import re
import sys
import gzip
import logging
import random
import string
from typing import *
from collections import defaultdict
from itertools import chain
from collections import OrderedDict
from collections import Counter
import velocyto as vcy
import h5py
import pysam
import numpy as np
import os
import sys
[docs]class ExInCounter:
""" Main class to do the counting of introns and exons """
def __init__(self, sampleid: str, logic: vcy.Logic, valid_bcset: Set[str]=None,
umi_extension: str="no", onefilepercell: bool=False, dump_option: str="0",
outputfolder: str="./", loom_numeric_dtype: str=vcy.LOOM_NUMERIC_DTYPE) -> None:
self.outputfolder = outputfolder
self.sampleid = sampleid
self.loom_numeric_dtype = loom_numeric_dtype
self.logic = logic()
# NOTE: maybe there shoulb be a self.logic.verify_inputs(args) step at the end of init
if valid_bcset is None:
self.valid_bcset: Set[str] = set()
self.filter_mode = False
self.counter = 0
else:
self.valid_bcset = valid_bcset
self.filter_mode = True
self.annotations_by_chrm_strand: Dict[str, Dict[str, vcy.TranscriptModel]] = {}
self.mask_ivls_by_chromstrand = defaultdict(list) # type: Dict[str, List]
self.geneid2ix: Dict[str, int] = {}
self.genes: Dict[str, vcy.GeneInfo] = {}
if umi_extension.lower() == "no":
self.umi_extract = self._no_extension
elif umi_extension.lower() == "chr":
self.umi_extract = self._extension_chr
elif umi_extension.lower() == "gene" or umi_extension.lower() == "gx":
self.umi_extract = self._extension_Gene
elif umi_extension[-2:] == "bp":
self.umi_bp = int(umi_extension[:-2])
self.umi_extract = self._extension_Nbp
elif umi_extension.lower() == "without_umi":
self.umi_extract = self._placeolder_umi
else:
raise ValueError(f"umi_extension {umi_extension} is not allowed. Use `no`, `Gene` or `[N]bp`")
if onefilepercell:
self.cell_barcode_get = self._bam_id_barcode
else:
self.cell_barcode_get = self._normal_cell_barcode_get
if self.logic.stranded:
if self.logic.accept_discordant:
self.count_cell_batch = self._count_cell_batch_stranded_discordant
else:
self.count_cell_batch = self._count_cell_batch_stranded
else:
self.count_cell_batch = self._count_cell_batch_non_stranded
# NOTE: by using a default dict and not logging access to keys that do not exist, we might miss bugs!!!
self.test_flag = None
if dump_option[0] == "p":
self.kind_of_report = "p"
self.report_state = 0
self.every_n_report = int(dump_option[1:])
else:
self.kind_of_report = "h"
self.report_state = 0
self.every_n_report = int(dump_option)
self.cellbarcode_str = "NULL_BC" # This value should never be used this is just to initialize it and detect if there are bugs downstream
self.umibarcode_str = "NULL_UB" # This value should never be used this is just to initialize it and detect if there are bugs downstream
# NOTE: not supported anymore because now we support variable length barcodes
# @property
# def bclen(self) -> int:
# try:
# return len(next(iter(self.valid_bcset)))
# except StopIteration:
# return None
[docs] @staticmethod
def parse_cigar_tuple(cigartuples: List[Tuple], pos: int) -> Tuple[List[Tuple[int, int]], bool, int, int]:
segments = []
hole_to_remove = set()
ref_skip = False
clip5 = clip3 = 0
p = pos
for i, (operation_id, length) in enumerate(cigartuples):
if operation_id == 0: # vcy.CIGAR[operation_id] == "BAM_CMATCH"
segments.append((p, p + length - 1))
p += length
elif operation_id == 3: # A splice || vcy.CIGAR[operation_id] == 'BAM_CREF_SKIP'
ref_skip = True
p += length
elif operation_id == 2: # A deletion || cy.CIGAR[operation_id] == 'BAM_CDEL'
if length <= vcy.PATCH_INDELS:
try:
if cigartuples[i + 1][0] == 0 and cigartuples[i - 1][0] == 0:
hole_to_remove.add(len(segments) - 1)
except IndexError:
pass
p += length
elif operation_id == 4: # bases at 5' or 3' are NOT part of the alignment || vcy.CIGAR[operation_id] == 'BAM_CSOFT_CLIP'
if p == pos:
clip5 = length # At start of alignment
else:
clip3 = length # Must be at end of alignment vcy.CIGAR[operation_id] in ["BAM_CINS", "BAM_CHARD_CLIP"]
p += length
elif operation_id == 1: # An insertion BAM_CINS
if length <= vcy.PATCH_INDELS:
try:
if cigartuples[i + 1][0] == 0 and cigartuples[i - 1][0] == 0:
hole_to_remove.add(len(segments) - 1)
except IndexError:
pass
# else do nothing
# NOTE: maybe we should make so that the reads get discarded
elif operation_id == 5: # BAM_CHARD_CLIP
logging.warn("Hard clip was encountered! All mapping are assumed soft clipped")
# Merge segments separated by small insertions and deletions
for a, b in enumerate(sorted(hole_to_remove)): # NOTE maybe sorted is not required realy
segments[b - a] = (segments.pop(b - a)[0], segments[b - a][1])
return segments, ref_skip, clip5, clip3
[docs] def peek(self, bamfile: str, lines: int=1000) -> None:
"""Peeks into the samfile to determine if it is a cellranger or dropseq file
"""
logging.debug(f"Peeking into {bamfile}")
fin = pysam.AlignmentFile(bamfile) # type: pysam.AlignmentFile
cellranger: int = 0
dropseq: int = 0
failed: int = 0
for i, read in enumerate(fin):
if read.is_unmapped:
continue
if read.has_tag("CB") and read.has_tag("UB"):
cellranger += 1
elif read.has_tag("XC") and read.has_tag("XM"):
dropseq += 1
else:
logging.warn(f"Not found cell and umi barcode in entry {i} of the bam file")
failed += 1
if cellranger > lines:
self.cellbarcode_str = "CB"
self.umibarcode_str = "UB"
break
elif dropseq > lines:
self.cellbarcode_str = "XC"
self.umibarcode_str = "XM"
break
elif failed > 5 * lines:
raise IOError("The bam file does not contain cell and umi barcodes appropriatelly formatted. If you are runnin UMI-less data you should use the -U flag.")
else:
pass
fin.close()
[docs] def peek_umi_only(self, bamfile: str, lines: int=30) -> None:
"""Peeks for umi into the samfile to determine if it is a cellranger or dropseq file
"""
logging.debug(f"Peeking into {bamfile}")
fin = pysam.AlignmentFile(bamfile) # type: pysam.AlignmentFile
cellranger: int = 0
dropseq: int = 0
failed: int = 0
for i, read in enumerate(fin):
if read.is_unmapped:
continue
if read.has_tag("UB"):
cellranger += 1
elif read.has_tag("XM"):
dropseq += 1
else:
logging.warn(f"Not found cell and umi barcode in entry {i} of the bam file")
failed += 1
if cellranger > lines:
self.umibarcode_str = "UB"
break
elif dropseq > lines:
self.umibarcode_str = "XM"
break
elif failed > 5 * lines:
raise IOError("The bam file does not contain umi barcodes appropriatelly formatted. If you are runnin UMI-less data you should use the -U flag.")
else:
pass
fin.close()
def _no_extension(self, read: pysam.AlignedSegment) -> str:
return read.get_tag(self.umibarcode_str)
def _extension_Nbp(self, read: pysam.AlignedSegment) -> str:
return read.get_tag(self.umibarcode_str) + read.query_alignment_sequence[:self.umi_bp]
def _extension_Gene(self, read: pysam.AlignedSegment) -> str:
try:
return read.get_tag(self.umibarcode_str) + "_" + read.get_tag("GX") # catch the error
except KeyError:
return read.get_tag(self.umibarcode_str) + "_withoutGX"
def _placeolder_umi(self, read: pysam.AlignedSegment) -> str:
return ''.join(random.choice(string.ascii_uppercase + string.digits) for _ in range(vcy.PLACEHOLDER_UMI_LEN))
def _extension_chr(self, read: pysam.AlignedSegment) -> str:
return read.get_tag(self.umibarcode_str) + f"_{read.rname}:{read.reference_start // 10000000}" # catch the error
def _normal_cell_barcode_get(self, read: pysam.AlignedSegment) -> str:
return read.get_tag(self.cellbarcode_str).split("-")[0]
def _bam_id_barcode(self, read: pysam.AlignedSegment) -> str:
return f"{self._current_bamfile}"
[docs] def iter_alignments(self, bamfiles: Tuple[str], unique: bool=True, yield_line: bool=False) -> Iterable:
"""Iterates over the bam/sam file and yield Read objects
Arguments
---------
bamfiles: Tuple[str]
path to the bam files
unique: bool
yield only unique alignments
yield_line: bool
whether to yield the raw sam line
Returns
-------
yields vcy.Read for each valid line of the bam file
or a Tuple (vcy.Read, sam_line) if ``yield_line==True``
NOTE: At the file change it yields a `None`
"""
# No peeking here, this will happen a layer above, and only once on the not sorted file! before it was self.peek(samfile, lines=10)
bamfile_name_seen: Set[str] = set()
counter_skipped_no_barcode = 0
if Counter(bamfiles).most_common(1)[0][1] != 1:
logging.warning("The bamfiles names are not unique. The full path to them will be used as unique identifier")
use_basename = False
else:
use_basename = True
for bamfile in bamfiles:
if use_basename:
self._current_bamfile = os.path.basename(bamfile)
else:
self._current_bamfile = str(bamfile)
logging.debug(f"Reading {bamfile}")
fin = pysam.AlignmentFile(bamfile) # type: pysam.AlignmentFile
for i, read in enumerate(fin):
if i % 10000000 == 0:
logging.debug(f"Read first {i // 1000000} million reads")
if read.is_unmapped:
continue
# If unique parameter is set to True, skip not unique alignments
if unique and read.get_tag("NH") != 1:
continue
try:
bc = self.cell_barcode_get(read) # NOTE: this rstrip is relevant only for cellranger, should not cause trouble in Dropseq
umi = self.umi_extract(read) # read.get_tag(self.umibarcode_str)
except KeyError:
if read.has_tag(self.cellbarcode_str) and read.has_tag(self.umibarcode_str):
raise KeyError(f"Some errors in parsing the cell barcode has occurred {self.cellbarcode_str}, {self.umibarcode_str}\n{read}")
counter_skipped_no_barcode += 1
continue # NOTE: Here errors could go unnoticed
if bc not in self.valid_bcset:
if self.filter_mode:
continue
else:
self.valid_bcset.add(bc)
strand = '-' if read.is_reverse else '+'
chrom = fin.get_reference_name(read.rname) # this is return a string otherwise read.rname for the integer
if chrom.startswith('chr'):
# I have to deal with incongruences with the .gft (what is cellranger doing???)
# NOTE Why this happens?
if "_" in chrom:
chrom = chrom.split("_")[1]
else:
chrom = chrom[3:]
if chrom == "M":
chrom = "MT"
pos = read.reference_start + 1 # reads in pysam are always 0-based, but 1-based is more convenient to wor with in bioinformatics
segments, ref_skipped, clip5, clip3 = self.parse_cigar_tuple(read.cigartuples, pos)
if segments == []:
logging.debug("No segments in read:%s" % read.qname)
read_object = vcy.Read(bc, umi, chrom, strand, pos, segments, clip5, clip3, ref_skipped)
if yield_line:
if read_object.span > 3000000: # Longest locus existing
logging.warn(f"Trashing read, too long span\n{read.tostring(fin)}")
else:
yield read_object, read.tostring(fin)
else:
if read_object.span > 3000000: # Longest locus existing
logging.warn(f"Trashing read, too long span\n{read.tostring(fin)}")
else:
yield read_object
fin.close()
# NOTE Yielding None counts as a flag that the file read has been changed
if yield_line:
yield None, None
else:
yield None
logging.debug(f"{counter_skipped_no_barcode} reads were skipped because no apropiate cell or umi barcode was found")
[docs] def read_repeats(self, gtf_file: str, tolerance: int=5) -> Dict[str, List[vcy.Feature]]:
"""Read repeats and merge close ones into highly repetitive areas
Arguments
---------
gtf_file: str
file to read
tolerance: int, default=5
if two repeats intervals to be masked are found closer than tolerance bases from each other they are fused in one bigger masked interval.
Notice that in the downstream analysis only reads that are fall inside mask intervals are discarded
Returns
-------
mask_ivls_by_chromstrand: Dict[str, List[vcy.Feature]]
A dictionary key: chromosome+strand value: list of features (repeat intervals)
(The reference is returned but an internal attribure self.self.masked_by_chrm_strand is kept)
"""
# Example code to sort the gtf file
# sorted_filename = gtffile.split(".")[0] + "_sorted.gtf"
# logging.debug(f"Sorting by `sort -k1,1 -k7,7 -k4,4n {gtffile} > {sorted_filename}`")
# with open(sorted_filename, "w") as f:
# p1 = subprocess.run(["sort", "-k1,1", "-k7,7", "-k4,4n", gtffile],
# stdout=f)
# Parse arguments
logging.debug(f'Reading {gtf_file}, the file will be sorted in memory')
# Read up skipping headers up to the first valid entry
repeat_ivls_list: List[vcy.Feature] = []
# fin = open(gtf_file)
gtf_lines = [line for line in open(gtf_file) if not line.startswith('#')]
def sorting_key(entry: str) -> Tuple[str, bool, int, str]:
"""This sorting strategy is equivalent to sort -k1,1 -k7,7 -k4,4n"""
x = entry.split("\t")
return (x[0], x[6] == "+", int(x[3]), entry) # The last element of the touple corresponds to the `last resort comparison`
gtf_lines = sorted(gtf_lines, key=sorting_key)
###
line = gtf_lines.pop(0)
fields = line.rstrip().split('\t')
chrom, feature_class, feature_type, start_str, end_str, junk, strand, junk, tags = fields
# Removing chr from the chromosome name to uniform different formats of gtf files, taht might or might not have the prefix "chr"
if chrom[:3].lower() == "chr":
chrom = chrom[3:]
start = int(start_str)
end = int(end_str)
chromstrand: str = chrom + strand
# Set this tu the current entry
curr_chrom: str = chrom
curr_feature_class: str = feature_class
curr_feature_type: str = feature_type
curr_start: int = start
curr_end: int = end
curr_n: int = 1
curr_strand: str = strand
curr_tags: str = tags
curr_chromstrand: str = chromstrand
for line in gtf_lines:
fields = line.rstrip().split('\t')
chrom, feature_class, feature_type, start_str, end_str, junk, strand, junk, tags = fields
# Removing chr from the chromosome name to uniform different formats of gtf files, taht might or might not have the prefix "chr"
if chrom[:3].lower() == "chr":
chrom = chrom[3:]
start = int(start_str)
end = int(end_str)
chromstrand = chrom + strand
# On chromosome or strand change: save and clean after yourself
if chromstrand != curr_chromstrand:
self.mask_ivls_by_chromstrand[curr_chromstrand] = repeat_ivls_list
repeat_ivls_list = []
curr_chrom = chrom
curr_strand = strand
curr_chromstrand = curr_chrom + curr_strand
# If there is no overlap/contiguity within a certain tolerance with the next
if start > curr_end + tolerance:
# Close and store the previus interval
repeat_ivls_list.append(vcy.Feature(start=curr_start, end=curr_end, kind=ord("r"), exin_no=curr_n))
# NOTE exin_no is being used to store curr_n, and kind could be used to store tag for some extra memory cost
# Remember the newly parsed interval
curr_start = start
curr_end = end
curr_n = 1 # number of original repeat regions included
# this is extra information that right now is not stored in vcy.Feature (not to waste memory) since is not used
curr_feature_class = feature_class
curr_feature_type = feature_type
curr_tags = tags
else:
# Extend the masked interval!
curr_end = end
curr_n += 1
# this is extra information that right now is not saved (not to waste memory) since is not used
gap = start - curr_end
curr_tags = f"{curr_tags} gap {gap}; {tags}" if gap > 0 else curr_tags + tags
n = 0
for chromstrand, feature_list in self.mask_ivls_by_chromstrand.items():
feature_list.sort() # relies on the __lt__ method of vcy.Feature
n += len(feature_list)
logging.debug(f'Processed masked annotation .gtf and generated {n} intervals to mask!')
return self.mask_ivls_by_chromstrand
[docs] def assign_indexes_to_genes(self, features: Dict[str, vcy.TranscriptModel]) -> None:
"""Assign to each newly encoutered gene an unique index corresponding to the output matrix column ix
"""
logging.debug("Assigning indexes to genes")
for name, trmodel in features.items():
if trmodel.geneid in self.geneid2ix:
if self.genes[trmodel.geneid].start > trmodel.start:
self.genes[trmodel.geneid].start = trmodel.start
if self.genes[trmodel.geneid].end < trmodel.end:
self.genes[trmodel.geneid].end = trmodel.end
else:
self.geneid2ix[trmodel.geneid] = len(self.geneid2ix)
self.genes[trmodel.geneid] = vcy.GeneInfo(trmodel.genename, trmodel.geneid, trmodel.chromstrand, trmodel.start, trmodel.end)
[docs] def read_transcriptmodels(self, gtf_file: str) -> Dict[str, Dict[str, vcy.TranscriptModel]]:
"""Reads transcript models from a sorted .gtf file
Arguments
---------
gtf_file: str
Path to the sorted gtf file
Returns
-------
annotations_by_chrm_strand: Dict[str, List[vcy.TrancriptModel]]
A dictionary key: chromosome+strand value: list of trascript models
(The reference is returned but an internal attribure self.annotations_by_chrm_strand is kept)
There will exist an object vcy.Features for the same exon appearing in a different vcy.TranscriptModel. (his is desired)
"""
# Define the regexes for the parsing
regex_trid = re.compile('transcript_id "([^"]+)"')
regex_trname = re.compile('transcript_name "([^"]+)"')
regex_geneid = re.compile('gene_id "([^"]+)"')
regex_genename = re.compile('gene_name "([^"]+)"')
regex_exonno = re.compile('exon_number "*?([\w]+)') # re.compile('exon_number "([^"]+)"')
# Initialize containers
# headerlines: List[str] = []
gtf_lines = [line for line in open(gtf_file) if not line.startswith('#')]
def sorting_key(entry: str) -> Tuple[str, bool, int, str]:
"""This sorting strategy is equivalent to sort -k1,1 -k7,7 -k4,4n"""
x = entry.split("\t")
return (x[0], x[6] == "+", int(x[3]), entry) # The last element of the touple corresponds to the `last resort comparison`
gtf_lines = self.peek_and_correct(gtf_lines)
gtf_lines = sorted(gtf_lines, key=sorting_key)
curr_chromstrand = None
features: Dict[str, vcy.TranscriptModel] = OrderedDict()
# Loop throug gtf file (assumes it is ordered)
for nth_line, line in enumerate(gtf_lines):
# Deal with headers
# if line.startswith('#'):
# headerlines.append(line)
# continue
fields = line.rstrip().split('\t')
chrom, feature_class, feature_type, start_str, end_str, junk, strand, junk, tags = fields
# Deal with possible incongruences between .gtf and bam file chromosome naming
# We keep the name of the chromosome removing the chr from the name and removing part after the . if present
if "chr" in chrom[:4]:
chrom = chrom[3:] # NOTE before it was chrom[3:].split(".")[0]
else:
pass # NOTE before it was chrom.split(".")[0]
# A new chromosome/strand encountered
if chrom + strand != curr_chromstrand:
if curr_chromstrand is not None: # Every time with exception with first and the last chromosome
if not chrom + strand not in self.annotations_by_chrm_strand:
# NOTE this is not enough as a check but it will detect with few checks if file is not sorted at all
raise IOError(f"Genome annotation gtf file is not sorted correctly! Run the following command:\nsort -k1,1 -k7,7 -k4,4n -o [GTF_OUTFILE] [GTF_INFILE]")
else:
logging.debug(f"Done with {curr_chromstrand} [line {nth_line-1}]")
self.assign_indexes_to_genes(features)
self.annotations_by_chrm_strand[curr_chromstrand] = features
logging.debug(f"Seen {len(self.geneid2ix)} genes until now")
features = OrderedDict()
logging.debug(f"Parsing Chromosome {chrom} strand {strand} [line {nth_line}]")
curr_chromstrand = chrom + strand
if feature_type in ("exon"):
trid = regex_trid.search(tags).group(1)
_trname_search = regex_trname.search(tags)
if _trname_search is None:
trname = trid
else:
trname = _trname_search.group(1)
geneid = regex_geneid.search(tags).group(1)
_genename_search = regex_genename.search(tags)
if _genename_search is None:
genename = geneid
else:
genename = _genename_search.group(1)
try:
exonno = regex_exonno.search(tags).group(1)
except AttributeError:
# NOTE: Don't try to release this constraint, velocyto relies on it for safe calculations! Rather make a utility script that does putative annotation separatelly.
raise IOError("The genome annotation .gtf file provided does not contain exon_number. `exon_number` is described as a mandatory field by GENCODE gtf file specification and we rely on it for easier processing")
start = int(start_str)
end = int(end_str)
chromstrand = chrom + strand
try:
features[trid].append_exon(vcy.Feature(start=start, end=end, kind=ord("e"), exin_no=exonno))
except KeyError:
features[trid] = vcy.TranscriptModel(trid=trid, trname=trname, geneid=geneid, genename=genename, chromstrand=chromstrand)
features[trid].append_exon(vcy.Feature(start=start, end=end, kind=ord("e"), exin_no=exonno))
# Do it for the last chromosome
self.assign_indexes_to_genes(features)
self.annotations_by_chrm_strand[curr_chromstrand] = features
logging.debug(f"Done with {curr_chromstrand} [line {nth_line-1}]")
logging.debug(f"Fixing corner cases of transcript models containg intron longer than {vcy.LONGEST_INTRON_ALLOWED//1000}Kbp")
# Fix corner cases of extremelly long introns ~1Mbp that would be masking genes that are found internally
for tmodels_orddict in self.annotations_by_chrm_strand.values():
for tm in tmodels_orddict.values():
tm.chop_if_long_intron() # Change it in place
# Respect the sorting it had before
# NOTE: not sure it is needed downstream anymore also not sure it guarantees exactly the same order
for chromstrand in self.annotations_by_chrm_strand.keys():
tmp = OrderedDict((i.trid, i) for i in sorted(self.annotations_by_chrm_strand[chromstrand].values()))
self.annotations_by_chrm_strand[chromstrand] = tmp
return self.annotations_by_chrm_strand
[docs] def peek_and_correct(self, gtf_lines: List[str]) -> List[str]:
"""Look at the first 20 instances of a list of lines of a gtf file to dermine if exon number is specified as it should.
If econ number is not contained it will infer the exon number sorting the list by lexicographic ordering tr_id, start, end
Arguments
---------
gtf_lines:
a list of the lines of a gtf file
Returns
-------
gtf_lines:
the same list or the list corrected with added a exon number (filtered to contain only exons)
"""
regex_exonno = re.compile('exon_number "*?([\w]+)')
flag = False
# Scan the first 500 lines for an occurrence of exon
for lin in gtf_lines[:500]:
chrom, feature_class, feature_type, start_str, end_str, junk, strand, junk, tags = lin.split("\t")
if feature_type == "exon":
exonno = regex_exonno.search(tags)
if exonno is None:
flag = True
# If at least one exon was missing the exon number create the entry for all the others
if flag:
logging.warning("The entry exon_number was not present in the gtf file. It will be infferred from the position.")
regex_trid = re.compile('transcript_id "([^"]+)"')
min_info_lines_minus: List[List] = []
min_info_lines_plus: List[List] = []
for lin in gtf_lines:
chrom, feature_class, feature_type, start_str, end_str, junk, strand, junk, tags = lin.split("\t")
if feature_type == "exon":
try:
trid = regex_trid.search(tags).group(1)
except AttributeError:
raise AttributeError(f"transcript_id entry not found in line: {lin}")
if strand == "-":
min_info_lines_minus.append([trid, int(start_str), int(end_str), lin])
else:
min_info_lines_plus.append([trid, int(start_str), int(end_str), lin])
min_info_lines_minus = sorted(min_info_lines_minus)
min_info_lines_plus = sorted(min_info_lines_plus)
current_trid = "None"
exon_n = 1
modified_lines_plus: List[str] = []
for i in min_info_lines_plus:
if current_trid != i[0]:
current_trid = i[0]
exon_n = 1
else:
exon_n += 1
modified_lines_plus.append(f'{i[3][:-1]} exon_number "{exon_n}";\n')
exon_n = 1
modified_lines_minus: List[str] = []
for i in min_info_lines_minus[::-1]:
if current_trid != i[0]:
current_trid = i[0]
exon_n = 1
else:
exon_n += 1
modified_lines_plus.append(f'{i[3][:-1]} exon_number "{exon_n}";\n')
return modified_lines_plus + modified_lines_minus
else:
return gtf_lines
[docs] def mark_up_introns(self, bamfile: Tuple[str], multimap: bool) -> None:
""" Mark up introns that have reads across exon-intron junctions
Arguments
---------
bamfile: Tuple[str]
path to the bam files to markup
logic: vcy.Logic
The logic object to use, changes in different techniques / levels of strictness
NOTE: Right now it is not used
Returns
-------
Nothing it just add to validation to the vcy.Features
Note
----
Situations not considered:
# If an the exon is so short that is possible to get both exonA-exonB junction and exonB-intronB boundary in the same read
"""
if not self.logic.perform_validation_markup:
return
else:
# Since I support multiple files (Smart seq2 it makes sense here to load the feature indexes into memory)
# NOTE this means that maybe I could do this once at a level above
# NOTE if this is not done in count then I need to bring it before the if/else statement
self.feature_indexes: DefaultDict[str, vcy.FeatureIndex] = defaultdict(vcy.FeatureIndex)
for chromstrand_key, annotions_ordered_dict in self.annotations_by_chrm_strand.items():
self.feature_indexes[chromstrand_key] = vcy.FeatureIndex(sorted(chain.from_iterable(annotions_ordered_dict.values())))
# VERBOSE: dump_list = []
# Read the file
currchrom = ""
set_chromosomes_seen: Set[str] = set()
for r in self.iter_alignments(bamfile, unique=not multimap):
# Don't consider spliced reads (exonic) in this step
# NOTE Can the exon be so short that we get splicing and exon-intron boundary
if r is None:
# This happens only when there is a change of file
currchrom = ""
set_chromosomes_seen = set()
# I need to reset indexes in position before the next file is restarted
# NOTE this is far from optimal for lots of cells
logging.debug("End of file. Reset index: start scanning from initial position.")
for chromstrand_key, annotions_ordered_dict in self.annotations_by_chrm_strand.items():
self.feature_indexes[chromstrand_key].reset()
continue
if r.is_spliced:
continue
# When the chromosome mapping of the read changes, change interval index.
if r.chrom != currchrom:
if r.chrom in set_chromosomes_seen:
raise IOError(f"Input .bam file should be chromosome-sorted. (Hint: use `samtools sort {bamfile}`)")
set_chromosomes_seen.add(r.chrom)
logging.debug(f"Marking up chromosome {r.chrom}")
currchrom = r.chrom
if currchrom + '+' not in self.annotations_by_chrm_strand:
logging.warn(f"The .bam file refers to a chromosome '{currchrom}+' not present in the annotation (.gtf) file")
iif = vcy.FeatureIndex([])
else:
iif = self.feature_indexes[currchrom + '+']
if currchrom + '-' not in self.annotations_by_chrm_strand:
logging.warn(f"The .bam file refers to a chromosome '{currchrom}-' not present in the annotation (.gtf) file")
iir = vcy.FeatureIndex([])
else:
iir = self.feature_indexes[currchrom + '-']
# Consider the correct strand
ii = iif if r.strand == '+' else iir
# VERBOSE: # Look for overlap between the intervals and the read
# VERBOSE: dump_list += ii.mark_overlapping_ivls(r)
ii.mark_overlapping_ivls(r)
# VERBOSE: import pickle
# VERBOSE: pickle.dump(dump_list, open("dump_mark_overlapping_ivls.pickle", "wb"))
[docs] def count(self, bamfile: Tuple[str], multimap: bool, cell_batch_size: int=100, molecules_report: bool=False) -> Tuple[Dict[str, List[np.ndarray]], List[str]]:
""" Do the counting of molecules
Arguments
---------
bamfile: str
path to the bam files to markup
cell_batch_size: int, default = 50
it defines whether to require or not exon-intron spanning read to consider an intron valid.
Returns
-------
dict_list_arrays, cell_bcs_order
Note
----
The memory footprint could be reduced allocating scipy sparse arrays
"""
# Initialize variables and containers
self.cell_batch: Set[str] = set()
self.reads_to_count: List[vcy.Read] = []
# self.cells_since_last_count = 0
# Analysis is cell wise so the Feature Index swapping is happening often and it is worth to preload everything in memory
# NOTE: for the features this was already done at markup time, maybe I should just reset them
self.feature_indexes: DefaultDict[str, vcy.FeatureIndex] = defaultdict(vcy.FeatureIndex)
for chromstrand_key, annotions_ordered_dict in self.annotations_by_chrm_strand.items():
self.feature_indexes[chromstrand_key] = vcy.FeatureIndex(sorted(chain.from_iterable(annotions_ordered_dict.values())))
self.mask_indexes: DefaultDict[str, vcy.FeatureIndex] = defaultdict(vcy.FeatureIndex)
for chromstrand_key, annotions_list in self.mask_ivls_by_chromstrand.items():
self.mask_indexes[chromstrand_key] = vcy.FeatureIndex(annotions_list) # This suould be sorted
logging.debug(f"Features available for chromosomes : {list(self.feature_indexes.keys())}")
logging.debug(f"Mask available for chromosomes : {list(self.mask_indexes.keys())}")
# Before counting, report how many features where validated
logging.debug(f"Summarizing the results of intron validation.")
n_is_intron = 0
n_is_intron_valid = 0
unique_valid = set()
for chromstrand_key, feature_index in self.feature_indexes.items():
for ivl in feature_index.ivls:
if ivl.kind == ord("i"):
n_is_intron += 1
if ivl.is_validated:
n_is_intron_valid += 1
unique_valid.add((ivl.start, ivl.end))
logging.debug(f"Validated {n_is_intron_valid} introns (of which unique intervals {len(unique_valid)}) out of {n_is_intron} total possible introns (considering each possible transcript models).")
cell_bcs_order: List[str] = []
dict_list_arrays: Dict[str, List[np.ndarray]] = {layer_name: [] for layer_name in self.logic.layers}
nth = 0
# Loop through the aligment of the bamfile
for r in self.iter_alignments(bamfile, unique=not multimap):
if (r is None) or (len(self.cell_batch) == cell_batch_size and r.bc not in self.cell_batch):
# Perfrom the molecule counting
nth += 1
logging.debug(f"Counting for batch {nth}, containing {len(self.cell_batch)} cells and {len(self.reads_to_count)} reads")
dict_layer_columns, list_bcs = self.count_cell_batch()
# This is to avoid crazy big matrix output if the barcode selection is not chosen
if not self.filter_mode:
logging.warning("The barcode selection mode is off, no cell events will be identified by <80 counts")
tot_mol = dict_layer_columns["spliced"].sum(0) + dict_layer_columns["unspliced"].sum(0)
cell_bcs_order += list(np.array(list_bcs)[tot_mol > 80])
for layer_name, layer_columns in dict_layer_columns.items():
dict_list_arrays[layer_name].append(layer_columns[:, tot_mol > 80])
logging.warning(f"{np.sum(tot_mol < 80)} of the barcodes where without cell")
else:
# The normal case
cell_bcs_order += list_bcs
for layer_name, layer_columns in dict_layer_columns.items():
dict_list_arrays[layer_name].append(layer_columns)
self.cell_batch = set()
# Drop the counted reads (If there are no other reference to it) and reset the indexes to 0
self.reads_to_count = []
for chromstrand_key, annotions_ordered_dict in self.annotations_by_chrm_strand.items():
self.feature_indexes[chromstrand_key].reset()
for chromstrand_key, annotions_list in self.mask_ivls_by_chromstrand.items():
self.mask_indexes[chromstrand_key].reset()
if r is not None:
self.cell_batch.add(r.bc)
self.reads_to_count.append(r)
# NOTE: Since iter_allignments is yielding None at each file change (even when only one bamfile) I do not need the following
# logging.debug(f"Counting molecule for last batch of {len(self.cell_batch)}, total reads {len(self.reads_to_count)}")
# spliced, unspliced, ambiguous, list_bcs = self.count_cell_batch()
# cell_bcs_order += list_bcs
# list_spliced_arrays.append(spliced)
# list_unspliced_arrays.append(unspliced)
# list_ambiguous_arrays.append(ambiguous)
# self.cell_batch = set()
# self.reads_to_count = []
logging.debug(f"Counting done!")
return dict_list_arrays, cell_bcs_order
def _count_cell_batch_stranded(self) -> Tuple[Dict[str, np.ndarray], List[str]]:
"""It performs molecule counting for the current batch of cells in the case of stranded method
Returns
-------
dict_layers_columns: Dict[str, np.ndarray]
name_layer->np.ndarray of the batch
idx2bc: List[str]
list of barcodes
NOTE This duplications of method is bad for code mantainance
"""
molitems: DefaultDict[str, vcy.Molitem] = defaultdict(vcy.Molitem)
# Sort similarly to what the sort linux command would do. (implemented using Read.__lt__)
# NOTE NOTE!!!! Here I changed the way to sort because it was using the strand causing to skip a lot of reads in SmartSeq2
self.reads_to_count.sort()
# NOTE: I could start by sorting the reads by chromosome, strand, position but for now let's see if it is fast without doing do
repeats_reads_count = 0
for r in self.reads_to_count:
# Consider the correct strand
ii = self.feature_indexes[r.chrom + r.strand]
iim = self.mask_indexes[r.chrom + r.strand]
# Check if read is fully inside a masked region, in that case skip it
if iim.has_ivls_enclosing(r):
repeats_reads_count += 1 # VERBOSE
continue
# Look for overlap between the intervals and the read
mappings_record = ii.find_overlapping_ivls(r)
if len(mappings_record):
# logging.debug("IN: Non empty mapping record")
bcumi = f"{r.bc}${r.umi}"
molitems[bcumi].add_mappings_record(mappings_record)
# if len(molitems[bcumi].mappings_record):
# logging.debug("OUT: Non empty")
# else:
# logging.debug("OUT: Empty")
logging.debug(f"{repeats_reads_count} reads not considered because fully enclosed in repeat masked regions") # VERBOSE
# initialize np.ndarray
shape = (len(self.geneid2ix), len(self.cell_batch))
dict_layers_columns: Dict[str, np.ndarray] = {}
for layer_name in self.logic.layers:
dict_layers_columns[layer_name] = np.zeros(shape, dtype=self.loom_numeric_dtype, order="C")
bc2idx: Dict[str, int] = dict(zip(self.cell_batch, range(len(self.cell_batch))))
# After the whole file has been read, do the actual counting
failures = 0
counter: Counter = Counter()
for bcumi, molitem in molitems.items():
bc = bcumi.split("$")[0] # extract the bc part from the bc+umi
bcidx = bc2idx[bc]
rcode = self.logic.count(molitem, bcidx, dict_layers_columns, self.geneid2ix)
if rcode:
failures += 1
counter[rcode] += 1
# before it was molitem.count(bcidx, spliced, unspliced, ambiguous, self.geneid2ix)
if failures > (0.25 * len(molitems)):
logging.warn(f"More than 20% ({(100*failures / len(molitems)):.1f}%) of molitems trashed, of those:")
logging.warn(f"A situation where many genes were compatible with the observation in {(100*counter[1] / len(molitems)):.1f} cases")
logging.warn(f"No gene is compatible with the observation in {(100*counter[2] / len(molitems)):.1f} cases")
logging.warn(f"Observation compatible with more genes {(100*counter[3] / len(molitems)):.1f} of the cases")
logging.warn(f"Situation that were not described by the logic in the {(100*counter[4] / len(molitems)):.1f} of the cases")
if self.every_n_report and ((self.report_state % self.every_n_report) == 0):
if self.kind_of_report == "p":
import pickle
first_cell_batch = next(iter(molitems.keys())).split("$")[0]
if not os.path.exists("pickle_dump"):
os.makedirs("pickle_dump")
pickle.dump(molitems, open(f"pickle_dump/molitems_dump_{first_cell_batch}.pickle", "wb"))
pickle.dump(self.reads_to_count, open(f"pickle_dump/reads_to_count{first_cell_batch}.pickle", "wb"))
else:
if not os.path.exists(os.path.join(self.outputfolder, "dump")):
os.makedirs(os.path.join(self.outputfolder, "dump"))
f = h5py.File(os.path.join(self.outputfolder, f"dump/{self.sampleid}.hdf5")) # From the docs: Read/write if exists, create otherwise (default)
if "info/tr_id" not in f:
logging.warning("The hdf5 report is less accurate in reporting exactly all the information than the pickle.")
info_tr_id = []
info_features_gene = []
info_is_last3prime = []
info_is_intron = []
info_start_end = []
info_exino = []
info_strandplus = []
info_chrm = []
for k, v_dict_tm in self.annotations_by_chrm_strand.items():
for v1_tm in v_dict_tm.values():
for v2_ivl in v1_tm:
info_tr_id.append(v2_ivl.transcript_model.trid) # “info/ivls/tr_id“,
info_features_gene.append(v2_ivl.transcript_model.genename) # “info/ivls/features_gene“,
info_is_last3prime.append(v2_ivl.is_last_3prime) # “info/ivls/is_last3prime“
info_is_intron.append(v2_ivl.kind == 105) # “info/ivls/is_intron“,
info_start_end.append((v2_ivl.start, v2_ivl.end)) # “info/ivls/feture_start_end“
info_exino.append(v2_ivl.exin_no) # “info/ivls/exino“
info_strandplus.append(v2_ivl.transcript_model.chromstrand[-1:] == "+") # “info/ivls/strandplus“
info_chrm.append(v2_ivl.transcript_model.chromstrand[:-1]) # “info/ivls/chrm“
self.inv_tridstart2ix: Dict[str, int] = {}
for i in range(len(info_tr_id)):
self.inv_tridstart2ix[f"{info_tr_id[i]}_{info_start_end[i][0]}"] = i
f.create_dataset("info/tr_id", data=np.array(info_tr_id, dtype="S24"),
maxshape=(len(info_tr_id), ), chunks=(500,), compression="gzip", shuffle=False, compression_opts=4)
f.create_dataset("info/features_gene", data=np.array(info_features_gene, dtype="S15"),
maxshape=(len(info_features_gene), ), chunks=(500,), compression="gzip", shuffle=False, compression_opts=4)
f.create_dataset("info/is_last3prime", data=np.array(info_is_last3prime, dtype=bool),
maxshape=(len(info_is_last3prime), ), chunks=(500,), compression="gzip", shuffle=False, compression_opts=4)
f.create_dataset("info/is_intron", data=np.array(info_is_intron, dtype=bool),
maxshape=(len(info_is_intron), ), chunks=(500,), compression="gzip", shuffle=False, compression_opts=4)
f.create_dataset("info/start_end", data=np.array(info_start_end, dtype=np.int64),
maxshape=(len(info_start_end), 2), chunks=(500, 2), compression="gzip", shuffle=False, compression_opts=4)
f.create_dataset("info/exino", data=np.array(info_exino, dtype=np.uint8),
maxshape=(len(info_exino), ), chunks=(500,), compression="gzip", shuffle=False, compression_opts=4)
f.create_dataset("info/strandplus", data=np.array(info_strandplus, dtype=np.bool),
maxshape=(len(info_strandplus), ), chunks=(500,), compression="gzip", shuffle=False, compression_opts=4)
f.create_dataset("info/chrm", data=np.array(info_chrm, dtype="S6"),
maxshape=(len(info_chrm), ), chunks=(500,), compression="gzip", shuffle=False, compression_opts=4)
# cell_name = next(iter(molitems.keys())).split("$")[0]
pos: DefaultDict[str, List[Tuple[int, int]]] = defaultdict(list)
mol: DefaultDict[str, List[int]] = defaultdict(list)
ixs: DefaultDict[str, List[int]] = defaultdict(list)
count_i: int = 0
for mol_bc, molitem in molitems.items():
cell_name = mol_bc.split("$")[0]
try:
for match in next(iter(molitem.mappings_record.items()))[1]:
mol[cell_name].append(count_i)
pos[cell_name].append(match.segment)
ixs[cell_name].append(self.inv_tridstart2ix[f"{match.feature.transcript_model.trid}_{match.feature.start}"])
count_i += 1
except StopIteration:
pass # An empty or chimeric molitem ?
# Do the last cell and close the file
for cell_name in mol.keys():
posA = np.array(pos[cell_name], dtype=np.int32)
ixsA = np.array(ixs[cell_name], dtype=np.intp)
molA = np.array(mol[cell_name], dtype=np.uint32)
f.create_dataset(f'cells/{self.sampleid}_{cell_name}/pos', data=posA, maxshape=posA.shape, chunks=(min(500, posA.shape[0]), 2), compression="gzip", shuffle=False, compression_opts=4)
f.create_dataset(f'cells/{self.sampleid}_{cell_name}/ixs', data=ixsA, maxshape=ixsA.shape, chunks=(min(500, ixsA.shape[0]),), compression="gzip", shuffle=False, compression_opts=4)
f.create_dataset(f'cells/{self.sampleid}_{cell_name}/mol', data=molA, maxshape=molA.shape, chunks=(min(500, molA.shape[0]),), compression="gzip", shuffle=False, compression_opts=4)
f.close()
self.report_state += 1
idx2bc = {v: k for k, v in bc2idx.items()}
return dict_layers_columns, [idx2bc[i] for i in range(len(idx2bc))]
def _count_cell_batch_stranded_discordant(self) -> Tuple[Dict[str, np.ndarray], List[str]]:
"""It performs molecule counting for the current batch of cells in the case of stranded method with discordant masking
Returns
-------
dict_layers_columns: Dict[str, np.ndarray]
name_layer->np.ndarray of the batch
idx2bc: List[str]
list of barcodes
NOTE This duplications of method is bad for code mantainance
"""
molitems: DefaultDict[str, vcy.Molitem] = defaultdict(vcy.Molitem)
# Sort similarly to what the sort linux command would do. (implemented using Read.__lt__)
self.reads_to_count.sort()
# NOTE: I could start by sorting the reads by chromosome, strand, position but for now let's see if it is fast without doing do
repeats_reads_count = 0
for r in self.reads_to_count:
# Consider the correct strand
ii = self.feature_indexes[f"{r.chrom}{r.strand}"]
iir = self.feature_indexes[f"{r.chrom}{reverse(r.strand)}"]
iim = self.mask_indexes[f"{r.chrom}{r.strand}"]
iimr = self.mask_indexes[f"{r.chrom}{reverse(r.strand)}"]
# Check if read is fully inside a masked region, in that case check if it is allowed in the reverse
if iim.has_ivls_enclosing(r):
repeats_reads_count += 1 # VERBOSE
if not iimr.has_ivls_enclosing(r):
mappings_record = iir.find_overlapping_ivls(r)
else:
continue
else:
# Look for overlap between the intervals and the read
mappings_record = ii.find_overlapping_ivls(r)
if len(mappings_record):
# logging.debug("IN: Non empty mapping record")
bcumi = f"{r.bc}${r.umi}"
molitems[bcumi].add_mappings_record(mappings_record)
# if len(molitems[bcumi].mappings_record):
# logging.debug("OUT: Non empty")
# else:
# logging.debug("OUT: Empty")
logging.debug(f"{repeats_reads_count} reads not considered because fully enclosed in repeat masked regions") # VERBOSE
# initialize np.ndarray
shape = (len(self.geneid2ix), len(self.cell_batch))
dict_layers_columns: Dict[str, np.ndarray] = {}
for layer_name in self.logic.layers:
dict_layers_columns[layer_name] = np.zeros(shape, dtype=self.loom_numeric_dtype, order="C")
bc2idx: Dict[str, int] = dict(zip(self.cell_batch, range(len(self.cell_batch))))
# After the whole file has been read, do the actual counting
for bcumi, molitem in molitems.items():
bc = bcumi.split("$")[0] # extract the bc part from the bc+umi
bcidx = bc2idx[bc]
self.logic.count(molitem, bcidx, dict_layers_columns, self.geneid2ix)
# NOTE I need to generalize this to any set of layers
# before it was molitem.count(bcidx, spliced, unspliced, ambiguous, self.geneid2ix)
if self.every_n_report and ((self.report_state % self.every_n_report) == 0):
if self.kind_of_report == "p":
import pickle
first_cell_batch = next(iter(molitems.keys())).split("$")[0]
if not os.path.exists("pickle_dump"):
os.makedirs("pickle_dump")
pickle.dump(molitems, open(f"pickle_dump/molitems_dump_{first_cell_batch}.pickle", "wb"))
pickle.dump(self.reads_to_count, open(f"pickle_dump/reads_to_count{first_cell_batch}.pickle", "wb"))
else:
if not os.path.exists(os.path.join(self.outputfolder, "dump")):
os.makedirs(os.path.join(self.outputfolder, "dump"))
f = h5py.File(os.path.join(self.outputfolder, f"dump/{self.sampleid}.hdf5")) # From the docs: Read/write if exists, create otherwise (default)
if "info/tr_id" not in f:
logging.warning("The hdf5 report is less accurate in reporting exactly all the information than the pickle.")
info_tr_id = []
info_features_gene = []
info_is_last3prime = []
info_is_intron = []
info_start_end = []
info_exino = []
info_strandplus = []
info_chrm = []
for k, v_dict_tm in self.annotations_by_chrm_strand.items():
for v1_tm in v_dict_tm.values():
for v2_ivl in v1_tm:
info_tr_id.append(v2_ivl.transcript_model.trid) # “info/ivls/tr_id“,
info_features_gene.append(v2_ivl.transcript_model.genename) # “info/ivls/features_gene“,
info_is_last3prime.append(v2_ivl.is_last_3prime) # “info/ivls/is_last3prime“
info_is_intron.append(v2_ivl.kind == 105) # “info/ivls/is_intron“,
info_start_end.append((v2_ivl.start, v2_ivl.end)) # “info/ivls/feture_start_end“
info_exino.append(v2_ivl.exin_no) # “info/ivls/exino“
info_strandplus.append(v2_ivl.transcript_model.chromstrand[-1:] == "+") # “info/ivls/strandplus“
info_chrm.append(v2_ivl.transcript_model.chromstrand[:-1]) # “info/ivls/chrm“
self.inv_tridstart2ix: Dict[str, int] = {}
for i in range(len(info_tr_id)):
self.inv_tridstart2ix[f"{info_tr_id[i]}_{info_start_end[i][0]}"] = i
f.create_dataset("info/tr_id", data=np.array(info_tr_id, dtype="S24"),
maxshape=(len(info_tr_id), ), chunks=(500,), compression="gzip", shuffle=False, compression_opts=4)
f.create_dataset("info/features_gene", data=np.array(info_features_gene, dtype="S15"),
maxshape=(len(info_features_gene), ), chunks=(500,), compression="gzip", shuffle=False, compression_opts=4)
f.create_dataset("info/is_last3prime", data=np.array(info_is_last3prime, dtype=bool),
maxshape=(len(info_is_last3prime), ), chunks=(500,), compression="gzip", shuffle=False, compression_opts=4)
f.create_dataset("info/is_intron", data=np.array(info_is_intron, dtype=bool),
maxshape=(len(info_is_intron), ), chunks=(500,), compression="gzip", shuffle=False, compression_opts=4)
f.create_dataset("info/start_end", data=np.array(info_start_end, dtype=np.int64),
maxshape=(len(info_start_end), 2), chunks=(500, 2), compression="gzip", shuffle=False, compression_opts=4)
f.create_dataset("info/exino", data=np.array(info_exino, dtype=np.uint8),
maxshape=(len(info_exino), ), chunks=(500,), compression="gzip", shuffle=False, compression_opts=4)
f.create_dataset("info/strandplus", data=np.array(info_strandplus, dtype=np.bool),
maxshape=(len(info_strandplus), ), chunks=(500,), compression="gzip", shuffle=False, compression_opts=4)
f.create_dataset("info/chrm", data=np.array(info_chrm, dtype="S6"),
maxshape=(len(info_chrm), ), chunks=(500,), compression="gzip", shuffle=False, compression_opts=4)
# cell_name = next(iter(molitems.keys())).split("$")[0]
pos: DefaultDict[str, List[Tuple[int, int]]] = defaultdict(list)
mol: DefaultDict[str, List[int]] = defaultdict(list)
ixs: DefaultDict[str, List[int]] = defaultdict(list)
count_i: int = 0
for mol_bc, molitem in molitems.items():
cell_name = mol_bc.split("$")[0]
try:
for match in next(iter(molitem.mappings_record.items()))[1]:
mol[cell_name].append(count_i)
pos[cell_name].append(match.segment)
ixs[cell_name].append(self.inv_tridstart2ix[f"{match.feature.transcript_model.trid}_{match.feature.start}"])
count_i += 1
except StopIteration:
pass # An empty or chimeric molitem ?
# Do the last cell and close the file
for cell_name in mol.keys():
posA = np.array(pos[cell_name], dtype=np.int32)
ixsA = np.array(ixs[cell_name], dtype=np.intp)
molA = np.array(mol[cell_name], dtype=np.uint32)
f.create_dataset(f'cells/{self.sampleid}_{cell_name}/pos', data=posA, maxshape=posA.shape, chunks=(min(500, posA.shape[0]), 2), compression="gzip", shuffle=False, compression_opts=4)
f.create_dataset(f'cells/{self.sampleid}_{cell_name}/ixs', data=ixsA, maxshape=ixsA.shape, chunks=(min(500, ixsA.shape[0]),), compression="gzip", shuffle=False, compression_opts=4)
f.create_dataset(f'cells/{self.sampleid}_{cell_name}/mol', data=molA, maxshape=molA.shape, chunks=(min(500, molA.shape[0]),), compression="gzip", shuffle=False, compression_opts=4)
f.close()
self.report_state += 1
idx2bc = {v: k for k, v in bc2idx.items()}
return dict_layers_columns, [idx2bc[i] for i in range(len(idx2bc))]
def _count_cell_batch_non_stranded(self) -> Tuple[Dict[str, np.ndarray], List[str]]:
"""It performs molecule counting for the current batch of cells in the case of non stranded method
Returns
-------
dict_layers_columns: Dict[str, np.ndarray]
name_layer->np.ndarray of the batch
idx2bc: List[str]
list of barcodes
"""
molitems: DefaultDict[str, vcy.Molitem] = defaultdict(vcy.Molitem)
# Sort similarly to what the sort linux command would do. (implemented using Read.__lt__)
# NOTE NOTE!!!! Here I changed the way to sort because it was using the strand causing to skip a lot of reads in SmartSeq2
self.reads_to_count.sort()
# NOTE: I could start by sorting the reads by chromosome, strand, position but for now let's see if it is fast without doing do
repeats_reads_count = 0
plus_reads_count = 0
minus_reads_count = 0
both_reads_count = 0
for r in self.reads_to_count:
# Consider the correct strand
ii = self.feature_indexes[f"{r.chrom}{r.strand}"]
iir = self.feature_indexes[f"{r.chrom}{reverse(r.strand)}"]
iim = self.mask_indexes[f"{r.chrom}{r.strand}"]
iimr = self.mask_indexes[f"{r.chrom}{reverse(r.strand)}"]
# Check if read is fully inside a masked region, in that case skip it
if iim.has_ivls_enclosing(r) or iimr.has_ivls_enclosing(r):
repeats_reads_count += 1 # VERBOSE
continue
# Look for overlap between the intervals and the read
mappings_record = ii.find_overlapping_ivls(r)
if len(mappings_record):
bcumi = f"{r.bc}${r.umi}"
molitems[bcumi].add_mappings_record(mappings_record)
if r.strand == "+":
plus_reads_count += 1
else:
minus_reads_count += 1
mappings_record_r = iir.find_overlapping_ivls(r)
if len(mappings_record_r):
bcumi = f"{r.bc}${r.umi}"
molitems[bcumi].add_mappings_record(mappings_record_r)
if r.strand == "-":
plus_reads_count += 1
else:
minus_reads_count += 1
if len(mappings_record) and len(mappings_record_r):
both_reads_count += 1
logging.debug(f"{repeats_reads_count} reads in repeat masked regions") # VERBOSE
logging.debug(f"{plus_reads_count} reads overlapping with features on plus strand") # VERBOSE
logging.debug(f"{minus_reads_count} reads overlapping with features on minus strand") # VERBOSE
logging.debug(f"{both_reads_count} reads overlapping with features on both strands") # VERBOSE
# initialize np.ndarray
shape = (len(self.geneid2ix), len(self.cell_batch))
dict_layers_columns: Dict[str, np.ndarray] = {}
for layer_name in self.logic.layers:
dict_layers_columns[layer_name] = np.zeros(shape, dtype=self.loom_numeric_dtype, order="C")
bc2idx: Dict[str, int] = dict(zip(self.cell_batch, range(len(self.cell_batch))))
# After the whole file has been read, do the actual counting
for bcumi, molitem in molitems.items():
bc = bcumi.split("$")[0] # extract the bc part from the bc+umi
bcidx = bc2idx[bc]
self.logic.count(molitem, bcidx, dict_layers_columns, self.geneid2ix)
# NOTE I need to generalize this to any set of layers
# before it was molitem.count(bcidx, spliced, unspliced, ambiguous, self.geneid2ix)
if self.every_n_report and ((self.report_state % self.every_n_report) == 0):
if self.kind_of_report == "p":
import pickle
first_cell_batch = next(iter(molitems.keys())).split("$")[0]
if not os.path.exists("pickle_dump"):
os.makedirs("pickle_dump")
pickle.dump(molitems, open(f"pickle_dump/molitems_dump_{first_cell_batch}.pickle", "wb"))
pickle.dump(self.reads_to_count, open(f"pickle_dump/reads_to_count{first_cell_batch}.pickle", "wb"))
else:
if not os.path.exists(os.path.join(self.outputfolder, "dump")):
os.makedirs(os.path.join(self.outputfolder, "dump"))
f = h5py.File(os.path.join(self.outputfolder, f"dump/{self.sampleid}.hdf5")) # From the docs: Read/write if exists, create otherwise (default)
if "info/tr_id" not in f:
logging.warning("The hdf5 report is less accurate than the pickle in the completeness of the info it is reporting.")
info_tr_id = []
info_features_gene = []
info_is_last3prime = []
info_is_intron = []
info_start_end = []
info_exino = []
info_strandplus = []
info_chrm = []
for k, v_dict_tm in self.annotations_by_chrm_strand.items():
for v1_tm in v_dict_tm.values():
for v2_ivl in v1_tm:
info_tr_id.append(v2_ivl.transcript_model.trid) # “info/ivls/tr_id“,
info_features_gene.append(v2_ivl.transcript_model.genename) # “info/ivls/features_gene“,
info_is_last3prime.append(v2_ivl.is_last_3prime) # “info/ivls/is_last3prime“
info_is_intron.append(v2_ivl.kind == 105) # “info/ivls/is_intron“,
info_start_end.append((v2_ivl.start, v2_ivl.end)) # “info/ivls/feture_start_end“
info_exino.append(v2_ivl.exin_no) # “info/ivls/exino“
info_strandplus.append(v2_ivl.transcript_model.chromstrand[-1:] == "+") # “info/ivls/strandplus“
info_chrm.append(v2_ivl.transcript_model.chromstrand[:-1]) # “info/ivls/chrm“
self.inv_tridstart2ix: Dict[str, int] = {}
for i in range(len(info_tr_id)):
self.inv_tridstart2ix[f"{info_tr_id[i]}_{info_start_end[i][0]}"] = i
f.create_dataset("info/tr_id", data=np.array(info_tr_id, dtype="S24"),
maxshape=(len(info_tr_id), ), chunks=(500,), compression="gzip", shuffle=False, compression_opts=4)
f.create_dataset("info/features_gene", data=np.array(info_features_gene, dtype="S15"),
maxshape=(len(info_features_gene), ), chunks=(500,), compression="gzip", shuffle=False, compression_opts=4)
f.create_dataset("info/is_last3prime", data=np.array(info_is_last3prime, dtype=bool),
maxshape=(len(info_is_last3prime), ), chunks=(500,), compression="gzip", shuffle=False, compression_opts=4)
f.create_dataset("info/is_intron", data=np.array(info_is_intron, dtype=bool),
maxshape=(len(info_is_intron), ), chunks=(500,), compression="gzip", shuffle=False, compression_opts=4)
f.create_dataset("info/start_end", data=np.array(info_start_end, dtype=np.int64),
maxshape=(len(info_start_end), 2), chunks=(500, 2), compression="gzip", shuffle=False, compression_opts=4)
f.create_dataset("info/exino", data=np.array(info_exino, dtype=np.uint8),
maxshape=(len(info_exino), ), chunks=(500,), compression="gzip", shuffle=False, compression_opts=4)
f.create_dataset("info/strandplus", data=np.array(info_strandplus, dtype=np.bool),
maxshape=(len(info_strandplus), ), chunks=(500,), compression="gzip", shuffle=False, compression_opts=4)
f.create_dataset("info/chrm", data=np.array(info_chrm, dtype="S6"),
maxshape=(len(info_chrm), ), chunks=(500,), compression="gzip", shuffle=False, compression_opts=4)
# cell_name = next(iter(molitems.keys())).split("$")[0]
pos: DefaultDict[str, List[Tuple[int, int]]] = defaultdict(list)
mol: DefaultDict[str, List[int]] = defaultdict(list)
ixs: DefaultDict[str, List[int]] = defaultdict(list)
count_i: int = 0
for mol_bc, molitem in molitems.items():
cell_name = mol_bc.split("$")[0]
try:
for match in next(iter(molitem.mappings_record.items()))[1]:
mol[cell_name].append(count_i)
pos[cell_name].append(match.segment)
ixs[cell_name].append(self.inv_tridstart2ix[f"{match.feature.transcript_model.trid}_{match.feature.start}"])
count_i += 1
except StopIteration:
pass # An empty or chimeric molitem ?
# Do the last cell and close the file
for cell_name in mol.keys():
posA = np.array(pos[cell_name], dtype=np.int32)
ixsA = np.array(ixs[cell_name], dtype=np.intp)
molA = np.array(mol[cell_name], dtype=np.uint32)
f.create_dataset(f'cells/{self.sampleid}_{cell_name}/pos', data=posA, maxshape=posA.shape, chunks=(min(500, posA.shape[0]), 2), compression="gzip", shuffle=False, compression_opts=4)
f.create_dataset(f'cells/{self.sampleid}_{cell_name}/ixs', data=ixsA, maxshape=ixsA.shape, chunks=(min(500, ixsA.shape[0]),), compression="gzip", shuffle=False, compression_opts=4)
f.create_dataset(f'cells/{self.sampleid}_{cell_name}/mol', data=molA, maxshape=molA.shape, chunks=(min(500, molA.shape[0]),), compression="gzip", shuffle=False, compression_opts=4)
f.close()
self.report_state += 1
idx2bc = {v: k for k, v in bc2idx.items()}
return dict_layers_columns, [idx2bc[i] for i in range(len(idx2bc))]
[docs] def pcount(self, samfile: str, cell_batch_size: int=50, molecules_report: bool=False, n_processes: int=4) -> Tuple[np.ndarray, np.ndarray, np.ndarray, List[str]]:
""" Do the counting of molecules in parallel using multiprocessing
"""
raise NotImplementedError("Implement this using multiprocessiong")
[docs] def pcount_cell_batch(self) -> Tuple[np.ndarray, np.ndarray, np.ndarray, List[str]]:
"""It performs molecule counting for the current batch of cells
"""
raise NotImplementedError("This will be a used by .pcount")
[docs]def reverse(strand: str) -> str:
if strand == "+":
return "-"
elif strand == "-":
return "+"
else:
raise ValueError(f"Unknown strand {strand}")