def poisson(self, op_id, log_overflow=1000.):
if bool(self):
from scipy.stats import poisson
const, measures, labeled = self._measure(op_id)
out = list()
with warnings.catch_warnings():
warnings.simplefilter('error')
for row in self.data_.itertuples(index=False):
# Calculate expectation
item = const * measures[labeled[0]][_map(row, labeled[0])]
for col in labeled[1:]:
item *= measures[col][_map(row, col)] / self.total_sum
# Log Fish
p, pexp = self.weight(row), item
if p < pexp:
# Xo < pois(Xexp)
try:
val = poisson.logcdf(p, pexp) # P(pois(Xexp)<Xo), Xo = pt[v], Xexp = pexp[v]
except RuntimeWarning:
val = log_overflow
else:
try:
val = -poisson.logsf(p, pexp) # P(pois(Xexp)>=Xo), Xo = pt[v], Xexp = pexp[v]
except RuntimeWarning:
val = -log_overflow
out.append(self.dump_row(row) + (val,))
return self._weighted_output(out)
else:
return self.__class__()
def logprob_not_dark_hot(cts, exp, ref, quantile=64):
"""Log-probabilities of pixels being not dark and not hot.
The cumulative Poisson distribution function (for dark pixels) and its
inverse, the survival function (for hot pixels) are calculated and
normalised to account for the number of dead pixels and the additional
scatter due to mismatch of the reference shape. The pixel ``quantile``
is used for shape normalisation.
Arguments
----------
cts : ndarray of ints
counts per pixel
exp : ndarray of floats
exposure per pixel
ref : ndarray of floats
reference intensity per pixel
Keyword arguments
-----------------
quantile
reference pixel for shape normalisation
Returns
-------
(d, h) : log-probabilities of pixels being not (dark, hot)
"""
if hasattr(cts, 'mask'):
invalid = cts.mask
valid = np.logical_not(invalid)
else:
invalid = []
valid = Ellipsis
ref_cts = ref / ref[valid].mean() * cts[valid].mean()
logcdf = poisson.logcdf(cts, ref_cts)
logsf = poisson.logsf(cts, ref_cts)
logcdf[invalid] = 0
logsf[invalid] = 0
logcdf_s = np.sort(logcdf[valid])
logsf_s = np.sort(logsf[valid])
logcdf, logsf = \
[f(cts, ref_cts) for f in (poisson.logcdf, poisson.logsf)]
for f in logcdf, logsf:
f[invalid] = 0
logcdf_s, logsf_s = \
[np.sort(f[valid]) for f in (logcdf, logsf)]
norm_cdf, norm_sf = \
[np.log((cts.count() - quantile) / quantile) / logf_s[quantile]
for logf_s in (logcdf_s, logsf_s)]
offset = np.log(1 - ((128**2 - cts.count()) / 128**2))
return (offset - logcdf * norm_cdf,
offset - logsf * norm_sf)
def computePvalue(tileCoverage, args):
"""
This function is called by the writeBedGraph workers for every
tile in the genome that is considered
"""
# if tileCoverage == (0,0):
# return np.nan
treatmentWindowTags = tileCoverage[0]
controlWindowTags = tileCoverage[1]
if controlWindowTags == 0:
return np.nan
treatmentExtraSignalTags = treatmentWindowTags - args['treatmentMean']
controlLambda = args['controlMean'] + (treatmentExtraSignalTags * args['controlSignalRatio'])
log10pvalue = -1* poisson.logcdf(controlWindowTags, controlLambda) / np.log(10)
# log10pvalue = -1* poisson.logsf(controlWindowTags, controlLambda) / np.log(10)
return log10pvalue
def log_polyg(lx_var, alpha_var, d_var):
"""
compute gumbel polylog
:param lx_var:
:param alpha_var:
:param d_var:
:return:
>>> lsum(np.transpose(ipsi_gumbel(np.array([
... [0.42873569, 0.18285458, 0.9514195],
... [0.25148149, 0.05617784, 0.3378213],
... [0.79410993, 0.76175687, 0.0709562],
... [0.02694249, 0.45788802, 0.6299574],
... [0.39522060, 0.02189511, 0.6332237],
... [0.66878367, 0.38075101, 0.5185625],
... [0.90365653, 0.19654621, 0.6809525],
... [0.28607729, 0.82713755, 0.7686878],
... [0.22437343, 0.16907646, 0.5740400],
... [0.66752741, 0.69487362, 0.3329266]
... ]),
... 1.2,
... is_log=True
... )))
array([1.00636964, 1.81365937, 1.27973155, 1.76000074, 1.84085744,
0.64075371, 0.77684883, 0.49862315, 1.41268535, 0.56279559])
>>> log_polyg(
... lsum(np.transpose(ipsi_gumbel(np.array([
... [0.42873569, 0.18285458, 0.9514195],
... [0.25148149, 0.05617784, 0.3378213],
... [0.79410993, 0.76175687, 0.0709562],
... [0.02694249, 0.45788802, 0.6299574],
... [0.39522060, 0.02189511, 0.6332237],
... [0.66878367, 0.38075101, 0.5185625],
... [0.90365653, 0.19654621, 0.6809525],
... [0.28607729, 0.82713755, 0.7686878],
... [0.22437343, 0.16907646, 0.5740400],
... [0.66752741, 0.69487362, 0.3329266]
... ]),
... 1.2,
... is_log=True
... ))) * 1/1.2, 1/1.2, 3
... )
array([2.24028738, 4.12345214, 2.86724735, 3.99567951, 4.1883271 ,
1.42510898, 1.72500906, 1.11696417, 3.17657604, 1.25542124])
>>> log_polyg(
... lsum(np.transpose(ipsi_gumbel(np.array([
... [0.42873569, 0.18285458, 0.9514195],
... [0.25148149, 0.05617784, 0.3378213],
... [0.79410993, 0.76175687, 0.0709562],
... [0.02694249, 0.45788802, 0.6299574],
... [0.39522060, 0.02189511, 0.6332237],
... [0.66878367, 0.38075101, 0.5185625],
... [0.90365653, 0.19654621, 0.6809525],
... [0.28607729, 0.82713755, 0.7686878],
... [0.22437343, 0.16907646, 0.5740400],
... [0.66752741, 0.69487362, 0.3329266]
... ]),
... 3.2,
... is_log=True
... ))) * 1/3.2, 1/3.2, 3
... )
array([ 0.35110025, 1.31419104, 1.07707314, 1.68854151, 1.80435943,
-0.43406987, 0.23166651, -0.18316099, 0.62329368, -0.35013782])
"""
k = np.linspace(start=1.0, stop=d_var, num=int(d_var))
x = np.exp(lx_var)
lppois = np.zeros(shape=[int(d_var), int(lx_var.shape[0])])
for i in range(int(d_var)):
lppois[i] = poisson.logcdf(d_var - k[i], x)
llx = np.dot(k.reshape(-1, 1), lx_var.reshape(1, -1))
labspoch = np.zeros(shape=int(d_var))
for i in range(int(d_var)):
with np.errstate(divide='ignore'):
labspoch[i] = np.sum(np.log(
abs(alpha_var * float(i + 1) - (k - 1.0))),
axis=0)
lfac = np.log(factorial(k))
lxabs = llx + lppois + np.tile(labspoch - lfac, int(
lx_var.shape[0])).reshape(
(int(d_var), int(lx_var.shape[0])), order='F') + np.repeat(
x, int(d_var)).reshape(
(int(d_var), int(lx_var.shape[0])), order='F')
return lssum(
x_values=lxabs,
x_sign=signff(alpha_var, k, d_var),
)
