def simulate(self, cterms=['intrinsic', 'environmental', 'interactions'], interactions_size=None):
# train gp with requested terms
model = Model1(self.Y, self.X, norm='quantile', oos_predictions=0.,
cov_terms=cterms, kin_from=self.kin_from)
model.reset_params()
model.train_gp(grid_size=10)
# simulate from gp after removing interactions term
k = model.covar_terms['intrinsic'].K() + \
model.covar_terms['environmental'].K() + \
model.covar_terms['noise'].K()
k *= covar_rescaling_factor_efficient(k)
# manually add a cross-talk term
if interactions_size is not None:
assert 0. < interactions_size < 1., 'interactions size must be between 0 and 1 '
tmp = model.covar_terms['interactions'].K()
tmp *= covar_rescaling_factor_efficient(tmp)
tmp *= (interactions_size / (1. - interactions_size))
k += tmp
res = np.random.multivariate_normal([0.]*k.shape[0], k)
return res
def simulate_env(self):
kin = self.exp.dot(self.exp.transpose())
kin *= covar_rescaling_factor_efficient(kin)
tmp = ZKZCov(self.X, kin)
tmp.length = self.l2
k = tmp.K()
k *= covar_rescaling_factor_efficient(k)
self.covar += k
def reset_params(self):
if self.init_from_previous:
self.reset_from_previous()
return
# import pdb; pdb.set_trace()
self.intrinsic_cov.scale = 1.
self.interactions_cov.scale = 1.
self.environmental_cov.scale = 1.
self.noise_cov.scale = 1.
used_covar = self.covar_terms.values()
n = len(used_covar)
for cov in used_covar:
k = covar_rescaling_factor_efficient(cov.K()) / n
# TODO: not so clean, would be good to have setInterParams ?
new_params = np.log(cov.getInterParams() * k)
cov.setParams(new_params)
if self.use_scale_down:
for term in self.scale_down:
self.covar_terms[term].scale *= 1e-6
def simulate_local(self):
tmp = SQExpCov(self.X)
tmp.length = self.l2
k = tmp.K()
k *= covar_rescaling_factor_efficient(k)
self.covar += k
def build_Kinship(self):
if self.kin_from is None:
assert self.types is not None, 'Provide a vector of cell types or a cell state matrix to compute kinship'
self.Kin = utils.build_cell_type_kinship(self.types)
else:
self.kin_from -= self.kin_from.mean(axis=0)
#self.kin_from /= self.kin_from.std(axis=0)
self.Kin = self.kin_from.dot(self.kin_from.transpose())
self.Kin *= covar_rescaling_factor_efficient(self.Kin)
def simulate_crowding(self):
k1 = np.ones([self.n_samples, self.n_samples])
# kin *= covar_rescaling_factor_efficient(kin)
tmp = ZKZCov(self.X, k1)
tmp.length = self.l2
k = tmp.K()
k *= covar_rescaling_factor_efficient(k)
self.covar += k
def reset_from_previous(self):
for cov_key in self.covar_terms.keys():
if cov_key in self.init_covs.keys():
self.covar_terms[cov_key].setParams(self.init_covs[cov_key])
else:
self.covar_terms[cov_key].scale = 1.
k = covar_rescaling_factor_efficient(
self.covar_terms[cov_key].K())
self.covar_terms[cov_key].scale = 0.5 * k
if self.use_scale_down:
for term in self.scale_down:
self.covar_terms[term].scale *= 1e-10
self.covar = SumCov(*self.covar_terms.values())
self.build_gp()
def simulate_intrinsic(self):
kin = self.exp.dot(self.exp.transpose())
kin *= covar_rescaling_factor_efficient(kin)
self.covar += kin
def write_results(trained_model, output_dir, file_prefix, n_types, cell_types, cell_types_names, by_effective_type=True):
# extracting covariance terms from the trained model
####################################################################
intrinsic_cov = trained_model['intrinsic_cov']
noise_covs = trained_model['noise_covs']
local_noise_cov = trained_model['local_noise_cov']
env_covs = trained_model['env_covs']
# putting parameters together
####################################################################
if by_effective_type:
parameters = np.zeros([1, int(2 * n_types**2.0 + n_types + 2 + 1)])
else:
parameters = np.zeros([1, int(2 * n_types + n_types + 2 + 1)])
try:
parameters[0,0] = 1./covar_rescaling_factor_efficient(intrinsic_cov.K())
except:
parameters[0,0] = np.nan
try:
parameters[0,1] = 1./covar_rescaling_factor_efficient(local_noise_cov.K())
except:
parameters[0,1] = np.nan
parameters[0,2] = local_noise_cov.length
count = 3
for type1 in np.unique(cell_types):
cell_filter = (cell_types == type1)
assert np.all(noise_covs[type1].K()[:, ~cell_filter][~cell_filter,:] == 0), 'problem cell filter'
K_tmp = noise_covs[type1].K()[:, cell_filter][cell_filter,:]
try:
parameters[0,count] = 1./covar_rescaling_factor_efficient(K_tmp)
except:
parameters[0,count] = np.nan
count += 1
for type1 in np.unique(cell_types):
cell_filter = (cell_types == type1)
if by_effective_type:
for type2 in np.unique(cell_types):
K_tmp = env_covs[type1][type2].K()[:, cell_filter][cell_filter,:]
assert np.all(env_covs[type1][type2].K()[:, ~cell_filter][~cell_filter,:] ==0), 'problem cell filter'
try:
parameters[0,count] = 1./covar_rescaling_factor_efficient(K_tmp)
except:
parameters[0,count] = np.nan
count += 1
else:
K_tmp = env_covs[type1].K()[:, cell_filter][cell_filter,:]
assert np.all(env_covs[type1].K()[:, ~cell_filter][~cell_filter,:] ==0), 'problem cell filter'
try:
parameters[0,count] = 1./covar_rescaling_factor_efficient(K_tmp)
except:
parameters[0,count] = np.nan
count += 1
for type1 in np.unique(cell_types):
if by_effective_type:
for type2 in np.unique(cell_types):
parameters[0,count] = env_covs[type1][type2].length
count += 1
else:
parameters[0,count] = env_covs[type1].length
count += 1
# Putting header together
####################################################################
if by_effective_type:
result_header_array = [None for i in range(int(2 * n_types**2.0 + n_types + 2 + 1))]
else:
result_header_array = [None for i in range(int(2 * n_types + n_types + 2 + 1))]
result_header_array[0] = 'inrtinsic'
result_header_array[1] = 'local_noise_scale'
result_header_array[2] = 'local_noise_length'
count = 3
for type in np.unique(cell_types):
result_header_array[count] = 'noise_'+cell_types_names[type]
count += 1
for type1 in np.unique(cell_types):
if by_effective_type:
for type2 in np.unique(cell_types):
result_header_array[count] = 'env_'+cell_types_names[type1]+'_'+cell_types_names[type2]
count += 1
else:
result_header_array[count] = 'env_'+cell_types_names[type1]
count += 1
for type1 in np.unique(cell_types):
if by_effective_type:
for type2 in np.unique(cell_types):
result_header_array[count] = 'env_length_'+cell_types_names[type1]+'_'+cell_types_names[type2]
count += 1
else:
result_header_array[count] = 'env_length_'+cell_types_names[type1]
count += 1
# TODO concatenate list into single string
result_header = ' '.join(result_header_array)
# Write file
####################################################################
output_file = output_dir + '/' + file_prefix
with open(output_file, 'w') as f:
np.savetxt(f,
parameters,
delimiter=' ',
header=result_header,
fmt='%s',
comments='')