def condition_clustering(distance_name, folder='/media/lalil0u/New/projects/drug_screen/results/', color_gradient='YlOrRd',
hit_only=False, compare_to='MITO',
level_row=0.4, level_column=0.5,show=False,
filename='Clusters_{}_{}.pkl',
#to avoid reloading distance files each time
distances=None, all_exposures=None,
row_method='ward'
):
'''
DOING CONDITION CLUSTERING (MEDIAN OF EXPERIMENTS FOR THIS CONDITION)
- compare_to: We can do drug clustering based on their distances to eachother (value 'DS')
but we can also do drug clustering based on their distances to Mitocheck (value 'MITO')
- 'hit_only': to do clustering considering hit distances only or no
'''
f=open(os.path.join(folder, 'DS_hits_1.5IQR.pkl'))
_, _, exposure_hits=pickle.load(f)
f.close()
d=Counter(exposure_hits)
d={el:d[el]/float(PASSED_QC_COND[el]) for el in d}
distinct_exposure=filter(lambda x:d[x]>0.5, d)
if distances is None:
distances, _, exposure_, _=_return_right_distance(distance_name, folder, filter_replicates=True,
hit_only=hit_only, compare_to=compare_to)
if not hit_only:
all_exposures=sorted(Counter(exposure_).keys())
distances=np.vstack((np.median(distances[np.where(exposure_==condition)],0) for condition in all_exposures))
else:
all_exposures=sorted(distinct_exposure)
distances=np.vstack((np.median(distances[np.where(exposure_==condition)],0) for condition in all_exposures))
if compare_to=='MITO':
column_header=[k for k in range(distances.shape[1])]
else:
column_header=all_exposures
print distances.shape
clusters=hierarchical_clustering.heatmap(distances, row_header=all_exposures, column_header=column_header,
row_method=row_method, column_method='ward',
row_metric='euclidean', column_metric='euclidean',
color_gradient=color_gradient, filename="{}{}".format(int(hit_only),distance_name),
folder='{}/inference_Freplicates'.format(folder),
level_row=level_row, level_column=level_column,
title=drug_screen_utils.DISTANCES[distance_name],
colorbar_ticks=[-2, 0, 2],
colorbar_ticklabels=[0, '', 1], show=show,
colorbar_title='Distance (arbitrary units)',
range_normalization=(scoreatpercentile(distances.flatten(),10), scoreatpercentile(distances.flatten(), per=90)))
global_=np.bincount(clusters)
if not hit_only:
hit_clusters = clusters[np.where(np.array([el in distinct_exposure for el in all_exposures]))]
all_exposures=distinct_exposure
else:
hit_clusters=clusters
print len(hit_clusters), len(distinct_exposure)
print global_, np.bincount(hit_clusters)
who_cluster_hits={k: Counter(np.array(all_exposures)[np.where(hit_clusters==k)]) for k in range(1,np.max(clusters)+1)}
if hit_only:
f=open(os.path.join(folder, 'inference_Freplicates', filename.format(distance_name, level_row)), 'w')
pickle.dump(who_cluster_hits, f); f.close()
return distances, all_exposures, who_cluster_hits
def experiment_clustering(distance_name, folder='/media/lalil0u/New/projects/drug_screen/results/', color_gradient='YlOrRd',
hit_only=False, compare_to='MITO',
level=0.4):
'''
DOING EXPERIMENT CLUSTERING AS OPPOSED TO CONDITION CLUSTERING
- compare_to: We can do drug clustering based on their distances to eachother (value 'DS')
but we can also do drug clustering based on their distances to Mitocheck (value 'MITO')
- 'hit_only': to do clustering considering hit distances only or no
'''
distances, who_, exposure_, mito_who=_return_right_distance(distance_name, folder,
filter_replicates=True,
hit_only=hit_only, compare_to=compare_to)
plates=np.array([int(el.split('--')[0].split('_')[1]) for el in who_])
exposure_wPL=np.array(['{}{:>10}'.format(exposure_[i], plates[i]) for i in range(len(exposure_))])
f=open(os.path.join(folder, 'DS_hits_1.5IQR.pkl'))
_, who_hits, exposure_hits=pickle.load(f)
f.close()
d=Counter(exposure_hits)
d={el:d[el]/float(PASSED_QC_COND[el]) for el in d}
distinct_exposure=filter(lambda x:d[x]>0.5, d)
if hit_only:
wh_=np.hstack((np.where(who_==who_hits[i])[0] for i in range(len(who_hits)) if exposure_hits[i] in distinct_exposure))
distances=distances[wh_]
if compare_to=='DS':
distances=distances[:,wh_]
print wh_
who_=who_[wh_]
exposure_wPL=exposure_wPL[wh_]
plates=plates[wh_]
if compare_to=='MITO':
column_header=mito_who
else:
column_header=who_
clusters=hierarchical_clustering.heatmap(distances, row_header=exposure_wPL, column_header=column_header,
row_method='ward', column_method='ward',
row_metric='euclidean', column_metric='euclidean',
color_gradient=color_gradient, filename="E{}".format(distance_name),
folder='{}/inference_Freplicates'.format(folder),
level=level,title=drug_screen_utils.DISTANCES[distance_name],
colorbar_ticks=[-2, 0, 2],
colorbar_ticklabels=[0, '', 1],
colorbar_title='Distance (arbitrary units)',
range_normalization=(scoreatpercentile(distances.flatten(),10), scoreatpercentile(distances.flatten(), per=90)))
global_=np.bincount(clusters)
if not hit_only:
hit_clusters = clusters[np.where(np.array([el in who_hits for el in who_]))]
hit_clusters = hit_clusters[np.where(np.array([exposure_hits[i] in distinct_exposure for i in range(len(who_hits))]))]
plates=plates[np.where(np.array([el in who_hits for el in who_]))]
plates=plates[np.where(np.array([exposure_hits[i] in distinct_exposure for i in range(len(who_hits))]))]
exposure_hits=exposure_hits[np.where(np.array([exposure_hits[i] in distinct_exposure for i in range(len(who_hits))]))]
else:
hit_clusters=clusters
print len(hit_clusters), len(exposure_hits), len(plates)
print global_, np.bincount(hit_clusters)
who_cluster_hits={k: Counter(exposure_hits[np.where(hit_clusters==k)]) for k in range(1,np.max(clusters)+1)}
plate_clusters={k: Counter(plates[np.where(hit_clusters==k)]) for k in range(1,np.max(clusters)+1)}
return hit_clusters, who_cluster_hits, plate_clusters
