forked from chenlingantelope/HarmonizationSCANVI
-
Notifications
You must be signed in to change notification settings - Fork 0
/
NoOverlapSCANVI.py
224 lines (190 loc) · 10.5 KB
/
NoOverlapSCANVI.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
use_cuda = True
from sklearn.neighbors import KNeighborsClassifier
from scvi.harmonization.utils_chenling import get_matrix_from_dir,SubsetGenes
from scvi.dataset.pbmc import PbmcDataset
from scvi.harmonization.utils_chenling import assign_label
import numpy as np
from scipy.stats import entropy
from sklearn.neighbors import NearestNeighbors
from scvi.dataset.dataset import GeneExpressionDataset
from scvi.harmonization.utils_chenling import run_model
from copy import deepcopy
import matplotlib
matplotlib.rcParams['pdf.fonttype'] = 42
matplotlib.rcParams['ps.fonttype'] = 42
import matplotlib.pyplot as plt
import seaborn as sns
from scvi.dataset.dataset import SubsetGenes
from scvi.metrics.clustering import select_indices_evenly
from umap import UMAP
from scvi.models.vae import VAE
from scvi.models.scanvi import SCANVI
from scvi.inference import UnsupervisedTrainer, AlternateSemiSupervisedTrainer
import torch
import os
plotname = 'NoOverlap'
dataset1 = PbmcDataset(filter_out_de_genes=False)
dataset1.update_cells(dataset1.batch_indices.ravel()==0)
newCellType = [k for i, k in enumerate(dataset1.cell_types) if k not in ['Other']]
dataset1.filter_cell_types(newCellType)
dataset1.subsample_genes(dataset1.nb_genes)
count, geneid, cellid = get_matrix_from_dir('cite')
count = count.T.tocsr()
seurat = np.genfromtxt('../cite/cite.seurat.labels', dtype='str', delimiter=',')
cellid = np.asarray([x.split('-')[0] for x in cellid])
labels_map = [0, 0, 1, 2, 3, 4, 5, 6]
labels = seurat[1:, 4]
cell_type = ['CD4 T cells', 'NK cells', 'CD14+ Monocytes', 'B cells','CD8 T cells', 'FCGR3A+ Monocytes', 'Other']
dataset2 = assign_label(cellid, geneid, labels_map, count, cell_type, seurat)
newCellType = [k for i, k in enumerate(dataset2.cell_types) if k not in ['Others']]
dataset2.filter_cell_types(newCellType)
dataset1.subsample_genes(dataset1.nb_genes)
dataset2.subsample_genes(dataset2.nb_genes)
def entropy_from_indices(indices):
return entropy(np.array(np.unique(indices, return_counts=True)[1].astype(np.int32)))
def entropy_batch_mixing_subsampled(latent, batches, labels, removed_type, n_neighbors=50, n_pools=50, n_samples_per_pool=100):
X = latent[labels == removed_type,:]
nbrs = NearestNeighbors(n_neighbors=n_neighbors + 1).fit(latent)
indices = nbrs.kneighbors(X, return_distance=False)[:, 1:]
batch_indices = np.vectorize(lambda i: batches[i])(indices)
entropies = np.apply_along_axis(entropy_from_indices, axis=1, arr=batch_indices)
if n_pools == 1:
res = np.mean(entropies)
else:
res = np.mean([
np.mean(entropies[np.random.choice(len(entropies), size=n_samples_per_pool)])
for _ in range(n_pools)
])
return res
def plotUMAP(latent,plotname,method,rmCellTypes):
colors = sns.color_palette('tab20')
# sample = select_indices_evenly(2000, labels)
sample=np.arange(0,latent.shape[0])
latent_s = latent[sample, :]
label_s = labels[sample]
batch_s = batch_indices[sample]
if latent_s.shape[1] != 2:
latent_s = UMAP(spread=2).fit_transform(latent_s)
fig, ax = plt.subplots(figsize=(10, 10))
key_order = np.argsort(keys)
for i, k in enumerate(key_order):
ax.scatter(latent_s[label_s == k, 0], latent_s[label_s == k, 1], c=colors[i % 20], label=keys[k],
edgecolors='none')
# ax.legend(bbox_to_anchor=(1.1, 0.5), borderaxespad=0, fontsize='x-large')
ax.axis("off")
fig.tight_layout()
plt.savefig('../%s/%s.%s.%s.labels.pdf' % (plotname, plotname, method, rmCellTypes))
plt.figure(figsize=(10, 10))
plt.scatter(latent_s[:, 0], latent_s[:, 1], c=batch_s, edgecolors='none')
plt.axis("off")
plt.tight_layout()
plt.savefig('../%s/%s.%s.%s.batches.pdf' % (plotname, plotname, method, rmCellTypes))
def KNNacc(latent, labels, keys):
neigh = KNeighborsClassifier(n_neighbors=10)
neigh = neigh.fit(latent, labels)
labels_pred = neigh.predict(latent)
acc = [np.mean(labels[labels_pred == i] == i) for i in np.unique(labels)]
cell_type = keys[np.unique(labels)]
return acc, cell_type
def BEbyType(keys,latent,labels,batch_indices):
rm_idx = np.arange(len(keys))[keys == rmCellTypes][0]
other_idx = np.arange(len(keys))[keys != rmCellTypes]
cell_type = [keys[rm_idx]] + list(keys[other_idx])
BE1 = entropy_batch_mixing_subsampled(latent, batch_indices, labels, removed_type=rm_idx)
BE2 = [entropy_batch_mixing_subsampled(latent, batch_indices, labels, removed_type=i) for i in other_idx]
res = [BE1] + BE2
return(res, cell_type)
from sklearn.neighbors import NearestNeighbors
def JaccardIndex(x1,x2):
intersection = np.sum(x1*x2)
union = np.sum((x1+x2)>0)
return intersection/union
def KNNpurity(latent1, latent2,latent,batchid,labels, keys, nn=50):
knn = NearestNeighbors(n_neighbors=nn, algorithm='auto')
nbrs1 = knn.fit(latent1)
nbrs1 = nbrs1.kneighbors_graph(latent1).toarray()
np.fill_diagonal(nbrs1,0)
nbrs2 = knn.fit(latent2)
nbrs2 = nbrs2.kneighbors_graph(latent2).toarray()
np.fill_diagonal(nbrs2,0)
nbrs_1 = knn.fit(latent[batchid==0,:])
nbrs_1 = nbrs_1.kneighbors_graph(latent[batchid==0,:]).toarray()
np.fill_diagonal(nbrs_1,0)
nbrs_2 = knn.fit(latent[batchid==1,:])
nbrs_2 = nbrs_2.kneighbors_graph(latent[batchid==1,:]).toarray()
np.fill_diagonal(nbrs_2,0)
JI1 = [JaccardIndex(x1, x2) for x1, x2 in zip(nbrs1, nbrs_1)]
JI1 = np.asarray(JI1)
JI2 = [JaccardIndex(x1, x2) for x1, x2 in zip(nbrs2, nbrs_2)]
JI2 = np.asarray(JI2)
res1 = [np.mean(JI1[labels[batchid==0] == i]) if np.sum(labels[batchid==0] == i)>0 else 0 for i in np.unique(labels)]
res2 = [np.mean(JI2[labels[batchid==1] == i]) if np.sum(labels[batchid==1] == i)>0 else 0 for i in np.unique(labels)]
res = (np.asarray(res1)+np.asarray(res2))/2
cell_type = keys[np.unique(labels)]
return res,cell_type
def trainVAE(gene_dataset, rmCellTypes,rep):
vae = VAE(gene_dataset.nb_genes, n_batch=gene_dataset.n_batches, n_labels=gene_dataset.n_labels,
n_hidden=128, n_latent=10, n_layers=2, dispersion='gene')
trainer = UnsupervisedTrainer(vae, gene_dataset, train_size=1.0)
if os.path.isfile('../NoOverlap/vae.%s%s.pkl' % (rmCellTypes,rep)):
trainer.model.load_state_dict(torch.load('../NoOverlap/vae.%s%s.pkl' % (rmCellTypes,rep)))
trainer.model.eval()
else:
trainer.train(n_epochs=150)
torch.save(trainer.model.state_dict(), '../NoOverlap/vae.%s%s.pkl' % (rmCellTypes,rep))
full = trainer.create_posterior(trainer.model, gene_dataset, indices=np.arange(len(gene_dataset)))
latent, batch_indices, labels = full.sequential().get_latent()
batch_indices = batch_indices.ravel()
return latent, batch_indices,labels,trainer
f = open('../'+plotname+'/'+plotname+'.kp.res.txt', "w")
g = open('../'+plotname+'/'+plotname+'.res.txt', "w")
# scp [email protected]:/data/yosef2/users/chenling/harmonization/Seurat_data/PopRemove* .
for rmCellTypes in dataset2.cell_types[:6]:
pbmc = deepcopy(dataset1)
newCellType = [k for i, k in enumerate(dataset1.cell_types) if k not in [rmCellTypes]]
pbmc.filter_cell_types(newCellType)
pbmc2 = deepcopy(dataset2)
pbmc2.filter_cell_types([rmCellTypes])
pbmc.subsample_genes(pbmc.nb_genes)
pbmc2.subsample_genes(pbmc2.nb_genes)
gene_dataset = GeneExpressionDataset.concat_datasets(pbmc, pbmc2)
# _,_,_,_,_ = run_model('writedata', gene_dataset, pbmc, pbmc2, filename=plotname+rmCellTypes.replace(' ',''))
latent1 = np.genfromtxt('../Seurat_data/' + plotname+rmCellTypes.replace(' ','') + '.1.CCA.txt')
latent2 = np.genfromtxt('../Seurat_data/' + plotname+rmCellTypes.replace(' ','') + '.2.CCA.txt')
latent, batch_indices, labels, keys, stats = run_model(
'readSeurat', gene_dataset, pbmc, pbmc2, filename=plotname + rmCellTypes.replace(' ', ''))
acc, cell_type = KNNpurity(latent1, latent2,latent,batch_indices.ravel(),labels,keys)
f.write('Seurat' + '\t' + rmCellTypes + ("\t%.4f" * 8 + "\t%s" * 8 + "\n") % tuple(list(acc) + list(cell_type)))
be, cell_type2 = BEbyType(keys, latent, labels, batch_indices)
g.write('Seurat' + '\t' + rmCellTypes + ("\t%.4f" * 8 + "\t%s" * 8 + "\n") % tuple(be + list(cell_type2)))
# plotUMAP(latent, plotname, 'Seurat', rmCellTypes)
pbmc, pbmc2, gene_dataset = SubsetGenes(pbmc, pbmc2, gene_dataset, plotname + rmCellTypes.replace(' ', ''))
# plotUMAP(latent, plotname, 'vae', rmCellTypes)
latent1, _, _, _ = trainVAE(pbmc, rmCellTypes, rep='1')
latent2, _, _, _ = trainVAE(pbmc2, rmCellTypes, rep='2')
latent, batch_indices, labels, trainer = trainVAE(gene_dataset, rmCellTypes, rep='')
acc, cell_type = KNNpurity(latent1, latent2,latent,batch_indices.ravel(),labels,keys)
f.write('vae' + '\t' + rmCellTypes + ("\t%.4f" * 8 + "\t%s" * 8 + "\n") % tuple(list(acc) + list(cell_type)))
be, cell_type2 = BEbyType(keys, latent, labels, batch_indices)
g.write('vae' + '\t' + rmCellTypes + ("\t%.4f" * 8 + "\t%s" * 8 + "\n") % tuple(be + list(cell_type2)))
scanvi = SCANVI(gene_dataset.nb_genes, gene_dataset.n_batches, (gene_dataset.n_labels), n_layers=2)
scanvi.load_state_dict(trainer.model.state_dict(), strict=False)
trainer_scanvi = AlternateSemiSupervisedTrainer(scanvi, gene_dataset, classification_ratio=50,
n_epochs_classifier=10, lr_classification=5 * 1e-3)
trainer_scanvi.labelled_set = trainer_scanvi.create_posterior(indices=gene_dataset.batch_indices.ravel() == 0)
trainer_scanvi.unlabelled_set = trainer_scanvi.create_posterior(indices=gene_dataset.batch_indices.ravel() == 1)
if os.path.isfile('../NoOverlap/scanvi.%s.pkl' % rmCellTypes):
trainer_scanvi.model.load_state_dict(torch.load('../NoOverlap/scanvi.%s.pkl' % rmCellTypes))
trainer_scanvi.model.eval()
else:
trainer_scanvi.train(n_epochs=10)
torch.save(trainer_scanvi.model.state_dict(), '../NoOverlap/scanvi.%s.pkl' % rmCellTypes)
scanvi_full = trainer_scanvi.create_posterior(trainer_scanvi.model, gene_dataset, indices=np.arange(len(gene_dataset)))
latent, _, _ = scanvi_full.sequential().get_latent()
acc, cell_type = KNNpurity(latent1, latent2,latent,batch_indices.ravel(),labels,keys)
f.write('scanvi' + '\t' + rmCellTypes + ("\t%.4f" * 8 + "\t%s" * 8 + "\n") % tuple(list(acc) + list(cell_type)))
be, cell_type2 = BEbyType(keys, latent, labels, batch_indices)
g.write('scanvi' + '\t' + rmCellTypes + ("\t%.4f" * 8 + "\t%s" * 8 + "\n") % tuple(be + list(cell_type2)))
# plotUMAP(latent, plotname, 'scanvi', rmCellTypes)
f.close()
g.close()