"""Genomic arrays"""
import hashlib
import os
from collections import OrderedDict
import h5py
import numpy as np
from pybedtools import Interval
from scipy import sparse
from janggu.utils import _get_output_data_location
from janggu.utils import _iv_to_str
from janggu.utils import _str_to_iv
def _get_iv_length(length, resolution):
"""obtain the chromosome length for a given resolution."""
if resolution is None:
return 1
return int(np.ceil(float(length)/resolution))
def get_collapser(method):
"""Get collapse method."""
if method is None:
return None
elif callable(method):
return method
elif method == 'mean':
return lambda x: x.mean(axis=1)
elif method == 'sum':
return lambda x: x.sum(axis=1)
elif method == 'max':
return lambda x: x.max(axis=1)
raise ValueError('Unknown method: {}'.format(method))
def create_sha256_cache(data, parameters):
"""Cache file determined from files and parameters."""
sha256_hash = hashlib.sha256()
# add file content to hash
for datum in data or []:
if isinstance(datum, str) and os.path.exists(datum):
with open(datum, 'rb') as file_:
for bblock in iter(lambda: file_.read(1024**2), b""):
sha256_hash.update(bblock)
elif isinstance(datum, np.ndarray):
sha256_hash.update(datum.tobytes())
else:
sha256_hash.update(str(datum).encode('utf-8'))
# add parameter settings to hash
sha256_hash.update(str(parameters).encode('utf-8'))
return sha256_hash.hexdigest()
def _get_cachefile(cachestr, tags, fileending):
""" Determine cache file location """
filename = None
if cachestr is not None:
memmap_dir = _get_output_data_location(tags)
if not os.path.exists(memmap_dir):
os.makedirs(memmap_dir)
filename = str(cachestr) + fileending
filename = os.path.join(memmap_dir, filename)
return filename
return None
def _load_data(cachestr, tags, fileending):
""" loading data from scratch or from cache """
filename = _get_cachefile(cachestr, tags, fileending)
if filename is not None and os.path.exists(filename):
return False
return True
class GenomicArray(object): # pylint: disable=too-many-instance-attributes
"""GenomicArray stores multi-dimensional genomic information.
It acts as a dataset for holding genomic data. For instance,
coverage along an entire genome composed of arbitrary length chromosomes
as well as for multiple cell-types and conditions simultaneously.
Inspired by the HTSeq analog, the array can hold the data in different
storage modes, including ndarray, hdf5 or as sparse dataset.
Parameters
----------
stranded : bool
Consider stranded profiles. Default: True.
conditions : list(str) or None
List of cell-type or condition labels associated with the corresponding
array dimensions. Default: None means a one-dimensional array is produced.
typecode : str
Datatype. Default: 'd'.
resolution : int or None
Resolution for storing the genomic array. Only relevant for the use
with Cover Datasets. Default: 1.
padding_value : float
Padding value. Default: 0.
order : int
Order of the alphabet size. Only relevant for Bioseq Datasets. Default: 1.
collapser : None or callable
Method to aggregate values along a given interval.
"""
handle = OrderedDict()
_condition = None
_resolution = None
_order = None
region2index = None
def __init__(self, stranded=True, conditions=None, typecode='d',
resolution=1, padding_value=0.,
order=1, store_whole_genome=True, collapser=None):
self.stranded = stranded
if conditions is None:
conditions = ['sample']
self.condition = conditions
self.order = order
self.padding_value = padding_value
if not isinstance(order, int) or order < 1:
raise Exception('order must be a positive integer')
self.resolution = resolution
self.typecode = typecode
self._full_genome_stored = store_whole_genome
self.collapser = collapser
def _get_indices(self, interval, arraylen):
"""Given the original genomic coordinates,
the array indices of the reference dataset (garray.handle)
and the array indices of the view are returned.
Parameters
----------
interval : Interval
Interval containing (chr, start, end)
arraylen : int
Length of the numpy target array.
Returns
-------
tuple
Tuple of indices corresponding to the slice in the arrays
(ref_start, ref_end, array_start, array_end).
ref_[start,end] indicate the slice in garray.handle
while array_[start,end] indicates the slice in the
target array / view.
"""
chrom = interval.chrom
start = self.get_iv_start(interval.start)
end = self.get_iv_end(interval.end) - self.order + 1
if start >= self.handle[chrom].shape[0]:
return 0, 0, 0, 0
else:
array_start = 0
ref_start = start
if end > self.handle[chrom].shape[0]:
array_end = arraylen - (end - self.handle[chrom].shape[0])
ref_end = self.handle[chrom].shape[0]
else:
array_end = arraylen
ref_end = end
return ref_start, ref_end, array_start, array_end
def __setitem__(self, index, value):
interval = index[0]
condition = index[1]
if isinstance(condition, slice) and value.ndim != 3:
raise ValueError('Expected 3D array with condition slice.')
if isinstance(condition, slice):
condition = slice(None, value.shape[-1], None)
if self.stranded and value.shape[1] != 2:
raise ValueError('If genomic array is in stranded mode, shape[-1] == 2 is expected')
if not self.stranded and value.shape[1] != 1:
value = value.sum(axis=1, keepdims=True)
if isinstance(interval, Interval) and isinstance(condition, (int, slice)):
start = self.get_iv_start(interval.start)
end = self.get_iv_end(interval.end)
length = end - start - self.order + 1
# value should be a 2 dimensional array
# it will be reshaped to a 2D array where the collapse operation is performed
# along the second dimension.
value = self._do_collapse(interval, value)
try:
self._setitem(interval, condition, length, value)
except KeyError:
# we end up here if the peak regions are not a subset of
# the regions of interest. that might be the case if
# peaks from the holdout proportion of the genome are tried
# to be added.
# unfortunately, it is also possible that store_whole_genome=False
# and the peaks and regions of interest are just not synchronized
# in which case nothing (or too few peaks) are added. in the latter
# case an error would help actually, but I am not sure how to
# check if the first or the second is the case here.
pass
else:
raise IndexError("Cannot interpret interval and condition: {}".format((interval, condition)))
def _setitem(self, interval, condition, length, value):
if not self._full_genome_stored:
idx = self.region2index[_iv_to_str(interval.chrom, interval.start,
interval.end)]
# correcting for the overshooting starts and ends is not necessary
# for partially loaded data
self.handle['data'][idx, :length, :, condition] = value
else:
ref_start, ref_end, array_start, \
array_end = self._get_indices(interval, value.shape[0])
self.handle[interval.chrom][ref_start:ref_end, :, condition] = \
value[array_start:array_end]
def _do_collapse(self, interval, value):
if self.collapser is not None:
if self.resolution is None and value.shape[0] == 1 or \
self.resolution is not None and \
value.shape[0] == interval.length//self.resolution:
# collapsing becomes obsolete, because the data has already
# the expected shape (after collapsing)
pass
else:
if self.resolution is None:
# collapse along the entire interval
value = value.reshape((1,) + value.shape)
else:
# if the array shape[0] is a multipe of resolution,
# it can simply be reshaped. otherwise,
# it needs to be resized before.
if value.shape[0] % self.resolution > 0:
value = np.resize(value, (
int(np.ceil(value.shape[0]/float(self.resolution))*self.resolution),) +
value.shape[1:])
# collapse in bins of size resolution
value = value.reshape((value.shape[0]//min(self.resolution,
value.shape[0]),
min(self.resolution, value.shape[0]),) + \
value.shape[1:])
value = self.collapser(value)
return value
def __getitem__(self, index):
# for now lets ignore everything except for chrom, start and end.
if isinstance(index, Interval):
interval = index
chrom = interval.chrom
start = self.get_iv_start(interval.start)
end = self.get_iv_end(interval.end)
# original length
length = end-start - self.order + 1
if not self._full_genome_stored:
idx = self.region2index[_iv_to_str(chrom, interval.start, interval.end)]
# correcting for the overshooting starts and ends is not necessary
# for partially loaded data
return self._reshape(self.handle['data'][idx],
(length, 2 if self.stranded else 1,
len(self.condition)))
if chrom not in self.handle:
return np.ones((length, 2 if self.stranded else 1,
len(self.condition)),
dtype=self.typecode) * self.padding_value
if start >= 0 and end <= self.handle[chrom].shape[0]:
end = end - self.order + 1
# this is a short-cut, which does not require zero-padding
return self._reshape(self.handle[chrom][start:end],
(end-start, 2 if self.stranded else 1,
len(self.condition)))
# below is some functionality for zero-padding, in case the region
# reaches out of the chromosome size
if self.padding_value == 0.0:
data = np.zeros((length, 2 if self.stranded else 1,
len(self.condition)),
dtype=self.typecode)
else:
data = np.ones((length, 2 if self.stranded else 1,
len(self.condition)),
dtype=self.typecode) * self.padding_value
ref_start, ref_end, array_start, array_end = self._get_indices(interval, data.shape[0])
data[array_start:array_end, :, :] = self._reshape(self.handle[chrom][ref_start:ref_end],
(ref_end - ref_start,
2 if self.stranded else 1,
len(self.condition)))
return data
raise IndexError("Cannot interpret interval: {}".format(index))
@property
def condition(self):
"""condition"""
return self._condition
@condition.setter
def condition(self, conditions):
self._condition = conditions
@property
def resolution(self):
"""resolution"""
return self._resolution
@resolution.setter
def resolution(self, value):
if value is not None and value <= 0:
raise ValueError('resolution>0 required')
self._resolution = value
def _reshape(self, data, shape):
# shape not necessary here,
# data should just fall through
return data
def interval_length(self, chrom):
"""Method returns the interval lengths."""
# extract the length by the interval length
# or by the array shape
locus = _str_to_iv(chrom)
if len(locus) > 1:
return locus[2] - locus[1]
return self.resolution
def scale_by_region_length(self):
""" This method scales the regions by the region length ."""
for chrom in self.handle:
if self._full_genome_stored:
self.handle[chrom][:] /= self.interval_length(chrom)
else:
for rstr, idx in self.region2index.items():
self.handle[chrom][idx] /= self.interval_length(rstr)
def weighted_mean(self):
""" Base pair resolution mean weighted by interval length
"""
# summing the signal
sums = [self.sum(chrom) for chrom in self.handle]
sums = np.asarray(sums).sum(axis=0)
# weights are determined by interval and chromosome length
weights = [np.prod(self.handle[chrom].shape[:-1]) \
for chrom in self.handle]
weights = np.asarray(weights).sum()
return sums / weights
def shift(self, means):
"""Centering the signal by the weighted mean"""
#means = self.weighted_mean()
for chrom in self.handle:
# adjust base pair resoltion mean to interval length
self.handle[chrom][:] -= means
def rescale(self, scale):
""" Method to rescale the signal """
for chrom in self.handle:
self.handle[chrom][:] /= scale
def sum(self, chrom=None):
"""Sum signal across chromosomes."""
if chrom is not None:
return self.handle[chrom][:]\
.sum(axis=tuple(range(self.handle[chrom].ndim - 1)))
return np.asarray([self.handle[chrom][:]\
.sum(axis=tuple(range(self.handle[chrom].ndim - 1)))
for chrom in self.handle])
def weighted_sd(self):
""" Interval scaled standard deviation """
# summing the squared signal signal
sums = [np.square(self.handle[chrom][:, :, :]).sum(
axis=tuple(range(self.handle[chrom].ndim - 1))) \
for chrom in self.handle]
sums = np.asarray(sums).sum(axis=0)
# weights are determined by interval and chromosome length
weights = [np.prod(self.handle[chrom].shape[:-1]) \
for chrom in self.handle]
weights = np.asarray(weights).sum()
return np.sqrt(sums / (weights - 1.))
@property
def order(self):
"""order"""
return self._order
@order.setter
def order(self, order):
if order <= 0:
raise ValueError('order>0 required')
self._order = order
def get_iv_end(self, end):
"""obtain the chromosome length for a given resolution."""
return _get_iv_length(end, self.resolution)
def get_iv_start(self, start):
"""obtain the chromosome length for a given resolution."""
if self.resolution is None:
return 0
return start // self.resolution
def init_with_padding_value(padding_value, shape, dtype):
""" create array with given padding value. """
if padding_value == 0.0:
return np.zeros(shape, dtype)
else:
return np.ones(shape, dtype) * padding_value
class HDF5GenomicArray(GenomicArray):
"""HDF5GenomicArray stores multi-dimensional genomic information.
Implements GenomicArray.
Parameters
----------
gsize : GenomicIndexer or callable
GenomicIndexer containing the genome sizes or a callable that
returns a GenomicIndexer to enable lazy loading.
stranded : bool
Consider stranded profiles. Default: True.
conditions : list(str) or None
List of cell-type or condition labels associated with the corresponding
array dimensions. Default: None means a one-dimensional array is produced.
typecode : str
Datatype. Default: 'd'.
datatags : list(str) or None
Tags describing the dataset. This is used to store the cache file.
resolution : int
Resolution for storing the genomic array. Only relevant for the use
with Cover Datasets. Default: 1.
order : int
Order of the alphabet size. Only relevant for Bioseq Datasets. Default: 1.
store_whole_genome : boolean
Whether to store the entire genome or only the regions of interest.
Default: True
padding_value : float
Padding value. Default: 0.
cache : str or None
Hash string of the data and parameters to cache the dataset. If None,
caching is deactivated. Default: None.
overwrite : boolean
Whether to overwrite the cache. Default: False
loader : callable or None
Function to be called for loading the genomic array.
collapser : None or callable
Method to aggregate values along a given interval.
verbose : boolean
Verbosity. Default: False
"""
def __init__(self, gsize, # pylint: disable=too-many-locals
stranded=True,
conditions=None,
typecode='d',
datatags=None,
resolution=1,
order=1,
padding_value=0.,
store_whole_genome=True,
cache=None,
overwrite=False, loader=None,
normalizer=None,
collapser=None,
verbose=False):
super(HDF5GenomicArray, self).__init__(stranded, conditions, typecode,
resolution,
order=order,
padding_value=padding_value,
store_whole_genome=store_whole_genome,
collapser=collapser)
if cache is None:
raise ValueError('cache=True required for HDF format')
gsize_ = None
if not store_whole_genome:
gsize_ = gsize() if callable(gsize) else gsize
self.region2index = {_iv_to_str(region.chrom,
region.start,
region.end): i \
for i, region in enumerate(gsize_)}
cachefile = _get_cachefile(cache, datatags, '.h5')
load_from_file = _load_data(cache, datatags, '.h5')
if load_from_file:
if gsize_ is None:
gsize_ = gsize() if callable(gsize) else gsize
h5file = h5py.File(cachefile, 'w')
if store_whole_genome:
for region in gsize_:
shape = (_get_iv_length(region.length - self.order + 1, self.resolution),
2 if stranded else 1, len(self.condition))
h5file.create_dataset(str(region.chrom), shape,
dtype=self.typecode,
data=init_with_padding_value(padding_value,
shape,
self.typecode))
self.handle = h5file
else:
shape = (len(gsize_),
_get_iv_length(gsize_.binsize + 2*gsize_.flank - self.order + 1,
self.resolution),
2 if stranded else 1, len(self.condition))
h5file.create_dataset('data', shape,
dtype=self.typecode,
data=init_with_padding_value(padding_value,
shape,
self.typecode))
self.handle = h5file
# invoke the loader
if loader:
loader(self)
for norm in normalizer or []:
get_normalizer(norm)(self)
h5file.close()
if verbose: print('reload {}'.format(cachefile))
h5file = h5py.File(cachefile, 'a', driver='stdio')
self.handle = h5file
class NPGenomicArray(GenomicArray):
"""NPGenomicArray stores multi-dimensional genomic information.
Implements GenomicArray.
Parameters
----------
gsize : GenomicIndexer or callable
GenomicIndexer containing the genome sizes or a callable that
returns a GenomicIndexer to enable lazy loading.
stranded : bool
Consider stranded profiles. Default: True.
conditions : list(str) or None
List of cell-type or condition labels associated with the corresponding
array dimensions. Default: None means a one-dimensional array is produced.
typecode : str
Datatype. Default: 'd'.
datatags : list(str) or None
Tags describing the dataset. This is used to store the cache file.
resolution : int
Resolution for storing the genomic array. Only relevant for the use
with Cover Datasets. Default: 1.
order : int
Order of the alphabet size. Only relevant for Bioseq Datasets. Default: 1.
store_whole_genome : boolean
Whether to store the entire genome or only the regions of interest.
Default: True
padding_value : float
Padding value. Default: 0.
cache : str or None
Hash string of the data and parameters to cache the dataset. If None,
caching is deactivated. Default: None.
overwrite : boolean
Whether to overwrite the cache. Default: False
loader : callable or None
Function to be called for loading the genomic array.
normalizer : callable or None
Normalization to be applied. This argumenet can be None,
if no normalization is applied, or a callable that takes
a garray and returns a normalized garray.
Default: None.
collapser : None or callable
Method to aggregate values along a given interval.
verbose : boolean
Verbosity. Default: False
"""
def __init__(self, gsize, # pylint: disable=too-many-locals
stranded=True,
conditions=None,
typecode='d',
datatags=None,
resolution=1,
order=1,
padding_value=0.0,
store_whole_genome=True,
cache=None,
overwrite=False, loader=None,
normalizer=None, collapser=None,
verbose=False):
super(NPGenomicArray, self).__init__(stranded, conditions, typecode,
resolution,
order=order,
padding_value=padding_value,
store_whole_genome=store_whole_genome,
collapser=collapser)
gsize_ = None
if not store_whole_genome:
gsize_ = gsize() if callable(gsize) else gsize
self.region2index = {_iv_to_str(region.chrom,
region.start,
region.end): i \
for i, region in enumerate(gsize_)}
cachefile = _get_cachefile(cache, datatags, '.npz')
load_from_file = _load_data(cache, datatags, '.npz')
if load_from_file:
if gsize_ is None:
gsize_ = gsize() if callable(gsize) else gsize
if store_whole_genome:
data = {str(region.chrom): init_with_padding_value(
padding_value,
shape=(_get_iv_length(region.length - self.order + 1,
self.resolution),
2 if stranded else 1,
len(self.condition)),
dtype=self.typecode) for region in gsize_}
names = [str(region.chrom) for region in gsize_]
self.handle = data
else:
data = {'data': init_with_padding_value(
padding_value,
shape=(len(gsize_),
_get_iv_length(gsize_.binsize + 2*gsize_.flank - self.order + 1,
self.resolution) if self.resolution is not None else 1,
2 if stranded else 1,
len(self.condition)),
dtype=self.typecode)}
names = ['data']
self.handle = data
# invoke the loader
if loader:
loader(self)
if cachefile is not None:
np.savez(cachefile, **data)
if cachefile is not None:
if verbose: print('reload {}'.format(cachefile))
data = np.load(cachefile)
names = [x for x in data]
# here we get either the freshly loaded data or the reloaded
# data from np.load.
self.handle = {key: data[key] for key in names}
for norm in normalizer or []:
get_normalizer(norm)(self)
class SparseGenomicArray(GenomicArray):
"""SparseGenomicArray stores multi-dimensional genomic information.
Implements GenomicArray.
Parameters
----------
gsize : GenomicIndexer or callable
GenomicIndexer containing the genome sizes or a callable that
returns a GenomicIndexer to enable lazy loading.
stranded : bool
Consider stranded profiles. Default: True.
conditions : list(str) or None
List of cell-type or condition labels associated with the corresponding
array dimensions. Default: None means a one-dimensional array is produced.
typecode : str
Datatype. Default: 'd'.
datatags : list(str) or None
Tags describing the dataset. This is used to store the cache file.
resolution : int
Resolution for storing the genomic array. Only relevant for the use
with Cover Datasets. Default: 1.
order : int
Order of the alphabet size. Only relevant for Bioseq Datasets. Default: 1.
store_whole_genome : boolean
Whether to store the entire genome or only the regions of interest.
Default: True
padding_value : float
Padding value. Default: 0.
cache : str or None
Hash string of the data and parameters to cache the dataset. If None,
caching is deactivated. Default: None.
overwrite : boolean
Whether to overwrite the cache. Default: False
loader : callable or None
Function to be called for loading the genomic array.
normalizer : callable or None
Normalization to be applied. This argumenet can be None,
if no normalization is applied, or a callable that takes
a garray and returns a normalized garray.
Default: None.
collapser : None or callable
Method to aggregate values along a given interval.
verbose : boolean
Verbosity. Default: False
"""
def __init__(self, gsize, # pylint: disable=too-many-locals
stranded=True,
conditions=None,
typecode='d',
datatags=None,
resolution=1,
order=1,
store_whole_genome=True,
cache=None,
padding_value=0.0,
overwrite=False,
loader=None,
collapser=None,
verbose=False):
super(SparseGenomicArray, self).__init__(stranded, conditions,
typecode,
resolution,
order=order,
padding_value=padding_value,
store_whole_genome=store_whole_genome,
collapser=collapser)
cachefile = _get_cachefile(cache, datatags, '.npz')
load_from_file = _load_data(cache, datatags, '.npz')
gsize_ = None
if not store_whole_genome:
gsize_ = gsize() if callable(gsize) else gsize
self.region2index = {_iv_to_str(region.chrom,
region.start,
region.end): i \
for i, region in enumerate(gsize_)}
if load_from_file:
if gsize_ is None:
gsize_ = gsize() if callable(gsize) else gsize
if store_whole_genome:
data = {str(region.chrom): sparse.dok_matrix(
(_get_iv_length(region.length - self.order + 1,
resolution),
(2 if stranded else 1) * len(self.condition)),
dtype=self.typecode)
for region in gsize_}
else:
data = {'data': sparse.dok_matrix(
(len(gsize_),
(_get_iv_length(gsize_.binsize + 2*gsize_.flank - self.order + 1,
self.resolution) if self.resolution is not None else 1) *
(2 if stranded else 1) * len(self.condition)),
dtype=self.typecode)}
self.handle = data
# invoke the loader
if loader:
loader(self)
data = self.handle
data = {chrom: data[chrom].tocoo() for chrom in data}
storage = {chrom: np.column_stack([data[chrom].data,
data[chrom].row,
data[chrom].col]) \
for chrom in data}
for region in gsize_:
if store_whole_genome:
storage[region.chrom + '__length__'] = region.length
names = [name for name in storage]
if cachefile is not None:
np.savez(cachefile, **storage)
if cachefile is not None:
if verbose: print('reload {}'.format(cachefile))
storage = np.load(cachefile)
names = [name for name in storage if '__length__' not in name]
if store_whole_genome:
self.handle = {name: sparse.coo_matrix(
(storage[name][:, 0],
(storage[name][:, 1].astype('int'),
storage[name][:, 2].astype('int'))),
shape=(_get_iv_length(storage[str(name)+'__length__'], resolution),
(2 if stranded else 1) * len(self.condition))).tocsr()
for name in names}
else:
# gsize_ is always available for store_whole_genome=False
self.handle = {name: sparse.coo_matrix(
(storage[name][:, 0],
(storage[name][:, 1].astype('int'),
storage[name][:, 2].astype('int'))),
shape=(len(gsize_),
(_get_iv_length(gsize_.binsize + 2*gsize_.flank, resolution)
if self.resolution is not None else 1) *
(2 if stranded else 1) * len(self.condition))).tocsr()
for name in names}
def _reshape(self, data, shape):
# what to do with zero padding
data = data.toarray()
if self._full_genome_stored:
return data.reshape(data.shape[0], data.shape[1]//(shape[-1]), shape[-1])
else:
return data.reshape(data.shape[1]//(shape[-2]*shape[-1]), shape[-2], shape[-1])
def _setitem(self, interval, condition, length, value):
if not self._full_genome_stored:
regidx = self.region2index[_iv_to_str(interval.chrom, interval.start, interval.end)]
nconditions = len(self.condition)
ncondstrand = len(self.condition) * value.shape[-1]
#end = end - self.order + 1
idxs = np.where(value > 0)
for idx in zip(*idxs):
basepos = idx[0] * ncondstrand
strand = idx[1] * nconditions
cond = condition if isinstance(condition, int) else idx[2]
self.handle['data'][regidx,
basepos + strand + cond] = value[idx]
else:
ref_start, ref_end, array_start, _ = self._get_indices(interval, value.shape[0])
idxs = np.where(value > 0)
iarray = np.arange(ref_start, ref_end)
for idx in zip(*idxs):
cond = condition if isinstance(condition, int) else idx[2]
self.handle[interval.chrom][iarray[idx[0]],
idx[1] * len(self.condition)
+ cond] = value[idx[0] + array_start][idx[1:]]
class PercentileTrimming(object):
"""Percentile trimming normalization.
This class performs percentile trimming of a GenomicArray to aleviate
the effect of outliers.
All values that exceed the value associated with the given percentile
are set to be equal to the percentile.
Parameters
----------
percentile : float
Percentile at which to perform chromosome-level trimming.
"""
def __init__(self, percentile):
self.percentile = percentile
def __call__(self, garray):
quants = np.percentile(np.concatenate(np.asarray([garray.handle[chrom] for \
chrom in garray.handle]), axis=0),
self.percentile, axis=(0, 1))
for icond, quant in enumerate(quants):
for chrom in garray.handle:
arr = garray.handle[chrom][:, :, icond]
arr[arr > quant] = quant
garray.handle[chrom][:, :, icond] = arr
return garray
def __str__(self): # pragma: no cover
return 'PercentileTrimming({})'.format(self.percentile)
def __repr__(self): # pragma: no cover
return str(self)
class RegionLengthNormalization(object):
""" Normalization for variable-region length.
This class performs region length normalization of a GenomicArray.
This is relevant when genomic features are of variable size, e.g.
enhancer regions of different width or when using variable length genes.
Parameters
----------
regionmask : str or GenomicIndexer, None
A bed file or a genomic indexer that contains the masking region
that is considered for the signal. For instance, when normalizing
gene expression to TPM, the mask contains exons. Otherwise, the
TPM would normalize for the full length gene annotation.
If None, no mask is included.
"""
def __init__(self, regionmask=None):
self.regionmask = regionmask
def __call__(self, garray):
# length scaling
garray.scale_by_region_length()
return garray
def __str__(self): # pragma: no cover
return 'RegionLengthNormalization({})'.format(self.regionmask)
def __repr__(self): # pragma: no cover
return str(self)
class ZScore(object):
"""ZScore normalization.
This class performs ZScore normalization of a GenomicArray.
It automatically adjusts for variable interval lenths.
Parameters
----------
means : float or None
Provided means will be applied for zero-centering.
If None, the means will be determined
from the GenomicArray and then applied.
Default: None.
stds : float or None
Provided standard deviations will be applied for scaling.
If None, the stds will be determined
from the GenomicArray and then applied.
Default: None.
"""
def __init__(self, mean=None, std=None):
self.mean = mean
self.std = std
def __call__(self, garray):
# determine the mean signal per condition
if self.mean is None:
self.mean = garray.weighted_mean()
# centering to zero-mean
garray.shift(self.mean)
# determines standard deviation per contition
if self.std is None:
self.std = garray.weighted_sd()
# rescale by standard deviation
garray.rescale(self.std)
return garray
def __str__(self): # pragma: no cover
return 'ZScore({},{})'.format(self.mean, self.std)
def __repr__(self): # pragma: no cover
return str(self)
class LogTransform(object):
"""Log transformation of intput signal.
This class performs log-transformation
of a GenomicArray using log(x + 1.) to avoid NAN's from zeros.
"""
def __init__(self):
pass
def __call__(self, garray):
for chrom in garray.handle:
garray.handle[chrom][:] = np.log(garray.handle[chrom][:] + 1.)
return garray
def __str__(self): # pragma: no cover
return 'Log'
def __repr__(self): # pragma: no cover
return str(self)
class ZScoreLog(object):
"""ZScore normalization after log transformation.
This class performs ZScore normalization after log-transformation
of a GenomicArray using log(x + 1.) to avoid NAN's from zeros.
It automatically adjusts for variable interval lenths.
Parameters
----------
means : float or None
Provided means will be applied for zero-centering.
If None, the means will be determined
from the GenomicArray and then applied.
Default: None.
stds : float or None
Provided standard deviations will be applied for scaling.
If None, the stds will be determined
from the GenomicArray and then applied.
Default: None.
"""
def __init__(self, mean=None, std=None):
self.logtr = LogTransform()
self.zscore = ZScore(mean, std)
def __call__(self, garray):
return self.zscore(self.logtr(garray))
def __str__(self): # pragma: no cover
return str(self.zscore) + str(self.logtr)
def __repr__(self): # pragma: no cover
return str(self)
def normalize_garray_tpm(garray):
"""This function performs TPM normalization
for a given GenomicArray.
"""
# rescale by region lengths in bp
garray = RegionLengthNormalization()(garray)
# recale to kb
garray.rescale(1e-3)
# compute scaling factor
scale = garray.sum() # per chromsome sum
scale = scale.sum(axis=0) # sum across chroms
scale /= 1e6 # rescale by million
# divide by scaling factor
garray.rescale(scale)
return garray
def get_normalizer(normalizer):
""" maps built-in normalizers by name and
returns the respective function """
if callable(normalizer):
return normalizer
elif normalizer == 'zscore':
return ZScore()
elif normalizer == 'zscorelog':
return ZScoreLog()
elif normalizer == 'binsizenorm':
return RegionLengthNormalization()
elif normalizer == 'perctrim':
return PercentileTrimming(99)
elif normalizer == 'tpm':
return normalize_garray_tpm
raise ValueError('unknown normalizer: {}'.format(normalizer))
def create_genomic_array(chroms, stranded=True, conditions=None, typecode='float32',
storage='hdf5', resolution=1,
order=1,
padding_value=0.0,
store_whole_genome=True,
datatags=None, cache=None, overwrite=False,
loader=None,
normalizer=None, collapser=None,
verbose=False):
"""Factory function for creating a GenomicArray.
This function creates a genomic array for a given storage mode.
Parameters
----------
chroms : dict
Dictionary with chromosome names as keys and chromosome lengths
as values.
stranded : bool
Consider stranded profiles. Default: True.
conditions : list(str) or None
List of cell-type or condition labels associated with the corresponding
array dimensions. Default: None means a one-dimensional array is produced.
typecode : str
Datatype. Default: 'float32'.
storage : str
Storage type can be 'ndarray', 'hdf5' or 'sparse'.
Numpy loads the entire dataset into the memory. HDF5 keeps
the data on disk and loads the mini-batches from disk.
Sparse maintains sparse matrix representation of the dataset
in the memory.
Usage of numpy will require high memory consumption, but allows fast
slicing operations on the dataset. HDF5 requires low memory consumption,
but fetching the data from disk might be time consuming.
sparse will be a good compromise if the data is indeed sparse. In this
case, memory consumption will be low while slicing will still be fast.
datatags : list(str) or None
Tags describing the dataset. This is used to store the cache file.
resolution : int
Resolution for storing the genomic array. Only relevant for the use
with Cover Datasets. Default: 1.
order : int
Order of the alphabet size. Only relevant for Bioseq Datasets. Default: 1.
padding_value : float
Padding value. Default: 0.0
store_whole_genome : boolean
Whether to store the entire genome or only the regions of interest.
Default: True
cache : str or None
Hash string of the data and parameters to cache the dataset. If None,
caching is deactivated. Default: None.
overwrite : boolean
Whether to overwrite the cache. Default: False
loader : callable or None
Function to be called for loading the genomic array.
normalizer : callable or None
Normalization to be applied. This argumenet can be None,
if no normalization is applied, or a callable that takes
a garray and returns a normalized garray.
Default: None.
collapser : str, callable or None
Collapse method defines how the signal is aggregated for resolution>1 or resolution=None.
For example, by summing the signal over a given interval.
verbose : boolean
Verbosity. Default: False
"""
# check if collapser available
if (resolution is None or resolution > 1) and collapser is None:
raise ValueError('Requiring collapser due to resolution=None or resolution>1')
# force store_whole_genome=False if resolution=None
if resolution is None and store_whole_genome:
print('store_whole_genome=True ignored, because it is not compatible'
'with resolution=None. store_whole_genome=False is used instead.')
store_whole_genome = False
if storage == 'hdf5':
return HDF5GenomicArray(chroms, stranded=stranded,
conditions=conditions,
typecode=typecode,
datatags=datatags,
resolution=resolution,
order=order,
store_whole_genome=store_whole_genome,
cache=cache,
padding_value=padding_value,
overwrite=overwrite,
loader=loader,
normalizer=normalizer,
collapser=get_collapser(collapser),
verbose=verbose)
elif storage == 'ndarray':
return NPGenomicArray(chroms, stranded=stranded,
conditions=conditions,
typecode=typecode,
datatags=datatags,
resolution=resolution,
order=order,
store_whole_genome=store_whole_genome,
cache=cache,
padding_value=padding_value,
overwrite=overwrite,
loader=loader,
normalizer=normalizer,
collapser=get_collapser(collapser),
verbose=verbose)
elif storage == 'sparse':
return SparseGenomicArray(chroms, stranded=stranded,
conditions=conditions,
typecode=typecode,
datatags=datatags,
resolution=resolution,
order=order,
store_whole_genome=store_whole_genome,
cache=cache,
padding_value=padding_value,
overwrite=overwrite,
loader=loader,
collapser=get_collapser(collapser),
verbose=verbose)
raise Exception("Storage type must be 'hdf5', 'ndarray' or 'sparse'")
Datasets
Models
