Source code for pytadbit.tadbit

"""
24 Oct 2012
"""
from __future__ import print_function

from os                           import path, listdir
from pytadbit.parsers.hic_parser  import read_matrix
from pytadbit.tadbit_py           import _tadbit_wrapper
from math                         import isnan, sqrt
from scipy.sparse.csr             import csr_matrix
from scipy.stats                  import mannwhitneyu
import numpy as np


[docs]def tadbit(x, remove=None, n_cpus=1, verbose=True, max_tad_size="max", no_heuristic=0, use_topdom=False, topdom_window=5, **kwargs): """ The TADbit algorithm works on raw chromosome interaction count data. The normalization is neither necessary nor recommended, since the data is assumed to be discrete counts. TADbit is a breakpoint detection algorithm that returns the optimal segmentation of the chromosome under BIC-penalized likelihood. The model assumes that counts have a Poisson distribution and that the expected value of the counts decreases like a power-law with the linear distance on the chromosome. This expected value of the counts at position (i,j) is corrected by the counts at diagonal positions (i,i) and (j,j). This normalizes for different restriction enzyme site densities and 'mappability' of the reads in case a bin contains repeated regions. :param x: a square matrix of interaction counts in the HI-C data or a list of such matrices for replicated experiments. The counts must be evenly sampled and not normalized. x might be either a list of list, a path to a file or a file handler :argument 'visibility' norm: kind of normalization to use. Choose between 'visibility' of 'Imakaev' :argument None remove: a python list of lists of booleans mapping positively columns to remove (if None only columns with a 0 in the diagonal will be removed) :param 1 n_cpus: The number of CPUs to allocate to TADbit. If n_cpus='max' the total number of CPUs will be used :param auto max_tad_size: an integer defining maximum size of TAD. Default (auto or max) defines it as the number of rows/columns :param False no_heuristic: whether to use or not some heuristics :param False use_topdom: whether to use TopDom algorithm to find tads or not (http://www.ncbi.nlm.nih.gov/pubmed/26704975, http://zhoulab.usc.edu/TopDom/) :param 5 topdom_window: the window size for topdom algorithm :param False get_weights: either to return the weights corresponding to the Hi-C count (weights are a normalization dependent of the count of each columns) :returns: the :py:func:`list` of topologically associated domains' boundaries, and the corresponding list associated log likelihoods. If no weights are given, it may also return calculated weights. """ nums = [hic_data for hic_data in read_matrix(x, one=False)] if not use_topdom: size = len(nums[0]) nums = [num.get_as_tuple() for num in nums] if not remove: # if not given just remove columns with zero in diagonal remove = tuple([0 if nums[0][i*size+i] else 1 for i in range(size)]) n_cpus = n_cpus if n_cpus != 'max' else 0 max_tad_size = size if max_tad_size in ["max", "auto"] else max_tad_size _, nbks, passages, _, _, bkpts = \ _tadbit_wrapper(nums, # list of lists of Hi-C data remove, # list of columns marking filtered size, # size of one row/column len(nums), # number of matrices n_cpus, # number of threads int(verbose), # verbose 0/1 max_tad_size, # max_tad_size kwargs.get('ntads', -1) + 1, int(no_heuristic),# heuristic 0/1 ) breaks = [i for i in range(size) if bkpts[i + nbks * size] == 1] scores = [p for p in passages if p > 0] result = {'start': [], 'end' : [], 'score': []} for brk in range(len(breaks)+1): result['start'].append((breaks[brk-1] + 1) if brk > 0 else 0) result['end' ].append(breaks[brk] if brk < len(breaks) else size - 1) result['score'].append(scores[brk] if brk < len(breaks) else None) else: result = {'start': [], 'end' : [], 'score': [], 'tag': []} ret = TopDom(nums[0],window_size=topdom_window) for key in sorted(ret): result['tag'].append(ret[key]['tag']) result['start'].append(ret[key]['start']) result['end'].append(ret[key]['end']) if ret[key]['tag'] == 'domain': result['score'].append(ret[key]['score']) else: result['score'].append(0) max_score = max(result['score']) for i in range(len(result['score'])): result['score'][i] = 1-int((result['score'][i]/max_score)*10) return result
[docs]def batch_tadbit(directory, parser=None, **kwargs): """ Use tadbit on directories of data files. All files in the specified directory will be considered data file. The presence of non data files will cause the function to either crash or produce aberrant results. Each file has to contain the data for a single unit/chromosome. The files can be separated in sub-directories corresponding to single experiments or any other organization. Data files that should be considered replicates have to start with the same characters, until the character sep. For instance, all replicates of the unit 'chr1' should start with 'chr1\_', using the default value of sep. The data files are read through read.delim. You can pass options to read.delim through the list read_options. For instance if the files have no header, use read_options=list(header=FALSE) and if they also have row names, read_options=list(header=FALSE, row.names=1). Other arguments such as max_size, n_CPU and verbose are passed to :func:`tadbit`. NOTE: only used externally, not from Chromosome :param directory: the directory containing the data files :param kwargs: arguments passed to :func:`tadbit` function :param None parser: a parser function that takes file name as input and returns a tuple representing the matrix of data. Tuple is a concatenation of column1 + column2 + column3 + ... :returns: A :py:func:`list` where each element has the name of the unit/chromosome, and is the output of :func:`tadbit` run on the corresponding files assumed to be replicates """ matrix = [] for f_name in sorted(listdir(directory)): if f_name.startswith('.'): continue f_name = path.join(directory, f_name) if parser: matrix.append(parser(f_name)) continue elif not path.isfile(f_name): continue matrix.append(f_name) return tadbit(matrix, **kwargs)
def TopDom(hic_data,window_size,statFilter=True): """ Python implementation of the algorithm TopDom for the identification of TADs. See http://www.ncbi.nlm.nih.gov/pubmed/26704975 and http://zhoulab.usc.edu/TopDom/ :param hic_data: a list corresponding to the Hi-C data :param window_size: window size parameter for the TopDom algorithm :param True statFilter: whether to apply or not statistical filtering for false detection of TADs :returns: the :py:func:`list` of topologically associated domains, boundaries and gaps. Domains include the mean value of computed p-values by Wilcox Ranksum Test as score while boundaries and gaps have a score of zero. """ n_bins = len(hic_data) mean_cf = np.zeros(n_bins) pvalue = np.ones(n_bins) local_ext = np.ones(n_bins)*(-0.5) #Step 1 csr_mat = hic_data.get_hic_data_as_csr() for i in range(n_bins): diamond_mean = Get_Diamond_Matrix_Mean(data=csr_mat, i=i, size=window_size) mean_cf[i] = diamond_mean #Step 2 gap_idx = Which_Gap_Region(data=csr_mat) proc_regions = Which_process_region(rmv_idx=gap_idx, n_bins=n_bins, min_size=3) for key in proc_regions: start = proc_regions[key]["start"] end = proc_regions[key]["end"] #print "Process Regions from " + str(start) + " to " + str(end) local_ext[start:end+1] = Detect_Local_Extreme(x=mean_cf[start:end+1]) if statFilter: #Step 3 lil_mat = csr_mat.todense() for k in range(1,(2*window_size)): mat_row = [] mat_column = [] my_range = list(range((n_bins*k), (n_bins*n_bins), 1+n_bins)) mat_values = np.empty(len(my_range)) col_arr = 0 for j in my_range: mat_row.append(int(round(j//lil_mat.shape[0]))) mat_column.append(j%lil_mat.shape[0]) mat_values[col_arr] = lil_mat[int(round(j//lil_mat.shape[0])),j%lil_mat.shape[0]] col_arr = col_arr + 1 scale_values = scale(mat_values) for i in range(len(mat_row)): lil_mat[mat_column[i],mat_row[i]] = scale_values[i] for key in proc_regions: start = proc_regions[key]['start'] end = proc_regions[key]['end'] pvalue[start:end] = Get_Pvalue(data=lil_mat[start:end+1, start:end+1], size=window_size, scale=1) for i in range(len(local_ext)): if local_ext[i] == -1 and pvalue[i] < 0.05: local_ext[i] = -2 local_ext[local_ext==-1] = 0 local_ext[local_ext==-2] = -1 pvalue_cut=0.05 else: pvalue = None pvalue_cut=None domains = Convert_Bin_To_Domain_TMP(n_bins=n_bins, signal_idx=np.where(local_ext==-1)[0], gap_idx=np.where(local_ext==-0.5)[0], pvalues=pvalue, pvalue_cut=pvalue_cut) return domains def Get_Diamond_Matrix_Mean(data, i, size): n_bins = data.shape[1] if i==n_bins-1: return lowerbound = max( 0, i-size+1 ) upperbound = min( i+size+1, n_bins) return (data[lowerbound:(i+1),(i+1):upperbound].mean()) def Which_Gap_Region(data): n_bins = data.shape[1] gap = np.zeros(n_bins) i=0 while i < n_bins: j = i + 1 while j < n_bins: if data[i:j+1, i:j+1].sum() == 0: gap[i:j+1] = -0.5 j = j+1 else: break i = j idx = np.where(gap==-0.5)[0] #return dict(zip(idx,idx)) return idx def Which_process_region(rmv_idx, n_bins, min_size): gap_idx = rmv_idx proc_regions = dict() proc_set = np.arange(n_bins) proc_set = np.setdiff1d(proc_set,gap_idx) n_proc_set = proc_set.shape[0] i=0 while i < n_proc_set: start = proc_set[i] j = i+1 while j < n_proc_set: if proc_set[j] - proc_set[j-1] <= 1: j = j + 1 else: tmp_dict = {'start':start,'end':proc_set[j-1]} if abs(proc_set[j-1]-start) >= min_size: proc_regions[start]=tmp_dict i = j break if j >= n_proc_set: tmp_dict = {'start':start,'end':proc_set[j-1]} if abs(proc_set[j-1]-start) >= min_size: proc_regions[start]=tmp_dict break return(proc_regions) def Detect_Local_Extreme(x): n_bins = len(x) ret = np.zeros(n_bins) x[np.isnan(x)]=0 if n_bins <= 3: ret[np.argmin(x)]=-1 ret[np.argmax(x)]=1 return ret # Norm##################################################3 new_point_x, new_point_y = Data_Norm(x=np.arange(n_bins), y=x) x=new_point_y cp,Fv,Ev = Change_Point(x=np.arange(n_bins), y=x) if len(cp) <= 2: return ret for i in range(1,len(cp)-1): if x[cp[i]] >= x[cp[i]-1] and x[cp[i]] >= x[cp[i]+1]: ret[cp[i]] = 1 else: if x[cp[i]] < x[cp[i]-1] and x[cp[i]] < x[cp[i]+1]: ret[cp[i]] = -1 min_val = min( x[ cp[i-1] ], x[ cp[i] ] ) max_val = max( x[ cp[i-1] ], x[ cp[i] ] ) if np.min( x[cp[i-1]:cp[i]+1] ) < min_val: ret[ cp[i-1] + np.argmin( x[cp[i-1]:cp[i]+1] ) ] = -1 if np.max( x[cp[i-1]:cp[i]+1] ) > max_val: ret[ cp[i-1] + np.argmax( x[cp[i-1]:cp[i]+1] ) ] = 1 return ret def Data_Norm(x, y): ret_x = np.zeros(len(x)) ret_y = np.zeros(len(y)) ret_x[0] = x[0] ret_y[0] = y[0] diff_x = np.diff(x) diff_y = np.diff(y) scale_x = 1 / ( np.abs(np.diff(x) ) ).mean() scale_y = 1 / ( np.abs(np.diff(y) ) ).mean() for i in range(1,len(x)): ret_x[i] = ret_x[i-1] + (diff_x[i-1]*scale_x) ret_y[i] = ret_y[i-1] + (diff_y[i-1]*scale_y) #return dict(zip(ret_x,ret_y)) return ret_x, ret_y def Change_Point(x, y): if len(x) != len(y): print("ERROR : The length of x and y should be the same") return 0 n_bins = len(x) Fv = np.empty(n_bins) Fv[:] = np.NAN Ev = np.empty(n_bins) Ev[:] = np.NAN cp = [] cp.append(0) #print x i=0 Fv[0]=0 while i < n_bins-1: j=i+1 Fv[j] = sqrt( (x[j]-x[i])**2 + (y[j] - y[i] )**2 ) #print Fv[j] while j < n_bins-1: j=j+1 #k=(i+1):(j-1) Ev[j] = ( ( np.abs( (y[j]-y[i] )*x[(i+1):j] - (x[j] -x[i])*y[(i+1):j] - (x[i]*y[j]) + (x[j]*y[i]) ) ).sum() / sqrt( (x[j]-x[i])**2 + (y[j] - y[i] )**2 ) ) #print Ev[j] #print x[(i+1):j] Fv[j] = sqrt( (x[j]-x[i])**2 + (y[j] - y[i])**2 ) - ( ( np.abs( (y[j]-y[i] )*x[(i+1):j] - (x[j] -x[i])*y[(i+1):j] - (x[i]*y[j]) + (x[j]*y[i]) ) ).sum() / sqrt( (x[j]-x[i])**2 + (y[j] - y[i] )**2 ) ) ################################################# #Not Original Code if isnan(Fv[j]) or isnan(Fv[j-1]): j = j-1 cp.append(j) break ####################################################3 if Fv[j] < Fv[j-1]: j = j - 1 cp.append(j) break i=j cp.append(n_bins-1) return cp, Fv, Ev def Convert_Bin_To_Domain_TMP(n_bins, signal_idx, gap_idx, pvalues=None, pvalue_cut=None): bins = dict() rmv_idx = np.setdiff1d(np.arange(n_bins),gap_idx) proc_region = Which_process_region(rmv_idx, n_bins, min_size=0) for key in proc_region: bins[proc_region[key]['start']] = {'start': proc_region[key]['start'], 'end' : (proc_region[key]['end']+1), 'score': 10, 'tag' : 'gap'} rmv_idx = np.union1d(signal_idx, gap_idx) proc_region = Which_process_region(rmv_idx, n_bins, min_size=0) for key in proc_region: bins[proc_region[key]['start']] = {'start': proc_region[key]['start'], 'end' : (proc_region[key]['end']+1), 'score': 10, 'tag' : 'domain'} rmv_idx = np.setdiff1d(np.arange(n_bins),signal_idx) proc_region = Which_process_region(rmv_idx, n_bins, min_size=1) for key in proc_region: bins[proc_region[key]['start']] = {'start': proc_region[key]['start'], 'end' : (proc_region[key]['end']+1), 'score': 10, 'tag' : 'boundary'} if pvalues is not None and pvalue_cut is not None: for key in bins: if bins[key]['tag'] == 'domain': start_id = bins[key]['start'] end_id = bins[key]['end'] p_value_constr = pvalues[start_id:end_id] bins[key]['score'] = p_value_constr.mean() p_value_constr = p_value_constr[p_value_constr < pvalue_cut] if end_id - start_id == len(p_value_constr): bins[key]['tag'] = "boundary" return bins def scale(y): x = y.copy() x -= np.mean(x) x /= np.std(x, axis = 0, ddof = 1) # WTF different in numpy and R http://stackoverflow.com/questions/6457755/standard-deviation-in-r-seems-to-be-returning-the-wrong-answer-am-i-doing-some return x def Get_Pvalue(data, size, scale): n_bins = data.shape[0] pvalue = np.ones(n_bins-1) for i in range(1,n_bins): dia = Get_Diamond_Matrix2(data, i, size=size) ups = Get_Upstream_Triangle(data, i, size=size) downs = Get_Downstream_Triangle(data, i, size=size) wil_test = mannwhitneyu(x=dia*scale, y=ups+downs, use_continuity=True, alternative='less') pvalue[i-1] = wil_test.pvalue pvalue[ np.isnan(pvalue) ] = 1 return(pvalue) def Get_Diamond_Matrix2(data, i, size): n_bins = data.shape[0] #new_mat = np.ones_like(data)*np.NaN new_mat = np.ones(shape=(size,size))*np.NaN for k in range(1,size+1): if i-(k-1) >= 1 and i < n_bins: lower = min(i+1, n_bins) upper = min(i+size, n_bins) new_mat[size-(k-1)-1,0:(upper-lower+1)] = data[i-(k-1)-1,lower-1:upper] new_mat = new_mat[np.logical_not(np.isnan(new_mat))] return ((new_mat.transpose()).flatten()).tolist() def Get_Upstream_Triangle(data, i, size): lower = max(1, i-size) tmp_mat = data[lower-1:i,lower-1:i] triag = (np.triu(tmp_mat,k=1).flatten()) return triag[triag!=0].tolist() def Get_Downstream_Triangle(data, i, size): n_bins = data.shape[0] if i==n_bins: return NaN upperbound = min(i+size, n_bins) tmp_mat = data[i:upperbound, i:upperbound] triag = (np.triu(tmp_mat,k=1).flatten()) return triag[triag!=0].tolist() def insulation_score(hic_data, dists, normalize=False, resolution=1, delta=0, silent=False, savedata=None, savedeltas=None): """ Compute insulation score. :param hic_dada: HiC_data object already normalized :param dists: list of pairs of distances between which to compute the insulation score. E.g. 4,5 means that for a given bin B(i), all interactions between B(i-4) to B(i-5) and B(i+4) to B(i+5) will be summed and used to compute the insulation score. :param False normalize: Normalize insulation score by the average in the chromosome, and log2 of this ratio. :param 1 resolution: :param 0 delta: to compute the delta for TAD detection (e.g. at 10kb use 10) :param False silent: :param None savedata: path to file where to save result :param None savedeltas: path to file where to save deltas :returns: dictionary with insulation score """ bias = hic_data.bias bads = hic_data.bads decay = hic_data.expected if not decay or not bias: raise Exception('ERROR: HiC_data should be normalized by visibility ' 'and by expected') insidx = {} deltas = {} for dist, end in dists: if not silent: print(' - computing insulation in band %d-%d' % (dist, end)) insidx[(dist, end)] = {} deltas[(dist, end)] = {} for crm in hic_data.chromosomes: if crm in decay: this_decay = decay[crm] else: this_decay = decay total = 0 count = 0 for pos in range(hic_data.section_pos[crm][0] + end, hic_data.section_pos[crm][1] - end): val = sum(hic_data[i, j] / bias[i] / bias[j] / this_decay[abs(j-i)] for i in range(pos - end, pos - dist + 1) if not i in bads for j in range(pos + dist, pos + end + 1) if not j in bads) total += val count += 1 insidx[(dist, end)][pos] = val if normalize: try: total /= float(count) except ZeroDivisionError: pass if total == 0: total = float('nan') for pos in range(hic_data.section_pos[crm][0] + end, hic_data.section_pos[crm][1] - end): try: with np.errstate(divide='ignore'): insidx[(dist, end)][pos] = np.log2(insidx[(dist, end)][pos] / total) except ZeroDivisionError: insidx[(dist, end)][pos] = float('nan') if deltas: for pos in range(hic_data.section_pos[crm][0] + end, hic_data.section_pos[crm][1] - end): up_vals = [] dw_vals = [] for spos in range(delta): try: up_vals.append(insidx[(dist, end)][pos - delta + spos]) except KeyError: pass try: dw_vals.append(insidx[(dist, end)][pos + delta - spos]) except KeyError: pass with np.errstate(invalid='ignore'): deltas[(dist, end)][pos] = (np.mean(up_vals) - np.mean(dw_vals)) if savedata: out = open(savedata, 'w') out.write('# CRM\tCOORD\t' + '\t'.join(['%d-%d' % (d1, d2) for d1, d2 in dists]) + '\n') for crm in hic_data.section_pos: for pos in range(*hic_data.section_pos[crm]): beg = (pos - hic_data.section_pos[crm][0]) * resolution out.write('{}\t{}-{}\t{}\n'.format( crm, beg + 1, beg + resolution, '\t'.join([str(insidx[dist].get(pos, 'NaN')) for dist in dists]))) out.close() if savedeltas: out = open(savedeltas, 'w') out.write('# CRM\tCOORD\t' + '\t'.join(['%d-%d' % (d1, d2) for d1, d2 in dists]) + '\n') for crm in hic_data.section_pos: for pos in range(*hic_data.section_pos[crm]): beg = (pos - hic_data.section_pos[crm][0]) * resolution out.write('{}\t{}-{}\t{}\n'.format( crm, beg + 1, beg + resolution, '\t'.join([str(deltas[dist].get(pos, 'NaN')) for dist in dists]))) out.close() if delta: return insidx, deltas return insidx def insulation_to_borders(ins_score, deltas, min_strength=0.1): """ Best (for human-like genome size) according to https://doi.org/10.1038/nature14450 is (at 10kb resolution) to use awindow size of 500 kb (use the function insulation_score with dist=(1,50)) and a delta of 100 kb (10 bins). :returns: the position in bin of each border, and the intensity of the border (sigmoid normalized, from 0 to 1) """ borders = [] for pos in range(max(ins_score)): if (ins_score.get(pos, 100) >= ins_score.get(pos + 1, -100) or ins_score.get(pos, 100) >= ins_score.get(pos - 1, -100)): continue if not (deltas.get(pos - 1, 100) > 0 and deltas.get(pos + 1, -100) < 0): continue # left lo = 1 prev_lv = deltas.get(pos - 1, 100) while pos - 1 - lo > 0: lv = deltas.get(pos - 1 - lo, 100) if lv < prev_lv: break prev_lv = lv lo += 1 # right ro = 1 prev_rv = deltas.get(pos + 1, -100) while pos + 1 + ro <= len(deltas): rv = deltas.get(pos + 1 + ro, -100) if rv > prev_rv: break prev_rv = rv ro += 1 strength = 1. / (1 + np.exp(-(prev_lv - prev_rv))) * 2 - 1 if strength > min_strength: borders.append((pos, strength)) return borders