forked from Kuhlman-Lab/ThermoMPNN
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathprotein_mpnn_utils.py
1587 lines (1386 loc) · 72.7 KB
/
protein_mpnn_utils.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
from __future__ import print_function
import json, time, os, sys, glob
import shutil
import numpy as np
import torch
from torch import optim
from torch.utils.data import DataLoader
from torch.utils.data.dataset import random_split, Subset
import copy
import torch.nn as nn
import torch.nn.functional as F
import random
import itertools
import re
# A number of functions/classes are adopted from: https://github.com/jingraham/neurips19-graph-protein-design
def _scores(S, log_probs, mask):
""" Negative log probabilities """
criterion = torch.nn.NLLLoss(reduction='none')
loss = criterion(
log_probs.contiguous().view(-1, log_probs.size(-1)),
S.contiguous().view(-1)
).view(S.size())
scores = torch.sum(loss * mask, dim=-1) / torch.sum(mask, dim=-1)
return scores
def _S_to_seq(S, mask):
alphabet = 'ACDEFGHIKLMNPQRSTVWYX'
seq = ''.join([alphabet[c] for c, m in zip(S.tolist(), mask.tolist()) if m > 0])
return seq
def parse_PDB_biounits(x, atoms=['N', 'CA', 'C'], chain=None):
'''
input: x = PDB filename
atoms = atoms to extract (optional)
output: (length, atoms, coords=(x,y,z)), sequence
'''
alpha_1 = list("ARNDCQEGHILKMFPSTWYV-")
states = len(alpha_1)
alpha_3 = ['ALA', 'ARG', 'ASN', 'ASP', 'CYS', 'GLN', 'GLU', 'GLY', 'HIS', 'ILE',
'LEU', 'LYS', 'MET', 'PHE', 'PRO', 'SER', 'THR', 'TRP', 'TYR', 'VAL', 'GAP']
aa_1_N = {a: n for n, a in enumerate(alpha_1)}
aa_3_N = {a: n for n, a in enumerate(alpha_3)}
aa_N_1 = {n: a for n, a in enumerate(alpha_1)}
aa_1_3 = {a: b for a, b in zip(alpha_1, alpha_3)}
aa_3_1 = {b: a for a, b in zip(alpha_1, alpha_3)}
def AA_to_N(x):
# ["ARND"] -> [[0,1,2,3]]
x = np.array(x);
if x.ndim == 0: x = x[None]
return [[aa_1_N.get(a, states - 1) for a in y] for y in x]
def N_to_AA(x):
# [[0,1,2,3]] -> ["ARND"]
x = np.array(x);
if x.ndim == 1: x = x[None]
return ["".join([aa_N_1.get(a, "-") for a in y]) for y in x]
xyz, seq, min_resn, max_resn = {}, {}, 1e6, -1e6
for line in open(x, "rb"):
line = line.decode("utf-8", "ignore").rstrip()
if line[:6] == "HETATM" and line[17:17 + 3] == "MSE":
line = line.replace("HETATM", "ATOM ")
line = line.replace("MSE", "MET")
if line[:4] == "ATOM":
ch = line[21:22]
if ch == chain or chain is None:
atom = line[12:12 + 4].strip()
resi = line[17:17 + 3]
resn = line[22:22 + 5].strip()
x, y, z = [float(line[i:(i + 8)]) for i in [30, 38, 46]]
if resn[-1].isalpha():
resa, resn = resn[-1], int(resn[:-1]) - 1
else:
resa, resn = "", int(resn) - 1
# resn = int(resn)
if resn < min_resn:
min_resn = resn
if resn > max_resn:
max_resn = resn
if resn not in xyz:
xyz[resn] = {}
if resa not in xyz[resn]:
xyz[resn][resa] = {}
if resn not in seq:
seq[resn] = {}
if resa not in seq[resn]:
seq[resn][resa] = resi
if atom not in xyz[resn][resa]:
xyz[resn][resa][atom] = np.array([x, y, z])
# convert to numpy arrays, fill in missing values
seq_, xyz_ = [], []
try:
for resn in range(min_resn, max_resn + 1):
if resn in seq:
for k in sorted(seq[resn]): seq_.append(aa_3_N.get(seq[resn][k], 20))
else:
seq_.append(20)
if resn in xyz:
for k in sorted(xyz[resn]):
for atom in atoms:
if atom in xyz[resn][k]:
xyz_.append(xyz[resn][k][atom])
else:
xyz_.append(np.full(3, np.nan))
else:
for atom in atoms: xyz_.append(np.full(3, np.nan))
return np.array(xyz_).reshape(-1, len(atoms), 3), N_to_AA(np.array(seq_))
except TypeError:
return 'no_chain', 'no_chain'
def parse_PDB(path_to_pdb, input_chain_list=None, ca_only=False, side_chains=False):
c = 0
pdb_dict_list = []
init_alphabet = ['A', 'B', 'C', 'D', 'E', 'F', 'G', 'H', 'I', 'J', 'K', 'L', 'M', 'N', 'O', 'P', 'Q', 'R', 'S', 'T',
'U', 'V', 'W', 'X', 'Y', 'Z', 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'm', 'n',
'o', 'p', 'q', 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z']
extra_alphabet = [str(item) for item in list(np.arange(300))]
chain_alphabet = init_alphabet + extra_alphabet
if input_chain_list:
chain_alphabet = input_chain_list
biounit_names = [path_to_pdb]
for biounit in biounit_names:
my_dict = {}
s = 0
concat_seq = ''
concat_N = []
concat_CA = []
concat_C = []
concat_O = []
concat_mask = []
coords_dict = {}
for letter in chain_alphabet:
if ca_only:
sidechain_atoms = ['CA']
elif side_chains:
sidechain_atoms = ["N", "CA", "C", "O", "CB",
"CG", "CG1", "OG1", "OG2", "CG2", "OG", "SG",
"CD", "SD", "CD1", "ND1", "CD2", "OD1", "OD2", "ND2",
"CE", "CE1", "NE1", "OE1", "NE2", "OE2", "NE", "CE2", "CE3",
"NZ", "CZ", "CZ2", "CZ3", "CH2", "OH", "NH1", "NH2"]
else:
sidechain_atoms = ['N', 'CA', 'C', 'O']
xyz, seq = parse_PDB_biounits(biounit, atoms=sidechain_atoms, chain=letter)
if type(xyz) != str:
concat_seq += seq[0]
my_dict['seq_chain_' + letter] = seq[0]
coords_dict_chain = {}
if ca_only:
coords_dict_chain['CA_chain_' + letter] = xyz.tolist()
else:
coords_dict_chain['N_chain_' + letter] = xyz[:, 0, :].tolist()
coords_dict_chain['CA_chain_' + letter] = xyz[:, 1, :].tolist()
coords_dict_chain['C_chain_' + letter] = xyz[:, 2, :].tolist()
coords_dict_chain['O_chain_' + letter] = xyz[:, 3, :].tolist()
my_dict['coords_chain_' + letter] = coords_dict_chain
s += 1
fi = biounit.rfind("/")
my_dict['name'] = biounit[(fi + 1):-4]
my_dict['num_of_chains'] = s
my_dict['seq'] = concat_seq
if s <= len(chain_alphabet):
pdb_dict_list.append(my_dict)
c += 1
return pdb_dict_list
def alt_parse_PDB_biounits(x, atoms=['N', 'CA', 'C'], chain=None):
'''
input: x = PDB filename
atoms = atoms to extract (optional)
output: (length, atoms, coords=(x,y,z)), sequence
'''
alpha_1 = list("ARNDCQEGHILKMFPSTWYV-")
states = len(alpha_1)
alpha_3 = ['ALA', 'ARG', 'ASN', 'ASP', 'CYS', 'GLN', 'GLU', 'GLY', 'HIS', 'ILE',
'LEU', 'LYS', 'MET', 'PHE', 'PRO', 'SER', 'THR', 'TRP', 'TYR', 'VAL', 'GAP']
aa_1_N = {a: n for n, a in enumerate(alpha_1)}
aa_3_N = {a: n for n, a in enumerate(alpha_3)}
aa_N_1 = {n: a for n, a in enumerate(alpha_1)}
aa_1_3 = {a: b for a, b in zip(alpha_1, alpha_3)}
aa_3_1 = {b: a for a, b in zip(alpha_1, alpha_3)}
def AA_to_N(x):
# ["ARND"] -> [[0,1,2,3]]
x = np.array(x);
if x.ndim == 0: x = x[None]
return [[aa_1_N.get(a, states - 1) for a in y] for y in x]
def N_to_AA(x):
# [[0,1,2,3]] -> ["ARND"]
x = np.array(x);
if x.ndim == 1: x = x[None]
return ["".join([aa_N_1.get(a, "-") for a in y]) for y in x]
xyz, seq, min_resn, max_resn = {}, {}, 1e6, -1e6
resn_list = []
for line in open(x, "rb"):
line = line.decode("utf-8", "ignore").rstrip()
if line[:6] == "HETATM" and line[17:17 + 3] == "MSE":
line = line.replace("HETATM", "ATOM ")
line = line.replace("MSE", "MET")
if line[:4] == "ATOM":
ch = line[21:22]
if ch == chain or chain is None:
atom = line[12:12 + 4].strip()
resi = line[17:17 + 3]
resn = line[22:22 + 5].strip()
# RAW resn is defined HERE
if resn not in resn_list:
resn_list.append(resn) # NEED to keep ins code here
x, y, z = [float(line[i:(i + 8)]) for i in [30, 38, 46]]
if resn[-1].isalpha():
resa, resn = resn[-1], int(resn[:-1]) - 1
else:
resa, resn = "", int(resn) - 1
if resn < min_resn:
min_resn = resn
if resn > max_resn:
max_resn = resn
if resn not in xyz:
xyz[resn] = {}
if resa not in xyz[resn]:
xyz[resn][resa] = {}
if resn not in seq:
seq[resn] = {}
if resa not in seq[resn]:
seq[resn][resa] = resi
if atom not in xyz[resn][resa]:
xyz[resn][resa][atom] = np.array([x, y, z])
# print(resn_list)
# convert to numpy arrays, fill in missing values
seq_, xyz_ = [], []
# print(seq)
try:
for resn in range(min_resn, max_resn + 1):
if resn in seq:
# CLEANED resn is defined HERE
# resn_list.append(str(resn + 1))
# print(resn)
for k in sorted(seq[resn]): seq_.append(aa_3_N.get(seq[resn][k], 20))
else:
seq_.append(20)
# if str(resn - 1) not in resn_list:
# print(resn, '**')
# resn_list.append('DASH')
# print(resn, resn_list[-1])
if resn in xyz:
for k in sorted(xyz[resn]):
for atom in atoms:
if atom in xyz[resn][k]:
xyz_.append(xyz[resn][k][atom])
else:
xyz_.append(np.full(3, np.nan))
else:
for atom in atoms: xyz_.append(np.full(3, np.nan))
# assert len(seq_) == len(list(dict.fromkeys(resn_list)))
return np.array(xyz_).reshape(-1, len(atoms), 3), N_to_AA(np.array(seq_)), list(dict.fromkeys(resn_list))
except TypeError:
return 'no_chain', 'no_chain', 'no_chain'
def alt_parse_PDB(path_to_pdb, input_chain_list=None, ca_only=False, side_chains=False):
c = 0
pdb_dict_list = []
init_alphabet = ['A', 'B', 'C', 'D', 'E', 'F', 'G', 'H', 'I', 'J', 'K', 'L', 'M', 'N', 'O', 'P', 'Q', 'R', 'S', 'T',
'U', 'V', 'W', 'X', 'Y', 'Z', 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'm', 'n',
'o', 'p', 'q', 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z']
extra_alphabet = [str(item) for item in list(np.arange(300))]
chain_alphabet = init_alphabet + extra_alphabet
if input_chain_list:
chain_alphabet = input_chain_list
biounit_names = [path_to_pdb]
for biounit in biounit_names:
my_dict = {}
my_dict['resn_list'] = []
s = 0
concat_seq = ''
concat_N = []
concat_CA = []
concat_C = []
concat_O = []
concat_mask = []
coords_dict = {}
for letter in chain_alphabet:
if ca_only:
sidechain_atoms = ['CA']
elif side_chains:
sidechain_atoms = ["N", "CA", "C", "O", "CB",
"CG", "CG1", "OG1", "OG2", "CG2", "OG", "SG",
"CD", "SD", "CD1", "ND1", "CD2", "OD1", "OD2", "ND2",
"CE", "CE1", "NE1", "OE1", "NE2", "OE2", "NE", "CE2", "CE3",
"NZ", "CZ", "CZ2", "CZ3", "CH2", "OH", "NH1", "NH2"]
else:
sidechain_atoms = ['N', 'CA', 'C', 'O']
xyz, seq, resn_list = alt_parse_PDB_biounits(biounit, atoms=sidechain_atoms, chain=letter)
# add chain resn set to list if exists
# new for multi-chain handling
# if len(resn_list) > 7:
# check_chain = ''.join(resn_list[:8])
# if check_chain != 'no_chain':
# my_dict['resn_list'] += resn_list
if type(xyz) != str:
concat_seq += seq[0]
my_dict['seq_chain_' + letter] = seq[0]
coords_dict_chain = {}
if ca_only:
coords_dict_chain['CA_chain_' + letter] = xyz.tolist()
else:
coords_dict_chain['N_chain_' + letter] = xyz[:, 0, :].tolist()
coords_dict_chain['CA_chain_' + letter] = xyz[:, 1, :].tolist()
coords_dict_chain['C_chain_' + letter] = xyz[:, 2, :].tolist()
coords_dict_chain['O_chain_' + letter] = xyz[:, 3, :].tolist()
my_dict['coords_chain_' + letter] = coords_dict_chain
s += 1
fi = biounit.rfind("/")
my_dict['name'] = biounit[(fi + 1):-4]
my_dict['num_of_chains'] = s
my_dict['seq'] = concat_seq
# old handling for resn lists
my_dict['resn_list'] = list(resn_list)
if s <= len(chain_alphabet):
pdb_dict_list.append(my_dict)
c += 1
return pdb_dict_list
def tied_featurize(batch, device, chain_dict, fixed_position_dict=None, omit_AA_dict=None, tied_positions_dict=None,
pssm_dict=None, bias_by_res_dict=None, ca_only=False):
""" Pack and pad batch into torch tensors """
alphabet = 'ACDEFGHIKLMNPQRSTVWYX'
B = len(batch)
lengths = np.array([len(b['seq']) for b in batch], dtype=np.int32) # sum of chain seq lengths
L_max = max([len(b['seq']) for b in batch])
if ca_only:
X = np.zeros([B, L_max, 1, 3])
else:
X = np.zeros([B, L_max, 4, 3])
residue_idx = -100 * np.ones([B, L_max], dtype=np.int32)
chain_M = np.zeros([B, L_max], dtype=np.int32) # 1.0 for the bits that need to be predicted
pssm_coef_all = np.zeros([B, L_max], dtype=np.float32) # 1.0 for the bits that need to be predicted
pssm_bias_all = np.zeros([B, L_max, 21], dtype=np.float32) # 1.0 for the bits that need to be predicted
pssm_log_odds_all = 10000.0 * np.ones([B, L_max, 21],
dtype=np.float32) # 1.0 for the bits that need to be predicted
chain_M_pos = np.zeros([B, L_max], dtype=np.int32) # 1.0 for the bits that need to be predicted
bias_by_res_all = np.zeros([B, L_max, 21], dtype=np.float32)
chain_encoding_all = np.zeros([B, L_max], dtype=np.int32) # 1.0 for the bits that need to be predicted
S = np.zeros([B, L_max], dtype=np.int32)
omit_AA_mask = np.zeros([B, L_max, len(alphabet)], dtype=np.int32)
# Build the batch
letter_list_list = []
visible_list_list = []
masked_list_list = []
masked_chain_length_list_list = []
tied_pos_list_of_lists_list = []
# shuffle all chains before the main loop
for i, b in enumerate(batch):
if chain_dict != None:
masked_chains, visible_chains = chain_dict[
b['name']] # masked_chains a list of chain letters to predict [A, D, F]
else:
masked_chains = [item[-1:] for item in list(b) if item[:10] == 'seq_chain_']
visible_chains = []
num_chains = b['num_of_chains']
all_chains = masked_chains + visible_chains
# random.shuffle(all_chains)
for i, b in enumerate(batch):
mask_dict = {}
a = 0
x_chain_list = []
chain_mask_list = []
chain_seq_list = []
chain_encoding_list = []
c = 1
letter_list = []
global_idx_start_list = [0]
visible_list = []
masked_list = []
masked_chain_length_list = []
fixed_position_mask_list = []
omit_AA_mask_list = []
pssm_coef_list = []
pssm_bias_list = []
pssm_log_odds_list = []
bias_by_res_list = []
l0 = 0
l1 = 0
for step, letter in enumerate(all_chains):
if letter in visible_chains:
letter_list.append(letter)
visible_list.append(letter)
chain_seq = b[f'seq_chain_{letter}']
chain_seq = ''.join([a if a != '-' else 'X' for a in chain_seq])
chain_length = len(chain_seq)
global_idx_start_list.append(global_idx_start_list[-1] + chain_length)
chain_coords = b[f'coords_chain_{letter}'] # this is a dictionary
chain_mask = np.zeros(chain_length) # 0.0 for visible chains
if ca_only:
x_chain = np.array(chain_coords[f'CA_chain_{letter}']) # [chain_lenght,1,3] #CA_diff
if len(x_chain.shape) == 2:
x_chain = x_chain[:, None, :]
else:
x_chain = np.stack([chain_coords[c] for c in
[f'N_chain_{letter}', f'CA_chain_{letter}', f'C_chain_{letter}',
f'O_chain_{letter}']], 1) # [chain_lenght,4,3]
x_chain_list.append(x_chain)
chain_mask_list.append(chain_mask)
chain_seq_list.append(chain_seq)
chain_encoding_list.append(c * np.ones(np.array(chain_mask).shape[0]))
l1 += chain_length
residue_idx[i, l0:l1] = 100 * (c - 1) + np.arange(l0, l1)
l0 += chain_length
c += 1
fixed_position_mask = np.ones(chain_length)
fixed_position_mask_list.append(fixed_position_mask)
omit_AA_mask_temp = np.zeros([chain_length, len(alphabet)], np.int32)
omit_AA_mask_list.append(omit_AA_mask_temp)
pssm_coef = np.zeros(chain_length)
pssm_bias = np.zeros([chain_length, 21])
pssm_log_odds = 10000.0 * np.ones([chain_length, 21])
pssm_coef_list.append(pssm_coef)
pssm_bias_list.append(pssm_bias)
pssm_log_odds_list.append(pssm_log_odds)
bias_by_res_list.append(np.zeros([chain_length, 21]))
if letter in masked_chains:
masked_list.append(letter)
letter_list.append(letter)
chain_seq = b[f'seq_chain_{letter}']
chain_seq = ''.join([a if a != '-' else 'X' for a in chain_seq])
chain_length = len(chain_seq)
global_idx_start_list.append(global_idx_start_list[-1] + chain_length)
masked_chain_length_list.append(chain_length)
chain_coords = b[f'coords_chain_{letter}'] # this is a dictionary
chain_mask = np.ones(chain_length) # 1.0 for masked
if ca_only:
x_chain = np.array(chain_coords[f'CA_chain_{letter}']) # [chain_lenght,1,3] #CA_diff
if len(x_chain.shape) == 2:
x_chain = x_chain[:, None, :]
else:
x_chain = np.stack([chain_coords[c] for c in
[f'N_chain_{letter}', f'CA_chain_{letter}', f'C_chain_{letter}',
f'O_chain_{letter}']], 1) # [chain_lenght,4,3]
x_chain_list.append(x_chain)
chain_mask_list.append(chain_mask)
chain_seq_list.append(chain_seq)
chain_encoding_list.append(c * np.ones(np.array(chain_mask).shape[0]))
l1 += chain_length
residue_idx[i, l0:l1] = 100 * (c - 1) + np.arange(l0, l1)
l0 += chain_length
c += 1
fixed_position_mask = np.ones(chain_length)
if fixed_position_dict != None:
fixed_pos_list = fixed_position_dict[b['name']][letter]
if fixed_pos_list:
fixed_position_mask[np.array(fixed_pos_list) - 1] = 0.0
fixed_position_mask_list.append(fixed_position_mask)
omit_AA_mask_temp = np.zeros([chain_length, len(alphabet)], np.int32)
if omit_AA_dict != None:
for item in omit_AA_dict[b['name']][letter]:
idx_AA = np.array(item[0]) - 1
AA_idx = np.array([np.argwhere(np.array(list(alphabet)) == AA)[0][0] for AA in item[1]]).repeat(
idx_AA.shape[0])
idx_ = np.array([[a, b] for a in idx_AA for b in AA_idx])
omit_AA_mask_temp[idx_[:, 0], idx_[:, 1]] = 1
omit_AA_mask_list.append(omit_AA_mask_temp)
pssm_coef = np.zeros(chain_length)
pssm_bias = np.zeros([chain_length, 21])
pssm_log_odds = 10000.0 * np.ones([chain_length, 21])
if pssm_dict:
if pssm_dict[b['name']][letter]:
pssm_coef = pssm_dict[b['name']][letter]['pssm_coef']
pssm_bias = pssm_dict[b['name']][letter]['pssm_bias']
pssm_log_odds = pssm_dict[b['name']][letter]['pssm_log_odds']
pssm_coef_list.append(pssm_coef)
pssm_bias_list.append(pssm_bias)
pssm_log_odds_list.append(pssm_log_odds)
if bias_by_res_dict:
bias_by_res_list.append(bias_by_res_dict[b['name']][letter])
else:
bias_by_res_list.append(np.zeros([chain_length, 21]))
letter_list_np = np.array(letter_list)
tied_pos_list_of_lists = []
tied_beta = np.ones(L_max)
if tied_positions_dict != None:
tied_pos_list = tied_positions_dict[b['name']]
if tied_pos_list:
set_chains_tied = set(list(itertools.chain(*[list(item) for item in tied_pos_list])))
for tied_item in tied_pos_list:
one_list = []
for k, v in tied_item.items():
start_idx = global_idx_start_list[np.argwhere(letter_list_np == k)[0][0]]
if isinstance(v[0], list):
for v_count in range(len(v[0])):
one_list.append(start_idx + v[0][v_count] - 1) # make 0 to be the first
tied_beta[start_idx + v[0][v_count] - 1] = v[1][v_count]
else:
for v_ in v:
one_list.append(start_idx + v_ - 1) # make 0 to be the first
tied_pos_list_of_lists.append(one_list)
tied_pos_list_of_lists_list.append(tied_pos_list_of_lists)
x = np.concatenate(x_chain_list, 0) # [L, 4, 3]
all_sequence = "".join(chain_seq_list)
m = np.concatenate(chain_mask_list, 0) # [L,], 1.0 for places that need to be predicted
chain_encoding = np.concatenate(chain_encoding_list, 0)
m_pos = np.concatenate(fixed_position_mask_list, 0) # [L,], 1.0 for places that need to be predicted
pssm_coef_ = np.concatenate(pssm_coef_list, 0) # [L,], 1.0 for places that need to be predicted
pssm_bias_ = np.concatenate(pssm_bias_list, 0) # [L,], 1.0 for places that need to be predicted
pssm_log_odds_ = np.concatenate(pssm_log_odds_list, 0) # [L,], 1.0 for places that need to be predicted
bias_by_res_ = np.concatenate(bias_by_res_list,
0) # [L,21], 0.0 for places where AA frequencies don't need to be tweaked
l = len(all_sequence)
x_pad = np.pad(x, [[0, L_max - l], [0, 0], [0, 0]], 'constant', constant_values=(np.nan,))
X[i, :, :, :] = x_pad
m_pad = np.pad(m, [[0, L_max - l]], 'constant', constant_values=(0.0,))
m_pos_pad = np.pad(m_pos, [[0, L_max - l]], 'constant', constant_values=(0.0,))
omit_AA_mask_pad = np.pad(np.concatenate(omit_AA_mask_list, 0), [[0, L_max - l]], 'constant',
constant_values=(0.0,))
chain_M[i, :] = m_pad
chain_M_pos[i, :] = m_pos_pad
omit_AA_mask[i,] = omit_AA_mask_pad
chain_encoding_pad = np.pad(chain_encoding, [[0, L_max - l]], 'constant', constant_values=(0.0,))
chain_encoding_all[i, :] = chain_encoding_pad
pssm_coef_pad = np.pad(pssm_coef_, [[0, L_max - l]], 'constant', constant_values=(0.0,))
pssm_bias_pad = np.pad(pssm_bias_, [[0, L_max - l], [0, 0]], 'constant', constant_values=(0.0,))
pssm_log_odds_pad = np.pad(pssm_log_odds_, [[0, L_max - l], [0, 0]], 'constant', constant_values=(0.0,))
pssm_coef_all[i, :] = pssm_coef_pad
pssm_bias_all[i, :] = pssm_bias_pad
pssm_log_odds_all[i, :] = pssm_log_odds_pad
bias_by_res_pad = np.pad(bias_by_res_, [[0, L_max - l], [0, 0]], 'constant', constant_values=(0.0,))
bias_by_res_all[i, :] = bias_by_res_pad
# Convert to labels
indices = np.asarray([alphabet.index(a) for a in all_sequence], dtype=np.int32)
S[i, :l] = indices
letter_list_list.append(letter_list)
visible_list_list.append(visible_list)
masked_list_list.append(masked_list)
masked_chain_length_list_list.append(masked_chain_length_list)
isnan = np.isnan(X)
mask = np.isfinite(np.sum(X, (2, 3))).astype(np.float32)
X[isnan] = 0.
# Conversion
pssm_coef_all = torch.from_numpy(pssm_coef_all).to(dtype=torch.float32, device=device)
pssm_bias_all = torch.from_numpy(pssm_bias_all).to(dtype=torch.float32, device=device)
pssm_log_odds_all = torch.from_numpy(pssm_log_odds_all).to(dtype=torch.float32, device=device)
tied_beta = torch.from_numpy(tied_beta).to(dtype=torch.float32, device=device)
jumps = ((residue_idx[:, 1:] - residue_idx[:, :-1]) == 1).astype(np.float32)
bias_by_res_all = torch.from_numpy(bias_by_res_all).to(dtype=torch.float32, device=device)
phi_mask = np.pad(jumps, [[0, 0], [1, 0]])
psi_mask = np.pad(jumps, [[0, 0], [0, 1]])
omega_mask = np.pad(jumps, [[0, 0], [0, 1]])
dihedral_mask = np.concatenate([phi_mask[:, :, None], psi_mask[:, :, None], omega_mask[:, :, None]], -1) # [B,L,3]
dihedral_mask = torch.from_numpy(dihedral_mask).to(dtype=torch.float32, device=device)
residue_idx = torch.from_numpy(residue_idx).to(dtype=torch.long, device=device)
S = torch.from_numpy(S).to(dtype=torch.long, device=device)
X = torch.from_numpy(X).to(dtype=torch.float32, device=device)
mask = torch.from_numpy(mask).to(dtype=torch.float32, device=device)
chain_M = torch.from_numpy(chain_M).to(dtype=torch.float32, device=device)
chain_M_pos = torch.from_numpy(chain_M_pos).to(dtype=torch.float32, device=device)
omit_AA_mask = torch.from_numpy(omit_AA_mask).to(dtype=torch.float32, device=device)
chain_encoding_all = torch.from_numpy(chain_encoding_all).to(dtype=torch.long, device=device)
if ca_only:
X_out = X[:, :, 0]
else:
X_out = X
return X_out, S, mask, lengths, chain_M, chain_encoding_all, letter_list_list, visible_list_list, masked_list_list, masked_chain_length_list_list, chain_M_pos, omit_AA_mask, residue_idx, dihedral_mask, tied_pos_list_of_lists_list, pssm_coef_all, pssm_bias_all, pssm_log_odds_all, bias_by_res_all, tied_beta
def loss_nll(S, log_probs, mask):
""" Negative log probabilities """
criterion = torch.nn.NLLLoss(reduction='none')
loss = criterion(
log_probs.contiguous().view(-1, log_probs.size(-1)), S.contiguous().view(-1)
).view(S.size())
loss_av = torch.sum(loss * mask) / torch.sum(mask)
return loss, loss_av
def loss_smoothed(S, log_probs, mask, weight=0.1):
""" Negative log probabilities """
S_onehot = torch.nn.functional.one_hot(S, 21).float()
# Label smoothing
S_onehot = S_onehot + weight / float(S_onehot.size(-1))
S_onehot = S_onehot / S_onehot.sum(-1, keepdim=True)
loss = -(S_onehot * log_probs).sum(-1)
loss_av = torch.sum(loss * mask) / torch.sum(mask)
return loss, loss_av
class StructureDataset():
def __init__(self, jsonl_file, verbose=True, truncate=None, max_length=100,
alphabet='ACDEFGHIKLMNPQRSTVWYX-'):
alphabet_set = set([a for a in alphabet])
discard_count = {
'bad_chars': 0,
'too_long': 0,
'bad_seq_length': 0
}
with open(jsonl_file) as f:
self.data = []
lines = f.readlines()
start = time.time()
for i, line in enumerate(lines):
entry = json.loads(line)
seq = entry['seq']
name = entry['name']
# Convert raw coords to np arrays
# for key, val in entry['coords'].items():
# entry['coords'][key] = np.asarray(val)
# Check if in alphabet
bad_chars = set([s for s in seq]).difference(alphabet_set)
if len(bad_chars) == 0:
if len(entry['seq']) <= max_length:
if True:
self.data.append(entry)
else:
discard_count['bad_seq_length'] += 1
else:
discard_count['too_long'] += 1
else:
print(name, bad_chars, entry['seq'])
discard_count['bad_chars'] += 1
# Truncate early
if truncate is not None and len(self.data) == truncate:
return
if verbose and (i + 1) % 1000 == 0:
elapsed = time.time() - start
print('{} entries ({} loaded) in {:.1f} s'.format(len(self.data), i + 1, elapsed))
print('discarded', discard_count)
def __len__(self):
return len(self.data)
def __getitem__(self, idx):
return self.data[idx]
class StructureDatasetPDB():
def __init__(self, pdb_dict_list, verbose=True, truncate=None, max_length=100,
alphabet='ACDEFGHIKLMNPQRSTVWYX-'):
alphabet_set = set([a for a in alphabet])
discard_count = {
'bad_chars': 0,
'too_long': 0,
'bad_seq_length': 0
}
self.data = []
start = time.time()
for i, entry in enumerate(pdb_dict_list):
seq = entry['seq']
name = entry['name']
bad_chars = set([s for s in seq]).difference(alphabet_set)
if len(bad_chars) == 0:
if len(entry['seq']) <= max_length:
self.data.append(entry)
else:
discard_count['too_long'] += 1
else:
discard_count['bad_chars'] += 1
# Truncate early
if truncate is not None and len(self.data) == truncate:
return
if verbose and (i + 1) % 1000 == 0:
elapsed = time.time() - start
# print('Discarded', discard_count)
def __len__(self):
return len(self.data)
def __getitem__(self, idx):
return self.data[idx]
class StructureLoader():
def __init__(self, dataset, batch_size=100, shuffle=True,
collate_fn=lambda x: x, drop_last=False):
self.dataset = dataset
self.size = len(dataset)
self.lengths = [len(dataset[i]['seq']) for i in range(self.size)]
self.batch_size = batch_size
sorted_ix = np.argsort(self.lengths)
# Cluster into batches of similar sizes
clusters, batch = [], []
batch_max = 0
for ix in sorted_ix:
size = self.lengths[ix]
if size * (len(batch) + 1) <= self.batch_size:
batch.append(ix)
batch_max = size
else:
clusters.append(batch)
batch, batch_max = [], 0
if len(batch) > 0:
clusters.append(batch)
self.clusters = clusters
def __len__(self):
return len(self.clusters)
def __iter__(self):
np.random.shuffle(self.clusters)
for b_idx in self.clusters:
batch = [self.dataset[i] for i in b_idx]
yield batch
# The following gather functions
def gather_edges(edges, neighbor_idx):
# Features [B,N,N,C] at Neighbor indices [B,N,K] => Neighbor features [B,N,K,C]
neighbors = neighbor_idx.unsqueeze(-1).expand(-1, -1, -1, edges.size(-1))
edge_features = torch.gather(edges, 2, neighbors)
return edge_features
def gather_nodes(nodes, neighbor_idx):
# Features [B,N,C] at Neighbor indices [B,N,K] => [B,N,K,C]
# Flatten and expand indices per batch [B,N,K] => [B,NK] => [B,NK,C]
neighbors_flat = neighbor_idx.view((neighbor_idx.shape[0], -1))
neighbors_flat = neighbors_flat.unsqueeze(-1).expand(-1, -1, nodes.size(2))
# Gather and re-pack
neighbor_features = torch.gather(nodes, 1, neighbors_flat)
neighbor_features = neighbor_features.view(list(neighbor_idx.shape)[:3] + [-1])
return neighbor_features
def gather_nodes_t(nodes, neighbor_idx):
# Features [B,N,C] at Neighbor index [B,K] => Neighbor features[B,K,C]
idx_flat = neighbor_idx.unsqueeze(-1).expand(-1, -1, nodes.size(2))
neighbor_features = torch.gather(nodes, 1, idx_flat)
return neighbor_features
def cat_neighbors_nodes(h_nodes, h_neighbors, E_idx):
h_nodes = gather_nodes(h_nodes, E_idx)
h_nn = torch.cat([h_neighbors, h_nodes], -1)
return h_nn
class EncLayer(nn.Module):
def __init__(self, num_hidden, num_in, dropout=0.1, num_heads=None, scale=30):
super(EncLayer, self).__init__()
self.num_hidden = num_hidden
self.num_in = num_in
self.scale = scale
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
self.dropout3 = nn.Dropout(dropout)
self.norm1 = nn.LayerNorm(num_hidden)
self.norm2 = nn.LayerNorm(num_hidden)
self.norm3 = nn.LayerNorm(num_hidden)
self.W1 = nn.Linear(num_hidden + num_in, num_hidden, bias=True)
self.W2 = nn.Linear(num_hidden, num_hidden, bias=True)
self.W3 = nn.Linear(num_hidden, num_hidden, bias=True)
self.W11 = nn.Linear(num_hidden + num_in, num_hidden, bias=True)
self.W12 = nn.Linear(num_hidden, num_hidden, bias=True)
self.W13 = nn.Linear(num_hidden, num_hidden, bias=True)
self.act = torch.nn.GELU()
self.dense = PositionWiseFeedForward(num_hidden, num_hidden * 4)
def forward(self, h_V, h_E, E_idx, mask_V=None, mask_attend=None):
""" Parallel computation of full transformer layer """
h_EV = cat_neighbors_nodes(h_V, h_E, E_idx)
h_V_expand = h_V.unsqueeze(-2).expand(-1, -1, h_EV.size(-2), -1)
h_EV = torch.cat([h_V_expand, h_EV], -1)
h_message = self.W3(self.act(self.W2(self.act(self.W1(h_EV)))))
if mask_attend is not None:
h_message = mask_attend.unsqueeze(-1) * h_message
dh = torch.sum(h_message, -2) / self.scale
h_V = self.norm1(h_V + self.dropout1(dh))
dh = self.dense(h_V)
h_V = self.norm2(h_V + self.dropout2(dh))
if mask_V is not None:
mask_V = mask_V.unsqueeze(-1)
h_V = mask_V * h_V
h_EV = cat_neighbors_nodes(h_V, h_E, E_idx)
h_V_expand = h_V.unsqueeze(-2).expand(-1, -1, h_EV.size(-2), -1)
h_EV = torch.cat([h_V_expand, h_EV], -1)
h_message = self.W13(self.act(self.W12(self.act(self.W11(h_EV)))))
h_E = self.norm3(h_E + self.dropout3(h_message))
return h_V, h_E
class DecLayer(nn.Module):
def __init__(self, num_hidden, num_in, dropout=0.1, num_heads=None, scale=30):
super(DecLayer, self).__init__()
self.num_hidden = num_hidden
self.num_in = num_in
self.scale = scale
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
self.norm1 = nn.LayerNorm(num_hidden)
self.norm2 = nn.LayerNorm(num_hidden)
self.W1 = nn.Linear(num_hidden + num_in, num_hidden, bias=True)
self.W2 = nn.Linear(num_hidden, num_hidden, bias=True)
self.W3 = nn.Linear(num_hidden, num_hidden, bias=True)
self.act = torch.nn.GELU()
self.dense = PositionWiseFeedForward(num_hidden, num_hidden * 4)
def forward(self, h_V, h_E, mask_V=None, mask_attend=None):
""" Parallel computation of full transformer layer """
# Concatenate h_V_i to h_E_ij
h_V_expand = h_V.unsqueeze(-2).expand(-1, -1, h_E.size(-2), -1)
h_EV = torch.cat([h_V_expand, h_E], -1)
h_message = self.W3(self.act(self.W2(self.act(self.W1(h_EV)))))
if mask_attend is not None:
h_message = mask_attend.unsqueeze(-1) * h_message
dh = torch.sum(h_message, -2) / self.scale
h_V = self.norm1(h_V + self.dropout1(dh))
# Position-wise feedforward
dh = self.dense(h_V)
h_V = self.norm2(h_V + self.dropout2(dh))
if mask_V is not None:
mask_V = mask_V.unsqueeze(-1)
h_V = mask_V * h_V
return h_V
class PositionWiseFeedForward(nn.Module):
def __init__(self, num_hidden, num_ff):
super(PositionWiseFeedForward, self).__init__()
self.W_in = nn.Linear(num_hidden, num_ff, bias=True)
self.W_out = nn.Linear(num_ff, num_hidden, bias=True)
self.act = torch.nn.GELU()
def forward(self, h_V):
h = self.act(self.W_in(h_V))
h = self.W_out(h)
return h
class PositionalEncodings(nn.Module):
def __init__(self, num_embeddings, max_relative_feature=32):
super(PositionalEncodings, self).__init__()
self.num_embeddings = num_embeddings
self.max_relative_feature = max_relative_feature
self.linear = nn.Linear(2 * max_relative_feature + 1 + 1, num_embeddings)
def forward(self, offset, mask):
d = torch.clip(offset + self.max_relative_feature, 0, 2 * self.max_relative_feature) * mask + (1 - mask) * (
2 * self.max_relative_feature + 1)
d_onehot = torch.nn.functional.one_hot(d, 2 * self.max_relative_feature + 1 + 1)
E = self.linear(d_onehot.float())
return E
class CA_ProteinFeatures(nn.Module):
def __init__(self, edge_features, node_features, num_positional_embeddings=16,
num_rbf=16, top_k=30, augment_eps=0., num_chain_embeddings=16):
""" Extract protein features """
super(CA_ProteinFeatures, self).__init__()
self.edge_features = edge_features
self.node_features = node_features
self.top_k = top_k
self.augment_eps = augment_eps
self.num_rbf = num_rbf
self.num_positional_embeddings = num_positional_embeddings
# Positional encoding
self.embeddings = PositionalEncodings(num_positional_embeddings)
# Normalization and embedding
node_in, edge_in = 3, num_positional_embeddings + num_rbf * 9 + 7
self.node_embedding = nn.Linear(node_in, node_features, bias=False) # NOT USED
self.edge_embedding = nn.Linear(edge_in, edge_features, bias=False)
self.norm_nodes = nn.LayerNorm(node_features)
self.norm_edges = nn.LayerNorm(edge_features)
def _quaternions(self, R):
""" Convert a batch of 3D rotations [R] to quaternions [Q]
R [...,3,3]
Q [...,4]
"""
# Simple Wikipedia version
# en.wikipedia.org/wiki/Rotation_matrix#Quaternion
# For other options see math.stackexchange.com/questions/2074316/calculating-rotation-axis-from-rotation-matrix
diag = torch.diagonal(R, dim1=-2, dim2=-1)
Rxx, Ryy, Rzz = diag.unbind(-1)
magnitudes = 0.5 * torch.sqrt(torch.abs(1 + torch.stack([
Rxx - Ryy - Rzz,
- Rxx + Ryy - Rzz,
- Rxx - Ryy + Rzz
], -1)))
_R = lambda i, j: R[:, :, :, i, j]
signs = torch.sign(torch.stack([
_R(2, 1) - _R(1, 2),
_R(0, 2) - _R(2, 0),
_R(1, 0) - _R(0, 1)
], -1))
xyz = signs * magnitudes
# The relu enforces a non-negative trace
w = torch.sqrt(F.relu(1 + diag.sum(-1, keepdim=True))) / 2.
Q = torch.cat((xyz, w), -1)
Q = F.normalize(Q, dim=-1)
return Q
def _orientations_coarse(self, X, E_idx, eps=1e-6):
dX = X[:, 1:, :] - X[:, :-1, :]
dX_norm = torch.norm(dX, dim=-1)
dX_mask = (3.6 < dX_norm) & (dX_norm < 4.0) # exclude CA-CA jumps
dX = dX * dX_mask[:, :, None]
U = F.normalize(dX, dim=-1)
u_2 = U[:, :-2, :]
u_1 = U[:, 1:-1, :]
u_0 = U[:, 2:, :]
# Backbone normals
n_2 = F.normalize(torch.cross(u_2, u_1), dim=-1)
n_1 = F.normalize(torch.cross(u_1, u_0), dim=-1)
# Bond angle calculation
cosA = -(u_1 * u_0).sum(-1)
cosA = torch.clamp(cosA, -1 + eps, 1 - eps)
A = torch.acos(cosA)
# Angle between normals
cosD = (n_2 * n_1).sum(-1)
cosD = torch.clamp(cosD, -1 + eps, 1 - eps)
D = torch.sign((u_2 * n_1).sum(-1)) * torch.acos(cosD)
# Backbone features
AD_features = torch.stack((torch.cos(A), torch.sin(A) * torch.cos(D), torch.sin(A) * torch.sin(D)), 2)
AD_features = F.pad(AD_features, (0, 0, 1, 2), 'constant', 0)
# Build relative orientations
o_1 = F.normalize(u_2 - u_1, dim=-1)
O = torch.stack((o_1, n_2, torch.cross(o_1, n_2)), 2)
O = O.view(list(O.shape[:2]) + [9])
O = F.pad(O, (0, 0, 1, 2), 'constant', 0)
O_neighbors = gather_nodes(O, E_idx)
X_neighbors = gather_nodes(X, E_idx)
# Re-view as rotation matrices
O = O.view(list(O.shape[:2]) + [3, 3])
O_neighbors = O_neighbors.view(list(O_neighbors.shape[:3]) + [3, 3])
# Rotate into local reference frames
dX = X_neighbors - X.unsqueeze(-2)
dU = torch.matmul(O.unsqueeze(2), dX.unsqueeze(-1)).squeeze(-1)
dU = F.normalize(dU, dim=-1)