def parse_inputs(
samp1_bm_fns, samp2_bm_fns, strand_offset, samp_names, valid_pos_fn,
out_fp):
# parse valid positions
valid_pos = None
if valid_pos_fn is not None:
valid_pos = mh.parse_beds(
valid_pos_fn, ignore_strand=strand_offset is not None)
# parse bed methyl files
samp1_cov, samp1_mod_cov = mh.parse_bed_methyls(
samp1_bm_fns, strand_offset=strand_offset, valid_pos=valid_pos)
samp1_all_cov = np.array([cov for ctg_cov in samp1_cov.values()
for cov in ctg_cov.values()])
samp2_cov, samp2_mod_cov = mh.parse_bed_methyls(
samp2_bm_fns, strand_offset=strand_offset, valid_pos=valid_pos)
samp2_all_cov = np.array([cov for ctg_cov in samp2_cov.values()
for cov in ctg_cov.values()])
out_fp.write(
'{} coverage median: {:.2f} mean: {:.2f} sd: {:.2f}\n'.format(
samp_names[0], np.median(samp1_all_cov),
np.mean(samp1_all_cov), np.std(samp1_all_cov)))
out_fp.write(
'{} coverage median: {:.2f} mean: {:.2f} sd: {:.2f}\n'.format(
samp_names[1], np.median(samp2_all_cov),
np.mean(samp2_all_cov), np.std(samp2_all_cov)))
return (samp1_cov, samp1_mod_cov, samp1_all_cov,
samp2_cov, samp2_mod_cov, samp2_all_cov)
def parse_inputs(samp1_bm_fns, samp2_bm_fns, strand_offset, samp_names,
valid_pos_fn, out_fp):
# parse valid positions
valid_pos = None
if valid_pos_fn is not None:
LOGGER.info("Parsing valid sites bed")
valid_pos = mh.parse_beds(
valid_pos_fn,
ignore_strand=strand_offset is not None,
show_prog_bar=False,
)
# parse bed methyl files
LOGGER.info("Parsing bedmethyl files")
samp1_cov, samp1_mod_cov = mh.parse_bed_methyls(
samp1_bm_fns,
strand_offset=strand_offset,
valid_pos=valid_pos,
show_prog_bar=False,
)
samp1_all_cov = np.array(
[cov for ctg_cov in samp1_cov.values() for cov in ctg_cov.values()])
samp2_cov, samp2_mod_cov = mh.parse_bed_methyls(
samp2_bm_fns,
strand_offset=strand_offset,
valid_pos=valid_pos,
show_prog_bar=False,
)
samp2_all_cov = np.array(
[cov for ctg_cov in samp2_cov.values() for cov in ctg_cov.values()])
out_fp.write(
"{} coverage median: {:.2f} mean: {:.2f} sd: {:.2f}\n".format(
samp_names[0],
np.median(samp1_all_cov),
np.mean(samp1_all_cov),
np.std(samp1_all_cov),
))
out_fp.write(
"{} coverage median: {:.2f} mean: {:.2f} sd: {:.2f}\n".format(
samp_names[1],
np.median(samp2_all_cov),
np.mean(samp2_all_cov),
np.std(samp2_all_cov),
))
return (
samp1_cov,
samp1_mod_cov,
samp1_all_cov,
samp2_cov,
samp2_mod_cov,
samp2_all_cov,
)
def parse_valid_sites(valid_sites_fns, gt_data_fn, include_strand):
if valid_sites_fns is None and gt_data_fn is None:
return None, None, None
# if ground truth file provided, parse first
if gt_data_fn is not None:
LOGGER.info("Reading ground truth file")
gt_mod_pos, gt_ctrl_pos = mh.parse_ground_truth_file(
gt_data_fn, include_strand=include_strand)
if valid_sites_fns is None:
# if ground truth provided, but not valid sites return parsed
# ground truth sites.
return (
[
gt_mod_pos,
],
None,
[
gt_ctrl_pos,
],
)
# parse valid sites files and intersect with ground truth (if provided)
LOGGER.info("Reading valid sites data")
valid_sites, vs_labs = [], []
ctrl_sites = None if gt_data_fn is None else []
for vs_lab, valid_sites_fn in valid_sites_fns:
try:
vs_i_sites = mh.parse_beds([
valid_sites_fn,
])
except FileNotFoundError:
LOGGER.warning(
"Could not find valid sites file: {}".format(valid_sites_fn))
continue
vs_i_sites = set(((chrm, strand, pos) if include_strand else
(chrm, pos)
for (chrm, strand), cs_pos in vs_i_sites.items()
for pos in cs_pos))
if gt_data_fn is None:
valid_sites.append(vs_i_sites)
else:
ctrl_sites.append(vs_i_sites.intersection(gt_ctrl_pos))
valid_sites.append(vs_i_sites.intersection(gt_mod_pos))
vs_labs.append(vs_lab)
if len(valid_sites) == 0:
return None, None, None
return valid_sites, vs_labs, ctrl_sites
def compute_val_metrics(mod_samp,
ctrl_samp,
gt_data,
out_fp,
pdf_fp,
balance_classes,
ignore_strand,
samp_name='sample',
valid_pos_fn=None):
# extract ground truth either from mod and control samples or ground truth
# data
if gt_data is None:
if valid_pos_fn is not None:
valid_pos = mh.parse_beds([
valid_pos_fn,
],
ignore_strand=ignore_strand)
mod_samp = mod_samp._replace(test_sites=dict(
(ctg, valid_pos[ctg].intersection(ctg_sites))
for ctg, ctg_sites in mod_samp.test_sites.items()
if ctg in valid_pos))
ctrl_samp = ctrl_samp._replace(test_sites=dict(
(ctg, valid_pos[ctg].intersection(ctg_sites))
for ctg, ctg_sites in ctrl_samp.test_sites.items()
if ctg in valid_pos))
mod_pct_mod = np.array([
100 * mod_samp.mod_cov[ctg][pos] / mod_samp.cov[ctg][pos]
for ctg, ctg_poss in mod_samp.test_sites.items()
for pos in ctg_poss
])
ctrl_pct_mod = np.array([
100 * ctrl_samp.mod_cov[ctg][pos] / ctrl_samp.cov[ctg][pos]
for ctg, ctg_poss in ctrl_samp.test_sites.items()
for pos in ctg_poss
])
else:
mod_pct_mod, ctrl_pct_mod = [], []
for ctg, pos_is_mod in gt_data.items():
try:
ctg_cov = mod_samp.cov[ctg]
ctg_mod_cov = mod_samp.mod_cov[ctg]
except KeyError:
continue
for pos, is_mod in pos_is_mod:
try:
pos_cov = ctg_cov[pos]
pos_mod_cov = ctg_mod_cov[pos]
except KeyError:
continue
if is_mod:
mod_pct_mod.append(100 * pos_mod_cov / pos_cov)
else:
ctrl_pct_mod.append(100 * pos_mod_cov / pos_cov)
mod_pct_mod = np.array(mod_pct_mod)
ctrl_pct_mod = np.array(ctrl_pct_mod)
if balance_classes:
if mod_pct_mod.shape[0] > ctrl_pct_mod.shape[0]:
mod_pct_mod = np.random.choice(mod_pct_mod,
ctrl_pct_mod.shape[0],
replace=False)
elif mod_pct_mod.shape[0] < ctrl_pct_mod.shape[0]:
ctrl_pct_mod = np.random.choice(ctrl_pct_mod,
mod_pct_mod.shape[0],
replace=False)
all_pct_mod = np.concatenate([mod_pct_mod, ctrl_pct_mod])
if all_pct_mod.shape[0] == 0:
LOGGER.info(
'Skipping "{}". No vaild sites available.'.format(samp_name))
return
is_mod = np.repeat((1, 0), (mod_pct_mod.shape[0], ctrl_pct_mod.shape[0]))
precision, recall, thresh = precision_recall_curve(is_mod, all_pct_mod)
prec_recall_sum = precision + recall
valid_idx = np.where(prec_recall_sum > 0)
all_f1 = (2 * precision[valid_idx] * recall[valid_idx] /
prec_recall_sum[valid_idx])
optim_f1_idx = np.argmax(all_f1)
optim_f1 = all_f1[optim_f1_idx]
optim_thresh = thresh[optim_f1_idx]
avg_prcn = average_precision_score(is_mod, all_pct_mod)
fpr, tpr, _ = roc_curve(is_mod, all_pct_mod)
roc_auc = auc(fpr, tpr)
out_fp.write(
MOD_VAL_METRICS_TMPLT.format(optim_f1, optim_thresh, avg_prcn, roc_auc,
mod_pct_mod.shape[0],
ctrl_pct_mod.shape[0], samp_name))
LOGGER.info('Plotting {}'.format(samp_name))
plt.figure(figsize=(11, 7))
try:
sns.kdeplot(mod_pct_mod,
shade=True,
bw_adjust=MOD_BANDWIDTH,
gridsize=MOD_GRIDSIZE,
label='Yes')
sns.kdeplot(ctrl_pct_mod,
shade=True,
bw_adjust=MOD_BANDWIDTH,
gridsize=MOD_GRIDSIZE,
label='No')
except AttributeError:
sns.kdeplot(mod_pct_mod,
shade=True,
bw=MOD_BANDWIDTH2,
gridsize=MOD_GRIDSIZE,
label='Yes')
sns.kdeplot(ctrl_pct_mod,
shade=True,
bw=MOD_BANDWIDTH2,
gridsize=MOD_GRIDSIZE,
label='No')
plt.legend(prop={'size': 16}, title='Is Modified?')
plt.xlabel('Percent Modified')
plt.ylabel('Density')
plt.title(samp_name)
pdf_fp.savefig(bbox_inches='tight')
plt.close()
plt.figure(figsize=(8, 7))
plt.step(recall, precision, where='post')
plt.ylim([-0.05, 1.05])
plt.xlim([-0.05, 1.05])
plt.xlabel('Recall')
plt.ylabel('Precision')
plt.title(('{} Precision-Recall curve: AP={:0.2f}').format(
samp_name, avg_prcn))
pdf_fp.savefig(bbox_inches='tight')
plt.close()
plt.figure(figsize=(8, 7))
plt.plot(fpr, tpr)
plt.xlim([-0.05, 1.05])
plt.ylim([-0.05, 1.05])
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title(('{} ROC curve: auc={:0.2f}').format(samp_name, roc_auc))
pdf_fp.savefig(bbox_inches='tight')
plt.close()