From 9e89b4fae7dcb1c0d37a89dfeedee9397c8d682e Mon Sep 17 00:00:00 2001
From: Tim O'Donnell <timodonnell@gmail.com>
Date: Thu, 23 Jan 2020 18:13:53 -0500
Subject: [PATCH] working on presentation model
---
.travis.yml | 1 +
.../class1_presentation_neural_network.py | 668 +++++++-----------
mhcflurry/class1_presentation_predictor.py | 41 +-
mhcflurry/multiallelic_refinement_command.py | 2 +-
mhcflurry/random_negative_peptides.py | 2 +-
...test_class1_presentation_neural_network.py | 307 ++------
test/test_class1_presentation_predictor.py | 2 +-
7 files changed, 342 insertions(+), 681 deletions(-)
diff --git a/.travis.yml b/.travis.yml
index 58b32a53..b4b2f9f5 100644
--- a/.travis.yml
+++ b/.travis.yml
@@ -56,6 +56,7 @@ script:
data_mass_spec_annotated
models_class1
models_class1_pan
+ models_class1_pan_variants
allele_sequences
--already-downloaded-dir /tmp/downloads
- mhcflurry-downloads info # just to test this command works
diff --git a/mhcflurry/class1_presentation_neural_network.py b/mhcflurry/class1_presentation_neural_network.py
index 32ab07bf..b8e30958 100644
--- a/mhcflurry/class1_presentation_neural_network.py
+++ b/mhcflurry/class1_presentation_neural_network.py
@@ -8,29 +8,19 @@ import numpy
import pandas
import mhcnames
import hashlib
+from copy import copy
from .hyperparameters import HyperparameterDefaults
from .class1_neural_network import Class1NeuralNetwork, DEFAULT_PREDICT_BATCH_SIZE
+from .class1_cleavage_neural_network import Class1CleavageNeuralNetwork
from .encodable_sequences import EncodableSequences
-from .regression_target import from_ic50, to_ic50
-from .random_negative_peptides import RandomNegativePeptides
from .allele_encoding import MultipleAlleleEncoding, AlleleEncoding
from .auxiliary_input import AuxiliaryInputEncoder
-from .batch_generator import BatchGenerator
-from .custom_loss import (
- MSEWithInequalities,
- TransformPredictionsLossWrapper,
- MultiallelicMassSpecLoss)
+from .flanking_encoding import FlankingEncoding
class Class1PresentationNeuralNetwork(object):
network_hyperparameter_defaults = HyperparameterDefaults(
- allele_amino_acid_encoding="BLOSUM62",
- peptide_encoding={
- 'vector_encoding_name': 'BLOSUM62',
- 'alignment_method': 'left_pad_centered_right_pad',
- 'max_length': 15,
- },
max_alleles=6,
)
"""
@@ -39,11 +29,12 @@ class Class1PresentationNeuralNetwork(object):
"""
fit_hyperparameter_defaults = HyperparameterDefaults(
+ trainable_cleavage_predictor=False,
+ trainable_affinity_predictor=False,
max_epochs=500,
+ validation_split=0.1,
early_stopping=True,
- random_negative_affinity_min=20000.0,).extend(
- RandomNegativePeptides.hyperparameter_defaults).extend(
- BatchGenerator.hyperparameter_defaults
+ minibatch_size=256,
)
"""
Hyperparameters for neural network training.
@@ -58,8 +49,7 @@ class Class1PresentationNeuralNetwork(object):
"""
compile_hyperparameter_defaults = HyperparameterDefaults(
- loss_multiallelic_mass_spec_delta=0.2,
- loss_multiallelic_mass_spec_multiplier=1.0,
+ loss="binary_crossentropy",
optimizer="rmsprop",
learning_rate=None,
)
@@ -71,6 +61,7 @@ class Class1PresentationNeuralNetwork(object):
auxiliary_input_hyperparameter_defaults = HyperparameterDefaults(
auxiliary_input_features=["gene"],
auxiliary_input_feature_parameters={},
+ include_cleavage=True,
)
"""
Allele feature hyperparameters.
@@ -88,54 +79,48 @@ class Class1PresentationNeuralNetwork(object):
self.network = None
self.fit_info = []
self.allele_representation_hash = None
+ self.affinity_model = None
+ self.cleavage_model = None
- def load_from_class1_neural_network(self, model):
+ def build(self, affinity_model, cleavage_model=None):
import keras.backend as K
- from keras.layers import (
- Input,
- TimeDistributed,
- Dense,
- Flatten,
- RepeatVector,
- Reshape,
- concatenate,
- Activation,
- Lambda,
- Add,
- Multiply,
- Embedding)
- from keras.models import Model
- from keras.initializers import Zeros
-
- assert isinstance(model, Class1NeuralNetwork), model
- affinity_network = model.network()
+ import keras.models
+
+ assert isinstance(affinity_model, Class1NeuralNetwork), affinity_model
+ affinity_model = copy(affinity_model)
+
+ self.affinity_model = affinity_model
+ affinity_network = affinity_model.network()
+
+ model_inputs = {}
peptide_shape = tuple(
int(x) for x in K.int_shape(affinity_network.inputs[0])[1:])
- input_alleles = Input(
- shape=(self.hyperparameters['max_alleles'],), name="allele")
- input_peptides = Input(
+ model_inputs['allele_set'] = keras.layers.Input(
+ shape=(self.hyperparameters['max_alleles'],), name="allele_set")
+ model_inputs['peptide'] = keras.layers.Input(
shape=peptide_shape,
dtype='float32',
name='peptide')
- peptides_flattened = Flatten()(input_peptides)
- peptides_repeated = RepeatVector(self.hyperparameters['max_alleles'])(
+ peptides_flattened = keras.layers.Flatten()(model_inputs['peptide'])
+ peptides_repeated = keras.layers.RepeatVector(
+ self.hyperparameters['max_alleles'])(
peptides_flattened)
- allele_representation = Embedding(
+ allele_representation = keras.layers.Embedding(
name="allele_representation",
input_dim=64, # arbitrary, how many alleles to have room for
output_dim=affinity_network.get_layer(
"allele_representation").output_shape[-1],
input_length=self.hyperparameters['max_alleles'],
trainable=False,
- mask_zero=False)(input_alleles)
+ mask_zero=False)(model_inputs['allele_set'])
allele_flat = allele_representation
- allele_peptide_merged = concatenate(
+ allele_peptide_merged = keras.layers.concatenate(
[peptides_repeated, allele_flat], name="allele_peptide_merged")
layer_names = [
@@ -159,56 +144,61 @@ class Class1PresentationNeuralNetwork(object):
"allele_peptide_merged") + 1:
]
node = allele_peptide_merged
- layer_name_to_new_node = {
+ affinity_predictor_layer_name_to_new_node = {
"allele_peptide_merged": allele_peptide_merged,
}
for layer in layers:
- assert layer.name not in layer_name_to_new_node
+ assert layer.name not in affinity_predictor_layer_name_to_new_node
input_layer_names = []
for inbound_node in layer._inbound_nodes:
for inbound_layer in inbound_node.inbound_layers:
input_layer_names.append(inbound_layer.name)
input_nodes = [
- layer_name_to_new_node[name]
+ affinity_predictor_layer_name_to_new_node[name]
for name in input_layer_names
]
if len(input_nodes) == 1:
- lifted = TimeDistributed(layer)
+ lifted = keras.layers.TimeDistributed(layer, name=layer.name)
node = lifted(input_nodes[0])
else:
node = layer(input_nodes)
- layer_name_to_new_node[layer.name] = node
+ affinity_predictor_layer_name_to_new_node[layer.name] = node
- node = Reshape(
- (self.hyperparameters['max_alleles'],),
- name="unmasked_affinity_matrix_output")(node)
-
- pre_mask_affinity_predictor_matrix_output = node
+ def logit(x):
+ import tensorflow as tf
+ return -tf.log(1. / x - 1.)
+
+ #node = keras.layers.Lambda(logit, name="logit")(node)
+ affinity_prediction_and_other_signals = [node]
+ if self.hyperparameters['include_cleavage']:
+ assert isinstance(cleavage_model, Class1CleavageNeuralNetwork)
+ cleavage_model = copy(cleavage_model)
+ self.cleavage_model = cleavage_model
+ cleavage_network = cleavage_model.network()
+
+ model_inputs['sequence'] = keras.layers.Input(
+ shape=cleavage_network.get_layer("sequence").output_shape[1:],
+ dtype='float32',
+ name='sequence')
+ model_inputs['peptide_length'] = keras.layers.Input(
+ shape=(1,),
+ dtype='int32',
+ name='peptide_length')
+ cleavage_network.name = "cleavage_predictor"
+ cleavage_prediction = cleavage_network([
+ model_inputs['peptide_length'],
+ model_inputs['sequence'],
+ ])
+ cleavage_prediction.trainable = False
+ cleavage_prediction_repeated = keras.layers.RepeatVector(
+ self.hyperparameters['max_alleles'])(cleavage_prediction)
+ affinity_prediction_and_other_signals.append(
+ cleavage_prediction_repeated)
- # Apply allele mask: zero out all outputs corresponding to alleles
- # with the special index 0.
- def alleles_to_mask(x):
- import keras.backend as K
- result = K.cast(K.not_equal(x, 0), "float32")
- return result
-
- allele_mask = Lambda(alleles_to_mask, name="allele_mask")(input_alleles)
-
- affinity_predictor_matrix_output = Multiply(
- name="affinity_matrix_output")([
- allele_mask,
- pre_mask_affinity_predictor_matrix_output
- ])
- node = Reshape(
- (self.hyperparameters['max_alleles'], 1),
- name="expand_dims_affinity_matrix_output")(
- affinity_predictor_matrix_output)
-
- auxiliary_input = None
if self.hyperparameters['auxiliary_input_features']:
- auxiliary_input = Input(
+ model_inputs['auxiliary'] = keras.layers.Input(
shape=(
self.hyperparameters['max_alleles'],
len(
@@ -218,116 +208,154 @@ class Class1PresentationNeuralNetwork(object):
'auxiliary_input_feature_parameters']))),
dtype="float32",
name="auxiliary")
- node = concatenate(
- [node, auxiliary_input], name="affinities_with_auxiliary")
-
- layer = Dense(8, activation="tanh")
- lifted = TimeDistributed(layer, name="presentation_adjustment_hidden1")
- node = lifted(node)
-
- # By initializing to zero we ensure that before training the
- # presentation output is the same as the affinity output.
- layer = Dense(
- 1,
- activation="tanh",
- kernel_initializer=Zeros(),
- bias_initializer=Zeros())
- lifted = TimeDistributed(layer, name="presentation_adjustment")
- presentation_adjustment = lifted(node)
- presentation_adjustment = Reshape(
- target_shape=(self.hyperparameters['max_alleles'],),
- name="reshaped_presentation_adjustment")(presentation_adjustment)
-
-
- def logit(x):
- import tensorflow as tf
- return - tf.math.log(
- tf.maximum(
- tf.math.divide_no_nan(1., x) - 1.,
- 0.0))
-
- presentation_output_pre_sigmoid = Add()([
- Lambda(logit, name="logit")(affinity_predictor_matrix_output),
- presentation_adjustment,
- ])
- pre_mask_presentation_output = Activation(
- "sigmoid", name="unmasked_presentation_output")(
- presentation_output_pre_sigmoid)
+ affinity_prediction_and_other_signals.append(
+ model_inputs['auxiliary'])
+
+ if len(affinity_prediction_and_other_signals) > 1:
+ node = keras.layers.concatenate(
+ affinity_prediction_and_other_signals,
+ name="affinity_prediction_and_other_signals")
+ layer = keras.layers.Dense(
+ 1,
+ activation="sigmoid",
+ kernel_initializer=keras.initializers.Ones(),
+ name="combine")
+ lifted = keras.layers.TimeDistributed(layer, name="per_allele_output")
+ node = lifted(node)
+ else:
+ (node,) = affinity_prediction_and_other_signals
# Apply allele mask: zero out all outputs corresponding to alleles
# with the special index 0.
- presentation_output = Multiply(name="presentation_output")([
- allele_mask,
- pre_mask_presentation_output
- ])
-
- self.network = Model(
- inputs=[
- input_peptides,
- input_alleles,
- ] + ([] if auxiliary_input is None else [auxiliary_input]),
- outputs=[
- affinity_predictor_matrix_output,
- presentation_output,
- ],
+ #def alleles_to_mask(x):
+ # import keras.backend as K
+ # result = K.expand_dims(
+ # K.cast(K.not_equal(x, 0), "float32"), axis=-1)
+ # return result
+
+ #allele_mask = keras.layers.Lambda(
+ # alleles_to_mask, name="allele_mask")(model_inputs['allele_set'])
+
+ #node = keras.layers.Multiply(
+ # name="masked_per_allele_outputs")(
+ # [allele_mask, node])
+
+ presentation_output = keras.layers.Reshape(
+ target_shape=(self.hyperparameters['max_alleles'],))(
+ node)
+
+ self.network = keras.models.Model(
+ inputs=list(model_inputs.values()),
+ outputs=presentation_output,
name="presentation",
)
- def copy_weights_to_affinity_model(self, model):
- # We assume that the other model's layers are a prefix of ours.
- self.clear_allele_representations()
- model.clear_allele_representations()
- model.network().set_weights(
- self.get_weights()[:len(model.get_weights())])
+ if not self.hyperparameters['trainable_cleavage_predictor']:
+ if self.hyperparameters['include_cleavage']:
+ self.network.get_layer("cleavage_predictor").trainable = False
- def peptides_to_network_input(self, peptides):
- """
- Encode peptides to the fixed-length encoding expected by the neural
- network (which depends on the architecture).
+ self.affinity_predictor_layer_names = list(
+ affinity_predictor_layer_name_to_new_node)
- Parameters
- ----------
- peptides : EncodableSequences or list of string
+ self.set_trainable(
+ trainable_affinity_predictor=(
+ self.hyperparameters['trainable_affinity_predictor']))
- Returns
- -------
- numpy.array
- """
- encoder = EncodableSequences.create(peptides)
- encoded = encoder.variable_length_to_fixed_length_vector_encoding(
- **self.hyperparameters['peptide_encoding'])
- assert len(encoded) == len(peptides)
- return encoded
+ def set_trainable(self, trainable_affinity_predictor=None):
+ if trainable_affinity_predictor is not None:
+ for name in self.affinity_predictor_layer_names:
+ self.network.get_layer(name).trainable = trainable_affinity_predictor
- def allele_encoding_to_network_input(self, allele_encoding):
+
+ @staticmethod
+ def loss(y_true, y_pred):
+ # Binary cross entropy
+ from keras import backend as K
+ import tensorflow as tf
+
+ y_pred = K.constant(y_pred) if not K.is_tensor(y_pred) else y_pred
+ y_true = K.cast(y_true, y_pred.dtype)
+
+ #y_pred = tf.Print(y_pred, [y_pred], message="y_pred", summarize=50)
+ #y_true = tf.Print(y_true, [y_true], message="y_true", summarize=50)
+
+ #logit_y_pred = -tf.log(1. / y_pred - 1.)
+ #logit_y_pred = tf.Print(logit_y_pred, [logit_y_pred], message="logit_y_pred", summarize=50)
+
+ #softmax = K.softmax(5 * logit_y_pred, axis=-1)
+ #softmax = tf.Print(softmax, [softmax], message="softmax", summarize=50)
+
+ #product = softmax * y_pred
+ #product = tf.Print(product, [product], message="product", summarize=50)
+
+ #result = tf.reduce_sum(product, axis=-1)
+ #result = tf.Print(result, [result], message="result", summarize=50)
+
+ #result = tf.reduce_max(y_pred, axis=-1)
+ result = tf.reduce_sum(y_pred, axis=-1)
+
+ return K.mean(
+ K.binary_crossentropy(y_true, result),
+ axis=-1)
+
+ def network_input(
+ self, peptides, allele_encoding, flanking_encoding=None):
"""
- Encode alleles to the fixed-length encoding expected by the neural
- network (which depends on the architecture).
Parameters
----------
+ peptides : EncodableSequences or list of string
+
allele_encoding : AlleleEncoding
- Returns
- -------
- (numpy.array, numpy.array)
+ flanking_encoding: Flank
- Indices and allele representations.
+ Returns
+ -------
+ numpy.array
"""
- return (
- allele_encoding.indices,
- allele_encoding.allele_representations(
- self.hyperparameters['allele_amino_acid_encoding']))
+ assert self.affinity_model is not None
+
+ (allele_input, allele_representations) = (
+ self.affinity_model.allele_encoding_to_network_input(
+ allele_encoding))
+ peptides = EncodableSequences.create(peptides)
+ x_dict = {
+ 'peptide': self.affinity_model.peptides_to_network_input(peptides),
+ 'allele_set': allele_input,
+ }
+ if self.hyperparameters['include_cleavage']:
+ assert self.cleavage_model is not None
+ numpy.testing.assert_array_equal(
+ peptides.sequences,
+ flanking_encoding.dataframe.peptide.values)
+ if flanking_encoding is None:
+ raise RuntimeError("flanking_encoding required")
+ cleavage_x_dict = self.cleavage_model.network_input(
+ flanking_encoding)
+ x_dict.update(cleavage_x_dict)
+ if self.hyperparameters['auxiliary_input_features']:
+ auxiliary_encoder = AuxiliaryInputEncoder(
+ alleles=allele_encoding.alleles,
+ peptides=peptides.sequences)
+ x_dict[
+ 'auxiliary'
+ ] = auxiliary_encoder.get_array(
+ features=self.hyperparameters['auxiliary_input_features'],
+ feature_parameters=self.hyperparameters[
+ 'auxiliary_input_feature_parameters']) * 0.01
+ #import ipdb;ipdb.set_trace()
+ return (x_dict, allele_representations)
def fit(
self,
+ targets,
peptides,
- labels,
allele_encoding,
- affinities_mask=None, # True when a peptide/label is actually a peptide and an affinity
- inequalities=None, # interpreted only for elements where affinities_mask is True, otherwise ignored
- validation_weights=None,
+ flanking_encoding=None,
+ sample_weights=None,
+ shuffle_permutation=None,
verbose=1,
progress_callback=None,
progress_preamble="",
@@ -340,260 +368,59 @@ class Class1PresentationNeuralNetwork(object):
allele_encoding.max_alleles_per_experiment ==
self.hyperparameters['max_alleles'])
- encodable_peptides = EncodableSequences.create(peptides)
-
- if labels is not None:
- labels = numpy.array(labels, copy=False)
- if inequalities is not None:
- inequalities = numpy.array(inequalities, copy=True)
- else:
- inequalities = numpy.tile("=", len(labels))
- if affinities_mask is not None:
- affinities_mask = numpy.array(affinities_mask, copy=False)
- else:
- affinities_mask = numpy.tile(False, len(labels))
- if validation_weights is None:
- validation_weights = numpy.tile(1.0, len(labels))
- else:
- validation_weights = numpy.array(validation_weights, copy=False)
- inequalities[~affinities_mask] = "="
-
- random_negatives_planner = RandomNegativePeptides(
- **RandomNegativePeptides.hyperparameter_defaults.subselect(
- self.hyperparameters))
- random_negatives_planner.plan(
- peptides=encodable_peptides.sequences,
- affinities=numpy.where(affinities_mask, labels, to_ic50(labels)),
- alleles=[
- numpy.random.choice(row[row != numpy.array(None)])
- for row in allele_encoding.alleles
- ],
- inequalities=inequalities)
-
- peptide_input = self.peptides_to_network_input(encodable_peptides)
-
- # Optional optimization
- (allele_encoding_input, allele_representations) = (
- self.allele_encoding_to_network_input(allele_encoding))
-
- x_dict_without_random_negatives = {
- 'peptide': peptide_input,
- 'allele': allele_encoding_input,
- }
- if self.hyperparameters['auxiliary_input_features']:
- auxiliary_encoder = AuxiliaryInputEncoder(
- alleles=allele_encoding.alleles,
- peptides=peptides)
- x_dict_without_random_negatives[
- 'auxiliary'
- ] = auxiliary_encoder.get_array(
- features=self.hyperparameters['auxiliary_input_features'],
- feature_parameters=self.hyperparameters[
- 'auxiliary_input_feature_parameters'])
-
- y1 = numpy.zeros(shape=len(labels))
- y1[affinities_mask] = from_ic50(labels[affinities_mask])
-
- random_negative_alleles = random_negatives_planner.get_alleles()
- random_negatives_allele_encoding = MultipleAlleleEncoding(
- experiment_names=random_negative_alleles,
- experiment_to_allele_list=dict(
- (a, [a]) for a in random_negative_alleles),
- max_alleles_per_experiment=(
- allele_encoding.max_alleles_per_experiment),
- borrow_from=allele_encoding.allele_encoding)
- num_random_negatives = random_negatives_planner.get_total_count()
-
- # Reverse inequalities because from_ic50() flips the direction
- # (i.e. lower affinity results in higher y values).
- adjusted_inequalities = pandas.Series(inequalities).map({
- "=": "=",
- ">": "<",
- "<": ">",
- }).values
- adjusted_inequalities[~affinities_mask] = ">"
-
- # Note: we are using "<" here not ">" because the inequalities are
- # now in target-space (0-1) not affinity-space.
- adjusted_inequalities_with_random_negative = numpy.concatenate([
- numpy.tile("<", num_random_negatives),
- adjusted_inequalities
- ])
- random_negative_ic50 = self.hyperparameters[
- 'random_negative_affinity_min'
- ]
- y1_with_random_negatives = numpy.concatenate([
- numpy.tile(
- from_ic50(random_negative_ic50), num_random_negatives),
- y1,
- ])
-
- def tensor_max(matrix):
- import keras.backend as K
- return K.max(matrix, axis=1)
-
- affinities_loss = TransformPredictionsLossWrapper(
- loss=MSEWithInequalities(),
- y_pred_transform=tensor_max)
- encoded_y1 = affinities_loss.encode_y(
- y1_with_random_negatives,
- inequalities=adjusted_inequalities_with_random_negative)
-
- mms_loss = MultiallelicMassSpecLoss(
- delta=self.hyperparameters['loss_multiallelic_mass_spec_delta'],
- multiplier=self.hyperparameters[
- 'loss_multiallelic_mass_spec_multiplier'])
- y2 = labels.copy()
- y2[affinities_mask] = -1
- y2_with_random_negatives = numpy.concatenate([
- numpy.tile(0.0, num_random_negatives),
- y2,
- ])
- encoded_y2 = mms_loss.encode_y(y2_with_random_negatives)
+ (x_dict, allele_representations) = (
+ self.network_input(
+ peptides=peptides,
+ allele_encoding=allele_encoding,
+ flanking_encoding=flanking_encoding))
+
+ # Shuffle
+ if shuffle_permutation is None:
+ shuffle_permutation = numpy.random.permutation(len(targets))
+ targets = numpy.array(targets)[shuffle_permutation]
+ assert numpy.isnan(targets).sum() == 0, targets
+ if sample_weights is not None:
+ sample_weights = numpy.array(sample_weights)[shuffle_permutation]
+ for key in list(x_dict):
+ x_dict[key] = x_dict[key][shuffle_permutation]
+ del peptides
+ del allele_encoding
+ del flanking_encoding
fit_info = collections.defaultdict(list)
allele_representations_hash = self.set_allele_representations(
allele_representations)
- loss_reduction = "sum"
+
self.network.compile(
- loss=[
- affinities_loss.get_keras_loss(reduction=loss_reduction),
- mms_loss.get_keras_loss(reduction=loss_reduction),
- ],
+ loss=self.loss,
optimizer=self.hyperparameters['optimizer'])
if self.hyperparameters['learning_rate'] is not None:
K.set_value(
self.network.optimizer.lr,
self.hyperparameters['learning_rate'])
- fit_info["learning_rate"] = float(
- K.get_value(self.network.optimizer.lr))
+ fit_info["learning_rate"] = float(K.get_value(self.network.optimizer.lr))
if verbose:
self.network.summary()
- batch_generator = BatchGenerator.create(
- hyperparameters=BatchGenerator.hyperparameter_defaults.subselect(
- self.hyperparameters))
- start = time.time()
- batch_generator.plan(
- num=len(peptides) + num_random_negatives,
- affinities_mask=numpy.concatenate([
- numpy.tile(True, num_random_negatives),
- affinities_mask
- ]),
- experiment_names=numpy.concatenate([
- numpy.tile(None, num_random_negatives),
- allele_encoding.experiment_names
- ]),
- alleles_matrix=numpy.concatenate([
- random_negatives_allele_encoding.alleles,
- allele_encoding.alleles,
- ]),
- is_binder=numpy.concatenate([
- numpy.tile(False, num_random_negatives),
- numpy.where(affinities_mask, labels, to_ic50(labels)) < 1000.0
- ]),
- validation_weights=numpy.concatenate([
- numpy.tile(0.0, num_random_negatives),
- validation_weights
- ]),
- )
- if verbose:
- print("Generated batch generation plan in %0.2f sec." % (
- time.time() - start))
- print(batch_generator.summary())
+ training_start = time.time()
min_val_loss_iteration = None
min_val_loss = None
last_progress_print = 0
- start = time.time()
- x_dict_with_random_negatives = {}
for i in range(self.hyperparameters['max_epochs']):
epoch_start = time.time()
-
- random_negative_peptides = EncodableSequences.create(
- random_negatives_planner.get_peptides())
- random_negative_peptides_encoding = (
- self.peptides_to_network_input(random_negative_peptides))
-
- if not x_dict_with_random_negatives:
- if len(random_negative_peptides) > 0:
- x_dict_with_random_negatives[
- "peptide"
- ] = numpy.concatenate([
- random_negative_peptides_encoding,
- x_dict_without_random_negatives['peptide'],
- ])
- x_dict_with_random_negatives[
- 'allele'
- ] = numpy.concatenate([
- self.allele_encoding_to_network_input(
- random_negatives_allele_encoding)[0],
- x_dict_without_random_negatives['allele']
- ])
- if 'auxiliary' in x_dict_without_random_negatives:
- random_negative_auxiliary_encoder = AuxiliaryInputEncoder(
- alleles=random_negatives_allele_encoding.alleles,
- #peptides=random_negative_peptides.sequences
- )
- x_dict_with_random_negatives['auxiliary'] = (
- numpy.concatenate([
- random_negative_auxiliary_encoder.get_array(
- features=self.hyperparameters[
- 'auxiliary_input_features'],
- feature_parameters=self.hyperparameters[
- 'auxiliary_input_feature_parameters']),
- x_dict_without_random_negatives['auxiliary']
- ]))
- else:
- x_dict_with_random_negatives = (
- x_dict_without_random_negatives)
- else:
- # Update x_dict_with_random_negatives in place.
- # This is more memory efficient than recreating it as above.
- if len(random_negative_peptides) > 0:
- x_dict_with_random_negatives[
- "peptide"
- ][:num_random_negatives] = random_negative_peptides_encoding
-
- if i == 0:
- (train_generator, test_generator) = (
- batch_generator.get_train_and_test_generators(
- x_dict=x_dict_with_random_negatives,
- y_list=[encoded_y1, encoded_y2],
- epochs=1))
- pairs = [
- ("train", train_generator, batch_generator.num_train_batches),
- ("test", test_generator, batch_generator.num_test_batches),
- ]
- for (kind, generator, steps) in pairs:
- self.assert_allele_representations_hash(
- allele_representations_hash)
- metrics = self.network.evaluate_generator(
- generator=generator,
- steps=steps,
- workers=0,
- use_multiprocessing=False)
- for (key, val) in zip(self.network.metrics_names, metrics):
- fit_info["pre_fit_%s_%s" % (kind, key)] = val
- (train_generator, test_generator) = (
- batch_generator.get_train_and_test_generators(
- x_dict=x_dict_with_random_negatives,
- y_list=[encoded_y1, encoded_y2],
- epochs=1))
self.assert_allele_representations_hash(allele_representations_hash)
- fit_history = self.network.fit_generator(
- train_generator,
- steps_per_epoch=batch_generator.num_train_batches,
+ fit_history = self.network.fit(
+ x_dict,
+ targets,
+ validation_split=self.hyperparameters['validation_split'],
+ batch_size=self.hyperparameters['minibatch_size'],
epochs=i + 1,
+ sample_weight=sample_weights,
initial_epoch=i,
- verbose=verbose,
- use_multiprocessing=False,
- workers=0,
- validation_data=test_generator,
- validation_steps=batch_generator.num_test_batches)
-
+ verbose=verbose)
epoch_time = time.time() - epoch_start
for (key, value) in fit_history.history.items():
@@ -608,21 +435,25 @@ class Class1PresentationNeuralNetwork(object):
time.time() - last_progress_print
> progress_print_interval)):
print((progress_preamble + " " +
- "Epoch %3d / %3d [%0.2f sec]: loss=%g. "
+ "Epoch %3d / %3d [%0.2f sec]: loss=%g val_loss=%g. "
"Min val loss (%s) at epoch %s" % (
i,
self.hyperparameters['max_epochs'],
epoch_time,
fit_info['loss'][-1],
+ (
+ fit_info['val_loss'][-1]
+ if 'val_loss' in fit_info else numpy.nan
+ ),
str(min_val_loss),
min_val_loss_iteration)).strip())
last_progress_print = time.time()
- if batch_generator.num_test_batches:
+ if self.hyperparameters['validation_split']:
val_loss = fit_info['val_loss'][-1]
+
if min_val_loss is None or (
- val_loss < min_val_loss -
- self.hyperparameters['min_delta']):
+ val_loss < min_val_loss - self.hyperparameters['min_delta']):
min_val_loss = val_loss
min_val_loss_iteration = i
@@ -647,51 +478,28 @@ class Class1PresentationNeuralNetwork(object):
if progress_callback:
progress_callback()
- fit_info["time"] = time.time() - start
- fit_info["num_points"] = len(labels)
+ fit_info["time"] = time.time() - training_start
+ fit_info["num_points"] = len(targets)
self.fit_info.append(dict(fit_info))
- return {
- 'batch_generator': batch_generator,
- 'last_x': x_dict_with_random_negatives,
- 'last_y': [encoded_y1, encoded_y2],
- 'fit_info': fit_info,
- }
-
- Predictions = collections.namedtuple(
- "ligandone_neural_network_predictions",
- "affinity score")
-
def predict(
self,
peptides,
- allele_encoding=None,
+ allele_encoding,
+ flanking_encoding=None,
batch_size=DEFAULT_PREDICT_BATCH_SIZE):
peptides = EncodableSequences.create(peptides)
assert isinstance(allele_encoding, MultipleAlleleEncoding)
- (allele_encoding_input, allele_representations) = (
- self.allele_encoding_to_network_input(allele_encoding))
- self.set_allele_representations(allele_representations)
- x_dict = {
- 'peptide': self.peptides_to_network_input(peptides),
- 'allele': allele_encoding_input,
- }
- if self.hyperparameters['auxiliary_input_features']:
- auxiliary_encoder = AuxiliaryInputEncoder(
- alleles=allele_encoding.alleles,
- peptides=peptides.sequences)
- x_dict[
- 'auxiliary'
- ] = auxiliary_encoder.get_array(
- features=self.hyperparameters['auxiliary_input_features'],
- feature_parameters=self.hyperparameters[
- 'auxiliary_input_feature_parameters'])
+ (x_dict, allele_representations) = self.network_input(
+ peptides=peptides,
+ allele_encoding=allele_encoding,
+ flanking_encoding=flanking_encoding)
- predictions = self.Predictions._make(
- self.network.predict(x_dict, batch_size=batch_size))
- return predictions
+ self.set_allele_representations(allele_representations)
+ raw_predictions = self.network.predict(x_dict, batch_size=batch_size)
+ return raw_predictions
def clear_allele_representations(self):
"""
diff --git a/mhcflurry/class1_presentation_predictor.py b/mhcflurry/class1_presentation_predictor.py
index cc266f22..fcea5053 100644
--- a/mhcflurry/class1_presentation_predictor.py
+++ b/mhcflurry/class1_presentation_predictor.py
@@ -26,6 +26,7 @@ from .allele_encoding import MultipleAlleleEncoding
from .downloads import get_default_class1_presentation_models_dir
from .class1_presentation_neural_network import Class1PresentationNeuralNetwork
from .common import save_weights, load_weights, NumpyJSONEncoder
+from .flanking_encoding import FlankingEncoding
class Class1PresentationPredictor(object):
@@ -104,16 +105,27 @@ class Class1PresentationPredictor(object):
"""
return join(models_dir, "weights_%s.npz" % model_name)
- def predict(self, peptides, alleles, batch_size=DEFAULT_PREDICT_BATCH_SIZE):
+ def predict(
+ self,
+ peptides,
+ alleles,
+ n_flanks=None,
+ c_flanks=None,
+ batch_size=DEFAULT_PREDICT_BATCH_SIZE):
return self.predict_to_dataframe(
peptides=peptides,
alleles=alleles,
+ n_flanks=n_flanks,
+ c_flanks=c_flanks,
batch_size=batch_size).score.values
def predict_to_dataframe(
self,
peptides,
alleles,
+ n_flanks=None,
+ c_flanks=None,
+ flanking_encoding=None,
include_details=False,
batch_size=DEFAULT_PREDICT_BATCH_SIZE):
@@ -146,31 +158,38 @@ class Class1PresentationPredictor(object):
allele_to_sequence=self.allele_to_sequence,
max_alleles_per_experiment=self.max_alleles)
+ if n_flanks is not None:
+ if flanking_encoding is not None:
+ raise ValueError(
+ "Specify either n_flanks/c_flanks or flanking_encoding, not"
+ "both.")
+ if c_flanks is None:
+ raise ValueError("Both flanks required")
+ flanking_encoding = FlankingEncoding(
+ peptides=peptides.sequences,
+ n_flanks=n_flanks,
+ c_flanks=c_flanks)
+
score_array = []
- affinity_array = []
for (i, network) in enumerate(self.models):
predictions = network.predict(
peptides=peptides,
allele_encoding=alleles,
+ flanking_encoding=flanking_encoding,
batch_size=batch_size)
- score_array.append(predictions.score)
- affinity_array.append(predictions.affinity)
+ score_array.append(predictions)
score_array = numpy.array(score_array)
- affinity_array = numpy.array(affinity_array)
ensemble_scores = numpy.mean(score_array, axis=0)
- ensemble_affinity = numpy.mean(affinity_array, axis=0)
top_allele_index = numpy.argmax(ensemble_scores, axis=-1)
top_allele_flat_indices = (
numpy.arange(len(peptides)) * self.max_alleles + top_allele_index)
top_score = ensemble_scores.flatten()[top_allele_flat_indices]
- top_affinity = ensemble_affinity.flatten()[top_allele_flat_indices]
result_df = pandas.DataFrame({"peptide": peptides.sequences})
result_df["allele"] = alleles.alleles.flatten()[top_allele_flat_indices]
result_df["score"] = top_score
- result_df["affinity"] = to_ic50(top_affinity)
if include_details:
for i in range(self.max_alleles):
@@ -180,12 +199,6 @@ class Class1PresentationPredictor(object):
score_array[:, :, i], 5.0, axis=0)
result_df["allele%d score high" % (i + 1)] = numpy.percentile(
score_array[:, :, i], 95.0, axis=0)
- result_df["allele%d affinity" % (i + 1)] = to_ic50(
- ensemble_affinity[:, i])
- result_df["allele%d affinity low" % (i + 1)] = to_ic50(
- numpy.percentile(affinity_array[:, :, i], 95.0, axis=0))
- result_df["allele%d affinity high" % (i + 1)] = to_ic50(
- numpy.percentile(affinity_array[:, :, i], 5.0, axis=0))
return result_df
def check_consistency(self):
diff --git a/mhcflurry/multiallelic_refinement_command.py b/mhcflurry/multiallelic_refinement_command.py
index 38cbb0e5..6165c4bc 100644
--- a/mhcflurry/multiallelic_refinement_command.py
+++ b/mhcflurry/multiallelic_refinement_command.py
@@ -310,7 +310,7 @@ def refine_model(
presentation_model.load_from_class1_neural_network(affinity_model)
presentation_model.fit(
peptides=combined_train_df.peptide.values,
- labels=combined_train_df.label.values,
+ targets=combined_train_df.label.values,
allele_encoding=allele_encoding,
affinities_mask=combined_train_df.is_affinity.values,
inequalities=combined_train_df.measurement_inequality.values,
diff --git a/mhcflurry/random_negative_peptides.py b/mhcflurry/random_negative_peptides.py
index 978a93e3..684a8aa8 100644
--- a/mhcflurry/random_negative_peptides.py
+++ b/mhcflurry/random_negative_peptides.py
@@ -16,7 +16,7 @@ class RandomNegativePeptides(object):
hyperparameter_defaults = HyperparameterDefaults(
random_negative_rate=0.0,
- random_negative_constant=25,
+ random_negative_constant=0,
random_negative_match_distribution=True,
random_negative_distribution_smoothing=0.0,
random_negative_method="recommended",
diff --git a/test/test_class1_presentation_neural_network.py b/test/test_class1_presentation_neural_network.py
index 08056889..cb52127b 100644
--- a/test/test_class1_presentation_neural_network.py
+++ b/test/test_class1_presentation_neural_network.py
@@ -18,6 +18,7 @@ from random import shuffle
from sklearn.metrics import roc_auc_score
from mhcflurry import Class1AffinityPredictor
+from mhcflurry.class1_cleavage_predictor import Class1CleavagePredictor
from mhcflurry.allele_encoding import MultipleAlleleEncoding
from mhcflurry.class1_presentation_neural_network import Class1PresentationNeuralNetwork
from mhcflurry.class1_presentation_predictor import Class1PresentationPredictor
@@ -32,7 +33,7 @@ from mhcflurry.regression_target import to_ic50
# disable
-sys.exit(0)
+#sys.exit(0)
###################################################
# SETUP
@@ -41,19 +42,26 @@ sys.exit(0)
COMMON_AMINO_ACIDS = sorted(COMMON_AMINO_ACIDS)
AFFINITY_PREDICTOR = None
+CLEAVAGE_PREDICTOR = None
+
def setup():
global AFFINITY_PREDICTOR
+ global CLEAVAGE_PREDICTOR
startup()
AFFINITY_PREDICTOR = Class1AffinityPredictor.load(
get_path("models_class1_pan_variants", "models.affinity_only"),
optimization_level=0,
max_models=1)
+ CLEAVAGE_PREDICTOR = Class1CleavagePredictor.load(max_models=1)
+
def teardown():
global AFFINITY_PREDICTOR
+ global CLEAVAGE_PREDICTOR
AFFINITY_PREDICTOR = None
+ CLEAVAGE_PREDICTOR = None
cleanup()
@@ -64,6 +72,9 @@ def data_path(name):
'''
return os.path.join(os.path.dirname(__file__), "data", name)
+#disable
+#sys.exit(0)
+
###################################################
# UTILITY FUNCTIONS
@@ -100,7 +111,22 @@ def make_motif(presentation_predictor, allele, peptides, frac=0.01, master_allel
# TESTS
###################################################
-def Xtest_synthetic_allele_refinement_max_affinity(include_affinities=True):
+def Xtest_build():
+ global AFFINITY_PREDICTOR
+ global CLEAVAGE_PREDICTOR
+
+ for include_cleavage in [False, True, False, True]:
+ print("Include cleavage: %s" % include_cleavage)
+ model = Class1PresentationNeuralNetwork(
+ include_cleavage=include_cleavage)
+ model.build(
+ affinity_model=AFFINITY_PREDICTOR.class1_pan_allele_models[0],
+ cleavage_model=CLEAVAGE_PREDICTOR.models[0])
+ network = model.network
+ print(network.summary())
+
+
+def test_synthetic_allele_refinement():
"""
Test that in a synthetic example the model is able to learn that HLA-C*01:02
prefers P at position 3.
@@ -140,79 +166,53 @@ def Xtest_synthetic_allele_refinement_max_affinity(include_affinities=True):
hits_df = pandas.DataFrame({"peptide": train_peptides})
hits_df["true_allele"] = train_true_alleles
hits_df["hit"] = 1.0
- hits_df["label"] = 500
- hits_df["measurement_inequality"] = "<"
decoys_df = hits_df.copy()
decoys_df["peptide"] = decoys_df.peptide.map(scramble_peptide)
decoys_df["true_allele"] = ""
decoys_df["hit"] = 0.0
- decoys_df["label"] = 500
- hits_df["measurement_inequality"] = ">"
-
- mms_train_df = pandas.concat([hits_df, decoys_df], ignore_index=True)
- mms_train_df["label"] = mms_train_df.hit
- mms_train_df["is_affinity"] = True
-
- if include_affinities:
- affinity_train_df = pandas.read_csv(get_path("models_class1_pan",
- "models.combined/train_data.csv.bz2"))
- affinity_train_df = affinity_train_df.loc[
- affinity_train_df.allele.isin(alleles),
- ["peptide", "allele", "measurement_inequality", "measurement_value"]
- ]
-
- affinity_train_df["label"] = affinity_train_df["measurement_value"]
- del affinity_train_df["measurement_value"]
- affinity_train_df["is_affinity"] = True
- else:
- affinity_train_df = None
+
+ train_df = pandas.concat(
+ [hits_df, decoys_df], ignore_index=True).sample(frac=1.0)
(affinity_model,) = AFFINITY_PREDICTOR.class1_pan_allele_models
presentation_model = Class1PresentationNeuralNetwork(
+ include_cleavage=False,
+ trainable_affinity_predictor=False,
auxiliary_input_features=["gene"],
- batch_generator_batch_size=1024,
+ minibatch_size=1024,
max_epochs=10,
learning_rate=0.001,
patience=5,
- min_delta=0.0,
- random_negative_rate=1.0,
- random_negative_constant=25)
- presentation_model.load_from_class1_neural_network(affinity_model)
+ min_delta=0.0)
+ presentation_model.build(affinity_model)
+ print(presentation_model.network.summary())
presentation_predictor = Class1PresentationPredictor(
models=[presentation_model],
allele_to_sequence=AFFINITY_PREDICTOR.allele_to_sequence)
- mms_allele_encoding = MultipleAlleleEncoding(
- experiment_names=["experiment1"] * len(mms_train_df),
+ allele_encoding = MultipleAlleleEncoding(
+ experiment_names=["experiment1"] * len(train_df),
experiment_to_allele_list={
"experiment1": alleles,
}, max_alleles_per_experiment=6,
allele_to_sequence=AFFINITY_PREDICTOR.allele_to_sequence, )
- allele_encoding = copy.deepcopy(mms_allele_encoding)
- if affinity_train_df is not None:
- allele_encoding.append_alleles(affinity_train_df.allele.values)
- train_df = pandas.concat([mms_train_df, affinity_train_df],
- ignore_index=True, sort=False)
- else:
- train_df = mms_train_df
allele_encoding = allele_encoding.compact()
- mms_allele_encoding = mms_allele_encoding.compact()
pre_predictions = presentation_model.predict(
- peptides=mms_train_df.peptide.values,
- allele_encoding=mms_allele_encoding).score
+ peptides=train_df.peptide.values,
+ allele_encoding=allele_encoding)
- expected_pre_predictions = from_ic50(affinity_model.predict(
- peptides=numpy.repeat(mms_train_df.peptide.values, len(alleles)),
- allele_encoding=mms_allele_encoding.allele_encoding, )).reshape(
- (-1, len(alleles)))
- assert_allclose(pre_predictions, expected_pre_predictions, rtol=1e-4)
+ #expected_pre_predictions = from_ic50(affinity_model.predict(
+ # peptides=numpy.repeat(train_df.peptide.values, len(alleles)),
+ # allele_encoding=mms_allele_encoding.allele_encoding, )).reshape(
+ # (-1, len(alleles)))
+ #assert_allclose(pre_predictions, expected_pre_predictions, rtol=1e-4)
random_peptides_encodable = EncodableSequences.create(
- random_peptides(10000, 9))
+ random_peptides(20000, 9))
original_motif = make_motif(
presentation_predictor=presentation_predictor,
@@ -229,209 +229,47 @@ def Xtest_synthetic_allele_refinement_max_affinity(include_affinities=True):
iteration_box[0] += 1
print("*** iteration ", label, "***")
predictions_df = presentation_predictor.predict_to_dataframe(
- peptides=mms_train_df.peptide.values,
- alleles=mms_allele_encoding)
- merged_df = pandas.merge(mms_train_df, predictions_df, on="peptide")
+ peptides=train_df.peptide.values,
+ alleles=allele_encoding)
+ merged_df = pandas.merge(train_df, predictions_df, on="peptide")
merged_hit_df = merged_df.loc[merged_df.hit == 1.0]
correct_allele_fraction = (
merged_hit_df.allele == merged_hit_df.true_allele).mean()
print("Correct allele fraction", correct_allele_fraction)
print(
- "Mean score/affinity for hit",
- merged_df.loc[merged_df.hit == 1.0].score.mean(),
- merged_df.loc[merged_df.hit == 1.0].affinity.mean())
+ "Mean score for hit",
+ merged_df.loc[merged_df.hit == 1.0].score.mean())
print(
- "Mean score/affinity for decoy",
- merged_df.loc[merged_df.hit == 0.0].score.mean(),
- merged_df.loc[merged_df.hit == 0.0].affinity.mean())
+ "Mean score for decoy",
+ merged_df.loc[merged_df.hit == 0.0].score.mean())
+ print("Scores for hit",
+ merged_df.loc[merged_df.hit == 1.0].score.values)
+ print("Scores for decoy",
+ merged_df.loc[merged_df.hit == 0.0].score.values)
+ print("Weights", presentation_model.network.get_layer("per_allele_output").get_weights())
auc = roc_auc_score(merged_df.hit.values, merged_df.score.values)
print("AUC", auc)
return (auc, correct_allele_fraction)
(pre_auc, pre_correct_allele_fraction) = progress(label="Pre fitting")
- presentation_model.fit(peptides=train_df.peptide.values,
- labels=train_df.label.values,
- inequalities=train_df.measurement_inequality.values,
- affinities_mask=train_df.is_affinity.values,
+ #import ipdb ; ipdb.set_trace()
+ presentation_model.fit(
+ peptides=train_df.peptide.values,
+ targets=train_df.hit.values,
allele_encoding=allele_encoding,
progress_callback=progress)
- (post_auc, post_correct_allele_fraction) = progress(label="Done fitting")
-
- final_motif = make_motif(
- presentation_predictor=presentation_predictor,
- peptides=random_peptides_encodable,
- allele=refine_allele)
- print("Final motif proline-3 rate: ", final_motif.loc[3, "P"])
-
- assert_greater(post_auc, pre_auc)
- assert_greater(
- post_correct_allele_fraction, pre_correct_allele_fraction - 0.05)
- assert_greater(final_motif.loc[3, "P"], original_motif.loc[3, "P"])
-
-
-
-def Xtest_synthetic_allele_refinement(include_affinities=True):
- """
- Test that in a synthetic example the model is able to learn that HLA-C*01:02
- prefers P at position 3.
- """
- refine_allele = "HLA-C*01:02"
- alleles = ["HLA-A*02:01", "HLA-B*27:01", "HLA-C*07:01", "HLA-A*03:01",
- "HLA-B*15:01", refine_allele]
- peptides_per_allele = [2000, 1000, 500, 1500, 1200, 800, ]
-
- allele_to_peptides = dict(zip(alleles, peptides_per_allele))
-
- length = 9
-
- train_with_ms = pandas.read_csv(get_path("data_curated",
- "curated_training_data.csv.bz2"))
- train_no_ms = pandas.read_csv(
- get_path("data_curated", "curated_training_data.affinity.csv.bz2"))
-
- def filter_df(df):
- return df.loc[
- (df.allele.isin(alleles)) & (df.peptide.str.len() == length)]
-
- train_with_ms = filter_df(train_with_ms)
- train_no_ms = filter_df(train_no_ms)
-
- ms_specific = train_with_ms.loc[
- ~train_with_ms.peptide.isin(train_no_ms.peptide)]
-
- train_peptides = []
- train_true_alleles = []
- for allele in alleles:
- peptides = ms_specific.loc[ms_specific.allele == allele].peptide.sample(
- n=allele_to_peptides[allele])
- train_peptides.extend(peptides)
- train_true_alleles.extend([allele] * len(peptides))
-
- hits_df = pandas.DataFrame({"peptide": train_peptides})
- hits_df["true_allele"] = train_true_alleles
- hits_df["hit"] = 1.0
-
- decoys_df = hits_df.copy()
- decoys_df["peptide"] = decoys_df.peptide.map(scramble_peptide)
- decoys_df["true_allele"] = ""
- decoys_df["hit"] = 0.0
-
- mms_train_df = pandas.concat([hits_df, decoys_df], ignore_index=True)
- mms_train_df["label"] = mms_train_df.hit
- mms_train_df["is_affinity"] = False
- mms_train_df["measurement_inequality"] = None
-
- if include_affinities:
- affinity_train_df = pandas.read_csv(get_path("models_class1_pan",
- "models.combined/train_data.csv.bz2"))
- affinity_train_df = affinity_train_df.loc[
- affinity_train_df.allele.isin(alleles), ["peptide", "allele",
- "measurement_inequality", "measurement_value"]]
-
- affinity_train_df["label"] = affinity_train_df["measurement_value"]
- del affinity_train_df["measurement_value"]
- affinity_train_df["is_affinity"] = True
- else:
- affinity_train_df = None
-
- (affinity_model,) = AFFINITY_PREDICTOR.class1_pan_allele_models
- presentation_model = Class1PresentationNeuralNetwork(
- #batch_generator="multiallelic_mass_spec",
- batch_generator="simple",
- auxiliary_input_features=["gene"],
- batch_generator_batch_size=1024,
- max_epochs=10,
- learning_rate=0.0001,
- patience=5,
- min_delta=0.0,
- random_negative_rate=0,
- random_negative_constant=0)
- presentation_model.load_from_class1_neural_network(affinity_model)
-
- presentation_predictor = Class1PresentationPredictor(
- models=[presentation_model],
- allele_to_sequence=AFFINITY_PREDICTOR.allele_to_sequence)
-
- mms_allele_encoding = MultipleAlleleEncoding(
- experiment_names=["experiment1"] * len(mms_train_df),
- experiment_to_allele_list={
- "experiment1": alleles,
- }, max_alleles_per_experiment=6,
- allele_to_sequence=AFFINITY_PREDICTOR.allele_to_sequence)
- allele_encoding = copy.deepcopy(mms_allele_encoding)
- if affinity_train_df is not None:
- allele_encoding.append_alleles(affinity_train_df.allele.values)
- train_df = pandas.concat([mms_train_df, affinity_train_df],
- ignore_index=True, sort=False)
- else:
- train_df = mms_train_df
-
- allele_encoding = allele_encoding.compact()
- mms_allele_encoding = mms_allele_encoding.compact()
-
- pre_predictions = presentation_model.predict(
- peptides=mms_train_df.peptide.values,
- allele_encoding=mms_allele_encoding).score
-
- expected_pre_predictions = from_ic50(affinity_model.predict(
- peptides=numpy.repeat(mms_train_df.peptide.values, len(alleles)),
- allele_encoding=mms_allele_encoding.allele_encoding, )).reshape(
- (-1, len(alleles)))
- assert_allclose(pre_predictions, expected_pre_predictions, rtol=1e-4)
-
- random_peptides_encodable = EncodableSequences.create(
- random_peptides(10000, 9))
- original_motif = make_motif(
- presentation_predictor=presentation_predictor,
- peptides=random_peptides_encodable,
- allele=refine_allele)
- print("Original motif proline-3 rate: ", original_motif.loc[3, "P"])
- assert_less(original_motif.loc[3, "P"], 0.1)
-
- iteration_box = [0]
+ progress(label="Done fitting first round")
- def progress(label = None):
- if label is None:
- label = str(iteration_box[0])
- iteration_box[0] += 1
- print("*** iteration ", label, "***")
- predictions_df = presentation_predictor.predict_to_dataframe(
- peptides=mms_train_df.peptide.values,
- alleles=mms_allele_encoding)
- merged_df = pandas.merge(mms_train_df, predictions_df, on="peptide")
- merged_hit_df = merged_df.loc[merged_df.hit == 1.0]
- correct_allele_fraction = (
- merged_hit_df.allele == merged_hit_df.true_allele).mean()
- print("Correct allele fraction", correct_allele_fraction)
- print(
- "Mean score/affinity for hit",
- merged_df.loc[merged_df.hit == 1.0].score.mean(),
- merged_df.loc[merged_df.hit == 1.0].affinity.mean())
- print(
- "Mean score/affinity for decoy",
- merged_df.loc[merged_df.hit == 0.0].score.mean(),
- merged_df.loc[merged_df.hit == 0.0].affinity.mean())
- auc = roc_auc_score(merged_df.hit.values, merged_df.score.values)
- print("AUC", auc)
-
- motif = make_motif(
- presentation_predictor=presentation_predictor,
- peptides=random_peptides_encodable,
- allele=refine_allele,
- master_allele_encoding=allele_encoding.allele_encoding)
- print("Proline-3 rate: ", motif.loc[3, "P"])
-
- return (auc, correct_allele_fraction)
-
- (pre_auc, pre_correct_allele_fraction) = progress(label="Pre fitting")
- presentation_model.fit(peptides=train_df.peptide.values,
- labels=train_df.label.values,
- inequalities=train_df.measurement_inequality.values,
- affinities_mask=train_df.is_affinity.values,
+ presentation_model.set_trainable(trainable_affinity_predictor=True)
+ presentation_model.hyperparameters['learning_rate'] = 1e-4
+ presentation_model.fit(
+ peptides=train_df.peptide.values,
+ targets=train_df.hit.values,
allele_encoding=allele_encoding,
progress_callback=progress)
- (post_auc, post_correct_allele_fraction) = progress(label="Done fitting")
+
+ (post_auc, post_correct_allele_fraction) = progress(label="Done fitting second round")
final_motif = make_motif(
presentation_predictor=presentation_predictor,
@@ -445,6 +283,7 @@ def Xtest_synthetic_allele_refinement(include_affinities=True):
assert_greater(final_motif.loc[3, "P"], original_motif.loc[3, "P"])
+
def Xtest_real_data_multiallelic_refinement(max_epochs=10):
"""
Test on real data that we can learn that HLA-A*02:20 has a preference K at
@@ -550,7 +389,7 @@ def Xtest_real_data_multiallelic_refinement(max_epochs=10):
presentation_model.fit(
peptides=combined_train_df.peptide.values,
- labels=combined_train_df.label.values,
+ targets=combined_train_df.label.values,
allele_encoding=allele_encoding,
affinities_mask=combined_train_df.is_affinity.values,
inequalities=combined_train_df.measurement_inequality.values,
diff --git a/test/test_class1_presentation_predictor.py b/test/test_class1_presentation_predictor.py
index 11e2a3e2..55d16014 100644
--- a/test/test_class1_presentation_predictor.py
+++ b/test/test_class1_presentation_predictor.py
@@ -130,7 +130,7 @@ def Xtest_basic():
train_df.peptide.values, alleles=["HLA-A*02:20"])
model.fit(
peptides=train_df.peptide.values,
- labels=train_df.label.values,
+ targets=train_df.label.values,
allele_encoding=allele_encoding)
train_df["updated_score"] = new_predictor.predict(
train_df.peptide.values,
--
GitLab