-
Notifications
You must be signed in to change notification settings - Fork 0
/
models.py
352 lines (310 loc) · 17.6 KB
/
models.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
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
import os
import time
import copy
import numpy as np
import torch.nn as nn
import torch
import torch.nn.functional as F
from torch.nn import Sequential, ModuleList, Linear, ReLU, BatchNorm1d, Dropout, LogSoftmax
from fingerprint.features import num_atom_features, num_bond_features
from fingerprint.models import NeuralFingerprint
import logging
from transformers import BertTokenizer, BertModel, BertForMaskedLM, BertConfig, BertForPreTraining
from transformers.tokenization_albert import AlbertTokenizer
from transformers.modeling_albert import AlbertModel
from transformers.modeling_albert import AlbertForMaskedLM
from transformers.configuration_albert import AlbertConfig
from transformers.modeling_albert import load_tf_weights_in_albert
from resnet import ResnetEncoderModel, ResnetEncoderSuperTiny
class EmbeddingRNNEncoder(nn.Module):
def __init__(self, vocab_size, embedding_size, hidden_size, num_layers, out_size,
input_dropout_p=0.2, rnn_dropout_p=0.3):
super(EmbeddingRNNEncoder, self).__init__()
self.input_dropout = nn.Dropout(p=input_dropout_p)
self.embed = nn.Embedding(vocab_size, embedding_size)
# dropout on RNN is not effective when it has only l1 layer
self.lstm = nn.LSTM(embedding_size, hidden_size, num_layers, batch_first=True, dropout=rnn_dropout_p)
def cuda(self, device_id=None):
self.lstm.cuda(device_id=device_id)
def forward(self, input_seq):
# TODO: variable length rnn
logging.debug("input_seq {}".format(input_seq.size()))
embedding = self.embed(input_seq)
logging.debug("embedding {}".format(embedding.size()))
embedding = self.input_dropout(embedding)
logging.debug("Apply LSTM...")
context, _ = self.lstm(embedding)
logging.debug("Done. LSTM context size {}".format(context.size()))
return context
class MolecularGraphCoupler(nn.Module):
#chemical embedding is by GraphDegreeCNN
#protein embedding type can be chosen by the user
def __init__(self,
protein_embedding_type='albert', #could be albert, LSTM,
prediction_mode='binary', #could be continuous (e.g. pKi, pIC50)
#protein features - albert
albertconfig=None,
tokenizer=None,
ckpt_path=None,
frozen_list=[22,23],
#protein features - LSTM
lstm_vocab_size=26,
lstm_embedding_size=128,
lstm_hidden_size=64,
lstm_num_layers=3,
lstm_out_size=32,
lstm_input_dropout_p=0.2,
lstm_output_dropout_p=0.3,
#chemical features
conv_layer_sizes=[20,20,20,20],
output_size=128,
degrees=[0,1,2,3,4,5],
#attentive pooler features
ap_hidden_size=64,
ap_dropout=0.1
):
super(MolecularGraphCoupler, self).__init__()
self.prediction_mode = prediction_mode
self.protein_embedding_type = protein_embedding_type
prot_hidden_size = 256
if self.protein_embedding_type == 'albert':
print('==================================================ALBERT')
self.proteinEmbedding = AlbertResnet(albertconfig=albertconfig,
tokenizer=tokenizer,
ckpt_path=ckpt_path,
frozen_list =frozen_list)
self.protein_embedding_type = 'albert'
else:
print('==========================================LSTM: VOCAB SIZE', lstm_vocab_size)
self.proteinEmbedding = EmbeddingRNNEncoder(lstm_vocab_size,
lstm_embedding_size,
lstm_hidden_size,
lstm_num_layers,
lstm_out_size,
input_dropout_p=lstm_input_dropout_p,
rnn_dropout_p=lstm_output_dropout_p)
self.linear_protein_pooler = EmbeddingTransform(210*lstm_hidden_size,128,lstm_hidden_size,dropout_p=ap_dropout)
self.linear_ligand_pooler = EmbeddingTransform(210*output_size,128,output_size,dropout_p=ap_dropout)
self.protein_embedding_type = 'lstm'
prot_hidden_size = lstm_hidden_size
logging.debug("Protein Embedding initialized: {}".format(protein_embedding_type.upper()))
self.ligandEmbedding = ChemicalGraphConv(conv_layer_sizes=conv_layer_sizes,
output_size=output_size,
degrees=degrees,
num_atom_features=num_atom_features(),
num_bond_features=num_bond_features())
self.attentive_interaction_pooler = AttentivePooling(chem_hidden_size=output_size,prot_hidden_size=prot_hidden_size)
self.linear_interaction_pooler = EmbeddingTransform(output_size+prot_hidden_size,128,ap_hidden_size,dropout_p=ap_dropout)
self.binary_predictor = EmbeddingTransform(ap_hidden_size,64,1,dropout_p=0.2)
self.continuous_predictor = EmbeddingTransform(ap_hidden_size,64,1,dropout_p=0.2)
def forward(self, batch_input, **kwargs):
logging.debug("MolecularGraphCoupler: input protein {}".format(batch_input['protein'].size()))
protein_vector = self.proteinEmbedding(batch_input['protein'])
if self.protein_embedding_type == 'lstm':
protein_vector = protein_vector
else:
protein_vector = protein_vector.reshape(batch_input['protein'].size()[0],1,-1)
logging.debug("MolecularGraphCoupler: protein_vector {}".format(protein_vector.size()))
ligand_vector = self.ligandEmbedding(batch_input['ligand'])
logging.debug("MolecularGraphCoupler: ligand_vector {}".format(ligand_vector.size()))
(ligand_vector, ligand_score),(protein_vector, protein_score) = self.attentive_interaction_pooler(ligand_vector,protein_vector)
logging.debug("Attentive pooled ligand {}, protein {}".format(ligand_vector.size(),protein_vector.size()))
if self.protein_embedding_type == 'lstm':
protein_vector = self.linear_protein_pooler(protein_vector.reshape(protein_vector.size()[0],-1))
ligand_vector = self.linear_ligand_pooler(ligand_vector.reshape(ligand_vector.size()[0],-1))
logging.debug("Only for LSTM: Additionally linear pooled ligand {}, protein {}".format(ligand_vector.size(),protein_vector.size()))
interaction_vector = self.linear_interaction_pooler(torch.cat((protein_vector.squeeze(),ligand_vector.squeeze()),1))
logging.debug("interaction_vector {}".format(interaction_vector.size()))
if self.prediction_mode.lower() == 'binary':
logits = self.binary_predictor(interaction_vector)
else:
logits = self.continuous_predictor(interaction_vector)
logging.debug("MolecularGraphCoupler: logits {}".format(logits.size()))
return logits
class AlbertResnet(nn.Module):
def __init__(self, albertconfig=None,
tokenizer=None,
ckpt_path=None, frozen_list=[22,23]):
super(AlbertResnet, self).__init__()
self.config = albertconfig
if self.config is None:
self.config = AlbertConfig.from_pretrained('data/albertdata/albertconfig/albert_config_tiny_google.json')
self.tokenizer = tokenizer
if self.tokenizer is None:
self.tokenizer = BertTokenizer.from_pretrained('data/albertdata/vocab/pfam_vocab_triplets.txt')
if ckpt_path is None:
ckpt_path = 'data/albertdata/pretrained_whole_pfam/model.ckpt-1500000'
self.config.output_hidden_states=False
model = AlbertForMaskedLM(config=self.config)
model = load_tf_weights_in_albert(model,self.config,ckpt_path)
logging.info("Pretrained Albert loaded from {}".format(ckpt_path))
self.albert = model.albert
ct = 0
for m in self.albert.modules():
ct += 1
if ct in frozen_list:
print(frozen_list)
for param in m.parameters():
param.requires_grad = False
else:
for param in m.parameters():
param.requires_grad = True
self.resnet = ResnetEncoderModel(1)
def forward(self, batch_input, **kwargs):
#batch_input is tensor for encoded tokens
logging.debug("AlbertResnet: batch_input {}".format(batch_input.size()))
albert_outputs = self.albert(batch_input)
logging.debug("AlbertResnet: albert_outputs[0] {}".format(albert_outputs[0].size()))
logits = self.resnet(albert_outputs[0].unsqueeze(1))
logging.debug("AlbertResnet: logits {}".format(logits.size()))
return logits
class ChemicalGraphConv(nn.Module):
def __init__(self, conv_layer_sizes=[20,20,20,20],
output_size=128,
degrees=[0,1,2,3,4,5],
num_atom_features=num_atom_features(),
num_bond_features=num_bond_features()):
super(ChemicalGraphConv, self).__init__()
type_map = dict(batch='molecule', node='atom', edge='bond')
self.model = NeuralFingerprint(
num_atom_features,
num_bond_features,
conv_layer_sizes,
output_size,
type_map,
degrees)
for param in self.model.parameters():
param.data.uniform_(-0.08, 0.08)
def forward(self, batch_input, **kwargs):
batch_embedding = self.model(batch_input)
return batch_embedding
class EmbeddingTransform(nn.Module):
def __init__(self, input_size, hidden_size, out_size,
dropout_p=0.1):
super(EmbeddingTransform, self).__init__()
self.dropout = nn.Dropout(p=dropout_p)
self.transform = nn.Sequential(
nn.Linear(input_size, hidden_size),
nn.BatchNorm1d(hidden_size),
nn.ReLU(inplace=True),
nn.Linear(hidden_size, out_size),
nn.BatchNorm1d(out_size)
)
def forward(self, embedding):
embedding = self.dropout(embedding)
hidden = self.transform(embedding)
return hidden
class AttentivePooling(nn.Module):
""" Attentive pooling network according to https://arxiv.org/pdf/1602.03609.pdf """
def __init__(self, chem_hidden_size=128,prot_hidden_size=256):
super(AttentivePooling, self).__init__()
self.chem_hidden_size = chem_hidden_size
self.prot_hidden_size = prot_hidden_size
self.param = nn.Parameter(torch.zeros(chem_hidden_size, prot_hidden_size))
def forward(self, first, second):
""" Calculate attentive pooling attention weighted representation and
attention scores for the two inputs.
Args:
first: output from one source with size (batch_size, length_1, hidden_size)
second: outputs from other sources with size (batch_size, length_2, hidden_size)
Returns:
(rep_1, attn_1): attention weighted representations and attention scores
for the first input
(rep_2, attn_2): attention weighted representations and attention scores
for the second input
"""
logging.debug("AttentivePooling first {0}, second {1}".format(first.size(), second.size()))
param = self.param.expand(first.size(0), self.chem_hidden_size,self.prot_hidden_size)
logging.debug("AttentivePooling params: {0}".format(param.size()))
wm1 = torch.tanh(torch.bmm(second,param.transpose(1,2)))
wm2 = torch.tanh(torch.bmm(first,param))
logging.debug("Wm1 {}, Wm2 {} before softmax".format(wm1.size(),wm2.size()))
score_m1 = F.softmax(wm1,dim=2)
score_m2 = F.softmax(wm2,dim=2)
logging.debug("score_m1 {}, score_m2 {}".format(score_m1.size(),score_m2.size()))
rep_first = first*score_m1
rep_second = second*score_m2
logging.debug("AttentivePooling reps: {0}, {1}".format(rep_first.size(), rep_second.size()))
return ((rep_first, score_m1), (rep_second, score_m2))
class Predict(nn.Module):
""" Prepare a similarity prediction model for each distinct pair of
entity types.
"""
def __init__(self, chem_hidden_size=128, prot_hidden_size=256,
hidden_size=64, attn_dropout=0.1):
super(Predict, self).__init__()
self.hidden_size = hidden_size
self.chem_hidden_size = chem_hidden_size
self.prot_hidden_size = prot_hidden_size
self.add_module('chemical', EmbeddingTransform(self.chem_hidden_size,
hidden_size, hidden_size, dropout_p=attn_dropout))
self.add_module('protein', EmbeddingTransform(self.prot_hidden_size,
hidden_size, hidden_size, dropout_p=attn_dropout)) #for temporal conv
self.add_module(" ".join(('chemical', 'protein')),
AttentivePooling(chem_hidden_size=self.chem_hidden_size,
prot_hidden_size=self.prot_hidden_size))
for param in self.parameters():
param.data.uniform_(-0.08, 0.08)
def forward(self, chem_batch, prot_batch):
""" Calculate the 'similarity' between two inputs, where the first input
is a matrix and the second batched matrices.
Args:
first: output from one source with size (length_1, hidden_size)
second: outputs from other sources with size (batch_size, length_2, hidden_size)
Returns:
prob: a `batch_size` vector that contains the probabilities that each
entity in the second input has association with the first input
"""
logging.debug("AttentivePooling inputs {}, {}".format(chem_batch.size(),prot_batch.size()))
#first, second = sorted((first, second), key=lambda x: x['type'])
attn_model = getattr(self, "chemical protein")
(rep_first, w_first), (rep_second, w_second) = attn_model(chem_batch, prot_batch)
logging.debug("rep_first {}, rep_second {}".format(rep_first.size(),rep_second.size()))
rep_first = getattr(self, 'chemical')(rep_first.squeeze()).unsqueeze(1)
rep_second = getattr(self, 'protein')(rep_second.squeeze()).unsqueeze(2)
logging.debug("Transformed representation vectors: {0}, {1}".format(rep_first.size(), rep_second.size()))
return torch.bmm(rep_first, rep_second).squeeze(), (w_first, w_second)
class PredictBinary(nn.Module):
def __init__(self, chem_hidden_size=128, prot_hidden_size=256,
hidden_size=64, attn_dropout=0.1):
super(PredictBinary, self).__init__()
self.hidden_size = hidden_size
self.chem_hidden_size = chem_hidden_size
self.prot_hidden_size = prot_hidden_size
self.add_module('chemical', EmbeddingTransform(self.chem_hidden_size,
hidden_size, hidden_size, dropout_p=attn_dropout))
self.add_module('protein', EmbeddingTransform(self.prot_hidden_size,
hidden_size, hidden_size, dropout_p=attn_dropout)) #for temporal conv
self.add_module(" ".join(('chemical', 'protein')),
AttentivePooling(chem_hidden_size=self.chem_hidden_size,
prot_hidden_size=self.prot_hidden_size))
self.transform = nn.Sequential(
nn.Linear(128, 128),
nn.BatchNorm1d(128),
nn.ReLU(inplace=True),
nn.Linear(128, 1),
nn.BatchNorm1d(1)
)
for param in self.parameters():
param.data.uniform_(-0.08, 0.08)
def forward(self, chem_batch, prot_batch):
""" Calculate the 'similarity' between two inputs, where the first input
is a matrix and the second batched matrices.
Args:
first: output from one source with size (length_1, hidden_size)
second: outputs from other sources with size (batch_size, length_2, hidden_size)
Returns:
prob: a `batch_size` vector that contains the probabilities that each
entity in the second input has association with the first input
"""
logging.debug("AttentivePooling inputs {}, {}".format(chem_batch.size(),prot_batch.size()))
#first, second = sorted((first, second), key=lambda x: x['type'])
attn_model = getattr(self, "chemical protein")
(rep_first, w_first), (rep_second, w_second) = attn_model(chem_batch, prot_batch)
logging.debug("rep_first {}, rep_second {}".format(rep_first.size(),rep_second.size()))
rep_first = getattr(self, 'chemical')(rep_first.squeeze()).unsqueeze(1)
rep_second = getattr(self, 'protein')(rep_second.squeeze()).unsqueeze(2)
logging.debug("Transformed representation vectors: {0}, {1}".format(rep_first.size(), rep_second.size()))
output = self.transform(torch.cat((rep_first.squeeze(),rep_second.squeeze()),1))
logging.debug("attentive pooling transformation result size {}".format(output.size()))
return output