def calc_occupancy_scores(self, factor, motif, tf_concs=[1e-6,]):
mean_rvs = []
max_rvs = []
trimmed_atacseq_cov = self.atacseq_cov[len(motif)+1:]
atacseq_weights = trimmed_atacseq_cov/trimmed_atacseq_cov.max()
for tf_conc in tf_concs:
log_tf_conc = numpy.log(tf_conc)
raw_occ = logistic(
log_tf_conc - self.score_cov[motif.name]/(R*T))
occ = raw_occ*atacseq_weights
mean_rv.append(occ.mean())
max_rv.append(occ.max())
return mean_rv, max_rvs
def find_optimal_GFE(motif, pks, chipseq_scores, atacseq_signal):
res = []
max_len = max(len(pk) for pk in pks)
scores = numpy.zeros((len(pks), max_len+1))
for GFE in numpy.arange(-20, 10, 1.0):
motif.build_occupancy_weights(4, GFE)
for pk_i, pk in enumerate(pks):
score_cov = numpy.array(
[score for pos, score in motif.iter_seq_score(pk.seq)])
scores[pk_i, len(motif)+1:] = score_cov
res.append((
spearmanr((logistic(-scores/(R*T))*atacseq_signal).mean(1), chipseq_scores)[0],
GFE))
print motif.name, GFE, res[-1]
max_cor = max(x[0] for x in res)
print motif.name, [x[1] for x in res if x[0] == max_cor][0], max_cor
print >> sys.stderr, motif.name, [x[1] for x in res if x[0] == max_cor][0], max_cor
return (motif.name, [x[1] for x in res if x[0] == max_cor][0], max_cor)
def calc_summary_stats(self):
header = []
rv = []
# add on the region and atacseq data
header.append('pk_length')
rv.append(self.stop - self.start)
header.append('ATAC_mean')
rv.append(self.atacseq_cov.mean())
header.append('ATAC_max')
rv.append(self.atacseq_cov.max())
# find all factors with motif and chip-seq data
factors = sorted(set(motif.factor for name, motif
in self.motifs.iteritems()
).intersection(self.chipseq_cov.iterkeys()))
percentiles = numpy.array(
[1e-3, 1e-2, 0.02, 0.05, 0.10, 0.25, 0.50])
for factor in sorted(factors):
for BSID, cov in self.chipseq_cov[factor].iteritems():
header.append('%s_%s_mean_ChIPseq_cov' % (factor, BSID))
rv.append(cov.mean())
for motif_name, motif in sorted(self.motifs.iteritems()):
# skip motifs that aren't the correct factor
if factor != motif.factor: continue
header.append('%s_mean_score' % motif_name)
rv.append(self.score_cov[motif_name].mean())
header.append('%s_max_score' % motif_name)
rv.append(self.score_cov[motif_name].min())
#for percentile, score in self.iter_upper_rank_means(
# self.score_cov[motif_name], percentiles):
# header.append('%s_q_%.2f_score' % (motif_name, percentile))
# rv.append(score)
trimmed_atacseq_cov = self.atacseq_cov[len(motif)+1:]
atacseq_weights = trimmed_atacseq_cov/trimmed_atacseq_cov.max()
#1000, trimmed_atacseq_cov.max())
w_pwm_scores = self.pwm_cov[motif_name]*atacseq_weights
header.append('%s_mean_w_pwm_score' % motif_name)
rv.append(w_pwm_scores.mean())
#for percentile, score in self.iter_upper_rank_means(
# w_pwm_scores, percentiles):
# header.append('%s_q_%.2f_w_pwm_score' % (motif_name, percentile))
# rv.append(score)
#header.append('%s_max_w_pwm_score' % motif_name)
#rv.append(w_pwm_scores.max())
"""
for tf_conc in [1e-30, 1e-20, 1e-15, 1e-10, 1e-7, 1e-5, 1e-2, 1e-1,
1, 1e2, 1e5, 1e7, 1e10, 1e15, 1e20, 1e30 ]:
log_tf_conc = numpy.log(tf_conc)
raw_occ = logistic(
log_tf_conc - self.score_cov[motif.name]/(R*T))
occ = raw_occ*atacseq_weights
header.append('%s_%e_mean_occ' % (motif_name, tf_conc))
rv.append(occ.mean())
#header.append('%s_%e_max_occ' % (motif_name, tf_conc))
#rv.append(occ.max())
"""
log_tf_conc = numpy.log(1e5)
raw_occ = logistic(
log_tf_conc - self.score_cov[motif_name]/(R*T))
occ = raw_occ*atacseq_weights
header.append('%s_mean_occ' % motif_name)
rv.append(occ.mean())
header.append('%s_max_occ' % motif_name)
rv.append(occ.max())
"""
# XXX
# find the raw occupancy that provies the best correpondence
# between the signals, and then try and predict these
# sequentially
unbnd_conc = self.estimate_unbnd_conc_in_region(motif_name)
#print self.score_cov[motif_name].mean(), unbnd_conc
#print unbnd_conc
#unbnd_conc = 0.0
raw_occ = logistic(
unbnd_conc + self.score_cov[motif_name]/(R*T))
occ = raw_occ*atacseq_weights
header.append('%s_weighted_occ' % motif_name)
rv.append(occ.mean())
header.append('%s_unbnd_conc' % motif_name)
rv.append(unbnd_conc)
"""
#for percentile, score in self.iter_upper_rank_means(
# occ, percentiles):
# header.append('%s_q_%.2f_occ' % (motif_name, percentile))
# rv.append(score)
#header.append('%s_max_occ' % motif_name)
#rv.append(occ.max())
return header, rv