|
115 |
tokenized_dataset = self.create_dataset(tokenized_cells, cell_metadata)
|
116 |
|
117 |
output_path = (Path(output_directory) / output_prefix).with_suffix(".dataset")
|
118 |
tokenized_dataset.save_to_disk(output_path)
|
119 |
|
120 |
+
def tokenize_files(
|
121 |
+
self, data_directory, file_format: Literal["loom", "h5ad"] = "loom"
|
122 |
+
):
|
123 |
tokenized_cells = []
|
124 |
loom_cell_attr = [attr_key for attr_key in self.custom_attr_name_dict.keys()]
|
125 |
+
cell_metadata = {
|
126 |
+
attr_key: [] for attr_key in self.custom_attr_name_dict.values()
|
127 |
+
}
|
128 |
|
129 |
+
# loops through directories to tokenize .loom or .h5ad files
|
130 |
+
tokenize_file_fn = (
|
131 |
+
self.tokenize_file if file_format == "loom" else self.tokenize_anndata
|
132 |
+
)
|
133 |
+
for file_path in data_directory.glob("*.{}".format(file_format)):
|
134 |
+
print(f"Tokenizing {file_path}")
|
135 |
+
file_tokenized_cells, file_cell_metadata = tokenize_file_fn(file_path)
|
136 |
tokenized_cells += file_tokenized_cells
|
137 |
for k in loom_cell_attr:
|
138 |
cell_metadata[self.custom_attr_name_dict[k]] += file_cell_metadata[k]
|
139 |
|
140 |
return tokenized_cells, cell_metadata
|
141 |
|
142 |
+
def tokenize_anndata(self, adata_file_path):
|
143 |
+
import anndata as ad
|
144 |
+
|
145 |
+
adata = ad.read(adata_file_path)
|
146 |
+
file_cell_metadata = {
|
147 |
+
attr_key: [] for attr_key in self.custom_attr_name_dict.keys()
|
148 |
+
}
|
149 |
+
|
150 |
+
coding_miRNA_loc = np.where(
|
151 |
+
[self.genelist_dict.get(i, False) for i in adata.var["ensembl_id"]]
|
152 |
+
)[0]
|
153 |
+
norm_factor_vector = np.array(
|
154 |
+
[
|
155 |
+
self.gene_median_dict[i]
|
156 |
+
for i in adata.var["ensembl_id"][coding_miRNA_loc]
|
157 |
+
]
|
158 |
+
)
|
159 |
+
coding_miRNA_ids = adata.var["ensembl_id"][coding_miRNA_loc]
|
160 |
+
coding_miRNA_tokens = np.array(
|
161 |
+
[self.gene_token_dict[i] for i in coding_miRNA_ids]
|
162 |
+
)
|
163 |
+
|
164 |
+
try:
|
165 |
+
adata.obs["filter_pass"]
|
166 |
+
except KeyError:
|
167 |
+
var_exists = False
|
168 |
+
else:
|
169 |
+
var_exists = True
|
170 |
+
|
171 |
+
if var_exists is True:
|
172 |
+
filter_pass_loc = np.where(
|
173 |
+
[True if i == 1 else False for i in adata.obs["filter_pass"]]
|
174 |
+
)[0]
|
175 |
+
elif var_exists is False:
|
176 |
+
print(
|
177 |
+
f"{adata_file_path} has no column attribute 'filter_pass'; tokenizing all cells."
|
178 |
+
)
|
179 |
+
filter_pass_loc = np.array([i for i in range(adata.shape[0])])
|
180 |
+
|
181 |
+
tokenized_cells = []
|
182 |
+
adata_filter = adata[
|
183 |
+
filter_pass_loc, coding_miRNA_loc # filter cells and genes
|
184 |
+
]
|
185 |
+
|
186 |
+
X_norm = (
|
187 |
+
adata_filter.X
|
188 |
+
/ adata.obs["n_counts"].values.reshape(-1, 1)
|
189 |
+
* 10_000
|
190 |
+
/ norm_factor_vector
|
191 |
+
).tocsr()
|
192 |
+
|
193 |
+
tokenized_cells += [
|
194 |
+
tokenize_cell(X_norm[i, ...].A.flatten(), coding_miRNA_tokens)
|
195 |
+
for i in range(X_norm.shape[0])
|
196 |
+
]
|
197 |
+
|
198 |
+
# add custom attributes for subview to dict
|
199 |
+
for k in file_cell_metadata.keys():
|
200 |
+
file_cell_metadata[k] += adata_filter.obs[k].tolist()
|
201 |
+
|
202 |
+
return tokenized_cells, file_cell_metadata
|
203 |
+
|
204 |
def tokenize_file(self, loom_file_path):
|
205 |
file_cell_metadata = {
|
206 |
attr_key: [] for attr_key in self.custom_attr_name_dict.keys()
|
|
|
242 |
|
243 |
# scan through .loom files and tokenize cells
|
244 |
tokenized_cells = []
|
245 |
+
for _ix, _selection, view in data.scan(
|
246 |
+
items=filter_pass_loc, axis=1, layers=""
|
247 |
+
):
|
248 |
# select subview with protein-coding and miRNA genes
|
249 |
subview = view.view[coding_miRNA_loc, :]
|
250 |
|
|
|
256 |
* 10_000
|
257 |
/ norm_factor_vector[:, None]
|
258 |
)
|
259 |
+
|
260 |
# tokenize subview gene vectors
|
261 |
tokenized_cells += [
|
262 |
tokenize_cell(subview_norm_array[:, i], coding_miRNA_tokens)
|