jglaser commited on
Commit
28974f4
1 Parent(s): 1ee751f

use openfold features

Browse files
README.md CHANGED
@@ -9,7 +9,7 @@ tags:
9
 
10
  ## How to use the data sets
11
 
12
- This dataset contains more about 36,000 unique pairs of protein sequences and ligand SMILES, and the coordinates
13
  of their complexes from the PDB.
14
 
15
  SMILES are assumed to be tokenized by the regex from P. Schwaller.
@@ -35,9 +35,9 @@ are considered.
35
  Load a test/train split using
36
 
37
  ```
38
- from datasets import load_dataset
39
- train = load_dataset("jglaser/pdb_protein_ligand_complexes",split='train[:90%]')
40
- validation = load_dataset("jglaser/pdb_protein_ligand_complexes",split='train[90%:]')
41
  ```
42
 
43
  ### Manual update from PDB
 
9
 
10
  ## How to use the data sets
11
 
12
+ This dataset contains about 36,000 unique pairs of protein sequences and ligand SMILES, and the coordinates
13
  of their complexes from the PDB.
14
 
15
  SMILES are assumed to be tokenized by the regex from P. Schwaller.
 
35
  Load a test/train split using
36
 
37
  ```
38
+ import pandas as pd
39
+ train = pd.read_pickle('data/pdb_train.p')
40
+ test = pd.read_pickle('data/pdb_test.p')
41
  ```
42
 
43
  ### Manual update from PDB
data/pdb_test.p ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f5c03e3bf9c84e79db19f566794945c3f9441f6797b8c7af9cf050a7db8de88c
3
+ size 1428764270
data/pdb_train.p ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f837b7c209aea0cfdadd44b9478ae5f2b0ea12fc138a07c3542ec26d42780667
3
+ size 13251374040
openfold/.gitignore ADDED
@@ -0,0 +1 @@
 
 
1
+ __pycache__/*
openfold/data_transforms.py ADDED
@@ -0,0 +1,1211 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2021 AlQuraishi Laboratory
2
+ # Copyright 2021 DeepMind Technologies Limited
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+
16
+ import itertools
17
+ from functools import reduce, wraps
18
+ from operator import add
19
+
20
+ import numpy as np
21
+ import torch
22
+
23
+ from . import residue_constants as rc
24
+ from .rigid_utils import Rotation, Rigid
25
+ from .tensor_utils import (
26
+ tree_map,
27
+ tensor_tree_map,
28
+ batched_gather,
29
+ )
30
+
31
+
32
+ MSA_FEATURE_NAMES = [
33
+ "msa",
34
+ "deletion_matrix",
35
+ "msa_mask",
36
+ "msa_row_mask",
37
+ "bert_mask",
38
+ "true_msa",
39
+ ]
40
+
41
+
42
+ def cast_to_64bit_ints(protein):
43
+ # We keep all ints as int64
44
+ for k, v in protein.items():
45
+ if v.dtype == torch.int32:
46
+ protein[k] = v.type(torch.int64)
47
+
48
+ return protein
49
+
50
+
51
+ def make_one_hot(x, num_classes):
52
+ x_one_hot = torch.zeros(*x.shape, num_classes, device=x.device)
53
+ x_one_hot.scatter_(-1, x.unsqueeze(-1), 1)
54
+ return x_one_hot
55
+
56
+
57
+ def make_seq_mask(protein):
58
+ protein["seq_mask"] = torch.ones(
59
+ protein["aatype"].shape, dtype=torch.float32
60
+ )
61
+ return protein
62
+
63
+
64
+ def make_template_mask(protein):
65
+ protein["template_mask"] = torch.ones(
66
+ protein["template_aatype"].shape[0], dtype=torch.float32
67
+ )
68
+ return protein
69
+
70
+
71
+ def curry1(f):
72
+ """Supply all arguments but the first."""
73
+ @wraps(f)
74
+ def fc(*args, **kwargs):
75
+ return lambda x: f(x, *args, **kwargs)
76
+
77
+ return fc
78
+
79
+
80
+ def make_all_atom_aatype(protein):
81
+ protein["all_atom_aatype"] = protein["aatype"]
82
+ return protein
83
+
84
+
85
+ def fix_templates_aatype(protein):
86
+ # Map one-hot to indices
87
+ num_templates = protein["template_aatype"].shape[0]
88
+ if(num_templates > 0):
89
+ protein["template_aatype"] = torch.argmax(
90
+ protein["template_aatype"], dim=-1
91
+ )
92
+ # Map hhsearch-aatype to our aatype.
93
+ new_order_list = rc.MAP_HHBLITS_AATYPE_TO_OUR_AATYPE
94
+ new_order = torch.tensor(
95
+ new_order_list, dtype=torch.int64, device=protein["aatype"].device,
96
+ ).expand(num_templates, -1)
97
+ protein["template_aatype"] = torch.gather(
98
+ new_order, 1, index=protein["template_aatype"]
99
+ )
100
+
101
+ return protein
102
+
103
+
104
+ def correct_msa_restypes(protein):
105
+ """Correct MSA restype to have the same order as rc."""
106
+ new_order_list = rc.MAP_HHBLITS_AATYPE_TO_OUR_AATYPE
107
+ new_order = torch.tensor(
108
+ [new_order_list] * protein["msa"].shape[1],
109
+ device=protein["msa"].device,
110
+ ).transpose(0, 1)
111
+ protein["msa"] = torch.gather(new_order, 0, protein["msa"])
112
+
113
+ perm_matrix = np.zeros((22, 22), dtype=np.float32)
114
+ perm_matrix[range(len(new_order_list)), new_order_list] = 1.0
115
+
116
+ for k in protein:
117
+ if "profile" in k:
118
+ num_dim = protein[k].shape.as_list()[-1]
119
+ assert num_dim in [
120
+ 20,
121
+ 21,
122
+ 22,
123
+ ], "num_dim for %s out of expected range: %s" % (k, num_dim)
124
+ protein[k] = torch.dot(protein[k], perm_matrix[:num_dim, :num_dim])
125
+
126
+ return protein
127
+
128
+
129
+ def squeeze_features(protein):
130
+ """Remove singleton and repeated dimensions in protein features."""
131
+ protein["aatype"] = torch.argmax(protein["aatype"], dim=-1)
132
+ for k in [
133
+ "domain_name",
134
+ "msa",
135
+ "num_alignments",
136
+ "seq_length",
137
+ "sequence",
138
+ "superfamily",
139
+ "deletion_matrix",
140
+ "resolution",
141
+ "between_segment_residues",
142
+ "residue_index",
143
+ "template_all_atom_mask",
144
+ ]:
145
+ if k in protein:
146
+ final_dim = protein[k].shape[-1]
147
+ if isinstance(final_dim, int) and final_dim == 1:
148
+ if torch.is_tensor(protein[k]):
149
+ protein[k] = torch.squeeze(protein[k], dim=-1)
150
+ else:
151
+ protein[k] = np.squeeze(protein[k], axis=-1)
152
+
153
+ for k in ["seq_length", "num_alignments"]:
154
+ if k in protein:
155
+ protein[k] = protein[k][0]
156
+
157
+ return protein
158
+
159
+
160
+ @curry1
161
+ def randomly_replace_msa_with_unknown(protein, replace_proportion):
162
+ """Replace a portion of the MSA with 'X'."""
163
+ msa_mask = torch.rand(protein["msa"].shape) < replace_proportion
164
+ x_idx = 20
165
+ gap_idx = 21
166
+ msa_mask = torch.logical_and(msa_mask, protein["msa"] != gap_idx)
167
+ protein["msa"] = torch.where(
168
+ msa_mask,
169
+ torch.ones_like(protein["msa"]) * x_idx,
170
+ protein["msa"]
171
+ )
172
+ aatype_mask = torch.rand(protein["aatype"].shape) < replace_proportion
173
+
174
+ protein["aatype"] = torch.where(
175
+ aatype_mask,
176
+ torch.ones_like(protein["aatype"]) * x_idx,
177
+ protein["aatype"],
178
+ )
179
+ return protein
180
+
181
+
182
+ @curry1
183
+ def sample_msa(protein, max_seq, keep_extra, seed=None):
184
+ """Sample MSA randomly, remaining sequences are stored are stored as `extra_*`."""
185
+ num_seq = protein["msa"].shape[0]
186
+ g = torch.Generator(device=protein["msa"].device)
187
+ if seed is not None:
188
+ g.manual_seed(seed)
189
+ shuffled = torch.randperm(num_seq - 1, generator=g) + 1
190
+ index_order = torch.cat(
191
+ (torch.tensor([0], device=shuffled.device), shuffled),
192
+ dim=0
193
+ )
194
+ num_sel = min(max_seq, num_seq)
195
+ sel_seq, not_sel_seq = torch.split(
196
+ index_order, [num_sel, num_seq - num_sel]
197
+ )
198
+
199
+ for k in MSA_FEATURE_NAMES:
200
+ if k in protein:
201
+ if keep_extra:
202
+ protein["extra_" + k] = torch.index_select(
203
+ protein[k], 0, not_sel_seq
204
+ )
205
+ protein[k] = torch.index_select(protein[k], 0, sel_seq)
206
+
207
+ return protein
208
+
209
+
210
+ @curry1
211
+ def add_distillation_flag(protein, distillation):
212
+ protein['is_distillation'] = distillation
213
+ return protein
214
+
215
+ @curry1
216
+ def sample_msa_distillation(protein, max_seq):
217
+ if(protein["is_distillation"] == 1):
218
+ protein = sample_msa(max_seq, keep_extra=False)(protein)
219
+ return protein
220
+
221
+
222
+ @curry1
223
+ def crop_extra_msa(protein, max_extra_msa):
224
+ num_seq = protein["extra_msa"].shape[0]
225
+ num_sel = min(max_extra_msa, num_seq)
226
+ select_indices = torch.randperm(num_seq)[:num_sel]
227
+ for k in MSA_FEATURE_NAMES:
228
+ if "extra_" + k in protein:
229
+ protein["extra_" + k] = torch.index_select(
230
+ protein["extra_" + k], 0, select_indices
231
+ )
232
+
233
+ return protein
234
+
235
+
236
+ def delete_extra_msa(protein):
237
+ for k in MSA_FEATURE_NAMES:
238
+ if "extra_" + k in protein:
239
+ del protein["extra_" + k]
240
+ return protein
241
+
242
+
243
+ # Not used in inference
244
+ @curry1
245
+ def block_delete_msa(protein, config):
246
+ num_seq = protein["msa"].shape[0]
247
+ block_num_seq = torch.floor(
248
+ torch.tensor(num_seq, dtype=torch.float32, device=protein["msa"].device)
249
+ * config.msa_fraction_per_block
250
+ ).to(torch.int32)
251
+
252
+ if config.randomize_num_blocks:
253
+ nb = torch.distributions.uniform.Uniform(
254
+ 0, config.num_blocks + 1
255
+ ).sample()
256
+ else:
257
+ nb = config.num_blocks
258
+
259
+ del_block_starts = torch.distributions.Uniform(0, num_seq).sample(nb)
260
+ del_blocks = del_block_starts[:, None] + torch.range(block_num_seq)
261
+ del_blocks = torch.clip(del_blocks, 0, num_seq - 1)
262
+ del_indices = torch.unique(torch.sort(torch.reshape(del_blocks, [-1])))[0]
263
+
264
+ # Make sure we keep the original sequence
265
+ combined = torch.cat((torch.range(1, num_seq)[None], del_indices[None]))
266
+ uniques, counts = combined.unique(return_counts=True)
267
+ difference = uniques[counts == 1]
268
+ intersection = uniques[counts > 1]
269
+ keep_indices = torch.squeeze(difference, 0)
270
+
271
+ for k in MSA_FEATURE_NAMES:
272
+ if k in protein:
273
+ protein[k] = torch.gather(protein[k], keep_indices)
274
+
275
+ return protein
276
+
277
+
278
+ @curry1
279
+ def nearest_neighbor_clusters(protein, gap_agreement_weight=0.0):
280
+ weights = torch.cat(
281
+ [
282
+ torch.ones(21, device=protein["msa"].device),
283
+ gap_agreement_weight * torch.ones(1, device=protein["msa"].device),
284
+ torch.zeros(1, device=protein["msa"].device)
285
+ ],
286
+ 0,
287
+ )
288
+
289
+ # Make agreement score as weighted Hamming distance
290
+ msa_one_hot = make_one_hot(protein["msa"], 23)
291
+ sample_one_hot = protein["msa_mask"][:, :, None] * msa_one_hot
292
+ extra_msa_one_hot = make_one_hot(protein["extra_msa"], 23)
293
+ extra_one_hot = protein["extra_msa_mask"][:, :, None] * extra_msa_one_hot
294
+
295
+ num_seq, num_res, _ = sample_one_hot.shape
296
+ extra_num_seq, _, _ = extra_one_hot.shape
297
+
298
+ # Compute tf.einsum('mrc,nrc,c->mn', sample_one_hot, extra_one_hot, weights)
299
+ # in an optimized fashion to avoid possible memory or computation blowup.
300
+ agreement = torch.matmul(
301
+ torch.reshape(extra_one_hot, [extra_num_seq, num_res * 23]),
302
+ torch.reshape(
303
+ sample_one_hot * weights, [num_seq, num_res * 23]
304
+ ).transpose(0, 1),
305
+ )
306
+
307
+ # Assign each sequence in the extra sequences to the closest MSA sample
308
+ protein["extra_cluster_assignment"] = torch.argmax(agreement, dim=1).to(
309
+ torch.int64
310
+ )
311
+
312
+ return protein
313
+
314
+
315
+ def unsorted_segment_sum(data, segment_ids, num_segments):
316
+ """
317
+ Computes the sum along segments of a tensor. Similar to
318
+ tf.unsorted_segment_sum, but only supports 1-D indices.
319
+
320
+ :param data: A tensor whose segments are to be summed.
321
+ :param segment_ids: The 1-D segment indices tensor.
322
+ :param num_segments: The number of segments.
323
+ :return: A tensor of same data type as the data argument.
324
+ """
325
+ assert (
326
+ len(segment_ids.shape) == 1 and
327
+ segment_ids.shape[0] == data.shape[0]
328
+ )
329
+ segment_ids = segment_ids.view(
330
+ segment_ids.shape[0], *((1,) * len(data.shape[1:]))
331
+ )
332
+ segment_ids = segment_ids.expand(data.shape)
333
+ shape = [num_segments] + list(data.shape[1:])
334
+ tensor = (
335
+ torch.zeros(*shape, device=segment_ids.device)
336
+ .scatter_add_(0, segment_ids, data.float())
337
+ )
338
+ tensor = tensor.type(data.dtype)
339
+ return tensor
340
+
341
+
342
+ @curry1
343
+ def summarize_clusters(protein):
344
+ """Produce profile and deletion_matrix_mean within each cluster."""
345
+ num_seq = protein["msa"].shape[0]
346
+
347
+ def csum(x):
348
+ return unsorted_segment_sum(
349
+ x, protein["extra_cluster_assignment"], num_seq
350
+ )
351
+
352
+ mask = protein["extra_msa_mask"]
353
+ mask_counts = 1e-6 + protein["msa_mask"] + csum(mask) # Include center
354
+
355
+ msa_sum = csum(mask[:, :, None] * make_one_hot(protein["extra_msa"], 23))
356
+ msa_sum += make_one_hot(protein["msa"], 23) # Original sequence
357
+ protein["cluster_profile"] = msa_sum / mask_counts[:, :, None]
358
+ del msa_sum
359
+
360
+ del_sum = csum(mask * protein["extra_deletion_matrix"])
361
+ del_sum += protein["deletion_matrix"] # Original sequence
362
+ protein["cluster_deletion_mean"] = del_sum / mask_counts
363
+ del del_sum
364
+
365
+ return protein
366
+
367
+
368
+ def make_msa_mask(protein):
369
+ """Mask features are all ones, but will later be zero-padded."""
370
+ protein["msa_mask"] = torch.ones(protein["msa"].shape, dtype=torch.float32)
371
+ protein["msa_row_mask"] = torch.ones(
372
+ (protein["msa"].shape[0]), dtype=torch.float32
373
+ )
374
+ return protein
375
+
376
+
377
+ def pseudo_beta_fn(aatype, all_atom_positions, all_atom_mask):
378
+ """Create pseudo beta features."""
379
+ is_gly = torch.eq(aatype, rc.restype_order["G"])
380
+ ca_idx = rc.atom_order["CA"]
381
+ cb_idx = rc.atom_order["CB"]
382
+ pseudo_beta = torch.where(
383
+ torch.tile(is_gly[..., None], [1] * len(is_gly.shape) + [3]),
384
+ all_atom_positions[..., ca_idx, :],
385
+ all_atom_positions[..., cb_idx, :],
386
+ )
387
+
388
+ if all_atom_mask is not None:
389
+ pseudo_beta_mask = torch.where(
390
+ is_gly, all_atom_mask[..., ca_idx], all_atom_mask[..., cb_idx]
391
+ )
392
+ return pseudo_beta, pseudo_beta_mask
393
+ else:
394
+ return pseudo_beta
395
+
396
+
397
+ @curry1
398
+ def make_pseudo_beta(protein, prefix=""):
399
+ """Create pseudo-beta (alpha for glycine) position and mask."""
400
+ assert prefix in ["", "template_"]
401
+ (
402
+ protein[prefix + "pseudo_beta"],
403
+ protein[prefix + "pseudo_beta_mask"],
404
+ ) = pseudo_beta_fn(
405
+ protein["template_aatype" if prefix else "aatype"],
406
+ protein[prefix + "all_atom_positions"],
407
+ protein["template_all_atom_mask" if prefix else "all_atom_mask"],
408
+ )
409
+ return protein
410
+
411
+
412
+ @curry1
413
+ def add_constant_field(protein, key, value):
414
+ protein[key] = torch.tensor(value, device=protein["msa"].device)
415
+ return protein
416
+
417
+
418
+ def shaped_categorical(probs, epsilon=1e-10):
419
+ ds = probs.shape
420
+ num_classes = ds[-1]
421
+ distribution = torch.distributions.categorical.Categorical(
422
+ torch.reshape(probs + epsilon, [-1, num_classes])
423
+ )
424
+ counts = distribution.sample()
425
+ return torch.reshape(counts, ds[:-1])
426
+
427
+
428
+ def make_hhblits_profile(protein):
429
+ """Compute the HHblits MSA profile if not already present."""
430
+ if "hhblits_profile" in protein:
431
+ return protein
432
+
433
+ # Compute the profile for every residue (over all MSA sequences).
434
+ msa_one_hot = make_one_hot(protein["msa"], 22)
435
+
436
+ protein["hhblits_profile"] = torch.mean(msa_one_hot, dim=0)
437
+ return protein
438
+
439
+
440
+ @curry1
441
+ def make_masked_msa(protein, config, replace_fraction):
442
+ """Create data for BERT on raw MSA."""
443
+ # Add a random amino acid uniformly.
444
+ random_aa = torch.tensor(
445
+ [0.05] * 20 + [0.0, 0.0],
446
+ dtype=torch.float32,
447
+ device=protein["aatype"].device
448
+ )
449
+
450
+ categorical_probs = (
451
+ config.uniform_prob * random_aa
452
+ + config.profile_prob * protein["hhblits_profile"]
453
+ + config.same_prob * make_one_hot(protein["msa"], 22)
454
+ )
455
+
456
+ # Put all remaining probability on [MASK] which is a new column
457
+ pad_shapes = list(
458
+ reduce(add, [(0, 0) for _ in range(len(categorical_probs.shape))])
459
+ )
460
+ pad_shapes[1] = 1
461
+ mask_prob = (
462
+ 1.0 - config.profile_prob - config.same_prob - config.uniform_prob
463
+ )
464
+ assert mask_prob >= 0.0
465
+
466
+ categorical_probs = torch.nn.functional.pad(
467
+ categorical_probs, pad_shapes, value=mask_prob
468
+ )
469
+
470
+ sh = protein["msa"].shape
471
+ mask_position = torch.rand(sh) < replace_fraction
472
+
473
+ bert_msa = shaped_categorical(categorical_probs)
474
+ bert_msa = torch.where(mask_position, bert_msa, protein["msa"])
475
+
476
+ # Mix real and masked MSA
477
+ protein["bert_mask"] = mask_position.to(torch.float32)
478
+ protein["true_msa"] = protein["msa"]
479
+ protein["msa"] = bert_msa
480
+
481
+ return protein
482
+
483
+
484
+ @curry1
485
+ def make_fixed_size(
486
+ protein,
487
+ shape_schema,
488
+ msa_cluster_size,
489
+ extra_msa_size,
490
+ num_res=0,
491
+ num_templates=0,
492
+ ):
493
+ """Guess at the MSA and sequence dimension to make fixed size."""
494
+ pad_size_map = {
495
+ NUM_RES: num_res,
496
+ NUM_MSA_SEQ: msa_cluster_size,
497
+ NUM_EXTRA_SEQ: extra_msa_size,
498
+ NUM_TEMPLATES: num_templates,
499
+ }
500
+
501
+ for k, v in protein.items():
502
+ # Don't transfer this to the accelerator.
503
+ if k == "extra_cluster_assignment":
504
+ continue
505
+ shape = list(v.shape)
506
+ schema = shape_schema[k]
507
+ msg = "Rank mismatch between shape and shape schema for"
508
+ assert len(shape) == len(schema), f"{msg} {k}: {shape} vs {schema}"
509
+ pad_size = [
510
+ pad_size_map.get(s2, None) or s1 for (s1, s2) in zip(shape, schema)
511
+ ]
512
+
513
+ padding = [(0, p - v.shape[i]) for i, p in enumerate(pad_size)]
514
+ padding.reverse()
515
+ padding = list(itertools.chain(*padding))
516
+ if padding:
517
+ protein[k] = torch.nn.functional.pad(v, padding)
518
+ protein[k] = torch.reshape(protein[k], pad_size)
519
+
520
+ return protein
521
+
522
+
523
+ @curry1
524
+ def make_msa_feat(protein):
525
+ """Create and concatenate MSA features."""
526
+ # Whether there is a domain break. Always zero for chains, but keeping for
527
+ # compatibility with domain datasets.
528
+ has_break = torch.clip(
529
+ protein["between_segment_residues"].to(torch.float32), 0, 1
530
+ )
531
+ aatype_1hot = make_one_hot(protein["aatype"], 21)
532
+
533
+ target_feat = [
534
+ torch.unsqueeze(has_break, dim=-1),
535
+ aatype_1hot, # Everyone gets the original sequence.
536
+ ]
537
+
538
+ msa_1hot = make_one_hot(protein["msa"], 23)
539
+ has_deletion = torch.clip(protein["deletion_matrix"], 0.0, 1.0)
540
+ deletion_value = torch.atan(protein["deletion_matrix"] / 3.0) * (
541
+ 2.0 / np.pi
542
+ )
543
+
544
+ msa_feat = [
545
+ msa_1hot,
546
+ torch.unsqueeze(has_deletion, dim=-1),
547
+ torch.unsqueeze(deletion_value, dim=-1),
548
+ ]
549
+
550
+ if "cluster_profile" in protein:
551
+ deletion_mean_value = torch.atan(
552
+ protein["cluster_deletion_mean"] / 3.0
553
+ ) * (2.0 / np.pi)
554
+ msa_feat.extend(
555
+ [
556
+ protein["cluster_profile"],
557
+ torch.unsqueeze(deletion_mean_value, dim=-1),
558
+ ]
559
+ )
560
+
561
+ if "extra_deletion_matrix" in protein:
562
+ protein["extra_has_deletion"] = torch.clip(
563
+ protein["extra_deletion_matrix"], 0.0, 1.0
564
+ )
565
+ protein["extra_deletion_value"] = torch.atan(
566
+ protein["extra_deletion_matrix"] / 3.0
567
+ ) * (2.0 / np.pi)
568
+
569
+ protein["msa_feat"] = torch.cat(msa_feat, dim=-1)
570
+ protein["target_feat"] = torch.cat(target_feat, dim=-1)
571
+ return protein
572
+
573
+
574
+ @curry1
575
+ def select_feat(protein, feature_list):
576
+ return {k: v for k, v in protein.items() if k in feature_list}
577
+
578
+
579
+ @curry1
580
+ def crop_templates(protein, max_templates):
581
+ for k, v in protein.items():
582
+ if k.startswith("template_"):
583
+ protein[k] = v[:max_templates]
584
+ return protein
585
+
586
+
587
+ def make_atom14_masks(protein):
588
+ """Construct denser atom positions (14 dimensions instead of 37)."""
589
+ restype_atom14_to_atom37 = []
590
+ restype_atom37_to_atom14 = []
591
+ restype_atom14_mask = []
592
+
593
+ for rt in rc.restypes:
594
+ atom_names = rc.restype_name_to_atom14_names[rc.restype_1to3[rt]]
595
+ restype_atom14_to_atom37.append(
596
+ [(rc.atom_order[name] if name else 0) for name in atom_names]
597
+ )
598
+ atom_name_to_idx14 = {name: i for i, name in enumerate(atom_names)}
599
+ restype_atom37_to_atom14.append(
600
+ [
601
+ (atom_name_to_idx14[name] if name in atom_name_to_idx14 else 0)
602
+ for name in rc.atom_types
603
+ ]
604
+ )
605
+
606
+ restype_atom14_mask.append(
607
+ [(1.0 if name else 0.0) for name in atom_names]
608
+ )
609
+
610
+ # Add dummy mapping for restype 'UNK'
611
+ restype_atom14_to_atom37.append([0] * 14)
612
+ restype_atom37_to_atom14.append([0] * 37)
613
+ restype_atom14_mask.append([0.0] * 14)
614
+
615
+ restype_atom14_to_atom37 = torch.tensor(
616
+ restype_atom14_to_atom37,
617
+ dtype=torch.int32,
618
+ device=protein["aatype"].device,
619
+ )
620
+ restype_atom37_to_atom14 = torch.tensor(
621
+ restype_atom37_to_atom14,
622
+ dtype=torch.int32,
623
+ device=protein["aatype"].device,
624
+ )
625
+ restype_atom14_mask = torch.tensor(
626
+ restype_atom14_mask,
627
+ dtype=torch.float32,
628
+ device=protein["aatype"].device,
629
+ )
630
+ protein_aatype = protein['aatype'].to(torch.long)
631
+
632
+ # create the mapping for (residx, atom14) --> atom37, i.e. an array
633
+ # with shape (num_res, 14) containing the atom37 indices for this protein
634
+ residx_atom14_to_atom37 = restype_atom14_to_atom37[protein_aatype]
635
+ residx_atom14_mask = restype_atom14_mask[protein_aatype]
636
+
637
+ protein["atom14_atom_exists"] = residx_atom14_mask
638
+ protein["residx_atom14_to_atom37"] = residx_atom14_to_atom37.long()
639
+
640
+ # create the gather indices for mapping back
641
+ residx_atom37_to_atom14 = restype_atom37_to_atom14[protein_aatype]
642
+ protein["residx_atom37_to_atom14"] = residx_atom37_to_atom14.long()
643
+
644
+ # create the corresponding mask
645
+ restype_atom37_mask = torch.zeros(
646
+ [21, 37], dtype=torch.float32, device=protein["aatype"].device
647
+ )
648
+ for restype, restype_letter in enumerate(rc.restypes):
649
+ restype_name = rc.restype_1to3[restype_letter]
650
+ atom_names = rc.residue_atoms[restype_name]
651
+ for atom_name in atom_names:
652
+ atom_type = rc.atom_order[atom_name]
653
+ restype_atom37_mask[restype, atom_type] = 1
654
+
655
+ residx_atom37_mask = restype_atom37_mask[protein_aatype]
656
+ protein["atom37_atom_exists"] = residx_atom37_mask
657
+
658
+ return protein
659
+
660
+
661
+ def make_atom14_masks_np(batch):
662
+ batch = tree_map(
663
+ lambda n: torch.tensor(n, device=batch["aatype"].device),
664
+ batch,
665
+ np.ndarray
666
+ )
667
+ out = make_atom14_masks(batch)
668
+ out = tensor_tree_map(lambda t: np.array(t), out)
669
+ return out
670
+
671
+
672
+ def make_atom14_positions(protein):
673
+ """Constructs denser atom positions (14 dimensions instead of 37)."""
674
+ residx_atom14_mask = protein["atom14_atom_exists"]
675
+ residx_atom14_to_atom37 = protein["residx_atom14_to_atom37"]
676
+
677
+ # Create a mask for known ground truth positions.
678
+ residx_atom14_gt_mask = residx_atom14_mask * batched_gather(
679
+ protein["all_atom_mask"],
680
+ residx_atom14_to_atom37,
681
+ dim=-1,
682
+ no_batch_dims=len(protein["all_atom_mask"].shape[:-1]),
683
+ )
684
+
685
+ # Gather the ground truth positions.
686
+ residx_atom14_gt_positions = residx_atom14_gt_mask[..., None] * (
687
+ batched_gather(
688
+ protein["all_atom_positions"],
689
+ residx_atom14_to_atom37,
690
+ dim=-2,
691
+ no_batch_dims=len(protein["all_atom_positions"].shape[:-2]),
692
+ )
693
+ )
694
+
695
+ protein["atom14_atom_exists"] = residx_atom14_mask
696
+ protein["atom14_gt_exists"] = residx_atom14_gt_mask
697
+ protein["atom14_gt_positions"] = residx_atom14_gt_positions
698
+
699
+ # As the atom naming is ambiguous for 7 of the 20 amino acids, provide
700
+ # alternative ground truth coordinates where the naming is swapped
701
+ restype_3 = [rc.restype_1to3[res] for res in rc.restypes]
702
+ restype_3 += ["UNK"]
703
+
704
+ # Matrices for renaming ambiguous atoms.
705
+ all_matrices = {
706
+ res: torch.eye(
707
+ 14,
708
+ dtype=protein["all_atom_mask"].dtype,
709
+ device=protein["all_atom_mask"].device,
710
+ )
711
+ for res in restype_3
712
+ }
713
+ for resname, swap in rc.residue_atom_renaming_swaps.items():
714
+ correspondences = torch.arange(
715
+ 14, device=protein["all_atom_mask"].device
716
+ )
717
+ for source_atom_swap, target_atom_swap in swap.items():
718
+ source_index = rc.restype_name_to_atom14_names[resname].index(
719
+ source_atom_swap
720
+ )
721
+ target_index = rc.restype_name_to_atom14_names[resname].index(
722
+ target_atom_swap
723
+ )
724
+ correspondences[source_index] = target_index
725
+ correspondences[target_index] = source_index
726
+ renaming_matrix = protein["all_atom_mask"].new_zeros((14, 14))
727
+ for index, correspondence in enumerate(correspondences):
728
+ renaming_matrix[index, correspondence] = 1.0
729
+ all_matrices[resname] = renaming_matrix
730
+
731
+ renaming_matrices = torch.stack(
732
+ [all_matrices[restype] for restype in restype_3]
733
+ )
734
+
735
+ # Pick the transformation matrices for the given residue sequence
736
+ # shape (num_res, 14, 14).
737
+ renaming_transform = renaming_matrices[protein["aatype"]]
738
+
739
+ # Apply it to the ground truth positions. shape (num_res, 14, 3).
740
+ alternative_gt_positions = torch.einsum(
741
+ "...rac,...rab->...rbc", residx_atom14_gt_positions, renaming_transform
742
+ )
743
+ protein["atom14_alt_gt_positions"] = alternative_gt_positions
744
+
745
+ # Create the mask for the alternative ground truth (differs from the
746
+ # ground truth mask, if only one of the atoms in an ambiguous pair has a
747
+ # ground truth position).
748
+ alternative_gt_mask = torch.einsum(
749
+ "...ra,...rab->...rb", residx_atom14_gt_mask, renaming_transform
750
+ )
751
+ protein["atom14_alt_gt_exists"] = alternative_gt_mask
752
+
753
+ # Create an ambiguous atoms mask. shape: (21, 14).
754
+ restype_atom14_is_ambiguous = protein["all_atom_mask"].new_zeros((21, 14))
755
+ for resname, swap in rc.residue_atom_renaming_swaps.items():
756
+ for atom_name1, atom_name2 in swap.items():
757
+ restype = rc.restype_order[rc.restype_3to1[resname]]
758
+ atom_idx1 = rc.restype_name_to_atom14_names[resname].index(
759
+ atom_name1
760
+ )
761
+ atom_idx2 = rc.restype_name_to_atom14_names[resname].index(
762
+ atom_name2
763
+ )
764
+ restype_atom14_is_ambiguous[restype, atom_idx1] = 1
765
+ restype_atom14_is_ambiguous[restype, atom_idx2] = 1
766
+
767
+ # From this create an ambiguous_mask for the given sequence.
768
+ protein["atom14_atom_is_ambiguous"] = restype_atom14_is_ambiguous[
769
+ protein["aatype"]
770
+ ]
771
+
772
+ return protein
773
+
774
+
775
+ def atom37_to_frames(protein, eps=1e-8):
776
+ aatype = protein["aatype"]
777
+ all_atom_positions = protein["all_atom_positions"]
778
+ all_atom_mask = protein["all_atom_mask"]
779
+
780
+ batch_dims = len(aatype.shape[:-1])
781
+
782
+ restype_rigidgroup_base_atom_names = np.full([21, 8, 3], "", dtype=object)
783
+ restype_rigidgroup_base_atom_names[:, 0, :] = ["C", "CA", "N"]
784
+ restype_rigidgroup_base_atom_names[:, 3, :] = ["CA", "C", "O"]
785
+
786
+ for restype, restype_letter in enumerate(rc.restypes):
787
+ resname = rc.restype_1to3[restype_letter]
788
+ for chi_idx in range(4):
789
+ if rc.chi_angles_mask[restype][chi_idx]:
790
+ names = rc.chi_angles_atoms[resname][chi_idx]
791
+ restype_rigidgroup_base_atom_names[
792
+ restype, chi_idx + 4, :
793
+ ] = names[1:]
794
+
795
+ restype_rigidgroup_mask = all_atom_mask.new_zeros(
796
+ (*aatype.shape[:-1], 21, 8),
797
+ )
798
+ restype_rigidgroup_mask[..., 0] = 1
799
+ restype_rigidgroup_mask[..., 3] = 1
800
+ restype_rigidgroup_mask[..., :20, 4:] = all_atom_mask.new_tensor(
801
+ rc.chi_angles_mask
802
+ )
803
+
804
+ lookuptable = rc.atom_order.copy()
805
+ lookuptable[""] = 0
806
+ lookup = np.vectorize(lambda x: lookuptable[x])
807
+ restype_rigidgroup_base_atom37_idx = lookup(
808
+ restype_rigidgroup_base_atom_names,
809
+ )
810
+ restype_rigidgroup_base_atom37_idx = aatype.new_tensor(
811
+ restype_rigidgroup_base_atom37_idx,
812
+ )
813
+ restype_rigidgroup_base_atom37_idx = (
814
+ restype_rigidgroup_base_atom37_idx.view(
815
+ *((1,) * batch_dims), *restype_rigidgroup_base_atom37_idx.shape
816
+ )
817
+ )
818
+
819
+ residx_rigidgroup_base_atom37_idx = batched_gather(
820
+ restype_rigidgroup_base_atom37_idx,
821
+ aatype,
822
+ dim=-3,
823
+ no_batch_dims=batch_dims,
824
+ )
825
+
826
+ base_atom_pos = batched_gather(
827
+ all_atom_positions,
828
+ residx_rigidgroup_base_atom37_idx,
829
+ dim=-2,
830
+ no_batch_dims=len(all_atom_positions.shape[:-2]),
831
+ )
832
+
833
+ gt_frames = Rigid.from_3_points(
834
+ p_neg_x_axis=base_atom_pos[..., 0, :],
835
+ origin=base_atom_pos[..., 1, :],
836
+ p_xy_plane=base_atom_pos[..., 2, :],
837
+ eps=eps,
838
+ )
839
+
840
+ group_exists = batched_gather(
841
+ restype_rigidgroup_mask,
842
+ aatype,
843
+ dim=-2,
844
+ no_batch_dims=batch_dims,
845
+ )
846
+
847
+ gt_atoms_exist = batched_gather(
848
+ all_atom_mask,
849
+ residx_rigidgroup_base_atom37_idx,
850
+ dim=-1,
851
+ no_batch_dims=len(all_atom_mask.shape[:-1]),
852
+ )
853
+ gt_exists = torch.min(gt_atoms_exist, dim=-1)[0] * group_exists
854
+
855
+ rots = torch.eye(3, dtype=all_atom_mask.dtype, device=aatype.device)
856
+ rots = torch.tile(rots, (*((1,) * batch_dims), 8, 1, 1))
857
+ rots[..., 0, 0, 0] = -1
858
+ rots[..., 0, 2, 2] = -1
859
+ rots = Rotation(rot_mats=rots)
860
+
861
+ gt_frames = gt_frames.compose(Rigid(rots, None))
862
+
863
+ restype_rigidgroup_is_ambiguous = all_atom_mask.new_zeros(
864
+ *((1,) * batch_dims), 21, 8
865
+ )
866
+ restype_rigidgroup_rots = torch.eye(
867
+ 3, dtype=all_atom_mask.dtype, device=aatype.device
868
+ )
869
+ restype_rigidgroup_rots = torch.tile(
870
+ restype_rigidgroup_rots,
871
+ (*((1,) * batch_dims), 21, 8, 1, 1),
872
+ )
873
+
874
+ for resname, _ in rc.residue_atom_renaming_swaps.items():
875
+ restype = rc.restype_order[rc.restype_3to1[resname]]
876
+ chi_idx = int(sum(rc.chi_angles_mask[restype]) - 1)
877
+ restype_rigidgroup_is_ambiguous[..., restype, chi_idx + 4] = 1
878
+ restype_rigidgroup_rots[..., restype, chi_idx + 4, 1, 1] = -1
879
+ restype_rigidgroup_rots[..., restype, chi_idx + 4, 2, 2] = -1
880
+
881
+ residx_rigidgroup_is_ambiguous = batched_gather(
882
+ restype_rigidgroup_is_ambiguous,
883
+ aatype,
884
+ dim=-2,
885
+ no_batch_dims=batch_dims,
886
+ )
887
+
888
+ residx_rigidgroup_ambiguity_rot = batched_gather(
889
+ restype_rigidgroup_rots,
890
+ aatype,
891
+ dim=-4,
892
+ no_batch_dims=batch_dims,
893
+ )
894
+
895
+ residx_rigidgroup_ambiguity_rot = Rotation(
896
+ rot_mats=residx_rigidgroup_ambiguity_rot
897
+ )
898
+ alt_gt_frames = gt_frames.compose(
899
+ Rigid(residx_rigidgroup_ambiguity_rot, None)
900
+ )
901
+
902
+ gt_frames_tensor = gt_frames.to_tensor_4x4()
903
+ alt_gt_frames_tensor = alt_gt_frames.to_tensor_4x4()
904
+
905
+ protein["rigidgroups_gt_frames"] = gt_frames_tensor
906
+ protein["rigidgroups_gt_exists"] = gt_exists
907
+ protein["rigidgroups_group_exists"] = group_exists
908
+ protein["rigidgroups_group_is_ambiguous"] = residx_rigidgroup_is_ambiguous
909
+ protein["rigidgroups_alt_gt_frames"] = alt_gt_frames_tensor
910
+
911
+ return protein
912
+
913
+
914
+ def get_chi_atom_indices():
915
+ """Returns atom indices needed to compute chi angles for all residue types.
916
+
917
+ Returns:
918
+ A tensor of shape [residue_types=21, chis=4, atoms=4]. The residue types are
919
+ in the order specified in rc.restypes + unknown residue type
920
+ at the end. For chi angles which are not defined on the residue, the
921
+ positions indices are by default set to 0.
922
+ """
923
+ chi_atom_indices = []
924
+ for residue_name in rc.restypes:
925
+ residue_name = rc.restype_1to3[residue_name]
926
+ residue_chi_angles = rc.chi_angles_atoms[residue_name]
927
+ atom_indices = []
928
+ for chi_angle in residue_chi_angles:
929
+ atom_indices.append([rc.atom_order[atom] for atom in chi_angle])
930
+ for _ in range(4 - len(atom_indices)):
931
+ atom_indices.append(
932
+ [0, 0, 0, 0]
933
+ ) # For chi angles not defined on the AA.
934
+ chi_atom_indices.append(atom_indices)
935
+
936
+ chi_atom_indices.append([[0, 0, 0, 0]] * 4) # For UNKNOWN residue.
937
+
938
+ return chi_atom_indices
939
+
940
+
941
+ @curry1
942
+ def atom37_to_torsion_angles(
943
+ protein,
944
+ prefix="",
945
+ ):
946
+ """
947
+ Convert coordinates to torsion angles.
948
+
949
+ This function is extremely sensitive to floating point imprecisions
950
+ and should be run with double precision whenever possible.
951
+
952
+ Args:
953
+ Dict containing:
954
+ * (prefix)aatype:
955
+ [*, N_res] residue indices
956
+ * (prefix)all_atom_positions:
957
+ [*, N_res, 37, 3] atom positions (in atom37
958
+ format)
959
+ * (prefix)all_atom_mask:
960
+ [*, N_res, 37] atom position mask
961
+ Returns:
962
+ The same dictionary updated with the following features:
963
+
964
+ "(prefix)torsion_angles_sin_cos" ([*, N_res, 7, 2])
965
+ Torsion angles
966
+ "(prefix)alt_torsion_angles_sin_cos" ([*, N_res, 7, 2])
967
+ Alternate torsion angles (accounting for 180-degree symmetry)
968
+ "(prefix)torsion_angles_mask" ([*, N_res, 7])
969
+ Torsion angles mask
970
+ """
971
+ aatype = protein[prefix + "aatype"]
972
+ all_atom_positions = protein[prefix + "all_atom_positions"]
973
+ all_atom_mask = protein[prefix + "all_atom_mask"]
974
+
975
+ aatype = torch.clamp(aatype, max=20)
976
+
977
+ pad = all_atom_positions.new_zeros(
978
+ [*all_atom_positions.shape[:-3], 1, 37, 3]
979
+ )
980
+ prev_all_atom_positions = torch.cat(
981
+ [pad, all_atom_positions[..., :-1, :, :]], dim=-3
982
+ )
983
+
984
+ pad = all_atom_mask.new_zeros([*all_atom_mask.shape[:-2], 1, 37])
985
+ prev_all_atom_mask = torch.cat([pad, all_atom_mask[..., :-1, :]], dim=-2)
986
+
987
+ pre_omega_atom_pos = torch.cat(
988
+ [prev_all_atom_positions[..., 1:3, :], all_atom_positions[..., :2, :]],
989
+ dim=-2,
990
+ )
991
+ phi_atom_pos = torch.cat(
992
+ [prev_all_atom_positions[..., 2:3, :], all_atom_positions[..., :3, :]],
993
+ dim=-2,
994
+ )
995
+ psi_atom_pos = torch.cat(
996
+ [all_atom_positions[..., :3, :], all_atom_positions[..., 4:5, :]],
997
+ dim=-2,
998
+ )
999
+
1000
+ pre_omega_mask = torch.prod(
1001
+ prev_all_atom_mask[..., 1:3], dim=-1
1002
+ ) * torch.prod(all_atom_mask[..., :2], dim=-1)
1003
+ phi_mask = prev_all_atom_mask[..., 2] * torch.prod(
1004
+ all_atom_mask[..., :3], dim=-1, dtype=all_atom_mask.dtype
1005
+ )
1006
+ psi_mask = (
1007
+ torch.prod(all_atom_mask[..., :3], dim=-1, dtype=all_atom_mask.dtype)
1008
+ * all_atom_mask[..., 4]
1009
+ )
1010
+
1011
+ chi_atom_indices = torch.as_tensor(
1012
+ get_chi_atom_indices(), device=aatype.device
1013
+ )
1014
+
1015
+ atom_indices = chi_atom_indices[..., aatype, :, :]
1016
+ chis_atom_pos = batched_gather(
1017
+ all_atom_positions, atom_indices, -2, len(atom_indices.shape[:-2])
1018
+ )
1019
+
1020
+ chi_angles_mask = list(rc.chi_angles_mask)
1021
+ chi_angles_mask.append([0.0, 0.0, 0.0, 0.0])
1022
+ chi_angles_mask = all_atom_mask.new_tensor(chi_angles_mask)
1023
+
1024
+ chis_mask = chi_angles_mask[aatype, :]
1025
+
1026
+ chi_angle_atoms_mask = batched_gather(
1027
+ all_atom_mask,
1028
+ atom_indices,
1029
+ dim=-1,
1030
+ no_batch_dims=len(atom_indices.shape[:-2]),
1031
+ )
1032
+ chi_angle_atoms_mask = torch.prod(
1033
+ chi_angle_atoms_mask, dim=-1, dtype=chi_angle_atoms_mask.dtype
1034
+ )
1035
+ chis_mask = chis_mask * chi_angle_atoms_mask
1036
+
1037
+ torsions_atom_pos = torch.cat(
1038
+ [
1039
+ pre_omega_atom_pos[..., None, :, :],
1040
+ phi_atom_pos[..., None, :, :],
1041
+ psi_atom_pos[..., None, :, :],
1042
+ chis_atom_pos,
1043
+ ],
1044
+ dim=-3,
1045
+ )
1046
+
1047
+ torsion_angles_mask = torch.cat(
1048
+ [
1049
+ pre_omega_mask[..., None],
1050
+ phi_mask[..., None],
1051
+ psi_mask[..., None],
1052
+ chis_mask,
1053
+ ],
1054
+ dim=-1,
1055
+ )
1056
+
1057
+ torsion_frames = Rigid.from_3_points(
1058
+ torsions_atom_pos[..., 1, :],
1059
+ torsions_atom_pos[..., 2, :],
1060
+ torsions_atom_pos[..., 0, :],
1061
+ eps=1e-8,
1062
+ )
1063
+
1064
+ fourth_atom_rel_pos = torsion_frames.invert().apply(
1065
+ torsions_atom_pos[..., 3, :]
1066
+ )
1067
+
1068
+ torsion_angles_sin_cos = torch.stack(
1069
+ [fourth_atom_rel_pos[..., 2], fourth_atom_rel_pos[..., 1]], dim=-1
1070
+ )
1071
+
1072
+ denom = torch.sqrt(
1073
+ torch.sum(
1074
+ torch.square(torsion_angles_sin_cos),
1075
+ dim=-1,
1076
+ dtype=torsion_angles_sin_cos.dtype,
1077
+ keepdims=True,
1078
+ )
1079
+ + 1e-8
1080
+ )
1081
+ torsion_angles_sin_cos = torsion_angles_sin_cos / denom
1082
+
1083
+ torsion_angles_sin_cos = torsion_angles_sin_cos * all_atom_mask.new_tensor(
1084
+ [1.0, 1.0, -1.0, 1.0, 1.0, 1.0, 1.0],
1085
+ )[((None,) * len(torsion_angles_sin_cos.shape[:-2])) + (slice(None), None)]
1086
+
1087
+ chi_is_ambiguous = torsion_angles_sin_cos.new_tensor(
1088
+ rc.chi_pi_periodic,
1089
+ )[aatype, ...]
1090
+
1091
+ mirror_torsion_angles = torch.cat(
1092
+ [
1093
+ all_atom_mask.new_ones(*aatype.shape, 3),
1094
+ 1.0 - 2.0 * chi_is_ambiguous,
1095
+ ],
1096
+ dim=-1,
1097
+ )
1098
+
1099
+ alt_torsion_angles_sin_cos = (
1100
+ torsion_angles_sin_cos * mirror_torsion_angles[..., None]
1101
+ )
1102
+
1103
+ protein[prefix + "torsion_angles_sin_cos"] = torsion_angles_sin_cos
1104
+ protein[prefix + "alt_torsion_angles_sin_cos"] = alt_torsion_angles_sin_cos
1105
+ protein[prefix + "torsion_angles_mask"] = torsion_angles_mask
1106
+
1107
+ return protein
1108
+
1109
+
1110
+ def get_backbone_frames(protein):
1111
+ # DISCREPANCY: AlphaFold uses tensor_7s here. I don't know why.
1112
+ protein["backbone_rigid_tensor"] = protein["rigidgroups_gt_frames"][
1113
+ ..., 0, :, :
1114
+ ]
1115
+ protein["backbone_rigid_mask"] = protein["rigidgroups_gt_exists"][..., 0]
1116
+
1117
+ return protein
1118
+
1119
+
1120
+ def get_chi_angles(protein):
1121
+ dtype = protein["all_atom_mask"].dtype
1122
+ protein["chi_angles_sin_cos"] = (
1123
+ protein["torsion_angles_sin_cos"][..., 3:, :]
1124
+ ).to(dtype)
1125
+ protein["chi_mask"] = protein["torsion_angles_mask"][..., 3:].to(dtype)
1126
+
1127
+ return protein
1128
+
1129
+
1130
+ @curry1
1131
+ def random_crop_to_size(
1132
+ protein,
1133
+ crop_size,
1134
+ max_templates,
1135
+ shape_schema,
1136
+ subsample_templates=False,
1137
+ seed=None,
1138
+ ):
1139
+ """Crop randomly to `crop_size`, or keep as is if shorter than that."""
1140
+ # We want each ensemble to be cropped the same way
1141
+ g = torch.Generator(device=protein["seq_length"].device)
1142
+ if seed is not None:
1143
+ g.manual_seed(seed)
1144
+
1145
+ seq_length = protein["seq_length"]
1146
+
1147
+ if "template_mask" in protein:
1148
+ num_templates = protein["template_mask"].shape[-1]
1149
+ else:
1150
+ num_templates = 0
1151
+
1152
+ # No need to subsample templates if there aren't any
1153
+ subsample_templates = subsample_templates and num_templates
1154
+
1155
+ num_res_crop_size = min(int(seq_length), crop_size)
1156
+
1157
+ def _randint(lower, upper):
1158
+ return int(torch.randint(
1159
+ lower,
1160
+ upper + 1,
1161
+ (1,),
1162
+ device=protein["seq_length"].device,
1163
+ generator=g,
1164
+ )[0])
1165
+
1166
+ if subsample_templates:
1167
+ templates_crop_start = _randint(0, num_templates)
1168
+ templates_select_indices = torch.randperm(
1169
+ num_templates, device=protein["seq_length"].device, generator=g
1170
+ )
1171
+ else:
1172
+ templates_crop_start = 0
1173
+
1174
+ num_templates_crop_size = min(
1175
+ num_templates - templates_crop_start, max_templates
1176
+ )
1177
+
1178
+ n = seq_length - num_res_crop_size
1179
+ if "use_clamped_fape" in protein and protein["use_clamped_fape"] == 1.:
1180
+ right_anchor = n
1181
+ else:
1182
+ x = _randint(0, n)
1183
+ right_anchor = n - x
1184
+
1185
+ num_res_crop_start = _randint(0, right_anchor)
1186
+
1187
+ for k, v in protein.items():
1188
+ if k not in shape_schema or (
1189
+ "template" not in k and NUM_RES not in shape_schema[k]
1190
+ ):
1191
+ continue
1192
+
1193
+ # randomly permute the templates before cropping them.
1194
+ if k.startswith("template") and subsample_templates:
1195
+ v = v[templates_select_indices]
1196
+
1197
+ slices = []
1198
+ for i, (dim_size, dim) in enumerate(zip(shape_schema[k], v.shape)):
1199
+ is_num_res = dim_size == NUM_RES
1200
+ if i == 0 and k.startswith("template"):
1201
+ crop_size = num_templates_crop_size
1202
+ crop_start = templates_crop_start
1203
+ else:
1204
+ crop_start = num_res_crop_start if is_num_res else 0
1205
+ crop_size = num_res_crop_size if is_num_res else dim
1206
+ slices.append(slice(crop_start, crop_start + crop_size))
1207
+ protein[k] = v[slices]
1208
+
1209
+ protein["seq_length"] = protein["seq_length"].new_tensor(num_res_crop_size)
1210
+
1211
+ return protein
openfold/protein.py ADDED
@@ -0,0 +1,429 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2021 AlQuraishi Laboratory
2
+ # Copyright 2021 DeepMind Technologies Limited
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+
16
+ """Protein data type."""
17
+ import dataclasses
18
+ import io
19
+ from typing import Any, Sequence, Mapping, Optional
20
+ import re
21
+ import string
22
+
23
+ from . import residue_constants
24
+ from Bio.PDB import PDBParser
25
+ import numpy as np
26
+
27
+
28
+ FeatureDict = Mapping[str, np.ndarray]
29
+ ModelOutput = Mapping[str, Any] # Is a nested dict.
30
+ PICO_TO_ANGSTROM = 0.01
31
+
32
+ @dataclasses.dataclass(frozen=True)
33
+ class Protein:
34
+ """Protein structure representation."""
35
+
36
+ # Cartesian coordinates of atoms in angstroms. The atom types correspond to
37
+ # residue_constants.atom_types, i.e. the first three are N, CA, CB.
38
+ atom_positions: np.ndarray # [num_res, num_atom_type, 3]
39
+
40
+ # Amino-acid type for each residue represented as an integer between 0 and
41
+ # 20, where 20 is 'X'.
42
+ aatype: np.ndarray # [num_res]
43
+
44
+ # Binary float mask to indicate presence of a particular atom. 1.0 if an atom
45
+ # is present and 0.0 if not. This should be used for loss masking.
46
+ atom_mask: np.ndarray # [num_res, num_atom_type]
47
+
48
+ # Residue index as used in PDB. It is not necessarily continuous or 0-indexed.
49
+ residue_index: np.ndarray # [num_res]
50
+
51
+ # B-factors, or temperature factors, of each residue (in sq. angstroms units),
52
+ # representing the displacement of the residue from its ground truth mean
53
+ # value.
54
+ b_factors: np.ndarray # [num_res, num_atom_type]
55
+
56
+ # Chain indices for multi-chain predictions
57
+ chain_index: Optional[np.ndarray] = None
58
+
59
+ # Optional remark about the protein. Included as a comment in output PDB
60
+ # files
61
+ remark: Optional[str] = None
62
+
63
+ # Templates used to generate this protein (prediction-only)
64
+ parents: Optional[Sequence[str]] = None
65
+
66
+ # Chain corresponding to each parent
67
+ parents_chain_index: Optional[Sequence[int]] = None
68
+
69
+
70
+ def from_pdb_string(pdb_str: str, chain_id: Optional[str] = None) -> Protein:
71
+ """Takes a PDB string and constructs a Protein object.
72
+ WARNING: All non-standard residue types will be converted into UNK. All
73
+ non-standard atoms will be ignored.
74
+ Args:
75
+ pdb_str: The contents of the pdb file
76
+ chain_id: If None, then the pdb file must contain a single chain (which
77
+ will be parsed). If chain_id is specified (e.g. A), then only that chain
78
+ is parsed.
79
+ Returns:
80
+ A new `Protein` parsed from the pdb contents.
81
+ """
82
+ pdb_fh = io.StringIO(pdb_str)
83
+ parser = PDBParser(QUIET=True)
84
+ structure = parser.get_structure("none", pdb_fh)
85
+ models = list(structure.get_models())
86
+ if len(models) != 1:
87
+ raise ValueError(
88
+ f"Only single model PDBs are supported. Found {len(models)} models."
89
+ )
90
+ model = models[0]
91
+
92
+ atom_positions = []
93
+ aatype = []
94
+ atom_mask = []
95
+ residue_index = []
96
+ chain_ids = []
97
+ b_factors = []
98
+
99
+ for chain in model:
100
+ if(chain_id is not None and chain.id != chain_id):
101
+ continue
102
+ for res in chain:
103
+ if res.id[2] != " ":
104
+ raise ValueError(
105
+ f"PDB contains an insertion code at chain {chain.id} and residue "
106
+ f"index {res.id[1]}. These are not supported."
107
+ )
108
+ res_shortname = residue_constants.restype_3to1.get(res.resname, "X")
109
+ restype_idx = residue_constants.restype_order.get(
110
+ res_shortname, residue_constants.restype_num
111
+ )
112
+ pos = np.zeros((residue_constants.atom_type_num, 3))
113
+ mask = np.zeros((residue_constants.atom_type_num,))
114
+ res_b_factors = np.zeros((residue_constants.atom_type_num,))
115
+ for atom in res:
116
+ if atom.name not in residue_constants.atom_types:
117
+ continue
118
+ pos[residue_constants.atom_order[atom.name]] = atom.coord
119
+ mask[residue_constants.atom_order[atom.name]] = 1.0
120
+ res_b_factors[
121
+ residue_constants.atom_order[atom.name]
122
+ ] = atom.bfactor
123
+ if np.sum(mask) < 0.5:
124
+ # If no known atom positions are reported for the residue then skip it.
125
+ continue
126
+ aatype.append(restype_idx)
127
+ atom_positions.append(pos)
128
+ atom_mask.append(mask)
129
+ residue_index.append(res.id[1])
130
+ chain_ids.append(chain.id)
131
+ b_factors.append(res_b_factors)
132
+
133
+ parents = None
134
+ parents_chain_index = None
135
+ if("PARENT" in pdb_str):
136
+ parents = []
137
+ parents_chain_index = []
138
+ chain_id = 0
139
+ for l in pdb_str.split("\n"):
140
+ if("PARENT" in l):
141
+ if(not "N/A" in l):
142
+ parent_names = l.split()[1:]
143
+ parents.extend(parent_names)
144
+ parents_chain_index.extend([
145
+ chain_id for _ in parent_names
146
+ ])
147
+ chain_id += 1
148
+
149
+ unique_chain_ids = np.unique(chain_ids)
150
+ chain_id_mapping = {cid: n for n, cid in enumerate(string.ascii_uppercase)}
151
+ chain_index = np.array([chain_id_mapping[cid] for cid in chain_ids])
152
+
153
+ return Protein(
154
+ atom_positions=np.array(atom_positions),
155
+ atom_mask=np.array(atom_mask),
156
+ aatype=np.array(aatype),
157
+ residue_index=np.array(residue_index),
158
+ chain_index=chain_index,
159
+ b_factors=np.array(b_factors),
160
+ parents=parents,
161
+ parents_chain_index=parents_chain_index,
162
+ )
163
+
164
+
165
+ def from_proteinnet_string(proteinnet_str: str) -> Protein:
166
+ tag_re = r'(\[[A-Z]+\]\n)'
167
+ tags = [
168
+ tag.strip() for tag in re.split(tag_re, proteinnet_str) if len(tag) > 0
169
+ ]
170
+ groups = zip(tags[0::2], [l.split('\n') for l in tags[1::2]])
171
+
172
+ atoms = ['N', 'CA', 'C']
173
+ aatype = None
174
+ atom_positions = None
175
+ atom_mask = None
176
+ for g in groups:
177
+ if("[PRIMARY]" == g[0]):
178
+ seq = g[1][0].strip()
179
+ for i in range(len(seq)):
180
+ if(seq[i] not in residue_constants.restypes):
181
+ seq[i] = 'X'
182
+ aatype = np.array([
183
+ residue_constants.restype_order.get(
184
+ res_symbol, residue_constants.restype_num
185
+ ) for res_symbol in seq
186
+ ])
187
+ elif("[TERTIARY]" == g[0]):
188
+ tertiary = []
189
+ for axis in range(3):
190
+ tertiary.append(list(map(float, g[1][axis].split())))
191
+ tertiary_np = np.array(tertiary)
192
+ atom_positions = np.zeros(
193
+ (len(tertiary[0])//3, residue_constants.atom_type_num, 3)
194
+ ).astype(np.float32)
195
+ for i, atom in enumerate(atoms):
196
+ atom_positions[:, residue_constants.atom_order[atom], :] = (
197
+ np.transpose(tertiary_np[:, i::3])
198
+ )
199
+ atom_positions *= PICO_TO_ANGSTROM
200
+ elif("[MASK]" == g[0]):
201
+ mask = np.array(list(map({'-': 0, '+': 1}.get, g[1][0].strip())))
202
+ atom_mask = np.zeros(
203
+ (len(mask), residue_constants.atom_type_num,)
204
+ ).astype(np.float32)
205
+ for i, atom in enumerate(atoms):
206
+ atom_mask[:, residue_constants.atom_order[atom]] = 1
207
+ atom_mask *= mask[..., None]
208
+
209
+ return Protein(
210
+ atom_positions=atom_positions,
211
+ atom_mask=atom_mask,
212
+ aatype=aatype,
213
+ residue_index=np.arange(len(aatype)),
214
+ b_factors=None,
215
+ )
216
+
217
+
218
+ def get_pdb_headers(prot: Protein, chain_id: int = 0) -> Sequence[str]:
219
+ pdb_headers = []
220
+
221
+ remark = prot.remark
222
+ if(remark is not None):
223
+ pdb_headers.append(f"REMARK {remark}")
224
+
225
+ parents = prot.parents
226
+ parents_chain_index = prot.parents_chain_index
227
+ if(parents_chain_index is not None):
228
+ parents = [
229
+ p for i, p in zip(parents_chain_index, parents) if i == chain_id
230
+ ]
231
+
232
+ if(parents is None or len(parents) == 0):
233
+ parents = ["N/A"]
234
+
235
+ pdb_headers.append(f"PARENT {' '.join(parents)}")
236
+
237
+ return pdb_headers
238
+
239
+
240
+ def add_pdb_headers(prot: Protein, pdb_str: str) -> str:
241
+ """ Add pdb headers to an existing PDB string. Useful during multi-chain
242
+ recycling
243
+ """
244
+ out_pdb_lines = []
245
+ lines = pdb_str.split('\n')
246
+
247
+ remark = prot.remark
248
+ if(remark is not None):
249
+ out_pdb_lines.append(f"REMARK {remark}")
250
+
251
+ parents_per_chain = None
252
+ if(prot.parents is not None and len(prot.parents) > 0):
253
+ parents_per_chain = []
254
+ if(prot.parents_chain_index is not None):
255
+ cur_chain = prot.parents_chain_index[0]
256
+ parent_dict = {}
257
+ for p, i in zip(prot.parents, prot.parents_chain_index):
258
+ parent_dict.setdefault(str(i), [])
259
+ parent_dict[str(i)].append(p)
260
+
261
+ max_idx = max([int(chain_idx) for chain_idx in parent_dict])
262
+ for i in range(max_idx + 1):
263
+ chain_parents = parent_dict.get(str(i), ["N/A"])
264
+ parents_per_chain.append(chain_parents)
265
+ else:
266
+ parents_per_chain.append(prot.parents)
267
+ else:
268
+ parents_per_chain = [["N/A"]]
269
+
270
+ make_parent_line = lambda p: f"PARENT {' '.join(p)}"
271
+
272
+ out_pdb_lines.append(make_parent_line(parents_per_chain[0]))
273
+
274
+ chain_counter = 0
275
+ for i, l in enumerate(lines):
276
+ if("PARENT" not in l and "REMARK" not in l):
277
+ out_pdb_lines.append(l)
278
+ if("TER" in l and not "END" in lines[i + 1]):
279
+ chain_counter += 1
280
+ if(not chain_counter >= len(parents_per_chain)):
281
+ chain_parents = parents_per_chain[chain_counter]
282
+ else:
283
+ chain_parents = ["N/A"]
284
+
285
+ out_pdb_lines.append(make_parent_line(chain_parents))
286
+
287
+ return '\n'.join(out_pdb_lines)
288
+
289
+
290
+ def to_pdb(prot: Protein) -> str:
291
+ """Converts a `Protein` instance to a PDB string.
292
+ Args:
293
+ prot: The protein to convert to PDB.
294
+ Returns:
295
+ PDB string.
296
+ """
297
+ restypes = residue_constants.restypes + ["X"]
298
+ res_1to3 = lambda r: residue_constants.restype_1to3.get(restypes[r], "UNK")
299
+ atom_types = residue_constants.atom_types
300
+
301
+ pdb_lines = []
302
+
303
+ atom_mask = prot.atom_mask
304
+ aatype = prot.aatype
305
+ atom_positions = prot.atom_positions
306
+ residue_index = prot.residue_index.astype(np.int32)
307
+ b_factors = prot.b_factors
308
+ chain_index = prot.chain_index
309
+
310
+ if np.any(aatype > residue_constants.restype_num):
311
+ raise ValueError("Invalid aatypes.")
312
+
313
+ headers = get_pdb_headers(prot)
314
+ if(len(headers) > 0):
315
+ pdb_lines.extend(headers)
316
+
317
+ n = aatype.shape[0]
318
+ atom_index = 1
319
+ prev_chain_index = 0
320
+ chain_tags = string.ascii_uppercase
321
+ # Add all atom sites.
322
+ for i in range(n):
323
+ res_name_3 = res_1to3(aatype[i])
324
+ for atom_name, pos, mask, b_factor in zip(
325
+ atom_types, atom_positions[i], atom_mask[i], b_factors[i]
326
+ ):
327
+ if mask < 0.5:
328
+ continue
329
+
330
+ record_type = "ATOM"
331
+ name = atom_name if len(atom_name) == 4 else f" {atom_name}"
332
+ alt_loc = ""
333
+ insertion_code = ""
334
+ occupancy = 1.00
335
+ element = atom_name[
336
+ 0
337
+ ] # Protein supports only C, N, O, S, this works.
338
+ charge = ""
339
+
340
+ chain_tag = "A"
341
+ if(chain_index is not None):
342
+ chain_tag = chain_tags[chain_index[i]]
343
+
344
+ # PDB is a columnar format, every space matters here!
345
+ atom_line = (
346
+ f"{record_type:<6}{atom_index:>5} {name:<4}{alt_loc:>1}"
347
+ f"{res_name_3:>3} {chain_tag:>1}"
348
+ f"{residue_index[i]:>4}{insertion_code:>1} "
349
+ f"{pos[0]:>8.3f}{pos[1]:>8.3f}{pos[2]:>8.3f}"
350
+ f"{occupancy:>6.2f}{b_factor:>6.2f} "
351
+ f"{element:>2}{charge:>2}"
352
+ )
353
+ pdb_lines.append(atom_line)
354
+ atom_index += 1
355
+
356
+ should_terminate = (i == n - 1)
357
+ if(chain_index is not None):
358
+ if(i != n - 1 and chain_index[i + 1] != prev_chain_index):
359
+ should_terminate = True
360
+ prev_chain_index = chain_index[i + 1]
361
+
362
+ if(should_terminate):
363
+ # Close the chain.
364
+ chain_end = "TER"
365
+ chain_termination_line = (
366
+ f"{chain_end:<6}{atom_index:>5} "
367
+ f"{res_1to3(aatype[i]):>3} "
368
+ f"{chain_tag:>1}{residue_index[i]:>4}"
369
+ )
370
+ pdb_lines.append(chain_termination_line)
371
+ atom_index += 1
372
+
373
+ if(i != n - 1):
374
+ # "prev" is a misnomer here. This happens at the beginning of
375
+ # each new chain.
376
+ pdb_lines.extend(get_pdb_headers(prot, prev_chain_index))
377
+
378
+ pdb_lines.append("END")
379
+ pdb_lines.append("")
380
+ return "\n".join(pdb_lines)
381
+
382
+
383
+ def ideal_atom_mask(prot: Protein) -> np.ndarray:
384
+ """Computes an ideal atom mask.
385
+ `Protein.atom_mask` typically is defined according to the atoms that are
386
+ reported in the PDB. This function computes a mask according to heavy atoms
387
+ that should be present in the given sequence of amino acids.
388
+ Args:
389
+ prot: `Protein` whose fields are `numpy.ndarray` objects.
390
+ Returns:
391
+ An ideal atom mask.
392
+ """
393
+ return residue_constants.STANDARD_ATOM_MASK[prot.aatype]
394
+
395
+
396
+ def from_prediction(
397
+ features: FeatureDict,
398
+ result: ModelOutput,
399
+ b_factors: Optional[np.ndarray] = None,
400
+ chain_index: Optional[np.ndarray] = None,
401
+ remark: Optional[str] = None,
402
+ parents: Optional[Sequence[str]] = None,
403
+ parents_chain_index: Optional[Sequence[int]] = None
404
+ ) -> Protein:
405
+ """Assembles a protein from a prediction.
406
+ Args:
407
+ features: Dictionary holding model inputs.
408
+ result: Dictionary holding model outputs.
409
+ b_factors: (Optional) B-factors to use for the protein.
410
+ chain_index: (Optional) Chain indices for multi-chain predictions
411
+ remark: (Optional) Remark about the prediction
412
+ parents: (Optional) List of template names
413
+ Returns:
414
+ A protein instance.
415
+ """
416
+ if b_factors is None:
417
+ b_factors = np.zeros_like(result["final_atom_mask"])
418
+
419
+ return Protein(
420
+ aatype=features["aatype"],
421
+ atom_positions=result["final_atom_positions"],
422
+ atom_mask=result["final_atom_mask"],
423
+ residue_index=features["residue_index"] + 1,
424
+ b_factors=b_factors,
425
+ chain_index=chain_index,
426
+ remark=remark,
427
+ parents=parents,
428
+ parents_chain_index=parents_chain_index,
429
+ )
openfold/residue_constants.py ADDED
@@ -0,0 +1,1310 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2021 AlQuraishi Laboratory
2
+ # Copyright 2021 DeepMind Technologies Limited
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+
16
+ """Constants used in AlphaFold."""
17
+
18
+ import collections
19
+ import functools
20
+ from typing import Mapping, List, Tuple
21
+ from importlib import resources
22
+
23
+ import numpy as np
24
+ import tree
25
+
26
+ # Internal import (35fd).
27
+
28
+
29
+ # Distance from one CA to next CA [trans configuration: omega = 180].
30
+ ca_ca = 3.80209737096
31
+
32
+ # Format: The list for each AA type contains chi1, chi2, chi3, chi4 in
33
+ # this order (or a relevant subset from chi1 onwards). ALA and GLY don't have
34
+ # chi angles so their chi angle lists are empty.
35
+ chi_angles_atoms = {
36
+ "ALA": [],
37
+ # Chi5 in arginine is always 0 +- 5 degrees, so ignore it.
38
+ "ARG": [
39
+ ["N", "CA", "CB", "CG"],
40
+ ["CA", "CB", "CG", "CD"],
41
+ ["CB", "CG", "CD", "NE"],
42
+ ["CG", "CD", "NE", "CZ"],
43
+ ],
44
+ "ASN": [["N", "CA", "CB", "CG"], ["CA", "CB", "CG", "OD1"]],
45
+ "ASP": [["N", "CA", "CB", "CG"], ["CA", "CB", "CG", "OD1"]],
46
+ "CYS": [["N", "CA", "CB", "SG"]],
47
+ "GLN": [
48
+ ["N", "CA", "CB", "CG"],
49
+ ["CA", "CB", "CG", "CD"],
50
+ ["CB", "CG", "CD", "OE1"],
51
+ ],
52
+ "GLU": [
53
+ ["N", "CA", "CB", "CG"],
54
+ ["CA", "CB", "CG", "CD"],
55
+ ["CB", "CG", "CD", "OE1"],
56
+ ],
57
+ "GLY": [],
58
+ "HIS": [["N", "CA", "CB", "CG"], ["CA", "CB", "CG", "ND1"]],
59
+ "ILE": [["N", "CA", "CB", "CG1"], ["CA", "CB", "CG1", "CD1"]],
60
+ "LEU": [["N", "CA", "CB", "CG"], ["CA", "CB", "CG", "CD1"]],
61
+ "LYS": [
62
+ ["N", "CA", "CB", "CG"],
63
+ ["CA", "CB", "CG", "CD"],
64
+ ["CB", "CG", "CD", "CE"],
65
+ ["CG", "CD", "CE", "NZ"],
66
+ ],
67
+ "MET": [
68
+ ["N", "CA", "CB", "CG"],
69
+ ["CA", "CB", "CG", "SD"],
70
+ ["CB", "CG", "SD", "CE"],
71
+ ],
72
+ "PHE": [["N", "CA", "CB", "CG"], ["CA", "CB", "CG", "CD1"]],
73
+ "PRO": [["N", "CA", "CB", "CG"], ["CA", "CB", "CG", "CD"]],
74
+ "SER": [["N", "CA", "CB", "OG"]],
75
+ "THR": [["N", "CA", "CB", "OG1"]],
76
+ "TRP": [["N", "CA", "CB", "CG"], ["CA", "CB", "CG", "CD1"]],
77
+ "TYR": [["N", "CA", "CB", "CG"], ["CA", "CB", "CG", "CD1"]],
78
+ "VAL": [["N", "CA", "CB", "CG1"]],
79
+ }
80
+
81
+ # If chi angles given in fixed-length array, this matrix determines how to mask
82
+ # them for each AA type. The order is as per restype_order (see below).
83
+ chi_angles_mask = [
84
+ [0.0, 0.0, 0.0, 0.0], # ALA
85
+ [1.0, 1.0, 1.0, 1.0], # ARG
86
+ [1.0, 1.0, 0.0, 0.0], # ASN
87
+ [1.0, 1.0, 0.0, 0.0], # ASP
88
+ [1.0, 0.0, 0.0, 0.0], # CYS
89
+ [1.0, 1.0, 1.0, 0.0], # GLN
90
+ [1.0, 1.0, 1.0, 0.0], # GLU
91
+ [0.0, 0.0, 0.0, 0.0], # GLY
92
+ [1.0, 1.0, 0.0, 0.0], # HIS
93
+ [1.0, 1.0, 0.0, 0.0], # ILE
94
+ [1.0, 1.0, 0.0, 0.0], # LEU
95
+ [1.0, 1.0, 1.0, 1.0], # LYS
96
+ [1.0, 1.0, 1.0, 0.0], # MET
97
+ [1.0, 1.0, 0.0, 0.0], # PHE
98
+ [1.0, 1.0, 0.0, 0.0], # PRO
99
+ [1.0, 0.0, 0.0, 0.0], # SER
100
+ [1.0, 0.0, 0.0, 0.0], # THR
101
+ [1.0, 1.0, 0.0, 0.0], # TRP
102
+ [1.0, 1.0, 0.0, 0.0], # TYR
103
+ [1.0, 0.0, 0.0, 0.0], # VAL
104
+ ]
105
+
106
+ # The following chi angles are pi periodic: they can be rotated by a multiple
107
+ # of pi without affecting the structure.
108
+ chi_pi_periodic = [
109
+ [0.0, 0.0, 0.0, 0.0], # ALA
110
+ [0.0, 0.0, 0.0, 0.0], # ARG
111
+ [0.0, 0.0, 0.0, 0.0], # ASN
112
+ [0.0, 1.0, 0.0, 0.0], # ASP
113
+ [0.0, 0.0, 0.0, 0.0], # CYS
114
+ [0.0, 0.0, 0.0, 0.0], # GLN
115
+ [0.0, 0.0, 1.0, 0.0], # GLU
116
+ [0.0, 0.0, 0.0, 0.0], # GLY
117
+ [0.0, 0.0, 0.0, 0.0], # HIS
118
+ [0.0, 0.0, 0.0, 0.0], # ILE
119
+ [0.0, 0.0, 0.0, 0.0], # LEU
120
+ [0.0, 0.0, 0.0, 0.0], # LYS
121
+ [0.0, 0.0, 0.0, 0.0], # MET
122
+ [0.0, 1.0, 0.0, 0.0], # PHE
123
+ [0.0, 0.0, 0.0, 0.0], # PRO
124
+ [0.0, 0.0, 0.0, 0.0], # SER
125
+ [0.0, 0.0, 0.0, 0.0], # THR
126
+ [0.0, 0.0, 0.0, 0.0], # TRP
127
+ [0.0, 1.0, 0.0, 0.0], # TYR
128
+ [0.0, 0.0, 0.0, 0.0], # VAL
129
+ [0.0, 0.0, 0.0, 0.0], # UNK
130
+ ]
131
+
132
+ # Atoms positions relative to the 8 rigid groups, defined by the pre-omega, phi,
133
+ # psi and chi angles:
134
+ # 0: 'backbone group',
135
+ # 1: 'pre-omega-group', (empty)
136
+ # 2: 'phi-group', (currently empty, because it defines only hydrogens)
137
+ # 3: 'psi-group',
138
+ # 4,5,6,7: 'chi1,2,3,4-group'
139
+ # The atom positions are relative to the axis-end-atom of the corresponding
140
+ # rotation axis. The x-axis is in direction of the rotation axis, and the y-axis
141
+ # is defined such that the dihedral-angle-definiting atom (the last entry in
142
+ # chi_angles_atoms above) is in the xy-plane (with a positive y-coordinate).
143
+ # format: [atomname, group_idx, rel_position]
144
+ rigid_group_atom_positions = {
145
+ "ALA": [
146
+ ["N", 0, (-0.525, 1.363, 0.000)],
147
+ ["CA", 0, (0.000, 0.000, 0.000)],
148
+ ["C", 0, (1.526, -0.000, -0.000)],
149
+ ["CB", 0, (-0.529, -0.774, -1.205)],
150
+ ["O", 3, (0.627, 1.062, 0.000)],
151
+ ],
152
+ "ARG": [
153
+ ["N", 0, (-0.524, 1.362, -0.000)],
154
+ ["CA", 0, (0.000, 0.000, 0.000)],
155
+ ["C", 0, (1.525, -0.000, -0.000)],
156
+ ["CB", 0, (-0.524, -0.778, -1.209)],
157
+ ["O", 3, (0.626, 1.062, 0.000)],
158
+ ["CG", 4, (0.616, 1.390, -0.000)],
159
+ ["CD", 5, (0.564, 1.414, 0.000)],
160
+ ["NE", 6, (0.539, 1.357, -0.000)],
161
+ ["NH1", 7, (0.206, 2.301, 0.000)],
162
+ ["NH2", 7, (2.078, 0.978, -0.000)],
163
+ ["CZ", 7, (0.758, 1.093, -0.000)],
164
+ ],
165
+ "ASN": [
166
+ ["N", 0, (-0.536, 1.357, 0.000)],
167
+ ["CA", 0, (0.000, 0.000, 0.000)],
168
+ ["C", 0, (1.526, -0.000, -0.000)],
169
+ ["CB", 0, (-0.531, -0.787, -1.200)],
170
+ ["O", 3, (0.625, 1.062, 0.000)],
171
+ ["CG", 4, (0.584, 1.399, 0.000)],
172
+ ["ND2", 5, (0.593, -1.188, 0.001)],
173
+ ["OD1", 5, (0.633, 1.059, 0.000)],
174
+ ],
175
+ "ASP": [
176
+ ["N", 0, (-0.525, 1.362, -0.000)],
177
+ ["CA", 0, (0.000, 0.000, 0.000)],
178
+ ["C", 0, (1.527, 0.000, -0.000)],
179
+ ["CB", 0, (-0.526, -0.778, -1.208)],
180
+ ["O", 3, (0.626, 1.062, -0.000)],
181
+ ["CG", 4, (0.593, 1.398, -0.000)],
182
+ ["OD1", 5, (0.610, 1.091, 0.000)],
183
+ ["OD2", 5, (0.592, -1.101, -0.003)],
184
+ ],
185
+ "CYS": [
186
+ ["N", 0, (-0.522, 1.362, -0.000)],
187
+ ["CA", 0, (0.000, 0.000, 0.000)],
188
+ ["C", 0, (1.524, 0.000, 0.000)],
189
+ ["CB", 0, (-0.519, -0.773, -1.212)],
190
+ ["O", 3, (0.625, 1.062, -0.000)],
191
+ ["SG", 4, (0.728, 1.653, 0.000)],
192
+ ],
193
+ "GLN": [
194
+ ["N", 0, (-0.526, 1.361, -0.000)],
195
+ ["CA", 0, (0.000, 0.000, 0.000)],
196
+ ["C", 0, (1.526, 0.000, 0.000)],
197
+ ["CB", 0, (-0.525, -0.779, -1.207)],
198
+ ["O", 3, (0.626, 1.062, -0.000)],
199
+ ["CG", 4, (0.615, 1.393, 0.000)],
200
+ ["CD", 5, (0.587, 1.399, -0.000)],
201
+ ["NE2", 6, (0.593, -1.189, -0.001)],
202
+ ["OE1", 6, (0.634, 1.060, 0.000)],
203
+ ],
204
+ "GLU": [
205
+ ["N", 0, (-0.528, 1.361, 0.000)],
206
+ ["CA", 0, (0.000, 0.000, 0.000)],
207
+ ["C", 0, (1.526, -0.000, -0.000)],
208
+ ["CB", 0, (-0.526, -0.781, -1.207)],
209
+ ["O", 3, (0.626, 1.062, 0.000)],
210
+ ["CG", 4, (0.615, 1.392, 0.000)],
211
+ ["CD", 5, (0.600, 1.397, 0.000)],
212
+ ["OE1", 6, (0.607, 1.095, -0.000)],
213
+ ["OE2", 6, (0.589, -1.104, -0.001)],
214
+ ],
215
+ "GLY": [
216
+ ["N", 0, (-0.572, 1.337, 0.000)],
217
+ ["CA", 0, (0.000, 0.000, 0.000)],
218
+ ["C", 0, (1.517, -0.000, -0.000)],
219
+ ["O", 3, (0.626, 1.062, -0.000)],
220
+ ],
221
+ "HIS": [
222
+ ["N", 0, (-0.527, 1.360, 0.000)],
223
+ ["CA", 0, (0.000, 0.000, 0.000)],
224
+ ["C", 0, (1.525, 0.000, 0.000)],
225
+ ["CB", 0, (-0.525, -0.778, -1.208)],
226
+ ["O", 3, (0.625, 1.063, 0.000)],
227
+ ["CG", 4, (0.600, 1.370, -0.000)],
228
+ ["CD2", 5, (0.889, -1.021, 0.003)],
229
+ ["ND1", 5, (0.744, 1.160, -0.000)],
230
+ ["CE1", 5, (2.030, 0.851, 0.002)],
231
+ ["NE2", 5, (2.145, -0.466, 0.004)],
232
+ ],
233
+ "ILE": [
234
+ ["N", 0, (-0.493, 1.373, -0.000)],
235
+ ["CA", 0, (0.000, 0.000, 0.000)],
236
+ ["C", 0, (1.527, -0.000, -0.000)],
237
+ ["CB", 0, (-0.536, -0.793, -1.213)],
238
+ ["O", 3, (0.627, 1.062, -0.000)],
239
+ ["CG1", 4, (0.534, 1.437, -0.000)],
240
+ ["CG2", 4, (0.540, -0.785, -1.199)],
241
+ ["CD1", 5, (0.619, 1.391, 0.000)],
242
+ ],
243
+ "LEU": [
244
+ ["N", 0, (-0.520, 1.363, 0.000)],
245
+ ["CA", 0, (0.000, 0.000, 0.000)],
246
+ ["C", 0, (1.525, -0.000, -0.000)],
247
+ ["CB", 0, (-0.522, -0.773, -1.214)],
248
+ ["O", 3, (0.625, 1.063, -0.000)],
249
+ ["CG", 4, (0.678, 1.371, 0.000)],
250
+ ["CD1", 5, (0.530, 1.430, -0.000)],
251
+ ["CD2", 5, (0.535, -0.774, 1.200)],
252
+ ],
253
+ "LYS": [
254
+ ["N", 0, (-0.526, 1.362, -0.000)],
255
+ ["CA", 0, (0.000, 0.000, 0.000)],
256
+ ["C", 0, (1.526, 0.000, 0.000)],
257
+ ["CB", 0, (-0.524, -0.778, -1.208)],
258
+ ["O", 3, (0.626, 1.062, -0.000)],
259
+ ["CG", 4, (0.619, 1.390, 0.000)],
260
+ ["CD", 5, (0.559, 1.417, 0.000)],
261
+ ["CE", 6, (0.560, 1.416, 0.000)],
262
+ ["NZ", 7, (0.554, 1.387, 0.000)],
263
+ ],
264
+ "MET": [
265
+ ["N", 0, (-0.521, 1.364, -0.000)],
266
+ ["CA", 0, (0.000, 0.000, 0.000)],
267
+ ["C", 0, (1.525, 0.000, 0.000)],
268
+ ["CB", 0, (-0.523, -0.776, -1.210)],
269
+ ["O", 3, (0.625, 1.062, -0.000)],
270
+ ["CG", 4, (0.613, 1.391, -0.000)],
271
+ ["SD", 5, (0.703, 1.695, 0.000)],
272
+ ["CE", 6, (0.320, 1.786, -0.000)],
273
+ ],
274
+ "PHE": [
275
+ ["N", 0, (-0.518, 1.363, 0.000)],
276
+ ["CA", 0, (0.000, 0.000, 0.000)],
277
+ ["C", 0, (1.524, 0.000, -0.000)],
278
+ ["CB", 0, (-0.525, -0.776, -1.212)],
279
+ ["O", 3, (0.626, 1.062, -0.000)],
280
+ ["CG", 4, (0.607, 1.377, 0.000)],
281
+ ["CD1", 5, (0.709, 1.195, -0.000)],
282
+ ["CD2", 5, (0.706, -1.196, 0.000)],
283
+ ["CE1", 5, (2.102, 1.198, -0.000)],
284
+ ["CE2", 5, (2.098, -1.201, -0.000)],
285
+ ["CZ", 5, (2.794, -0.003, -0.001)],
286
+ ],
287
+ "PRO": [
288
+ ["N", 0, (-0.566, 1.351, -0.000)],
289
+ ["CA", 0, (0.000, 0.000, 0.000)],
290
+ ["C", 0, (1.527, -0.000, 0.000)],
291
+ ["CB", 0, (-0.546, -0.611, -1.293)],
292
+ ["O", 3, (0.621, 1.066, 0.000)],
293
+ ["CG", 4, (0.382, 1.445, 0.0)],
294
+ # ['CD', 5, (0.427, 1.440, 0.0)],
295
+ ["CD", 5, (0.477, 1.424, 0.0)], # manually made angle 2 degrees larger
296
+ ],
297
+ "SER": [
298
+ ["N", 0, (-0.529, 1.360, -0.000)],
299
+ ["CA", 0, (0.000, 0.000, 0.000)],
300
+ ["C", 0, (1.525, -0.000, -0.000)],
301
+ ["CB", 0, (-0.518, -0.777, -1.211)],
302
+ ["O", 3, (0.626, 1.062, -0.000)],
303
+ ["OG", 4, (0.503, 1.325, 0.000)],
304
+ ],
305
+ "THR": [
306
+ ["N", 0, (-0.517, 1.364, 0.000)],
307
+ ["CA", 0, (0.000, 0.000, 0.000)],
308
+ ["C", 0, (1.526, 0.000, -0.000)],
309
+ ["CB", 0, (-0.516, -0.793, -1.215)],
310
+ ["O", 3, (0.626, 1.062, 0.000)],
311
+ ["CG2", 4, (0.550, -0.718, -1.228)],
312
+ ["OG1", 4, (0.472, 1.353, 0.000)],
313
+ ],
314
+ "TRP": [
315
+ ["N", 0, (-0.521, 1.363, 0.000)],
316
+ ["CA", 0, (0.000, 0.000, 0.000)],
317
+ ["C", 0, (1.525, -0.000, 0.000)],
318
+ ["CB", 0, (-0.523, -0.776, -1.212)],
319
+ ["O", 3, (0.627, 1.062, 0.000)],
320
+ ["CG", 4, (0.609, 1.370, -0.000)],
321
+ ["CD1", 5, (0.824, 1.091, 0.000)],
322
+ ["CD2", 5, (0.854, -1.148, -0.005)],
323
+ ["CE2", 5, (2.186, -0.678, -0.007)],
324
+ ["CE3", 5, (0.622, -2.530, -0.007)],
325
+ ["NE1", 5, (2.140, 0.690, -0.004)],
326
+ ["CH2", 5, (3.028, -2.890, -0.013)],
327
+ ["CZ2", 5, (3.283, -1.543, -0.011)],
328
+ ["CZ3", 5, (1.715, -3.389, -0.011)],
329
+ ],
330
+ "TYR": [
331
+ ["N", 0, (-0.522, 1.362, 0.000)],
332
+ ["CA", 0, (0.000, 0.000, 0.000)],
333
+ ["C", 0, (1.524, -0.000, -0.000)],
334
+ ["CB", 0, (-0.522, -0.776, -1.213)],
335
+ ["O", 3, (0.627, 1.062, -0.000)],
336
+ ["CG", 4, (0.607, 1.382, -0.000)],
337
+ ["CD1", 5, (0.716, 1.195, -0.000)],
338
+ ["CD2", 5, (0.713, -1.194, -0.001)],
339
+ ["CE1", 5, (2.107, 1.200, -0.002)],
340
+ ["CE2", 5, (2.104, -1.201, -0.003)],
341
+ ["OH", 5, (4.168, -0.002, -0.005)],
342
+ ["CZ", 5, (2.791, -0.001, -0.003)],
343
+ ],
344
+ "VAL": [
345
+ ["N", 0, (-0.494, 1.373, -0.000)],
346
+ ["CA", 0, (0.000, 0.000, 0.000)],
347
+ ["C", 0, (1.527, -0.000, -0.000)],
348
+ ["CB", 0, (-0.533, -0.795, -1.213)],
349
+ ["O", 3, (0.627, 1.062, -0.000)],
350
+ ["CG1", 4, (0.540, 1.429, -0.000)],
351
+ ["CG2", 4, (0.533, -0.776, 1.203)],
352
+ ],
353
+ }
354
+
355
+ # A list of atoms (excluding hydrogen) for each AA type. PDB naming convention.
356
+ residue_atoms = {
357
+ "ALA": ["C", "CA", "CB", "N", "O"],
358
+ "ARG": ["C", "CA", "CB", "CG", "CD", "CZ", "N", "NE", "O", "NH1", "NH2"],
359
+ "ASP": ["C", "CA", "CB", "CG", "N", "O", "OD1", "OD2"],
360
+ "ASN": ["C", "CA", "CB", "CG", "N", "ND2", "O", "OD1"],
361
+ "CYS": ["C", "CA", "CB", "N", "O", "SG"],
362
+ "GLU": ["C", "CA", "CB", "CG", "CD", "N", "O", "OE1", "OE2"],
363
+ "GLN": ["C", "CA", "CB", "CG", "CD", "N", "NE2", "O", "OE1"],
364
+ "GLY": ["C", "CA", "N", "O"],
365
+ "HIS": ["C", "CA", "CB", "CG", "CD2", "CE1", "N", "ND1", "NE2", "O"],
366
+ "ILE": ["C", "CA", "CB", "CG1", "CG2", "CD1", "N", "O"],
367
+ "LEU": ["C", "CA", "CB", "CG", "CD1", "CD2", "N", "O"],
368
+ "LYS": ["C", "CA", "CB", "CG", "CD", "CE", "N", "NZ", "O"],
369
+ "MET": ["C", "CA", "CB", "CG", "CE", "N", "O", "SD"],
370
+ "PHE": ["C", "CA", "CB", "CG", "CD1", "CD2", "CE1", "CE2", "CZ", "N", "O"],
371
+ "PRO": ["C", "CA", "CB", "CG", "CD", "N", "O"],
372
+ "SER": ["C", "CA", "CB", "N", "O", "OG"],
373
+ "THR": ["C", "CA", "CB", "CG2", "N", "O", "OG1"],
374
+ "TRP": [
375
+ "C",
376
+ "CA",
377
+ "CB",
378
+ "CG",
379
+ "CD1",
380
+ "CD2",
381
+ "CE2",
382
+ "CE3",
383
+ "CZ2",
384
+ "CZ3",
385
+ "CH2",
386
+ "N",
387
+ "NE1",
388
+ "O",
389
+ ],
390
+ "TYR": [
391
+ "C",
392
+ "CA",
393
+ "CB",
394
+ "CG",
395
+ "CD1",
396
+ "CD2",
397
+ "CE1",
398
+ "CE2",
399
+ "CZ",
400
+ "N",
401
+ "O",
402
+ "OH",
403
+ ],
404
+ "VAL": ["C", "CA", "CB", "CG1", "CG2", "N", "O"],
405
+ }
406
+
407
+ # Naming swaps for ambiguous atom names.
408
+ # Due to symmetries in the amino acids the naming of atoms is ambiguous in
409
+ # 4 of the 20 amino acids.
410
+ # (The LDDT paper lists 7 amino acids as ambiguous, but the naming ambiguities
411
+ # in LEU, VAL and ARG can be resolved by using the 3d constellations of
412
+ # the 'ambiguous' atoms and their neighbours)
413
+ # TODO: ^ interpret this
414
+ residue_atom_renaming_swaps = {
415
+ "ASP": {"OD1": "OD2"},
416
+ "GLU": {"OE1": "OE2"},
417
+ "PHE": {"CD1": "CD2", "CE1": "CE2"},
418
+ "TYR": {"CD1": "CD2", "CE1": "CE2"},
419
+ }
420
+
421
+ # Van der Waals radii [Angstroem] of the atoms (from Wikipedia)
422
+ van_der_waals_radius = {
423
+ "C": 1.7,
424
+ "N": 1.55,
425
+ "O": 1.52,
426
+ "S": 1.8,
427
+ }
428
+
429
+ Bond = collections.namedtuple(
430
+ "Bond", ["atom1_name", "atom2_name", "length", "stddev"]
431
+ )
432
+ BondAngle = collections.namedtuple(
433
+ "BondAngle",
434
+ ["atom1_name", "atom2_name", "atom3name", "angle_rad", "stddev"],
435
+ )
436
+
437
+
438
+ @functools.lru_cache(maxsize=None)
439
+ def load_stereo_chemical_props() -> Tuple[
440
+ Mapping[str, List[Bond]],
441
+ Mapping[str, List[Bond]],
442
+ Mapping[str, List[BondAngle]],
443
+ ]:
444
+ """Load stereo_chemical_props.txt into a nice structure.
445
+
446
+ Load literature values for bond lengths and bond angles and translate
447
+ bond angles into the length of the opposite edge of the triangle
448
+ ("residue_virtual_bonds").
449
+
450
+ Returns:
451
+ residue_bonds: dict that maps resname --> list of Bond tuples
452
+ residue_virtual_bonds: dict that maps resname --> list of Bond tuples
453
+ residue_bond_angles: dict that maps resname --> list of BondAngle tuples
454
+ """
455
+ # TODO: this file should be downloaded in a setup script
456
+ stereo_chemical_props = resources.read_text("openfold.resources", "stereo_chemical_props.txt")
457
+
458
+ lines_iter = iter(stereo_chemical_props.splitlines())
459
+ # Load bond lengths.
460
+ residue_bonds = {}
461
+ next(lines_iter) # Skip header line.
462
+ for line in lines_iter:
463
+ if line.strip() == "-":
464
+ break
465
+ bond, resname, length, stddev = line.split()
466
+ atom1, atom2 = bond.split("-")
467
+ if resname not in residue_bonds:
468
+ residue_bonds[resname] = []
469
+ residue_bonds[resname].append(
470
+ Bond(atom1, atom2, float(length), float(stddev))
471
+ )
472
+ residue_bonds["UNK"] = []
473
+
474
+ # Load bond angles.
475
+ residue_bond_angles = {}
476
+ next(lines_iter) # Skip empty line.
477
+ next(lines_iter) # Skip header line.
478
+ for line in lines_iter:
479
+ if line.strip() == "-":
480
+ break
481
+ bond, resname, angle_degree, stddev_degree = line.split()
482
+ atom1, atom2, atom3 = bond.split("-")
483
+ if resname not in residue_bond_angles:
484
+ residue_bond_angles[resname] = []
485
+ residue_bond_angles[resname].append(
486
+ BondAngle(
487
+ atom1,
488
+ atom2,
489
+ atom3,
490
+ float(angle_degree) / 180.0 * np.pi,
491
+ float(stddev_degree) / 180.0 * np.pi,
492
+ )
493
+ )
494
+ residue_bond_angles["UNK"] = []
495
+
496
+ def make_bond_key(atom1_name, atom2_name):
497
+ """Unique key to lookup bonds."""
498
+ return "-".join(sorted([atom1_name, atom2_name]))
499
+
500
+ # Translate bond angles into distances ("virtual bonds").
501
+ residue_virtual_bonds = {}
502
+ for resname, bond_angles in residue_bond_angles.items():
503
+ # Create a fast lookup dict for bond lengths.
504
+ bond_cache = {}
505
+ for b in residue_bonds[resname]:
506
+ bond_cache[make_bond_key(b.atom1_name, b.atom2_name)] = b
507
+ residue_virtual_bonds[resname] = []
508
+ for ba in bond_angles:
509
+ bond1 = bond_cache[make_bond_key(ba.atom1_name, ba.atom2_name)]
510
+ bond2 = bond_cache[make_bond_key(ba.atom2_name, ba.atom3name)]
511
+
512
+ # Compute distance between atom1 and atom3 using the law of cosines
513
+ # c^2 = a^2 + b^2 - 2ab*cos(gamma).
514
+ gamma = ba.angle_rad
515
+ length = np.sqrt(
516
+ bond1.length ** 2
517
+ + bond2.length ** 2
518
+ - 2 * bond1.length * bond2.length * np.cos(gamma)
519
+ )
520
+
521
+ # Propagation of uncertainty assuming uncorrelated errors.
522
+ dl_outer = 0.5 / length
523
+ dl_dgamma = (
524
+ 2 * bond1.length * bond2.length * np.sin(gamma)
525
+ ) * dl_outer
526
+ dl_db1 = (
527
+ 2 * bond1.length - 2 * bond2.length * np.cos(gamma)
528
+ ) * dl_outer
529
+ dl_db2 = (
530
+ 2 * bond2.length - 2 * bond1.length * np.cos(gamma)
531
+ ) * dl_outer
532
+ stddev = np.sqrt(
533
+ (dl_dgamma * ba.stddev) ** 2
534
+ + (dl_db1 * bond1.stddev) ** 2
535
+ + (dl_db2 * bond2.stddev) ** 2
536
+ )
537
+ residue_virtual_bonds[resname].append(
538
+ Bond(ba.atom1_name, ba.atom3name, length, stddev)
539
+ )
540
+
541
+ return (residue_bonds, residue_virtual_bonds, residue_bond_angles)
542
+
543
+
544
+ # Between-residue bond lengths for general bonds (first element) and for Proline
545
+ # (second element).
546
+ between_res_bond_length_c_n = [1.329, 1.341]
547
+ between_res_bond_length_stddev_c_n = [0.014, 0.016]
548
+
549
+ # Between-residue cos_angles.
550
+ between_res_cos_angles_c_n_ca = [-0.5203, 0.0353] # degrees: 121.352 +- 2.315
551
+ between_res_cos_angles_ca_c_n = [-0.4473, 0.0311] # degrees: 116.568 +- 1.995
552
+
553
+ # This mapping is used when we need to store atom data in a format that requires
554
+ # fixed atom data size for every residue (e.g. a numpy array).
555
+ atom_types = [
556
+ "N",
557
+ "CA",
558
+ "C",
559
+ "CB",
560
+ "O",
561
+ "CG",
562
+ "CG1",
563
+ "CG2",
564
+ "OG",
565
+ "OG1",
566
+ "SG",
567
+ "CD",
568
+ "CD1",
569
+ "CD2",
570
+ "ND1",
571
+ "ND2",
572
+ "OD1",
573
+ "OD2",
574
+ "SD",
575
+ "CE",
576
+ "CE1",
577
+ "CE2",
578
+ "CE3",
579
+ "NE",
580
+ "NE1",
581
+ "NE2",
582
+ "OE1",
583
+ "OE2",
584
+ "CH2",
585
+ "NH1",
586
+ "NH2",
587
+ "OH",
588
+ "CZ",
589
+ "CZ2",
590
+ "CZ3",
591
+ "NZ",
592
+ "OXT",
593
+ ]
594
+ atom_order = {atom_type: i for i, atom_type in enumerate(atom_types)}
595
+ atom_type_num = len(atom_types) # := 37.
596
+
597
+ # A compact atom encoding with 14 columns
598
+ # pylint: disable=line-too-long
599
+ # pylint: disable=bad-whitespace
600
+ restype_name_to_atom14_names = {
601
+ "ALA": ["N", "CA", "C", "O", "CB", "", "", "", "", "", "", "", "", ""],
602
+ "ARG": [
603
+ "N",
604
+ "CA",
605
+ "C",
606
+ "O",
607
+ "CB",
608
+ "CG",
609
+ "CD",
610
+ "NE",
611
+ "CZ",
612
+ "NH1",
613
+ "NH2",
614
+ "",
615
+ "",
616
+ "",
617
+ ],
618
+ "ASN": [
619
+ "N",
620
+ "CA",
621
+ "C",
622
+ "O",
623
+ "CB",
624
+ "CG",
625
+ "OD1",
626
+ "ND2",
627
+ "",
628
+ "",
629
+ "",
630
+ "",
631
+ "",
632
+ "",
633
+ ],
634
+ "ASP": [
635
+ "N",
636
+ "CA",
637
+ "C",
638
+ "O",
639
+ "CB",
640
+ "CG",
641
+ "OD1",
642
+ "OD2",
643
+ "",
644
+ "",
645
+ "",
646
+ "",
647
+ "",
648
+ "",
649
+ ],
650
+ "CYS": ["N", "CA", "C", "O", "CB", "SG", "", "", "", "", "", "", "", ""],
651
+ "GLN": [
652
+ "N",
653
+ "CA",
654
+ "C",
655
+ "O",
656
+ "CB",
657
+ "CG",
658
+ "CD",
659
+ "OE1",
660
+ "NE2",
661
+ "",
662
+ "",
663
+ "",
664
+ "",
665
+ "",
666
+ ],
667
+ "GLU": [
668
+ "N",
669
+ "CA",
670
+ "C",
671
+ "O",
672
+ "CB",
673
+ "CG",
674
+ "CD",
675
+ "OE1",
676
+ "OE2",
677
+ "",
678
+ "",
679
+ "",
680
+ "",
681
+ "",
682
+ ],
683
+ "GLY": ["N", "CA", "C", "O", "", "", "", "", "", "", "", "", "", ""],
684
+ "HIS": [
685
+ "N",
686
+ "CA",
687
+ "C",
688
+ "O",
689
+ "CB",
690
+ "CG",
691
+ "ND1",
692
+ "CD2",
693
+ "CE1",
694
+ "NE2",
695
+ "",
696
+ "",
697
+ "",
698
+ "",
699
+ ],
700
+ "ILE": [
701
+ "N",
702
+ "CA",
703
+ "C",
704
+ "O",
705
+ "CB",
706
+ "CG1",
707
+ "CG2",
708
+ "CD1",
709
+ "",
710
+ "",
711
+ "",
712
+ "",
713
+ "",
714
+ "",
715
+ ],
716
+ "LEU": [
717
+ "N",
718
+ "CA",
719
+ "C",
720
+ "O",
721
+ "CB",
722
+ "CG",
723
+ "CD1",
724
+ "CD2",
725
+ "",
726
+ "",
727
+ "",
728
+ "",
729
+ "",
730
+ "",
731
+ ],
732
+ "LYS": [
733
+ "N",
734
+ "CA",
735
+ "C",
736
+ "O",
737
+ "CB",
738
+ "CG",
739
+ "CD",
740
+ "CE",
741
+ "NZ",
742
+ "",
743
+ "",
744
+ "",
745
+ "",
746
+ "",
747
+ ],
748
+ "MET": [
749
+ "N",
750
+ "CA",
751
+ "C",
752
+ "O",
753
+ "CB",
754
+ "CG",
755
+ "SD",
756
+ "CE",
757
+ "",
758
+ "",
759
+ "",
760
+ "",
761
+ "",
762
+ "",
763
+ ],
764
+ "PHE": [
765
+ "N",
766
+ "CA",
767
+ "C",
768
+ "O",
769
+ "CB",
770
+ "CG",
771
+ "CD1",
772
+ "CD2",
773
+ "CE1",
774
+ "CE2",
775
+ "CZ",
776
+ "",
777
+ "",
778
+ "",
779
+ ],
780
+ "PRO": ["N", "CA", "C", "O", "CB", "CG", "CD", "", "", "", "", "", "", ""],
781
+ "SER": ["N", "CA", "C", "O", "CB", "OG", "", "", "", "", "", "", "", ""],
782
+ "THR": [
783
+ "N",
784
+ "CA",
785
+ "C",
786
+ "O",
787
+ "CB",
788
+ "OG1",
789
+ "CG2",
790
+ "",
791
+ "",
792
+ "",
793
+ "",
794
+ "",
795
+ "",
796
+ "",
797
+ ],
798
+ "TRP": [
799
+ "N",
800
+ "CA",
801
+ "C",
802
+ "O",
803
+ "CB",
804
+ "CG",
805
+ "CD1",
806
+ "CD2",
807
+ "NE1",
808
+ "CE2",
809
+ "CE3",
810
+ "CZ2",
811
+ "CZ3",
812
+ "CH2",
813
+ ],
814
+ "TYR": [
815
+ "N",
816
+ "CA",
817
+ "C",
818
+ "O",
819
+ "CB",
820
+ "CG",
821
+ "CD1",
822
+ "CD2",
823
+ "CE1",
824
+ "CE2",
825
+ "CZ",
826
+ "OH",
827
+ "",
828
+ "",
829
+ ],
830
+ "VAL": [
831
+ "N",
832
+ "CA",
833
+ "C",
834
+ "O",
835
+ "CB",
836
+ "CG1",
837
+ "CG2",
838
+ "",
839
+ "",
840
+ "",
841
+ "",
842
+ "",
843
+ "",
844
+ "",
845
+ ],
846
+ "UNK": ["", "", "", "", "", "", "", "", "", "", "", "", "", ""],
847
+ }
848
+ # pylint: enable=line-too-long
849
+ # pylint: enable=bad-whitespace
850
+
851
+
852
+ # This is the standard residue order when coding AA type as a number.
853
+ # Reproduce it by taking 3-letter AA codes and sorting them alphabetically.
854
+ restypes = [
855
+ "A",
856
+ "R",
857
+ "N",
858
+ "D",
859
+ "C",
860
+ "Q",
861
+ "E",
862
+ "G",
863
+ "H",
864
+ "I",
865
+ "L",
866
+ "K",
867
+ "M",
868
+ "F",
869
+ "P",
870
+ "S",
871
+ "T",
872
+ "W",
873
+ "Y",
874
+ "V",
875
+ ]
876
+ restype_order = {restype: i for i, restype in enumerate(restypes)}
877
+ restype_num = len(restypes) # := 20.
878
+ unk_restype_index = restype_num # Catch-all index for unknown restypes.
879
+
880
+ restypes_with_x = restypes + ["X"]
881
+ restype_order_with_x = {restype: i for i, restype in enumerate(restypes_with_x)}
882
+
883
+
884
+ def sequence_to_onehot(
885
+ sequence: str, mapping: Mapping[str, int], map_unknown_to_x: bool = False
886
+ ) -> np.ndarray:
887
+ """Maps the given sequence into a one-hot encoded matrix.
888
+
889
+ Args:
890
+ sequence: An amino acid sequence.
891
+ mapping: A dictionary mapping amino acids to integers.
892
+ map_unknown_to_x: If True, any amino acid that is not in the mapping will be
893
+ mapped to the unknown amino acid 'X'. If the mapping doesn't contain
894
+ amino acid 'X', an error will be thrown. If False, any amino acid not in
895
+ the mapping will throw an error.
896
+
897
+ Returns:
898
+ A numpy array of shape (seq_len, num_unique_aas) with one-hot encoding of
899
+ the sequence.
900
+
901
+ Raises:
902
+ ValueError: If the mapping doesn't contain values from 0 to
903
+ num_unique_aas - 1 without any gaps.
904
+ """
905
+ num_entries = max(mapping.values()) + 1
906
+
907
+ if sorted(set(mapping.values())) != list(range(num_entries)):
908
+ raise ValueError(
909
+ "The mapping must have values from 0 to num_unique_aas-1 "
910
+ "without any gaps. Got: %s" % sorted(mapping.values())
911
+ )
912
+
913
+ one_hot_arr = np.zeros((len(sequence), num_entries), dtype=np.int32)
914
+
915
+ for aa_index, aa_type in enumerate(sequence):
916
+ if map_unknown_to_x:
917
+ if aa_type.isalpha() and aa_type.isupper():
918
+ aa_id = mapping.get(aa_type, mapping["X"])
919
+ else:
920
+ raise ValueError(
921
+ f"Invalid character in the sequence: {aa_type}"
922
+ )
923
+ else:
924
+ aa_id = mapping[aa_type]
925
+ one_hot_arr[aa_index, aa_id] = 1
926
+
927
+ return one_hot_arr
928
+
929
+
930
+ restype_1to3 = {
931
+ "A": "ALA",
932
+ "R": "ARG",
933
+ "N": "ASN",
934
+ "D": "ASP",
935
+ "C": "CYS",
936
+ "Q": "GLN",
937
+ "E": "GLU",
938
+ "G": "GLY",
939
+ "H": "HIS",
940
+ "I": "ILE",
941
+ "L": "LEU",
942
+ "K": "LYS",
943
+ "M": "MET",
944
+ "F": "PHE",
945
+ "P": "PRO",
946
+ "S": "SER",
947
+ "T": "THR",
948
+ "W": "TRP",
949
+ "Y": "TYR",
950
+ "V": "VAL",
951
+ }
952
+
953
+
954
+ # NB: restype_3to1 differs from Bio.PDB.protein_letters_3to1 by being a simple
955
+ # 1-to-1 mapping of 3 letter names to one letter names. The latter contains
956
+ # many more, and less common, three letter names as keys and maps many of these
957
+ # to the same one letter name (including 'X' and 'U' which we don't use here).
958
+ restype_3to1 = {v: k for k, v in restype_1to3.items()}
959
+
960
+ # Define a restype name for all unknown residues.
961
+ unk_restype = "UNK"
962
+
963
+ resnames = [restype_1to3[r] for r in restypes] + [unk_restype]
964
+ resname_to_idx = {resname: i for i, resname in enumerate(resnames)}
965
+
966
+
967
+ # The mapping here uses hhblits convention, so that B is mapped to D, J and O
968
+ # are mapped to X, U is mapped to C, and Z is mapped to E. Other than that the
969
+ # remaining 20 amino acids are kept in alphabetical order.
970
+ # There are 2 non-amino acid codes, X (representing any amino acid) and
971
+ # "-" representing a missing amino acid in an alignment. The id for these
972
+ # codes is put at the end (20 and 21) so that they can easily be ignored if
973
+ # desired.
974
+ HHBLITS_AA_TO_ID = {
975
+ "A": 0,
976
+ "B": 2,
977
+ "C": 1,
978
+ "D": 2,
979
+ "E": 3,
980
+ "F": 4,
981
+ "G": 5,
982
+ "H": 6,
983
+ "I": 7,
984
+ "J": 20,
985
+ "K": 8,
986
+ "L": 9,
987
+ "M": 10,
988
+ "N": 11,
989
+ "O": 20,
990
+ "P": 12,
991
+ "Q": 13,
992
+ "R": 14,
993
+ "S": 15,
994
+ "T": 16,
995
+ "U": 1,
996
+ "V": 17,
997
+ "W": 18,
998
+ "X": 20,
999
+ "Y": 19,
1000
+ "Z": 3,
1001
+ "-": 21,
1002
+ }
1003
+
1004
+ # Partial inversion of HHBLITS_AA_TO_ID.
1005
+ ID_TO_HHBLITS_AA = {
1006
+ 0: "A",
1007
+ 1: "C", # Also U.
1008
+ 2: "D", # Also B.
1009
+ 3: "E", # Also Z.
1010
+ 4: "F",
1011
+ 5: "G",
1012
+ 6: "H",
1013
+ 7: "I",
1014
+ 8: "K",
1015
+ 9: "L",
1016
+ 10: "M",
1017
+ 11: "N",
1018
+ 12: "P",
1019
+ 13: "Q",
1020
+ 14: "R",
1021
+ 15: "S",
1022
+ 16: "T",
1023
+ 17: "V",
1024
+ 18: "W",
1025
+ 19: "Y",
1026
+ 20: "X", # Includes J and O.
1027
+ 21: "-",
1028
+ }
1029
+
1030
+ restypes_with_x_and_gap = restypes + ["X", "-"]
1031
+ MAP_HHBLITS_AATYPE_TO_OUR_AATYPE = tuple(
1032
+ restypes_with_x_and_gap.index(ID_TO_HHBLITS_AA[i])
1033
+ for i in range(len(restypes_with_x_and_gap))
1034
+ )
1035
+
1036
+
1037
+ def _make_standard_atom_mask() -> np.ndarray:
1038
+ """Returns [num_res_types, num_atom_types] mask array."""
1039
+ # +1 to account for unknown (all 0s).
1040
+ mask = np.zeros([restype_num + 1, atom_type_num], dtype=np.int32)
1041
+ for restype, restype_letter in enumerate(restypes):
1042
+ restype_name = restype_1to3[restype_letter]
1043
+ atom_names = residue_atoms[restype_name]
1044
+ for atom_name in atom_names:
1045
+ atom_type = atom_order[atom_name]
1046
+ mask[restype, atom_type] = 1
1047
+ return mask
1048
+
1049
+
1050
+ STANDARD_ATOM_MASK = _make_standard_atom_mask()
1051
+
1052
+
1053
+ # A one hot representation for the first and second atoms defining the axis
1054
+ # of rotation for each chi-angle in each residue.
1055
+ def chi_angle_atom(atom_index: int) -> np.ndarray:
1056
+ """Define chi-angle rigid groups via one-hot representations."""
1057
+ chi_angles_index = {}
1058
+ one_hots = []
1059
+
1060
+ for k, v in chi_angles_atoms.items():
1061
+ indices = [atom_types.index(s[atom_index]) for s in v]
1062
+ indices.extend([-1] * (4 - len(indices)))
1063
+ chi_angles_index[k] = indices
1064
+
1065
+ for r in restypes:
1066
+ res3 = restype_1to3[r]
1067
+ one_hot = np.eye(atom_type_num)[chi_angles_index[res3]]
1068
+ one_hots.append(one_hot)
1069
+
1070
+ one_hots.append(np.zeros([4, atom_type_num])) # Add zeros for residue `X`.
1071
+ one_hot = np.stack(one_hots, axis=0)
1072
+ one_hot = np.transpose(one_hot, [0, 2, 1])
1073
+
1074
+ return one_hot
1075
+
1076
+
1077
+ chi_atom_1_one_hot = chi_angle_atom(1)
1078
+ chi_atom_2_one_hot = chi_angle_atom(2)
1079
+
1080
+ # An array like chi_angles_atoms but using indices rather than names.
1081
+ chi_angles_atom_indices = [chi_angles_atoms[restype_1to3[r]] for r in restypes]
1082
+ chi_angles_atom_indices = tree.map_structure(
1083
+ lambda atom_name: atom_order[atom_name], chi_angles_atom_indices
1084
+ )
1085
+ chi_angles_atom_indices = np.array(
1086
+ [
1087
+ chi_atoms + ([[0, 0, 0, 0]] * (4 - len(chi_atoms)))
1088
+ for chi_atoms in chi_angles_atom_indices
1089
+ ]
1090
+ )
1091
+
1092
+ # Mapping from (res_name, atom_name) pairs to the atom's chi group index
1093
+ # and atom index within that group.
1094
+ chi_groups_for_atom = collections.defaultdict(list)
1095
+ for res_name, chi_angle_atoms_for_res in chi_angles_atoms.items():
1096
+ for chi_group_i, chi_group in enumerate(chi_angle_atoms_for_res):
1097
+ for atom_i, atom in enumerate(chi_group):
1098
+ chi_groups_for_atom[(res_name, atom)].append((chi_group_i, atom_i))
1099
+ chi_groups_for_atom = dict(chi_groups_for_atom)
1100
+
1101
+
1102
+ def _make_rigid_transformation_4x4(ex, ey, translation):
1103
+ """Create a rigid 4x4 transformation matrix from two axes and transl."""
1104
+ # Normalize ex.
1105
+ ex_normalized = ex / np.linalg.norm(ex)
1106
+
1107
+ # make ey perpendicular to ex
1108
+ ey_normalized = ey - np.dot(ey, ex_normalized) * ex_normalized
1109
+ ey_normalized /= np.linalg.norm(ey_normalized)
1110
+
1111
+ # compute ez as cross product
1112
+ eznorm = np.cross(ex_normalized, ey_normalized)
1113
+ m = np.stack(
1114
+ [ex_normalized, ey_normalized, eznorm, translation]
1115
+ ).transpose()
1116
+ m = np.concatenate([m, [[0.0, 0.0, 0.0, 1.0]]], axis=0)
1117
+ return m
1118
+
1119
+
1120
+ # create an array with (restype, atomtype) --> rigid_group_idx
1121
+ # and an array with (restype, atomtype, coord) for the atom positions
1122
+ # and compute affine transformation matrices (4,4) from one rigid group to the
1123
+ # previous group
1124
+ restype_atom37_to_rigid_group = np.zeros([21, 37], dtype=np.int)
1125
+ restype_atom37_mask = np.zeros([21, 37], dtype=np.float32)
1126
+ restype_atom37_rigid_group_positions = np.zeros([21, 37, 3], dtype=np.float32)
1127
+ restype_atom14_to_rigid_group = np.zeros([21, 14], dtype=np.int)
1128
+ restype_atom14_mask = np.zeros([21, 14], dtype=np.float32)
1129
+ restype_atom14_rigid_group_positions = np.zeros([21, 14, 3], dtype=np.float32)
1130
+ restype_rigid_group_default_frame = np.zeros([21, 8, 4, 4], dtype=np.float32)
1131
+
1132
+
1133
+ def _make_rigid_group_constants():
1134
+ """Fill the arrays above."""
1135
+ for restype, restype_letter in enumerate(restypes):
1136
+ resname = restype_1to3[restype_letter]
1137
+ for atomname, group_idx, atom_position in rigid_group_atom_positions[
1138
+ resname
1139
+ ]:
1140
+ atomtype = atom_order[atomname]
1141
+ restype_atom37_to_rigid_group[restype, atomtype] = group_idx
1142
+ restype_atom37_mask[restype, atomtype] = 1
1143
+ restype_atom37_rigid_group_positions[
1144
+ restype, atomtype, :
1145
+ ] = atom_position
1146
+
1147
+ atom14idx = restype_name_to_atom14_names[resname].index(atomname)
1148
+ restype_atom14_to_rigid_group[restype, atom14idx] = group_idx
1149
+ restype_atom14_mask[restype, atom14idx] = 1
1150
+ restype_atom14_rigid_group_positions[
1151
+ restype, atom14idx, :
1152
+ ] = atom_position
1153
+
1154
+ for restype, restype_letter in enumerate(restypes):
1155
+ resname = restype_1to3[restype_letter]
1156
+ atom_positions = {
1157
+ name: np.array(pos)
1158
+ for name, _, pos in rigid_group_atom_positions[resname]
1159
+ }
1160
+
1161
+ # backbone to backbone is the identity transform
1162
+ restype_rigid_group_default_frame[restype, 0, :, :] = np.eye(4)
1163
+
1164
+ # pre-omega-frame to backbone (currently dummy identity matrix)
1165
+ restype_rigid_group_default_frame[restype, 1, :, :] = np.eye(4)
1166
+
1167
+ # phi-frame to backbone
1168
+ mat = _make_rigid_transformation_4x4(
1169
+ ex=atom_positions["N"] - atom_positions["CA"],
1170
+ ey=np.array([1.0, 0.0, 0.0]),
1171
+ translation=atom_positions["N"],
1172
+ )
1173
+ restype_rigid_group_default_frame[restype, 2, :, :] = mat
1174
+
1175
+ # psi-frame to backbone
1176
+ mat = _make_rigid_transformation_4x4(
1177
+ ex=atom_positions["C"] - atom_positions["CA"],
1178
+ ey=atom_positions["CA"] - atom_positions["N"],
1179
+ translation=atom_positions["C"],
1180
+ )
1181
+ restype_rigid_group_default_frame[restype, 3, :, :] = mat
1182
+
1183
+ # chi1-frame to backbone
1184
+ if chi_angles_mask[restype][0]:
1185
+ base_atom_names = chi_angles_atoms[resname][0]
1186
+ base_atom_positions = [
1187
+ atom_positions[name] for name in base_atom_names
1188
+ ]
1189
+ mat = _make_rigid_transformation_4x4(
1190
+ ex=base_atom_positions[2] - base_atom_positions[1],
1191
+ ey=base_atom_positions[0] - base_atom_positions[1],
1192
+ translation=base_atom_positions[2],
1193
+ )
1194
+ restype_rigid_group_default_frame[restype, 4, :, :] = mat
1195
+
1196
+ # chi2-frame to chi1-frame
1197
+ # chi3-frame to chi2-frame
1198
+ # chi4-frame to chi3-frame
1199
+ # luckily all rotation axes for the next frame start at (0,0,0) of the
1200
+ # previous frame
1201
+ for chi_idx in range(1, 4):
1202
+ if chi_angles_mask[restype][chi_idx]:
1203
+ axis_end_atom_name = chi_angles_atoms[resname][chi_idx][2]
1204
+ axis_end_atom_position = atom_positions[axis_end_atom_name]
1205
+ mat = _make_rigid_transformation_4x4(
1206
+ ex=axis_end_atom_position,
1207
+ ey=np.array([-1.0, 0.0, 0.0]),
1208
+ translation=axis_end_atom_position,
1209
+ )
1210
+ restype_rigid_group_default_frame[
1211
+ restype, 4 + chi_idx, :, :
1212
+ ] = mat
1213
+
1214
+
1215
+ _make_rigid_group_constants()
1216
+
1217
+
1218
+ def make_atom14_dists_bounds(
1219
+ overlap_tolerance=1.5, bond_length_tolerance_factor=15
1220
+ ):
1221
+ """compute upper and lower bounds for bonds to assess violations."""
1222
+ restype_atom14_bond_lower_bound = np.zeros([21, 14, 14], np.float32)
1223
+ restype_atom14_bond_upper_bound = np.zeros([21, 14, 14], np.float32)
1224
+ restype_atom14_bond_stddev = np.zeros([21, 14, 14], np.float32)
1225
+ residue_bonds, residue_virtual_bonds, _ = load_stereo_chemical_props()
1226
+ for restype, restype_letter in enumerate(restypes):
1227
+ resname = restype_1to3[restype_letter]
1228
+ atom_list = restype_name_to_atom14_names[resname]
1229
+
1230
+ # create lower and upper bounds for clashes
1231
+ for atom1_idx, atom1_name in enumerate(atom_list):
1232
+ if not atom1_name:
1233
+ continue
1234
+ atom1_radius = van_der_waals_radius[atom1_name[0]]
1235
+ for atom2_idx, atom2_name in enumerate(atom_list):
1236
+ if (not atom2_name) or atom1_idx == atom2_idx:
1237
+ continue
1238
+ atom2_radius = van_der_waals_radius[atom2_name[0]]
1239
+ lower = atom1_radius + atom2_radius - overlap_tolerance
1240
+ upper = 1e10
1241
+ restype_atom14_bond_lower_bound[
1242
+ restype, atom1_idx, atom2_idx
1243
+ ] = lower
1244
+ restype_atom14_bond_lower_bound[
1245
+ restype, atom2_idx, atom1_idx
1246
+ ] = lower
1247
+ restype_atom14_bond_upper_bound[
1248
+ restype, atom1_idx, atom2_idx
1249
+ ] = upper
1250
+ restype_atom14_bond_upper_bound[
1251
+ restype, atom2_idx, atom1_idx
1252
+ ] = upper
1253
+
1254
+ # overwrite lower and upper bounds for bonds and angles
1255
+ for b in residue_bonds[resname] + residue_virtual_bonds[resname]:
1256
+ atom1_idx = atom_list.index(b.atom1_name)
1257
+ atom2_idx = atom_list.index(b.atom2_name)
1258
+ lower = b.length - bond_length_tolerance_factor * b.stddev
1259
+ upper = b.length + bond_length_tolerance_factor * b.stddev
1260
+ restype_atom14_bond_lower_bound[
1261
+ restype, atom1_idx, atom2_idx
1262
+ ] = lower
1263
+ restype_atom14_bond_lower_bound[
1264
+ restype, atom2_idx, atom1_idx
1265
+ ] = lower
1266
+ restype_atom14_bond_upper_bound[
1267
+ restype, atom1_idx, atom2_idx
1268
+ ] = upper
1269
+ restype_atom14_bond_upper_bound[
1270
+ restype, atom2_idx, atom1_idx
1271
+ ] = upper
1272
+ restype_atom14_bond_stddev[restype, atom1_idx, atom2_idx] = b.stddev
1273
+ restype_atom14_bond_stddev[restype, atom2_idx, atom1_idx] = b.stddev
1274
+ return {
1275
+ "lower_bound": restype_atom14_bond_lower_bound, # shape (21,14,14)
1276
+ "upper_bound": restype_atom14_bond_upper_bound, # shape (21,14,14)
1277
+ "stddev": restype_atom14_bond_stddev, # shape (21,14,14)
1278
+ }
1279
+
1280
+
1281
+ restype_atom14_ambiguous_atoms = np.zeros((21, 14), dtype=np.float32)
1282
+ restype_atom14_ambiguous_atoms_swap_idx = np.tile(
1283
+ np.arange(14, dtype=np.int), (21, 1)
1284
+ )
1285
+
1286
+
1287
+ def _make_atom14_ambiguity_feats():
1288
+ for res, pairs in residue_atom_renaming_swaps.items():
1289
+ res_idx = restype_order[restype_3to1[res]]
1290
+ for atom1, atom2 in pairs.items():
1291
+ atom1_idx = restype_name_to_atom14_names[res].index(atom1)
1292
+ atom2_idx = restype_name_to_atom14_names[res].index(atom2)
1293
+ restype_atom14_ambiguous_atoms[res_idx, atom1_idx] = 1
1294
+ restype_atom14_ambiguous_atoms[res_idx, atom2_idx] = 1
1295
+ restype_atom14_ambiguous_atoms_swap_idx[
1296
+ res_idx, atom1_idx
1297
+ ] = atom2_idx
1298
+ restype_atom14_ambiguous_atoms_swap_idx[
1299
+ res_idx, atom2_idx
1300
+ ] = atom1_idx
1301
+
1302
+
1303
+ _make_atom14_ambiguity_feats()
1304
+
1305
+
1306
+ def aatype_to_str_sequence(aatype):
1307
+ return ''.join([
1308
+ restypes_with_x[aatype[i]]
1309
+ for i in range(len(aatype))
1310
+ ])
openfold/rigid_utils.py ADDED
@@ -0,0 +1,1368 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2021 AlQuraishi Laboratory
2
+ # Copyright 2021 DeepMind Technologies Limited
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+
16
+ from __future__ import annotations
17
+ from typing import Tuple, Any, Sequence, Callable, Optional
18
+
19
+ import numpy as np
20
+ import torch
21
+
22
+
23
+ def rot_matmul(
24
+ a: torch.Tensor,
25
+ b: torch.Tensor
26
+ ) -> torch.Tensor:
27
+ """
28
+ Performs matrix multiplication of two rotation matrix tensors. Written
29
+ out by hand to avoid AMP downcasting.
30
+
31
+ Args:
32
+ a: [*, 3, 3] left multiplicand
33
+ b: [*, 3, 3] right multiplicand
34
+ Returns:
35
+ The product ab
36
+ """
37
+ def row_mul(i):
38
+ return torch.stack(
39
+ [
40
+ a[..., i, 0] * b[..., 0, 0]
41
+ + a[..., i, 1] * b[..., 1, 0]
42
+ + a[..., i, 2] * b[..., 2, 0],
43
+ a[..., i, 0] * b[..., 0, 1]
44
+ + a[..., i, 1] * b[..., 1, 1]
45
+ + a[..., i, 2] * b[..., 2, 1],
46
+ a[..., i, 0] * b[..., 0, 2]
47
+ + a[..., i, 1] * b[..., 1, 2]
48
+ + a[..., i, 2] * b[..., 2, 2],
49
+ ],
50
+ dim=-1,
51
+ )
52
+
53
+ return torch.stack(
54
+ [
55
+ row_mul(0),
56
+ row_mul(1),
57
+ row_mul(2),
58
+ ],
59
+ dim=-2
60
+ )
61
+
62
+
63
+ def rot_vec_mul(
64
+ r: torch.Tensor,
65
+ t: torch.Tensor
66
+ ) -> torch.Tensor:
67
+ """
68
+ Applies a rotation to a vector. Written out by hand to avoid transfer
69
+ to avoid AMP downcasting.
70
+
71
+ Args:
72
+ r: [*, 3, 3] rotation matrices
73
+ t: [*, 3] coordinate tensors
74
+ Returns:
75
+ [*, 3] rotated coordinates
76
+ """
77
+ x, y, z = torch.unbind(t, dim=-1)
78
+ return torch.stack(
79
+ [
80
+ r[..., 0, 0] * x + r[..., 0, 1] * y + r[..., 0, 2] * z,
81
+ r[..., 1, 0] * x + r[..., 1, 1] * y + r[..., 1, 2] * z,
82
+ r[..., 2, 0] * x + r[..., 2, 1] * y + r[..., 2, 2] * z,
83
+ ],
84
+ dim=-1,
85
+ )
86
+
87
+
88
+ def identity_rot_mats(
89
+ batch_dims: Tuple[int],
90
+ dtype: Optional[torch.dtype] = None,
91
+ device: Optional[torch.device] = None,
92
+ requires_grad: bool = True,
93
+ ) -> torch.Tensor:
94
+ rots = torch.eye(
95
+ 3, dtype=dtype, device=device, requires_grad=requires_grad
96
+ )
97
+ rots = rots.view(*((1,) * len(batch_dims)), 3, 3)
98
+ rots = rots.expand(*batch_dims, -1, -1)
99
+ rots = rots.contiguous()
100
+
101
+ return rots
102
+
103
+
104
+ def identity_trans(
105
+ batch_dims: Tuple[int],
106
+ dtype: Optional[torch.dtype] = None,
107
+ device: Optional[torch.device] = None,
108
+ requires_grad: bool = True,
109
+ ) -> torch.Tensor:
110
+ trans = torch.zeros(
111
+ (*batch_dims, 3),
112
+ dtype=dtype,
113
+ device=device,
114
+ requires_grad=requires_grad
115
+ )
116
+ return trans
117
+
118
+
119
+ def identity_quats(
120
+ batch_dims: Tuple[int],
121
+ dtype: Optional[torch.dtype] = None,
122
+ device: Optional[torch.device] = None,
123
+ requires_grad: bool = True,
124
+ ) -> torch.Tensor:
125
+ quat = torch.zeros(
126
+ (*batch_dims, 4),
127
+ dtype=dtype,
128
+ device=device,
129
+ requires_grad=requires_grad
130
+ )
131
+
132
+ with torch.no_grad():
133
+ quat[..., 0] = 1
134
+
135
+ return quat
136
+
137
+
138
+ _quat_elements = ["a", "b", "c", "d"]
139
+ _qtr_keys = [l1 + l2 for l1 in _quat_elements for l2 in _quat_elements]
140
+ _qtr_ind_dict = {key: ind for ind, key in enumerate(_qtr_keys)}
141
+
142
+
143
+ def _to_mat(pairs):
144
+ mat = np.zeros((4, 4))
145
+ for pair in pairs:
146
+ key, value = pair
147
+ ind = _qtr_ind_dict[key]
148
+ mat[ind // 4][ind % 4] = value
149
+
150
+ return mat
151
+
152
+
153
+ _QTR_MAT = np.zeros((4, 4, 3, 3))
154
+ _QTR_MAT[..., 0, 0] = _to_mat([("aa", 1), ("bb", 1), ("cc", -1), ("dd", -1)])
155
+ _QTR_MAT[..., 0, 1] = _to_mat([("bc", 2), ("ad", -2)])
156
+ _QTR_MAT[..., 0, 2] = _to_mat([("bd", 2), ("ac", 2)])
157
+ _QTR_MAT[..., 1, 0] = _to_mat([("bc", 2), ("ad", 2)])
158
+ _QTR_MAT[..., 1, 1] = _to_mat([("aa", 1), ("bb", -1), ("cc", 1), ("dd", -1)])
159
+ _QTR_MAT[..., 1, 2] = _to_mat([("cd", 2), ("ab", -2)])
160
+ _QTR_MAT[..., 2, 0] = _to_mat([("bd", 2), ("ac", -2)])
161
+ _QTR_MAT[..., 2, 1] = _to_mat([("cd", 2), ("ab", 2)])
162
+ _QTR_MAT[..., 2, 2] = _to_mat([("aa", 1), ("bb", -1), ("cc", -1), ("dd", 1)])
163
+
164
+
165
+ def quat_to_rot(quat: torch.Tensor) -> torch.Tensor:
166
+ """
167
+ Converts a quaternion to a rotation matrix.
168
+
169
+ Args:
170
+ quat: [*, 4] quaternions
171
+ Returns:
172
+ [*, 3, 3] rotation matrices
173
+ """
174
+ # [*, 4, 4]
175
+ quat = quat[..., None] * quat[..., None, :]
176
+
177
+ # [4, 4, 3, 3]
178
+ mat = quat.new_tensor(_QTR_MAT, requires_grad=False)
179
+
180
+ # [*, 4, 4, 3, 3]
181
+ shaped_qtr_mat = mat.view((1,) * len(quat.shape[:-2]) + mat.shape)
182
+ quat = quat[..., None, None] * shaped_qtr_mat
183
+
184
+ # [*, 3, 3]
185
+ return torch.sum(quat, dim=(-3, -4))
186
+
187
+
188
+ def rot_to_quat(
189
+ rot: torch.Tensor,
190
+ ):
191
+ if(rot.shape[-2:] != (3, 3)):
192
+ raise ValueError("Input rotation is incorrectly shaped")
193
+
194
+ rot = [[rot[..., i, j] for j in range(3)] for i in range(3)]
195
+ [[xx, xy, xz], [yx, yy, yz], [zx, zy, zz]] = rot
196
+
197
+ k = [
198
+ [ xx + yy + zz, zy - yz, xz - zx, yx - xy,],
199
+ [ zy - yz, xx - yy - zz, xy + yx, xz + zx,],
200
+ [ xz - zx, xy + yx, yy - xx - zz, yz + zy,],
201
+ [ yx - xy, xz + zx, yz + zy, zz - xx - yy,]
202
+ ]
203
+
204
+ k = (1./3.) * torch.stack([torch.stack(t, dim=-1) for t in k], dim=-2)
205
+
206
+ _, vectors = torch.linalg.eigh(k)
207
+ return vectors[..., -1]
208
+
209
+
210
+ _QUAT_MULTIPLY = np.zeros((4, 4, 4))
211
+ _QUAT_MULTIPLY[:, :, 0] = [[ 1, 0, 0, 0],
212
+ [ 0,-1, 0, 0],
213
+ [ 0, 0,-1, 0],
214
+ [ 0, 0, 0,-1]]
215
+
216
+ _QUAT_MULTIPLY[:, :, 1] = [[ 0, 1, 0, 0],
217
+ [ 1, 0, 0, 0],
218
+ [ 0, 0, 0, 1],
219
+ [ 0, 0,-1, 0]]
220
+
221
+ _QUAT_MULTIPLY[:, :, 2] = [[ 0, 0, 1, 0],
222
+ [ 0, 0, 0,-1],
223
+ [ 1, 0, 0, 0],
224
+ [ 0, 1, 0, 0]]
225
+
226
+ _QUAT_MULTIPLY[:, :, 3] = [[ 0, 0, 0, 1],
227
+ [ 0, 0, 1, 0],
228
+ [ 0,-1, 0, 0],
229
+ [ 1, 0, 0, 0]]
230
+
231
+ _QUAT_MULTIPLY_BY_VEC = _QUAT_MULTIPLY[:, 1:, :]
232
+
233
+
234
+ def quat_multiply(quat1, quat2):
235
+ """Multiply a quaternion by another quaternion."""
236
+ mat = quat1.new_tensor(_QUAT_MULTIPLY)
237
+ reshaped_mat = mat.view((1,) * len(quat1.shape[:-1]) + mat.shape)
238
+ return torch.sum(
239
+ reshaped_mat *
240
+ quat1[..., :, None, None] *
241
+ quat2[..., None, :, None],
242
+ dim=(-3, -2)
243
+ )
244
+
245
+
246
+ def quat_multiply_by_vec(quat, vec):
247
+ """Multiply a quaternion by a pure-vector quaternion."""
248
+ mat = quat.new_tensor(_QUAT_MULTIPLY_BY_VEC)
249
+ reshaped_mat = mat.view((1,) * len(quat.shape[:-1]) + mat.shape)
250
+ return torch.sum(
251
+ reshaped_mat *
252
+ quat[..., :, None, None] *
253
+ vec[..., None, :, None],
254
+ dim=(-3, -2)
255
+ )
256
+
257
+
258
+ def invert_rot_mat(rot_mat: torch.Tensor):
259
+ return rot_mat.transpose(-1, -2)
260
+
261
+
262
+ def invert_quat(quat: torch.Tensor):
263
+ quat_prime = quat.clone()
264
+ quat_prime[..., 1:] *= -1
265
+ inv = quat_prime / torch.sum(quat ** 2, dim=-1, keepdim=True)
266
+ return inv
267
+
268
+
269
+ class Rotation:
270
+ """
271
+ A 3D rotation. Depending on how the object is initialized, the
272
+ rotation is represented by either a rotation matrix or a
273
+ quaternion, though both formats are made available by helper functions.
274
+ To simplify gradient computation, the underlying format of the
275
+ rotation cannot be changed in-place. Like Rigid, the class is designed
276
+ to mimic the behavior of a torch Tensor, almost as if each Rotation
277
+ object were a tensor of rotations, in one format or another.
278
+ """
279
+ def __init__(self,
280
+ rot_mats: Optional[torch.Tensor] = None,
281
+ quats: Optional[torch.Tensor] = None,
282
+ normalize_quats: bool = True,
283
+ ):
284
+ """
285
+ Args:
286
+ rot_mats:
287
+ A [*, 3, 3] rotation matrix tensor. Mutually exclusive with
288
+ quats
289
+ quats:
290
+ A [*, 4] quaternion. Mutually exclusive with rot_mats. If
291
+ normalize_quats is not True, must be a unit quaternion
292
+ normalize_quats:
293
+ If quats is specified, whether to normalize quats
294
+ """
295
+ if((rot_mats is None and quats is None) or
296
+ (rot_mats is not None and quats is not None)):
297
+ raise ValueError("Exactly one input argument must be specified")
298
+
299
+ if((rot_mats is not None and rot_mats.shape[-2:] != (3, 3)) or
300
+ (quats is not None and quats.shape[-1] != 4)):
301
+ raise ValueError(
302
+ "Incorrectly shaped rotation matrix or quaternion"
303
+ )
304
+
305
+ # Force full-precision
306
+ if(quats is not None):
307
+ quats = quats.to(dtype=torch.float32)
308
+ if(rot_mats is not None):
309
+ rot_mats = rot_mats.to(dtype=torch.float32)
310
+
311
+ if(quats is not None and normalize_quats):
312
+ quats = quats / torch.linalg.norm(quats, dim=-1, keepdim=True)
313
+
314
+ self._rot_mats = rot_mats
315
+ self._quats = quats
316
+
317
+ @staticmethod
318
+ def identity(
319
+ shape,
320
+ dtype: Optional[torch.dtype] = None,
321
+ device: Optional[torch.device] = None,
322
+ requires_grad: bool = True,
323
+ fmt: str = "quat",
324
+ ) -> Rotation:
325
+ """
326
+ Returns an identity Rotation.
327
+
328
+ Args:
329
+ shape:
330
+ The "shape" of the resulting Rotation object. See documentation
331
+ for the shape property
332
+ dtype:
333
+ The torch dtype for the rotation
334
+ device:
335
+ The torch device for the new rotation
336
+ requires_grad:
337
+ Whether the underlying tensors in the new rotation object
338
+ should require gradient computation
339
+ fmt:
340
+ One of "quat" or "rot_mat". Determines the underlying format
341
+ of the new object's rotation
342
+ Returns:
343
+ A new identity rotation
344
+ """
345
+ if(fmt == "rot_mat"):
346
+ rot_mats = identity_rot_mats(
347
+ shape, dtype, device, requires_grad,
348
+ )
349
+ return Rotation(rot_mats=rot_mats, quats=None)
350
+ elif(fmt == "quat"):
351
+ quats = identity_quats(shape, dtype, device, requires_grad)
352
+ return Rotation(rot_mats=None, quats=quats, normalize_quats=False)
353
+ else:
354
+ raise ValueError(f"Invalid format: f{fmt}")
355
+
356
+ # Magic methods
357
+
358
+ def __getitem__(self, index: Any) -> Rotation:
359
+ """
360
+ Allows torch-style indexing over the virtual shape of the rotation
361
+ object. See documentation for the shape property.
362
+
363
+ Args:
364
+ index:
365
+ A torch index. E.g. (1, 3, 2), or (slice(None,))
366
+ Returns:
367
+ The indexed rotation
368
+ """
369
+ if type(index) != tuple:
370
+ index = (index,)
371
+
372
+ if(self._rot_mats is not None):
373
+ rot_mats = self._rot_mats[index + (slice(None), slice(None))]
374
+ return Rotation(rot_mats=rot_mats)
375
+ elif(self._quats is not None):
376
+ quats = self._quats[index + (slice(None),)]
377
+ return Rotation(quats=quats, normalize_quats=False)
378
+ else:
379
+ raise ValueError("Both rotations are None")
380
+
381
+ def __mul__(self,
382
+ right: torch.Tensor,
383
+ ) -> Rotation:
384
+ """
385
+ Pointwise left multiplication of the rotation with a tensor. Can be
386
+ used to e.g. mask the Rotation.
387
+
388
+ Args:
389
+ right:
390
+ The tensor multiplicand
391
+ Returns:
392
+ The product
393
+ """
394
+ if not(isinstance(right, torch.Tensor)):
395
+ raise TypeError("The other multiplicand must be a Tensor")
396
+
397
+ if(self._rot_mats is not None):
398
+ rot_mats = self._rot_mats * right[..., None, None]
399
+ return Rotation(rot_mats=rot_mats, quats=None)
400
+ elif(self._quats is not None):
401
+ quats = self._quats * right[..., None]
402
+ return Rotation(rot_mats=None, quats=quats, normalize_quats=False)
403
+ else:
404
+ raise ValueError("Both rotations are None")
405
+
406
+ def __rmul__(self,
407
+ left: torch.Tensor,
408
+ ) -> Rotation:
409
+ """
410
+ Reverse pointwise multiplication of the rotation with a tensor.
411
+
412
+ Args:
413
+ left:
414
+ The left multiplicand
415
+ Returns:
416
+ The product
417
+ """
418
+ return self.__mul__(left)
419
+
420
+ # Properties
421
+
422
+ @property
423
+ def shape(self) -> torch.Size:
424
+ """
425
+ Returns the virtual shape of the rotation object. This shape is
426
+ defined as the batch dimensions of the underlying rotation matrix
427
+ or quaternion. If the Rotation was initialized with a [10, 3, 3]
428
+ rotation matrix tensor, for example, the resulting shape would be
429
+ [10].
430
+
431
+ Returns:
432
+ The virtual shape of the rotation object
433
+ """
434
+ s = None
435
+ if(self._quats is not None):
436
+ s = self._quats.shape[:-1]
437
+ else:
438
+ s = self._rot_mats.shape[:-2]
439
+
440
+ return s
441
+
442
+ @property
443
+ def dtype(self) -> torch.dtype:
444
+ """
445
+ Returns the dtype of the underlying rotation.
446
+
447
+ Returns:
448
+ The dtype of the underlying rotation
449
+ """
450
+ if(self._rot_mats is not None):
451
+ return self._rot_mats.dtype
452
+ elif(self._quats is not None):
453
+ return self._quats.dtype
454
+ else:
455
+ raise ValueError("Both rotations are None")
456
+
457
+ @property
458
+ def device(self) -> torch.device:
459
+ """
460
+ The device of the underlying rotation
461
+
462
+ Returns:
463
+ The device of the underlying rotation
464
+ """
465
+ if(self._rot_mats is not None):
466
+ return self._rot_mats.device
467
+ elif(self._quats is not None):
468
+ return self._quats.device
469
+ else:
470
+ raise ValueError("Both rotations are None")
471
+
472
+ @property
473
+ def requires_grad(self) -> bool:
474
+ """
475
+ Returns the requires_grad property of the underlying rotation
476
+
477
+ Returns:
478
+ The requires_grad property of the underlying tensor
479
+ """
480
+ if(self._rot_mats is not None):
481
+ return self._rot_mats.requires_grad
482
+ elif(self._quats is not None):
483
+ return self._quats.requires_grad
484
+ else:
485
+ raise ValueError("Both rotations are None")
486
+
487
+ def get_rot_mats(self) -> torch.Tensor:
488
+ """
489
+ Returns the underlying rotation as a rotation matrix tensor.
490
+
491
+ Returns:
492
+ The rotation as a rotation matrix tensor
493
+ """
494
+ rot_mats = self._rot_mats
495
+ if(rot_mats is None):
496
+ if(self._quats is None):
497
+ raise ValueError("Both rotations are None")
498
+ else:
499
+ rot_mats = quat_to_rot(self._quats)
500
+
501
+ return rot_mats
502
+
503
+ def get_quats(self) -> torch.Tensor:
504
+ """
505
+ Returns the underlying rotation as a quaternion tensor.
506
+
507
+ Depending on whether the Rotation was initialized with a
508
+ quaternion, this function may call torch.linalg.eigh.
509
+
510
+ Returns:
511
+ The rotation as a quaternion tensor.
512
+ """
513
+ quats = self._quats
514
+ if(quats is None):
515
+ if(self._rot_mats is None):
516
+ raise ValueError("Both rotations are None")
517
+ else:
518
+ quats = rot_to_quat(self._rot_mats)
519
+
520
+ return quats
521
+
522
+ def get_cur_rot(self) -> torch.Tensor:
523
+ """
524
+ Return the underlying rotation in its current form
525
+
526
+ Returns:
527
+ The stored rotation
528
+ """
529
+ if(self._rot_mats is not None):
530
+ return self._rot_mats
531
+ elif(self._quats is not None):
532
+ return self._quats
533
+ else:
534
+ raise ValueError("Both rotations are None")
535
+
536
+ # Rotation functions
537
+
538
+ def compose_q_update_vec(self,
539
+ q_update_vec: torch.Tensor,
540
+ normalize_quats: bool = True
541
+ ) -> Rotation:
542
+ """
543
+ Returns a new quaternion Rotation after updating the current
544
+ object's underlying rotation with a quaternion update, formatted
545
+ as a [*, 3] tensor whose final three columns represent x, y, z such
546
+ that (1, x, y, z) is the desired (not necessarily unit) quaternion
547
+ update.
548
+
549
+ Args:
550
+ q_update_vec:
551
+ A [*, 3] quaternion update tensor
552
+ normalize_quats:
553
+ Whether to normalize the output quaternion
554
+ Returns:
555
+ An updated Rotation
556
+ """
557
+ quats = self.get_quats()
558
+ new_quats = quats + quat_multiply_by_vec(quats, q_update_vec)
559
+ return Rotation(
560
+ rot_mats=None,
561
+ quats=new_quats,
562
+ normalize_quats=normalize_quats,
563
+ )
564
+
565
+ def compose_r(self, r: Rotation) -> Rotation:
566
+ """
567
+ Compose the rotation matrices of the current Rotation object with
568
+ those of another.
569
+
570
+ Args:
571
+ r:
572
+ An update rotation object
573
+ Returns:
574
+ An updated rotation object
575
+ """
576
+ r1 = self.get_rot_mats()
577
+ r2 = r.get_rot_mats()
578
+ new_rot_mats = rot_matmul(r1, r2)
579
+ return Rotation(rot_mats=new_rot_mats, quats=None)
580
+
581
+ def compose_q(self, r: Rotation, normalize_quats: bool = True) -> Rotation:
582
+ """
583
+ Compose the quaternions of the current Rotation object with those
584
+ of another.
585
+
586
+ Depending on whether either Rotation was initialized with
587
+ quaternions, this function may call torch.linalg.eigh.
588
+
589
+ Args:
590
+ r:
591
+ An update rotation object
592
+ Returns:
593
+ An updated rotation object
594
+ """
595
+ q1 = self.get_quats()
596
+ q2 = r.get_quats()
597
+ new_quats = quat_multiply(q1, q2)
598
+ return Rotation(
599
+ rot_mats=None, quats=new_quats, normalize_quats=normalize_quats
600
+ )
601
+
602
+ def apply(self, pts: torch.Tensor) -> torch.Tensor:
603
+ """
604
+ Apply the current Rotation as a rotation matrix to a set of 3D
605
+ coordinates.
606
+
607
+ Args:
608
+ pts:
609
+ A [*, 3] set of points
610
+ Returns:
611
+ [*, 3] rotated points
612
+ """
613
+ rot_mats = self.get_rot_mats()
614
+ return rot_vec_mul(rot_mats, pts)
615
+
616
+ def invert_apply(self, pts: torch.Tensor) -> torch.Tensor:
617
+ """
618
+ The inverse of the apply() method.
619
+
620
+ Args:
621
+ pts:
622
+ A [*, 3] set of points
623
+ Returns:
624
+ [*, 3] inverse-rotated points
625
+ """
626
+ rot_mats = self.get_rot_mats()
627
+ inv_rot_mats = invert_rot_mat(rot_mats)
628
+ return rot_vec_mul(inv_rot_mats, pts)
629
+
630
+ def invert(self) -> Rotation:
631
+ """
632
+ Returns the inverse of the current Rotation.
633
+
634
+ Returns:
635
+ The inverse of the current Rotation
636
+ """
637
+ if(self._rot_mats is not None):
638
+ return Rotation(
639
+ rot_mats=invert_rot_mat(self._rot_mats),
640
+ quats=None
641
+ )
642
+ elif(self._quats is not None):
643
+ return Rotation(
644
+ rot_mats=None,
645
+ quats=invert_quat(self._quats),
646
+ normalize_quats=False,
647
+ )
648
+ else:
649
+ raise ValueError("Both rotations are None")
650
+
651
+ # "Tensor" stuff
652
+
653
+ def unsqueeze(self,
654
+ dim: int,
655
+ ) -> Rigid:
656
+ """
657
+ Analogous to torch.unsqueeze. The dimension is relative to the
658
+ shape of the Rotation object.
659
+
660
+ Args:
661
+ dim: A positive or negative dimension index.
662
+ Returns:
663
+ The unsqueezed Rotation.
664
+ """
665
+ if dim >= len(self.shape):
666
+ raise ValueError("Invalid dimension")
667
+
668
+ if(self._rot_mats is not None):
669
+ rot_mats = self._rot_mats.unsqueeze(dim if dim >= 0 else dim - 2)
670
+ return Rotation(rot_mats=rot_mats, quats=None)
671
+ elif(self._quats is not None):
672
+ quats = self._quats.unsqueeze(dim if dim >= 0 else dim - 1)
673
+ return Rotation(rot_mats=None, quats=quats, normalize_quats=False)
674
+ else:
675
+ raise ValueError("Both rotations are None")
676
+
677
+ @staticmethod
678
+ def cat(
679
+ rs: Sequence[Rotation],
680
+ dim: int,
681
+ ) -> Rigid:
682
+ """
683
+ Concatenates rotations along one of the batch dimensions. Analogous
684
+ to torch.cat().
685
+
686
+ Note that the output of this operation is always a rotation matrix,
687
+ regardless of the format of input rotations.
688
+
689
+ Args:
690
+ rs:
691
+ A list of rotation objects
692
+ dim:
693
+ The dimension along which the rotations should be
694
+ concatenated
695
+ Returns:
696
+ A concatenated Rotation object in rotation matrix format
697
+ """
698
+ rot_mats = [r.get_rot_mats() for r in rs]
699
+ rot_mats = torch.cat(rot_mats, dim=dim if dim >= 0 else dim - 2)
700
+
701
+ return Rotation(rot_mats=rot_mats, quats=None)
702
+
703
+ def map_tensor_fn(self,
704
+ fn: Callable[torch.Tensor, torch.Tensor]
705
+ ) -> Rotation:
706
+ """
707
+ Apply a Tensor -> Tensor function to underlying rotation tensors,
708
+ mapping over the rotation dimension(s). Can be used e.g. to sum out
709
+ a one-hot batch dimension.
710
+
711
+ Args:
712
+ fn:
713
+ A Tensor -> Tensor function to be mapped over the Rotation
714
+ Returns:
715
+ The transformed Rotation object
716
+ """
717
+ if(self._rot_mats is not None):
718
+ rot_mats = self._rot_mats.view(self._rot_mats.shape[:-2] + (9,))
719
+ rot_mats = torch.stack(
720
+ list(map(fn, torch.unbind(rot_mats, dim=-1))), dim=-1
721
+ )
722
+ rot_mats = rot_mats.view(rot_mats.shape[:-1] + (3, 3))
723
+ return Rotation(rot_mats=rot_mats, quats=None)
724
+ elif(self._quats is not None):
725
+ quats = torch.stack(
726
+ list(map(fn, torch.unbind(self._quats, dim=-1))), dim=-1
727
+ )
728
+ return Rotation(rot_mats=None, quats=quats, normalize_quats=False)
729
+ else:
730
+ raise ValueError("Both rotations are None")
731
+
732
+ def cuda(self) -> Rotation:
733
+ """
734
+ Analogous to the cuda() method of torch Tensors
735
+
736
+ Returns:
737
+ A copy of the Rotation in CUDA memory
738
+ """
739
+ if(self._rot_mats is not None):
740
+ return Rotation(rot_mats=self._rot_mats.cuda(), quats=None)
741
+ elif(self._quats is not None):
742
+ return Rotation(
743
+ rot_mats=None,
744
+ quats=self._quats.cuda(),
745
+ normalize_quats=False
746
+ )
747
+ else:
748
+ raise ValueError("Both rotations are None")
749
+
750
+ def to(self,
751
+ device: Optional[torch.device],
752
+ dtype: Optional[torch.dtype]
753
+ ) -> Rotation:
754
+ """
755
+ Analogous to the to() method of torch Tensors
756
+
757
+ Args:
758
+ device:
759
+ A torch device
760
+ dtype:
761
+ A torch dtype
762
+ Returns:
763
+ A copy of the Rotation using the new device and dtype
764
+ """
765
+ if(self._rot_mats is not None):
766
+ return Rotation(
767
+ rot_mats=self._rot_mats.to(device=device, dtype=dtype),
768
+ quats=None,
769
+ )
770
+ elif(self._quats is not None):
771
+ return Rotation(
772
+ rot_mats=None,
773
+ quats=self._quats.to(device=device, dtype=dtype),
774
+ normalize_quats=False,
775
+ )
776
+ else:
777
+ raise ValueError("Both rotations are None")
778
+
779
+ def detach(self) -> Rotation:
780
+ """
781
+ Returns a copy of the Rotation whose underlying Tensor has been
782
+ detached from its torch graph.
783
+
784
+ Returns:
785
+ A copy of the Rotation whose underlying Tensor has been detached
786
+ from its torch graph
787
+ """
788
+ if(self._rot_mats is not None):
789
+ return Rotation(rot_mats=self._rot_mats.detach(), quats=None)
790
+ elif(self._quats is not None):
791
+ return Rotation(
792
+ rot_mats=None,
793
+ quats=self._quats.detach(),
794
+ normalize_quats=False,
795
+ )
796
+ else:
797
+ raise ValueError("Both rotations are None")
798
+
799
+
800
+ class Rigid:
801
+ """
802
+ A class representing a rigid transformation. Little more than a wrapper
803
+ around two objects: a Rotation object and a [*, 3] translation
804
+ Designed to behave approximately like a single torch tensor with the
805
+ shape of the shared batch dimensions of its component parts.
806
+ """
807
+ def __init__(self,
808
+ rots: Optional[Rotation],
809
+ trans: Optional[torch.Tensor],
810
+ ):
811
+ """
812
+ Args:
813
+ rots: A [*, 3, 3] rotation tensor
814
+ trans: A corresponding [*, 3] translation tensor
815
+ """
816
+ # (we need device, dtype, etc. from at least one input)
817
+
818
+ batch_dims, dtype, device, requires_grad = None, None, None, None
819
+ if(trans is not None):
820
+ batch_dims = trans.shape[:-1]
821
+ dtype = trans.dtype
822
+ device = trans.device
823
+ requires_grad = trans.requires_grad
824
+ elif(rots is not None):
825
+ batch_dims = rots.shape
826
+ dtype = rots.dtype
827
+ device = rots.device
828
+ requires_grad = rots.requires_grad
829
+ else:
830
+ raise ValueError("At least one input argument must be specified")
831
+
832
+ if(rots is None):
833
+ rots = Rotation.identity(
834
+ batch_dims, dtype, device, requires_grad,
835
+ )
836
+ elif(trans is None):
837
+ trans = identity_trans(
838
+ batch_dims, dtype, device, requires_grad,
839
+ )
840
+
841
+ if((rots.shape != trans.shape[:-1]) or
842
+ (rots.device != trans.device)):
843
+ raise ValueError("Rots and trans incompatible")
844
+
845
+ # Force full precision. Happens to the rotations automatically.
846
+ trans = trans.to(dtype=torch.float32)
847
+
848
+ self._rots = rots
849
+ self._trans = trans
850
+
851
+ @staticmethod
852
+ def identity(
853
+ shape: Tuple[int],
854
+ dtype: Optional[torch.dtype] = None,
855
+ device: Optional[torch.device] = None,
856
+ requires_grad: bool = True,
857
+ fmt: str = "quat",
858
+ ) -> Rigid:
859
+ """
860
+ Constructs an identity transformation.
861
+
862
+ Args:
863
+ shape:
864
+ The desired shape
865
+ dtype:
866
+ The dtype of both internal tensors
867
+ device:
868
+ The device of both internal tensors
869
+ requires_grad:
870
+ Whether grad should be enabled for the internal tensors
871
+ Returns:
872
+ The identity transformation
873
+ """
874
+ return Rigid(
875
+ Rotation.identity(shape, dtype, device, requires_grad, fmt=fmt),
876
+ identity_trans(shape, dtype, device, requires_grad),
877
+ )
878
+
879
+ def __getitem__(self,
880
+ index: Any,
881
+ ) -> Rigid:
882
+ """
883
+ Indexes the affine transformation with PyTorch-style indices.
884
+ The index is applied to the shared dimensions of both the rotation
885
+ and the translation.
886
+
887
+ E.g.::
888
+
889
+ r = Rotation(rot_mats=torch.rand(10, 10, 3, 3), quats=None)
890
+ t = Rigid(r, torch.rand(10, 10, 3))
891
+ indexed = t[3, 4:6]
892
+ assert(indexed.shape == (2,))
893
+ assert(indexed.get_rots().shape == (2,))
894
+ assert(indexed.get_trans().shape == (2, 3))
895
+
896
+ Args:
897
+ index: A standard torch tensor index. E.g. 8, (10, None, 3),
898
+ or (3, slice(0, 1, None))
899
+ Returns:
900
+ The indexed tensor
901
+ """
902
+ if type(index) != tuple:
903
+ index = (index,)
904
+
905
+ return Rigid(
906
+ self._rots[index],
907
+ self._trans[index + (slice(None),)],
908
+ )
909
+
910
+ def __mul__(self,
911
+ right: torch.Tensor,
912
+ ) -> Rigid:
913
+ """
914
+ Pointwise left multiplication of the transformation with a tensor.
915
+ Can be used to e.g. mask the Rigid.
916
+
917
+ Args:
918
+ right:
919
+ The tensor multiplicand
920
+ Returns:
921
+ The product
922
+ """
923
+ if not(isinstance(right, torch.Tensor)):
924
+ raise TypeError("The other multiplicand must be a Tensor")
925
+
926
+ new_rots = self._rots * right
927
+ new_trans = self._trans * right[..., None]
928
+
929
+ return Rigid(new_rots, new_trans)
930
+
931
+ def __rmul__(self,
932
+ left: torch.Tensor,
933
+ ) -> Rigid:
934
+ """
935
+ Reverse pointwise multiplication of the transformation with a
936
+ tensor.
937
+
938
+ Args:
939
+ left:
940
+ The left multiplicand
941
+ Returns:
942
+ The product
943
+ """
944
+ return self.__mul__(left)
945
+
946
+ @property
947
+ def shape(self) -> torch.Size:
948
+ """
949
+ Returns the shape of the shared dimensions of the rotation and
950
+ the translation.
951
+
952
+ Returns:
953
+ The shape of the transformation
954
+ """
955
+ s = self._trans.shape[:-1]
956
+ return s
957
+
958
+ @property
959
+ def device(self) -> torch.device:
960
+ """
961
+ Returns the device on which the Rigid's tensors are located.
962
+
963
+ Returns:
964
+ The device on which the Rigid's tensors are located
965
+ """
966
+ return self._trans.device
967
+
968
+ def get_rots(self) -> Rotation:
969
+ """
970
+ Getter for the rotation.
971
+
972
+ Returns:
973
+ The rotation object
974
+ """
975
+ return self._rots
976
+
977
+ def get_trans(self) -> torch.Tensor:
978
+ """
979
+ Getter for the translation.
980
+
981
+ Returns:
982
+ The stored translation
983
+ """
984
+ return self._trans
985
+
986
+ def compose_q_update_vec(self,
987
+ q_update_vec: torch.Tensor,
988
+ ) -> Rigid:
989
+ """
990
+ Composes the transformation with a quaternion update vector of
991
+ shape [*, 6], where the final 6 columns represent the x, y, and
992
+ z values of a quaternion of form (1, x, y, z) followed by a 3D
993
+ translation.
994
+
995
+ Args:
996
+ q_vec: The quaternion update vector.
997
+ Returns:
998
+ The composed transformation.
999
+ """
1000
+ q_vec, t_vec = q_update_vec[..., :3], q_update_vec[..., 3:]
1001
+ new_rots = self._rots.compose_q_update_vec(q_vec)
1002
+
1003
+ trans_update = self._rots.apply(t_vec)
1004
+ new_translation = self._trans + trans_update
1005
+
1006
+ return Rigid(new_rots, new_translation)
1007
+
1008
+ def compose(self,
1009
+ r: Rigid,
1010
+ ) -> Rigid:
1011
+ """
1012
+ Composes the current rigid object with another.
1013
+
1014
+ Args:
1015
+ r:
1016
+ Another Rigid object
1017
+ Returns:
1018
+ The composition of the two transformations
1019
+ """
1020
+ new_rot = self._rots.compose_r(r._rots)
1021
+ new_trans = self._rots.apply(r._trans) + self._trans
1022
+ return Rigid(new_rot, new_trans)
1023
+
1024
+ def apply(self,
1025
+ pts: torch.Tensor,
1026
+ ) -> torch.Tensor:
1027
+ """
1028
+ Applies the transformation to a coordinate tensor.
1029
+
1030
+ Args:
1031
+ pts: A [*, 3] coordinate tensor.
1032
+ Returns:
1033
+ The transformed points.
1034
+ """
1035
+ rotated = self._rots.apply(pts)
1036
+ return rotated + self._trans
1037
+
1038
+ def invert_apply(self,
1039
+ pts: torch.Tensor
1040
+ ) -> torch.Tensor:
1041
+ """
1042
+ Applies the inverse of the transformation to a coordinate tensor.
1043
+
1044
+ Args:
1045
+ pts: A [*, 3] coordinate tensor
1046
+ Returns:
1047
+ The transformed points.
1048
+ """
1049
+ pts = pts - self._trans
1050
+ return self._rots.invert_apply(pts)
1051
+
1052
+ def invert(self) -> Rigid:
1053
+ """
1054
+ Inverts the transformation.
1055
+
1056
+ Returns:
1057
+ The inverse transformation.
1058
+ """
1059
+ rot_inv = self._rots.invert()
1060
+ trn_inv = rot_inv.apply(self._trans)
1061
+
1062
+ return Rigid(rot_inv, -1 * trn_inv)
1063
+
1064
+ def map_tensor_fn(self,
1065
+ fn: Callable[torch.Tensor, torch.Tensor]
1066
+ ) -> Rigid:
1067
+ """
1068
+ Apply a Tensor -> Tensor function to underlying translation and
1069
+ rotation tensors, mapping over the translation/rotation dimensions
1070
+ respectively.
1071
+
1072
+ Args:
1073
+ fn:
1074
+ A Tensor -> Tensor function to be mapped over the Rigid
1075
+ Returns:
1076
+ The transformed Rigid object
1077
+ """
1078
+ new_rots = self._rots.map_tensor_fn(fn)
1079
+ new_trans = torch.stack(
1080
+ list(map(fn, torch.unbind(self._trans, dim=-1))),
1081
+ dim=-1
1082
+ )
1083
+
1084
+ return Rigid(new_rots, new_trans)
1085
+
1086
+ def to_tensor_4x4(self) -> torch.Tensor:
1087
+ """
1088
+ Converts a transformation to a homogenous transformation tensor.
1089
+
1090
+ Returns:
1091
+ A [*, 4, 4] homogenous transformation tensor
1092
+ """
1093
+ tensor = self._trans.new_zeros((*self.shape, 4, 4))
1094
+ tensor[..., :3, :3] = self._rots.get_rot_mats()
1095
+ tensor[..., :3, 3] = self._trans
1096
+ tensor[..., 3, 3] = 1
1097
+ return tensor
1098
+
1099
+ @staticmethod
1100
+ def from_tensor_4x4(
1101
+ t: torch.Tensor
1102
+ ) -> Rigid:
1103
+ """
1104
+ Constructs a transformation from a homogenous transformation
1105
+ tensor.
1106
+
1107
+ Args:
1108
+ t: [*, 4, 4] homogenous transformation tensor
1109
+ Returns:
1110
+ T object with shape [*]
1111
+ """
1112
+ if(t.shape[-2:] != (4, 4)):
1113
+ raise ValueError("Incorrectly shaped input tensor")
1114
+
1115
+ rots = Rotation(rot_mats=t[..., :3, :3], quats=None)
1116
+ trans = t[..., :3, 3]
1117
+
1118
+ return Rigid(rots, trans)
1119
+
1120
+ def to_tensor_7(self) -> torch.Tensor:
1121
+ """
1122
+ Converts a transformation to a tensor with 7 final columns, four
1123
+ for the quaternion followed by three for the translation.
1124
+
1125
+ Returns:
1126
+ A [*, 7] tensor representation of the transformation
1127
+ """
1128
+ tensor = self._trans.new_zeros((*self.shape, 7))
1129
+ tensor[..., :4] = self._rots.get_quats()
1130
+ tensor[..., 4:] = self._trans
1131
+
1132
+ return tensor
1133
+
1134
+ @staticmethod
1135
+ def from_tensor_7(
1136
+ t: torch.Tensor,
1137
+ normalize_quats: bool = False,
1138
+ ) -> Rigid:
1139
+ if(t.shape[-1] != 7):
1140
+ raise ValueError("Incorrectly shaped input tensor")
1141
+
1142
+ quats, trans = t[..., :4], t[..., 4:]
1143
+
1144
+ rots = Rotation(
1145
+ rot_mats=None,
1146
+ quats=quats,
1147
+ normalize_quats=normalize_quats
1148
+ )
1149
+
1150
+ return Rigid(rots, trans)
1151
+
1152
+ @staticmethod
1153
+ def from_3_points(
1154
+ p_neg_x_axis: torch.Tensor,
1155
+ origin: torch.Tensor,
1156
+ p_xy_plane: torch.Tensor,
1157
+ eps: float = 1e-8
1158
+ ) -> Rigid:
1159
+ """
1160
+ Implements algorithm 21. Constructs transformations from sets of 3
1161
+ points using the Gram-Schmidt algorithm.
1162
+
1163
+ Args:
1164
+ p_neg_x_axis: [*, 3] coordinates
1165
+ origin: [*, 3] coordinates used as frame origins
1166
+ p_xy_plane: [*, 3] coordinates
1167
+ eps: Small epsilon value
1168
+ Returns:
1169
+ A transformation object of shape [*]
1170
+ """
1171
+ p_neg_x_axis = torch.unbind(p_neg_x_axis, dim=-1)
1172
+ origin = torch.unbind(origin, dim=-1)
1173
+ p_xy_plane = torch.unbind(p_xy_plane, dim=-1)
1174
+
1175
+ e0 = [c1 - c2 for c1, c2 in zip(origin, p_neg_x_axis)]
1176
+ e1 = [c1 - c2 for c1, c2 in zip(p_xy_plane, origin)]
1177
+
1178
+ denom = torch.sqrt(sum((c * c for c in e0)) + eps)
1179
+ e0 = [c / denom for c in e0]
1180
+ dot = sum((c1 * c2 for c1, c2 in zip(e0, e1)))
1181
+ e1 = [c2 - c1 * dot for c1, c2 in zip(e0, e1)]
1182
+ denom = torch.sqrt(sum((c * c for c in e1)) + eps)
1183
+ e1 = [c / denom for c in e1]
1184
+ e2 = [
1185
+ e0[1] * e1[2] - e0[2] * e1[1],
1186
+ e0[2] * e1[0] - e0[0] * e1[2],
1187
+ e0[0] * e1[1] - e0[1] * e1[0],
1188
+ ]
1189
+
1190
+ rots = torch.stack([c for tup in zip(e0, e1, e2) for c in tup], dim=-1)
1191
+ rots = rots.reshape(rots.shape[:-1] + (3, 3))
1192
+
1193
+ rot_obj = Rotation(rot_mats=rots, quats=None)
1194
+
1195
+ return Rigid(rot_obj, torch.stack(origin, dim=-1))
1196
+
1197
+ def unsqueeze(self,
1198
+ dim: int,
1199
+ ) -> Rigid:
1200
+ """
1201
+ Analogous to torch.unsqueeze. The dimension is relative to the
1202
+ shared dimensions of the rotation/translation.
1203
+
1204
+ Args:
1205
+ dim: A positive or negative dimension index.
1206
+ Returns:
1207
+ The unsqueezed transformation.
1208
+ """
1209
+ if dim >= len(self.shape):
1210
+ raise ValueError("Invalid dimension")
1211
+ rots = self._rots.unsqueeze(dim)
1212
+ trans = self._trans.unsqueeze(dim if dim >= 0 else dim - 1)
1213
+
1214
+ return Rigid(rots, trans)
1215
+
1216
+ @staticmethod
1217
+ def cat(
1218
+ ts: Sequence[Rigid],
1219
+ dim: int,
1220
+ ) -> Rigid:
1221
+ """
1222
+ Concatenates transformations along a new dimension.
1223
+
1224
+ Args:
1225
+ ts:
1226
+ A list of T objects
1227
+ dim:
1228
+ The dimension along which the transformations should be
1229
+ concatenated
1230
+ Returns:
1231
+ A concatenated transformation object
1232
+ """
1233
+ rots = Rotation.cat([t._rots for t in ts], dim)
1234
+ trans = torch.cat(
1235
+ [t._trans for t in ts], dim=dim if dim >= 0 else dim - 1
1236
+ )
1237
+
1238
+ return Rigid(rots, trans)
1239
+
1240
+ def apply_rot_fn(self, fn: Callable[Rotation, Rotation]) -> Rigid:
1241
+ """
1242
+ Applies a Rotation -> Rotation function to the stored rotation
1243
+ object.
1244
+
1245
+ Args:
1246
+ fn: A function of type Rotation -> Rotation
1247
+ Returns:
1248
+ A transformation object with a transformed rotation.
1249
+ """
1250
+ return Rigid(fn(self._rots), self._trans)
1251
+
1252
+ def apply_trans_fn(self, fn: Callable[torch.Tensor, torch.Tensor]) -> Rigid:
1253
+ """
1254
+ Applies a Tensor -> Tensor function to the stored translation.
1255
+
1256
+ Args:
1257
+ fn:
1258
+ A function of type Tensor -> Tensor to be applied to the
1259
+ translation
1260
+ Returns:
1261
+ A transformation object with a transformed translation.
1262
+ """
1263
+ return Rigid(self._rots, fn(self._trans))
1264
+
1265
+ def scale_translation(self, trans_scale_factor: float) -> Rigid:
1266
+ """
1267
+ Scales the translation by a constant factor.
1268
+
1269
+ Args:
1270
+ trans_scale_factor:
1271
+ The constant factor
1272
+ Returns:
1273
+ A transformation object with a scaled translation.
1274
+ """
1275
+ fn = lambda t: t * trans_scale_factor
1276
+ return self.apply_trans_fn(fn)
1277
+
1278
+ def stop_rot_gradient(self) -> Rigid:
1279
+ """
1280
+ Detaches the underlying rotation object
1281
+
1282
+ Returns:
1283
+ A transformation object with detached rotations
1284
+ """
1285
+ fn = lambda r: r.detach()
1286
+ return self.apply_rot_fn(fn)
1287
+
1288
+ @staticmethod
1289
+ def make_transform_from_reference(n_xyz, ca_xyz, c_xyz, eps=1e-20):
1290
+ """
1291
+ Returns a transformation object from reference coordinates.
1292
+
1293
+ Note that this method does not take care of symmetries. If you
1294
+ provide the atom positions in the non-standard way, the N atom will
1295
+ end up not at [-0.527250, 1.359329, 0.0] but instead at
1296
+ [-0.527250, -1.359329, 0.0]. You need to take care of such cases in
1297
+ your code.
1298
+
1299
+ Args:
1300
+ n_xyz: A [*, 3] tensor of nitrogen xyz coordinates.
1301
+ ca_xyz: A [*, 3] tensor of carbon alpha xyz coordinates.
1302
+ c_xyz: A [*, 3] tensor of carbon xyz coordinates.
1303
+ Returns:
1304
+ A transformation object. After applying the translation and
1305
+ rotation to the reference backbone, the coordinates will
1306
+ approximately equal to the input coordinates.
1307
+ """
1308
+ translation = -1 * ca_xyz
1309
+ n_xyz = n_xyz + translation
1310
+ c_xyz = c_xyz + translation
1311
+
1312
+ c_x, c_y, c_z = [c_xyz[..., i] for i in range(3)]
1313
+ norm = torch.sqrt(eps + c_x ** 2 + c_y ** 2)
1314
+ sin_c1 = -c_y / norm
1315
+ cos_c1 = c_x / norm
1316
+ zeros = sin_c1.new_zeros(sin_c1.shape)
1317
+ ones = sin_c1.new_ones(sin_c1.shape)
1318
+
1319
+ c1_rots = sin_c1.new_zeros((*sin_c1.shape, 3, 3))
1320
+ c1_rots[..., 0, 0] = cos_c1
1321
+ c1_rots[..., 0, 1] = -1 * sin_c1
1322
+ c1_rots[..., 1, 0] = sin_c1
1323
+ c1_rots[..., 1, 1] = cos_c1
1324
+ c1_rots[..., 2, 2] = 1
1325
+
1326
+ norm = torch.sqrt(eps + c_x ** 2 + c_y ** 2 + c_z ** 2)
1327
+ sin_c2 = c_z / norm
1328
+ cos_c2 = torch.sqrt(c_x ** 2 + c_y ** 2) / norm
1329
+
1330
+ c2_rots = sin_c2.new_zeros((*sin_c2.shape, 3, 3))
1331
+ c2_rots[..., 0, 0] = cos_c2
1332
+ c2_rots[..., 0, 2] = sin_c2
1333
+ c2_rots[..., 1, 1] = 1
1334
+ c2_rots[..., 2, 0] = -1 * sin_c2
1335
+ c2_rots[..., 2, 2] = cos_c2
1336
+
1337
+ c_rots = rot_matmul(c2_rots, c1_rots)
1338
+ n_xyz = rot_vec_mul(c_rots, n_xyz)
1339
+
1340
+ _, n_y, n_z = [n_xyz[..., i] for i in range(3)]
1341
+ norm = torch.sqrt(eps + n_y ** 2 + n_z ** 2)
1342
+ sin_n = -n_z / norm
1343
+ cos_n = n_y / norm
1344
+
1345
+ n_rots = sin_c2.new_zeros((*sin_c2.shape, 3, 3))
1346
+ n_rots[..., 0, 0] = 1
1347
+ n_rots[..., 1, 1] = cos_n
1348
+ n_rots[..., 1, 2] = -1 * sin_n
1349
+ n_rots[..., 2, 1] = sin_n
1350
+ n_rots[..., 2, 2] = cos_n
1351
+
1352
+ rots = rot_matmul(n_rots, c_rots)
1353
+
1354
+ rots = rots.transpose(-1, -2)
1355
+ translation = -1 * translation
1356
+
1357
+ rot_obj = Rotation(rot_mats=rots, quats=None)
1358
+
1359
+ return Rigid(rot_obj, translation)
1360
+
1361
+ def cuda(self) -> Rigid:
1362
+ """
1363
+ Moves the transformation object to GPU memory
1364
+
1365
+ Returns:
1366
+ A version of the transformation on GPU
1367
+ """
1368
+ return Rigid(self._rots.cuda(), self._trans.cuda())
openfold/tensor_utils.py ADDED
@@ -0,0 +1,121 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2021 AlQuraishi Laboratory
2
+ # Copyright 2021 DeepMind Technologies Limited
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+
16
+ from functools import partial
17
+ import logging
18
+ from typing import Tuple, List, Callable, Any, Dict, Sequence, Optional
19
+
20
+ import torch
21
+ import torch.nn as nn
22
+
23
+
24
+ def add(m1, m2, inplace):
25
+ # The first operation in a checkpoint can't be in-place, but it's
26
+ # nice to have in-place addition during inference. Thus...
27
+ if(not inplace):
28
+ m1 = m1 + m2
29
+ else:
30
+ m1 += m2
31
+
32
+ return m1
33
+
34
+
35
+ def permute_final_dims(tensor: torch.Tensor, inds: List[int]):
36
+ zero_index = -1 * len(inds)
37
+ first_inds = list(range(len(tensor.shape[:zero_index])))
38
+ return tensor.permute(first_inds + [zero_index + i for i in inds])
39
+
40
+
41
+ def flatten_final_dims(t: torch.Tensor, no_dims: int):
42
+ return t.reshape(t.shape[:-no_dims] + (-1,))
43
+
44
+
45
+ def masked_mean(mask, value, dim, eps=1e-4):
46
+ mask = mask.expand(*value.shape)
47
+ return torch.sum(mask * value, dim=dim) / (eps + torch.sum(mask, dim=dim))
48
+
49
+
50
+ def pts_to_distogram(pts, min_bin=2.3125, max_bin=21.6875, no_bins=64):
51
+ boundaries = torch.linspace(
52
+ min_bin, max_bin, no_bins - 1, device=pts.device
53
+ )
54
+ dists = torch.sqrt(
55
+ torch.sum((pts.unsqueeze(-2) - pts.unsqueeze(-3)) ** 2, dim=-1)
56
+ )
57
+ return torch.bucketize(dists, boundaries)
58
+
59
+
60
+ def dict_multimap(fn, dicts):
61
+ first = dicts[0]
62
+ new_dict = {}
63
+ for k, v in first.items():
64
+ all_v = [d[k] for d in dicts]
65
+ if type(v) is dict:
66
+ new_dict[k] = dict_multimap(fn, all_v)
67
+ else:
68
+ new_dict[k] = fn(all_v)
69
+
70
+ return new_dict
71
+
72
+
73
+ def one_hot(x, v_bins):
74
+ reshaped_bins = v_bins.view(((1,) * len(x.shape)) + (len(v_bins),))
75
+ diffs = x[..., None] - reshaped_bins
76
+ am = torch.argmin(torch.abs(diffs), dim=-1)
77
+ return nn.functional.one_hot(am, num_classes=len(v_bins)).float()
78
+
79
+
80
+ def batched_gather(data, inds, dim=0, no_batch_dims=0):
81
+ ranges = []
82
+ for i, s in enumerate(data.shape[:no_batch_dims]):
83
+ r = torch.arange(s)
84
+ r = r.view(*(*((1,) * i), -1, *((1,) * (len(inds.shape) - i - 1))))
85
+ ranges.append(r)
86
+
87
+ remaining_dims = [
88
+ slice(None) for _ in range(len(data.shape) - no_batch_dims)
89
+ ]
90
+ remaining_dims[dim - no_batch_dims if dim >= 0 else dim] = inds
91
+ ranges.extend(remaining_dims)
92
+ return data[ranges]
93
+
94
+
95
+ # With tree_map, a poor man's JAX tree_map
96
+ def dict_map(fn, dic, leaf_type):
97
+ new_dict = {}
98
+ for k, v in dic.items():
99
+ if type(v) is dict:
100
+ new_dict[k] = dict_map(fn, v, leaf_type)
101
+ else:
102
+ new_dict[k] = tree_map(fn, v, leaf_type)
103
+
104
+ return new_dict
105
+
106
+
107
+ def tree_map(fn, tree, leaf_type):
108
+ if isinstance(tree, dict):
109
+ return dict_map(fn, tree, leaf_type)
110
+ elif isinstance(tree, list):
111
+ return [tree_map(fn, x, leaf_type) for x in tree]
112
+ elif isinstance(tree, tuple):
113
+ return tuple([tree_map(fn, x, leaf_type) for x in tree])
114
+ elif isinstance(tree, leaf_type):
115
+ return fn(tree)
116
+ else:
117
+ print(type(tree))
118
+ raise ValueError("Not supported")
119
+
120
+
121
+ tensor_tree_map = partial(tree_map, leaf_type=torch.Tensor)
parse_complexes.py CHANGED
@@ -22,6 +22,12 @@ from rdkit.Chem.Descriptors import ExactMolWt
22
  import numpy as np
23
 
24
  import os
 
 
 
 
 
 
25
 
26
  # minimum molecular weight to consider sth a ligand
27
  mol_wt_cutoff = 100
@@ -35,9 +41,6 @@ punctuation_regex = r"""(\(|\)|\.|=|#|-|\+|\\|\/|:|~|@|\?|>>?|\*|\$|\%[0-9]{2}|
35
  # tokenization regex (Schwaller)
36
  molecule_regex = r"""(\[[^\]]+]|Br?|Cl?|N|O|S|P|F|I|b|c|n|o|s|p|\(|\)|\.|=|#|-|\+|\\|\/|:|~|@|\?|>>?|\*|\$|\%[0-9]{2}|[0-9])"""
37
 
38
- max_seq = 2046 # = 2048 - 2 (accounting for [CLS] and [SEP])
39
- max_smiles = 510 # = 512 - 2
40
-
41
  # filter out these common additives which occur in more than 75 complexes in the PDB
42
  ubiquitous_ligands = ['PEG', 'ADP', 'FAD', 'NAD', 'ATP', 'MPD', 'NAP', 'GDP', 'MES',
43
  'GTP', 'FMN', 'HEC', 'TRS', 'CIT', 'PGE', 'ANP', 'SAH', 'NDP',
@@ -60,11 +63,36 @@ ubiquitous_ligands = ['PEG', 'ADP', 'FAD', 'NAD', 'ATP', 'MPD', 'NAP', 'GDP', 'M
60
  'PQQ', '9TY', 'DUR', 'PPV', 'SPM', 'SIA', 'DUP', 'GTX', '1PG',
61
  'GUN', 'ETF', 'FDP', 'MFU', 'G2P', 'PC', 'DST', 'INI']
62
 
63
- def get_protein_sequence_and_coords(receptor):
64
- calpha = receptor.select('calpha')
65
- xyz = calpha.getCoords()
66
- seq = calpha.getSequence()
67
- return seq, xyz.tolist()
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
68
 
69
  def tokenize_ligand(mol):
70
  # convert to SMILES and map atoms
@@ -91,14 +119,17 @@ def tokenize_ligand(mol):
91
  k = 0
92
  conf_2d = AllChem.Compute2DCoords(mol)
93
  token_pos_2d = []
 
94
  for i,token in enumerate(masked_tokens):
95
  if token != '':
96
  token_pos_2d.append(tuple(mol.GetConformer(conf_2d).GetAtomPosition(atom_order[k])))
 
97
  k += 1
98
  else:
99
  token_pos_2d.append((0.,0.,0.))
 
100
 
101
- return smi, token_pos, token_pos_2d
102
 
103
  def read_ligand_expo():
104
  """
@@ -109,9 +140,10 @@ def read_ligand_expo():
109
  """
110
  file_name = "Components-smiles-stereo-oe.smi"
111
  try:
112
- df = pd.read_csv(file_name, sep="\t",
113
  header=None,
114
- names=["SMILES", "ID", "Name"])
 
115
  except FileNotFoundError:
116
  url = f"http://ligand-expo.rcsb.org/dictionaries/{file_name}"
117
  print(url)
@@ -119,9 +151,9 @@ def read_ligand_expo():
119
  open('Components-smiles-stereo-oe.smi', 'wb').write(r.content)
120
  df = pd.read_csv(file_name, sep="\t",
121
  header=None,
122
- names=["SMILES", "ID", "Name"])
123
- df.set_index("ID", inplace=True)
124
- return df.to_dict()
125
 
126
 
127
  def get_pdb_components(pdb_id):
@@ -138,7 +170,7 @@ def get_pdb_components(pdb_id):
138
  return protein, ligand
139
 
140
 
141
- def process_ligand(ligand, res_name, expo_dict):
142
  """
143
  Add bond orders to a pdb ligand
144
  1. Select the ligand component with name "res_name"
@@ -149,10 +181,10 @@ def process_ligand(ligand, res_name, expo_dict):
149
  6. Assign the bond orders from the template from step 3
150
  :param ligand: ligand as generated by prody
151
  :param res_name: residue name of ligand to extract
152
- :param expo_dict: dictionary with LigandExpo
153
  :return: molecule with bond orders assigned
154
  """
155
- sub_smiles = expo_dict['SMILES'][res_name]
156
  template = AllChem.MolFromSmiles(sub_smiles)
157
 
158
  allres = ligand.select(f"resname {res_name}")
@@ -167,12 +199,36 @@ def process_ligand(ligand, res_name, expo_dict):
167
  mols.append(AllChem.AssignBondOrdersFromTemplate(template, rd_mol))
168
  return mols, template
169
 
170
- def process_entry(df_dict, pdb_fn):
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
171
  try:
172
  """
173
  Slit pdb into protein and ligands,
174
  parse protein sequence and ligand tokens
175
- :param df_dict: ligand expo data
176
  :param pdb_fn: pdb entry file name
177
  :return:
178
  """
@@ -182,70 +238,103 @@ def process_entry(df_dict, pdb_fn):
182
 
183
  ligand_mols = []
184
  ligand_names = []
185
-
186
  if ligand is not None:
187
  # filter ligands by molecular weight
188
  res_name_list = list(set(ligand.getResnames()))
189
  for res in res_name_list:
190
- mols, template = process_ligand(ligand, res, df_dict)
 
 
191
 
192
  mol_wt = ExactMolWt(template)
193
  natoms = template.GetNumAtoms()
194
 
195
- if mol_wt >= mol_wt_cutoff and natoms >= min_atoms and res not in ubiquitous_ligands:
196
  # only use first copy of ligand
197
  mols = mols[:1]
198
  ligand_mols += mols
199
  ligand_names += [res]*len(mols)
200
 
 
 
 
 
 
201
 
202
  ligand_smiles = []
203
  ligand_xyz = []
204
  ligand_xyz_2d = []
205
-
206
- for mol, name in zip(ligand_mols, ligand_names):
207
  print('Processing {} and {}'.format(pdb_name, name))
208
- smi, xyz, xyz_2d = tokenize_ligand(mol)
209
  ligand_smiles.append(smi)
210
  ligand_xyz.append(xyz)
211
  ligand_xyz_2d.append(xyz_2d)
212
 
213
- seq, receptor_xyz = get_protein_sequence_and_coords(protein)
214
- return pdb_name, seq, receptor_xyz, ligand_names, ligand_smiles, ligand_xyz, ligand_xyz_2d
 
 
 
 
 
 
 
215
  except Exception as e:
 
216
  print(repr(e))
217
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
218
  if __name__ == '__main__':
219
  import glob
220
 
221
  filenames = glob.glob('pdb/*/*.gz')
222
  filenames = sorted(filenames)
 
 
 
 
 
 
 
 
223
  comm = MPI.COMM_WORLD
224
  with MPICommExecutor(comm, root=0) as executor:
225
  # with MPIPoolExecutor() as executor:
226
  if executor is not None:
227
  # read ligand table
228
- df_dict = read_ligand_expo()
 
 
 
 
 
229
 
230
- result = executor.map(partial(process_entry, df_dict), filenames, chunksize=512)
231
  result = list(result)
232
 
233
- # expand sequences and ligands
234
- pdb_id = [r[0] for r in result if r is not None for ligand in r[3]]
235
- seq = [r[1] for r in result if r is not None for ligand in r[3]]
236
- receptor_xyz = [r[2] for r in result if r is not None for ligand in r[3]]
237
- lig_id = [l for r in result if r is not None for l in r[3]]
238
- lig_smiles = [s for r in result if r is not None for s in r[4]]
239
- lig_xyz = [xyz for r in result if r is not None for xyz in r[5]]
240
- lig_xyz_2d = [xyz for r in result if r is not None for xyz in r[6]]
241
-
242
- import pandas as pd
243
- df = pd.DataFrame({
244
- 'pdb_id': pdb_id,
245
- 'lig_id': lig_id,
246
- 'seq': seq,
247
- 'smiles': lig_smiles,
248
- 'receptor_xyz': receptor_xyz,
249
- 'ligand_xyz': lig_xyz,
250
- 'ligand_xyz_2d': lig_xyz_2d})
251
- df.to_parquet('data/pdb.parquet',index=False)
 
22
  import numpy as np
23
 
24
  import os
25
+ import random
26
+ import traceback
27
+
28
+ from openfold import data_transforms, protein
29
+ from openfold.residue_constants import aatype_to_str_sequence
30
+ import torch
31
 
32
  # minimum molecular weight to consider sth a ligand
33
  mol_wt_cutoff = 100
 
41
  # tokenization regex (Schwaller)
42
  molecule_regex = r"""(\[[^\]]+]|Br?|Cl?|N|O|S|P|F|I|b|c|n|o|s|p|\(|\)|\.|=|#|-|\+|\\|\/|:|~|@|\?|>>?|\*|\$|\%[0-9]{2}|[0-9])"""
43
 
 
 
 
44
  # filter out these common additives which occur in more than 75 complexes in the PDB
45
  ubiquitous_ligands = ['PEG', 'ADP', 'FAD', 'NAD', 'ATP', 'MPD', 'NAP', 'GDP', 'MES',
46
  'GTP', 'FMN', 'HEC', 'TRS', 'CIT', 'PGE', 'ANP', 'SAH', 'NDP',
 
63
  'PQQ', '9TY', 'DUR', 'PPV', 'SPM', 'SIA', 'DUP', 'GTX', '1PG',
64
  'GUN', 'ETF', 'FDP', 'MFU', 'G2P', 'PC', 'DST', 'INI']
65
 
66
+ def get_protein_sequence_and_coords(receptor, pdb_str):
67
+ chains = [chain.getChid() for chain in receptor.getHierView()]
68
+
69
+ aatype = []
70
+ atom_positions = []
71
+ atom_mask = []
72
+ for chain in chains:
73
+ p = protein.from_pdb_string(pdb_str, chain)
74
+ aatype.append(p.aatype)
75
+ atom_positions.append(p.atom_positions)
76
+ atom_mask.append(p.atom_mask)
77
+
78
+ # concatenate chains
79
+ aatype = np.concatenate(aatype)
80
+ atom_positions = np.concatenate(atom_positions)
81
+ atom_mask = np.concatenate(atom_mask)
82
+
83
+ # determine torsion angles
84
+ features = {'aatype': torch.tensor(aatype),
85
+ 'all_atom_positions': torch.tensor(atom_positions),
86
+ 'all_atom_mask': torch.tensor(atom_mask)}
87
+ features = data_transforms.atom37_to_torsion_angles()(features)
88
+ features = data_transforms.atom37_to_frames(features)
89
+ features = data_transforms.make_atom14_masks(features)
90
+ features = data_transforms.make_atom14_positions(features)
91
+ features = {k: v.numpy() for k, v in features.items() if isinstance(v, torch.Tensor)}
92
+
93
+ seq = aatype_to_str_sequence(aatype)
94
+
95
+ return seq, features
96
 
97
  def tokenize_ligand(mol):
98
  # convert to SMILES and map atoms
 
119
  k = 0
120
  conf_2d = AllChem.Compute2DCoords(mol)
121
  token_pos_2d = []
122
+ atom_idx = []
123
  for i,token in enumerate(masked_tokens):
124
  if token != '':
125
  token_pos_2d.append(tuple(mol.GetConformer(conf_2d).GetAtomPosition(atom_order[k])))
126
+ atom_idx.append(atom_order[k])
127
  k += 1
128
  else:
129
  token_pos_2d.append((0.,0.,0.))
130
+ atom_idx.append(None)
131
 
132
+ return smi, token_pos, token_pos_2d, atom_idx
133
 
134
  def read_ligand_expo():
135
  """
 
140
  """
141
  file_name = "Components-smiles-stereo-oe.smi"
142
  try:
143
+ df = pd.read_csv(file_name, sep=r"[\t]+",
144
  header=None,
145
+ names=["SMILES", "ID", "Name"],
146
+ engine='python')
147
  except FileNotFoundError:
148
  url = f"http://ligand-expo.rcsb.org/dictionaries/{file_name}"
149
  print(url)
 
151
  open('Components-smiles-stereo-oe.smi', 'wb').write(r.content)
152
  df = pd.read_csv(file_name, sep="\t",
153
  header=None,
154
+ names=["SMILES", "ID", "Name"],
155
+ na_filter=False)
156
+ return df
157
 
158
 
159
  def get_pdb_components(pdb_id):
 
170
  return protein, ligand
171
 
172
 
173
+ def process_ligand(ligand, res_name, df_expo):
174
  """
175
  Add bond orders to a pdb ligand
176
  1. Select the ligand component with name "res_name"
 
181
  6. Assign the bond orders from the template from step 3
182
  :param ligand: ligand as generated by prody
183
  :param res_name: residue name of ligand to extract
184
+ :param df_expo: dictionary with LigandExpo
185
  :return: molecule with bond orders assigned
186
  """
187
+ sub_smiles = df_expo[df_expo['ID'].values == res_name]['SMILES'].values[0]
188
  template = AllChem.MolFromSmiles(sub_smiles)
189
 
190
  allres = ligand.select(f"resname {res_name}")
 
199
  mols.append(AllChem.AssignBondOrdersFromTemplate(template, rd_mol))
200
  return mols, template
201
 
202
+ def rot_from_two_vecs(e0_unnormalized, e1_unnormalized):
203
+ """Create rotation matrices from unnormalized vectors for the x and y-axes.
204
+ This creates a rotation matrix from two vectors using Gram-Schmidt
205
+ orthogonalization.
206
+ Args:
207
+ e0_unnormalized: vectors lying along x-axis of resulting rotation
208
+ e1_unnormalized: vectors lying in xy-plane of resulting rotation
209
+ Returns:
210
+ Rotations resulting from Gram-Schmidt procedure.
211
+ """
212
+ # Normalize the unit vector for the x-axis, e0.
213
+ e0 = e0_unnormalized / np.linalg.norm(e0_unnormalized)
214
+
215
+ # make e1 perpendicular to e0.
216
+ c = np.dot(e1_unnormalized, e0)
217
+ e1 = e1_unnormalized - c * e0
218
+ e1 = e1 / np.linalg.norm(e1)
219
+
220
+ # Compute e2 as cross product of e0 and e1.
221
+ e2 = np.cross(e0, e1)
222
+
223
+ # local to space frame
224
+ return np.stack([e0,e1,e2]).T
225
+
226
+ def process_entry(df, pdb_fn):
227
  try:
228
  """
229
  Slit pdb into protein and ligands,
230
  parse protein sequence and ligand tokens
231
+ :param df: ligand expo data
232
  :param pdb_fn: pdb entry file name
233
  :return:
234
  """
 
238
 
239
  ligand_mols = []
240
  ligand_names = []
241
+ ligand_bonds = []
242
  if ligand is not None:
243
  # filter ligands by molecular weight
244
  res_name_list = list(set(ligand.getResnames()))
245
  for res in res_name_list:
246
+ if res in ubiquitous_ligands:
247
+ continue
248
+ mols, template = process_ligand(ligand, res, df)
249
 
250
  mol_wt = ExactMolWt(template)
251
  natoms = template.GetNumAtoms()
252
 
253
+ if mol_wt >= mol_wt_cutoff and natoms >= min_atoms:
254
  # only use first copy of ligand
255
  mols = mols[:1]
256
  ligand_mols += mols
257
  ligand_names += [res]*len(mols)
258
 
259
+ bonds = []
260
+ for b in template.GetBonds():
261
+ bonds.append((b.GetBeginAtomIdx(), b.GetEndAtomIdx()))
262
+ ligand_bonds.append(bonds)
263
+
264
 
265
  ligand_smiles = []
266
  ligand_xyz = []
267
  ligand_xyz_2d = []
268
+ ligand_token_bonds = []
269
+ for mol, name, bonds in zip(ligand_mols, ligand_names, ligand_bonds):
270
  print('Processing {} and {}'.format(pdb_name, name))
271
+ smi, xyz, xyz_2d, atom_idx = tokenize_ligand(mol)
272
  ligand_smiles.append(smi)
273
  ligand_xyz.append(xyz)
274
  ligand_xyz_2d.append(xyz_2d)
275
 
276
+ ligand_token_bonds.append([ (atom_idx.index(b[0]), atom_idx.index(b[1])) for b in bonds ])
277
+
278
+ pdb_str = StringIO()
279
+ writePDBStream(pdb_str, protein)
280
+
281
+ seq, features = get_protein_sequence_and_coords(protein, pdb_str.getvalue())
282
+ features = { 'rigidgroups_gt_frames': features['rigidgroups_gt_frames'],
283
+ 'torsion_angles_sin_cos': features['torsion_angles_sin_cos']}
284
+ return pdb_name, seq, features, ligand_names, ligand_smiles, ligand_xyz, ligand_xyz_2d, ligand_token_bonds
285
  except Exception as e:
286
+ print(traceback.format_exc())
287
  print(repr(e))
288
 
289
+ def write_result(fn, data):
290
+ # expand sequences and ligands
291
+ pdb_id = [r[0] for r in data if r is not None for ligand in r[3]]
292
+ seq = [r[1] for r in data if r is not None for ligand in r[3]]
293
+ receptor_features = [r[2] for r in data if r is not None for ligand in r[3]]
294
+ lig_id = [l for r in data if r is not None for l in r[3]]
295
+ lig_smiles = [s for r in data if r is not None for s in r[4]]
296
+ lig_xyz = [xyz for r in data if r is not None for xyz in r[5]]
297
+ lig_xyz_2d = [xyz for r in data if r is not None for xyz in r[6]]
298
+ lig_bonds = [b for r in data if r is not None for b in r[7]]
299
+
300
+ import pandas as pd
301
+ df = pd.DataFrame({
302
+ 'pdb_id': pdb_id,
303
+ 'lig_id': lig_id,
304
+ 'seq': seq,
305
+ 'smiles': lig_smiles,
306
+ 'receptor_features': receptor_features,
307
+ 'ligand_xyz': lig_xyz,
308
+ 'ligand_xyz_2d': lig_xyz_2d,
309
+ 'ligand_bonds': lig_bonds})
310
+ df.to_pickle(fn)
311
+
312
  if __name__ == '__main__':
313
  import glob
314
 
315
  filenames = glob.glob('pdb/*/*.gz')
316
  filenames = sorted(filenames)
317
+
318
+ random.seed(42)
319
+ random.shuffle(filenames)
320
+
321
+ split_idx = int(0.9*len(filenames))
322
+ train = filenames[:split_idx]
323
+ test = filenames[split_idx:]
324
+
325
  comm = MPI.COMM_WORLD
326
  with MPICommExecutor(comm, root=0) as executor:
327
  # with MPIPoolExecutor() as executor:
328
  if executor is not None:
329
  # read ligand table
330
+ df = read_ligand_expo()
331
+
332
+ result = executor.map(partial(process_entry, df), train, chunksize=128)
333
+ result = list(result)
334
+
335
+ write_result('data/pdb_train.p', result)
336
 
337
+ result = executor.map(partial(process_entry, df), test, chunksize=128)
338
  result = list(result)
339
 
340
+ write_result('data/pdb_test.p', result)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
pdb.slurm CHANGED
@@ -3,7 +3,7 @@
3
  #SBATCH -p batch
4
  #SBATCH -A STF006
5
  #SBATCH -t 3:00:00
6
- #SBATCH -N 36
7
  #SBATCH --ntasks-per-node=16
8
 
9
  export PYTHONUNBUFFERED=1
 
3
  #SBATCH -p batch
4
  #SBATCH -A STF006
5
  #SBATCH -t 3:00:00
6
+ #SBATCH -N 128
7
  #SBATCH --ntasks-per-node=16
8
 
9
  export PYTHONUNBUFFERED=1
pdb_protein_ligand_complexes.py DELETED
@@ -1,131 +0,0 @@
1
- # coding=utf-8
2
- # Copyright 2020 The HuggingFace Datasets Authors and the current dataset script contributor.
3
- #
4
- # Licensed under the Apache License, Version 2.0 (the "License");
5
- # you may not use this file except in compliance with the License.
6
- # You may obtain a copy of the License at
7
- #
8
- # http://www.apache.org/licenses/LICENSE-2.0
9
- #
10
- # Unless required by applicable law or agreed to in writing, software
11
- # distributed under the License is distributed on an "AS IS" BASIS,
12
- # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
- # See the License for the specific language governing permissions and
14
- # limitations under the License.
15
- """TODO: A dataset of protein sequences, ligand SMILES, and complex coordinates."""
16
-
17
- import huggingface_hub
18
- import os
19
- import pyarrow.parquet as pq
20
- import datasets
21
-
22
-
23
- # TODO: Add BibTeX citation
24
- # Find for instance the citation on arxiv or on the dataset repo/website
25
- _CITATION = """\
26
- @InProceedings{huggingface:dataset,
27
- title = {jglaser/pdb_protein_ligand_complexes},
28
- author={Jens Glaser, ORNL
29
- },
30
- year={2022}
31
- }
32
- """
33
-
34
- # TODO: Add description of the dataset here
35
- # You can copy an official description
36
- _DESCRIPTION = """\
37
- A dataset to fine-tune language models on protein-ligand complex structures
38
- """
39
-
40
- # TODO: Add a link to an official homepage for the dataset here
41
- _HOMEPAGE = ""
42
-
43
- # TODO: Add the licence for the dataset here if you can find it
44
- _LICENSE = "BSD two-clause"
45
-
46
- # TODO: Add link to the official dataset URLs here
47
- # The HuggingFace dataset library don't host the datasets but only point to the original files
48
- # This can be an arbitrary nested dict/list of URLs (see below in `_split_generators` method)
49
- _URL = "https://huggingface.co/datasets/jglaser/pdb_protein_ligand_complexes/resolve/main/"
50
- _data_dir = "data/"
51
- _file_names = {'default': _data_dir+'pdb.parquet'}
52
-
53
- _URLs = {name: _URL+_file_names[name] for name in _file_names}
54
-
55
-
56
- # TODO: Name of the dataset usually match the script name with CamelCase instead of snake_case
57
- class PDBProteinLigandComplexes(datasets.ArrowBasedBuilder):
58
- """List of protein sequences, ligand SMILES, and complex coordinates."""
59
-
60
- VERSION = datasets.Version("1.3.0")
61
-
62
- def _info(self):
63
- # TODO: This method specifies the datasets.DatasetInfo object which contains informations and typings for the dataset
64
- #if self.config.name == "first_domain": # This is the name of the configuration selected in BUILDER_CONFIGS above
65
- # features = datasets.Features(
66
- # {
67
- # "sentence": datasets.Value("string"),
68
- # "option1": datasets.Value("string"),
69
- # "answer": datasets.Value("string")
70
- # # These are the features of your dataset like images, labels ...
71
- # }
72
- # )
73
- #else: # This is an example to show how to have different features for "first_domain" and "second_domain"
74
- features = datasets.Features(
75
- {
76
- "pdb_id": datasets.Value("string"),
77
- "lig_id": datasets.Value("string"),
78
- "seq": datasets.Value("string"),
79
- "smiles": datasets.Value("string"),
80
- "ligand_xyz": datasets.Sequence(datasets.Sequence(datasets.Value('float32'))),
81
- "ligand_xyz_2d": datasets.Sequence(datasets.Sequence(datasets.Value('float32'))),
82
- "receptor_xyz": datasets.Sequence(datasets.Sequence(datasets.Value('float32'))),
83
- # These are the features of your dataset like images, labels ...
84
- }
85
- )
86
- return datasets.DatasetInfo(
87
- # This is the description that will appear on the datasets page.
88
- description=_DESCRIPTION,
89
- # This defines the different columns of the dataset and their types
90
- features=features, # Here we define them above because they are different between the two configurations
91
- # If there's a common (input, target) tuple from the features,
92
- # specify them here. They'll be used if as_supervised=True in
93
- # builder.as_dataset.
94
- supervised_keys=None,
95
- # Homepage of the dataset for documentation
96
- homepage=_HOMEPAGE,
97
- # License for the dataset if available
98
- license=_LICENSE,
99
- # Citation for the dataset
100
- citation=_CITATION,
101
- )
102
-
103
- def _split_generators(self, dl_manager):
104
- """Returns SplitGenerators."""
105
- # TODO: This method is tasked with downloading/extracting the data and defining the splits depending on the configuration
106
- # If several configurations are possible (listed in BUILDER_CONFIGS), the configuration selected by the user is in self.config.name
107
-
108
- # dl_manager is a datasets.download.DownloadManager that can be used to download and extract URLs
109
- # It can accept any type or nested list/dict and will give back the same structure with the url replaced with path to local files.
110
- # By default the archives will be extracted and a path to a cached folder where they are extracted is returned instead of the archive
111
- files = dl_manager.download_and_extract(_URLs)
112
-
113
- return [
114
- datasets.SplitGenerator(
115
- # These kwargs will be passed to _generate_examples
116
- name=datasets.Split.TRAIN,
117
- gen_kwargs={
118
- 'filepath': files["default"],
119
- },
120
- ),
121
-
122
- ]
123
-
124
- def _generate_tables(
125
- self, filepath
126
- ):
127
- from pyarrow import fs
128
- local = fs.LocalFileSystem()
129
-
130
- for i, f in enumerate([filepath]):
131
- yield i, pq.read_table(f,filesystem=local)