def get_patient_set(self, filters):
f1 = list(filters)
filter_df = pd.concat(f1, axis=1)
clinical_filter = filter_df.dropna().sum(1) == 0
keepers_o = H.true_index(clinical_filter)
keepers_o = keepers_o.intersection(self.mut_df.columns)
keepers_o = keepers_o.intersection(self.cna_df.columns)
return keepers_o
def mut_filter(df, rate, binary_cutoff=12):
'''
Filter out mutation features, ensuring that a feature
is not entirely an artifact of mutation rate.
'''
df = df[df.sum(1) >= binary_cutoff]
cc = H.screen_feature(rate, rev_kruskal, df)
fc_apply = lambda s: fc(s, rate)
direction = df.apply(fc_apply, axis=1)
direction.name = 'direction'
cc = cc.join(direction)
cc = cc[cc.direction==False]
df = df.ix[H.true_index(cc.p > .01)]
df = df.dropna(axis=1)
return df
def process_real(df):
'''
Process real valued feature into binary feature.
'''
df_c = df.copy()
df_c = df_c.apply(lambda s: H.to_quants(s, std=1), axis=1)
df_c = df_c > 0
if type(df.index) == pd.MultiIndex:
df_c.index = map(lambda s: '_'.join(s), df_c.index)
return df_c.T
def process_real(df):
"""
Process real valued feature into binary feature.
"""
df_c = df.copy()
df_c = df_c.apply(lambda s: H.to_quants(s, std=1), axis=1)
df_c = df_c > 0
if type(df.index) == pd.MultiIndex:
df_c.index = map(lambda s: '_'.join(s), df_c.index)
return df_c.T
def get_patient_set(self, filters):
f1 = list(filters)
filter_df = pd.concat(f1, axis=1)
clinical_filter = filter_df.dropna().sum(1) == 0
keepers_o = H.true_index(clinical_filter)
keepers_o = keepers_o.intersection(self.mut_df.columns)
keepers_o = keepers_o.intersection(self.cna_df.columns)
keepers_o = keepers_o.intersection(self.surv.unstack().index)
keepers_o = keepers_o.intersection(self.rna_df.columns)
keepers_o = keepers_o.intersection(self.mirna_df.columns)
return keepers_o
def mut_filter(df, rate, binary_cutoff=12):
"""
Filter out mutation features, ensuring that a feature
is not entirely an artifact of mutation rate.
"""
get_min_count = lambda s: s.value_counts().min() if len(s.unique()) > 1 else -1
df = df[df.apply(get_min_count, axis=1) > binary_cutoff]
cc = H.screen_feature(rate, rev_kruskal, df)
fc_apply = lambda s: fc(s, rate)
direction = df.apply(fc_apply, axis=1)
direction.name = 'direction'
cc = cc.join(direction)
#cc = cc[cc.direction == False]
#return cc
df = df.ix[H.true_index((cc.p > .01) | (cc.direction == True))]
df = df.dropna(axis=1)
return df
def mut_filter(df, rate, binary_cutoff=12):
"""
Filter out mutation features, ensuring that a feature
is not entirely an artifact of mutation rate.
"""
get_min_count = lambda s: s.value_counts().min() if len(s.unique()
) > 1 else -1
df = df[df.apply(get_min_count, axis=1) > binary_cutoff]
cc = H.screen_feature(rate, rev_kruskal, df)
fc_apply = lambda s: fc(s, rate)
direction = df.apply(fc_apply, axis=1)
direction.name = 'direction'
cc = cc.join(direction)
#cc = cc[cc.direction == False]
#return cc
df = df.ix[H.true_index((cc.p > .01) | (cc.direction == True))]
df = df.dropna(axis=1)
return df
def corrections(vec):
'''
Correct p-values multiple ways along multi-index.
'''
bonf_all = vec * len(vec)
bonf_within = vec.groupby(level=0).apply(lambda s: s*len(s))
bh_all = H.bhCorrection(vec)
bh_within = vec.groupby(level=0).apply(H.bhCorrection).order()
two_step = bh_within * len(vec.groupby(level=0).size())
q = pd.concat([vec, bh_within, bh_all, bonf_all, bonf_within, two_step],
keys=['uncorrected', 'bh_within', 'bh_all', 'bonf_all', 'bonf_within',
'two_step'], axis=1)
return q
def corrections(vec):
"""
Correct p-values multiple ways along multi-index.
"""
bonf_all = vec * len(vec)
bonf_within = vec.groupby(level=0).apply(lambda s: s * len(s))
bh_all = H.bhCorrection(vec)
bh_within = vec.groupby(level=0).apply(H.bhCorrection).order()
two_step = bh_within * len(vec.groupby(level=0).size())
q = pd.concat([vec, bh_within, bh_all, bonf_all, bonf_within, two_step],
keys=[
'uncorrected', 'bh_within', 'bh_all', 'bonf_all',
'bonf_within', 'two_step'
],
axis=1)
return q
def remove_redundant_pathways(pathways, background, cutoff=.7,
binarize=False):
'''
Screens out redundant pathways with high correlation above _cutoff_.
Pathways are ranked based on lack of correlation to the background signal.
Then if two pathways have high correlation the lower ranked pathway is
removed.
'''
bg = H.screen_feature(background, spearman_pandas, pathways)
dd = pathways.ix[bg.index[::-1]].T.corr()
dd = pd.DataFrame(np.triu(dd, 1), dd.index, dd.index)
dd = dd.replace(0, np.nan).stack()
drop = dd[dd.abs() > cutoff].index.get_level_values(1)
pathways_to_keep = pathways.index.diff(drop.unique())
pathways = pathways.ix[pathways_to_keep]
if binarize is False:
return pathways
else:
binary_pathways = pathways.apply(binarize_feature, 1)
return binary_pathways
