Skip to content

GitLab

Commit a1c29028 authored by zhangqha's avatar zhangqha
Browse files

update uni-fold

parents
Pipeline #183 canceled with stages
Hide whitespace changes
Inline Side-by-side
Showing with 3396 additions and 0 deletions
unifold/msa/mmcif.py 0 → 100755
View file @ a1c29028
# 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.
"""Parses the mmCIF file format."""
import collections
import dataclasses
import functools
import io
from typing import Any, Mapping, Optional, Sequence, Tuple
from absl import logging
from Bio import PDB
from Bio.PDB.MMCIFParser import MMCIFParser
from Bio.Data import SCOPData
# Type aliases:
ChainId = str
PdbHeader = Mapping[str, Any]
PdbStructure = PDB.Structure.Structure
SeqRes = str
MmCIFDict = Mapping[str, Sequence[str]]
@dataclasses.dataclass(frozen=True)
class Monomer:
id: str
num: int
# Note - mmCIF format provides no guarantees on the type of author-assigned
# sequence numbers. They need not be integers.
@dataclasses.dataclass(frozen=True)
class AtomSite:
residue_name: str
author_chain_id: str
mmcif_chain_id: str
author_seq_num: str
mmcif_seq_num: int
insertion_code: str
hetatm_atom: str
model_num: int
# Used to map SEQRES index to a residue in the structure.
@dataclasses.dataclass(frozen=True)
class ResiduePosition:
chain_id: str
residue_number: int
insertion_code: str
@dataclasses.dataclass(frozen=True)
class ResidueAtPosition:
position: Optional[ResiduePosition]
name: str
is_missing: bool
hetflag: str
@dataclasses.dataclass(frozen=True)
class MmcifObject:
"""Representation of a parsed mmCIF file.
Contains:
file_id: A meaningful name, e.g. a pdb_id. Should be unique amongst all
files being processed.
header: Biopython header.
structure: Biopython structure.
chain_to_seqres: Dict mapping chain_id to 1 letter amino acid sequence. E.g.
{'A': 'ABCDEFG'}
seqres_to_structure: Dict; for each chain_id contains a mapping between
SEQRES index and a ResidueAtPosition. e.g. {'A': {0: ResidueAtPosition, 1: ResidueAtPosition, ...}}
raw_string: The raw string used to construct the MmcifObject.
"""
file_id: str
header: PdbHeader
structure: PdbStructure
chain_to_seqres: Mapping[ChainId, SeqRes]
seqres_to_structure: Mapping[ChainId, Mapping[int, ResidueAtPosition]]
raw_string: Any
mmcif_to_author_chain_id: Mapping[ChainId, ChainId]
valid_chains: Mapping[ChainId, str]
@dataclasses.dataclass(frozen=True)
class ParsingResult:
"""Returned by the parse function.
Contains:
mmcif_object: A MmcifObject, may be None if no chain could be successfully
parsed.
errors: A dict mapping (file_id, chain_id) to any exception generated.
"""
mmcif_object: Optional[MmcifObject]
errors: Mapping[Tuple[str, str], Any]
class ParseError(Exception):
"""An error indicating that an mmCIF file could not be parsed."""
def mmcif_loop_to_list(
prefix: str, parsed_info: MmCIFDict
) -> Sequence[Mapping[str, str]]:
"""Extracts loop associated with a prefix from mmCIF data as a list.
Reference for loop_ in mmCIF:
http://mmcif.wwpdb.org/docs/tutorials/mechanics/pdbx-mmcif-syntax.html
Args:
prefix: Prefix shared by each of the data items in the loop.
e.g. '_entity_poly_seq.', where the data items are _entity_poly_seq.num,
_entity_poly_seq.mon_id. Should include the trailing period.
parsed_info: A dict of parsed mmCIF data, e.g. _mmcif_dict from a Biopython
parser.
Returns:
Returns a list of dicts; each dict represents 1 entry from an mmCIF loop.
"""
cols = []
data = []
for key, value in parsed_info.items():
if key.startswith(prefix):
cols.append(key)
data.append(value)
assert all([len(xs) == len(data[0]) for xs in data]), (
"mmCIF error: Not all loops are the same length: %s" % cols
)
return [dict(zip(cols, xs)) for xs in zip(*data)]
def mmcif_loop_to_dict(
prefix: str,
index: str,
parsed_info: MmCIFDict,
) -> Mapping[str, Mapping[str, str]]:
"""Extracts loop associated with a prefix from mmCIF data as a dictionary.
Args:
prefix: Prefix shared by each of the data items in the loop.
e.g. '_entity_poly_seq.', where the data items are _entity_poly_seq.num,
_entity_poly_seq.mon_id. Should include the trailing period.
index: Which item of loop data should serve as the key.
parsed_info: A dict of parsed mmCIF data, e.g. _mmcif_dict from a Biopython
parser.
Returns:
Returns a dict of dicts; each dict represents 1 entry from an mmCIF loop,
indexed by the index column.
"""
entries = mmcif_loop_to_list(prefix, parsed_info)
return {entry[index]: entry for entry in entries}
@functools.lru_cache(16, typed=False)
def fast_parse(
*, file_id: str, mmcif_string: str, catch_all_errors: bool = True
) -> ParsingResult:
"""Entry point, parses an mmcif_string.
Args:
file_id: A string identifier for this file. Should be unique within the
collection of files being processed.
mmcif_string: Contents of an mmCIF file.
catch_all_errors: If True, all exceptions are caught and error messages are
returned as part of the ParsingResult. If False exceptions will be allowed
to propagate.
Returns:
A ParsingResult.
"""
errors = {}
try:
parser = MMCIFParser(QUIET=True)
handle = io.StringIO(mmcif_string)
full_structure = parser.get_structure("", handle)
parsed_info = parser._mmcif_dict # pylint:disable=protected-access
# Ensure all values are lists, even if singletons.
for key, value in parsed_info.items():
if not isinstance(value, list):
parsed_info[key] = [value]
header = _get_header(parsed_info)
# Determine the protein chains, and their start numbers according to the
# internal mmCIF numbering scheme (likely but not guaranteed to be 1).
valid_chains = _get_protein_chains(parsed_info=parsed_info)
if not valid_chains:
return ParsingResult(
None, {(file_id, ""): "No protein chains found in this file."}
)
mmcif_to_author_chain_id = {}
seq_to_structure_mappings = {}
for atom in _get_atom_site_list(parsed_info):
if atom.model_num != "1":
# We only process the first model at the moment.
continue
mmcif_to_author_chain_id[atom.mmcif_chain_id] = atom.author_chain_id
mmcif_object = MmcifObject(
file_id=file_id,
header=header,
structure=None,
chain_to_seqres=None,
seqres_to_structure=None,
raw_string=parsed_info,
mmcif_to_author_chain_id=mmcif_to_author_chain_id,
valid_chains=valid_chains,
)
return ParsingResult(mmcif_object=mmcif_object, errors=errors)
except Exception as e: # pylint:disable=broad-except
errors[(file_id, "")] = e
if not catch_all_errors:
raise
return ParsingResult(mmcif_object=None, errors=errors)
@functools.lru_cache(16, typed=False)
def parse(
*, file_id: str, mmcif_string: str, catch_all_errors: bool = True
) -> ParsingResult:
"""Entry point, parses an mmcif_string.
Args:
file_id: A string identifier for this file. Should be unique within the
collection of files being processed.
mmcif_string: Contents of an mmCIF file.
catch_all_errors: If True, all exceptions are caught and error messages are
returned as part of the ParsingResult. If False exceptions will be allowed
to propagate.
Returns:
A ParsingResult.
"""
errors = {}
try:
parser = PDB.MMCIFParser(QUIET=True)
handle = io.StringIO(mmcif_string)
full_structure = parser.get_structure("", handle)
first_model_structure = _get_first_model(full_structure)
# Extract the _mmcif_dict from the parser, which contains useful fields not
# reflected in the Biopython structure.
parsed_info = parser._mmcif_dict # pylint:disable=protected-access
# Ensure all values are lists, even if singletons.
for key, value in parsed_info.items():
if not isinstance(value, list):
parsed_info[key] = [value]
header = _get_header(parsed_info)
# Determine the protein chains, and their start numbers according to the
# internal mmCIF numbering scheme (likely but not guaranteed to be 1).
valid_chains = _get_protein_chains(parsed_info=parsed_info)
if not valid_chains:
return ParsingResult(
None, {(file_id, ""): "No protein chains found in this file."}
)
seq_start_num = {
chain_id: min([monomer.num for monomer in seq])
for chain_id, seq in valid_chains.items()
}
# Loop over the atoms for which we have coordinates. Populate two mappings:
# -mmcif_to_author_chain_id (maps internal mmCIF chain ids to chain ids used
# the authors / Biopython).
# -seq_to_structure_mappings (maps idx into sequence to ResidueAtPosition).
mmcif_to_author_chain_id = {}
seq_to_structure_mappings = {}
for atom in _get_atom_site_list(parsed_info):
if atom.model_num != "1":
# We only process the first model at the moment.
continue
mmcif_to_author_chain_id[atom.mmcif_chain_id] = atom.author_chain_id
if atom.mmcif_chain_id in valid_chains:
hetflag = " "
if atom.hetatm_atom == "HETATM":
# Water atoms are assigned a special hetflag of W in Biopython. We
# need to do the same, so that this hetflag can be used to fetch
# a residue from the Biopython structure by id.
if atom.residue_name in ("HOH", "WAT"):
hetflag = "W"
else:
hetflag = "H_" + atom.residue_name
insertion_code = atom.insertion_code
if not _is_set(atom.insertion_code):
insertion_code = " "
position = ResiduePosition(
chain_id=atom.author_chain_id,
residue_number=int(atom.author_seq_num),
insertion_code=insertion_code,
)
seq_idx = int(atom.mmcif_seq_num) - seq_start_num[atom.mmcif_chain_id]
current = seq_to_structure_mappings.get(atom.author_chain_id, {})
current[seq_idx] = ResidueAtPosition(
position=position,
name=atom.residue_name,
is_missing=False,
hetflag=hetflag,
)
seq_to_structure_mappings[atom.author_chain_id] = current
# Add missing residue information to seq_to_structure_mappings.
for chain_id, seq_info in valid_chains.items():
author_chain = mmcif_to_author_chain_id[chain_id]
current_mapping = seq_to_structure_mappings[author_chain]
for idx, monomer in enumerate(seq_info):
if idx not in current_mapping:
current_mapping[idx] = ResidueAtPosition(
position=None, name=monomer.id, is_missing=True, hetflag=" "
)
author_chain_to_sequence = {}
for chain_id, seq_info in valid_chains.items():
author_chain = mmcif_to_author_chain_id[chain_id]
seq = []
for monomer in seq_info:
code = SCOPData.protein_letters_3to1.get(monomer.id, "X")
seq.append(code if len(code) == 1 else "X")
seq = "".join(seq)
author_chain_to_sequence[author_chain] = seq
mmcif_object = MmcifObject(
file_id=file_id,
header=header,
structure=first_model_structure,
chain_to_seqres=author_chain_to_sequence,
seqres_to_structure=seq_to_structure_mappings,
raw_string=parsed_info,
mmcif_to_author_chain_id=mmcif_to_author_chain_id,
valid_chains=valid_chains,
)
return ParsingResult(mmcif_object=mmcif_object, errors=errors)
except Exception as e: # pylint:disable=broad-except
errors[(file_id, "")] = e
if not catch_all_errors:
raise
return ParsingResult(mmcif_object=None, errors=errors)
def _get_first_model(structure: PdbStructure) -> PdbStructure:
"""Returns the first model in a Biopython structure."""
return next(structure.get_models())
_MIN_LENGTH_OF_CHAIN_TO_BE_COUNTED_AS_PEPTIDE = 21
def get_release_date(parsed_info: MmCIFDict) -> str:
"""Returns the oldest revision date."""
revision_dates = parsed_info["_pdbx_audit_revision_history.revision_date"]
return min(revision_dates)
def _get_header(parsed_info: MmCIFDict) -> PdbHeader:
"""Returns a basic header containing method, release date and resolution."""
header = {}
experiments = mmcif_loop_to_list("_exptl.", parsed_info)
header["structure_method"] = ",".join(
[experiment["_exptl.method"].lower() for experiment in experiments]
)
# Note: The release_date here corresponds to the oldest revision. We prefer to
# use this for dataset filtering over the deposition_date.
if "_pdbx_audit_revision_history.revision_date" in parsed_info:
header["release_date"] = get_release_date(parsed_info)
else:
logging.warning(
"Could not determine release_date: %s", parsed_info["_entry.id"]
)
header["resolution"] = 0.00
for res_key in (
"_refine.ls_d_res_high",
"_em_3d_reconstruction.resolution",
"_reflns.d_resolution_high",
):
if res_key in parsed_info:
try:
raw_resolution = parsed_info[res_key][0]
header["resolution"] = float(raw_resolution)
except ValueError:
logging.debug("Invalid resolution format: %s", parsed_info[res_key])
return header
def _get_atom_site_list(parsed_info: MmCIFDict) -> Sequence[AtomSite]:
"""Returns list of atom sites; contains data not present in the structure."""
return [
AtomSite(*site)
for site in zip( # pylint:disable=g-complex-comprehension
parsed_info["_atom_site.label_comp_id"],
parsed_info["_atom_site.auth_asym_id"],
parsed_info["_atom_site.label_asym_id"],
parsed_info["_atom_site.auth_seq_id"],
parsed_info["_atom_site.label_seq_id"],
parsed_info["_atom_site.pdbx_PDB_ins_code"],
parsed_info["_atom_site.group_PDB"],
parsed_info["_atom_site.pdbx_PDB_model_num"],
)
]
def _get_protein_chains(
*, parsed_info: Mapping[str, Any]
) -> Mapping[ChainId, Sequence[Monomer]]:
"""Extracts polymer information for protein chains only.
Args:
parsed_info: _mmcif_dict produced by the Biopython parser.
Returns:
A dict mapping mmcif chain id to a list of Monomers.
"""
# Get polymer information for each entity in the structure.
entity_poly_seqs = mmcif_loop_to_list("_entity_poly_seq.", parsed_info)
polymers = collections.defaultdict(list)
for entity_poly_seq in entity_poly_seqs:
polymers[entity_poly_seq["_entity_poly_seq.entity_id"]].append(
Monomer(
id=entity_poly_seq["_entity_poly_seq.mon_id"],
num=int(entity_poly_seq["_entity_poly_seq.num"]),
)
)
# Get chemical compositions. Will allow us to identify which of these polymers
# are proteins.
chem_comps = mmcif_loop_to_dict("_chem_comp.", "_chem_comp.id", parsed_info)
# Get chains information for each entity. Necessary so that we can return a
# dict keyed on chain id rather than entity.
struct_asyms = mmcif_loop_to_list("_struct_asym.", parsed_info)
entity_to_mmcif_chains = collections.defaultdict(list)
for struct_asym in struct_asyms:
chain_id = struct_asym["_struct_asym.id"]
entity_id = struct_asym["_struct_asym.entity_id"]
entity_to_mmcif_chains[entity_id].append(chain_id)
# Identify and return the valid protein chains.
valid_chains = {}
for entity_id, seq_info in polymers.items():
chain_ids = entity_to_mmcif_chains[entity_id]
# Reject polymers without any peptide-like components, such as DNA/RNA.
if any(
[
"peptide" in chem_comps[monomer.id]["_chem_comp.type"]
for monomer in seq_info
]
):
for chain_id in chain_ids:
valid_chains[chain_id] = seq_info
return valid_chains
def _is_set(data: str) -> bool:
"""Returns False if data is a special mmCIF character indicating 'unset'."""
return data not in (".", "?")
unifold/msa/msa_identifiers.py 0 → 100755
View file @ a1c29028
# 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.
"""Utilities for extracting identifiers from MSA sequence descriptions."""
import dataclasses
import re
from typing import Optional
# Sequences coming from UniProtKB database come in the
# `db|UniqueIdentifier|EntryName` format, e.g. `tr|A0A146SKV9|A0A146SKV9_FUNHE`
# or `sp|P0C2L1|A3X1_LOXLA` (for TREMBL/Swiss-Prot respectively).
_UNIPROT_PATTERN = re.compile(
r"""
^
# UniProtKB/TrEMBL or UniProtKB/Swiss-Prot
(?:tr|sp)
\|
# A primary accession number of the UniProtKB entry.
(?P<AccessionIdentifier>[A-Za-z0-9]{6,10})
# Occasionally there is a _0 or _1 isoform suffix, which we ignore.
(?:_\d)?
\|
# TREMBL repeats the accession ID here. Swiss-Prot has a mnemonic
# protein ID code.
(?:[A-Za-z0-9]+)
_
# A mnemonic species identification code.
(?P<SpeciesIdentifier>([A-Za-z0-9]){1,5})
# Small BFD uses a final value after an underscore, which we ignore.
(?:_\d+)?
$
""",
re.VERBOSE,
)
@dataclasses.dataclass(frozen=True)
class Identifiers:
species_id: str = ""
def _parse_sequence_identifier(msa_sequence_identifier: str) -> Identifiers:
"""Gets accession id and species from an msa sequence identifier.
The sequence identifier has the format specified by
_UNIPROT_TREMBL_ENTRY_NAME_PATTERN or _UNIPROT_SWISSPROT_ENTRY_NAME_PATTERN.
An example of a sequence identifier: `tr|A0A146SKV9|A0A146SKV9_FUNHE`
Args:
msa_sequence_identifier: a sequence identifier.
Returns:
An `Identifiers` instance with a species_id. These
can be empty in the case where no identifier was found.
"""
matches = re.search(_UNIPROT_PATTERN, msa_sequence_identifier.strip())
if matches:
return Identifiers(species_id=matches.group("SpeciesIdentifier"))
return Identifiers()
def _extract_sequence_identifier(description: str) -> Optional[str]:
"""Extracts sequence identifier from description. Returns None if no match."""
split_description = description.split()
if split_description:
return split_description[0].partition("/")[0]
else:
return None
def get_identifiers(description: str) -> Identifiers:
"""Computes extra MSA features from the description."""
sequence_identifier = _extract_sequence_identifier(description)
if sequence_identifier is None:
return Identifiers()
else:
return _parse_sequence_identifier(sequence_identifier)
unifold/msa/parsers.py 0 → 100755
View file @ a1c29028
# 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 parsing various file formats."""
import collections
import dataclasses
import itertools
import re
import string
from typing import Dict, Iterable, List, Optional, Sequence, Tuple, Set
DeletionMatrix = Sequence[Sequence[int]]
@dataclasses.dataclass(frozen=True)
class Msa:
"""Class representing a parsed MSA file."""
sequences: Sequence[str]
deletion_matrix: DeletionMatrix
descriptions: Sequence[str]
def __post_init__(self):
if not (
len(self.sequences) == len(self.deletion_matrix) == len(self.descriptions)
):
raise ValueError(
"All fields for an MSA must have the same length. "
f"Got {len(self.sequences)} sequences, "
f"{len(self.deletion_matrix)} rows in the deletion matrix and "
f"{len(self.descriptions)} descriptions."
)
def __len__(self):
return len(self.sequences)
def truncate(self, max_seqs: int):
return Msa(
sequences=self.sequences[:max_seqs],
deletion_matrix=self.deletion_matrix[:max_seqs],
descriptions=self.descriptions[:max_seqs],
)
@dataclasses.dataclass(frozen=True)
class TemplateHit:
"""Class representing a template hit."""
index: int
name: str
aligned_cols: int
sum_probs: Optional[float]
query: str
hit_sequence: str
indices_query: List[int]
indices_hit: List[int]
def parse_fasta(fasta_string: str) -> Tuple[Sequence[str], Sequence[str]]:
"""Parses FASTA string and returns list of strings with amino-acid sequences.
Arguments:
fasta_string: The string contents of a FASTA file.
Returns:
A tuple of two lists:
* A list of sequences.
* A list of sequence descriptions taken from the comment lines. In the
same order as the sequences.
"""
sequences = []
descriptions = []
index = -1
for line in fasta_string.splitlines():
line = line.strip()
if line.startswith(">"):
index += 1
descriptions.append(line[1:]) # Remove the '>' at the beginning.
sequences.append("")
continue
elif not line:
continue # Skip blank lines.
sequences[index] += line
return sequences, descriptions
def parse_stockholm(stockholm_string: str) -> Msa:
"""Parses sequences and deletion matrix from stockholm format alignment.
Args:
stockholm_string: The string contents of a stockholm file. The first
sequence in the file should be the query sequence.
Returns:
A tuple of:
* A list of sequences that have been aligned to the query. These
might contain duplicates.
* The deletion matrix for the alignment as a list of lists. The element
at `deletion_matrix[i][j]` is the number of residues deleted from
the aligned sequence i at residue position j.
* The names of the targets matched, including the jackhmmer subsequence
suffix.
"""
name_to_sequence = collections.OrderedDict()
for line in stockholm_string.splitlines():
line = line.strip()
if not line or line.startswith(("#", "//")):
continue
name, sequence = line.split()
if name not in name_to_sequence:
name_to_sequence[name] = ""
name_to_sequence[name] += sequence
msa = []
deletion_matrix = []
query = ""
keep_columns = []
for seq_index, sequence in enumerate(name_to_sequence.values()):
if seq_index == 0:
# Gather the columns with gaps from the query
query = sequence
keep_columns = [i for i, res in enumerate(query) if res != "-"]
# Remove the columns with gaps in the query from all sequences.
aligned_sequence = "".join([sequence[c] for c in keep_columns])
msa.append(aligned_sequence)
# Count the number of deletions w.r.t. query.
deletion_vec = []
deletion_count = 0
for seq_res, query_res in zip(sequence, query):
if seq_res != "-" or query_res != "-":
if query_res == "-":
deletion_count += 1
else:
deletion_vec.append(deletion_count)
deletion_count = 0
deletion_matrix.append(deletion_vec)
return Msa(
sequences=msa,
deletion_matrix=deletion_matrix,
descriptions=list(name_to_sequence.keys()),
)
def parse_a3m(a3m_string: str) -> Msa:
"""Parses sequences and deletion matrix from a3m format alignment.
Args:
a3m_string: The string contents of a a3m file. The first sequence in the
file should be the query sequence.
Returns:
A tuple of:
* A list of sequences that have been aligned to the query. These
might contain duplicates.
* The deletion matrix for the alignment as a list of lists. The element
at `deletion_matrix[i][j]` is the number of residues deleted from
the aligned sequence i at residue position j.
* A list of descriptions, one per sequence, from the a3m file.
"""
sequences, descriptions = parse_fasta(a3m_string)
deletion_matrix = []
for msa_sequence in sequences:
deletion_vec = []
deletion_count = 0
for j in msa_sequence:
if j.islower():
deletion_count += 1
else:
deletion_vec.append(deletion_count)
deletion_count = 0
deletion_matrix.append(deletion_vec)
# Make the MSA matrix out of aligned (deletion-free) sequences.
deletion_table = str.maketrans("", "", string.ascii_lowercase)
aligned_sequences = [s.translate(deletion_table) for s in sequences]
return Msa(
sequences=aligned_sequences,
deletion_matrix=deletion_matrix,
descriptions=descriptions,
)
def _convert_sto_seq_to_a3m(
query_non_gaps: Sequence[bool], sto_seq: str
) -> Iterable[str]:
for is_query_res_non_gap, sequence_res in zip(query_non_gaps, sto_seq):
if is_query_res_non_gap:
yield sequence_res
elif sequence_res != "-":
yield sequence_res.lower()
def convert_stockholm_to_a3m(
stockholm_format: str,
max_sequences: Optional[int] = None,
remove_first_row_gaps: bool = True,
) -> str:
"""Converts MSA in Stockholm format to the A3M format."""
descriptions = {}
sequences = {}
reached_max_sequences = False
for line in stockholm_format.splitlines():
reached_max_sequences = max_sequences and len(sequences) >= max_sequences
if line.strip() and not line.startswith(("#", "//")):
# Ignore blank lines, markup and end symbols - remainder are alignment
# sequence parts.
seqname, aligned_seq = line.split(maxsplit=1)
if seqname not in sequences:
if reached_max_sequences:
continue
sequences[seqname] = ""
sequences[seqname] += aligned_seq
for line in stockholm_format.splitlines():
if line[:4] == "#=GS":
# Description row - example format is:
# #=GS UniRef90_Q9H5Z4/4-78 DE [subseq from] cDNA: FLJ22755 ...
columns = line.split(maxsplit=3)
seqname, feature = columns[1:3]
value = columns[3] if len(columns) == 4 else ""
if feature != "DE":
continue
if reached_max_sequences and seqname not in sequences:
continue
descriptions[seqname] = value
if len(descriptions) == len(sequences):
break
# Convert sto format to a3m line by line
a3m_sequences = {}
if remove_first_row_gaps:
# query_sequence is assumed to be the first sequence
query_sequence = next(iter(sequences.values()))
query_non_gaps = [res != "-" for res in query_sequence]
for seqname, sto_sequence in sequences.items():
# Dots are optional in a3m format and are commonly removed.
out_sequence = sto_sequence.replace(".", "")
if remove_first_row_gaps:
out_sequence = "".join(
_convert_sto_seq_to_a3m(query_non_gaps, out_sequence)
)
a3m_sequences[seqname] = out_sequence
fasta_chunks = (
f">{k} {descriptions.get(k, '')}\n{a3m_sequences[k]}" for k in a3m_sequences
)
return "\n".join(fasta_chunks) + "\n" # Include terminating newline.
def _keep_line(line: str, seqnames: Set[str]) -> bool:
"""Function to decide which lines to keep."""
if not line.strip():
return True
if line.strip() == "//": # End tag
return True
if line.startswith("# STOCKHOLM"): # Start tag
return True
if line.startswith("#=GC RF"): # Reference Annotation Line
return True
if line[:4] == "#=GS": # Description lines - keep if sequence in list.
_, seqname, _ = line.split(maxsplit=2)
return seqname in seqnames
elif line.startswith("#"): # Other markup - filter out
return False
else: # Alignment data - keep if sequence in list.
seqname = line.partition(" ")[0]
return seqname in seqnames
def truncate_stockholm_msa(stockholm_msa: str, max_sequences: int) -> str:
"""Truncates a stockholm file to a maximum number of sequences."""
seqnames = set()
filtered_lines = []
for line in stockholm_msa.splitlines():
if line.strip() and not line.startswith(("#", "//")):
# Ignore blank lines, markup and end symbols - remainder are alignment
# sequence parts.
seqname = line.partition(" ")[0]
seqnames.add(seqname)
if len(seqnames) >= max_sequences:
break
for line in stockholm_msa.splitlines():
if _keep_line(line, seqnames):
filtered_lines.append(line)
return "\n".join(filtered_lines) + "\n"
def remove_empty_columns_from_stockholm_msa(stockholm_msa: str) -> str:
"""Removes empty columns (dashes-only) from a Stockholm MSA."""
processed_lines = {}
unprocessed_lines = {}
for i, line in enumerate(stockholm_msa.splitlines()):
if line.startswith("#=GC RF"):
reference_annotation_i = i
reference_annotation_line = line
# Reached the end of this chunk of the alignment. Process chunk.
_, _, first_alignment = line.rpartition(" ")
mask = []
for j in range(len(first_alignment)):
for _, unprocessed_line in unprocessed_lines.items():
prefix, _, alignment = unprocessed_line.rpartition(" ")
if alignment[j] != "-":
mask.append(True)
break
else: # Every row contained a hyphen - empty column.
mask.append(False)
# Add reference annotation for processing with mask.
unprocessed_lines[reference_annotation_i] = reference_annotation_line
if not any(mask): # All columns were empty. Output empty lines for chunk.
for line_index in unprocessed_lines:
processed_lines[line_index] = ""
else:
for line_index, unprocessed_line in unprocessed_lines.items():
prefix, _, alignment = unprocessed_line.rpartition(" ")
masked_alignment = "".join(itertools.compress(alignment, mask))
processed_lines[line_index] = f"{prefix} {masked_alignment}"
# Clear raw_alignments.
unprocessed_lines = {}
elif line.strip() and not line.startswith(("#", "//")):
unprocessed_lines[i] = line
else:
processed_lines[i] = line
return "\n".join((processed_lines[i] for i in range(len(processed_lines))))
def deduplicate_stockholm_msa(stockholm_msa: str) -> str:
"""Remove duplicate sequences (ignoring insertions wrt query)."""
sequence_dict = collections.defaultdict(str)
# First we must extract all sequences from the MSA.
for line in stockholm_msa.splitlines():
# Only consider the alignments - ignore reference annotation, empty lines,
# descriptions or markup.
if line.strip() and not line.startswith(("#", "//")):
line = line.strip()
seqname, alignment = line.split()
sequence_dict[seqname] += alignment
seen_sequences = set()
seqnames = set()
# First alignment is the query.
query_align = next(iter(sequence_dict.values()))
mask = [c != "-" for c in query_align] # Mask is False for insertions.
for seqname, alignment in sequence_dict.items():
# Apply mask to remove all insertions from the string.
masked_alignment = "".join(itertools.compress(alignment, mask))
if masked_alignment in seen_sequences:
continue
else:
seen_sequences.add(masked_alignment)
seqnames.add(seqname)
filtered_lines = []
for line in stockholm_msa.splitlines():
if _keep_line(line, seqnames):
filtered_lines.append(line)
return "\n".join(filtered_lines) + "\n"
def _get_hhr_line_regex_groups(
regex_pattern: str, line: str
) -> Sequence[Optional[str]]:
match = re.match(regex_pattern, line)
if match is None:
raise RuntimeError(f"Could not parse query line {line}")
return match.groups()
def _update_hhr_residue_indices_list(
sequence: str, start_index: int, indices_list: List[int]
):
"""Computes the relative indices for each residue with respect to the original sequence."""
counter = start_index
for symbol in sequence:
if symbol == "-":
indices_list.append(-1)
else:
indices_list.append(counter)
counter += 1
def _parse_hhr_hit(detailed_lines: Sequence[str]) -> TemplateHit:
"""Parses the detailed HMM HMM comparison section for a single Hit.
This works on .hhr files generated from both HHBlits and HHSearch.
Args:
detailed_lines: A list of lines from a single comparison section between 2
sequences (which each have their own HMM's)
Returns:
A dictionary with the information from that detailed comparison section
Raises:
RuntimeError: If a certain line cannot be processed
"""
# Parse first 2 lines.
number_of_hit = int(detailed_lines[0].split()[-1])
name_hit = detailed_lines[1][1:]
# Parse the summary line.
pattern = (
"Probab=(.*)[\t ]*E-value=(.*)[\t ]*Score=(.*)[\t ]*Aligned_cols=(.*)[\t"
" ]*Identities=(.*)%[\t ]*Similarity=(.*)[\t ]*Sum_probs=(.*)[\t "
"]*Template_Neff=(.*)"
)
match = re.match(pattern, detailed_lines[2])
if match is None:
raise RuntimeError(
"Could not parse section: %s. Expected this: \n%s to contain summary."
% (detailed_lines, detailed_lines[2])
)
(_, _, _, aligned_cols, _, _, sum_probs, _) = [float(x) for x in match.groups()]
# The next section reads the detailed comparisons. These are in a 'human
# readable' format which has a fixed length. The strategy employed is to
# assume that each block starts with the query sequence line, and to parse
# that with a regexp in order to deduce the fixed length used for that block.
query = ""
hit_sequence = ""
indices_query = []
indices_hit = []
length_block = None
for line in detailed_lines[3:]:
# Parse the query sequence line
if (
line.startswith("Q ")
and not line.startswith("Q ss_dssp")
and not line.startswith("Q ss_pred")
and not line.startswith("Q Consensus")
):
# Thus the first 17 characters must be 'Q <query_name> ', and we can parse
# everything after that.
# start sequence end total_sequence_length
patt = r"[\t ]*([0-9]*) ([A-Z-]*)[\t ]*([0-9]*) \([0-9]*\)"
groups = _get_hhr_line_regex_groups(patt, line[17:])
# Get the length of the parsed block using the start and finish indices,
# and ensure it is the same as the actual block length.
start = int(groups[0]) - 1 # Make index zero based.
delta_query = groups[1]
end = int(groups[2])
num_insertions = len([x for x in delta_query if x == "-"])
length_block = end - start + num_insertions
assert length_block == len(delta_query)
# Update the query sequence and indices list.
query += delta_query
_update_hhr_residue_indices_list(delta_query, start, indices_query)
elif line.startswith("T "):
# Parse the hit sequence.
if (
not line.startswith("T ss_dssp")
and not line.startswith("T ss_pred")
and not line.startswith("T Consensus")
):
# Thus the first 17 characters must be 'T <hit_name> ', and we can
# parse everything after that.
# start sequence end total_sequence_length
patt = r"[\t ]*([0-9]*) ([A-Z-]*)[\t ]*[0-9]* \([0-9]*\)"
groups = _get_hhr_line_regex_groups(patt, line[17:])
start = int(groups[0]) - 1 # Make index zero based.
delta_hit_sequence = groups[1]
assert length_block == len(delta_hit_sequence)
# Update the hit sequence and indices list.
hit_sequence += delta_hit_sequence
_update_hhr_residue_indices_list(delta_hit_sequence, start, indices_hit)
return TemplateHit(
index=number_of_hit,
name=name_hit,
aligned_cols=int(aligned_cols),
sum_probs=sum_probs,
query=query,
hit_sequence=hit_sequence,
indices_query=indices_query,
indices_hit=indices_hit,
)
def parse_hhr(hhr_string: str) -> Sequence[TemplateHit]:
"""Parses the content of an entire HHR file."""
lines = hhr_string.splitlines()
# Each .hhr file starts with a results table, then has a sequence of hit
# "paragraphs", each paragraph starting with a line 'No <hit number>'. We
# iterate through each paragraph to parse each hit.
block_starts = [i for i, line in enumerate(lines) if line.startswith("No ")]
hits = []
if block_starts:
block_starts.append(len(lines)) # Add the end of the final block.
for i in range(len(block_starts) - 1):
hits.append(_parse_hhr_hit(lines[block_starts[i] : block_starts[i + 1]]))
return hits
def parse_e_values_from_tblout(tblout: str) -> Dict[str, float]:
"""Parse target to e-value mapping parsed from Jackhmmer tblout string."""
e_values = {"query": 0}
lines = [line for line in tblout.splitlines() if line[0] != "#"]
# As per http://eddylab.org/software/hmmer/Userguide.pdf fields are
# space-delimited. Relevant fields are (1) target name: and
# (5) E-value (full sequence) (numbering from 1).
for line in lines:
fields = line.split()
e_value = fields[4]
target_name = fields[0]
e_values[target_name] = float(e_value)
return e_values
def _get_indices(sequence: str, start: int) -> List[int]:
"""Returns indices for non-gap/insert residues starting at the given index."""
indices = []
counter = start
for symbol in sequence:
# Skip gaps but add a placeholder so that the alignment is preserved.
if symbol == "-":
indices.append(-1)
# Skip deleted residues, but increase the counter.
elif symbol.islower():
counter += 1
# Normal aligned residue. Increase the counter and append to indices.
else:
indices.append(counter)
counter += 1
return indices
@dataclasses.dataclass(frozen=True)
class HitMetadata:
pdb_id: str
chain: str
start: int
end: int
length: int
text: str
def _parse_hmmsearch_description(description: str) -> HitMetadata:
"""Parses the hmmsearch A3M sequence description line."""
# Example 1: >4pqx_A/2-217 [subseq from] mol:protein length:217 Free text
# Example 2: >5g3r_A/1-55 [subseq from] mol:protein length:352
match = re.match(
r"^>?([a-z0-9]+)_(\w+)/([0-9]+)-([0-9]+).*protein length:([0-9]+) *(.*)$",
description.strip(),
)
if not match:
raise ValueError(f'Could not parse description: "{description}".')
return HitMetadata(
pdb_id=match[1],
chain=match[2],
start=int(match[3]),
end=int(match[4]),
length=int(match[5]),
text=match[6],
)
def parse_hmmsearch_a3m(
query_sequence: str, a3m_string: str, skip_first: bool = True
) -> Sequence[TemplateHit]:
"""Parses an a3m string produced by hmmsearch.
Args:
query_sequence: The query sequence.
a3m_string: The a3m string produced by hmmsearch.
skip_first: Whether to skip the first sequence in the a3m string.
Returns:
A sequence of `TemplateHit` results.
"""
# Zip the descriptions and MSAs together, skip the first query sequence.
parsed_a3m = list(zip(*parse_fasta(a3m_string)))
if skip_first:
parsed_a3m = parsed_a3m[1:]
indices_query = _get_indices(query_sequence, start=0)
hits = []
for i, (hit_sequence, hit_description) in enumerate(parsed_a3m, start=1):
if "mol:protein" not in hit_description:
continue # Skip non-protein chains.
metadata = _parse_hmmsearch_description(hit_description)
# Aligned columns are only the match states.
aligned_cols = sum([r.isupper() and r != "-" for r in hit_sequence])
indices_hit = _get_indices(hit_sequence, start=metadata.start - 1)
hit = TemplateHit(
index=i,
name=f"{metadata.pdb_id}_{metadata.chain}",
aligned_cols=aligned_cols,
sum_probs=None,
query=query_sequence,
hit_sequence=hit_sequence.upper(),
indices_query=indices_query,
indices_hit=indices_hit,
)
hits.append(hit)
return hits
unifold/msa/pipeline.py 0 → 100755
View file @ a1c29028
# 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 unifold model."""
import os
from typing import Any, Mapping, MutableMapping, Optional, Sequence, Union
from absl import logging
from unifold.data import residue_constants
from unifold.msa import msa_identifiers
from unifold.msa import parsers
from unifold.msa import templates
from unifold.msa.tools import hhblits
from unifold.msa.tools import hhsearch
from unifold.msa.tools import hmmsearch
from unifold.msa.tools import jackhmmer
import numpy as np
FeatureDict = MutableMapping[str, np.ndarray]
TemplateSearcher = Union[hhsearch.HHSearch, hmmsearch.Hmmsearch]
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
def make_msa_features(msas: Sequence[parsers.Msa]) -> FeatureDict:
"""Constructs a feature dict of MSA features."""
if not msas:
raise ValueError("At least one MSA must be provided.")
int_msa = []
deletion_matrix = []
species_ids = []
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.")
for sequence_index, sequence in enumerate(msa.sequences):
if sequence in seen_sequences:
continue
seen_sequences.add(sequence)
int_msa.append(
[residue_constants.HHBLITS_AA_TO_ID[res] for res in sequence]
)
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])
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)
features["msa_species_identifiers"] = np.array(species_ids, dtype=np.object_)
return features
def run_msa_tool(
msa_runner,
input_fasta_path: str,
msa_out_path: str,
msa_format: str,
use_precomputed_msas: bool,
) -> 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):
result = msa_runner.query(input_fasta_path)[0]
with open(msa_out_path, "w") as f:
f.write(result[msa_format])
else:
logging.warning("Reading MSA from file %s", msa_out_path)
with open(msa_out_path, "r") as f:
result = {msa_format: f.read()}
return result
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,
bfd_database_path: Optional[str],
uniclust30_database_path: Optional[str],
small_bfd_database_path: Optional[str],
uniprot_database_path: Optional[str],
template_searcher: TemplateSearcher,
template_featurizer: templates.TemplateHitFeaturizer,
use_small_bfd: bool,
mgnify_max_hits: int = 501,
uniref_max_hits: int = 10000,
use_precomputed_msas: bool = False,
):
"""Initializes the data pipeline."""
self._use_small_bfd = use_small_bfd
self.jackhmmer_uniref90_runner = jackhmmer.Jackhmmer(
binary_path=jackhmmer_binary_path, database_path=uniref90_database_path
)
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],
)
self.jackhmmer_mgnify_runner = jackhmmer.Jackhmmer(
binary_path=jackhmmer_binary_path, database_path=mgnify_database_path
)
self.jackhmmer_uniprot_runner = jackhmmer.Jackhmmer(
binary_path=jackhmmer_binary_path, database_path=uniprot_database_path
)
self.template_searcher = template_searcher
self.template_featurizer = template_featurizer
self.mgnify_max_hits = mgnify_max_hits
self.uniref_max_hits = uniref_max_hits
self.use_precomputed_msas = use_precomputed_msas
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")
jackhmmer_uniref90_result = run_msa_tool(
self.jackhmmer_uniref90_runner,
input_fasta_path,
uniref90_out_path,
"sto",
self.use_precomputed_msas,
)
mgnify_out_path = os.path.join(msa_output_dir, "mgnify_hits.sto")
jackhmmer_mgnify_result = run_msa_tool(
self.jackhmmer_mgnify_runner,
input_fasta_path,
mgnify_out_path,
"sto",
self.use_precomputed_msas,
)
msa_for_templates = jackhmmer_uniref90_result["sto"]
msa_for_templates = parsers.truncate_stockholm_msa(
msa_for_templates, max_sequences=self.uniref_max_hits
)
msa_for_templates = parsers.deduplicate_stockholm_msa(msa_for_templates)
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}"
)
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)
uniref90_msa = parsers.parse_stockholm(jackhmmer_uniref90_result["sto"])
uniref90_msa = uniref90_msa.truncate(max_seqs=self.uniref_max_hits)
mgnify_msa = parsers.parse_stockholm(jackhmmer_mgnify_result["sto"])
mgnify_msa = mgnify_msa.truncate(max_seqs=self.mgnify_max_hits)
pdb_template_hits = self.template_searcher.get_template_hits(
output_string=pdb_templates_result, input_sequence=input_sequence
)
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(
self.jackhmmer_small_bfd_runner,
input_fasta_path,
bfd_out_path,
"sto",
self.use_precomputed_msas,
)
bfd_msa = parsers.parse_stockholm(jackhmmer_small_bfd_result["sto"])
else:
bfd_out_path = os.path.join(msa_output_dir, "bfd_uniclust_hits.a3m")
hhblits_bfd_uniclust_result = run_msa_tool(
self.hhblits_bfd_uniclust_runner,
input_fasta_path,
bfd_out_path,
"a3m",
self.use_precomputed_msas,
)
bfd_msa = parsers.parse_a3m(hhblits_bfd_uniclust_result["a3m"])
templates_result = self.template_featurizer.get_templates(
query_sequence=input_sequence, hits=pdb_template_hits
)
sequence_features = make_sequence_features(
sequence=input_sequence, description=input_description, num_res=num_res
)
msa_features = make_msa_features((uniref90_msa, bfd_msa, mgnify_msa))
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],
)
return {**sequence_features, **msa_features, **templates_result.features}
def process_uniprot(
self, input_fasta_path: str, msa_output_dir: str
) -> FeatureDict:
uniprot_path = os.path.join(msa_output_dir, "uniprot_hits.sto")
uniprot_result = run_msa_tool(
self.jackhmmer_uniprot_runner,
input_fasta_path,
uniprot_path,
"sto",
self.use_precomputed_msas,
)
msa = parsers.parse_stockholm(uniprot_result["sto"])
msa = msa.truncate(max_seqs=50000)
all_seq_dict = make_msa_features([msa])
return all_seq_dict
unifold/msa/templates.py 0 → 100755
View file @ a1c29028
# 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 getting templates and calculating template features."""
import abc
import dataclasses
import datetime
import functools
import glob
import os
import re
from typing import Any, Dict, Mapping, Optional, Sequence, Tuple
from absl import logging
from unifold.data import residue_constants
from unifold.msa import mmcif, parsers
from unifold.msa.tools import kalign
import numpy as np
class Error(Exception):
"""Base class for exceptions."""
class NoChainsError(Error):
"""An error indicating that template mmCIF didn't have any chains."""
class SequenceNotInTemplateError(Error):
"""An error indicating that template mmCIF didn't contain the sequence."""
class NoAtomDataInTemplateError(Error):
"""An error indicating that template mmCIF didn't contain atom positions."""
class TemplateAtomMaskAllZerosError(Error):
"""An error indicating that template mmCIF had all atom positions masked."""
class QueryToTemplateAlignError(Error):
"""An error indicating that the query can't be aligned to the template."""
class CaDistanceError(Error):
"""An error indicating that a CA atom distance exceeds a threshold."""
class MultipleChainsError(Error):
"""An error indicating that multiple chains were found for a given ID."""
# Prefilter exceptions.
class PrefilterError(Exception):
"""A base class for template prefilter exceptions."""
class DateError(PrefilterError):
"""An error indicating that the hit date was after the max allowed date."""
class AlignRatioError(PrefilterError):
"""An error indicating that the hit align ratio to the query was too small."""
class DuplicateError(PrefilterError):
"""An error indicating that the hit was an exact subsequence of the query."""
class LengthError(PrefilterError):
"""An error indicating that the hit was too short."""
TEMPLATE_FEATURES = {
"template_aatype": np.float32,
"template_all_atom_mask": np.float32,
"template_all_atom_positions": np.float32,
"template_domain_names": np.object_,
"template_sequence": np.object_,
"template_sum_probs": np.float32,
}
def _get_pdb_id_and_chain(hit: parsers.TemplateHit) -> Tuple[str, str]:
"""Returns PDB id and chain id for an HHSearch Hit."""
# PDB ID: 4 letters. Chain ID: 1+ alphanumeric letters or "." if unknown.
id_match = re.match(r"[a-zA-Z\d]{4}_[a-zA-Z0-9.]+", hit.name)
if not id_match:
raise ValueError(f"hit.name did not start with PDBID_chain: {hit.name}")
pdb_id, chain_id = id_match.group(0).split("_")
return pdb_id.lower(), chain_id
def _is_after_cutoff(
pdb_id: str,
release_dates: Mapping[str, datetime.datetime],
release_date_cutoff: Optional[datetime.datetime],
) -> bool:
"""Checks if the template date is after the release date cutoff.
Args:
pdb_id: 4 letter pdb code.
release_dates: Dictionary mapping PDB ids to their structure release dates.
release_date_cutoff: Max release date that is valid for this query.
Returns:
True if the template release date is after the cutoff, False otherwise.
"""
if release_date_cutoff is None:
raise ValueError("The release_date_cutoff must not be None.")
if pdb_id in release_dates:
return release_dates[pdb_id] > release_date_cutoff
else:
# Since this is just a quick prefilter to reduce the number of mmCIF files
# we need to parse, we don't have to worry about returning True here.
return False
def _parse_obsolete(obsolete_file_path: str) -> Mapping[str, Optional[str]]:
"""Parses the data file from PDB that lists which pdb_ids are obsolete."""
with open(obsolete_file_path) as f:
result = {}
for line in f:
line = line.strip()
# Format: Date From To
# 'OBSLTE 06-NOV-19 6G9Y' - Removed, rare
# 'OBSLTE 31-JUL-94 116L 216L' - Replaced, common
# 'OBSLTE 26-SEP-06 2H33 2JM5 2OWI' - Replaced by multiple, rare
if line.startswith("OBSLTE"):
if len(line) > 30:
# Replaced by at least one structure.
from_id = line[20:24].lower()
to_id = line[29:33].lower()
result[from_id] = to_id
elif len(line) == 24:
# Removed.
from_id = line[20:24].lower()
result[from_id] = None
return result
def _parse_release_dates(path: str) -> Mapping[str, datetime.datetime]:
"""Parses release dates file, returns a mapping from PDBs to release dates."""
if path.endswith("txt"):
release_dates = {}
with open(path, "r") as f:
for line in f:
pdb_id, date = line.split(":")
date = date.strip()
# Python 3.6 doesn't have datetime.date.fromisoformat() which is about
# 90x faster than strptime. However, splitting the string manually is
# about 10x faster than strptime.
release_dates[pdb_id.strip()] = datetime.datetime(
year=int(date[:4]), month=int(date[5:7]), day=int(date[8:10])
)
return release_dates
else:
raise ValueError("Invalid format of the release date file %s." % path)
def _assess_hhsearch_hit(
hit: parsers.TemplateHit,
hit_pdb_code: str,
query_sequence: str,
release_dates: Mapping[str, datetime.datetime],
release_date_cutoff: datetime.datetime,
max_subsequence_ratio: float = 0.95,
min_align_ratio: float = 0.1,
) -> bool:
"""Determines if template is valid (without parsing the template mmcif file).
Args:
hit: HhrHit for the template.
hit_pdb_code: The 4 letter pdb code of the template hit. This might be
different from the value in the actual hit since the original pdb might
have become obsolete.
query_sequence: Amino acid sequence of the query.
release_dates: Dictionary mapping pdb codes to their structure release
dates.
release_date_cutoff: Max release date that is valid for this query.
max_subsequence_ratio: Exclude any exact matches with this much overlap.
min_align_ratio: Minimum overlap between the template and query.
Returns:
True if the hit passed the prefilter. Raises an exception otherwise.
Raises:
DateError: If the hit date was after the max allowed date.
AlignRatioError: If the hit align ratio to the query was too small.
DuplicateError: If the hit was an exact subsequence of the query.
LengthError: If the hit was too short.
"""
aligned_cols = hit.aligned_cols
align_ratio = aligned_cols / len(query_sequence)
template_sequence = hit.hit_sequence.replace("-", "")
length_ratio = float(len(template_sequence)) / len(query_sequence)
# Check whether the template is a large subsequence or duplicate of original
# query. This can happen due to duplicate entries in the PDB database.
duplicate = (
template_sequence in query_sequence and length_ratio > max_subsequence_ratio
)
if _is_after_cutoff(hit_pdb_code, release_dates, release_date_cutoff):
raise DateError(
f"Date ({release_dates[hit_pdb_code]}) > max template date "
f"({release_date_cutoff})."
)
if align_ratio <= min_align_ratio:
raise AlignRatioError(
"Proportion of residues aligned to query too small. "
f"Align ratio: {align_ratio}."
)
if duplicate:
raise DuplicateError(
"Template is an exact subsequence of query with large "
f"coverage. Length ratio: {length_ratio}."
)
if len(template_sequence) < 10:
raise LengthError(f"Template too short. Length: {len(template_sequence)}.")
return True
def _find_template_in_pdb(
template_chain_id: str, template_sequence: str, mmcif_object: mmcif.MmcifObject
) -> Tuple[str, str, int]:
"""Tries to find the template chain in the given pdb file.
This method tries the three following things in order:
1. Tries if there is an exact match in both the chain ID and the sequence.
If yes, the chain sequence is returned. Otherwise:
2. Tries if there is an exact match only in the sequence.
If yes, the chain sequence is returned. Otherwise:
3. Tries if there is a fuzzy match (X = wildcard) in the sequence.
If yes, the chain sequence is returned.
If none of these succeed, a SequenceNotInTemplateError is thrown.
Args:
template_chain_id: The template chain ID.
template_sequence: The template chain sequence.
mmcif_object: The PDB object to search for the template in.
Returns:
A tuple with:
* The chain sequence that was found to match the template in the PDB object.
* The ID of the chain that is being returned.
* The offset where the template sequence starts in the chain sequence.
Raises:
SequenceNotInTemplateError: If no match is found after the steps described
above.
"""
# Try if there is an exact match in both the chain ID and the (sub)sequence.
pdb_id = mmcif_object.file_id
chain_sequence = mmcif_object.chain_to_seqres.get(template_chain_id)
if chain_sequence and (template_sequence in chain_sequence):
logging.info("Found an exact template match %s_%s.", pdb_id, template_chain_id)
mapping_offset = chain_sequence.find(template_sequence)
return chain_sequence, template_chain_id, mapping_offset
# Try if there is an exact match in the (sub)sequence only.
for chain_id, chain_sequence in mmcif_object.chain_to_seqres.items():
if chain_sequence and (template_sequence in chain_sequence):
logging.info("Found a sequence-only match %s_%s.", pdb_id, chain_id)
mapping_offset = chain_sequence.find(template_sequence)
return chain_sequence, chain_id, mapping_offset
# Return a chain sequence that fuzzy matches (X = wildcard) the template.
# Make parentheses unnamed groups (?:_) to avoid the 100 named groups limit.
regex = ["." if aa == "X" else "(?:%s|X)" % aa for aa in template_sequence]
regex = re.compile("".join(regex))
for chain_id, chain_sequence in mmcif_object.chain_to_seqres.items():
match = re.search(regex, chain_sequence)
if match:
logging.info("Found a fuzzy sequence-only match %s_%s.", pdb_id, chain_id)
mapping_offset = match.start()
return chain_sequence, chain_id, mapping_offset
# No hits, raise an error.
raise SequenceNotInTemplateError(
"Could not find the template sequence in %s_%s. Template sequence: %s, "
"chain_to_seqres: %s"
% (pdb_id, template_chain_id, template_sequence, mmcif_object.chain_to_seqres)
)
def _realign_pdb_template_to_query(
old_template_sequence: str,
template_chain_id: str,
mmcif_object: mmcif.MmcifObject,
old_mapping: Mapping[int, int],
kalign_binary_path: str,
) -> Tuple[str, Mapping[int, int]]:
"""Aligns template from the mmcif_object to the query.
In case PDB70 contains a different version of the template sequence, we need
to perform a realignment to the actual sequence that is in the mmCIF file.
This method performs such realignment, but returns the new sequence and
mapping only if the sequence in the mmCIF file is 90% identical to the old
sequence.
Note that the old_template_sequence comes from the hit, and contains only that
part of the chain that matches with the query while the new_template_sequence
is the full chain.
Args:
old_template_sequence: The template sequence that was returned by the PDB
template search (typically done using HHSearch).
template_chain_id: The template chain id was returned by the PDB template
search (typically done using HHSearch). This is used to find the right
chain in the mmcif_object chain_to_seqres mapping.
mmcif_object: A mmcif_object which holds the actual template data.
old_mapping: A mapping from the query sequence to the template sequence.
This mapping will be used to compute the new mapping from the query
sequence to the actual mmcif_object template sequence by aligning the
old_template_sequence and the actual template sequence.
kalign_binary_path: The path to a kalign executable.
Returns:
A tuple (new_template_sequence, new_query_to_template_mapping) where:
* new_template_sequence is the actual template sequence that was found in
the mmcif_object.
* new_query_to_template_mapping is the new mapping from the query to the
actual template found in the mmcif_object.
Raises:
QueryToTemplateAlignError:
* If there was an error thrown by the alignment tool.
* Or if the actual template sequence differs by more than 10% from the
old_template_sequence.
"""
aligner = kalign.Kalign(binary_path=kalign_binary_path)
new_template_sequence = mmcif_object.chain_to_seqres.get(template_chain_id, "")
# Sometimes the template chain id is unknown. But if there is only a single
# sequence within the mmcif_object, it is safe to assume it is that one.
if not new_template_sequence:
if len(mmcif_object.chain_to_seqres) == 1:
logging.info(
"Could not find %s in %s, but there is only 1 sequence, so "
"using that one.",
template_chain_id,
mmcif_object.file_id,
)
new_template_sequence = list(mmcif_object.chain_to_seqres.values())[0]
else:
raise QueryToTemplateAlignError(
f"Could not find chain {template_chain_id} in {mmcif_object.file_id}. "
"If there are no mmCIF parsing errors, it is possible it was not a "
"protein chain."
)
try:
parsed_a3m = parsers.parse_a3m(
aligner.align([old_template_sequence, new_template_sequence])
)
old_aligned_template, new_aligned_template = parsed_a3m.sequences
except Exception as e:
raise QueryToTemplateAlignError(
"Could not align old template %s to template %s (%s_%s). Error: %s"
% (
old_template_sequence,
new_template_sequence,
mmcif_object.file_id,
template_chain_id,
str(e),
)
)
logging.info(
"Old aligned template: %s\nNew aligned template: %s",
old_aligned_template,
new_aligned_template,
)
old_to_new_template_mapping = {}
old_template_index = -1
new_template_index = -1
num_same = 0
for old_template_aa, new_template_aa in zip(
old_aligned_template, new_aligned_template
):
if old_template_aa != "-":
old_template_index += 1
if new_template_aa != "-":
new_template_index += 1
if old_template_aa != "-" and new_template_aa != "-":
old_to_new_template_mapping[old_template_index] = new_template_index
if old_template_aa == new_template_aa:
num_same += 1
# Require at least 90 % sequence identity wrt to the shorter of the sequences.
if (
float(num_same) / min(len(old_template_sequence), len(new_template_sequence))
< 0.9
):
raise QueryToTemplateAlignError(
"Insufficient similarity of the sequence in the database: %s to the "
"actual sequence in the mmCIF file %s_%s: %s. We require at least "
"90 %% similarity wrt to the shorter of the sequences. This is not a "
"problem unless you think this is a template that should be included."
% (
old_template_sequence,
mmcif_object.file_id,
template_chain_id,
new_template_sequence,
)
)
new_query_to_template_mapping = {}
for query_index, old_template_index in old_mapping.items():
new_query_to_template_mapping[query_index] = old_to_new_template_mapping.get(
old_template_index, -1
)
new_template_sequence = new_template_sequence.replace("-", "")
return new_template_sequence, new_query_to_template_mapping
def _check_residue_distances(
all_positions: np.ndarray, all_positions_mask: np.ndarray, max_ca_ca_distance: float
):
"""Checks if the distance between unmasked neighbor residues is ok."""
ca_position = residue_constants.atom_order["CA"]
prev_is_unmasked = False
prev_calpha = None
for i, (coords, mask) in enumerate(zip(all_positions, all_positions_mask)):
this_is_unmasked = bool(mask[ca_position])
if this_is_unmasked:
this_calpha = coords[ca_position]
if prev_is_unmasked:
distance = np.linalg.norm(this_calpha - prev_calpha)
if distance > max_ca_ca_distance:
raise CaDistanceError(
"The distance between residues %d and %d is %f > limit %f."
% (i, i + 1, distance, max_ca_ca_distance)
)
prev_calpha = this_calpha
prev_is_unmasked = this_is_unmasked
def _get_atom_positions(
mmcif_object: mmcif.MmcifObject, auth_chain_id: str, max_ca_ca_distance: float
) -> Tuple[np.ndarray, np.ndarray]:
"""Gets atom positions and mask from a list of Biopython Residues."""
num_res = len(mmcif_object.chain_to_seqres[auth_chain_id])
relevant_chains = [
c for c in mmcif_object.structure.get_chains() if c.id == auth_chain_id
]
if len(relevant_chains) != 1:
raise MultipleChainsError(
f"Expected exactly one chain in structure with id {auth_chain_id}."
)
chain = relevant_chains[0]
all_positions = np.zeros([num_res, residue_constants.atom_type_num, 3])
all_positions_mask = np.zeros(
[num_res, residue_constants.atom_type_num], dtype=np.int64
)
for res_index in range(num_res):
pos = np.zeros([residue_constants.atom_type_num, 3], dtype=np.float32)
mask = np.zeros([residue_constants.atom_type_num], dtype=np.float32)
res_at_position = mmcif_object.seqres_to_structure[auth_chain_id][res_index]
if not res_at_position.is_missing:
res = chain[
(
res_at_position.hetflag,
res_at_position.position.residue_number,
res_at_position.position.insertion_code,
)
]
for atom in res.get_atoms():
atom_name = atom.get_name()
x, y, z = atom.get_coord()
if atom_name in residue_constants.atom_order.keys():
pos[residue_constants.atom_order[atom_name]] = [x, y, z]
mask[residue_constants.atom_order[atom_name]] = 1.0
elif atom_name.upper() == "SE" and res.get_resname() == "MSE":
# Put the coordinates of the selenium atom in the sulphur column.
pos[residue_constants.atom_order["SD"]] = [x, y, z]
mask[residue_constants.atom_order["SD"]] = 1.0
# Fix naming errors in arginine residues where NH2 is incorrectly
# assigned to be closer to CD than NH1.
cd = residue_constants.atom_order["CD"]
nh1 = residue_constants.atom_order["NH1"]
nh2 = residue_constants.atom_order["NH2"]
if (
res.get_resname() == "ARG"
and all(mask[atom_index] for atom_index in (cd, nh1, nh2))
and (
np.linalg.norm(pos[nh1] - pos[cd])
> np.linalg.norm(pos[nh2] - pos[cd])
)
):
pos[nh1], pos[nh2] = pos[nh2].copy(), pos[nh1].copy()
mask[nh1], mask[nh2] = mask[nh2].copy(), mask[nh1].copy()
all_positions[res_index] = pos
all_positions_mask[res_index] = mask
_check_residue_distances(all_positions, all_positions_mask, max_ca_ca_distance)
return all_positions, all_positions_mask
def _extract_template_features(
mmcif_object: mmcif.MmcifObject,
pdb_id: str,
mapping: Mapping[int, int],
template_sequence: str,
query_sequence: str,
template_chain_id: str,
kalign_binary_path: str,
) -> Tuple[Dict[str, Any], Optional[str]]:
"""Parses atom positions in the target structure and aligns with the query.
Atoms for each residue in the template structure are indexed to coincide
with their corresponding residue in the query sequence, according to the
alignment mapping provided.
Args:
mmcif_object: mmcif_parsing.MmcifObject representing the template.
pdb_id: PDB code for the template.
mapping: Dictionary mapping indices in the query sequence to indices in
the template sequence.
template_sequence: String describing the amino acid sequence for the
template protein.
query_sequence: String describing the amino acid sequence for the query
protein.
template_chain_id: String ID describing which chain in the structure proto
should be used.
kalign_binary_path: The path to a kalign executable used for template
realignment.
Returns:
A tuple with:
* A dictionary containing the extra features derived from the template
protein structure.
* A warning message if the hit was realigned to the actual mmCIF sequence.
Otherwise None.
Raises:
NoChainsError: If the mmcif object doesn't contain any chains.
SequenceNotInTemplateError: If the given chain id / sequence can't
be found in the mmcif object.
QueryToTemplateAlignError: If the actual template in the mmCIF file
can't be aligned to the query.
NoAtomDataInTemplateError: If the mmcif object doesn't contain
atom positions.
TemplateAtomMaskAllZerosError: If the mmcif object doesn't have any
unmasked residues.
"""
if mmcif_object is None or not mmcif_object.chain_to_seqres:
raise NoChainsError("No chains in PDB: %s_%s" % (pdb_id, template_chain_id))
warning = None
try:
seqres, chain_id, mapping_offset = _find_template_in_pdb(
template_chain_id=template_chain_id,
template_sequence=template_sequence,
mmcif_object=mmcif_object,
)
except SequenceNotInTemplateError:
# If PDB70 contains a different version of the template, we use the sequence
# from the mmcif_object.
chain_id = template_chain_id
warning = (
f"The exact sequence {template_sequence} was not found in "
f"{pdb_id}_{chain_id}. Realigning the template to the actual sequence."
)
logging.warning(warning)
# This throws an exception if it fails to realign the hit.
seqres, mapping = _realign_pdb_template_to_query(
old_template_sequence=template_sequence,
template_chain_id=template_chain_id,
mmcif_object=mmcif_object,
old_mapping=mapping,
kalign_binary_path=kalign_binary_path,
)
logging.info(
"Sequence in %s_%s: %s successfully realigned to %s",
pdb_id,
chain_id,
template_sequence,
seqres,
)
# The template sequence changed.
template_sequence = seqres
# No mapping offset, the query is aligned to the actual sequence.
mapping_offset = 0
try:
# Essentially set to infinity - we don't want to reject templates unless
# they're really really bad.
all_atom_positions, all_atom_mask = _get_atom_positions(
mmcif_object, chain_id, max_ca_ca_distance=150.0
)
except (CaDistanceError, KeyError) as ex:
raise NoAtomDataInTemplateError(
"Could not get atom data (%s_%s): %s" % (pdb_id, chain_id, str(ex))
) from ex
all_atom_positions = np.split(all_atom_positions, all_atom_positions.shape[0])
all_atom_masks = np.split(all_atom_mask, all_atom_mask.shape[0])
output_templates_sequence = []
templates_all_atom_positions = []
templates_all_atom_masks = []
for _ in query_sequence:
# Residues in the query_sequence that are not in the template_sequence:
templates_all_atom_positions.append(
np.zeros((residue_constants.atom_type_num, 3))
)
templates_all_atom_masks.append(np.zeros(residue_constants.atom_type_num))
output_templates_sequence.append("-")
for k, v in mapping.items():
template_index = v + mapping_offset
templates_all_atom_positions[k] = all_atom_positions[template_index][0]
templates_all_atom_masks[k] = all_atom_masks[template_index][0]
output_templates_sequence[k] = template_sequence[v]
# Alanine (AA with the lowest number of atoms) has 5 atoms (C, CA, CB, N, O).
if np.sum(templates_all_atom_masks) < 5:
raise TemplateAtomMaskAllZerosError(
"Template all atom mask was all zeros: %s_%s. Residue range: %d-%d"
% (
pdb_id,
chain_id,
min(mapping.values()) + mapping_offset,
max(mapping.values()) + mapping_offset,
)
)
output_templates_sequence = "".join(output_templates_sequence)
templates_aatype = residue_constants.sequence_to_onehot(
output_templates_sequence, residue_constants.HHBLITS_AA_TO_ID
)
return (
{
"template_all_atom_positions": np.array(templates_all_atom_positions),
"template_all_atom_mask": np.array(templates_all_atom_masks),
"template_sequence": output_templates_sequence.encode(),
"template_aatype": np.array(templates_aatype),
"template_domain_names": f"{pdb_id.lower()}_{chain_id}".encode(),
},
warning,
)
def _build_query_to_hit_index_mapping(
hit_query_sequence: str,
hit_sequence: str,
indices_hit: Sequence[int],
indices_query: Sequence[int],
original_query_sequence: str,
) -> Mapping[int, int]:
"""Gets mapping from indices in original query sequence to indices in the hit.
hit_query_sequence and hit_sequence are two aligned sequences containing gap
characters. hit_query_sequence contains only the part of the original query
sequence that matched the hit. When interpreting the indices from the .hhr, we
need to correct for this to recover a mapping from original query sequence to
the hit sequence.
Args:
hit_query_sequence: The portion of the query sequence that is in the .hhr
hit
hit_sequence: The portion of the hit sequence that is in the .hhr
indices_hit: The indices for each aminoacid relative to the hit sequence
indices_query: The indices for each aminoacid relative to the original query
sequence
original_query_sequence: String describing the original query sequence.
Returns:
Dictionary with indices in the original query sequence as keys and indices
in the hit sequence as values.
"""
# If the hit is empty (no aligned residues), return empty mapping
if not hit_query_sequence:
return {}
# Remove gaps and find the offset of hit.query relative to original query.
hhsearch_query_sequence = hit_query_sequence.replace("-", "")
hit_sequence = hit_sequence.replace("-", "")
hhsearch_query_offset = original_query_sequence.find(hhsearch_query_sequence)
# Index of -1 used for gap characters. Subtract the min index ignoring gaps.
min_idx = min(x for x in indices_hit if x > -1)
fixed_indices_hit = [x - min_idx if x > -1 else -1 for x in indices_hit]
min_idx = min(x for x in indices_query if x > -1)
fixed_indices_query = [x - min_idx if x > -1 else -1 for x in indices_query]
# Zip the corrected indices, ignore case where both seqs have gap characters.
mapping = {}
for q_i, q_t in zip(fixed_indices_query, fixed_indices_hit):
if q_t != -1 and q_i != -1:
if q_t >= len(hit_sequence) or q_i + hhsearch_query_offset >= len(
original_query_sequence
):
continue
mapping[q_i + hhsearch_query_offset] = q_t
return mapping
@dataclasses.dataclass(frozen=True)
class SingleHitResult:
features: Optional[Mapping[str, Any]]
error: Optional[str]
warning: Optional[str]
@functools.lru_cache(16, typed=False)
def _read_file(path):
with open(path, "r") as f:
file_data = f.read()
return file_data
def _process_single_hit(
query_sequence: str,
hit: parsers.TemplateHit,
mmcif_dir: str,
max_template_date: datetime.datetime,
release_dates: Mapping[str, datetime.datetime],
obsolete_pdbs: Mapping[str, Optional[str]],
kalign_binary_path: str,
strict_error_check: bool = False,
) -> SingleHitResult:
"""Tries to extract template features from a single HHSearch hit."""
# Fail hard if we can't get the PDB ID and chain name from the hit.
hit_pdb_code, hit_chain_id = _get_pdb_id_and_chain(hit)
# This hit has been removed (obsoleted) from PDB, skip it.
if hit_pdb_code in obsolete_pdbs and obsolete_pdbs[hit_pdb_code] is None:
return SingleHitResult(
features=None, error=None, warning=f"Hit {hit_pdb_code} is obsolete."
)
if hit_pdb_code not in release_dates:
if hit_pdb_code in obsolete_pdbs:
hit_pdb_code = obsolete_pdbs[hit_pdb_code]
# Pass hit_pdb_code since it might have changed due to the pdb being obsolete.
try:
_assess_hhsearch_hit(
hit=hit,
hit_pdb_code=hit_pdb_code,
query_sequence=query_sequence,
release_dates=release_dates,
release_date_cutoff=max_template_date,
)
except PrefilterError as e:
msg = f"hit {hit_pdb_code}_{hit_chain_id} did not pass prefilter: {str(e)}"
logging.info(msg)
if strict_error_check and isinstance(e, (DateError, DuplicateError)):
# In strict mode we treat some prefilter cases as errors.
return SingleHitResult(features=None, error=msg, warning=None)
return SingleHitResult(features=None, error=None, warning=None)
mapping = _build_query_to_hit_index_mapping(
hit.query, hit.hit_sequence, hit.indices_hit, hit.indices_query, query_sequence
)
# The mapping is from the query to the actual hit sequence, so we need to
# remove gaps (which regardless have a missing confidence score).
template_sequence = hit.hit_sequence.replace("-", "")
cif_path = os.path.join(mmcif_dir, hit_pdb_code + ".cif")
logging.debug(
"Reading PDB entry from %s. Query: %s, template: %s",
cif_path,
query_sequence,
template_sequence,
)
# Fail if we can't find the mmCIF file.
cif_string = _read_file(cif_path)
parsing_result = mmcif.parse(file_id=hit_pdb_code, mmcif_string=cif_string)
if parsing_result.mmcif_object is not None:
hit_release_date = datetime.datetime.strptime(
parsing_result.mmcif_object.header["release_date"], "%Y-%m-%d"
)
if hit_release_date > max_template_date:
error = "Template %s date (%s) > max template date (%s)." % (
hit_pdb_code,
hit_release_date,
max_template_date,
)
if strict_error_check:
return SingleHitResult(features=None, error=error, warning=None)
else:
logging.debug(error)
return SingleHitResult(features=None, error=None, warning=None)
try:
features, realign_warning = _extract_template_features(
mmcif_object=parsing_result.mmcif_object,
pdb_id=hit_pdb_code,
mapping=mapping,
template_sequence=template_sequence,
query_sequence=query_sequence,
template_chain_id=hit_chain_id,
kalign_binary_path=kalign_binary_path,
)
if hit.sum_probs is None:
features["template_sum_probs"] = [0]
else:
features["template_sum_probs"] = [hit.sum_probs]
# It is possible there were some errors when parsing the other chains in the
# mmCIF file, but the template features for the chain we want were still
# computed. In such case the mmCIF parsing errors are not relevant.
return SingleHitResult(features=features, error=None, warning=realign_warning)
except (
NoChainsError,
NoAtomDataInTemplateError,
TemplateAtomMaskAllZerosError,
) as e:
# These 3 errors indicate missing mmCIF experimental data rather than a
# problem with the template search, so turn them into warnings.
warning = (
"%s_%s (sum_probs: %s, rank: %s): feature extracting errors: "
"%s, mmCIF parsing errors: %s"
% (
hit_pdb_code,
hit_chain_id,
hit.sum_probs,
hit.index,
str(e),
parsing_result.errors,
)
)
if strict_error_check:
return SingleHitResult(features=None, error=warning, warning=None)
else:
return SingleHitResult(features=None, error=None, warning=warning)
except Error as e:
error = (
"%s_%s (sum_probs: %.2f, rank: %d): feature extracting errors: "
"%s, mmCIF parsing errors: %s"
% (
hit_pdb_code,
hit_chain_id,
hit.sum_probs,
hit.index,
str(e),
parsing_result.errors,
)
)
return SingleHitResult(features=None, error=error, warning=None)
@dataclasses.dataclass(frozen=True)
class TemplateSearchResult:
features: Mapping[str, Any]
errors: Sequence[str]
warnings: Sequence[str]
class TemplateHitFeaturizer(abc.ABC):
"""An abstract base class for turning template hits to template features."""
def __init__(
self,
mmcif_dir: str,
max_template_date: str,
max_hits: int,
kalign_binary_path: str,
release_dates_path: Optional[str],
obsolete_pdbs_path: Optional[str],
strict_error_check: bool = False,
):
"""Initializes the Template Search.
Args:
mmcif_dir: Path to a directory with mmCIF structures. Once a template ID
is found by HHSearch, this directory is used to retrieve the template
data.
max_template_date: The maximum date permitted for template structures. No
template with date higher than this date will be returned. In ISO8601
date format, YYYY-MM-DD.
max_hits: The maximum number of templates that will be returned.
kalign_binary_path: The path to a kalign executable used for template
realignment.
release_dates_path: An optional path to a file with a mapping from PDB IDs
to their release dates. Thanks to this we don't have to redundantly
parse mmCIF files to get that information.
obsolete_pdbs_path: An optional path to a file containing a mapping from
obsolete PDB IDs to the PDB IDs of their replacements.
strict_error_check: If True, then the following will be treated as errors:
* If any template date is after the max_template_date.
* If any template has identical PDB ID to the query.
* If any template is a duplicate of the query.
* Any feature computation errors.
"""
self._mmcif_dir = mmcif_dir
if not glob.glob(os.path.join(self._mmcif_dir, "*.cif")):
logging.error("Could not find CIFs in %s", self._mmcif_dir)
raise ValueError(f"Could not find CIFs in {self._mmcif_dir}")
try:
self._max_template_date = datetime.datetime.strptime(
max_template_date, "%Y-%m-%d"
)
except ValueError:
raise ValueError(
"max_template_date must be set and have format YYYY-MM-DD."
)
self._max_hits = max_hits
self._kalign_binary_path = kalign_binary_path
self._strict_error_check = strict_error_check
if release_dates_path:
logging.info("Using precomputed release dates %s.", release_dates_path)
self._release_dates = _parse_release_dates(release_dates_path)
else:
self._release_dates = {}
if obsolete_pdbs_path:
logging.info("Using precomputed obsolete pdbs %s.", obsolete_pdbs_path)
self._obsolete_pdbs = _parse_obsolete(obsolete_pdbs_path)
else:
self._obsolete_pdbs = {}
@abc.abstractmethod
def get_templates(
self, query_sequence: str, hits: Sequence[parsers.TemplateHit]
) -> TemplateSearchResult:
"""Computes the templates for given query sequence."""
class HhsearchHitFeaturizer(TemplateHitFeaturizer):
"""A class for turning a3m hits from hhsearch to template features."""
def get_templates(
self, query_sequence: str, hits: Sequence[parsers.TemplateHit]
) -> TemplateSearchResult:
"""Computes the templates for given query sequence (more details above)."""
logging.info("Searching for template for: %s", query_sequence)
template_features = {}
for template_feature_name in TEMPLATE_FEATURES:
template_features[template_feature_name] = []
num_hits = 0
errors = []
warnings = []
for hit in sorted(hits, key=lambda x: x.sum_probs, reverse=True):
# We got all the templates we wanted, stop processing hits.
if num_hits >= self._max_hits:
break
result = _process_single_hit(
query_sequence=query_sequence,
hit=hit,
mmcif_dir=self._mmcif_dir,
max_template_date=self._max_template_date,
release_dates=self._release_dates,
obsolete_pdbs=self._obsolete_pdbs,
strict_error_check=self._strict_error_check,
kalign_binary_path=self._kalign_binary_path,
)
if result.error:
errors.append(result.error)
# There could be an error even if there are some results, e.g. thrown by
# other unparsable chains in the same mmCIF file.
if result.warning:
warnings.append(result.warning)
if result.features is None:
logging.info(
"Skipped invalid hit %s, error: %s, warning: %s",
hit.name,
result.error,
result.warning,
)
else:
# Increment the hit counter, since we got features out of this hit.
num_hits += 1
for k in template_features:
template_features[k].append(result.features[k])
for name in template_features:
if num_hits > 0:
template_features[name] = np.stack(
template_features[name], axis=0
).astype(TEMPLATE_FEATURES[name])
else:
# Make sure the feature has correct dtype even if empty.
template_features[name] = np.array([], dtype=TEMPLATE_FEATURES[name])
return TemplateSearchResult(
features=template_features, errors=errors, warnings=warnings
)
class HmmsearchHitFeaturizer(TemplateHitFeaturizer):
"""A class for turning a3m hits from hmmsearch to template features."""
def get_templates(
self, query_sequence: str, hits: Sequence[parsers.TemplateHit]
) -> TemplateSearchResult:
"""Computes the templates for given query sequence (more details above)."""
logging.info("Searching for template for: %s", query_sequence)
template_features = {}
for template_feature_name in TEMPLATE_FEATURES:
template_features[template_feature_name] = []
already_seen = set()
errors = []
warnings = []
if not hits or hits[0].sum_probs is None:
sorted_hits = hits
else:
sorted_hits = sorted(hits, key=lambda x: x.sum_probs, reverse=True)
for hit in sorted_hits:
# We got all the templates we wanted, stop processing hits.
if len(already_seen) >= self._max_hits:
break
result = _process_single_hit(
query_sequence=query_sequence,
hit=hit,
mmcif_dir=self._mmcif_dir,
max_template_date=self._max_template_date,
release_dates=self._release_dates,
obsolete_pdbs=self._obsolete_pdbs,
strict_error_check=self._strict_error_check,
kalign_binary_path=self._kalign_binary_path,
)
if result.error:
errors.append(result.error)
# There could be an error even if there are some results, e.g. thrown by
# other unparsable chains in the same mmCIF file.
if result.warning:
warnings.append(result.warning)
if result.features is None:
logging.debug(
"Skipped invalid hit %s, error: %s, warning: %s",
hit.name,
result.error,
result.warning,
)
else:
already_seen_key = result.features["template_sequence"]
if already_seen_key in already_seen:
continue
# Increment the hit counter, since we got features out of this hit.
already_seen.add(already_seen_key)
for k in template_features:
template_features[k].append(result.features[k])
if already_seen:
for name in template_features:
template_features[name] = np.stack(
template_features[name], axis=0
).astype(TEMPLATE_FEATURES[name])
else:
num_res = len(query_sequence)
# Construct a default template with all zeros.
template_features = {
"template_aatype": np.zeros(
(1, num_res, len(residue_constants.restypes_with_x_and_gap)),
np.float32,
),
"template_all_atom_mask": np.zeros(
(1, num_res, residue_constants.atom_type_num), np.float32
),
"template_all_atom_positions": np.zeros(
(1, num_res, residue_constants.atom_type_num, 3), np.float32
),
"template_domain_names": np.array(["".encode()], dtype=np.object),
"template_sequence": np.array(["".encode()], dtype=np.object),
"template_sum_probs": np.array([0], dtype=np.float32),
}
return TemplateSearchResult(
features=template_features, errors=errors, warnings=warnings
)
unifold/msa/tools/__init__.py 0 → 100755
View file @ a1c29028
# 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.
"""Python wrappers for third party tools."""
\ No newline at end of file
unifold/msa/tools/hhblits.py 0 → 100755
View file @ a1c29028
# 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.
"""Library to run HHblits from Python."""
import glob
import os
import subprocess
from typing import Any, List, Mapping, Optional, Sequence
from absl import logging
from . import utils
_HHBLITS_DEFAULT_P = 20
_HHBLITS_DEFAULT_Z = 500
class HHBlits:
"""Python wrapper of the HHblits binary."""
def __init__(
self,
*,
binary_path: str,
databases: Sequence[str],
n_cpu: int = 4,
n_iter: int = 3,
e_value: float = 0.001,
maxseq: int = 1_000_000,
realign_max: int = 100_000,
maxfilt: int = 100_000,
min_prefilter_hits: int = 1000,
all_seqs: bool = False,
alt: Optional[int] = None,
p: int = _HHBLITS_DEFAULT_P,
z: int = _HHBLITS_DEFAULT_Z,
):
"""Initializes the Python HHblits wrapper.
Args:
binary_path: The path to the HHblits executable.
databases: A sequence of HHblits database paths. This should be the
common prefix for the database files (i.e. up to but not including
_hhm.ffindex etc.)
n_cpu: The number of CPUs to give HHblits.
n_iter: The number of HHblits iterations.
e_value: The E-value, see HHblits docs for more details.
maxseq: The maximum number of rows in an input alignment. Note that this
parameter is only supported in HHBlits version 3.1 and higher.
realign_max: Max number of HMM-HMM hits to realign. HHblits default: 500.
maxfilt: Max number of hits allowed to pass the 2nd prefilter.
HHblits default: 20000.
min_prefilter_hits: Min number of hits to pass prefilter.
HHblits default: 100.
all_seqs: Return all sequences in the MSA / Do not filter the result MSA.
HHblits default: False.
alt: Show up to this many alternative alignments.
p: Minimum Prob for a hit to be included in the output hhr file.
HHblits default: 20.
z: Hard cap on number of hits reported in the hhr file.
HHblits default: 500. NB: The relevant HHblits flag is -Z not -z.
Raises:
RuntimeError: If HHblits binary not found within the path.
"""
self.binary_path = binary_path
self.databases = databases
for database_path in self.databases:
if not glob.glob(database_path + "_*"):
logging.error("Could not find HHBlits database %s", database_path)
raise ValueError(f"Could not find HHBlits database {database_path}")
self.n_cpu = n_cpu
self.n_iter = n_iter
self.e_value = e_value
self.maxseq = maxseq
self.realign_max = realign_max
self.maxfilt = maxfilt
self.min_prefilter_hits = min_prefilter_hits
self.all_seqs = all_seqs
self.alt = alt
self.p = p
self.z = z
def query(self, input_fasta_path: str) -> List[Mapping[str, Any]]:
"""Queries the database using HHblits."""
with utils.tmpdir_manager() as query_tmp_dir:
a3m_path = os.path.join(query_tmp_dir, "output.a3m")
db_cmd = []
for db_path in self.databases:
db_cmd.append("-d")
db_cmd.append(db_path)
cmd = [
self.binary_path,
"-i",
input_fasta_path,
"-cpu",
str(self.n_cpu),
"-oa3m",
a3m_path,
"-o",
"/dev/null",
"-n",
str(self.n_iter),
"-e",
str(self.e_value),
"-maxseq",
str(self.maxseq),
"-realign_max",
str(self.realign_max),
"-maxfilt",
str(self.maxfilt),
"-min_prefilter_hits",
str(self.min_prefilter_hits),
]
if self.all_seqs:
cmd += ["-all"]
if self.alt:
cmd += ["-alt", str(self.alt)]
if self.p != _HHBLITS_DEFAULT_P:
cmd += ["-p", str(self.p)]
if self.z != _HHBLITS_DEFAULT_Z:
cmd += ["-Z", str(self.z)]
cmd += db_cmd
logging.info('Launching subprocess "%s"', " ".join(cmd))
process = subprocess.Popen(
cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE
)
with utils.timing("HHblits query"):
stdout, stderr = process.communicate()
retcode = process.wait()
if retcode:
# Logs have a 15k character limit, so log HHblits error line by line.
logging.error("HHblits failed. HHblits stderr begin:")
for error_line in stderr.decode("utf-8").splitlines():
if error_line.strip():
logging.error(error_line.strip())
logging.error("HHblits stderr end")
raise RuntimeError(
"HHblits failed\nstdout:\n%s\n\nstderr:\n%s\n"
% (stdout.decode("utf-8"), stderr[:500_000].decode("utf-8"))
)
with open(a3m_path) as f:
a3m = f.read()
raw_output = dict(
a3m=a3m,
output=stdout,
stderr=stderr,
n_iter=self.n_iter,
e_value=self.e_value,
)
return [raw_output]
unifold/msa/tools/hhsearch.py 0 → 100755
View file @ a1c29028
# 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.
"""Library to run HHsearch from Python."""
import glob
import os
import subprocess
from typing import Sequence
from absl import logging
from unifold.msa import parsers
from . import utils
class HHSearch:
"""Python wrapper of the HHsearch binary."""
def __init__(
self, *, binary_path: str, databases: Sequence[str], maxseq: int = 1_000_000
):
"""Initializes the Python HHsearch wrapper.
Args:
binary_path: The path to the HHsearch executable.
databases: A sequence of HHsearch database paths. This should be the
common prefix for the database files (i.e. up to but not including
_hhm.ffindex etc.)
maxseq: The maximum number of rows in an input alignment. Note that this
parameter is only supported in HHBlits version 3.1 and higher.
Raises:
RuntimeError: If HHsearch binary not found within the path.
"""
self.binary_path = binary_path
self.databases = databases
self.maxseq = maxseq
for database_path in self.databases:
if not glob.glob(database_path + "_*"):
logging.error("Could not find HHsearch database %s", database_path)
raise ValueError(f"Could not find HHsearch database {database_path}")
@property
def output_format(self) -> str:
return "hhr"
@property
def input_format(self) -> str:
return "a3m"
def query(self, a3m: str) -> str:
"""Queries the database using HHsearch using a given a3m."""
with utils.tmpdir_manager() as query_tmp_dir:
input_path = os.path.join(query_tmp_dir, "query.a3m")
hhr_path = os.path.join(query_tmp_dir, "output.hhr")
with open(input_path, "w") as f:
f.write(a3m)
db_cmd = []
for db_path in self.databases:
db_cmd.append("-d")
db_cmd.append(db_path)
cmd = [
self.binary_path,
"-i",
input_path,
"-o",
hhr_path,
"-maxseq",
str(self.maxseq),
] + db_cmd
logging.info('Launching subprocess "%s"', " ".join(cmd))
process = subprocess.Popen(
cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE
)
with utils.timing("HHsearch query"):
stdout, stderr = process.communicate()
retcode = process.wait()
if retcode:
# Stderr is truncated to prevent proto size errors in Beam.
raise RuntimeError(
"HHSearch failed:\nstdout:\n%s\n\nstderr:\n%s\n"
% (stdout.decode("utf-8"), stderr[:100_000].decode("utf-8"))
)
with open(hhr_path) as f:
hhr = f.read()
return hhr
def get_template_hits(
self, output_string: str, input_sequence: str
) -> Sequence[parsers.TemplateHit]:
"""Gets parsed template hits from the raw string output by the tool."""
del input_sequence # Used by hmmseach but not needed for hhsearch.
return parsers.parse_hhr(output_string)
unifold/msa/tools/hmmbuild.py 0 → 100755
View file @ a1c29028
# 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.
"""A Python wrapper for hmmbuild - construct HMM profiles from MSA."""
import os
import re
import subprocess
from absl import logging
from . import utils
class Hmmbuild(object):
"""Python wrapper of the hmmbuild binary."""
def __init__(self, *, binary_path: str, singlemx: bool = False):
"""Initializes the Python hmmbuild wrapper.
Args:
binary_path: The path to the hmmbuild executable.
singlemx: Whether to use --singlemx flag. If True, it forces HMMBuild to
just use a common substitution score matrix.
Raises:
RuntimeError: If hmmbuild binary not found within the path.
"""
self.binary_path = binary_path
self.singlemx = singlemx
def build_profile_from_sto(self, sto: str, model_construction="fast") -> str:
"""Builds a HHM for the aligned sequences given as an A3M string.
Args:
sto: A string with the aligned sequences in the Stockholm format.
model_construction: Whether to use reference annotation in the msa to
determine consensus columns ('hand') or default ('fast').
Returns:
A string with the profile in the HMM format.
Raises:
RuntimeError: If hmmbuild fails.
"""
return self._build_profile(sto, model_construction=model_construction)
def build_profile_from_a3m(self, a3m: str) -> str:
"""Builds a HHM for the aligned sequences given as an A3M string.
Args:
a3m: A string with the aligned sequences in the A3M format.
Returns:
A string with the profile in the HMM format.
Raises:
RuntimeError: If hmmbuild fails.
"""
lines = []
for line in a3m.splitlines():
if not line.startswith(">"):
line = re.sub("[a-z]+", "", line) # Remove inserted residues.
lines.append(line + "\n")
msa = "".join(lines)
return self._build_profile(msa, model_construction="fast")
def _build_profile(self, msa: str, model_construction: str = "fast") -> str:
"""Builds a HMM for the aligned sequences given as an MSA string.
Args:
msa: A string with the aligned sequences, in A3M or STO format.
model_construction: Whether to use reference annotation in the msa to
determine consensus columns ('hand') or default ('fast').
Returns:
A string with the profile in the HMM format.
Raises:
RuntimeError: If hmmbuild fails.
ValueError: If unspecified arguments are provided.
"""
if model_construction not in {"hand", "fast"}:
raise ValueError(
f"Invalid model_construction {model_construction} - only"
"hand and fast supported."
)
with utils.tmpdir_manager() as query_tmp_dir:
input_query = os.path.join(query_tmp_dir, "query.msa")
output_hmm_path = os.path.join(query_tmp_dir, "output.hmm")
with open(input_query, "w") as f:
f.write(msa)
cmd = [self.binary_path]
# If adding flags, we have to do so before the output and input:
if model_construction == "hand":
cmd.append(f"--{model_construction}")
if self.singlemx:
cmd.append("--singlemx")
cmd.extend(
[
"--amino",
output_hmm_path,
input_query,
]
)
logging.info("Launching subprocess %s", cmd)
process = subprocess.Popen(
cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE
)
with utils.timing("hmmbuild query"):
stdout, stderr = process.communicate()
retcode = process.wait()
logging.info(
"hmmbuild stdout:\n%s\n\nstderr:\n%s\n",
stdout.decode("utf-8"),
stderr.decode("utf-8"),
)
if retcode:
raise RuntimeError(
"hmmbuild failed\nstdout:\n%s\n\nstderr:\n%s\n"
% (stdout.decode("utf-8"), stderr.decode("utf-8"))
)
with open(output_hmm_path, encoding="utf-8") as f:
hmm = f.read()
return hmm
unifold/msa/tools/hmmsearch.py 0 → 100755
View file @ a1c29028
# 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.
"""A Python wrapper for hmmsearch - search profile against a sequence db."""
import os
import subprocess
from typing import Optional, Sequence
from absl import logging
from unifold.msa import parsers
from . import hmmbuild, utils
class Hmmsearch(object):
"""Python wrapper of the hmmsearch binary."""
def __init__(
self,
*,
binary_path: str,
hmmbuild_binary_path: str,
database_path: str,
flags: Optional[Sequence[str]] = None,
):
"""Initializes the Python hmmsearch wrapper.
Args:
binary_path: The path to the hmmsearch executable.
hmmbuild_binary_path: The path to the hmmbuild executable. Used to build
an hmm from an input a3m.
database_path: The path to the hmmsearch database (FASTA format).
flags: List of flags to be used by hmmsearch.
Raises:
RuntimeError: If hmmsearch binary not found within the path.
"""
self.binary_path = binary_path
self.hmmbuild_runner = hmmbuild.Hmmbuild(binary_path=hmmbuild_binary_path)
self.database_path = database_path
if flags is None:
# Default hmmsearch run settings.
flags = [
"--F1",
"0.1",
"--F2",
"0.1",
"--F3",
"0.1",
"--incE",
"100",
"-E",
"100",
"--domE",
"100",
"--incdomE",
"100",
]
self.flags = flags
if not os.path.exists(self.database_path):
logging.error("Could not find hmmsearch database %s", database_path)
raise ValueError(f"Could not find hmmsearch database {database_path}")
@property
def output_format(self) -> str:
return "sto"
@property
def input_format(self) -> str:
return "sto"
def query(self, msa_sto: str) -> str:
"""Queries the database using hmmsearch using a given stockholm msa."""
hmm = self.hmmbuild_runner.build_profile_from_sto(
msa_sto, model_construction="hand"
)
return self.query_with_hmm(hmm)
def query_with_hmm(self, hmm: str) -> str:
"""Queries the database using hmmsearch using a given hmm."""
with utils.tmpdir_manager() as query_tmp_dir:
hmm_input_path = os.path.join(query_tmp_dir, "query.hmm")
out_path = os.path.join(query_tmp_dir, "output.sto")
with open(hmm_input_path, "w") as f:
f.write(hmm)
cmd = [
self.binary_path,
"--noali", # Don't include the alignment in stdout.
"--cpu",
"8",
]
# If adding flags, we have to do so before the output and input:
if self.flags:
cmd.extend(self.flags)
cmd.extend(
[
"-A",
out_path,
hmm_input_path,
self.database_path,
]
)
logging.info("Launching sub-process %s", cmd)
process = subprocess.Popen(
cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE
)
with utils.timing(
f"hmmsearch ({os.path.basename(self.database_path)}) query"
):
stdout, stderr = process.communicate()
retcode = process.wait()
if retcode:
raise RuntimeError(
"hmmsearch failed:\nstdout:\n%s\n\nstderr:\n%s\n"
% (stdout.decode("utf-8"), stderr.decode("utf-8"))
)
with open(out_path) as f:
out_msa = f.read()
return out_msa
def get_template_hits(
self, output_string: str, input_sequence: str
) -> Sequence[parsers.TemplateHit]:
"""Gets parsed template hits from the raw string output by the tool."""
a3m_string = parsers.convert_stockholm_to_a3m(
output_string, remove_first_row_gaps=False
)
template_hits = parsers.parse_hmmsearch_a3m(
query_sequence=input_sequence, a3m_string=a3m_string, skip_first=False
)
return template_hits
unifold/msa/tools/jackhmmer.py 0 → 100755
View file @ a1c29028
# 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.
"""Library to run Jackhmmer from Python."""
from concurrent import futures
import glob
import os
import subprocess
from typing import Any, Callable, Mapping, Optional, Sequence
from urllib import request
from absl import logging
from . import utils
class Jackhmmer:
"""Python wrapper of the Jackhmmer binary."""
def __init__(
self,
*,
binary_path: str,
database_path: str,
n_cpu: int = 8,
n_iter: int = 1,
e_value: float = 0.0001,
z_value: Optional[int] = None,
get_tblout: bool = False,
filter_f1: float = 0.0005,
filter_f2: float = 0.00005,
filter_f3: float = 0.0000005,
incdom_e: Optional[float] = None,
dom_e: Optional[float] = None,
num_streamed_chunks: Optional[int] = None,
streaming_callback: Optional[Callable[[int], None]] = None,
):
"""Initializes the Python Jackhmmer wrapper.
Args:
binary_path: The path to the jackhmmer executable.
database_path: The path to the jackhmmer database (FASTA format).
n_cpu: The number of CPUs to give Jackhmmer.
n_iter: The number of Jackhmmer iterations.
e_value: The E-value, see Jackhmmer docs for more details.
z_value: The Z-value, see Jackhmmer docs for more details.
get_tblout: Whether to save tblout string.
filter_f1: MSV and biased composition pre-filter, set to >1.0 to turn off.
filter_f2: Viterbi pre-filter, set to >1.0 to turn off.
filter_f3: Forward pre-filter, set to >1.0 to turn off.
incdom_e: Domain e-value criteria for inclusion of domains in MSA/next
round.
dom_e: Domain e-value criteria for inclusion in tblout.
num_streamed_chunks: Number of database chunks to stream over.
streaming_callback: Callback function run after each chunk iteration with
the iteration number as argument.
"""
self.binary_path = binary_path
self.database_path = database_path
self.num_streamed_chunks = num_streamed_chunks
if not os.path.exists(self.database_path) and num_streamed_chunks is None:
logging.error("Could not find Jackhmmer database %s", database_path)
raise ValueError(f"Could not find Jackhmmer database {database_path}")
self.n_cpu = n_cpu
self.n_iter = n_iter
self.e_value = e_value
self.z_value = z_value
self.filter_f1 = filter_f1
self.filter_f2 = filter_f2
self.filter_f3 = filter_f3
self.incdom_e = incdom_e
self.dom_e = dom_e
self.get_tblout = get_tblout
self.streaming_callback = streaming_callback
def _query_chunk(
self, input_fasta_path: str, database_path: str
) -> Mapping[str, Any]:
"""Queries the database chunk using Jackhmmer."""
with utils.tmpdir_manager() as query_tmp_dir:
sto_path = os.path.join(query_tmp_dir, "output.sto")
# The F1/F2/F3 are the expected proportion to pass each of the filtering
# stages (which get progressively more expensive), reducing these
# speeds up the pipeline at the expensive of sensitivity. They are
# currently set very low to make querying Mgnify run in a reasonable
# amount of time.
cmd_flags = [
# Don't pollute stdout with Jackhmmer output.
"-o",
"/dev/null",
"-A",
sto_path,
"--noali",
"--F1",
str(self.filter_f1),
"--F2",
str(self.filter_f2),
"--F3",
str(self.filter_f3),
"--incE",
str(self.e_value),
# Report only sequences with E-values <= x in per-sequence output.
"-E",
str(self.e_value),
"--cpu",
str(self.n_cpu),
"-N",
str(self.n_iter),
]
if self.get_tblout:
tblout_path = os.path.join(query_tmp_dir, "tblout.txt")
cmd_flags.extend(["--tblout", tblout_path])
if self.z_value:
cmd_flags.extend(["-Z", str(self.z_value)])
if self.dom_e is not None:
cmd_flags.extend(["--domE", str(self.dom_e)])
if self.incdom_e is not None:
cmd_flags.extend(["--incdomE", str(self.incdom_e)])
cmd = [self.binary_path] + cmd_flags + [input_fasta_path, database_path]
logging.info('Launching subprocess "%s"', " ".join(cmd))
process = subprocess.Popen(
cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE
)
with utils.timing(f"Jackhmmer ({os.path.basename(database_path)}) query"):
_, stderr = process.communicate()
retcode = process.wait()
if retcode:
raise RuntimeError(
"Jackhmmer failed\nstderr:\n%s\n" % stderr.decode("utf-8")
)
# Get e-values for each target name
tbl = ""
if self.get_tblout:
with open(tblout_path) as f:
tbl = f.read()
with open(sto_path) as f:
sto = f.read()
raw_output = dict(
sto=sto, tbl=tbl, stderr=stderr, n_iter=self.n_iter, e_value=self.e_value
)
return raw_output
def query(self, input_fasta_path: str) -> Sequence[Mapping[str, Any]]:
"""Queries the database using Jackhmmer."""
if self.num_streamed_chunks is None:
return [self._query_chunk(input_fasta_path, self.database_path)]
db_basename = os.path.basename(self.database_path)
db_remote_chunk = lambda db_idx: f"{self.database_path}.{db_idx}"
db_local_chunk = lambda db_idx: f"/tmp/ramdisk/{db_basename}.{db_idx}"
# Remove existing files to prevent OOM
for f in glob.glob(db_local_chunk("[0-9]*")):
try:
os.remove(f)
except OSError:
print(f"OSError while deleting {f}")
# Download the (i+1)-th chunk while Jackhmmer is running on the i-th chunk
with futures.ThreadPoolExecutor(max_workers=2) as executor:
chunked_output = []
for i in range(1, self.num_streamed_chunks + 1):
# Copy the chunk locally
if i == 1:
future = executor.submit(
request.urlretrieve, db_remote_chunk(i), db_local_chunk(i)
)
if i < self.num_streamed_chunks:
next_future = executor.submit(
request.urlretrieve,
db_remote_chunk(i + 1),
db_local_chunk(i + 1),
)
# Run Jackhmmer with the chunk
future.result()
chunked_output.append(
self._query_chunk(input_fasta_path, db_local_chunk(i))
)
# Remove the local copy of the chunk
os.remove(db_local_chunk(i))
# Do not set next_future for the last chunk so that this works even for
# databases with only 1 chunk.
if i < self.num_streamed_chunks:
future = next_future
if self.streaming_callback:
self.streaming_callback(i)
return chunked_output
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment