Skip to content
Snippets Groups Projects
Commit a3328a63 authored by Tim O'Donnell's avatar Tim O'Donnell
Browse files

Better random negative peptides generation

parent 2bd4c3af
No related branches found
No related tags found
No related merge requests found
Hide whitespace changes
Inline Side-by-side
mhcflurry/class1_neural_network.py
+ 157
− 60
View file @ a3328a63
...@@ -14,6 +14,7 @@ import pandas ...@@ -14,6 +14,7 @@ import pandas
from .hyperparameters import HyperparameterDefaults from .hyperparameters import HyperparameterDefaults
from .encodable_sequences import EncodableSequences, EncodingError from .encodable_sequences import EncodableSequences, EncodingError
from .allele_encoding import AlleleEncoding
from .regression_target import to_ic50, from_ic50 from .regression_target import to_ic50, from_ic50
from .common import random_peptides, amino_acid_distribution from .common import random_peptides, amino_acid_distribution
from .custom_loss import get_loss from .custom_loss import get_loss
...@@ -97,7 +98,10 @@ class Class1NeuralNetwork(object): ...@@ -97,7 +98,10 @@ class Class1NeuralNetwork(object):
random_negative_affinity_max=50000.0, random_negative_affinity_max=50000.0,
random_negative_match_distribution=True, random_negative_match_distribution=True,
random_negative_distribution_smoothing=0.0, random_negative_distribution_smoothing=0.0,
random_negative_output_indices=None) random_negative_output_indices=None,
random_negative_method="by_length",
random_negative_binder_threshold=None,
random_negative_lengths=[8,9,10,11,12,13,14,15])
""" """
Hyperparameters for neural network training. Hyperparameters for neural network training.
""" """
...@@ -674,6 +678,132 @@ class Class1NeuralNetwork(object): ...@@ -674,6 +678,132 @@ class Class1NeuralNetwork(object):
fit_info["num_points"] = mutable_generator_state["yielded_values"] fit_info["num_points"] = mutable_generator_state["yielded_values"]
self.fit_info.append(dict(fit_info)) self.fit_info.append(dict(fit_info))
def random_negatives_generator(
self,
encodable_peptides,
affinities,
allele_encoding,
inequalities):
random_negative_lengths = self.hyperparameters['random_negative_lengths']
df = pandas.DataFrame({
"peptide": encodable_peptides.sequences,
"affinity": affinities,
})
if allele_encoding is not None:
df["allele"] = allele_encoding.alleles
df["length"] = df.peptide.str.len()
if inequalities is None:
df["inequality"] = "="
else:
df["inequality"] = inequalities
if self.hyperparameters['random_negative_binder_threshold']:
df = df.loc[
(df.inequality != ">") &
(df.affinity < self.hyperparameters[
'random_negative_binder_threshold'
])
]
aa_distribution = None
if self.hyperparameters['random_negative_match_distribution']:
aa_distribution = amino_acid_distribution(
encodable_peptides.sequences,
smoothing=self.hyperparameters[
'random_negative_distribution_smoothing'
])
logging.info(
"Using amino acid distribution for random negative:\n%s" % (
str(aa_distribution.to_dict())))
random_negative_alleles = None
if self.hyperparameters["random_negative_method"] == "by_length":
# Different numbers of random negatives per length. Alleles are
# sampled proportionally to the number of times they are used in
# the training data.
length_to_num_random_negative = {}
random_negative_lengths = self.hyperparameters[
'random_negative_lengths'
]
length_counts = df.length.value_counts().to_dict()
for length in random_negative_lengths:
length_to_num_random_negative[length] = int(
length_counts.get(length, 0) *
self.hyperparameters['random_negative_rate'] +
self.hyperparameters['random_negative_constant'])
length_to_num_random_negative = pandas.Series(
length_to_num_random_negative)
total_random_negatives = length_to_num_random_negative.sum()
logging.info("Random negative counts per length:\n%s" % (
str(length_to_num_random_negative.to_dict())))
if allele_encoding is not None:
random_negative_alleles = df.allele.sample(
n=total_random_negatives, replace=True).values
def sample_peptides():
peptides = []
for (length, count) in length_to_num_random_negative.items():
peptides.extend(
random_peptides(
count,
length=length,
distribution=aa_distribution))
random.shuffle(peptides) # important
return EncodableSequences.create(peptides)
elif self.hyperparameters["random_negative_method"] == "by_allele":
# For each allele, a particular number of random negatives are used
# for all lengths. Across alleles, the number of random negatives
# varies; within an allele, the number of random negatives for each
# length is a constant
allele_to_num_per_length = {}
total_random_peptides_per_length = 0
for (allele, sub_df) in df.groupby("allele"):
num_for_allele = len(sub_df) * (
self.hyperparameters['random_negative_rate']
) + self.hyperparameters['random_negative_constant']
num_per_length = int(
num_for_allele / len(random_negative_lengths))
total_random_peptides_per_length += num_per_length
allele_to_num_per_length[allele] = num_per_length
for _ in random_negative_lengths:
for (allele, num) in allele_to_num_per_length.items():
random_negative_alleles.append([allele] * num)
numpy.testing.assert_equal(
len(random_negative_alleles),
total_random_peptides_per_length * len(random_negative_lengths))
logging.info(
"Random negative counts for each length by allele:\n%s" % (
str(allele_to_num_per_length)))
def sample_peptides():
peptides = []
for length in random_negative_lengths:
peptides.extend(
random_peptides(
total_random_peptides_per_length,
length=length,
distribution=aa_distribution))
# important NOT to shuffle peptides.
return EncodableSequences.create(peptides)
else:
raise NotImplementedError(
self.hyperparameters["random_negative_method"])
random_negative_allele_encoding = None
if random_negative_alleles:
random_negative_allele_encoding = AlleleEncoding(
random_negative_alleles, borrow_from=allele_encoding)
yield random_negative_allele_encoding
while True:
yield sample_peptides()
def fit( def fit(
self, self,
peptides, peptides,
...@@ -738,29 +868,14 @@ class Class1NeuralNetwork(object): ...@@ -738,29 +868,14 @@ class Class1NeuralNetwork(object):
peptide_encoding = self.peptides_to_network_input(encodable_peptides) peptide_encoding = self.peptides_to_network_input(encodable_peptides)
fit_info = collections.defaultdict(list) fit_info = collections.defaultdict(list)
length_counts = ( random_negatives_generator = self.random_negatives_generator(
pandas.Series(encodable_peptides.sequences) encodable_peptides=encodable_peptides,
.str.len().value_counts().to_dict()) affinities=affinities,
allele_encoding=allele_encoding,
num_random_negative = {} inequalities=inequalities)
for length in range(8, 16): random_negatives_allele_encoding = next(random_negatives_generator)
num_random_negative[length] = int( num_random_negatives = len(
length_counts.get(length, 0) * next(random_negatives_generator).sequences)
self.hyperparameters['random_negative_rate'] +
self.hyperparameters['random_negative_constant'])
num_random_negative = pandas.Series(num_random_negative)
logging.info("Random negative counts per length:\n%s" % (
str(num_random_negative.to_dict())))
aa_distribution = None
if self.hyperparameters['random_negative_match_distribution']:
aa_distribution = amino_acid_distribution(
encodable_peptides.sequences,
smoothing=self.hyperparameters[
'random_negative_distribution_smoothing'])
logging.info(
"Using amino acid distribution for random negative:\n%s" % (
str(aa_distribution.to_dict())))
y_values = from_ic50(numpy.array(affinities, copy=False)) y_values = from_ic50(numpy.array(affinities, copy=False))
assert numpy.isnan(y_values).sum() == 0, y_values assert numpy.isnan(y_values).sum() == 0, y_values
...@@ -853,15 +968,14 @@ class Class1NeuralNetwork(object): ...@@ -853,15 +968,14 @@ class Class1NeuralNetwork(object):
y_dict_with_random_negatives = { y_dict_with_random_negatives = {
"output": numpy.concatenate([ "output": numpy.concatenate([
numpy.tile( numpy.tile(
random_negative_target, int(num_random_negative.sum())), random_negative_target, num_random_negatives),
y_values, y_values,
]), ]),
} }
# Note: we are using "<" here not ">" because the inequalities are # Note: we are using "<" here not ">" because the inequalities are
# now in target-space (0-1) not affinity-space. # now in target-space (0-1) not affinity-space.
adjusted_inequalities_with_random_negatives = ( adjusted_inequalities_with_random_negatives = (
["<"] * int(num_random_negative.sum()) + ["<"] * num_random_negatives + list(adjusted_inequalities))
list(adjusted_inequalities))
else: else:
# Randomly sample random negative affinities # Randomly sample random negative affinities
y_dict_with_random_negatives = { y_dict_with_random_negatives = {
...@@ -872,7 +986,7 @@ class Class1NeuralNetwork(object): ...@@ -872,7 +986,7 @@ class Class1NeuralNetwork(object):
'random_negative_affinity_min'], 'random_negative_affinity_min'],
self.hyperparameters[ self.hyperparameters[
'random_negative_affinity_max'], 'random_negative_affinity_max'],
int(num_random_negative.sum()))), num_random_negatives)),
y_values, y_values,
]), ]),
} }
...@@ -881,7 +995,7 @@ class Class1NeuralNetwork(object): ...@@ -881,7 +995,7 @@ class Class1NeuralNetwork(object):
y_dict_with_random_negatives) y_dict_with_random_negatives)
if sample_weights is not None: if sample_weights is not None:
sample_weights_with_random_negatives = numpy.concatenate([ sample_weights_with_random_negatives = numpy.concatenate([
numpy.ones(int(num_random_negative.sum())), numpy.ones(num_random_negatives),
sample_weights]) sample_weights])
else: else:
sample_weights_with_random_negatives = None sample_weights_with_random_negatives = None
...@@ -893,7 +1007,7 @@ class Class1NeuralNetwork(object): ...@@ -893,7 +1007,7 @@ class Class1NeuralNetwork(object):
else list(range(0, self.hyperparameters['num_outputs']))) else list(range(0, self.hyperparameters['num_outputs'])))
output_indices_with_random_negatives = numpy.concatenate([ output_indices_with_random_negatives = numpy.concatenate([
pandas.Series(random_negative_output_indices, dtype=int).sample( pandas.Series(random_negative_output_indices, dtype=int).sample(
n=int(num_random_negative.sum()), replace=True).values, n=num_random_negatives, replace=True).values,
output_indices output_indices
]) ])
else: else:
...@@ -924,32 +1038,24 @@ class Class1NeuralNetwork(object): ...@@ -924,32 +1038,24 @@ class Class1NeuralNetwork(object):
last_progress_print = None last_progress_print = None
x_dict_with_random_negatives = {} x_dict_with_random_negatives = {}
for i in range(self.hyperparameters['max_epochs']): for i in range(self.hyperparameters['max_epochs']):
random_negative_peptides_list = [] random_negative_peptides = next(random_negatives_generator)
for (length, count) in num_random_negative.iteritems():
random_negative_peptides_list.extend(
random_peptides(
count,
length=length,
distribution=aa_distribution))
random.shuffle(random_negative_peptides_list)
random_negative_peptides = EncodableSequences.create(
random_negative_peptides_list)
random_negative_peptides_encoding = ( random_negative_peptides_encoding = (
self.peptides_to_network_input(random_negative_peptides)) self.peptides_to_network_input(random_negative_peptides))
if not x_dict_with_random_negatives: if not x_dict_with_random_negatives:
if len(random_negative_peptides) > 0: if len(random_negative_peptides) > 0:
x_dict_with_random_negatives["peptide"] = numpy.concatenate([ x_dict_with_random_negatives[
"peptide"
] = numpy.concatenate([
random_negative_peptides_encoding, random_negative_peptides_encoding,
peptide_encoding, x_dict_without_random_negatives['peptide'],
]) ])
if 'allele' in x_dict_without_random_negatives: if 'allele' in x_dict_without_random_negatives:
x_dict_with_random_negatives['allele'] = numpy.concatenate([ x_dict_with_random_negatives[
x_dict_without_random_negatives['allele'][ 'allele'
numpy.random.choice( ] = numpy.concatenate([
x_dict_without_random_negatives[ self.allele_encoding_to_network_input(
'allele'].shape[0], random_negatives_allele_encoding)[0],
size=len(random_negative_peptides_list))],
x_dict_without_random_negatives['allele'] x_dict_without_random_negatives['allele']
]) ])
else: else:
...@@ -959,18 +1065,9 @@ class Class1NeuralNetwork(object): ...@@ -959,18 +1065,9 @@ class Class1NeuralNetwork(object):
# Update x_dict_with_random_negatives in place. # Update x_dict_with_random_negatives in place.
# This is more memory efficient than recreating it as above. # This is more memory efficient than recreating it as above.
if len(random_negative_peptides) > 0: if len(random_negative_peptides) > 0:
x_dict_with_random_negatives["peptide"][:len(random_negative_peptides)] = ( x_dict_with_random_negatives[
random_negative_peptides_encoding "peptide"
) ][:num_random_negatives] = random_negative_peptides_encoding
if 'allele' in x_dict_with_random_negatives:
x_dict_with_random_negatives['allele'][:len(random_negative_peptides)] = (
x_dict_with_random_negatives['allele'][
len(random_negative_peptides) + numpy.random.choice(
x_dict_with_random_negatives['allele'].shape[0] -
len(random_negative_peptides),
size=len(random_negative_peptides))
]
)
if needs_initialization: if needs_initialization:
self.data_dependent_weights_initialization( self.data_dependent_weights_initialization(
......
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment