def __init__(self,
snps_test,
phenotype,
K=None,
snps_K=None,
covariates=None,
h2=None,
interact_with_snp=None,
nGridH2=10,
standardizer=None,
add_bias=True,
normalize_K=True,
blocksize=10000):
if standardizer is None:
standardizer = pysnptools.standardizer.Unit()
self.standardizer = standardizer
self.phenotype = GWAS.check_pheno_format(phenotype=phenotype)
self.covariates = GWAS.check_pheno_format(phenotype=covariates)
self.snps_test = snps_test
self.snps_K = snps_K
self.K = K
self.linreg = False
if (self.K is None) and (self.snps_K is None):
G = np.zeros((self.sample_count, 0))
self.linreg = True
elif self.K is None:
G = None
self.K = snps_K.kernel(standardizer=standardizer,
blocksize=blocksize)
if normalize_K:
self.K /= self.K.diagonal().mean()
elif snps_K is None:
G = None
if normalize_K:
self.K = self.K / self.K.diagonal().mean()
else:
raise NotImplementedError("either K or snps_K has to be None")
if add_bias and (self.covariates is None):
pass # this case is treated in LMM()
elif add_bias:
bias = pd.Series(np.ones((self.covariates.shape[0])),
index=self.covariates.index)
self.covariates['bias'] = bias
elif (add_bias == False) and self.covariates is None:
raise NotImplementedError(
"currently a model with neither bias term, nor covariates is not supported."
)
self.interact_with_snp = interact_with_snp
self.nGridH2 = nGridH2
if self.covariates is not None:
self.lmm = lmm_cov.LMM(X=self.covariates.values,
Y=None,
G=G,
K=self.K,
inplace=True)
else:
self.lmm = lmm_cov.LMM(X=None, Y=None, G=G, K=self.K, inplace=True)
self._find_h2()
def __init__(self,
Y,
X=None,
appendbias=False,
forcefullrank=False,
G0=None,
K0=None,
nullModel=None,
altModel=None):
association.varcomp_test.__init__(self,
Y=Y,
X=X,
appendbias=appendbias)
N = self.Y.shape[0]
self.forcefullrank = forcefullrank
self.nullModel = nullModel
self.altModel = altModel
self.G0 = G0
self.K0 = K0
self.__testGcalled = False
self.lmm = lmm.LMM(forcefullrank=self.forcefullrank,
X=self.X,
linreg=None,
Y=self.Y[:, np.newaxis],
G=self.G0,
K=self.K0,
regressX=True)
self.model0 = self.lmm.findH2(
) # The null model only has a single kernel and only needs to find h2
self.model1 = None
def est_h2(Y, K, covariates=None, nGridH2=10000, plot=True, verbose=True):
"""
This function implements the Bayesian heritability estimate from Furlotte et al., 2014
Furlotte, Nicholas A., David Heckerman, and Christoph Lippert.
"Quantifying the uncertainty in heritability." Journal of human genetics 59.5 (2014): 269-275.
Args:
Y: [N x 1] np.ndarray of phenotype values
K: [N x N] np.ndarray of kinship values
covariates: [N x D] np.ndarray of covariate values [default: None]
plot: Boolean, create a plot? [default: True]
verbose: print results? [default: True]
returns:
REML estimate of h^2 (as in Yang et al. 2010)
posterior mean of h^2
posterior variance of h^2
h2 values on a grid
posterior for h2 values on a grid
"""
lmm = lmm_cov.LMM(forcefullrank=False,
X=covariates,
linreg=None,
Y=Y,
G=None,
K=K,
regressX=True,
inplace=False)
h2 = lmm.findH2()
h2_posterior = lmm.posterior_h2(nGridH2=nGridH2)
logp = -h2_posterior[2]
grid = h2_posterior[1]
post_h2 = np.exp(logp -
logp.max()) / np.exp(logp -
logp.max()).sum() * logp.shape[0]
h2_mean = (np.exp(logp - logp.max()) *
grid[:, np.newaxis]).sum() / np.exp(logp - logp.max()).sum()
h2_var = (np.exp(logp - logp.max()) * (grid[:, np.newaxis] - h2_mean)**
2.0).sum() / np.exp(logp - logp.max()).sum()
if plot:
import pylab as plt
plt.figure()
plt.plot([h2['h2'], h2['h2']], [0, 1], "r")
plt.plot([h2_mean, h2_mean], [0, 1], "g")
plt.legend([
"REML-estimate = %.3f" % h2['h2'],
"posterior mean = %.3f" % h2_mean
])
plt.plot(grid.flatten(), post_h2.flatten())
plt.xlabel("$h^2$")
plt.ylabel("$p( h^2 | Data)$")
if verbose:
print("max[h^2] = %.5f E[h^2] = %.5f +- %.5f" %
(h2['h2'], h2_mean, np.sqrt(h2_var)))
return h2, h2_mean, h2_var, grid, post_h2
def assoc_scan(Y, X, covariates=None, K=None):
lmm = lmm_cov.LMM(X=covariates, Y=Y, G=X, K=K)
# opt = lmm.find_log_delta(X.shape[1], nGridH2=100, REML=False)
opt = lmm.findH2(nGridH2=100)
h2 = opt['h2']
res = lmm.nLLeval(h2=h2,
delta=None,
dof=None,
scale=1.0,
penalty=0.0,
snps=X)
chi2stats = res['beta'] * res['beta'] / res['variance_beta']
p_values = st.chi2.sf(chi2stats, 1)[:, 0]
return p_values, h2