def network_lasso(data, response_type='level', ground_truth=None, inhib_targets=None, perfect=True, group_stimuli=False):
'''
do lasso. automatically do CV to find best alpha.
input:
data
response_type : (level, rate)
ground_truth : adjacency matrix
group_stimuli : binary
'''
from sklearn import preprocessing, linear_model, cross_validation, metrics
# model interventions if supplied an inhib_targets dict
if inhib_targets:
training_dict = prepare_markov_data(introduce_inhibs(data, inhib_targets=inhib_targets, perfect=perfect), response_type, group_stimuli)
else:
training_dict = prepare_markov_data(data, response_type, group_stimuli)
antibodies = [col for col in data.columns if col not in ['Cell Line', 'Inhibitor', 'Stimulus', 'Timepoint']]
stims = set(data['Stimulus'])
# fit lasso for each (X,Y) pair
A = {}
for key in training_dict:
X = training_dict[key][0]
Y = training_dict[key][1]
preprocessing.StandardScaler().fit_transform(X)
A[key] = pd.DataFrame(np.zeros((X.shape[1], X.shape[1])), columns=X.columns, index=X.columns)
for col in Y.columns:
#print col
# check if col is not all the identical
if len(set(Y[col])) > 1:
rgn = linear_model.LassoCV(verbose=False).fit(X, Y[col])
if np.max(rgn.coef_) != 0:
A[key].ix[:,col] = np.abs(rgn.coef_) / np.abs(rgn.coef_).max()
else:
A[key].ix[:,col] = np.zeros((X.shape[1],))
if ground_truth:
auc = {}
for key in training_dict:
auc[key] = score_network(A[key], ground_truth)
return A, auc
else:
return A
'TestInhib3' : ['mTOR_pS2448'],
'TestInhib4' : ['EGFR_pY1068', 'EGFR_pY1173', 'EGFR_pY992'],
'TestInhib5' : []}
print '----------- ' + cell_line + ' ------------'
data = pd.read_csv('data/{0}_main.csv'.format(cell_line), header=0)
inhibs = set(data['Inhibitor'])
stims = set(data['Stimulus'])
node_list = data.columns[4:]
inhib_targets = {'GSK690693' : ['AKT_pT308','AKT_pS473'],
'GSK690693_GSK1120212' : ['AKT_pT308','AKT_pS473','MEK1_pS217_S221']}
regGBR= {}
scalar = {}
td = utilities.prepare_markov_data(utilities.introduce_inhibs(data, inhib_targets=inhib_targets, perfect=True), 'level', group_stimuli=False)
for stim in stims:
X, Y = td[stim]
scalar[stim] = preprocessing.StandardScaler()
scalar[stim].fit_transform(X)
X.ix[X.ix[:,'Inhib_GSK690693']>0,'Inhib_GSK690693'] = 1
X.ix[X.ix[:,'Inhib_GSK690693_GSK1120212']>0,'Inhib_GSK690693_GSK1120212'] = 1
X.ix[X.ix[:,'Inhib_GSK690693']<0,'Inhib_GSK690693'] = 0
X.ix[X.ix[:,'Inhib_GSK690693_GSK1120212']<0,'Inhib_GSK690693_GSK1120212'] = 0
n_estimators = 100
max_depth = 3
regGBR[stim] = do_gbr(X, Y, n_estimators=n_estimators, max_depth=max_depth)
# predict
uacc812 = pd.read_csv('data/UACC812_main.csv')
# load in the network prior curated from the literature
prior = pd.read_csv('data/experimental_prior.csv', index_col=0, header=0)
## BT20
##############################################
print '----------- ' + 'BT20' + ' ------------'
inhibs = set(bt20['Inhibitor'])
stims = set(bt20['Stimulus'])
node_list = bt20.columns[4:]
inhib_targets = {'GSK690693' : ['AKT_pT308','AKT_pS473'],
'GSK690693_GSK1120212' : ['AKT_pT308','AKT_pS473','MEK1_pS217_S221']}
regGBR= {}
scalar = {}
td_bt20 = utilities.prepare_markov_data(utilities.introduce_inhibs(bt20, inhib_targets=inhib_targets, perfect=True), 'level', group_stimuli=False)
for stim in td_bt20:
X, Y = td_bt20[stim]
scalar[stim] = preprocessing.StandardScaler()
scalar[stim].fit_transform(X)
X.ix[X.ix[:,'Inhib_GSK690693']>0,'Inhib_GSK690693'] = 1
X.ix[X.ix[:,'Inhib_GSK690693_GSK1120212']>0,'Inhib_GSK690693_GSK1120212'] = 1
X.ix[X.ix[:,'Inhib_GSK690693']<0,'Inhib_GSK690693'] = 0
X.ix[X.ix[:,'Inhib_GSK690693_GSK1120212']<0,'Inhib_GSK690693_GSK1120212'] = 0
n_estimators = 100
max_depth = 3
regGBR[stim] = do_gbr(X, Y, n_estimators=n_estimators, max_depth=max_depth)
adj = build_adj_matrix(regGBR, node_list, stims)
A_true_bt20 = prior
for stim in adj: