fixes

downloads-generation/data_curated/curate.py

@@ -4,9 +4,6 @@ Train single allele models
 """
 import sys
 import argparse
-import json
-import os
-import pickle
 
 import pandas
 
@@ -37,13 +34,20 @@ parser.add_argument(
     required=True,
     help="Result file")
 
-QUALITATIVE_TO_AFFINITY = {
-    "Negative": 50000.0,
-    "Positive": 100.0,
-    "Positive-High": 50.0,
-    "Positive-Intermediate": 500.0,
-    "Positive-Low": 5000.0,
+QUALITATIVE_TO_AFFINITY_AND_INEQUALITY = {
+    "Negative": (20000.0, ">"),
+    "Positive": (500.0, "<"),
+    "Positive-High": (100.0, "<"),
+    "Positive-Intermediate": (1000.0, "<"),
+    "Positive-Low": (5000.0, "<"),
 }
+QUALITATIVE_TO_AFFINITY = dict(
+    (key, value[0]) for (key, value)
+    in QUALITATIVE_TO_AFFINITY_AND_INEQUALITY.items())
+QUALITATIVE_TO_INEQUALITY = dict(
+    (key, value[1]) for (key, value)
+    in QUALITATIVE_TO_AFFINITY_AND_INEQUALITY.items())
+
 
 EXCLUDE_IEDB_ALLELES = [
     "HLA class I",
@@ -60,6 +64,7 @@ def load_data_kim2014(filename):
         True: "quantitative",
         False: "qualitative",
     })
+    df["measurement_inequality"] = df.inequality
     df["original_allele"] = df.mhc
     df["peptide"] = df.sequence
     df["allele"] = df.mhc.map(normalize_allele_name)
@@ -99,24 +104,28 @@ def load_data_iedb(iedb_csv, include_qualitative=True):
 
     quantitative = iedb_df.ix[iedb_df["Units"] == "nM"].copy()
     quantitative["measurement_type"] = "quantitative"
+    quantitative["measurement_inequality"] = "="
     print("Quantitative measurements: %d" % len(quantitative))
 
     qualitative = iedb_df.ix[iedb_df["Units"] != "nM"].copy()
     qualitative["measurement_type"] = "qualitative"
     print("Qualitative measurements: %d" % len(qualitative))
-    non_mass_spec_qualitative = qualitative.ix[
-        (~qualitative["Method/Technique"].str.contains("mass spec"))
-    ].copy()
-    non_mass_spec_qualitative["Quantitative measurement"] = (
-        non_mass_spec_qualitative["Qualitative Measure"].map(
-            QUALITATIVE_TO_AFFINITY))
-    print("Qualitative measurements after dropping MS: %d" % (
-        len(non_mass_spec_qualitative)))
+    #qualitative = qualitative.ix[
+    #    (~qualitative["Method/Technique"].str.contains("mass spec"))
+    #].copy()
+
+    qualitative["Quantitative measurement"] = (
+        qualitative["Qualitative Measure"].map(QUALITATIVE_TO_AFFINITY))
+    qualitative["measurement_inequality"] = (
+        qualitative["Qualitative Measure"].map(QUALITATIVE_TO_INEQUALITY))
+
+    print("Qualitative measurements (possibly after dropping MS): %d" % (
+        len(qualitative)))
 
     iedb_df = pandas.concat(
         (
             ([quantitative]) +
-            ([non_mass_spec_qualitative] if include_qualitative else [])),
+            ([qualitative] if include_qualitative else [])),
         ignore_index=True)
 
     print("IEDB measurements per allele:\n%s" % iedb_df.allele.value_counts())
@@ -145,6 +154,7 @@ def load_data_iedb(iedb_csv, include_qualitative=True):
         "Quantitative measurement"
     ].values
     train_data["measurement_source"] = iedb_df.category.values
+    train_data["measurement_inequality"] = iedb_df.measurement_inequality.values
 
     train_data["allele"] = iedb_df["allele"].values
     train_data["original_allele"] = iedb_df["Allele Name"].values
@@ -181,6 +191,7 @@ def run():
         "allele",
         "peptide",
         "measurement_value",
+        "measurement_inequality",
         "measurement_type",
         "measurement_source",
         "original_allele",

mhcflurry/class1_affinity_predictor.py

@@ -14,12 +14,12 @@ import pandas
 from numpy.testing import assert_equal
 from six import string_types
 
-from mhcflurry.class1_neural_network import Class1NeuralNetwork
-from mhcflurry.common import random_peptides
-from mhcflurry.downloads import get_path
-from mhcflurry.encodable_sequences import EncodableSequences
-from mhcflurry.percent_rank_transform import PercentRankTransform
-from mhcflurry.regression_target import to_ic50
+from .class1_neural_network import Class1NeuralNetwork
+from .common import random_peptides
+from .downloads import get_path
+from .encodable_sequences import EncodableSequences
+from .percent_rank_transform import PercentRankTransform
+from .regression_target import to_ic50
 
 
 class Class1AffinityPredictor(object):
@@ -368,7 +368,7 @@ class Class1AffinityPredictor(object):
             affinities,
             inequalities=None,
             models_dir_for_save=None,
-            verbose=1,
+            verbose=0,
             progress_preamble=""):
         """
         Fit one or more allele specific predictors for a single allele using a
@@ -425,7 +425,7 @@ class Class1AffinityPredictor(object):
             if n_architectures > 1:
                 pieces.append(
                     "Architecture {architecture_num:2d} / {n_architectures:2d}"
-                    " (best so far: {best_num:2d)")
+                    " (best so far: {best_num})")
             progress_preamble_template = "[ %s ] {user_progress_preamble}" % (
                 ", ".join(pieces))
         else:
@@ -450,13 +450,12 @@ class Class1AffinityPredictor(object):
                     verbose=verbose,
                     progress_preamble=progress_preamble_template.format(
                         user_progress_preamble=progress_preamble,
-                        best_num=best_num,
-                        model_num=model_num,
+                        best_num="n/a" if best_num is None else best_num + 1,
+                        model_num=model_num + 1,
                         n_models=n_models,
-                        architecture_num=architecture_num,
+                        architecture_num=architecture_num + 1,
                         n_architectures=n_architectures))
 
-
                 if n_architectures > 1:
                     # We require val_loss (i.e. a validation set) if we have
                     # multiple architectures.
@@ -464,11 +463,14 @@ class Class1AffinityPredictor(object):
                 else:
                     loss = None
                 if loss is None or best_loss is None or best_loss > loss:
-                    best_loss = best_loss
+                    best_loss = loss
                     best_num = architecture_num
                     best_model = model
                 del model
 
+            if n_architectures > 1:
+                print("Selected architecture %d." % (best_num + 1))
+
             model_name = self.model_name(allele, model_num)
             row = pandas.Series(collections.OrderedDict([
                 ("model_name", model_name),

mhcflurry/class1_neural_network.py

@@ -444,9 +444,17 @@ class Class1NeuralNetwork(object):
         if sample_weights is not None:
             sample_weights = sample_weights[shuffle_permutation]
 
-        if self.hyperparameters['loss'] in LOSSES:
+        if self.hyperparameters['loss'].startswith("custom:"):
             # Using a custom loss that supports inequalities
-            loss_name_or_function = LOSSES[self.hyperparameters['loss']]
+            try:
+                loss_name_or_function = LOSSES[
+                    self.hyperparameters['loss'].replace("custom:", "")
+                ]
+            except KeyError:
+                raise ValueError(
+                    "No such custom loss function: %s. Supported losses are: %s" % (
+                        self.hyperparameters['loss'],
+                        ", ".join(["custom:" + loss_name for loss_name in LOSSES])))
             loss_supports_inequalities = True
         else:
             # Using a regular keras loss. No inequalities supported.

mhcflurry/loss_with_inequalities.py

@@ -25,14 +25,14 @@ between 4 - 5:
 from keras import backend as K
 
 import pandas
-import numpy
+from numpy import isnan, array
 
 LOSSES = {}
 
 
 def encode_y(y, inequalities=None):
-    y = numpy.array(y, dtype="float32")
-    if y.isnan().any():
+    y = array(y, dtype="float32")
+    if isnan(y).any():
         raise ValueError("y contains NaN")
     if (y > 1.0).any():
         raise ValueError("y contains values > 1.0")
@@ -47,14 +47,14 @@ def encode_y(y, inequalities=None):
             '<': 2,
             '>': 4,
         }).values
-        if offsets.isnan().any():
+        if isnan(offsets).any():
             raise ValueError("Invalid inequality. Must be =, <, or >")
         encoded = y + offsets
-    assert not encoded.isnan().any()
+    assert not isnan(encoded).any()
     return encoded
 
 
-def mse_with_ineqalities(y_true, y_pred):
+def mse_with_inequalities(y_true, y_pred):
     # Handle (=) inequalities
     diff1 = y_pred - y_true
     diff1 *= K.cast(y_true >= 0.0, "float32")
@@ -75,4 +75,4 @@ def mse_with_ineqalities(y_true, y_pred):
         K.sum(K.square(diff1), axis=-1) +
         K.sum(K.square(diff2), axis=-1) +
         K.sum(K.square(diff3), axis=-1))
-LOSSES["mse_with_ineqalities"] = mse_with_ineqalities
\ No newline at end of file
+LOSSES["mse_with_inequalities"] = mse_with_inequalities
\ No newline at end of file