def from_upper_triu(vector_repr, matrix_len, num_diags):
z = np.zeros((matrix_len, matrix_len))
triu_tup = np.triu_indices(matrix_len, num_diags)
z[triu_tup] = vector_repr
for i in range(-num_diags + 1, num_diags):
set_diag(z, np.nan, i)
return z + z.T
def plot_juicebox_from_predicted_array(mat,
binsize,
interval,
out_dir,
diagonal_offset,
chr_dict,
use_control=False,
**kwargs):
predicted_data = mat_to_pandas_df(mat=mat,
binsize=binsize,
interval=interval)
print(predicted_data.isna().sum())
print(predicted_data)
mp = MatrixPlotter()
mp.set_data(predicted_data)
if not use_control:
mp.set_control(predicted_data)
else:
if 'genome_cool_file' not in kwargs:
print("please add path to control cool file")
raise Exception
# process hic data
print("open and process control cool file")
genome_hic_cool = kwargs['ghc']
mseq_str = '%s:%d-%d' % (interval.chr, interval.start, interval.end)
seq_hic_raw = genome_hic_cool.matrix(balance=True).fetch(mseq_str)
print("seq_hic from cool file shape:", seq_hic_raw.shape,
"predicted matrix shape:", mat.shape)
assert seq_hic_raw.shape == mat.shape
clipval = np.nanmedian(np.diag(seq_hic_raw, diagonal_offset))
for i in range(-diagonal_offset + 1, diagonal_offset):
set_diag(seq_hic_raw, clipval, i)
seq_hic_raw = np.clip(seq_hic_raw, 0, clipval)
# adaptively coarsegrain based on raw counts
seq_hic_smoothed = adaptive_coarsegrain(
seq_hic_raw,
genome_hic_cool.matrix(balance=False).fetch(mseq_str),
cutoff=2,
max_levels=8)
control_data = mat_to_pandas_df(mat=seq_hic_smoothed,
binsize=binsize,
interval=interval)
print(len(control_data))
# choose only contacts <= seqlen
control_data_merge = pd.merge(predicted_data,
control_data,
on=["chr", "contact_st", "contact_en"])
control_data = control_data_merge[[
'chr', 'contact_st', 'contact_en', 'contact_count_y'
]]
control_data.rename(columns={'contact_count_y': 'contact_count'},
inplace=True)
print(control_data)
print(len(control_data))
mp.set_control(control_data)
# mp.set_apply_log(self.apply_log)
MatPlot2HiC(
mp, interval.chr + "_" + str(interval.start) + '_' + str(interval.end),
out_dir, chr_dict)
def from_oe_to_contacts(seq_hic_obsexp,
genome_hic_expected_file,
interval,
seq_len_pool,
no_log=False):
if no_log == False:
seq_hic_obsexp = np.exp(seq_hic_obsexp)
genome_hic_expected = pd.read_csv(genome_hic_expected_file, sep='\t')
exp_chr = genome_hic_expected.iloc[genome_hic_expected['chrom'].values ==
interval.chr][0:seq_len_pool]
if len(exp_chr) == 0:
raise ValueError('no expected values found for chr:' + interval.chr)
exp_map = np.zeros((seq_len_pool, seq_len_pool))
print(exp_map.shape)
for i in range(seq_len_pool):
set_diag(exp_map, exp_chr['balanced.avg'].values[i], i)
set_diag(exp_map, exp_chr['balanced.avg'].values[i], -i)
seq_hic_smoothed = exp_map * seq_hic_obsexp
return seq_hic_smoothed
def _insul_diamond_dense(mat, window=10, ignore_diags=2, norm_by_median=True):
"""
Calculates the insulation score of a Hi-C interaction matrix.
Parameters
----------
mat : numpy.array
A dense square matrix of Hi-C interaction frequencies.
May contain nans, e.g. in rows/columns excluded from the analysis.
window : int
The width of the window to calculate the insulation score.
ignore_diags : int
If > 0, the interactions at separations < `ignore_diags` are ignored
when calculating the insulation score. Typically, a few first diagonals
of the Hi-C map should be ignored due to contamination with Hi-C
artifacts.
norm_by_median : bool
If True, normalize the insulation score by its NaN-median.
"""
if (ignore_diags):
mat = mat.copy()
for i in range(-ignore_diags + 1, ignore_diags):
numutils.set_diag(mat, np.nan, i)
with warnings.catch_warnings():
warnings.simplefilter("ignore")
N = mat.shape[0]
score = np.nan * np.ones(N)
for i in range(0, N):
lo = max(0, i + 1 - window)
hi = min(i + window, N)
# nanmean of interactions to reduce the effect of bad bins
score[i] = np.nanmean(mat[lo:i + 1, i:hi])
if norm_by_median:
score /= np.nanmedian(score)
return score
def main():
usage = 'usage: %prog [options] <genome_hic_file> <seqs_bed_file> <seqs_hic_file>'
parser = OptionParser(usage)
parser.add_option('-b', dest='blacklist_bed',
help='Set blacklist nucleotides to a baseline value.')
parser.add_option('--clip', dest='clip',
default=None, type='float',
help='Clip values post-summary to a maximum [Default: %default]')
parser.add_option('--crop', dest='crop_bp',
default=0, type='int',
help='Crop bp off each end [Default: %default]')
parser.add_option('-d', dest='diagonal_offset',
default=2, type='int',
help='Positions on the diagonal to ignore [Default: %default]')
parser.add_option('-k', dest='kernel_stddev',
default=0, type='int',
help='Gaussian kernel stddev to smooth values [Default: %default]')
# parser.add_option('-s', dest='scale',
# default=1., type='float',
# help='Scale values by [Default: %default]')
parser.add_option('-w',dest='pool_width',
default=1, type='int',
help='Average pooling width [Default: %default]')
parser.add_option('--as_obsexp',dest='as_obsexp',
default=False,action="store_true",
help='save targets as obsexp profiles')
parser.add_option('--global_obsexp',dest='global_obsexp',
default=False,action="store_true",
help='use global obs/exp')
parser.add_option('--no_log',dest='no_log',
default=False,action="store_true",
help='no not take log for obs/exp')
(options, args) = parser.parse_args()
if len(args) != 3:
parser.error('')
else:
genome_hic_file = args[0]
seqs_bed_file = args[1]
seqs_hic_file = args[2]
# read model sequences
model_seqs = []
for line in open(seqs_bed_file):
a = line.split()
model_seqs.append(ModelSeq(a[0],int(a[1]),int(a[2]),None))
# read blacklist regions
black_chr_trees = read_blacklist(options.blacklist_bed)
# compute dimensions
num_seqs = len(model_seqs)
seq_len_nt = model_seqs[0].end - model_seqs[0].start
seq_len_pool = seq_len_nt // options.pool_width
if options.crop_bp == 0:
seq_len_crop = seq_len_pool
else:
crop_start = options.crop_bp // options.pool_width
crop_end = seq_len_pool - crop_start
seq_len_crop = seq_len_pool - 2*crop_start
# compute upper triangular indexes
triu_tup = np.triu_indices(seq_len_crop, options.diagonal_offset)
seq_len_nodiag = seq_len_crop - options.diagonal_offset
seq_len_hic = seq_len_nodiag*(seq_len_nodiag + 1) // 2
# initialize sequences coverage file
seqs_hic_open = h5py.File(seqs_hic_file, 'w')
seqs_hic_open.create_dataset('targets', shape=(num_seqs, seq_len_hic), dtype='float16')
if options.kernel_stddev > 0:
# initialize Gaussian kernel
kernel = Gaussian2DKernel(x_stddev=options.kernel_stddev)
else:
kernel = None
# open genome coverage file
genome_hic_cool = cooler.Cooler(genome_hic_file)
if options.global_obsexp:
try:
print('loading by-chromosome expected')
genome_hic_expected = pd.read_csv(genome_hic_file.replace('.cool','.expected'), sep='\t')
except:
print('not found: '+genome_hic_file.replace('cool','expected'))
raise ValueError('invalid expected file')
# check for "chr" prefix
chr_pre = 'chr1' in genome_hic_cool.chromnames
# assert that resolution matches
assert(options.pool_width == genome_hic_cool.info['bin-size'])
# for each model sequence
for si in range(num_seqs):
mseq = model_seqs[si]
try:
# pull hic values
if chr_pre:
mseq_str = '%s:%d-%d' % (mseq.chr, mseq.start, mseq.end)
else:
mseq_str = '%s:%d-%d' % (mseq.chr[3:], mseq.start, mseq.end)
#print('mseq_str:', mseq_str)
seq_hic_raw = genome_hic_cool.matrix(balance=True).fetch(mseq_str)
seq_hic_nan = np.isnan(seq_hic_raw)
num_filtered_bins = np.sum(np.sum(seq_hic_nan,axis=0) == len(seq_hic_nan))
if num_filtered_bins > (.5*len(seq_hic_nan)):
print("WARNING: %s >50% bins filtered, check: %s. " % (genome_hic_file, mseq_str))
# set blacklist to NaNs
if mseq.chr in black_chr_trees:
for black_interval in black_chr_trees[mseq.chr][mseq.start:mseq.end]:
# adjust for sequence indexes
black_seq_start = (black_interval.begin - mseq.start)// options.pool_width
black_seq_end = int( np.ceil( (black_interval.end - mseq.start)/ options.pool_width ) )
seq_hic_raw[:,black_seq_start:black_seq_end] = np.nan
seq_hic_raw[black_seq_start:black_seq_end,:] = np.nan
seq_hic_nan = np.isnan(seq_hic_raw)
# clip first diagonals and high values
clipval = np.nanmedian(np.diag(seq_hic_raw,options.diagonal_offset))
for i in range(-options.diagonal_offset+1,options.diagonal_offset):
set_diag(seq_hic_raw, clipval, i)
seq_hic_raw = np.clip(seq_hic_raw, 0, clipval)
seq_hic_raw[seq_hic_nan] = np.nan
# adaptively coarsegrain based on raw counts
seq_hic_smoothed = adaptive_coarsegrain(
seq_hic_raw,
genome_hic_cool.matrix(balance=False).fetch(mseq_str),
cutoff=2, max_levels=8)
seq_hic_nan = np.isnan(seq_hic_smoothed)
#todo: pass an option to add a certain pseudocount value, or the minimum nonzero value
if options.as_obsexp:
# compute obs/exp
if options.global_obsexp: # compute global obs/exp
exp_chr = genome_hic_expected.iloc[ genome_hic_expected['chrom'].values ==mseq.chr][0:seq_len_pool]
if len(exp_chr) ==0:
raise ValueError('no expected values found for chr:'+mseq.chr)
exp_map= np.zeros((seq_len_pool,seq_len_pool))
for i in range(seq_len_pool):
set_diag(exp_map,exp_chr['balanced.avg'].values[i],i)
set_diag(exp_map,exp_chr['balanced.avg'].values[i],-i)
seq_hic_obsexp = seq_hic_smoothed / exp_map
for i in range(-options.diagonal_offset+1,options.diagonal_offset): set_diag(seq_hic_obsexp,1.0,i)
seq_hic_obsexp[seq_hic_nan] = np.nan
else: # compute local obs/exp
seq_hic_obsexp = observed_over_expected(seq_hic_smoothed, ~seq_hic_nan)[0]
# log
if options.no_log==False:
seq_hic_obsexp = np.log(seq_hic_obsexp)
if options.clip is not None:
seq_hic_obsexp = np.clip(seq_hic_obsexp, -options.clip, options.clip)
seq_hic_obsexp = interp_nan(seq_hic_obsexp)
for i in range(-options.diagonal_offset+1, options.diagonal_offset): set_diag(seq_hic_obsexp, 0,i)
else:
if options.clip is not None:
seq_hic_obsexp = np.clip(seq_hic_obsexp, 0, options.clip)
seq_hic_obsexp = interp_nan(seq_hic_obsexp)
for i in range(-options.diagonal_offset+1, options.diagonal_offset): set_diag(seq_hic_obsexp, 1,i)
# apply kernel
if kernel is not None:
seq_hic = convolve(seq_hic_obsexp, kernel)
else:
seq_hic = seq_hic_obsexp
else:
# interpolate all missing bins
seq_hic_interpolated = interp_nan(seq_hic_smoothed)
# rescale, reclip
seq_hic = 100000*seq_hic_interpolated
clipval = np.nanmedian(np.diag(seq_hic,options.diagonal_offset))
for i in range(-options.diagonal_offset+1, options.diagonal_offset):
set_diag(seq_hic,clipval,i)
seq_hic = np.clip(seq_hic, 0, clipval)
#extra smoothing. todo pass kernel specs
if kernel is not None:
seq_hic = convolve(seq_hic, kernel)
except ValueError:
print("WARNING: %s doesn't see %s. Setting to all zeros." % (genome_hic_file, mseq_str))
seq_hic = np.zeros((seq_len_pool,seq_len_pool), dtype='float16')
# crop
if options.crop_bp > 0:
seq_hic = seq_hic[crop_start:crop_end,:]
seq_hic = seq_hic[:,crop_start:crop_end]
# unroll upper triangular
seq_hic = seq_hic[triu_tup]
# write
seqs_hic_open['targets'][si,:] = seq_hic.astype('float16')
# close sequences coverage file
seqs_hic_open.close()
def main():
usage = 'usage: %prog [options] <genome_hic_file> <seqs_bed_file> <seqs_hic_file>'
parser = OptionParser(usage)
parser.add_option('-b',
dest='blacklist_bed',
help='Set blacklist nucleotides to a baseline value.')
parser.add_option(
'-c',
dest='clip',
default=None,
type='float',
help='Clip values post-summary to a maximum [Default: %default]')
parser.add_option('-s',
dest='scale',
default=1.,
type='float',
help='Scale values by [Default: %default]')
parser.add_option(
'--soft',
dest='soft_clip',
default=False,
action='store_true',
help=
'Soft clip values, applying sqrt to the execess above the threshold [Default: %default]'
)
parser.add_option(
'-u',
dest='sum_stat',
default='sum',
help='Summary statistic to compute in windows [Default: %default]')
parser.add_option('-w',
dest='pool_width',
default=1,
type='int',
help='Average pooling width [Default: %default]')
parser.add_option('--as_obsexp',
dest='as_obsexp',
default=False,
action="store_true",
help='save targets as obsexp profiles')
(options, args) = parser.parse_args()
if len(args) != 3:
parser.error('')
else:
genome_hic_file = args[0]
seqs_bed_file = args[1]
seqs_hic_file = args[2]
print('saving TFRs as obsexp:', options.as_obsexp)
# read model sequences
model_seqs = []
for line in open(seqs_bed_file):
a = line.split()
model_seqs.append(ModelSeq(a[0], int(a[1]), int(a[2]), None))
# read blacklist regions
black_chr_trees = read_blacklist(options.blacklist_bed)
# compute dimensions
num_seqs = len(model_seqs)
seq_len_nt = model_seqs[0].end - model_seqs[0].start
seq_len_pool = seq_len_nt // options.pool_width
# initialize sequences coverage file
seqs_hic_open = h5py.File(seqs_hic_file, 'w')
seqs_hic_open.create_dataset('seqs_hic',
shape=(num_seqs, seq_len_pool, seq_len_pool),
dtype='float16')
# open genome coverage file
genome_hic_cool = cooler.Cooler(genome_hic_file)
# check for "chr" prefix
chr_pre = 'chr1' in genome_hic_cool.chromnames
# assert that resolution matches
assert (options.pool_width == genome_hic_cool.info['bin-size'])
# for each model sequence
for si in range(num_seqs):
mseq = model_seqs[si]
try:
# pull hic values
if chr_pre:
mseq_str = '%s:%d-%d' % (mseq.chr, mseq.start, mseq.end)
else:
mseq_str = '%s:%d-%d' % (mseq.chr[3:], mseq.start, mseq.end)
#print('mseq_str:', mseq_str)
seq_hic_raw = genome_hic_cool.matrix(balance=True).fetch(mseq_str)
seq_hic_nan = np.isnan(seq_hic_raw)
if np.sum(seq_hic_nan[len(seq_hic_nan) // 2 -
1:len(seq_hic_nan) // 2 + 1,
len(seq_hic_nan) // 2 -
2:len(seq_hic_nan) // 2 + 2]) > 4:
print(
"WARNING: %s lots of zeros, check that umap_midpoint is correct %s. "
% (genome_hic_file, mseq_str))
# set blacklist to NaNs
if mseq.chr in black_chr_trees:
for black_interval in black_chr_trees[
mseq.chr][mseq.start:mseq.end]:
# adjust for sequence indexes
black_seq_start = (black_interval.begin -
mseq.start) // options.pool_width
black_seq_end = int(
np.ceil((black_interval.end - mseq.start) /
options.pool_width))
seq_hic_raw[:, black_seq_start:black_seq_end] = np.nan
seq_hic_raw[black_seq_start:black_seq_end, :] = np.nan
seq_hic_nan = np.isnan(seq_hic_raw)
# clip first diagonals and high values
clipval = np.nanmedian(np.diag(seq_hic_raw, 2))
for i in [-1, 0, 1]:
set_diag(seq_hic_raw, clipval, i)
seq_hic_raw = np.clip(seq_hic_raw, 0, seq_hic_raw)
seq_hic_raw[seq_hic_nan] = np.nan
# adaptively coarsegrain based on raw counts
seq_hic_smoothed = adaptive_coarsegrain(
seq_hic_raw,
genome_hic_cool.matrix(balance=False).fetch(mseq_str),
cutoff=2,
max_levels=8)
#todo: pass an option to add a certain pseudocount value, or the minimum nonzero value
if options.as_obsexp == True:
# interpolate single missing bins
seq_hic_interpolated = interpolate_bad_singletons(
seq_hic_smoothed,
mask=(~seq_hic_nan),
fillDiagonal=True,
returnMask=False,
secondPass=True,
verbose=False)
seq_hic_nan = np.isnan(seq_hic_interpolated)
# compute observed/expected
seq_hic_obsexp = observed_over_expected(
seq_hic_interpolated, ~seq_hic_nan)[0]
# todo: allow passing a global expected rather than computing locally
# log
seq_hic_obsexp = np.log(seq_hic_obsexp)
# set nan to 0
seq_hic_obsexp = np.nan_to_num(seq_hic_obsexp)
# todo: make obsexp_clip an option for obs/exp
seq_hic = np.clip(seq_hic_obsexp, -2, 2)
else:
# interpolate all missing bins
seq_hic_interpolated = interp_nan(seq_hic_smoothed)
# rescale
seq_hic = 100000 * seq_hic_interpolated
# todo add extra smoothing
except ValueError:
print("WARNING: %s doesn't see %s. Setting to all zeros." %
(genome_hic_file, mseq_str))
seq_hic = np.zeros((seq_len_pool, seq_len_pool), dtype='float16')
# write
seqs_hic_open['seqs_hic'][si, :, :] = seq_hic.astype('float16')
# close sequences coverage file
seqs_hic_open.close()
