def get_guess(self):
"""Random guess to initialise optimisation"""
params = {}
m = self.data.min()
M = self.data.max()
range_ = M - m
mus = [m + range_ / (self.k + 1.) * i for i in range(1, self.k + 1)]
params['mus'] = mus
sigma = range_ / float(self.k + 1) / 2.
params['sigmas'] = [sigma] * self.k
params['pis'] = [1. / self.k] * self.k
params = [[mu, sigma, pi] for mu, sigma, pi in zip(
params['mus'], params['sigmas'], params['pis'])]
params = list(pylab.flatten(params))
return params
def _get_df(self):
# When scanning the BAM, we can extract the length, SNR of ACGT (still
# need to know how to use it). The GC content (note there is no
# ambiguity so no S character). The ZMW. Also, from the tags we could
# get more
# In each alignement, there are lots of information to retrieve.
# One could for instance introspect the tags.
# - cx: subread local context flags
# - ip: vector of length qlen from 0 to 250. This is the IPD (raw frames
# or codec V1)
# - np: number of passes (1 for subread, variable for CCS)
# - pw: vector of length qlen from 0 to 128? This is the PulseWidth (raw
# frames or codec V1)
# - qs: 0-based start of query in the ZMW read (absent in CCS)
# - qe: 0-based end of query in the ZMW read (absent in CCS)
# - zm: position/ID of the ZMW
# - sn: list of ACGT SNRs. A, C, G, T in that order
# - rq: float encoding exepted accuracy
# - dq: DeletionQV
# - dt: deletion Tag
# - iq: insertionQV
# - mq: mergeQV
# - sq: substituionQV
# - st: substituion tag
# - RG: ?
# See http://pacbiofileformats.readthedocs.io/en/3.0/BAM.html
if self._df is None:
logger.info("Scanning input file. Please wait")
self.reset()
N = 0
all_results = []
# This takes 60% of the time...could use cython ?
for i, read in enumerate(self.data):
tags = dict(read.tags)
res = []
# count reads
N += 1
if (N % 10000) == 0:
logger.info("Read %d sequences" % N)
# res[0] = read length
res.append(read.query_length
) # also stored in tags["qe"] - tags["qs"]
res.append(read.reference_length
) # also stored in tags["qe"] - tags["qs"]
# collections.counter is slow, let us do it ourself
if read.query_length and read.query_sequence:
res.append(100. / read.query_length * sum([
read.query_sequence.count(letter)
for letter in "CGcgSs"
]))
else:
res.append(None)
# res[1:4] contains SNR stored in tags['sn'] in the order A, C, G, T
try:
snr = list(tags['sn'])
except:
snr = [None] * 4
res = res + snr
# res[6] = ZMW name, also stored in tags["zm"]
try:
res.append(int(read.qname.split('/')[1]))
except: # simulated data may not have the ZMW info, in which
#case, we store just a unique ID
res.append(i)
# aggregate results
all_results.append(res)
if self._sample and N >= self._sample:
break
self._df = pd.DataFrame(all_results,
columns=[
'read_length', "reference_length",
'GC_content', 'snr_A', 'snr_C',
'snr_G', 'snr_T', 'ZMW'
])
# populate the nb passes from the ZMW
grouped = self._df.groupby("ZMW")
agg = grouped.agg({"read_length": len})
ZMW = self._df.ZMW.unique()
aa = list(
pylab.flatten([[agg.loc[this][0]] * agg.loc[this][0]
for this in ZMW]))
self._df['nb_passes'] = aa
self._df['nb_passes'] -= 1 # nb passes starts at 0
self.reset()
return self._df
def _get_df(self):
# When scanning the BAM, we can extract the length, SNR of ACGT (still
# need to know how to use it). The GC content (note there is no
# ambiguity so no S character). The ZMW. Also, from the tags we could
# get more
# In each alignement, there are lots of information to retrieve.
# One could for instance introspect the tags.
# - cx: subread local context flags
# - ip: vector of length qlen from 0 to 250. This is the IPD (raw frames
# or codec V1)
# - np: number of passes (1 for subread, variable for CCS)
# - pw: vector of length qlen from 0 to 128? This is the PulseWidth (raw
# frames or codec V1)
# - qs: 0-based start of query in the ZMW read (absent in CCS)
# - qe: 0-based end of query in the ZMW read (absent in CCS)
# - zm: position/ID of the ZMW
# - sn: list of ACGT SNRs. A, C, G, T in that order
# - rq: float encoding exepted accuracy
# - dq: DeletionQV
# - dt: deletion Tag
# - iq: insertionQV
# - mq: mergeQV
# - sq: substituionQV
# - st: substituion tag
# - RG: ?
# See http://pacbiofileformats.readthedocs.io/en/3.0/BAM.html
if self._df is None:
logger.info("Scanning input file. Please wait")
self.reset()
N = 0
all_results = []
# This takes 60% of the time...could use cython ?
for i, read in enumerate(self.data):
tags = dict(read.tags)
res = []
# count reads
N += 1
if (N % 10000) == 0:
logger.info("Read %d sequences" %N)
# res[0] = read length
res.append(read.query_length) # also stored in tags["qe"] - tags["qs"]
res.append(read.reference_length) # also stored in tags["qe"] - tags["qs"]
# collections.counter is slow, let us do it ourself
if read.query_length and read.query_sequence:
res.append( 100. / read.query_length * sum(
[read.query_sequence.count(letter) for letter in "CGcgSs"]))
else:
res.append(None)
# res[1:4] contains SNR stored in tags['sn'] in the order A, C, G, T
try:
snr = list(tags['sn'])
except:
snr = [None] * 4
res = res + snr
# res[6] = ZMW name, also stored in tags["zm"]
try:
res.append(int(read.qname.split('/')[1]))
except: # simulated data may not have the ZMW info, in which
#case, we store just a unique ID
res.append(i)
# aggregate results
all_results.append(res)
if self._sample and N >= self._sample:
break
self._df = pd.DataFrame(all_results,
columns=['read_length', "reference_length", 'GC_content',
'snr_A','snr_C','snr_G','snr_T','ZMW'])
# populate the nb passes from the ZMW
grouped = self._df.groupby("ZMW")
agg = grouped.agg({"read_length": len})
ZMW = self._df.ZMW.unique()
aa = list(pylab.flatten([[agg.loc[this][0]] * agg.loc[this][0] for this in ZMW]))
self._df['nb_passes'] = aa
self._df['nb_passes'] -= 1 # nb passes starts at 0
self.reset()
return self._df