From f1d17bf092a4e37205a81740c8b949bc5ba6ebb4 Mon Sep 17 00:00:00 2001
From: Tim O'Donnell <timodonnell@gmail.com>
Date: Thu, 26 Sep 2019 15:07:50 -0400
Subject: [PATCH] better random negative peptides
---
mhcflurry/class1_neural_network.py | 229 ++------------------------
mhcflurry/random_negative_peptides.py | 179 ++++++++++++++++++++
test/test_random_negative_peptides.py | 90 +++++-----
3 files changed, 235 insertions(+), 263 deletions(-)
create mode 100644 mhcflurry/random_negative_peptides.py
diff --git a/mhcflurry/class1_neural_network.py b/mhcflurry/class1_neural_network.py
index e6d33a57..1299dd8d 100644
--- a/mhcflurry/class1_neural_network.py
+++ b/mhcflurry/class1_neural_network.py
@@ -18,6 +18,7 @@ from .regression_target import to_ic50, from_ic50
from .common import random_peptides, amino_acid_distribution
from .custom_loss import get_loss
from .data_dependent_weights_initialization import lsuv_init
+from .random_negative_peptides import RandomNegativePeptides
DEFAULT_PREDICT_BATCH_SIZE = 4096
@@ -91,16 +92,10 @@ class Class1NeuralNetwork(object):
early_stopping=True,
minibatch_size=128,
data_dependent_initialization_method=None,
- random_negative_rate=0.0,
- random_negative_constant=25,
random_negative_affinity_min=20000.0,
random_negative_affinity_max=50000.0,
- random_negative_match_distribution=True,
- random_negative_distribution_smoothing=0.0,
- 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])
+ random_negative_output_indices=None).extend(
+ RandomNegativePeptides.hyperparameter_defaults)
"""
Hyperparameters for neural network training.
"""
@@ -677,202 +672,6 @@ class Class1NeuralNetwork(object):
fit_info["num_points"] = mutable_generator_state["yielded_values"]
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
- df_nonbinders = None
- if self.hyperparameters['random_negative_binder_threshold']:
- df_nonbinders = df.loc[
- (df.inequality != "<") &
- (df.affinity > 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(math.ceil(
- num_for_allele / len(random_negative_lengths)))
- total_random_peptides_per_length += num_per_length
- allele_to_num_per_length[allele] = num_per_length
-
- if allele_encoding is not None:
- random_negative_alleles = []
- for _ in random_negative_lengths:
- for (allele, num) in allele_to_num_per_length.items():
- random_negative_alleles.extend([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, since they correspond with
- # specific alleles.
- return EncodableSequences.create(peptides)
- elif self.hyperparameters["random_negative_method"] == "by_allele_equalize_nonbinders":
- assert df_nonbinders is not None
-
- allele_and_length_to_num_random_negative = {}
-
- # First pass
- allele_to_num_per_length_first_pass = {}
- 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(math.ceil(
- num_for_allele / len(random_negative_lengths)))
- allele_to_num_per_length_first_pass[allele] = num_per_length
-
- # Second pass
- for (allele, sub_df) in df_nonbinders.groupby("allele"):
- nonbinders_by_length = sub_df.peptide.str.len().value_counts()
- nonbinders_by_length = nonbinders_by_length.reindex(
- random_negative_lengths
- ).fillna(0)
-
- total_nonbinders_by_length = (
- nonbinders_by_length +
- allele_to_num_per_length_first_pass[allele])
- max_total_nonbinders_by_length = total_nonbinders_by_length.max()
- for (length, total) in total_nonbinders_by_length.items():
- allele_and_length_to_num_random_negative[
- (allele, length)
- ] = math.ceil(
- allele_to_num_per_length_first_pass[allele] +
- max_total_nonbinders_by_length -
- total)
-
- plan_df = []
- for (
- (allele, length), num_random_negative) in (
- allele_and_length_to_num_random_negative.items()):
- plan_df.append((allele, length, num_random_negative))
- plan_df = pandas.DataFrame(
- plan_df, columns=["allele", "length", "num"])
-
- logging.info("Random negative plan:\n%s", plan_df)
-
- if allele_encoding is not None:
- random_negative_alleles = []
- for _, row in plan_df.iterrows():
- random_negative_alleles.extend([row.allele] * int(row.num))
-
- def sample_peptides():
- peptides = []
- for _, row in plan_df.iterrows():
- peptides.extend(
- random_peptides(
- row.num,
- length=row.length,
- distribution=aa_distribution))
- # important NOT to shuffle peptides, since they correspond with
- # specific alleles.
- return EncodableSequences.create(peptides)
- else:
- raise NotImplementedError(
- self.hyperparameters["random_negative_method"])
-
- random_negative_allele_encoding = None
- if random_negative_alleles is not None:
- random_negative_allele_encoding = AlleleEncoding(
- random_negative_alleles, borrow_from=allele_encoding)
-
- yield random_negative_allele_encoding
- while True:
- yield sample_peptides()
-
def fit(
self,
peptides,
@@ -937,14 +736,21 @@ class Class1NeuralNetwork(object):
peptide_encoding = self.peptides_to_network_input(encodable_peptides)
fit_info = collections.defaultdict(list)
- random_negatives_generator = self.random_negatives_generator(
- encodable_peptides=encodable_peptides,
+ random_negatives_planner = RandomNegativePeptides(
+ **RandomNegativePeptides.hyperparameter_defaults.subselect(
+ self.hyperparameters))
+ random_negatives_planner.plan(
+ peptides=encodable_peptides.sequences,
affinities=affinities,
- allele_encoding=allele_encoding,
+ alleles=allele_encoding.alleles if allele_encoding else None,
inequalities=inequalities)
- random_negatives_allele_encoding = next(random_negatives_generator)
- num_random_negatives = len(
- next(random_negatives_generator).sequences)
+
+ random_negatives_allele_encoding = None
+ if allele_encoding is not None:
+ random_negatives_allele_encoding = AlleleEncoding(
+ random_negatives_planner.get_alleles(),
+ borrow_from=allele_encoding)
+ num_random_negatives = random_negatives_planner.get_total_count()
y_values = from_ic50(numpy.array(affinities, copy=False))
assert numpy.isnan(y_values).sum() == 0, y_values
@@ -1107,7 +913,8 @@ class Class1NeuralNetwork(object):
last_progress_print = None
x_dict_with_random_negatives = {}
for i in range(self.hyperparameters['max_epochs']):
- random_negative_peptides = next(random_negatives_generator)
+ random_negative_peptides = EncodableSequences.create(
+ random_negatives_planner.get_peptides())
random_negative_peptides_encoding = (
self.peptides_to_network_input(random_negative_peptides))
diff --git a/mhcflurry/random_negative_peptides.py b/mhcflurry/random_negative_peptides.py
new file mode 100644
index 00000000..80fe7ab4
--- /dev/null
+++ b/mhcflurry/random_negative_peptides.py
@@ -0,0 +1,179 @@
+import logging
+import math
+
+import numpy
+import pandas
+
+from .hyperparameters import HyperparameterDefaults
+from .common import amino_acid_distribution, random_peptides
+
+class RandomNegativePeptides(object):
+ hyperparameter_defaults = HyperparameterDefaults(
+ random_negative_rate=0.0,
+ random_negative_constant=25,
+ random_negative_match_distribution=True,
+ random_negative_distribution_smoothing=0.0,
+ random_negative_method="by_length",
+ random_negative_binder_threshold=None,
+ random_negative_lengths=[8,9,10,11,12,13,14,15])
+
+
+ def __init__(self, **hyperparameters):
+ self.hyperparameters = self.hyperparameter_defaults.with_defaults(
+ hyperparameters)
+ self.plan_df = None
+ self.aa_distribution = None
+
+ def plan(self, peptides, affinities, alleles=None, inequalities=None):
+ peptides = pandas.Series(peptides, copy=False)
+ peptide_lengths = peptides.str.len()
+
+ if self.hyperparameters['random_negative_match_distribution']:
+ self.aa_distribution = amino_acid_distribution(
+ peptides.values,
+ smoothing=self.hyperparameters[
+ 'random_negative_distribution_smoothing'
+ ])
+ logging.info(
+ "Using amino acid distribution for random negative:\n%s" % (
+ str(self.aa_distribution.to_dict())))
+
+ df_all = pandas.DataFrame({
+ 'length': peptide_lengths,
+ 'affinity': affinities,
+ })
+ df_all["allele"] = "" if alleles is None else alleles
+ df_all["inequality"] = "=" if inequalities is None else inequalities
+
+ df_binders = None
+ df_nonbinders = None
+ if self.hyperparameters['random_negative_binder_threshold']:
+ df_nonbinders = df_all.loc[
+ (df_all.inequality != "<") &
+ (df_all.affinity > self.hyperparameters[
+ 'random_negative_binder_threshold'
+ ])
+ ]
+ df_binders = df_all.loc[
+ (df_all.inequality != ">") &
+ (df_all.affinity <= self.hyperparameters[
+ 'random_negative_binder_threshold'
+ ])
+ ]
+
+ method = self.hyperparameters['random_negative_method']
+ function = {
+ 'by_length': self.plan_by_length,
+ 'by_allele': self.plan_by_allele,
+ 'by_allele_equalize_nonbinders':
+ self.plan_by_allele_equalize_nonbinders,
+ }[method]
+ function(df_all, df_binders, df_nonbinders)
+ assert self.plan_df is not None
+ logging.info("Random negative plan [%s]:\n%s", method, self.plan_df)
+ return self.plan_df
+
+ def plan_by_length(self, df_all, df_binders=None, df_nonbinders=None):
+ """
+ Different numbers of random negatives per length. Alleles are sampled
+ proportionally to the number of times they are used in the training
+ data.
+
+ Used for allele-specific predictors. Does not work well for pan-allele.
+ """
+ assert list(df_all.allele.unique()) == [""], (
+ "by_length only recommended for allele specific prediction")
+
+ df = df_all if df_binders is None else df_binders
+ lengths = self.hyperparameters['random_negative_lengths']
+
+ length_to_num_random_negative = {}
+ length_counts = df.length.value_counts().to_dict()
+ for length in lengths:
+ length_to_num_random_negative[length] = int(
+ length_counts.get(length, 0) *
+ self.hyperparameters['random_negative_rate'] +
+ self.hyperparameters['random_negative_constant'])
+
+ plan_df = pandas.DataFrame(index=sorted(df.allele.unique()))
+ for length in lengths:
+ plan_df[length] = length_to_num_random_negative[length]
+ self.plan_df = plan_df.astype(int)
+
+ def plan_by_allele(self, df_all, df_binders=None, df_nonbinders=None):
+ """
+ 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
+ df = df_all if df_binders is None else df_binders
+ lengths = self.hyperparameters['random_negative_lengths']
+ all_alleles = df_all.allele.unique()
+ for allele in all_alleles:
+ sub_df = df.loc[df.allele == allele]
+ num_for_allele = len(sub_df) * (
+ self.hyperparameters['random_negative_rate']
+ ) + self.hyperparameters['random_negative_constant']
+ num_per_length = int(math.ceil(
+ num_for_allele / len(lengths)))
+ total_random_peptides_per_length += num_per_length
+ allele_to_num_per_length[allele] = num_per_length
+
+ plan_df = pandas.DataFrame(index=sorted(df.allele.unique()))
+ for length in lengths:
+ plan_df[length] = plan_df.index.map(allele_to_num_per_length)
+ self.plan_df = plan_df.astype(int)
+
+ def plan_by_allele_equalize_nonbinders(
+ self, df_all, df_binders, df_nonbinders):
+ """
+ """
+ assert df_binders is not None
+ assert df_nonbinders is not None
+
+ lengths = self.hyperparameters['random_negative_lengths']
+
+ self.plan_by_allele(df_all, df_binders, df_nonbinders)
+ first_pass_plan = self.plan_df
+ self.plan_df = None
+
+ new_plan = first_pass_plan.copy()
+ new_plan[:] = numpy.nan
+
+ for (allele, first_pass_per_length) in first_pass_plan.iterrows():
+ real_nonbinders_by_length = df_nonbinders.loc[
+ df_nonbinders.allele == allele
+ ].length.value_counts().reindex(lengths).fillna(0)
+ total_nonbinders_by_length = (
+ real_nonbinders_by_length + first_pass_per_length)
+ new_plan.loc[allele] = first_pass_per_length + (
+ total_nonbinders_by_length.max() - total_nonbinders_by_length)
+
+ self.plan_df = new_plan.astype(int)
+
+ def get_alleles(self):
+ assert self.plan_df is not None, "Call plan() first"
+ alleles = []
+ for allele, row in self.plan_df.iterrows():
+ alleles.extend([allele] * int(row.sum()))
+ assert len(alleles) == self.get_total_count()
+ return alleles
+
+ def get_peptides(self):
+ assert self.plan_df is not None, "Call plan() first"
+ peptides = []
+ for allele, row in self.plan_df.iterrows():
+ for (length, num) in row.items():
+ peptides.extend(
+ random_peptides(
+ num,
+ length=length,
+ distribution=self.aa_distribution))
+ assert len(peptides) == self.get_total_count()
+ return peptides
+
+ def get_total_count(self):
+ return self.plan_df.sum().sum()
\ No newline at end of file
diff --git a/test/test_random_negative_peptides.py b/test/test_random_negative_peptides.py
index c3c978f2..06765cb4 100644
--- a/test/test_random_negative_peptides.py
+++ b/test/test_random_negative_peptides.py
@@ -1,3 +1,7 @@
+import logging
+logging.getLogger('matplotlib').disabled = True
+logging.getLogger('tensorflow').disabled = True
+
from mhcflurry import amino_acid
from nose.tools import eq_
from numpy.testing import assert_equal
@@ -9,20 +13,16 @@ from mhcflurry import Class1NeuralNetwork
from mhcflurry.encodable_sequences import EncodableSequences
from mhcflurry.allele_encoding import AlleleEncoding
from mhcflurry.common import random_peptides
+from mhcflurry.random_negative_peptides import RandomNegativePeptides
def test_random_negative_peptides_by_allele():
- network = Class1NeuralNetwork(
+ planner = RandomNegativePeptides(
random_negative_method="by_allele",
random_negative_binder_threshold=500,
random_negative_rate=1.0,
random_negative_constant=2)
- allele_to_sequence = {
- "HLA-A*02:01": "AAAAA",
- "HLA-B*44:02": "CCCCC",
- "HLA-C*07:02": "EEEEE",
- }
data_rows = [
("HLA-A*02:01", "SIINFEKL", 400, "="),
("HLA-A*02:01", "SIINFEKLL", 300, "="),
@@ -42,25 +42,18 @@ def test_random_negative_peptides_by_allele():
columns=["allele", "peptide", "affinity", "inequality"])
data["length"] = data.peptide.str.len()
- peptides = EncodableSequences.create(data.peptide.values)
- allele_encoding = AlleleEncoding(data.allele.values, allele_to_sequence)
-
- results = network.random_negatives_generator(
- peptides,
- data.affinity.values,
- allele_encoding,
- data.inequality.values
- )
- result_allele_encoding = next(results)
- first_peptide_sample = next(results)
- second_peptide_sample = next(results)
-
- result_df1 = pandas.DataFrame({
- "allele": result_allele_encoding.alleles,
- "peptide": first_peptide_sample.sequences,
+ planner.plan(
+ peptides=data.peptide.values,
+ affinities=data.affinity.values,
+ alleles=data.allele.values,
+ inequalities=data.inequality.values)
+ result_df = pandas.DataFrame({
+ "allele": planner.get_alleles(),
+ "peptide": planner.get_peptides(),
})
- result_df1["length"] = result_df1.peptide.str.len()
- random_negatives = result_df1.groupby(["allele", "length"]).peptide.count().unstack()
+
+ result_df["length"] = result_df.peptide.str.len()
+ random_negatives = result_df.groupby(["allele", "length"]).peptide.count().unstack()
real_data = data.groupby(["allele", "length"]).peptide.count().unstack().fillna(0)
real_binders = data.loc[
data.affinity <= 500
@@ -77,23 +70,20 @@ def test_random_negative_peptides_by_allele():
def test_random_negative_peptides_by_allele():
- network = Class1NeuralNetwork(
+ planner = RandomNegativePeptides(
random_negative_method="by_allele_equalize_nonbinders",
random_negative_binder_threshold=500,
random_negative_rate=1.0,
random_negative_constant=2)
-
- allele_to_sequence = {
- "HLA-A*02:01": "AAAAA",
- "HLA-B*44:02": "CCCCC",
- "HLA-C*07:02": "EEEEE",
- }
data_rows = [
("HLA-A*02:01", "SIINFEKL", 400, "="),
("HLA-A*02:01", "SIINFEKLL", 300, "="),
("HLA-A*02:01", "SIINFEKLL", 300, "="),
("HLA-A*02:01", "SIINFEKLQ", 1000, "="),
("HLA-A*02:01", "SIINFEKLZZ", 12000, ">"),
+ ("HLA-C*01:02", "SIINFEKLQ", 100, "="), # only binders
+ ("HLA-C*07:02", "SIINFEKLL", 1000, "=") # only non-binders
+
]
for peptide in random_peptides(1000, length=9):
data_rows.append(("HLA-B*44:02", peptide, 100, "="))
@@ -107,25 +97,17 @@ def test_random_negative_peptides_by_allele():
columns=["allele", "peptide", "affinity", "inequality"])
data["length"] = data.peptide.str.len()
- peptides = EncodableSequences.create(data.peptide.values)
- allele_encoding = AlleleEncoding(data.allele.values, allele_to_sequence)
-
- results = network.random_negatives_generator(
- peptides,
- data.affinity.values,
- allele_encoding,
- data.inequality.values
- )
- result_allele_encoding = next(results)
- first_peptide_sample = next(results)
- second_peptide_sample = next(results)
-
- result_df1 = pandas.DataFrame({
- "allele": result_allele_encoding.alleles,
- "peptide": first_peptide_sample.sequences,
+ planner.plan(
+ peptides=data.peptide.values,
+ affinities=data.affinity.values,
+ alleles=data.allele.values,
+ inequalities=data.inequality.values)
+ result_df = pandas.DataFrame({
+ "allele": planner.get_alleles(),
+ "peptide": planner.get_peptides(),
})
- result_df1["length"] = result_df1.peptide.str.len()
- random_negatives = result_df1.groupby(["allele", "length"]).peptide.count().unstack()
+ result_df["length"] = result_df.peptide.str.len()
+ random_negatives = result_df.groupby(["allele", "length"]).peptide.count().unstack()
real_data = data.groupby(["allele", "length"]).peptide.count().unstack().fillna(0)
real_binders = data.loc[
data.affinity <= 500
@@ -136,7 +118,11 @@ def test_random_negative_peptides_by_allele():
for length in random_negatives.columns:
if length not in real_nonbinders.columns:
real_nonbinders[length] = 0
- total_nonbinders = random_negatives + real_nonbinders
+ total_nonbinders = (
+ random_negatives.reindex(real_data.index).fillna(0) +
+ real_nonbinders.reindex(real_data.index).fillna(0))
+
+ assert (total_nonbinders.loc["HLA-A*02:01"] == 2.0).all(), total_nonbinders
+ assert (total_nonbinders.loc["HLA-B*44:02"] == 1126).all(), total_nonbinders
- assert (total_nonbinders.loc["HLA-A*02:01"] == 2.0).all()
- assert (total_nonbinders.loc["HLA-B*44:02"] == 1126).all()
+ assert not total_nonbinders.isnull().any().any()
--
GitLab