class Base(PresencePipe):
"""Linear regression classifiers with elastic net regularization.
The `Base` model takes a symmetric sample over `l1_ratio` regardless of
the performance of the classifier at each value of this parameter in order
for us to gain a better understanding of how mixing the LASSO and ridge
regularization penalties affects classifier behaviour.
Models that implement tuning grids that favour better-performing values of
`l1_ratio` are implemented below.
"""
tune_priors = (
('fit__alpha', tuple(10**np.linspace(-4, -1 / 3, 12))),
('fit__l1_ratio', (0.25, 0.5, 0.75)),
)
feat_inst = SelectMeanVar(mean_perc=95, var_perc=95)
norm_inst = StandardScaler()
fit_inst = SGDClassifier(loss='log',
penalty='elasticnet',
max_iter=1000,
class_weight='balanced')
def __init__(self):
super().__init__([('feat', self.feat_inst), ('norm', self.norm_inst),
('fit', self.fit_inst)])
class SVCrbf(Base, Kernel):
feat_inst = SelectMeanVar(mean_perc=90, var_perc=100)
tune_priors = (
('fit__C', tuple(10 ** np.linspace(-3, 4, 8))),
)
test_count = 8
fit_inst = SVC(kernel='rbf', gamma='scale', probability=True,
cache_size=500, class_weight='balanced')
class Forests(Base, Trees):
feat_inst = SelectMeanVar(mean_perc=90, var_perc=100)
tune_priors = (
('fit__min_samples_leaf', (1, 2, 3, 4, 6, 8, 10, 15)),
)
test_count = 8
fit_inst = RandomForestClassifier(n_estimators=5000,
class_weight='balanced')
class Ridge(Base, LinearPipe):
feat_inst = SelectMeanVar(mean_perc=90, var_perc=100)
tune_priors = (
('fit__C', tuple(10 ** np.linspace(-7, 0, 8))),
)
test_count = 8
fit_inst = LogisticRegression(solver='liblinear', penalty='l2',
max_iter=200, class_weight='balanced')
class StanPipe(PresencePipe):
tune_priors = (('fit__alpha', tuple(10**np.linspace(-3, 0.68, 24))), )
feat_inst = SelectMeanVar(mean_perc=75)
norm_inst = StandardScaler()
fit_inst = OptimModel(model_code=gauss_model)
def __init__(self):
super().__init__([('feat', self.feat_inst), ('norm', self.norm_inst),
('fit', self.fit_inst)])
class Base(MultiPipe, PresencePipe):
tune_priors = (
('fit__margin', (0.4, 0.6, 0.8, 1.0, 1.2, 1.4)),
('fit__sigma_h', (0.04, 0.08, 0.1, 0.12, 0.16, 0.24)),
)
feat_inst = SelectMeanVar(mean_perc=95, var_perc=95)
norm_inst = StandardScaler()
fit_inst = SingleDomain(latent_features=5, max_iter=500, stop_tol=0.05)
def __init__(self):
super().__init__([('feat', self.feat_inst), ('norm', self.norm_inst),
('fit', self.fit_inst)])
class Base(PresencePipe):
tune_priors = (
('fit__max_depth', (2, 3, 4, 5)),
('fit__min_samples_split', tuple(np.linspace(0.003, 0.051, 9))),
)
feat_inst = SelectMeanVar(mean_perc=95, var_perc=95)
norm_inst = StandardScaler()
fit_inst = GradientBoostingClassifier()
def __init__(self):
super().__init__([('feat', self.feat_inst), ('norm', self.norm_inst),
('fit', self.fit_inst)])
class Base(PresencePipe):
tune_priors = (
('fit__alpha', tuple(10**np.linspace(-4, -2.35, 12))),
('fit__gamma', (0.5, 1., 2.)),
)
feat_inst = SelectMeanVar(mean_perc=95, var_perc=95)
norm_inst = StandardScaler()
fit_inst = OptimModel(model_code=base_model)
def __init__(self):
super().__init__([('feat', self.feat_inst), ('norm', self.norm_inst),
('fit', self.fit_inst)])
class MultiTransfer(TransferPipe, PresencePipe):
tune_priors = (
('fit__margin', (2. / 3, 24. / 23)),
('fit__sigma_h', (1. / 11, 1. / 7)),
)
feat_inst = SelectMeanVar(mean_perc=80, var_perc=90)
norm_inst = StandardScaler()
fit_inst = MultiDomain(latent_features=3, max_iter=50)
def __init__(self):
super().__init__([('feat', self.feat_inst), ('norm', self.norm_inst),
('fit', self.fit_inst)])
class Base(PresencePipe):
tune_priors = (
('fit__max_features', tuple(10**np.linspace(-3, -2 / 3, 6))),
('fit__min_samples_leaf', (1, 2, 3, 4, 6, 8)),
)
feat_inst = SelectMeanVar(mean_perc=95, var_perc=95)
norm_inst = StandardScaler()
fit_inst = RandomForestClassifier(n_estimators=500,
class_weight='balanced')
def __init__(self):
super().__init__([('feat', self.feat_inst), ('norm', self.norm_inst),
('fit', self.fit_inst)])
class Base(OmicPipe):
"""An abstract class for the set of standard transformers.
The transformers in this module are designed to use all available -omic
features save those with very low expression in order to get the fullest
possible picture of the features that can be used to cluster a given task.
"""
feat_inst = SelectMeanVar(mean_perc=90, var_perc=100)
norm_inst = StandardScaler()
def __init__(self):
super().__init__([('feat', self.feat_inst), ('norm', self.norm_inst),
('fit', self.fit_inst)])
class Meanvar(Base):
"""Ridge regression with tuning over feature selection thresholds.
We fix a value of `C` that tends to work well across a wide variety of
mutation prediction tasks, and then tune over different filter cutoffs for
removing expression features with low mean or variance.
"""
tune_priors = (
('feat__mean_perc', (100. / 3, 50, 75, 90, 98, 100)),
('feat__var_perc', (100. / 3, 50, 75, 90, 98, 100)),
)
feat_inst = SelectMeanVar()
fit_inst = LogisticRegression(C=0.002,
penalty='l2',
class_weight='balanced')
class Meanvar(Base):
"""LASSO regression with tuning over feature selection thresholds.
We fix a value of `C` that tends to work well across a wide variety of
mutation prediction tasks, and then tune over different filter cutoffs for
removing expression features with low mean or variance.
"""
tune_priors = (
('feat__mean_perc', (50, 65, 75, 85, 90, 99)),
('feat__var_perc', (50, 65, 75, 85, 90, 99)),
)
feat_inst = SelectMeanVar()
fit_inst = LogisticRegression(penalty='l1',
max_iter=200,
C=np.exp(1),
class_weight='balanced')
class Base(PresencePipe):
"""Linear regression classifiers with the ridge regularization penalty.
Note that the `C` regularization strength parameter should have this same
testing value grid in all cases where it is tuned over. This reflects that
optimal values of `C` tend to always fall well within this selected range
for observed mutation prediction tasks, and also that past a certain point
all large values of `C` will result in no regularization.
"""
tune_priors = (('fit__C', tuple(10**np.linspace(-7.1, 3.4, 36))), )
feat_inst = SelectMeanVar(mean_perc=95, var_perc=95)
norm_inst = StandardScaler()
fit_inst = LogisticRegression(penalty='l2', class_weight='balanced')
def __init__(self):
super().__init__([('feat', self.feat_inst), ('norm', self.norm_inst),
('fit', self.fit_inst)])
class Meanvar(Base):
"""Elastic net regression with tuning over feature selection thresholds.
We fix values of `alpha` and `l1_ratio` that tend to work well across a
wide variety of mutation prediction tasks, and then tune over different
filter cutoffs for removing expression features with low mean or variance.
"""
tune_priors = (
('feat__mean_perc', (50, 65, 75, 85, 90, 99)),
('feat__var_perc', (50, 65, 75, 85, 90, 99)),
)
feat_inst = SelectMeanVar()
fit_inst = SGDClassifier(loss='log',
penalty='elasticnet',
max_iter=1000,
l1_ratio=0.5,
alpha=0.01,
class_weight='balanced')
class Base(PresencePipe):
"""Linear regression classifiers with the LASSO regularization penalty.
Note that the `C` regularization strength parameter should have this same
testing value grid in all cases where it is tuned over. The selected range
of values reflects the finding that setting `C` to less than 0.01 doesn't
appear to ever work in the context of predicting mutation status in any of
the variants of LASSO regression given below, and also that past a certain
point all large values of `C` will simply result in no regularization.
"""
tune_priors = (('fit__C', tuple(10**np.linspace(-4.25, 8, 36))), )
feat_inst = SelectMeanVar(mean_perc=95, var_perc=95)
norm_inst = StandardScaler()
fit_inst = LogisticRegression(penalty='l1',
max_iter=200,
class_weight='balanced')
def __init__(self):
super().__init__([('feat', self.feat_inst), ('norm', self.norm_inst),
('fit', self.fit_inst)])
class Base(PresencePipe):
"""Support Vector Classifiers with various kernels.
The `Base` model corresponds to the simplest linear kernel; other choices
of kernels are implemented below.
Note that a unique tuning grid for the `C` regularization parameter needs
to be specified in each version of this model due to the differences in
characteristics associated with each kernel.
"""
tune_priors = (('fit__C', tuple(10**np.linspace(-6.3, -2.8, 36))), )
feat_inst = SelectMeanVar(mean_perc=95, var_perc=95)
norm_inst = StandardScaler()
fit_inst = SVC(kernel='linear',
probability=True,
cache_size=500,
class_weight='balanced')
def __init__(self):
super().__init__([('feat', self.feat_inst), ('norm', self.norm_inst),
('fit', self.fit_inst)])
class Cauchy(Base):
feat_inst = SelectMeanVar(mean_perc=200. / 3, var_perc=200. / 3)
fit_inst = OptimModel(model_code=cauchy_model)
class Select_few(Base):
feat_inst = SelectMeanVar(mean_perc=200. / 3, var_perc=200. / 3)