def run_linear_regression(genotype_df,
phenotype_df,
covariate_df,
add_intercept=True):
phenotype_names = phenotype_df.columns.astype('str').to_series()
C = covariate_df.to_numpy('float64', copy=True)
if add_intercept:
C = gwas_fx._add_intercept(C, genotype_df.shape[0])
if not C.size:
C = np.zeros((genotype_df.shape[0], 1))
Y = phenotype_df.to_numpy('float64', copy=True)
Y_mask = ~np.isnan(Y)
Y[~Y_mask] = 0
Y -= Y.mean(axis=0)
Q = np.linalg.qr(C)[0]
Y = gwas_fx._residualize_in_place(Y, Q) * Y_mask
Y_scale = np.ones(Y.shape[1])
Y_state = lr._create_YState(Y, phenotype_df, pd.DataFrame({}), Y_mask,
np.float64, None)
dof = C.shape[0] - C.shape[1] - 1
pdf = pd.DataFrame(
{lr._VALUES_COLUMN_NAME: list(genotype_df.to_numpy('float64').T)})
return lr._linear_regression_inner(pdf, Y_state, Y_mask.astype('float64'),
Y_scale, Q, dof, phenotype_names, None,
None)
def run_score_test(genotype_df,
phenotype_df,
covariate_df,
correction=lr.correction_none,
add_intercept=True):
C = covariate_df.to_numpy(copy=True)
if add_intercept:
C = gwas_fx._add_intercept(C, phenotype_df.shape[0])
Y = phenotype_df.to_numpy(copy=True)
Y_mask = ~np.isnan(Y)
Y[~Y_mask] = 0
state_rows = [
lr._prepare_one_phenotype(
C, pd.Series({
'label': p,
'values': phenotype_df[p]
}), correction, add_intercept) for p in phenotype_df
]
phenotype_names = phenotype_df.columns.to_series().astype('str')
state = lr._pdf_to_log_reg_state(pd.DataFrame(state_rows), phenotype_names,
C.shape[1])
values_df = pd.DataFrame(
{gwas_fx._VALUES_COLUMN_NAME: list(genotype_df.to_numpy().T)})
return lr._logistic_regression_inner(
values_df, state, C, Y, Y_mask, None, lr.correction_none, 0.05,
phenotype_df.columns.to_series().astype('str'))
def statsmodels_baseline(genotype_df,
phenotype_df,
covariate_df,
offset_dfs=None,
add_intercept=True):
# Project out covariates from genotypes and phenotypes
C = covariate_df.to_numpy('float64')
num_samples = C.shape[0] if C.size else genotype_df.shape[0]
if add_intercept:
C = gwas_fx._add_intercept(C, num_samples)
Y = phenotype_df.to_numpy('float64')
X = genotype_df.to_numpy('float64')
phenotype_df.columns = phenotype_df.columns.astype('str')
dof = C.shape[0] - C.shape[1] - 1
effects = []
errors = []
tvalues = []
pvalues = []
for phenotype_idx in range(Y.shape[1]):
for genotype_idx in range(X.shape[1]):
phenotype = Y[:, phenotype_idx].copy()
phenotype_mask = ~np.isnan(phenotype)
phenotype[~phenotype_mask] = 0
phenotype -= phenotype.mean()
phenotype = residualize(phenotype, C) * phenotype_mask
phenotype_scale = np.sqrt(
(phenotype**2).sum() / (phenotype_mask.sum() - C.shape[1]))
phenotype /= phenotype_scale
if offset_dfs:
offset = offset_dfs[genotype_idx].iloc[:,
phenotype_idx].to_numpy(
'float64')
phenotype = phenotype - offset
phenotype[~phenotype_mask] = np.nan
genotype = residualize(X[:, genotype_idx], C)
genotype = pd.Series(genotype, name='genotype')
model = sm.OLS(phenotype, genotype, missing='drop')
model.df_resid = dof
results = model.fit()
effects.append(results.params.genotype * phenotype_scale)
errors.append(results.bse.genotype * phenotype_scale)
tvalues.append(results.tvalues.genotype)
pvalues.append(results.pvalues.genotype)
return pd.DataFrame({
'effect':
effects,
'standard_error':
errors,
'tvalue':
tvalues,
'pvalue':
pvalues,
'phenotype':
phenotype_df.columns.to_series().repeat(genotype_df.shape[1])
})