-
Notifications
You must be signed in to change notification settings - Fork 5
/
prott5_batch_predictor.py
1406 lines (1151 loc) · 62.2 KB
/
prott5_batch_predictor.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
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Mon Jun 20 11:17:06 2022
@author: mheinzinger
"""
import argparse
import time
from pathlib import Path
from typing import Union
import numpy as np
import torch
from torch import nn
from urllib import request
import shutil
# Huggingface Transformer models:
from transformers import T5EncoderModel, T5Tokenizer
from transformers import GPT2LMHeadModel , GPT2Tokenizer
# TMbed specifics:
from tmbed_predictor import Predictor
from tmbed_viterbi import Decoder
from tqdm import tqdm
def get_device(device: Union[None, str, torch.device] = None) -> torch.device:
"""Returns what the user specified, or defaults to the GPU,
with a fallback to CPU if no GPU is available."""
if isinstance(device, torch.device):
return device
elif device:
return torch.device(device)
elif torch.cuda.is_available():
return torch.device("cuda")
else:
return torch.device("cpu")
# At the beginning, there was no device
device = None
#### Architectures
# ProtTucker predicts structural classes according to CATH
# https://github.com/Rostlab/EAT/blob/main/eat.py#L25
class TuckerFNN(nn.Module):
def __init__(self):
super(TuckerFNN, self).__init__()
self.tucker = nn.Sequential(
nn.Linear(1024, 256),
nn.Tanh(),
nn.Linear(256, 128),
)
def single_pass(self, x):
return self.tucker(x)
def forward(self, X): # only needed during training
ancor = self.single_pass(X[:, 0, :])
pos = self.single_pass(X[:, 1, :])
neg = self.single_pass(X[:, 2, :])
return (ancor, pos, neg)
# LightAttention for subcellular localization prediction:
# https://github.com/HannesStark/protein-localization/blob/master/models/light_attention.py
class LightAttention(nn.Module):
def __init__(self, embeddings_dim=1024, output_dim=11, dropout=0.25, kernel_size=9, conv_dropout: float = 0.25):
super(LightAttention, self).__init__()
self.feature_convolution = nn.Conv1d(embeddings_dim, embeddings_dim, kernel_size, stride=1,
padding=kernel_size // 2)
self.attention_convolution = nn.Conv1d(embeddings_dim, embeddings_dim, kernel_size, stride=1,
padding=kernel_size // 2)
self.softmax = nn.Softmax(dim=-1)
self.dropout = nn.Dropout(conv_dropout)
self.linear = nn.Sequential(
nn.Linear(2 * embeddings_dim, 32),
nn.Dropout(dropout),
nn.ReLU(),
nn.BatchNorm1d(32)
)
self.output = nn.Linear(32, output_dim)
def forward(self, x: torch.Tensor, mask, **kwargs) -> torch.Tensor:
"""
Args:
x: [batch_size, embeddings_dim, sequence_length] embedding tensor that should be classified
mask: [batch_size, sequence_length] mask corresponding to the zero padding used for the shorter sequecnes in the batch. All values corresponding to padding are False and the rest is True.
Returns:
classification: [batch_size,output_dim] tensor with logits
"""
o = self.feature_convolution(x) # [batch_size, embeddings_dim, sequence_length]
o = self.dropout(o) # [batch_gsize, embeddings_dim, sequence_length]
attention = self.attention_convolution(x) # [batch_size, embeddings_dim, sequence_length]
# mask out the padding to which we do not want to pay any attention (we have the padding because the sequences have different lenghts).
# This padding is added by the dataloader when using the padded_permuted_collate function in utils/general.py
attention = attention.masked_fill(mask[:, None, :] == False, torch.tensor(-1e+4))
# code used for extracting embeddings for UMAP visualizations
# extraction = torch.sum(x * self.softmax(attention), dim=-1)
# extraction = self.id0(extraction)
o1 = torch.sum(o * self.softmax(attention), dim=-1) # [batchsize, embeddings_dim]
o2, _ = torch.max(o, dim=-1) # [batchsize, embeddings_dim]
o = torch.cat([o1, o2], dim=-1) # [batchsize, 2*embeddings_dim]
o = self.linear(o) # [batchsize, 32]
return self.output(o) # [batchsize, output_dim]
class BindPredictCNN(torch.nn.Module):
# Parameters:
# https://github.com/Rostlab/bindPredict/blob/e9f1f33c5b614966fbf7d85b79f856b68ca495ad/develop_bindEmbed21DL.py#L34
def __init__(self, in_channels=1024, feature_channels=128, kernel_size=5, stride=1, padding=2, dropout=0.7):
super(BindPredictCNN, self).__init__()
self.conv1 = torch.nn.Sequential(
torch.nn.Conv1d(in_channels=in_channels, out_channels=feature_channels, kernel_size=kernel_size,
stride=stride, padding=padding),
torch.nn.ELU(),
torch.nn.Dropout(dropout),
torch.nn.Conv1d(in_channels=feature_channels, out_channels=3, kernel_size=kernel_size,
stride=stride, padding=padding),
)
def forward(self, x):
x = self.conv1(x)
return torch.squeeze(x)
# CNN for secondary structure prediction in 3- and 8-states (see ProtTrans)
# 3- and 8-states refer to DSSP definitions
class SecStructCNN( nn.Module ):
def __init__( self ):
super(SecStructCNN, self).__init__()
# This is only called "elmo_feature_extractor" for historic reason
# CNN weights are trained on ProtT5 embeddings
self.elmo_feature_extractor = nn.Sequential(
nn.Conv2d( 1024, 32, kernel_size=(7,1), padding=(3,0) ), # 7x32
nn.ReLU(),
nn.Dropout( 0.25 ),
)
n_final_in = 32
self.dssp3_classifier = torch.nn.Sequential(
nn.Conv2d( n_final_in, 3, kernel_size=(7,1), padding=(3,0)) # 7
)
self.dssp8_classifier = torch.nn.Sequential(
nn.Conv2d( n_final_in, 8, kernel_size=(7,1), padding=(3,0))
)
self.diso_classifier = torch.nn.Sequential(
nn.Conv2d( n_final_in, 2, kernel_size=(7,1), padding=(3,0))
)
def forward( self, x):
# IN: X = (B x L x F); OUT: (B x F x L, 1)
x = x.permute(0,2,1).unsqueeze(dim=-1)
x = self.elmo_feature_extractor(x) # OUT: (B x 32 x L x 1)
d3_Yhat = self.dssp3_classifier( x ).squeeze(dim=-1).permute(0,2,1) # OUT: (B x L x 3)
d8_Yhat = self.dssp8_classifier( x ).squeeze(dim=-1).permute(0,2,1) # OUT: (B x L x 8)
return d3_Yhat, d8_Yhat
# CNN for conservation prediction in 9 classes [0-8] (see VESPA).
# See ConSeq for conservation class definitions. 0=highly_variable; 8=highly_conserved
class ConservationCNN( nn.Module ):
def __init__( self, pretrained_model=None ):
super(ConservationCNN, self).__init__()
n_features= 1024
bottleneck_dim = 32
n_classes = 9
dropout_rate = 0.25
self.classifier = nn.Sequential(
nn.Conv2d( n_features, bottleneck_dim, kernel_size=(7,1), padding=(3,0) ), # 7x32
nn.ReLU(),
nn.Dropout( dropout_rate ),
nn.Conv2d( bottleneck_dim, n_classes, kernel_size=(7,1), padding=(3,0))
)
def forward( self, x):
'''
L = protein length
B = batch-size
F = number of features (1024 for embeddings)
N = number of classes (9 for conservation)
'''
# IN: X = (B x L x F); OUT: (B x F x L, 1)
x = x.permute(0,2,1).unsqueeze(dim=-1)
Yhat_consurf = self.classifier(x) # OUT: Yhat_consurf = (B x N x L x 1)
# IN: (B x N x L x 1); OUT: ( B x L x N )
Yhat_consurf = Yhat_consurf.squeeze(dim=-1).permute(0,2,1)
return Yhat_consurf
# Define CNN structure for SETH continuous disorder prediction (CheZOD-scores)
class DisorderCNN( nn.Module ):
def __init__( self ):
super(DisorderCNN, self).__init__()
self.n_classes = 1
bottleneck_dim = 28
self.classifier = nn.Sequential(
#summarize information from 5 neighbouring amino acids (AAs)
#padding: dimension corresponding to AA number does not change
nn.Conv2d( 1024, bottleneck_dim, kernel_size=(5,1), padding=(2,0) ),
nn.Tanh(),
nn.Conv2d( bottleneck_dim, self.n_classes, kernel_size=(5,1), padding=(2,0))
)
def forward( self, x):
'''
L = protein length
B = batch-size
F = number of features (1024 for embeddings)
N = number of output nodes (1 for disorder, since predict one continuous number)
'''
# IN: X = (B x L x F); OUT: (B x F x L, 1)
x = x.permute(0,2,1).unsqueeze(dim=-1)
Yhat = self.classifier(x) # OUT: Yhat_consurf = (B x N x L x 1)
# IN: (B x N x L x 1); OUT: ( B x L x N )
Yhat = Yhat.squeeze(dim=-1).permute(0,2,1)
return Yhat
#### Residue-level Predictors
# https://github.com/Rostlab/VESPA
class ConservationPredictor():
def __init__(self, model_dir):
self.model = self.load_model(model_dir)
def load_model(self, model_dir):
checkpoint_p = model_dir / "conservation_checkpoint.pt"
weights_link = "http://data.bioembeddings.com/public/embeddings/feature_models/prott5cons/checkpoint.pt"
model = ConservationCNN()
return load_model(model, weights_link, checkpoint_p)
def write_predictions(self, predictions, out_dir):
out_p = out_dir / "conservation_pred.txt"
with open(out_p, 'w+') as out_f:
out_f.write( '\n'.join(
[ "{}".format(
''.join( [str(j) for j in yhat] ))
for yhat in predictions
]
) )
return None
# https://github.com/BernhoferM/TMbed
class TMbed():
def __init__(self,model_dir):
model_dir = model_dir / "tmbed"
if not model_dir.is_dir():
model_dir.mkdir()
self.model = self.load_model(model_dir)
self.decoder = Decoder().to(device)
def load_model(self,model_dir):
model_names = ["cv_0.pt","cv_1.pt","cv_2.pt","cv_3.pt","cv_4.pt"]
models = []
for model_file in model_names:
model = Predictor()
weights_link = "http://data.bioembeddings.com/public/embeddings/feature_models/tmbed/" + model_file
checkpoint_p = model_dir / model_file
# TMbed has a normalization layer --> cast to fp32 for stability
models.append( load_model(model,weights_link,checkpoint_p,state_dict="model").float() )
return models
def pred2label(self):
return {0: 'B', 1: 'b', 2: 'H', 3: 'h', 4: 'S', 5: 'i', 6: 'o'}
def write_predictions(self,predictions,out_dir):
out_p = out_dir / "membrane_tmbed.txt"
class_mapping = self.pred2label()
with open(out_p, 'w+') as out_f:
out_f.write( '\n'.join(
[ "{}".format(
''.join( [class_mapping[j] for j in yhat] ))
for yhat in predictions
]
) )
return None
# https://github.com/DagmarIlz/SETH
class SETH():
def __init__(self,model_dir):
model_dir = model_dir / "seth"
if not model_dir.is_dir():
model_dir.mkdir()
self.model = self.load_model(model_dir)
def load_model(self,model_dir):
checkpoint_p = model_dir / "seth_checkpoint.pt"
weights_link = "https://rostlab.org/~deepppi/SETH_CNN.pt"
model = DisorderCNN()
return load_model(model, weights_link, checkpoint_p)
def write_predictions(self,predictions,out_dir):
out_p = out_dir / "seth_disorder_pred.csv"
with open(out_p, 'w+') as out_f:
out_f.write( '\n'.join(
[ "{}".format( ', '.join( [str(j) for j in Zscore] ))
for Zscore in predictions
]
)
)
return None
# https://github.com/Rostlab/bindPredict
class BindPredict():
def __init__(self,model_dir):
model_dir = model_dir / "bindpredict"
if not model_dir.is_dir():
model_dir.mkdir()
self.model = self.load_model(model_dir)
def load_model(self,model_dir):
model_names = ["checkpoint1.pt","checkpoint2.pt","checkpoint3.pt","checkpoint4.pt","checkpoint5.pt"]
models = []
for model_file in model_names:
model = BindPredictCNN()
weights_link = "http://data.bioembeddings.com/public/embeddings/feature_models/bindembed21/" + model_file
checkpoint_p = model_dir / model_file
models.append( load_model(model,weights_link,checkpoint_p) )
return models
def binding_classes(self):
return {0: ('metal',"M"), 1: ('nucleic',"N"), 2: ('small',"S")}
def write_predictions(self,predictions,out_dir):
for idx, (binding_type, bind_short) in self.binding_classes().items():
out_p = out_dir / "binding_bindEmbed_{}_pred.txt".format(binding_type)
with open(out_p, 'w+') as out_f:
out_f.write( '\n'.join(
[ "{}".format(
''.join( [ bind_short if j==1 else "-" for j in yhat[:,idx]] ))
for yhat in predictions
]
) )
return None
class SecStructPredictor():
def __init__(self, model_dir):
model_dir = model_dir / "prott5_sec_struct"
if not model_dir.is_dir():
model_dir.mkdir()
self.model = self.load_model(model_dir)
def load_model(self, model_dir):
checkpoint_p = model_dir / "secstruct_checkpoint.pt"
weights_link = "http://data.bioembeddings.com/public/embeddings/feature_models/t5/secstruct_checkpoint.pt"
model = SecStructCNN()
return load_model(model, weights_link, checkpoint_p)
def label_mapping_3states(self):
return {0:"H",1:"E",2:"L"}
def label_mapping_8states(self):
return { idx : state for idx, state in enumerate("GHIBESTC") }
def write_predictions(self, predictions, out_dir, dssp3=True):
class_mapping=self.label_mapping_3states() if dssp3 else self.label_mapping_8states()
out_name = "dssp3_pred.txt" if dssp3 else "dssp8_pred.txt"
out_p = out_dir / out_name
with open(out_p, 'w+') as out_f:
out_f.write( '\n'.join(
[ "{}".format(
''.join( [class_mapping[j] for j in yhat] ))
for yhat in predictions
]
) )
return None
#### EAT-based methods (embedding-based annotation transfer)
class GoPredSim():
def __init__(self,model_dir,onto):
model_dir = model_dir / "goPredSim"
if not model_dir.is_dir():
model_dir.mkdir()
ONTOS = { "bpo" : "goa_annotations_2022_bpo.txt",
"mfo" : "goa_annotations_2022_mfo.txt",
"cco" : "goa_annotations_2022_cco.txt"
}
self.onto=onto
file_name = ONTOS[onto]
label_p = model_dir / file_name
self.lookup_labels = self.read_annotation(label_p)
self.lookup_ids, self.lookup_embs = self.read_embeddings(model_dir)
def read_annotation(self, label_p):
go_annotations=dict()
if not label_p.is_file():
embedding_link = "http://data.bioembeddings.com/public/embeddings/reference/goa/" + label_p.name
download_file(embedding_link,label_p)
with open(label_p,'r') as read_in:
next(read_in) # skip header
for line in read_in:
splitted_line = line.strip().split(";")
identifier = splitted_line[0]
go_terms = splitted_line[1]
if identifier not in go_annotations:
go_annotations[identifier]=set()
go_annotations[identifier].add(go_terms)
return go_annotations
def write_predictions(self,predictions,out_dir):
out_p = out_dir / "goPredSim_GO_{}_pred.csv".format(self.onto)
with open(out_p,"w+") as out_f:
out_f.write("\n".join( [ lookup_id + "\t" + ",".join(go_anno) + "\t{:.3f}".format( nn_dist )
for lookup_id, go_anno, nn_dist in predictions
]
)
)
return None
def read_embeddings(self,goPredSim_dir):
emb_file_name = "prott5_reference_embeddings.npy"
emb_local_p = goPredSim_dir / emb_file_name
if not emb_local_p.is_file():
weights_link = "https://rostlab.org/~deepppi/goPredSim/" + emb_file_name
download_file(weights_link,emb_local_p)
embs = torch.from_numpy( np.load(emb_local_p) )
id_file_name = "prott5_reference_embeddings.txt"
id_local_p = goPredSim_dir / id_file_name
if not id_local_p.is_file():
weights_link = "https://rostlab.org/~deepppi/goPredSim/" + id_file_name
download_file(weights_link,id_local_p)
with open(id_local_p,'r') as in_f:
ids = [line.strip() for line in in_f ]
# extract only those embeddings that have annotations for a given part of the GO-ontologies
filtered_ids = list()
filtered_idxs = list()
for i, identifier in enumerate(ids):
if identifier in self.lookup_labels:
filtered_ids.append(identifier)
filtered_idxs.append(i)
# add extra empty dimension to embeddings to allow for batch-comp of pairwise distances
return filtered_ids, embs[filtered_idxs].unsqueeze(dim=0).to(device)
class ProtTucker():
def __init__(self,model_dir):
model_dir = model_dir / "prottucker"
if not model_dir.is_dir():
model_dir.mkdir()
self.model = self.load_model(model_dir)
self.lookup_labels = self.read_annotation(model_dir)
self.lookup_ids, self.lookup_embs = self.read_embeddings(model_dir)
def load_model(self,model_dir):
checkpoint_p = model_dir / "tucker_weights.pt"
weights_link = "http://rostlab.org/~deepppi/embedding_repo/embedding_models/ProtTucker/ProtTucker_ProtT5.pt"
model = TuckerFNN()
return load_model(model,weights_link,checkpoint_p)
def read_annotation(self,model_dir):
label_p = model_dir / "cath_v430_dom_seqs_S100_161121_labels.txt"
if not label_p.is_file():
embedding_link = "https://rostlab.org/~deepppi/eat_dbs/cath_v430_dom_seqs_S100_161121_labels.txt"
download_file(embedding_link,label_p)
with open(label_p, 'r') as in_f:
# protein-ID : label
label_mapping = {
line.strip().split(',')[0] : line.strip().split(',')[1]
for line in in_f
}
return label_mapping
def read_embeddings(self,model_dir):
emb_local_p = model_dir / "cath_v430_dom_seqs_S100_161121.npy"
if not emb_local_p.is_file():
embeddings_link = "https://rostlab.org/~deepppi/eat_dbs/cath_v430_dom_seqs_S100_161121.npy"
download_file(embeddings_link,emb_local_p)
embeddings = torch.from_numpy( np.load(emb_local_p) ).to(device)
embeddings = self.model.single_pass(embeddings)
id_file_name = "cath_v430_dom_seqs_S100_161121.txt"
id_local_p = model_dir / id_file_name
if not id_local_p.is_file():
weights_link = "https://rostlab.org/~deepppi/eat_dbs/" + id_file_name
download_file(weights_link,id_local_p)
with open(id_local_p,'r') as in_f:
# cath|4_3_0|126lA00_1-162 --> 126lA00
ids = [ line.strip().split("|")[-1].split("_")[0] for line in in_f ]
return ids, embeddings.unsqueeze(dim=0)
def write_predictions(self, predictions, out_dir):
out_p = out_dir / "prottucker_CATH_pred.csv"
with open(out_p,"w+") as out_f:
out_f.write("\n".join( [ lookup_id + "\t" + cath_anno + "\t{:.3f}".format(nn_dist)
for lookup_id, cath_anno, nn_dist in predictions
]
)
)
return None
#### Protein-level predictors
class LA():
def __init__(self, model_dir, output_dim):
model_dir = model_dir / "light_attention"
if not model_dir.is_dir():
model_dir.mkdir()
self.subcell=True if output_dim==10 else False
self.model = self.load_model(model_dir, output_dim)
def load_model(self, model_dir, output_dim):
if self.subcell:
checkpoint_p = model_dir / "la_prott5_subcellular_localization.pt"
weights_link = 'http://data.bioembeddings.com/public/embeddings/feature_models/light_attention/la_prott5_subcellular_localization.pt'
else:
checkpoint_p = model_dir / "la_prott5_subcellular_solubility.pt"
weights_link = 'http://data.bioembeddings.com/public/embeddings/feature_models/light_attention/la_prott5_subcellular_solubility.pt'
model = LightAttention(output_dim=output_dim)
# LA has a normalization layer --> cast to fp32 for stability
return load_model(model,weights_link,checkpoint_p).float()
# https://github.com/sacdallago/bio_embeddings/blob/develop/bio_embeddings/extract/light_attention/light_attention_annotation_extractor.py#L15
def class2label(self):
if self.subcell:
return {
0: "Cell_membrane",
1: "Cytoplasm",
2: "Endoplasmatic_reticulum",
3: "Golgi_apparatus",
4: "Lysosome_or_Vacuole",
5: "Mitochondrion",
6: "Nucleus",
7: "Peroxisome",
8: "Plastid",
9: "Extracellular"
}
else:
return {
0: "Membrane",
1: "Soluble"
}
def write_predictions(self,predictions,out_dir):
class2label=self.class2label()
if self.subcell:
out_p = out_dir / "la_subcell_pred.txt"
else:
out_p = out_dir / "la_mem_pred.txt"
with open(out_p, 'w+') as out_f:
out_f.write( '\n'.join( [ class2label[pred] for pred in predictions] ) )
return None
#### 3D Structure prediction:
class Ember3D:
def __init__(self, model_dir):
model_dir = model_dir / "ember3D"
if not model_dir.is_dir():
model_dir.mkdir()
self.model, self.Ember3D_Result = self.load_model(model_dir)
def load_model(self, model_dir):
# EMBER3D specifics
# Import here to avoid unnecessary dependencies if 3D output is not wanted
import sys
sys.path.insert(0, './EMBER3D')
from model import RF_1I1F
from Ember3D import Ember3D_Result
checkpoint_p = model_dir / "EMBER3D.model"
weights_link = 'https://github.com/kWeissenow/EMBER3D/raw/main/model/EMBER3D.model'
if not checkpoint_p.exists():
download_file(weights_link,checkpoint_p)
model = RF_1I1F()
model = model.to(device)
model.load_state_dict(torch.load(checkpoint_p))
model.eval()
return model, Ember3D_Result
# original one-hot implementation of Konstantin
def sequence_to_onehot(self, seq):
aa_list = list("ACDEFGHIKLMNPQRSTVWY")
# replace unknown/ambigious/rare AAs (X) by Glycin for 3D reconstruction (simply assume no side-chain in this case)
encoded = torch.tensor([aa_list.index(c) if c not in "XOUZB" else aa_list.index("G") for c in seq])
return torch.nn.functional.one_hot(encoded, num_classes=20)
def write_predictions(self,predictions,ids,out_dir):
out_dir = out_dir / "ember3d_pdbs"
if not out_dir.is_dir():
out_dir.mkdir()
# write one PDB file for each protein
for (prot_id, result) in zip(ids,predictions):
result.save_pdb(prot_id, str(out_dir/(prot_id+".pdb")))
return None
def predict(self, seq, emb_1d, emb_2d):
with torch.no_grad():
emb_1d = torch.unsqueeze(emb_1d, dim=0)
emb_1d = torch.unsqueeze(emb_1d, dim=0)
emb_2d = torch.permute(emb_2d, (1, 2, 0))
emb_2d = torch.unsqueeze(emb_2d, dim=0)
seq1hot = self.sequence_to_onehot(seq)
seq1hot = torch.unsqueeze(seq1hot, dim=0).to(device)
idx = torch.arange(len(seq))
idx = torch.unsqueeze(idx, dim=0).to(device)
with torch.cuda.amp.autocast(): # actual 3D prediction
pair_pred, coords_pred, lddt_pred = self.model.forward(seq1hot, idx, emb_1d, emb_2d)
pair_pred = list(pair_pred)
for i in range(len(pair_pred)):
pair_pred[i] = pair_pred[i].detach()
result = self.Ember3D_Result(seq, pair_pred, coords_pred.detach(), lddt_pred.squeeze().detach())
return result
#### Generators:
# ProtGPT2 will generate new proteins if no FASTA file is passed to the script
# Adapted from: https://huggingface.co/spaces/sdnf/GradioFold/blob/main/app.py#L409
class ProtGPT2():
def __init__(self,model_dir):
model_dir = model_dir / "protgpt2"
if not model_dir.is_dir():
model_dir.mkdir()
# Load ProTGPT2
self.model, self.tokenizer = self.load_protGPT2(model_dir)
self.bad_words_ids = self.get_bad_word_ids()
# the seed used as starting point for generating new random sequences
self.start_seq = "<|endoftext|>"
# the token indicating the end of a generated sequence.
# currently identical to start_sequences but can be different if a different starting_point is known (prior knowledge)
self.eos = "<|endoftext|>"
# Necessary for perplexity computation
self.ce_loss = nn.CrossEntropyLoss(reduction="none")
def get_bad_word_ids(self):
# avoid generating proteins with unknown/ambigious/rare AAs
kmer_X = [ kmer for kmer in self.tokenizer.get_vocab().keys()
if "X" in kmer or "Z" in kmer or "B" in kmer or "O" in kmer ]
print(kmer_X)
bad_words_ids = self.tokenizer(kmer_X, add_special_tokens=False).input_ids
return bad_words_ids
def load_protGPT2(self,model_dir):
print("Start loading ProtGPT...")
start=time.time()
transformer_name="nferruz/ProtGPT2"
model = GPT2LMHeadModel.from_pretrained(transformer_name,cache_dir=model_dir)
tokenizer = GPT2Tokenizer.from_pretrained(transformer_name,cache_dir=model_dir)
# Artificially add padding token (necessary for batch-processing of perplexity computation)
tokenizer.pad_token = tokenizer.eos_token
model = model.half() # TODO: double check with Noelia whether that's ok
model = model.to(device)
model = model.eval()
print("Finished loading: {} in {:.1f}[s]".format(transformer_name,time.time()-start))
return model, tokenizer
# Mostly taken/modified from: https://huggingface.co/spaces/evaluate-metric/perplexity/blob/main/perplexity.py
def batch_compute_perplexity(self, seqs, add_start_token=True):
token_encoding = self.tokenizer.batch_encode_plus(seqs,padding="longest")
input_ids = torch.tensor(token_encoding['input_ids']).to(device)
attention_mask = torch.tensor(token_encoding['attention_mask']).to(device)
if add_start_token: # add start token to each sequence to mimick training as closely as possible
bos_tokens_tensor = torch.tensor([[self.tokenizer.bos_token_id]] * input_ids.size(dim=0)).to(device)
input_ids = torch.cat([bos_tokens_tensor, input_ids], dim=1)
attention_mask = torch.cat(
[torch.ones(bos_tokens_tensor.size(), dtype=torch.int64).to(device), attention_mask], dim=1
)
with torch.no_grad():
out_logits = self.model(input_ids, attention_mask=attention_mask).logits
# generate targets to compute ppl. (in this case: copy of inputs that gets shifted to the right)
target_ids = input_ids
#target_ids[~attention_mask] = -100 # mask padded tokens from loss/ppl computation
# for the last token, we can not predict any next token --> remove
shift_logits = out_logits[..., :-1, :].contiguous()
# removing the first token from targets effectively shifts everything to the right
shift_labels = target_ids[..., 1:].contiguous()
# same goes for attention-mask: removing first element, effectively shifts right
shift_attention_mask_batch = attention_mask[..., 1:].contiguous()
perplexity_batch = torch.exp2(
(self.ce_loss(shift_logits.transpose(1, 2), shift_labels) * shift_attention_mask_batch).sum(1)
/ shift_attention_mask_batch.sum(1)
)
ppls = perplexity_batch.tolist()
return ppls
# parameters directly taken from: https://huggingface.co/spaces/sdnf/GradioFold/blob/main/app.py#L409
def run_protgpt2(self, n_gen, length=100, repetitionPenalty=1.2, top_k_poolsize=950):
print("Start generating proteins with ProtGPT2 ...")
start=time.time()
n_seqs_to_select = 50 # max. number of sequences to select from each batch
n_seqs_to_sample = n_seqs_to_select*10 # how many samples to generate per batch; only subset of generated are used (filtering based on PPL)
seq_dict = dict()
input_ids = self.tokenizer.encode(self.start_seq, return_tensors='pt').to(device)
while len(seq_dict)<n_gen: # generate new sequences batch-wise until desired size is reached
hallucinated_proteins = self.model.generate(input_ids,
max_length=length,
do_sample=True,
top_k=top_k_poolsize,
repetition_penalty=repetitionPenalty,
num_return_sequences=n_seqs_to_sample,
eos_token_id=0,
pad_token_id=0, # this avoids raising a warning
bad_words_ids=self.bad_words_ids
)
clean_seqs = list()
filtered_seqs = list()
# "clean up" newly generated sequences IF
# new-lines are generated after few tokens or
# End-of-sentence is reached too early/often
for rnd_seq in hallucinated_proteins:
decoded_seq = self.tokenizer.decode(rnd_seq)
if (decoded_seq[0:60].count("\n")==1 and # No newlines allowed in first line and avoid truncation
decoded_seq.count(self.eos)>=2 and # minimally, two EOS tokens (one for start, one for end)
decoded_seq.count("X")<10 # avoid sequences with too many unknowns
):
# in case multiple sequences were generated in one pass, use only the first one
clean_seq = decoded_seq.split(self.eos)[1]
if len(clean_seq)<5: # remove sequences that are too short
continue
clean_seqs.append(clean_seq)
# get perplexity for batch of proteins
# TODO: add perplexity threshold here to filter out low-complexity hallucinations
ppls = self.batch_compute_perplexity(clean_seqs)
filtered_seqs += [ (seq, ppl) for seq, ppl in zip(clean_seqs,ppls) ]
filtered_seqs.sort(key = lambda x: x[1])
# choose only the 10% generated sequences with best perplexity
# TODO: Rather in-efficient but can be skipped once we have a proper threshold; see filtering above
sel_seqs = filtered_seqs[:n_seqs_to_select]
# dictionary style is required to be compatible with the output of my read_fasta function
for i, seq in enumerate(sel_seqs,start=len(seq_dict)):
s = seq[0].replace("\n","")
seq_len = len(s)
identifier=">seq{:.0f}, L={:.0f}, ppl={:.3f}".format(i,seq_len,seq[1])
seq_dict[identifier] = s
if len(seq_dict)==n_gen:
print("Generated and filtered {} sequences in {:.1f}[s]".format(len(seq_dict),time.time()-start))
return seq_dict # return once desired number of sequences is reached
print("Generated and filtered {} sequences in {:.1f}[s]".format(len(seq_dict),time.time()-start))
return seq_dict
#### Embedding/Inference
# ProtT5 acts as a microscope with different lenses/predictors to investigate
# various protein properties
class ProtT5Microscope():
def __init__( self, seq_dict, model_dir, fmt ):
self.predictors, self.results = self.register_predictors(model_dir,fmt)
# only output attention heads if 3D output is requested
output_attentions=True if "EMBER3D" in self.predictors else False
self.prott5, self.tokenizer = self.get_T5_model(model_dir,output_attentions)
self.seq_dict = seq_dict
self.ids = list()
self.seqs = list()
self.sigm = nn.Sigmoid()
def register_predictors(self,model_dir,fmt):
'''
Parameters
----------
model_dir : PATH
Path to the checkpoint directory
fmt : LIST
list of predictors to run
Returns
-------
p : DICT
dictionary of predictors.
Key refers to the name of the predictor. One predictor can have multiple outputs
r : DICT
dictionary of results (lists) to store predictions.
Some predictors have multiple outputs and need multiple output containers
'''
p = {} # dictionary holding predictor classes
r = {} # dict for storing lists of results results
# currently supported: ss,cons,dis,mem,bind, go,subcell,tucker,ember,emb
for f in fmt:
if f=="ss":
print("Loading secondary structure predictor")
p["ProtT5_SecStruct"] = SecStructPredictor(model_dir)
r["SecStruct3"] = list() # 3-state secondary structure
r["SecStruct8"] = list() # 3-state secondary structure
elif f=="cons":
print("Loading conservation predictor")
p["VESPA_Conservation"] = ConservationPredictor(model_dir)
r["Conservation"] = list()
elif f=="dis":
print("Loading disorder predictor")
p["SETH"] = SETH(model_dir)
r["Disorder"] = list()
elif f=="mem":
print("Loading transmembrane predictor")
p["TMbed"] = TMbed(model_dir)
r["Membrane"] = list()
elif f=="bind":
print("Loading bindinx predictor")
p["BindEmbed21DL"] = BindPredict(model_dir)
r["Binding"] = list()
elif f=="go":
print("Loading GO predictor")
p["goPredSim-mfo"] = GoPredSim(model_dir, "mfo")
p["goPredSim-bpo"] = GoPredSim(model_dir, "bpo")
p["goPredSim-cco"] = GoPredSim(model_dir, "cco")
r["GO-mfo"] = list()
r["GO-bpo"] = list()
r["GO-cco"] = list()
elif f=="subcell":
print("Loading subcellular location predictor")
p["LA-subcell"] = LA(model_dir,output_dim=10)
p["LA-mem"] = LA(model_dir,output_dim=2)
r["Subcell"] = list()
r["LA-mem"] = list()
elif f=="tucker":
print("Loading CATH predictor")
p["ProtTucker"] = ProtTucker(model_dir)
r["ProtTucker"] = list()
elif f=="ember3D":
print("Loading 3D structure predictor")
p["EMBER3D"] = Ember3D(model_dir)
r["3DStructure"] = list()
elif f=="emb":
p["ProtEmbs"] = None # flag to highlight embeddings are requested
r["ProtEmbs"] = list() # list of per-protein embeddings
else:
print("The requested predictor is not implemted. Check list of available predictors (-fmt parameter)")
raise NotImplementedError
return p, r
def write_list(self,data,out_p):
with open(out_p,"w+") as out_f:
out_f.write("\n".join(data))
return None
def write_protEmbs(self,out_dir):
out_p = out_dir / "protein_embeddings.npy"
np.save(str(out_p),self.results["ProtEmbs"])
return None
def get_T5_model(self,model_dir,output_attentions):
# Load your checkpoint here
# Currently, only the encoder-part of ProtT5 is loaded in half-precision
print("Start loading ProtT5...")
start=time.time()
transformer_name = "Rostlab/prot_t5_xl_half_uniref50-enc"
prott5_dir = model_dir / "prott5"
model = T5EncoderModel.from_pretrained(transformer_name, torch_dtype=torch.float16,
cache_dir=prott5_dir, output_attentions=output_attentions)
model = model.to(device)
model = model.eval()
tokenizer = T5Tokenizer.from_pretrained(transformer_name, do_lower_case=False, cache_dir=prott5_dir)
print("Finished loading: {} in {:.1f}[s]".format(transformer_name,time.time()-start))
return model, tokenizer
def batch_predict_residues( self, max_batch_size=100, max_residues=4000):
sorted_seqs = sorted( self.seq_dict.items(), key=lambda kv: len( self.seq_dict[kv[0]] ), reverse=True )
batch = list()
print("Start predicting protein properties ...")
start = time.time()
# for each sequence in the batch
for seq_idx, (pdb_id, seq) in tqdm(enumerate(sorted_seqs,1)):
seq_len = len(seq)
seq = ' '.join(list(seq))
batch.append((pdb_id,seq,seq_len))
n_res_batch = sum([ s_len for _, _, s_len in batch ]) + seq_len
# if enough sequences were accumulated to process one batch
if ( len(batch)>=max_batch_size or # max. number of sequences per batch
n_res_batch>=max_residues or # max. number of residues per batch
seq_idx==len(sorted_seqs) # special case: last batch
):
pdb_ids, seqs, seq_lens = zip(*batch)
batch = list()
# tokenize protein sequence and generate attention_mask for batch_processing
token_encoding = self.tokenizer.batch_encode_plus(seqs, add_special_tokens=True, padding='longest')
input_ids = torch.tensor(token_encoding['input_ids']).to(device)
attention_mask = torch.tensor(token_encoding['attention_mask']).to(device)
try:
with torch.no_grad():
prott5_output = self.prott5( input_ids, attention_mask=attention_mask )
except RuntimeError:
print("RuntimeError for {} (L={})".format(pdb_id, seq_lens))
print("Cleaning up ...")
del(token_encoding)
del(input_ids)
del(attention_mask)
continue
# extract last hidden states (=embeddings)
residue_embedding = prott5_output.last_hidden_state.detach()
# mask out padded elements in the attention output (can be non-zero) for further processing/prediction
residue_embedding = residue_embedding*attention_mask.unsqueeze(dim=-1)
# cast to fp32 (important to ascertain performance of downstream predictors, i.e., TMbed's layerNorm)
residue_embedding = residue_embedding
protein_embeddings = list()
# ProtT5 appends a special tokens at the end of each sequence
# Mask this also out during inference
for idx, s_len in enumerate(seq_lens):
attention_mask[idx,s_len] = 0
# get per-protein embeddings here (take padding into account)
protein_embeddings.append( residue_embedding[idx,:s_len].mean(dim=0) )
# stack them again in one tensor for batch-processing
protein_embeddings=torch.vstack(protein_embeddings)
# some predictors require dimensions to be swapped; prepare only once here
# B x L x N --> B x N x L
residue_embedding_transpose = torch.permute(residue_embedding, (0,2,1))
with torch.no_grad():
# batch-predict all properties but 3D structure
# 3D structure predictions works only without batching right now
for predictor_name, predictor in self.predictors.items(): # for each registered predictor
### Residue-level predictions ###
# predict 3- and 8-state sec. struct
if predictor_name=="ProtT5_SecStruct":
d3_Yhat, d8_Yhat = predictor.model(residue_embedding)
d3_Yhat = toCPU(torch.max( d3_Yhat, dim=-1, keepdim=True )[1] ).astype(np.byte)
d8_Yhat = toCPU(torch.max( d8_Yhat, dim=-1, keepdim=True )[1] ).astype(np.byte)