def make_ngs_error_examples(ref_path, vcf_path, bam_path):
""" Yields tf.Example for training a ML model.
Each tf.Example contains
relevant features aboout the ngs read.
Args:
ref_path: str. A path to an indexed fasta file.
vcf_path: str. A path to an indexed VCF file.
bam_path: str. A path to an SAM/BAM file.
Yields:
A tuple (example, ngs_read_length, has_error), where example is a
tf.Example, ngs_read_length is the length of the read generated by the
sequencer, and has_error is a boolean specifying whether the example
contains a read error.
"""
# Create a ref_reader backed by ref.
ref_reader = fasta.IndexedFastaReader(ref_path)
# Create a vcf_reader backed by vcf.
vcf_reader = vcf.VcfReader(vcf_path)
# Create a sam_reader backed by bam. Provide an empty ReadRequirements
# proto to the reader so it enables standard filtering based on the default
# values of ReadRequirements. Also explicitly allow the reader to access an
# unindexed BAM, so only the iterate() function is enabled.
read_requirements = reads_pb2.ReadRequirements()
sam_reader = sam.SamReader(bam_path, read_requirements=read_requirements)
# All our readers and writers are context managers, so use the `with`
# construct to open all of the inputs/outputs and close them when we are done
# looping over our reads.
with ref_reader, vcf_reader, sam_reader:
# Loop over the reads in our BAM file:
for read in sam_reader.iterate():
# Get the Range proto describing the chrom/start/stop spanned by our read.
assert len(read.alignment.cigar) > 0
first_cigar = read.alignment.cigar[0]
# If the first cigar is a CLIP_SOFT, the start of sequence is the cigar
# operation length before the alignment position.
start = read.alignment.position.position
if first_cigar.operation == cigar_pb2.CigarUnit.CLIP_SOFT:
start -= first_cigar.operation_length
read_range = ranges.make_range(read.alignment.position.reference_name,
start, start + len(read.aligned_sequence))
# Get all of the variants that overlap our read range.
variants = list(vcf_reader.query(read_range))
# Get the reference bases spanned by our read.
ref_bases = ref_reader.query(read_range)
# Check that we can use our read for generating an example.
if is_usable_training_example(read, variants, ref_bases):
# Convert read and ref_bases to a tf.Example with make_example.
yield make_example(read, ref_bases), len(read.aligned_sequence), (
read.aligned_sequence != ref_bases)
def make_ngs_examples(hparams):
"""Generator function that yields training, evaluation and test examples."""
ref_reader = fasta.IndexedFastaReader(input_path=hparams.ref_path)
vcf_reader = vcf.VcfReader(input_path=hparams.vcf_path)
read_requirements = reads_pb2.ReadRequirements()
sam_reader = sam.SamReader(input_path=hparams.bam_path,
read_requirements=read_requirements)
# Use a separate SAM reader to query for reads falling in the pileup range.
sam_query_reader = sam.SamReader(input_path=hparams.bam_path,
read_requirements=read_requirements)
used_pileup_ranges = set()
with ref_reader, vcf_reader, sam_reader, sam_query_reader:
for read in sam_reader:
# Check that read has cigar string present and allowed alignment.
if not read.alignment.cigar:
print('Skipping read, no cigar alignment found')
continue
if not has_allowed_alignment(read):
continue
# Obtain window that will be used to construct an example.
read_range = utils.read_range(read)
ref = ref_reader.query(region=read_range)
pileup_range = get_pileup_range(hparams, read, read_range, ref)
# Do not construct multiple examples with the same pileup range.
pileup_range_serialized = pileup_range.SerializeToString()
if pileup_range_serialized in used_pileup_ranges:
continue
used_pileup_ranges.add(pileup_range_serialized)
# Get reference sequence, reads, and truth variants for the pileup range.
pileup_reads = list(sam_query_reader.query(region=pileup_range))
pileup_ref = ref_reader.query(region=pileup_range)
pileup_variants = list(vcf_reader.query(region=pileup_range))
if is_usable_example(pileup_reads, pileup_variants, pileup_ref):
yield make_example(hparams, pileup_reads, pileup_ref,
pileup_range)
def make_ngs_error_examples(ref_path,
vcf_path,
bam_path,
examples_out_path,
max_reads=None):
"""Driver program for ngs_errors.
See module description for details.
Args:
ref_path: str. A path to an indexed fasta file.
vcf_path: str. A path to an indexed VCF file.
bam_path: str. A path to an SAM/BAM file.
examples_out_path: str. A path where we will write out examples.
max_reads: int or None. If not None, we will emit at most max_reads examples
to examples_out_path.
"""
# Create a ref_reader backed by ref.
ref_reader = fasta.IndexedFastaReader(ref_path)
# Create a vcf_reader backed by vcf.
vcf_reader = vcf.VcfReader(vcf_path)
# Create a sam_reader backed by bam. Provide an empty ReadRequirements
# proto to the reader so it enables standard filtering based on the default
# values of ReadRequirements. Also explicitly allow the reader to access an
# unindexed BAM, so only the iterate() function is enabled.
read_requirements = reads_pb2.ReadRequirements()
sam_reader = sam.SamReader(bam_path, read_requirements=read_requirements)
# Create our TFRecordWriter where we'll send our tf.Examples.
examples_out = genomics_writer.TFRecordWriter(examples_out_path)
# All our readers and writers are context managers, so use the `with`
# construct to open all of the inputs/outputs and close them when we are done
# looping over our reads.
n_examples = 0
with ref_reader, vcf_reader, sam_reader, examples_out:
# Loop over the reads in our BAM file:
for i, read in enumerate(sam_reader.iterate(), start=1):
# Get the Range proto describing the chrom/start/stop spanned by our read.
read_range = utils.read_range(read)
# Get all of the variants that overlap our read range.
variants = list(vcf_reader.query(read_range))
# Get the reference bases spanned by our read.
ref_bases = ref_reader.query(read_range)
# Check that we can use our read for generating an example.
if is_usable_training_example(read, variants, ref_bases):
n_examples += 1
# Convert read and ref_bases to a tf.Example with make_example.
example = make_example(read, ref_bases)
# And write it out to our TFRecord output file.
examples_out.write(example)
# Do a bit of convenient logging. This is very verbose if we convert a
# lot of reads...
logging.info((
'Added an example for read %s (span=%s) with cigar %s [%d added '
'of %d total reads]'), read.fragment_name,
ranges.to_literal(read_range),
cigar.format_cigar_units(read.alignment.cigar),
n_examples, i)
if max_reads is not None and n_examples >= max_reads:
return