def fn(dist):
position = dist + 500
sstart, send = motif_coords(side_motif, position)
seqlets = [Seqlet(None, cstart, cend, "center", ""),
Seqlet(None, sstart, send, "side", "")]
# TODO - add also contribution scores
du = {"p": p[position], "contrib": contrib[position]}
# TODO - order them correctly
d = OrderedDict([(f"{prefix}/{task}", du[prefix][task])
for task in p[position]
for prefix in ['p', 'contrib']])
ylims = []
for k in d:
if k.startswith("p"):
ylims.append((0, ymax))
else:
ylims.append((0, ymax_contrib))
plot_tracks(d,
seqlets,
title=dist, ylim=ylims)
def plot_regions(self, regions, ds=None, variants=None,
seqlets=[],
pred_summary='profile/wn',
contrib_method='grad',
batch_size=128,
# ylim=None,
xlim=None,
# seq_height=1,
rotate_y=0,
add_title=True,
fig_height_per_track=2,
same_ylim=False,
fig_width=20):
"""Plot predictions
Args:
regions: list of pybedtools.Interval
variant: a single instance or a list of bpnet.extractors.Variant
ds: DataSpec. If provided, the ground truth will be added to the plot
pred_summary: 'mean' or 'max', summary function name for the profile gradients
"""
out = self.predict_regions(regions,
variants=variants,
contrib_method=contrib_method,
# pred_summary=pred_summary,
batch_size=batch_size)
figs = []
if xlim is None:
xmin = 0
else:
xmin = xlim[0]
shifted_seqlets = [s.shift(-xmin) for s in seqlets]
for i in range(len(out)):
pred = out[i]
interval = out[i]['interval']
if ds is not None:
obs = {task: ds.task_specs[task].load_counts([interval])[0] for task in self.tasks}
else:
obs = None
title = "{i.chrom}:{i.start}-{i.end}, {i.name} {v}".format(i=interval, v=pred.get('variant', ''))
# handle the DNase case
if isinstance(pred['seq'], dict):
seq = pred['seq']['seq']
else:
seq = pred['seq']
if obs is None:
# TODO - simplify?
viz_dict = OrderedDict(flatten_list([[
(f"{task} Pred", pred['pred'][task]),
(f"{task} Contrib profile", pred['contrib_score'][f"{task}/{pred_summary}"] * seq),
# (f"{task} Contrib counts", sum(pred['grads'][task_idx]['counts'].values()) / 2 * seq),
] for task_idx, task in enumerate(self.tasks)]))
else:
viz_dict = OrderedDict(flatten_list([[
(f"{task} Pred", pred['pred'][task]),
(f"{task} Obs", obs[task]),
(f"{task} Contrib profile", pred['contrib_score'][f"{task}/{pred_summary}"] * seq),
# (f"{task} Contrib counts", sum(pred['grads'][task_idx]['counts'].values()) / 2 * seq),
] for task_idx, task in enumerate(self.tasks)]))
if add_title:
title = "{i.chrom}:{i.start}-{i.end}, {i.name} {v}".format(i=interval, v=pred.get('variant', '')),
else:
title = None
if same_ylim:
fmax = {feature: max([np.abs(viz_dict[f"{task} {feature}"]).max() for task in self.tasks])
for feature in ['Pred', 'Contrib profile', 'Obs']}
ylim = []
for k in viz_dict:
f = k.split(" ", 1)[1]
if "Contrib" in f:
ylim.append((-fmax[f], fmax[f]))
else:
ylim.append((0, fmax[f]))
else:
ylim = None
fig = plot_tracks(filter_tracks(viz_dict, xlim),
seqlets=shifted_seqlets,
title=title,
fig_height_per_track=fig_height_per_track,
rotate_y=rotate_y,
fig_width=fig_width,
ylim=ylim,
legend=True)
figs.append(fig)
return figs
from bpnet.plot.tracks import plot_tracks, to_neg
import seaborn as sns
import matplotlib.pyplot as plt
cf = ContribFile(contrib_file)
profiles = cf.get_profiles()
contrib_scores = cf.get_contrib()
examples = list({v.max(axis=-2).mean(axis=-1).argmax() for k,v in profiles.items()})
examples
tasks = ['Oct4', 'Sox2', 'Nanog']
fig = plt.figure(figsize=[8,8])
xrange = slice(50, 150)
for idx in examples:
plot_tracks({**{'profile/' + k: to_neg(v[idx,xrange]) for k,v in profiles.items()},
**{'contrib/' + k:v[idx,xrange] for k,v in contrib_scores.items()}},
title=idx,
rotate_y=0,
fig_width=10,
fig_height_per_track=1);
sns.despine(top=True, right=True, bottom=True)
fig.savefig(f"/groups/lackgrp/ll_members/berkay/SYNTAX/exampleRun/results/train/bpnet/examples/chip-nexus/output/contributions.pdf")
## convert dfi to Seqlet objects
xrange = slice(50, 150)
seqlets = [s.shift(-xrange.start)
for s in dfi2seqlets(dfi[dfi.example_idx == idx], short_name=True)]
## Visualize the locus with motif instances highlighted
## get the contribution scores and profile score for that example idx
xrange = slice(300, 700)
cf = ContribFile(contrib_file)
profiles = cf.get_profiles(idx=idx)
contrib_scores = cf.get_contrib(idx=idx)
## Let's focus only on the best match per track
dfi_best = dfi[dfi.example_idx == idx].sort_values("match_weighted_p", ascending=False).groupby('tf').first()
seqlets = [s.shift(-xrange.start)
for s in dfi2seqlets(dfi_best, short_name=True)]
fig = plot_tracks({**{'profile/' + k: to_neg(v[xrange]) for k,v in profiles.items()},
**{'contrib/' + k:v[xrange] for k,v in contrib_scores.items()}},
title=idx,
rotate_y=0,
fig_width=20,
seqlets=[s.set_seqname('contrib/' + s.name.split("/")[0]) for s in seqlets], # plot seqlets to the 'contrib/Nanog' track
fig_height_per_track=1);
sns.despine(top=True, right=True, bottom=True)
fig.savefig(output_file, bbox_inches = "tight")