def fitFactor(self,idx=None,X0=None,k=1,standardize=False, use_ard=False, interaction=True, initMethod='fast'):
"""
Args:
idx: index of the genes involved
(e.g., for capturing cell cycle, index of cell cycle genes)
X0: known factor(s) on which to condition on when fitting the GPLVM
k: number of latent factors
standardize: if True, rescale gene expression by std prior to fitting GPLVM
(data are always mean-centered)
use_ard: use automatic relevance detection (switch off unimportant factors)
Returns:
X: hidden variable
Kconf: similarity matrix based on the confounding effect (XX.T)
varGPLVM: variance contributions of latent factors and residual biological noise
"""
assert idx!=None, 'scLVM:: specify idx'
if use_ard==True and k<2:
warnings.formatwarning = warning_on_one_line
warnings.warn('when using ARD consider choosing k>1')
if X0!=None:
assert use_ard == False, 'scLVM:: when fitting conditional GPLVM, use_ard has to be False'
Yconf = self.Y[:,idx]
Yconf-= Yconf.mean(0)
Kint = None
if X0==None:
#use PANAMA
# prepare data
# fit gplvm
panama = PANAMA.PANAMA(Y=Yconf,use_Kpop=False,standardize=standardize)
panama.train(rank=k,LinearARD=use_ard)
X = panama.get_Xpanama()
Kconf = panama.get_Kpanama()
var = panama.get_varianceComps()
if use_ard==False:
varGPLVM = {'K':var['Kpanama'],'noise':var['noise']}
else:
varGPLVM = {'X_ARD':var['LinearARD'],'noise':var['noise']}
else:
# use gpCLVM
gp = gpCLVM(Y=Yconf,X0=X0,k=k,standardize=standardize,interaction=interaction)
params0 = gp.initParams(method=initMethod)
conv = gp.optimize(params0)
X = gp.getX()
Kconf = gp.getK()
if interaction==True:
Kint = gp.getKi()
varGPLVM = gp.getVarianceComps()
return X,Kconf,Kint,varGPLVM
def fitFactor(self,idx=None,X0=None,k=1,standardize=False, use_ard=False, interaction=True, initMethod='fast'):
"""
Args:
idx: index of the genes involved
(e.g., for capturing cell cycle, index of cell cycle genes)
X0: known factor(s) on which to condition on when fitting the GPLVM
k: number of latent factors
standardize: if True, rescale gene expression by std prior to fitting GPLVM
(data are always mean-centered)
use_ard: use automatic relevance detection (switch off unimportant factors)
Returns:
X: hidden variable
Kconf: similarity matrix based on the confounding effect (XX.T)
varGPLVM: variance contributions of latent factors and residual biological noise
"""
assert idx!=None, 'scLVM:: specify idx'
if use_ard==True and k<2:
warnings.formatwarning = warning_on_one_line
warnings.warn('when using ARD consider choosing k>1')
if X0!=None:
assert use_ard == False, 'scLVM:: when fitting conditional GPLVM, use_ard has to be False'
Yconf = self.Y[:,idx]
Yconf-= Yconf.mean(0)
Kint = None
if X0==None:
#use PANAMA
# prepare data
# fit gplvm
panama = PANAMA.PANAMA(Y=Yconf,use_Kpop=False,standardize=standardize)
panama.train(rank=k,LinearARD=use_ard)
X = panama.get_Xpanama()
Kconf = panama.get_Kpanama()
var = panama.get_varianceComps()
if use_ard==False:
varGPLVM = {'K':var['Kpanama'],'noise':var['noise']}
else:
varGPLVM = {'X_ARD':var['LinearARD'],'noise':var['noise']}
else:
# use gpCLVM
gp = gpCLVM(Y=Yconf,X0=X0,k=k,standardize=standardize,interaction=interaction)
params0 = gp.initParams(method=initMethod)
conv = gp.optimize(params0)
X = gp.getX()
Kconf = gp.getK()
if interaction==True:
Kint = gp.getKi()
varGPLVM = gp.getVarianceComps()
return X,Kconf,Kint,varGPLVM
pdb.set_trace()
# initialization method
init_method = 'null' #['fast','regressOut','random','null']
Ycc = Y_X0+1e-2*SP.randn(N,G) # cell cycle contributions for regressOut
X_init = X_true+1e-2*SP.randn(N,k) # initial Xs for regressOut
varXX = 0.3 # variance explained by the normalized XX.T
varX0X0 = 0.3 # variance explained by known factor
nois = 0.7 # variance explained by noise
if 1:
"""
gp with interaction term
"""
gp = gpCLVM(Y=Y,X0=X0,k=k)
params0 = gp.initParams(method=init_method,Ycc=Ycc,X=X_init,varXX=varXX,varX0X0=varX0X0,nois=nois)
# gp.fix_a(flag=True)
conv = gp.optimize(params0)
X = gp.getX() # inferred hidden factor
K = gp.getK() # matrix from inferred inferred factor
Ki = gp.getKi() # interaction matrix
var = gp.getVarianceComps() # variance components
if 1:
""" plot """
PL.subplot(2,2,1)
PL.plot(X_true.ravel(),X.ravel(),'.k')
PL.xlabel('true X')
PL.ylabel('fitted X')