def check_ml(data, n_run, knn, n_fold, n_proportion, n_subset, model_type, prediction_type, features, recalculate_similarity, disjoint_cv, split_both=False, output_file = None, model_fun = None, verbose=False, n_seed = None):
drugs, disease_to_index, drug_to_values, se_to_index, drug_to_values_se, drug_to_values_structure, drug_to_values_target, drug_interaction_to_index, drug_to_values_interaction = data
if prediction_type == "disease":
# For drug repurposing (feature phenotype: side effect)
disease_to_drugs, pairs, classes = utilities.get_drug_disease_mapping(drugs, drug_to_values, disease_to_index)
elif prediction_type == "side effect":
# For side effect prediction (feature phenotype: disease indication)
disease_to_drugs, pairs, classes = utilities.get_drug_disease_mapping(drugs, drug_to_values_se, se_to_index)
drug_to_values_se = drug_to_values
se_to_index = disease_to_index
elif prediction_type == "drug interaction":
if drug_interaction_to_index is None:
raise ValueError("Drug interaction information is missing!")
# For drug interaction prediction (feature phenotype: side effect)
disease_to_drugs, pairs, classes = utilities.get_drug_disease_mapping(drugs, drug_to_values_interaction, drug_interaction_to_index)
else:
raise ValueError("Uknown prediction_type: " + prediction_type)
list_M_similarity = []
if "phenotype" in features:
drug_to_index, M_similarity_se = utilities.get_similarity(drugs, drug_to_values_se)
list_M_similarity.append(M_similarity_se)
if "chemical" in features:
drug_to_index, M_similarity_chemical = utilities.get_similarity(drugs, drug_to_values_structure)
list_M_similarity.append(M_similarity_chemical)
if "target" in features:
drug_to_index, M_similarity_target = utilities.get_similarity(drugs, drug_to_values_target)
list_M_similarity.append(M_similarity_target)
if output_file is not None:
output_f = open(output_file, 'a')
output_f.write("n_fold\tn_proportion\tn_subset\tmodel type\tprediction type\tfeatures\trecalculate\tdisjoint\tpairwise\tvariable\tauc.mean\tauc.sd\tauprc.mean\tauprc.sd\n")
else:
output_f = None
values = []
values2 = []
for i in xrange(n_run):
if n_seed is not None:
n_seed += i
random.seed(n_seed)
numpy.random.seed(n_seed)
pairs_, classes_, cv = utilities.balance_data_and_get_cv(pairs, classes, n_fold, n_proportion, n_subset, disjoint = disjoint_cv, split_both=split_both, n_seed = n_seed)
val, val2 = check_ml_helper(drugs, disease_to_drugs, drug_to_index, list_M_similarity, pairs_, classes_, cv, knn, n_fold, n_proportion, n_subset, model_type, prediction_type, features, recalculate_similarity, disjoint_cv, split_both, output_f, model_fun, verbose, n_seed)
values.append(val)
values2.append(val2)
print "AUC over runs: %.1f (+/-%.1f):" % (numpy.mean(values), numpy.std(values)), map(lambda x: round(x, ndigits=1), values)
if output_f is not None:
output_f.write("%d\t%d\t%d\t%s\t%s\t%s\t%s\t%s\t%s\t%s\t%f\t%f\t%f\t%f\n" % (n_fold, n_proportion, n_subset, model_type, prediction_type, "|".join(features), recalculate_similarity, disjoint_cv, split_both, "avg", numpy.mean(values), numpy.std(values), numpy.mean(values2), numpy.std(values2)))
output_f.close()
return "AUC: %.1f" % numpy.mean(values), "AUPRC: %.1f" % numpy.mean(values2)
def check_ml(data, n_run, knn, n_fold, n_proportion, n_subset, model_type, prediction_type, features, recalculate_similarity, disjoint_cv, output_file = None, model_fun = None, verbose=False, n_seed = None):
drugs, disease_to_index, drug_to_values, se_to_index, drug_to_values_se, drug_to_values_structure, drug_to_values_target = data
if prediction_type == "disease":
disease_to_drugs, pairs, classes = utilities.get_drug_disease_mapping(drugs, drug_to_values, disease_to_index)
elif prediction_type == "side effect":
# For side effect prediction
disease_to_drugs, pairs, classes = utilities.get_drug_disease_mapping(drugs, drug_to_values_se, se_to_index)
drug_to_values_se = drug_to_values
se_to_index = disease_to_index
else:
raise ValueError("Uknown prediction_type: " + prediction_type)
list_M_similarity = []
if "phenotype" in features:
drug_to_index, M_similarity_se = utilities.get_similarity(drugs, drug_to_values_se)
list_M_similarity.append(M_similarity_se)
if "chemical" in features:
drug_to_index, M_similarity_chemical = utilities.get_similarity(drugs, drug_to_values_structure)
list_M_similarity.append(M_similarity_chemical)
if "target" in features:
drug_to_index, M_similarity_target = utilities.get_similarity(drugs, drug_to_values_target)
list_M_similarity.append(M_similarity_target)
if output_file is not None:
output_f = open(output_file, 'a')
output_f.write("n_fold\tn_proportion\tn_subset\tmodel type\tprediction type\tfeatures\trecalculate\tdisjoint\tvariable\tauc.mean\tauc.sd\tauprc.mean\tauprc.sd\n")
else:
output_f = None
values = []
values2 = []
for i in xrange(n_run):
if n_seed is not None:
n_seed += i
random.seed(n_seed)
numpy.random.seed(n_seed)
pairs_, classes_, cv = utilities.balance_data_and_get_cv(pairs, classes, n_fold, n_proportion, n_subset, disjoint = disjoint_cv, n_seed = n_seed)
val, val2 = check_ml_helper(drugs, disease_to_drugs, drug_to_index, list_M_similarity, pairs_, classes_, cv, knn, n_fold, n_proportion, n_subset, model_type, prediction_type, features, recalculate_similarity, disjoint_cv, output_f, model_fun, verbose, n_seed)
values.append(val)
values2.append(val2)
print "AUC over runs: %.1f (+/-%.1f):" % (numpy.mean(values), numpy.std(values)), map(lambda x: round(x, ndigits=1), values)
if output_f is not None:
output_f.write("%d\t%d\t%d\t%s\t%s\t%s\t%s\t%s\t%s\t%f\t%f\t%f\t%f\n" % (n_fold, n_proportion, n_subset, model_type, prediction_type, "|".join(features), recalculate_similarity, disjoint_cv, "avg", numpy.mean(values), numpy.std(values), numpy.mean(values2), numpy.std(values2)))
output_f.close()
return "AUC: %.1f" % numpy.mean(values), "AUPRC: %.1f" % numpy.mean(values2)
def test_get_similarity_based_scores_knn2(self):
# c1 is used for p1, c2 is used for p2
disease_to_drugs = dict([("p1", set(["c1"])), ("p2", set(["c2"]))])
# Get similarity between all pair of drugs
drug_to_index, M = get_similarity(self.drugs, self.drug_to_values)
list_M_similarity = [ M ]
knn = 2
drug_to_disease_to_scores = get_similarity_based_scores(self.drugs, disease_to_drugs, drug_to_index, list_M_similarity, knn)
# c1's similarity to c2, c3, c4: 0.5, 0.5, -0.5
# c1's score for p2: 0.5*label(c2) + 0.5*label(c3) = 0.5 * 1 + 0 = 0.5
self.assertAlmostEqual(drug_to_disease_to_scores["c1"]["p2"][0], 0.5)
def test_get_similarity(self):
drug_to_index, M = get_similarity(self.drugs, self.drug_to_values)
drug1 = "c1"
# c1-c2 similarity cor([1,0,0], [1,1,0]) == 0.5
drug2 = "c2"
self.assertEqual(M[drug_to_index[drug1]][drug_to_index[drug2]], 0.5)
# c1-c3 similarity cor([1,0,0], [1,0,1]) == 0.5
drug2 = "c3"
self.assertEqual(M[drug_to_index[drug1]][drug_to_index[drug2]], 0.5)
# c1-c4 similarity cor([1,0,0], [0,0,1]) == -0.5
drug2 = "c4"
self.assertEqual(M[drug_to_index[drug1]][drug_to_index[drug2]], -0.5)
def match(self, otherchunk, partialmatching, mismatch_penalty=0):
"""
Check partial match (given bound variables in boundvars).
"""
similarity = 0
if self == otherchunk:
return similarity
# below starts the check that self is proper part of otherchunk. __emptyvalue is ignored. 4 cases have to be checked separately, =x, ~=x, !1, ~!1. Also, variables and their values have to be saved in boundvars. When self is not part of otherchunk the loop adds to (dis)similarity.
for x in self:
try:
matching_val = getattr(otherchunk.actrchunk,
x[0] + "_") # get the value of attr
except AttributeError:
matching_val = None # if it is missing, it must be None
if isinstance(matching_val, utilities.VarvalClass):
matching_val = matching_val.values # the value might be written using _variablesvalues namedtuple; in that case, get it out
varval = utilities.splitting(x[1])
# if otherchunk.typename == "IDENTIFY_LINKS":
# print("xxx:",x)
# print(matching_val)
# print(varval.values)
# print(type(varval.values))
# checking variables, e.g., =x
if varval.variables:
# if matching_val == self.__emptyvalue:
# similarity -= 1 #these two lines would require that variables are matched only to existing values; uncomment if you want that
var = varval.variables
for each in self.boundvars.get("~=" + var, set()):
if each == matching_val:
if partialmatching:
similarity += utilities.get_similarity(
self._similarities, each, matching_val,
mismatch_penalty
) # False if otherchunk's value among the values of ~=x
else:
return False
try:
if self.boundvars["=" + var] != matching_val:
if partialmatching:
similarity += utilities.get_similarity(
self._similarities, self.boundvars["=" + var],
matching_val, mismatch_penalty
) # False if =x does not match otherchunks' value
else:
return False
except KeyError:
self.boundvars.update({
"=" + var: matching_val
}) # if boundvars lack =x, update and proceed
# checking negvariables, e.g., ~=x
if varval.negvariables:
for var in varval.negvariables:
try:
if self.boundvars["=" + var] == matching_val:
if partialmatching:
similarity += utilities.get_similarity(
self._similarities,
self.boundvars["=" + var], matching_val,
mismatch_penalty
) # False if =x does not match otherchunks' value
else:
return False
except KeyError:
pass
self.boundvars.setdefault("~=" + var,
set([])).add(matching_val)
# checking values, e.g., 10 or !10
if varval.values:
val = varval.values
if val != None and val != matching_val: # None is the misssing value of the attribute
if partialmatching:
similarity += utilities.get_similarity(
self._similarities, val, matching_val,
mismatch_penalty)
else:
return False
# checking negvalues, e.g., ~!10
if varval.negvalues:
for negval in varval.negvalues:
if negval == matching_val or (
negval in {self.__emptyvalue, 'None'}
and matching_val == self.__emptyvalue):
if partialmatching:
similarity += utilities.get_similarity(
self._similarities, negval, matching_val,
mismatch_penalty)
else:
return False
if partialmatching:
return similarity
else:
return True