-
Notifications
You must be signed in to change notification settings - Fork 0
/
construct_data.py
270 lines (208 loc) · 14.3 KB
/
construct_data.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Constructs per-cell training and training/validation data sets for regular training and cross-validation
@author: Christian Guckelsberger, Lassi Paavolainen
"""
import re
import helpers
import shutil
import numpy as np
import pandas as pd
from pathlib import Path
from sklearn import preprocessing
'''
Constructs data for leave-one-out cross-validation.
Takes input feature data from path_data, and outputs the cross-validation files to path_output. These paths should be different in order to not mix up the two types of data.
'''
def construct_data_cross_validation(path_data, path_output, classes, df_controls, df_meta, normalisation_type, antibody, dilution_antibody):
# Accumulate feature data from multiple plates and normalise using per-plate controls
path_list = Path(path_data).glob('**/*.hdf5')
gt_plate = dict()
plate_ids = []
for filepath in path_list:
# Read plate
pat_plate = re.compile(r'([SNRM]_\d+_\d+)')
plate_name = filepath.name.split('.')[0]
plate_id = pat_plate.search(plate_name).groups()[0]
plate_ids.append(plate_id)
print("Reading plate {}".format(plate_id))
df_features = pd.read_hdf(filepath)
# Drop features for consistency
df_features.drop(['NUCLEUS.ELLIPSE-ORIENT'],axis=1,inplace=True,errors='ignore')
# Select metadata for this plate and merge into feature data
df_meta_plate = df_meta[df_meta["Destination Plate Barcode"]==plate_name]
df_features = pd.merge(df_features, df_meta_plate, how='left', on='well')
# Subselect training features. Should have a label assigned to at least one channel.
df_features_labelled = df_features[(df_features['CLASS.IgM']+df_features['CLASS.IgG']+df_features['CLASS.IgA'])>0].copy()
# Filter out rows with NaN's if any
df_features_labelled.dropna(inplace=True)
def sample_count(df, antibody):
col_class = 'CLASS.{}'.format(antibody)
labels, counts = np.unique(df[df[col_class]>0][col_class], return_counts=True)
for i in range(0, len(counts)):
print("%s: %d" % (classes[labels[i]], counts[i]))
# Filtering for specific dilutions
# Print info on class counts for each channel for all and specific dilution
df_features_labelled_dil = df_features_labelled.copy()
dil_antibody = dilution_antibody[antibody]
col_class = 'CLASS.{}'.format(antibody)
df_features_labelled_dil.loc[df_features_labelled_dil.Dilution!=dil_antibody, col_class] = 0
df_features_labelled_dil = df_features_labelled_dil[(df_features_labelled_dil['CLASS.IgM']+df_features_labelled_dil['CLASS.IgG']+df_features_labelled_dil['CLASS.IgA'])>0]
print("\nLabelled samples for antibody {} for dilution {}:".format(antibody, dil_antibody))
sample_count(df_features_labelled_dil, antibody)
features = [c for c in df_features.columns if ("NUCLEUS" in c or "CELL" in c or "DONUT" in c)]
#
# Normalise data for dilutions 25 and 100 separately
print("\nNormalising for dilution {}".format(dil_antibody))
# Flag wells that do not fulfill QC check
flagged_wells = helpers.well_qc(df_features, dilution=dil_antibody)
# Check if sufficient number of controls have been transferred
controls_pos = df_controls[(df_controls.Dilution == dil_antibody) & (df_controls.Type == 'positive')].Content.to_list()
controls_pos_transferred = df_meta_plate.loc[df_meta_plate.Content.isin(controls_pos),['Transferred','well']]
controls_neg = df_controls[(df_controls.Dilution == dil_antibody) & (df_controls.Type == 'negative')].Content.to_list()
controls_neg_transferred = df_meta_plate.loc[df_meta_plate.Content.isin(controls_neg),['Transferred','well']]
helpers.check_sufficient_transfer(controls_pos_transferred, controls_neg_transferred, flagged_wells)
# Fit normalisers on control wells for dilution (i.e. labelled and unlabelled samples)
normalisation_controls_wells = helpers.select_normalisation_wells(normalisation_type, controls_pos_transferred, controls_neg_transferred)
df_features_normalisation = df_features[df_features['well'].isin(normalisation_controls_wells)]
scaler = preprocessing.StandardScaler().fit(df_features_normalisation[features])
# Select labelled data for specific dilution and normalise
df_features_labelled_norm = df_features_labelled_dil[df_features_labelled_dil.Dilution==dil_antibody].copy()
df_features_labelled_norm[features] = scaler.transform(df_features_labelled_norm[features])
# Store for this specific plate and dilution
gt_plate[plate_id] = df_features_labelled_norm
print()
# Construct and export leave-one-plate-out-cross-validation data
# For N input data plates, we have N folds of N-1 training and 1 validation set
print("Exporting cv data for dilution {} and antibody {}".format(str(dil_antibody), antibody))
# Determine relevant featurees for channel and dilution: DAPI, per-channel ALEXA, class label
# Also drop NUCLEUS.NUMBER.OF.COMPONENTS column as 0 for all samples
features_channel = [col for col in df_features.columns if (('NUCLEUS' in col or 'CELL' in col or 'DONUT' in col) and not ('Alexa' in col) and not ('NUCLEUS.NUMBER.OF.COMPONENTS' in col))]
features_channel = features_channel + [col for col in df_features.columns if str(antibody) in col]
# Sort by columns alphabetically. To make sure that columns have same order everywhere (training, prediction)
features_channel.sort()
# Leave one plate out for validation, use others for training
col_class = 'CLASS.{}'.format(antibody)
for validation_plate in plate_ids:
df_validation_dil = gt_plate[validation_plate]
df_validation_dil = df_validation_dil[df_validation_dil[col_class]!=0]
df_train_dil = pd.DataFrame()
train_plates = []
for plate_id in gt_plate.keys():
if plate_id != validation_plate:
train_plates.append(plate_id)
df_train_plate = gt_plate[plate_id]
df_train_plate = df_train_plate[df_train_plate[col_class]!=0]
df_train_dil = df_train_dil.append(df_train_plate)
# Filter for relevant features for this channel
df_validation_dil = df_validation_dil[features_channel]
df_train_dil = df_train_dil[features_channel]
# Split up data into classes and drop class label colum
df_validation_dil_class = dict()
df_train_dil_class = dict()
for label in classes.keys():
df_validation_dil_class[label] = df_validation_dil[df_validation_dil[col_class]==label].drop([col_class], axis=1)
df_train_dil_class[label] = df_train_dil[df_train_dil[col_class]==label].drop([col_class], axis=1)
# Combine classes into datasets
X_validate, X_train = pd.DataFrame(), pd.DataFrame()
y_validate, y_train = pd.DataFrame(), pd.DataFrame()
for label in classes.keys():
X_validate = pd.concat([X_validate, df_validation_dil_class[label]], axis=0, ignore_index=True)
X_train = pd.concat([X_train, df_train_dil_class[label]], axis=0, ignore_index=True)
y_validate = pd.concat([y_validate, pd.DataFrame({'Label':np.full(df_validation_dil_class[label].shape[0], label, dtype=np.int8)})])
y_train = pd.concat([y_train, pd.DataFrame({'Label':np.full(df_train_dil_class[label].shape[0], label, dtype=np.int8)})])
print("Leaving out plate {}. Size of train and validation sets: {}, {}".format(validation_plate, X_train.shape[0], X_validate.shape[0]))
# Export data. Produce two hdf files (X,y) with 2 groups each (train, validate)
Path(path_output).mkdir(parents=True, exist_ok=True)
filename = path_output / 'leave_{}_out_train_on_{}_ab_{}_and_DAPI_dil_{}.hdf5'.format(validation_plate, "_and_".join(train_plates), antibody, dil_antibody)
X_train.to_hdf(filename, key='X_train', mode='w')
X_validate.to_hdf(filename, key='X_validate')
y_train.to_hdf(filename, key='y_train')
y_validate.to_hdf(filename, key='y_validate')
'''
Constructs a single training data set (no train/validation split)
Takes input feature data from path_data, and outputs the training data files to path_output. These paths should be different in order to not mix up the two types of data.
'''
def construct_training_data(path_data, path_output, classes, df_controls, df_meta, normalisation_type, antibody, dilution_antibody):
# Accumulate feature data from multiple plates and normalise using per-plate controls
path_list = Path(path_data).glob('**/*.hdf5')
gt = pd.DataFrame()
for filepath in path_list:
# Read plate
pat_plate = re.compile(r'([SNRM]_\d+_\d+)')
plate_name = filepath.name.split('.')[0]
plate_id = pat_plate.search(plate_name).groups()[0]
print("Reading plate {}".format(plate_id))
df_features = pd.read_hdf(filepath)
# Drop features for consistency
df_features.drop(['NUCLEUS.ELLIPSE-ORIENT'],axis=1,inplace=True,errors='ignore')
# Select metadata for this plate and merge into feature data
df_meta_plate = df_meta[df_meta["Destination Plate Barcode"]==plate_id]
df_features = pd.merge(df_features, df_meta_plate, how='left', on='well')
# Subselect training features. Should have a label assigned to at least one channel.
df_features_labelled = df_features[(df_features['CLASS.IgM']+df_features['CLASS.IgG']+df_features['CLASS.IgA'])>0].copy()
# Filter out rows with NaN's if any
df_features_labelled.dropna(inplace=True)
def sample_count(df, antibody):
col_class = 'CLASS.{}'.format(antibody)
labels, counts = np.unique(df[df[col_class]>0][col_class], return_counts=True)
for i in range(0, len(counts)):
print("%s: %d" % (classes[labels[i]], counts[i]))
# Filtering for specific dilutions
# Print info on class counts for each channel for all and specific dilution
df_features_labelled_dil = df_features_labelled.copy()
dil_antibody = dilution_antibody[antibody]
col_class = 'CLASS.{}'.format(antibody)
df_features_labelled_dil.loc[df_features_labelled_dil.Dilution!=dil_antibody, col_class] = 0
df_features_labelled_dil = df_features_labelled_dil[(df_features_labelled_dil['CLASS.IgM']+df_features_labelled_dil['CLASS.IgG']+df_features_labelled_dil['CLASS.IgA'])>0]
print("\nLabelled samples for antibody {} for dilution {}:".format(antibody, dil_antibody))
sample_count(df_features_labelled_dil, antibody)
features = [c for c in df_features.columns if ("NUCLEUS" in c or "CELL" in c or "DONUT" in c)]
# Normalise data for dilutions 25 and 100 separately
d = dilution_antibody[antibody]
print("\nNormalising for dilution {}".format(d))
# Flag wells that do not fulfill QC check
flagged_wells = helpers.well_qc(df_features, dilution=d)
# Check if sufficient number of controls have been transferred
controls_pos = df_controls[(df_controls.Dilution == d) & (df_controls.Type == 'positive')].Content.to_list()
controls_pos_transferred = df_meta_plate.loc[df_meta_plate.Content.isin(controls_pos),['Transferred','well']]
controls_neg = df_controls[(df_controls.Dilution == d) & (df_controls.Type == 'negative')].Content.to_list()
controls_neg_transferred = df_meta_plate.loc[df_meta_plate.Content.isin(controls_neg),['Transferred','well']]
helpers.check_sufficient_transfer(controls_pos_transferred, controls_neg_transferred, flagged_wells)
# Fit normalisers on control wells for dilution (i.e. labelled and unlabelled samples)
normalisation_controls_wells = helpers.select_normalisation_wells(normalisation_type, controls_pos_transferred, controls_neg_transferred)
df_features_normalisation = df_features[df_features['well'].isin(normalisation_controls_wells)]
scaler = preprocessing.StandardScaler().fit(df_features_normalisation[features])
# Select labelled data for specific dilution and normalise
df_features_labelled_norm = df_features_labelled_dil[df_features_labelled_dil.Dilution==d].copy()
df_features_labelled_norm[features] = scaler.transform(df_features_labelled_norm[features])
# Append to existing data
gt = gt.append(df_features_labelled_norm)
print()
# Construct and export training data
print("Exporting training data for dilution {} and antibody {}".format(str(dil_antibody), antibody))
# Extract labelled data for channel and dilution: DAPI, per-channel ALEXA, class label
# Also drop NUCLEUS.NUMBER.OF.COMPONENTS column as 0 for all samples
features_channel = [col for col in df_features.columns if (('NUCLEUS' in col or 'CELL' in col or 'DONUT' in col) and not ('Alexa' in col) and not ('NUCLEUS.NUMBER.OF.COMPONENTS' in col))]
features_channel = features_channel + [col for col in df_features.columns if str(antibody) in col]
df_features_labels = gt[features_channel]
# Split up data into classes and drop class label column
col_class = 'CLASS.{}'.format(antibody)
df_class = dict()
for label in classes.keys():
df_class[label] = df_features_labels[df_features_labels[col_class]==label].drop([col_class], axis=1)
# Combine classes into datasets
X = pd.DataFrame()
y = pd.DataFrame()
for label in classes.keys():
X = pd.concat([X, df_class[label]], axis=0, ignore_index=True)
y = pd.concat([y, pd.DataFrame({'Label':np.full(df_class[label].shape[0], label, dtype=np.int8)})])
# Sort by columns alphabetically. To make sure that columns have same order everywhere (training, prediction)
X.sort_index(axis=1, inplace=True)
# Save to harddrive
Path(path_output).mkdir(parents=True, exist_ok=True)
filename = path_output / 'train_ab_{}_and_DAPI_dil_{}.hdf5'.format(antibody, dil_antibody)
X.to_hdf(filename, key='X', mode='w')
y.to_hdf(filename, key='y')
print()