diff --git a/mhcflurry/class1_binding_predictor.py b/mhcflurry/class1_binding_predictor.py
index 27f8fc4cf1c1777be47ac0f3edc766e05db201ff..366a5a40b5850f33b3d89e246f04efa6e338f232 100644
--- a/mhcflurry/class1_binding_predictor.py
+++ b/mhcflurry/class1_binding_predictor.py
@@ -34,7 +34,7 @@ from .serialization_helpers import (
load_keras_model_from_disk,
save_keras_model_to_disk
)
-
+from .peptide_encoding import check_valid_index_encoding_array
from .class1_allele_specific_hyperparameters import MAX_IC50
_allele_predictor_cache = {}
@@ -386,14 +386,7 @@ class Class1BindingPredictor(PredictorBase):
Given an encoded array of amino acid indices, returns a vector
of predicted log IC50 values.
"""
- X = np.asarray(X)
- if len(X.shape) != 2:
- raise ValueError("Expected 2d input, got array with shape %s" % (
- X.shape,))
- max_expected_index = 20 if self.allow_unknown_amino_acids else 19
- if X.max() > max_expected_index:
- raise ValueError(
- "Got index %d in peptide encoding, max expected %d" % (
- X.max(),
- max_expected_index))
+ X = check_valid_index_encoding_array(
+ X,
+ allow_unknown_amino_acids=self.allow_unknown_amino_acids)
return self.model.predict(X, verbose=False).flatten()
diff --git a/mhcflurry/data.py b/mhcflurry/data.py
index bbf2e76c98044c29b08f21b5e6e3c379973a7167..57bc2c58dbecb06754ea9ff2dd6fd171bd604fc4 100644
--- a/mhcflurry/data.py
+++ b/mhcflurry/data.py
@@ -26,7 +26,8 @@ from .common import normalize_allele_name
from .amino_acid import common_amino_acids
from .peptide_encoding import (
indices_to_hotshot_encoding,
- fixed_length_from_many_peptides
+ fixed_length_index_encoding,
+ check_valid_index_encoding_array,
)
from .class1_allele_specific_hyperparameters import MAX_IC50
@@ -196,10 +197,32 @@ def load_allele_dicts(
def encode_peptide_to_affinity_dict(
peptide_to_affinity_dict,
peptide_length=9,
- flatten_binary_encoding=True):
+ flatten_binary_encoding=True,
+ allow_unknown_amino_acids=True):
"""
- Given a dictionary mapping from peptide sequences to affinity values,
- returns tuple with the following fields:
+ Given a dictionary mapping from peptide sequences to affinity values, return
+ both index and binary encodings of fixed length peptides, and
+ a vector of their affinities.
+
+ Parameters
+ ----------
+ peptide_to_affinity_dict : dict
+ Keys are peptide strings (of multiple lengths), each mapping to a
+ continuous affinity value.
+
+ peptide_length : int
+ Length of vector encoding
+
+ flatten_binary_encoding : bool
+ Should the binary encoding of a peptide be two-dimensional (9x20)
+ or a flattened 1d vector
+
+ allow_unknown_amino_acids : bool
+ When extending a short vector to the desired peptide length, should
+ we insert every possible amino acid or a designated character "X"
+ indicating an unknown amino acid.
+
+ Returns tuple with the following fields:
- kmer_peptides: fixed length peptide strings
- original_peptides: variable length peptide strings
- counts: how many fixed length peptides were made from this original
@@ -208,24 +231,32 @@ def encode_peptide_to_affinity_dict(
- Y: affinity values associated with original peptides
"""
raw_peptides = list(sorted(peptide_to_affinity_dict.keys()))
- kmer_peptides, original_peptides, counts = \
- fixed_length_from_many_peptides(
+ X_index, kmer_peptides, original_peptides, counts = \
+ fixed_length_index_encoding(
peptides=raw_peptides,
desired_length=peptide_length,
start_offset_shorten=0,
end_offset_shorten=0,
start_offset_extend=0,
- end_offset_extend=0)
+ end_offset_extend=0,
+ allow_unknown_amino_acids=allow_unknown_amino_acids)
+
n_samples = len(kmer_peptides)
+
assert n_samples == len(original_peptides), \
"Mismatch between # of samples (%d) and # of peptides (%d)" % (
n_samples, len(original_peptides))
assert n_samples == len(counts), \
"Mismatch between # of samples (%d) and # of counts (%d)" % (
n_samples, len(counts))
-
- X_index = index_encoding(kmer_peptides, peptide_length)
- X_binary = indices_to_hotshot_encoding(X_index, n_indices=20)
+ assert n_samples == len(X_index), \
+ "Mismatch between # of sample (%d) and index feature vectors (%d)" % (
+ n_samples, len(X_index))
+ X_index = check_valid_index_encoding_array(X_index, allow_unknown_amino_acids)
+ n_indices = 20 + allow_unknown_amino_acids
+ X_binary = indices_to_hotshot_encoding(
+ X_index,
+ n_indices=n_indices)
assert X_binary.shape[0] == X_index.shape[0], \
("Mismatch between number of samples for index encoding (%d)"
@@ -235,7 +266,7 @@ def encode_peptide_to_affinity_dict(
if flatten_binary_encoding:
# collapse 3D input into 2D matrix
- n_binary_features = peptide_length * 20
+ n_binary_features = peptide_length * n_indices
X_binary = X_binary.reshape((n_samples, n_binary_features))
# easier to work with counts when they're an array instead of list
@@ -260,7 +291,7 @@ def load_allele_datasets(
peptide_column_name=None,
peptide_length_column_name="peptide_length",
ic50_column_name="meas",
- only_human=True):
+ only_human=False):
"""
Loads an IEDB dataset, extracts "hot-shot" encoding of fixed length peptides
and log-transforms the IC50 measurement. Returns dictionary mapping allele
diff --git a/mhcflurry/peptide_encoding.py b/mhcflurry/peptide_encoding.py
index d07a41b8407179e62e4d8aebd39784be4c62b6b6..59bfaed5a3a94a3b7eb742d2c058ffe74b6bba56 100644
--- a/mhcflurry/peptide_encoding.py
+++ b/mhcflurry/peptide_encoding.py
@@ -273,12 +273,16 @@ def fixed_length_index_encoding(
refers to the position *before* the start of a peptide and, similarly,
`end_offset_extend` = 0 refers to the position *after* the peptide.
- Returns feature matrix X, a list of original peptides for each feature
- vector, and a list of integer counts indicating how many rows share a
- particular original peptide. When two rows are expanded out of a single
- original peptide, they will both have a count of 2. These counts can
- be useful for down-weighting the importance of multiple feature vectors
- which originate from the same sample.
+ Returns tuple with the following fields:
+ - index encoded feature matrix X
+ - list of fixed length peptides
+ - list of "original" peptides of varying lengths
+ - list of integer counts indicating how many rows came from
+ that original peptide.
+
+ When two rows are expanded out of a single original peptide, they will both
+ have a count of 2. These counts can be useful for down-weighting the
+ importance of multiple feature vectors which originate from the same sample.
"""
if allow_unknown_amino_acids:
insert_letters = ["X"]
@@ -295,5 +299,18 @@ def fixed_length_index_encoding(
start_offset_extend=start_offset_extend,
end_offset_extend=end_offset_extend,
insert_amino_acid_letters=insert_letters)
- X = index_encoding(fixed_length, desired_length)
- return X, original, counts
+ X_index = index_encoding(fixed_length, desired_length)
+ return X_index, fixed_length, original, counts
+
+def check_valid_index_encoding_array(X, allow_unknown_amino_acids=True):
+ X = np.asarray(X)
+ if len(X.shape) != 2:
+ raise ValueError("Expected 2d input, got array with shape %s" % (
+ X.shape,))
+ max_expected_index = 20 if allow_unknown_amino_acids else 19
+ if X.max() > max_expected_index:
+ raise ValueError(
+ "Got index %d in peptide encoding, max expected %d" % (
+ X.max(),
+ max_expected_index))
+ return X
diff --git a/mhcflurry/predictor_base.py b/mhcflurry/predictor_base.py
index 1c6ed73a5aeedbaf1b174601dbfb7e061782e324..a82192fed8f1400532e287c7e937a092b0b7a385 100644
--- a/mhcflurry/predictor_base.py
+++ b/mhcflurry/predictor_base.py
@@ -67,7 +67,7 @@ class PredictorBase(object):
indices = []
encoded_matrices = []
for i, peptide in enumerate(peptides):
- matrix, _, _ = fixed_length_index_encoding(
+ matrix, _, _, _ = fixed_length_index_encoding(
peptides=[peptide],
desired_length=9,
allow_unknown_amino_acids=self.allow_unknown_amino_acids)