/
deepnd_st.py
316 lines (269 loc) · 18 KB
/
deepnd_st.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
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
"""
deepnd_st.py
Training and testing processes for DeepND ST
Bilkent University, Department of Computer Engineering
Ankara, 2020
"""
import numpy as np
import pandas as pd
import csv
import pickle
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch_geometric.nn import GCNConv
from torch_geometric.data import Data
from torch.autograd import Variable
from sklearn.metrics import average_precision_score, roc_auc_score
import time
from models import *
from utils import *
def deepnd_st(root, path, input_size, mode, l_rate, trial, k, disordername , devices, pfcgpumask, mdcbcgpumask, shagpumask, v1cgpumask, state, experiment, networks):
init_time = time.time()
network_count = len(networks)
geneNames_all = pd.read_csv(root + "Data/row-genes.txt", header = None)
geneNames_all = geneNames_all[0].tolist()
geneDict = constructGeneDictionary(root + "Data/hugogenes_entrez.txt")
gene_names_list = [str(item) for item in geneNames_all]
# GOLD STANDARDS
# Following section loads gold standard genes
# To use other standards, following section needs to be changed
if disordername == "ID":
# ID Validation
g_bs_tada_intersect_indices, n_bs_tada_intersect_indices, y, pos_gold_std_genes, neg_gold_std_genes, pos_gold_std_evidence, gold_evidence = load_goldstandards(root, geneNames_all, geneDict, disordername = "ID")
else:
# ASD Validation
g_bs_tada_intersect_indices, n_bs_tada_intersect_indices, y, pos_gold_std_genes, neg_gold_std_genes, pos_gold_std_evidence, gold_evidence = load_goldstandards(root, geneNames_all, geneDict, disordername = "ASD")
# VALIDATION SETS
e1_gene_indices, e1_perm, e2_gene_indices, e2_perm, e3e4_gene_indices, e3e4_perm, neg_perm, counts = create_validation_set( g_bs_tada_intersect_indices, n_bs_tada_intersect_indices, gold_evidence, k = 5, state = state)
# FEATURES
if diseasename == "ID":
data, features = loadFeatures(root, y, geneNames_all, devices, disordername = "ID")
else:
data, features = loadFeatures(root, y, geneNames_all, devices, disordername = "ASD")
# NETWORKS
pfcnetworks, pfcnetworkweights, mdcbcnetworks, mdcbcnetworkweights, v1cnetworks, v1cnetworkweights, shanetworks, shanetworkweights = load_networks(root, devices, pfcgpumask, mdcbcgpumask, shagpumask, v1cgpumask, mask = networks)
# MODEL CONSTRUCTION
model = DeepND_ST(devices, pfcgpumask, mdcbcgpumask, shagpumask, v1cgpumask, featsize=input_size, unit=input_size)
average_att = []
stddev_att = []
average_att_gold = []
stddev_att_gold = []
average_att_gold_e1 = []
average_att_gold_e1e2 = []
average_att_gold_neg = []
all_att = []
pre_att = []
for i in range(network_count * 4):
average_att.append(0.0)
average_att_gold.append(0.0)
stddev_att.append(0.0)
stddev_att_gold.append(0.0)
average_att_gold_e1.append(0.0)
average_att_gold_e1e2.append(0.0)
average_att_gold_neg.append(0.0)
all_att.append(0.0)
pre_att.append(0.0)
aucs = []
aupr = []
predictions = torch.zeros((len(geneNames_all),1), dtype = torch.float)
usage, cached = memoryUpdate()
# Early Stop Configuration
early_stop_enabled = True
old_loss = 100
early_stop_window = 7
epoch_count = []
for j in range(trial): # 10 here means Run count. Run given times and calculate average AUC found from each run.
print("Trial:", j+1)
# Losses
tloss=[]
vloss=[]
fpr = dict()
tpr = dict()
usege, cached = memoryUpdate(usage, cached)
for k1 in range(k):
e1mask = [e1_gene_indices[index] for index in e1_perm[k1 * math.ceil(counts[0]/k): min(counts[0], (k1 + 1) * math.ceil(counts[0]/k)) ] ]
data.e1mask = e1mask.copy()
negmask = [n_bs_tada_intersect_indices[item] for item in neg_perm[k1 * math.ceil(counts[3]/k) : min(counts[3] , (k1 + 1) * math.ceil(counts[3]/k))] ]
data.negmask = negmask.copy()
test_mask = [e1_gene_indices[index] for index in e1_perm[k1 * math.ceil(counts[0]/k): min(counts[0], (k1 + 1) * math.ceil(counts[0]/k)) ] ]
test_mask += [n_bs_tada_intersect_indices[item] for item in neg_perm[k1 * math.ceil(counts[3]/k) : min(counts[3] , (k1 + 1) * math.ceil(counts[3]/k))] ]
data.test_mask = test_mask.copy()
print("Test Mask Length After E1:", len(test_mask))
print('Test Gene(s):', [gene_names_list[i] for i in test_mask])
test_mask += [e2_gene_indices[index] for index in e2_perm[(k1) * math.ceil(counts[1]/k): min(counts[1], (k1 + 1) * math.ceil(counts[1]/k)) ] ]
test_mask += [e3e4_gene_indices[index] for index in e3e4_perm[(k1) * math.ceil(counts[2]/k): min(counts[2], (k1 + 1) * math.ceil(counts[2]/k)) ] ]
k_e1_perm = np.delete(e1_perm,np.s_[k1*math.ceil(counts[0]/k):min(counts[0],(k1 + 1) * math.ceil(counts[0]/k))],axis=0)
k_neg_perm = np.delete(neg_perm,np.s_[k1 * math.ceil(counts[3]/k): min(counts[3], (k1 + 1) * math.ceil(counts[3]/k)) ],axis=0)
k_e2_perm = np.delete(e2_perm,np.s_[k1 * math.ceil(counts[1]/k): min(counts[1], (k1 + 1) * math.ceil(counts[1]/k)) ],axis=0)
k_e3e4_perm = np.delete(e3e4_perm,np.s_[k1 * math.ceil(counts[2]/k): min(counts[2], (k1 + 1) * math.ceil(counts[2]/k)) ],axis=0)
for k2 in range(k-1): # K-FOLD Cross Validation
print("Fold", k1+1, "_", k2+1, "of Trial", j+1)
validation_mask = [e1_gene_indices[index] for index in k_e1_perm[k2 * math.ceil(counts[0]/k): min(counts[0], (k2 + 1) * math.ceil(counts[0]/k)) ] ]
# Add negative genes to validation mask
validation_mask += [n_bs_tada_intersect_indices[item] for item in k_neg_perm[k2 * math.ceil(counts[3]/k) : min(counts[3] , (k2 + 1) * math.ceil(counts[3]/k))] ]
data.auc_mask = validation_mask.copy()
print('Validation Gene(s):', [gene_names_list[i] for i in validation_mask])
validation_mask += [e2_gene_indices[index] for index in k_e2_perm[(i) * math.ceil(counts[1]/k): min(counts[1], (i + 1) * math.ceil(counts[1]/k)) ] ]
validation_mask += [e3e4_gene_indices[index] for index in k_e3e4_perm[(i) * math.ceil(counts[2]/k): min(counts[2], (i + 1) * math.ceil(counts[2]/k)) ] ]
# Construct Train Mask
train_mask = g_bs_tada_intersect_indices + n_bs_tada_intersect_indices
print("Total Gene Count:", len(train_mask))
train_mask = [item for item in train_mask if item not in sorted(validation_mask + test_mask)]
print("Final Validation Mask Length:", len(validation_mask))
print("Final AUC Mask Length:", len(data.auc_mask))
print("Final Train Mask Length:", len(train_mask))
print("AUC Mask Length:", len(data.auc_mask))
#Uncomment line below if you want to use sample weights
sample_weights = torch.ones((len(train_mask)), dtype = torch.float).to(devices[0])
# Uncomment the loop below if you want to use sample weights
for index, value in enumerate(train_mask):
if value in g_bs_tada_intersect_indices:
index2 = g_bs_tada_intersect_indices.index(value)
evidence = pos_gold_std_evidence[index2]
if evidence == "E2":
sample_weights[index] = 0.5
elif evidence == "E3" or evidence == "E4":
sample_weights[index] = 0.25
data.train_mask = torch.tensor(train_mask, dtype= torch.long)
data.validation_mask = torch.tensor(validation_mask, dtype=torch.long)
if mode:
#Test Mode
model.load_state_dict(torch.load(root + disordername + "Exp" + str(experiment) + "/deepND_ST_" + disordername + "_trial"+str(j+1)+"_fold"+str(k1+1)+"_"+str(k2+1)+".pth"))
model = model.eval()
with torch.no_grad():
out = model(features, features, pfcnetworks, mdcbcnetworks, v1cnetworks, shanetworks, pfcnetworkweights, mdcbcnetworkweights, v1cnetworkweights, shanetworkweights, devices, pfcgpumask, mdcbcgpumask, v1cgpumask, shagpumask)
_, pred = out.max(dim=1)
correct = pred[data.auc_mask].eq(data.y[data.auc_mask]).sum().item()
correctTrain = pred[data.train_mask].eq(data.y[data.train_mask]).sum().item()
acc = correct / len(data.auc_mask)
accTrain = correctTrain / len(data.train_mask)
valLoss = (F.nll_loss(out[data.auc_mask], data.y[data.auc_mask])).to(devices[0])
vloss.append(valLoss.cpu().item())
else:
model.apply(weight_reset)
optimizer = torch.optim.Adam(model.parameters(), lr=l_rate, weight_decay=0.0001)
for epoch in range(1000):
model = model.train()
optimizer.zero_grad()
out = model(features, features, pfcnetworks, mdcbcnetworks, v1cnetworks, shanetworks, pfcnetworkweights, mdcbcnetworkweights, v1cnetworkweights, shanetworkweights, devices, pfcgpumask, mdcbcgpumask, v1cgpumask, shagpumask)
loss = (F.nll_loss(out[data.train_mask], data.y[data.train_mask], weight = torch.FloatTensor([1.0, 1.0]).to(devices[0]))).to(devices[0]) # You can adjust class weights using values in FloatTensor
#Uncomment section below and comment out 3 lines above to enable sample weights.
loss = loss * sample_weights
loss.mean().backward()
tloss.append(loss.mean().cpu().item())
optimizer.step()
model = model.eval()
with torch.no_grad():
out = model(features, features, pfcnetworks, mdcbcnetworks, v1cnetworks, shanetworks, pfcnetworkweights, mdcbcnetworkweights, v1cnetworkweights, shanetworkweights, devices, pfcgpumask, mdcbcgpumask, v1cgpumask, shagpumask)
_, pred = out.max(dim=1)
correct = pred[data.auc_mask].eq(data.y[data.auc_mask]).sum().item()
correctTrain = pred[data.train_mask].eq(data.y[data.train_mask]).sum().item()
acc = correct / len(data.auc_mask)
accTrain = correctTrain / len(data.train_mask)
valLoss = (F.nll_loss(out[data.auc_mask], data.y[data.auc_mask])).to(devices[0])
vloss.append(valLoss.cpu().item())
if epoch != 0 and epoch % 25 == 0:
print('Validation Accuracy: {:.4f}, Validation Loss: {:.4f} Train Accuracy: {:.4f} Train Loss: {:.4f}'.format(acc,valLoss, accTrain, loss.mean().item()))
# Early Stop Checks
if early_stop_enabled:
if valLoss < old_loss:
early_stop_count = 0
old_loss = valLoss
else:
early_stop_count += 1
if early_stop_count == early_stop_window:
print("Epoch:", epoch, ", Loss:",loss.mean().item())
break
torch.save(model.state_dict(), path + "/deepND_ST_" + disordername + "_trial"+str(j+1)+"_fold"+str(k1+1)+"_"+str(k2+1)+".pth")
# -------------------------------------------------------------
adjusted_mean_scores = (torch.exp(out.cpu()))[:,1]
adjusted_mean_scores[data.train_mask] = 0.0
adjusted_mean_scores[data.auc_mask] = 0.0
predictions[:,0] += adjusted_mean_scores
# -------------------------------------------------------------
area_under_roc = roc_auc_score(data.y.cpu()[data.test_mask],(F.softmax(out.cpu()[data.test_mask, :],dim=1))[:,1])
aucs.append(area_under_roc)
print("AUC", area_under_roc)
average_precision = average_precision_score(data.y.cpu()[data.test_mask],(F.softmax(out.cpu()[data.test_mask, :],dim=1))[:,1])
aupr.append(average_precision)
print('Average precision-recall score: {0:0.2f}'.format(average_precision))
for i in range(network_count * 4):
average_att[i] += torch.mean(model.experts[:,i]).item()
stddev_att[i] += model.experts[:,i].std().item()
average_att_gold[i] += torch.mean(model.experts[g_bs_tada_intersect_indices,i]).item()
stddev_att_gold[i] += model.experts[g_bs_tada_intersect_indices,i].std().item()
average_att_gold_e1[i] += torch.mean(model.experts[g_bs_tada_intersect_indices[0:18],i]).item()
average_att_gold_e1e2[i] += torch.mean(model.experts[g_bs_tada_intersect_indices[0:49],i]).item()
average_att_gold_neg[i] += torch.mean(model.experts[n_bs_tada_intersect_indices,i]).item()
att_leak_prevention = model.experts[:,i]
att_leak_prevention[data.train_mask] = 0.0
att_leak_prevention[data.validation_mask] = 0.0
all_att[i] += att_leak_prevention
pre_att_buffer = model.expert_results[i]
pre_att_buffer[data.train_mask] = 0.0
pre_att_buffer[data.validation_mask] = 0.0
pre_att[i] += pre_att_buffer
# -------------------------------------------------------------
print("-"*10)
print(disordername+" Trial Median AUC:" + str(np.median(aucs[-int((k*k-1)):])))
print(disordername+" Trial Median AUPR:" + str(np.median(aupr[-(k*k-1):])))
print("-"*10)
print(disordername+" Current Median AUC:" + str(np.median(aucs)))
print(disordername+" Current Median AUPR:" + str(np.median(aupr)))
print("-"*80)
###############################################################################################################################################
# Writing Final Result of the Session
writePrediction(predictions, g_bs_tada_intersect_indices, n_bs_tada_intersect_indices, pos_gold_std_genes, neg_gold_std_genes, geneDict, geneNames_all, path = path, disordername = disordername, trial = trial, k = k)
writeExperimentStats( aucs, aupr, path = path, disordername = disordername, trial = trial, k = k, init_time = init_time, network_count = network_count , mode = mode)
# HEATMAPS
heatmap = torch.zeros(4, network_count,dtype=torch.float)
heatmap2 = torch.zeros(4, network_count,dtype=torch.float)
heatmap3 = torch.zeros(4, network_count,dtype=torch.float)
heatmap4 = torch.zeros(4, network_count,dtype=torch.float)
heatmap5 = torch.zeros(4, network_count,dtype=torch.float)
heatmap6 = torch.zeros(4, network_count,25825, dtype=torch.float)
heatmap7 = torch.zeros(25825, network_count * 4, dtype=torch.float)
heatmap8 = torch.zeros(25825, network_count * 4, dtype=torch.float)
#heatmap[0,:] = average_att[:7]
for i in range(network_count):
heatmap[0,i] = average_att[i]
heatmap2[0,i] = average_att_gold[i]
heatmap3[0,i] = average_att_gold_e1[i]
heatmap4[0,i] = average_att_gold_e1e2[i]
heatmap5[0,i] = average_att_gold_neg[i]
heatmap6[0,i,:] = all_att[i]
#heatmap[1,2:] = average_att[7:12]
for i in range(network_count, network_count * 2):
heatmap[1,i - network_count] = average_att[i]
heatmap2[1,i - network_count] = average_att_gold[i]
heatmap3[1,i - network_count] = average_att_gold_e1[i]
heatmap4[1,i - network_count] = average_att_gold_e1e2[i]
heatmap5[1,i - network_count] = average_att_gold_neg[i]
heatmap6[1,i-network_count,:] = all_att[i]
#heatmap[2,:] = average_att[12:19]
for i in range(network_count * 2, network_count * 3):
heatmap[2,i - network_count * 2] = average_att[i]
heatmap2[2,i - network_count * 2] = average_att_gold[i]
heatmap3[2,i - network_count * 2] = average_att_gold_e1[i]
heatmap4[2,i - network_count * 2] = average_att_gold_e1e2[i]
heatmap5[2,i - network_count * 2] = average_att_gold_neg[i]
heatmap6[2,i-network_count*2,:] = all_att[i]
#heatmap[3,:] = average_att[19:26]
for i in range(network_count * 3,network_count * 4):
heatmap[3,i - network_count * 3] = average_att[i]
heatmap2[3,i - network_count * 3] = average_att_gold[i]
heatmap3[3,i - network_count * 3] = average_att_gold_e1[i]
heatmap4[3,i - network_count * 3] = average_att_gold_e1e2[i]
heatmap5[3,i - network_count * 3] = average_att_gold_neg[i]
heatmap6[3,i-network_count*3,:] = all_att[i]
torch.save(heatmap, path + "/heatmap_tensor.pt");
torch.save(heatmap2, path + "/heatmap_gold_tensor.pt");
torch.save(heatmap3, path + "/heatmap_gold_e1_tensor.pt");
torch.save(heatmap4, path + "/heatmap_gold_e1e2_tensor.pt");
torch.save(heatmap5, path + "/heatmap_gold_neg_tensor.pt");
torch.save(heatmap6, path + "/heatmap_all.pt");
heatmap7 = all_att
torch.save(heatmap7, path + "/heatmap_flat_all.pt");
heatmap8 = pre_att
torch.save(heatmap8, path + "/heatmap_pre_att_all.pt");