def test_init(self, gpm):
model = EpistasisLinearRegression(order=self.order, model_type="local")
model.add_gpm(gpm)
# Checks
check1 = model.order
check2 = model.model_type
assert check1 == self.order
assert check2 == "local"
def _fit_additive(self, X=None, y=None):
# Construct an additive model.
self.Additive = EpistasisLinearRegression(
order=1, model_type=self.model_type)
self.Additive.add_gpm(self.gpm)
# Prepare a high-order model
self.Additive.epistasis = EpistasisMap(
sites=self.Additive.Xcolumns,
)
# Fit the additive model and infer additive phenotypes
self.Additive.fit(X=X, y=y)
return self
def test_lnlikelihood(self, gpm):
model = EpistasisLinearRegression(order=self.order, model_type="local")
model.add_gpm(gpm)
model.fit()
# Calculate lnlikelihood
lnlike = model.lnlikelihood()
assert lnlike.dtype == float
def test_score(self, gpm):
model = EpistasisLinearRegression(order=self.order, model_type="local")
model.add_gpm(gpm)
model.fit()
score = model.score()
# Tests
assert score >= 0
assert score <= 1
def test_hypothesis(self, gpm):
model = EpistasisLinearRegression(order=self.order, model_type="local")
model.add_gpm(gpm)
model.fit()
# Checks
check1 = model.hypothesis(thetas=model.coef_)
# Tests
np.testing.assert_almost_equal(sorted(check1),
sorted(model.gpm.phenotype))
def test_predict(self, gpm):
model = EpistasisLinearRegression(order=self.order, model_type="local")
model.add_gpm(gpm)
model.fit()
check1 = model.predict()
# Tests
np.testing.assert_almost_equal(sorted(check1),
sorted(model.gpm.phenotype))
def __init__(self, function, model_type="global", **p0):
# Set up the function for fitting.
self.function = function
self.minimizer = FunctionMinimizer(self.function, **p0)
self.parameters = self.minimizer.parameters
self.order = 1
self.model_type = model_type
# Necessary for sklearn 0.24 +
self.positive = False
# Store model specs.
self.model_specs = dict(function=self.function,
model_type=self.model_type,
**p0)
# Set up additive and high-order linear model
self.Additive = EpistasisLinearRegression(order=1,
model_type=self.model_type)
def __init__(self,
function,
model_type="global",
**p0):
# Set up the function for fitting.
self.function = function
self.minimizer = FunctionMinimizer(self.function, **p0)
self.parameters = self.minimizer.parameters
self.order = 1
self.Xbuilt = {}
# Construct parameters object
self.set_params(model_type=model_type)
# Store model specs.
self.model_specs = dict(
function=self.function,
model_type=self.model_type,
**p0)
# Set up additive and high-order linear model
self.Additive = EpistasisLinearRegression(
order=1, model_type=self.model_type)
def __init__(self, order=1, threshold=5, model_type='global', **kwargs):
self.model_type = model_type
self.order = 1
self.threshold = threshold
super(self.__class__, self).__init__(**kwargs)
# Store model specs.
self.model_specs = dict(priors=None,
threshold=threshold,
model_type=self.model_type,
**kwargs)
# Set up additive linear model for pre-classifying
self.Additive = EpistasisLinearRegression(order=1,
model_type=self.model_type)
def test_fit(self, gpm):
model = EpistasisLinearRegression(order=self.order, model_type="local")
model.add_gpm(gpm)
model.fit()
# Checks
check2 = hasattr(model, "coef_")
check3 = hasattr(model, "epistasis")
# Tests
assert check2 is True
assert check3 is True
class EpistasisClassifierMixin:
"""A Mixin class for epistasis classifiers
"""
def _fit_additive(self, X=None, y=None):
# Construct an additive model.
self.Additive = EpistasisLinearRegression(
order=1, model_type=self.model_type)
self.Additive.add_gpm(self.gpm)
# Prepare a high-order model
self.Additive.epistasis = EpistasisMap(
sites=self.Additive.Xcolumns,
)
# Fit the additive model and infer additive phenotypes
self.Additive.fit(X=X, y=y)
return self
def _fit_classifier(self, X=None, y=None):
# This method builds x and y from data.
add_coefs = self.Additive.epistasis.values
add_X = self.Additive._X(data=X)
# Project X into padd space.
X = add_X * add_coefs
# Label X.
y = binarize(y.reshape(1, -1), self.threshold)[0]
self.classes = y
# Fit classifier.
super().fit(X=X, y=y)
return self
def fit_transform(self, X=None, y=None, **kwargs):
self.fit(X=X, y=y, **kwargs)
ypred = self.predict(X=X)
# Transform map.
gpm = GenotypePhenotypeMap.read_dataframe(
dataframe=self.gpm.data[ypred==1],
wildtype=self.gpm.wildtype,
mutations=self.gpm.mutations
)
return gpm
def predict(self, X=None):
Xadd = self.Additive._X(data=X)
X = Xadd * self.Additive.epistasis.values
return super().predict(X=X)
def predict_transform(self, X=None, y=None):
x = self.predict(X=X)
y[x <= 0.5] = self.threshold
return y
def predict_log_proba(self, X=None):
Xadd = self.Additive._X(data=X)
X = Xadd * self.Additive.epistasis.values
return super().predict_log_proba(X)
def predict_proba(self, X=None):
Xadd = self.Additive._X(data=X)
X = Xadd * self.Additive.epistasis.values
return super().predict_proba(X=X)
class EpistasisNonlinearRegression(BaseModel):
"""Use nonlinear least-squares regression to estimate epistatic coefficients
and nonlinear scale in a nonlinear genotype-phenotype map.
This models has two steps:
1. Fit an additive, linear regression to approximate the average effect
of individual mutations.
2. Fit the nonlinear function to the observed phenotypes vs. the
additive phenotypes estimated in step 1.
Methods are described in the following publication:
Sailer, Z. R. & Harms, M. J. 'Detecting High-Order Epistasis in
Nonlinear Genotype-Phenotype Maps'. Genetics 205, 1079-1088 (2017).
Parameters
----------
function : callable
Nonlinear function between Pobs and Padd
model_type : str (default: global)
type of epistasis model to use. See paper above for more information.
Keyword Arguments
-----------------
Keyword arguments are interpreted as intial guesses for the nonlinear
function parameters. Must have the same name as parameters in the
nonlinear function.
Attributes
----------
Additive : EpistasisLinearRegression
Linear regression object for fitting additive model
parameters : Parameters object
Mapping object for nonlinear coefficients
minimizer :
Object that fits data using the function and a least squares minimization.
"""
def __init__(self,
function,
model_type="global",
**p0):
# Set up the function for fitting.
self.function = function
self.minimizer = FunctionMinimizer(self.function, **p0)
self.parameters = self.minimizer.parameters
self.order = 1
self.Xbuilt = {}
# Construct parameters object
self.set_params(model_type=model_type)
# Store model specs.
self.model_specs = dict(
function=self.function,
model_type=self.model_type,
**p0)
# Set up additive and high-order linear model
self.Additive = EpistasisLinearRegression(
order=1, model_type=self.model_type)
def add_gpm(self, gpm):
super(EpistasisNonlinearRegression, self).add_gpm(gpm)
# Add gpm to other models.
self.Additive.add_gpm(gpm)
return self
@property
def thetas(self):
return np.concatenate((list(self.parameters.values()),
self.Additive.coef_))
@property
def num_of_params(self):
n = 0
n += len(self.parameters) + len(self.Additive.coef_)
return n
@arghandler
def transform(self, X=None, y=None):
# Use a first order matrix only.
if type(X) == np.ndarray or type(X) == pd.DataFrame:
Xadd = X[:, :self.Additive.epistasis.n]
else:
Xadd = X
# Predict additive model.
x = self.Additive.predict(X=Xadd)
# Transform y onto x scale
return self.minimizer.transform(x, y)
def fit(self,
X=None,
y=None,
**kwargs):
# Fit linear portion
self._fit_additive(X=X, y=y)
# Step 2: fit nonlinear function
self._fit_nonlinear(X=X, y=y, **kwargs)
return self
def _fit_additive(self, X=None, y=None, **kwargs):
# Fit with an additive model
self.Additive.epistasis = EpistasisMap(
sites=self.Additive.Xcolumns,
)
# Use a first order matrix only.
if type(X) == np.ndarray or type(X) == pd.DataFrame:
Xadd = X[:, :self.Additive.epistasis.n]
else:
Xadd = X
# Fit Additive model
self.Additive.fit(X=Xadd, y=y)
self.Additive.epistasis.values = self.Additive.coef_
return self
@arghandler
def _fit_nonlinear(self, X=None, y=None, **kwargs):
"""Estimate the scale of multiple mutations in a genotype-phenotype
map."""
# Use a first order matrix only.
if type(X) == np.ndarray or type(X) == pd.DataFrame:
Xadd = X[:, :self.Additive.epistasis.n]
else:
Xadd = X
# Predict additive phenotypes.
x = self.Additive.predict(X='fit')
# Fit function
self.minimizer.fit(x, y)
self.parameters = self.minimizer.parameters
@arghandler
def fit_transform(self, X=None, y=None, **kwargs):
self.fit(X=X, y=y, **kwargs)
linear_phenotypes = self.transform(X=X, y=y)
# Transform map.
gpm = GenotypePhenotypeMap.read_dataframe(
dataframe=self.gpm.data,
wildtype=self.gpm.wildtype,
mutations=self.gpm.mutations
)
gpm.data['phenotypes'] = linear_phenotypes
return gpm
def predict(self, X=None):
x = self.Additive.predict(X=X)
y = self.minimizer.predict(x)
return y
def predict_transform(self, X=None, y=None):
if y is None:
x = self.Additive.predict(X=X)
else:
x = y
return self.minimizer.predict(x)
@arghandler
def hypothesis(self, X=None, thetas=None):
# ----------------------------------------------------------------------
# Part 0: Break up thetas
# ----------------------------------------------------------------------
i, j = len(self.parameters.valuesdict()), self.Additive.epistasis.n
parameters = thetas[:i]
epistasis = thetas[i:i + j]
# Part 1: Linear portion
x = np.dot(X, epistasis)
# Part 2: Nonlinear portion
ynonlin = self.minimizer.function(x, *parameters)
return ynonlin
def hypothesis_transform(self, X=None, y=None, thetas=None):
# Break up thetas
i, j = len(self.parameters.valuesdict()), self.Additive.epistasis.n
parameters = thetas[:i]
epistasis = thetas[i:i + j]
if y is None:
x = self.Additive.hypothesis(X=X, thetas=epistasis)
else:
x = y
y_transform = self.minimizer.function(x, *parameters)
return y_transform
@arghandler
def score(self, X=None, y=None):
x = self.Additive.predict(X=X)
ypred = self.minimizer.predict(x)
return pearson(y, ypred)**2
@arghandler
def lnlike_of_data(self, X=None, y=None, yerr=None, thetas=None):
# Calculate ymodel
ymodel = self.hypothesis(X=X, thetas=thetas)
# Likelihood of data given model
return (- 0.5 * np.log(2 * np.pi * yerr**2) -
(0.5 * ((y - ymodel)**2 / yerr**2)))
@arghandler
def lnlike_transform(
self,
X=None,
y=None,
yerr=None,
lnprior=None,
thetas=None):
# Update likelihood.
lnlike = self.lnlike_of_data(X=X, y=y, yerr=yerr, thetas=thetas)
return lnlike + lnprior