pipeline.py 10.2 KB
Newer Older
Augustin-Zidek's avatar
Augustin-Zidek committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
# Copyright 2021 DeepMind Technologies Limited
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#      http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

"""Functions for building the input features for the AlphaFold model."""

import os
18
from typing import Any, Mapping, MutableMapping, Optional, Sequence, Union
19
from absl import logging
Augustin-Zidek's avatar
Augustin-Zidek committed
20
from alphafold.common import residue_constants
21
from alphafold.data import msa_identifiers
Augustin-Zidek's avatar
Augustin-Zidek committed
22
23
24
25
from alphafold.data import parsers
from alphafold.data import templates
from alphafold.data.tools import hhblits
from alphafold.data.tools import hhsearch
26
from alphafold.data.tools import hmmsearch
Augustin-Zidek's avatar
Augustin-Zidek committed
27
from alphafold.data.tools import jackhmmer
Tom Ward's avatar
Tom Ward committed
28
29
30
import numpy as np

# Internal import (7716).
Augustin-Zidek's avatar
Augustin-Zidek committed
31

32
33
FeatureDict = MutableMapping[str, np.ndarray]
TemplateSearcher = Union[hhsearch.HHSearch, hmmsearch.Hmmsearch]
Augustin-Zidek's avatar
Augustin-Zidek committed
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52


def make_sequence_features(
    sequence: str, description: str, num_res: int) -> FeatureDict:
  """Constructs a feature dict of sequence features."""
  features = {}
  features['aatype'] = residue_constants.sequence_to_onehot(
      sequence=sequence,
      mapping=residue_constants.restype_order_with_x,
      map_unknown_to_x=True)
  features['between_segment_residues'] = np.zeros((num_res,), dtype=np.int32)
  features['domain_name'] = np.array([description.encode('utf-8')],
                                     dtype=np.object_)
  features['residue_index'] = np.array(range(num_res), dtype=np.int32)
  features['seq_length'] = np.array([num_res] * num_res, dtype=np.int32)
  features['sequence'] = np.array([sequence.encode('utf-8')], dtype=np.object_)
  return features


53
def make_msa_features(msas: Sequence[parsers.Msa]) -> FeatureDict:
Augustin-Zidek's avatar
Augustin-Zidek committed
54
55
56
57
58
59
  """Constructs a feature dict of MSA features."""
  if not msas:
    raise ValueError('At least one MSA must be provided.')

  int_msa = []
  deletion_matrix = []
60
  species_ids = []
Augustin-Zidek's avatar
Augustin-Zidek committed
61
62
63
64
  seen_sequences = set()
  for msa_index, msa in enumerate(msas):
    if not msa:
      raise ValueError(f'MSA {msa_index} must contain at least one sequence.')
65
    for sequence_index, sequence in enumerate(msa.sequences):
Augustin-Zidek's avatar
Augustin-Zidek committed
66
67
68
69
70
      if sequence in seen_sequences:
        continue
      seen_sequences.add(sequence)
      int_msa.append(
          [residue_constants.HHBLITS_AA_TO_ID[res] for res in sequence])
71
72
73
74
75
76
      deletion_matrix.append(msa.deletion_matrix[sequence_index])
      identifiers = msa_identifiers.get_identifiers(
          msa.descriptions[sequence_index])
      species_ids.append(identifiers.species_id.encode('utf-8'))

  num_res = len(msas[0].sequences[0])
Augustin-Zidek's avatar
Augustin-Zidek committed
77
78
79
80
81
82
  num_alignments = len(int_msa)
  features = {}
  features['deletion_matrix_int'] = np.array(deletion_matrix, dtype=np.int32)
  features['msa'] = np.array(int_msa, dtype=np.int32)
  features['num_alignments'] = np.array(
      [num_alignments] * num_res, dtype=np.int32)
83
  features['msa_species_identifiers'] = np.array(species_ids, dtype=np.object_)
Augustin-Zidek's avatar
Augustin-Zidek committed
84
85
86
  return features


87
88
def run_msa_tool(msa_runner, input_fasta_path: str, msa_out_path: str,
                 msa_format: str, use_precomputed_msas: bool,
Augustin Zidek's avatar
Augustin Zidek committed
89
                 max_sto_sequences: Optional[int] = None
90
91
92
                 ) -> Mapping[str, Any]:
  """Runs an MSA tool, checking if output already exists first."""
  if not use_precomputed_msas or not os.path.exists(msa_out_path):
Augustin Zidek's avatar
Augustin Zidek committed
93
94
95
96
    if msa_format == 'sto' and max_sto_sequences is not None:
      result = msa_runner.query(input_fasta_path, max_sto_sequences)[0]  # pytype: disable=wrong-arg-count
    else:
      result = msa_runner.query(input_fasta_path)[0]
97
98
99
100
    with open(msa_out_path, 'w') as f:
      f.write(result[msa_format])
  else:
    logging.warning('Reading MSA from file %s', msa_out_path)
Augustin Zidek's avatar
Augustin Zidek committed
101
102
103
104
105
106
107
    if msa_format == 'sto' and max_sto_sequences is not None:
      precomputed_msa = parsers.truncate_stockholm_msa(
          msa_out_path, max_sto_sequences)
      result = {'sto': precomputed_msa}
    else:
      with open(msa_out_path, 'r') as f:
        result = {msa_format: f.read()}
108
109
110
  return result


Augustin-Zidek's avatar
Augustin-Zidek committed
111
112
113
114
115
116
117
118
class DataPipeline:
  """Runs the alignment tools and assembles the input features."""

  def __init__(self,
               jackhmmer_binary_path: str,
               hhblits_binary_path: str,
               uniref90_database_path: str,
               mgnify_database_path: str,
119
120
121
               bfd_database_path: Optional[str],
               uniclust30_database_path: Optional[str],
               small_bfd_database_path: Optional[str],
122
               template_searcher: TemplateSearcher,
Augustin-Zidek's avatar
Augustin-Zidek committed
123
               template_featurizer: templates.TemplateHitFeaturizer,
124
               use_small_bfd: bool,
Augustin-Zidek's avatar
Augustin-Zidek committed
125
               mgnify_max_hits: int = 501,
126
127
128
               uniref_max_hits: int = 10000,
               use_precomputed_msas: bool = False):
    """Initializes the data pipeline."""
129
    self._use_small_bfd = use_small_bfd
Augustin-Zidek's avatar
Augustin-Zidek committed
130
131
132
    self.jackhmmer_uniref90_runner = jackhmmer.Jackhmmer(
        binary_path=jackhmmer_binary_path,
        database_path=uniref90_database_path)
133
134
135
136
137
138
139
140
    if use_small_bfd:
      self.jackhmmer_small_bfd_runner = jackhmmer.Jackhmmer(
          binary_path=jackhmmer_binary_path,
          database_path=small_bfd_database_path)
    else:
      self.hhblits_bfd_uniclust_runner = hhblits.HHBlits(
          binary_path=hhblits_binary_path,
          databases=[bfd_database_path, uniclust30_database_path])
Augustin-Zidek's avatar
Augustin-Zidek committed
141
142
143
    self.jackhmmer_mgnify_runner = jackhmmer.Jackhmmer(
        binary_path=jackhmmer_binary_path,
        database_path=mgnify_database_path)
144
    self.template_searcher = template_searcher
Augustin-Zidek's avatar
Augustin-Zidek committed
145
146
147
    self.template_featurizer = template_featurizer
    self.mgnify_max_hits = mgnify_max_hits
    self.uniref_max_hits = uniref_max_hits
148
    self.use_precomputed_msas = use_precomputed_msas
Augustin-Zidek's avatar
Augustin-Zidek committed
149
150
151
152
153
154
155
156
157
158
159
160
161
162

  def process(self, input_fasta_path: str, msa_output_dir: str) -> FeatureDict:
    """Runs alignment tools on the input sequence and creates features."""
    with open(input_fasta_path) as f:
      input_fasta_str = f.read()
    input_seqs, input_descs = parsers.parse_fasta(input_fasta_str)
    if len(input_seqs) != 1:
      raise ValueError(
          f'More than one input sequence found in {input_fasta_path}.')
    input_sequence = input_seqs[0]
    input_description = input_descs[0]
    num_res = len(input_sequence)

    uniref90_out_path = os.path.join(msa_output_dir, 'uniref90_hits.sto')
163
    jackhmmer_uniref90_result = run_msa_tool(
Augustin Zidek's avatar
Augustin Zidek committed
164
165
166
167
168
169
        msa_runner=self.jackhmmer_uniref90_runner,
        input_fasta_path=input_fasta_path,
        msa_out_path=uniref90_out_path,
        msa_format='sto',
        use_precomputed_msas=self.use_precomputed_msas,
        max_sto_sequences=self.uniref_max_hits)
Augustin-Zidek's avatar
Augustin-Zidek committed
170
    mgnify_out_path = os.path.join(msa_output_dir, 'mgnify_hits.sto')
171
    jackhmmer_mgnify_result = run_msa_tool(
Augustin Zidek's avatar
Augustin Zidek committed
172
173
174
175
176
177
        msa_runner=self.jackhmmer_mgnify_runner,
        input_fasta_path=input_fasta_path,
        msa_out_path=mgnify_out_path,
        msa_format='sto',
        use_precomputed_msas=self.use_precomputed_msas,
        max_sto_sequences=self.mgnify_max_hits)
178
179

    msa_for_templates = jackhmmer_uniref90_result['sto']
Augustin Zidek's avatar
Augustin Zidek committed
180
    msa_for_templates = parsers.deduplicate_stockholm_msa(msa_for_templates)
181
182
183
184
185
186
187
188
189
190
191
    msa_for_templates = parsers.remove_empty_columns_from_stockholm_msa(
        msa_for_templates)

    if self.template_searcher.input_format == 'sto':
      pdb_templates_result = self.template_searcher.query(msa_for_templates)
    elif self.template_searcher.input_format == 'a3m':
      uniref90_msa_as_a3m = parsers.convert_stockholm_to_a3m(msa_for_templates)
      pdb_templates_result = self.template_searcher.query(uniref90_msa_as_a3m)
    else:
      raise ValueError('Unrecognized template input format: '
                       f'{self.template_searcher.input_format}')
192

193
194
195
196
    pdb_hits_out_path = os.path.join(
        msa_output_dir, f'pdb_hits.{self.template_searcher.output_format}')
    with open(pdb_hits_out_path, 'w') as f:
      f.write(pdb_templates_result)
Augustin-Zidek's avatar
Augustin-Zidek committed
197

198
199
    uniref90_msa = parsers.parse_stockholm(jackhmmer_uniref90_result['sto'])
    mgnify_msa = parsers.parse_stockholm(jackhmmer_mgnify_result['sto'])
Augustin-Zidek's avatar
Augustin-Zidek committed
200

201
202
    pdb_template_hits = self.template_searcher.get_template_hits(
        output_string=pdb_templates_result, input_sequence=input_sequence)
Augustin-Zidek's avatar
Augustin-Zidek committed
203

204
205
206
    if self._use_small_bfd:
      bfd_out_path = os.path.join(msa_output_dir, 'small_bfd_hits.sto')
      jackhmmer_small_bfd_result = run_msa_tool(
Augustin Zidek's avatar
Augustin Zidek committed
207
208
209
210
211
          msa_runner=self.jackhmmer_small_bfd_runner,
          input_fasta_path=input_fasta_path,
          msa_out_path=bfd_out_path,
          msa_format='sto',
          use_precomputed_msas=self.use_precomputed_msas)
212
      bfd_msa = parsers.parse_stockholm(jackhmmer_small_bfd_result['sto'])
213
214
    else:
      bfd_out_path = os.path.join(msa_output_dir, 'bfd_uniclust_hits.a3m')
215
      hhblits_bfd_uniclust_result = run_msa_tool(
Augustin Zidek's avatar
Augustin Zidek committed
216
217
218
219
220
          msa_runner=self.hhblits_bfd_uniclust_runner,
          input_fasta_path=input_fasta_path,
          msa_out_path=bfd_out_path,
          msa_format='a3m',
          use_precomputed_msas=self.use_precomputed_msas)
221
      bfd_msa = parsers.parse_a3m(hhblits_bfd_uniclust_result['a3m'])
Augustin-Zidek's avatar
Augustin-Zidek committed
222
223
224

    templates_result = self.template_featurizer.get_templates(
        query_sequence=input_sequence,
225
        hits=pdb_template_hits)
Augustin-Zidek's avatar
Augustin-Zidek committed
226
227
228
229
230
231

    sequence_features = make_sequence_features(
        sequence=input_sequence,
        description=input_description,
        num_res=num_res)

232
    msa_features = make_msa_features((uniref90_msa, bfd_msa, mgnify_msa))
Augustin-Zidek's avatar
Augustin-Zidek committed
233

234
235
236
237
238
239
240
241
242
    logging.info('Uniref90 MSA size: %d sequences.', len(uniref90_msa))
    logging.info('BFD MSA size: %d sequences.', len(bfd_msa))
    logging.info('MGnify MSA size: %d sequences.', len(mgnify_msa))
    logging.info('Final (deduplicated) MSA size: %d sequences.',
                 msa_features['num_alignments'][0])
    logging.info('Total number of templates (NB: this can include bad '
                 'templates and is later filtered to top 4): %d.',
                 templates_result.features['template_domain_names'].shape[0])

Augustin-Zidek's avatar
Augustin-Zidek committed
243
    return {**sequence_features, **msa_features, **templates_result.features}