def test_age_weight(self):
# Test age based weight distribution and best_match_weighted_average calculation
terms_a = ['HP:0001251', 'HP:0001263', 'HP:0001290',
'HP:0004322'] # ATAX, DD, HYP, SS
terms_b = ['HP:0001263', 'HP:0001249', 'HP:0001290'] # DD, ID, HYP
self.hpo_network = annotate(
self.hpo_network,
self.phenotype_to_diseases,
self.num_diseases_annotated,
self.alt2prim,
ages_distribution_file=self.ages_distribution_file)
age_a = 9.0
age_b = 4.0
# calculate weights based on patients age
weights_a = {
'age': calculate_age_weights(terms_a, age_b, self.hpo_network)
}
weights_b = {
'age': calculate_age_weights(terms_b, age_a, self.hpo_network)
}
# make pairwise scores matrix
df = pd.DataFrame([[4.22595743e-02, 3.92122308e-02, 3.04851573e-04],
[1.07473687e-01, 5.05101655e-01, 3.78305515e-04],
[3.69780479e-04, 3.78305515e-04, 4.64651944e-01],
[4.17139800e-04, 4.12232546e-04, 3.67984322e-04]],
index=pd.Index(terms_a, name='a'),
columns=pd.MultiIndex.from_arrays(
[['score'] * len(terms_b), terms_b],
names=[None, 'b']))
# compute pairwise best match weighted average
score_bmwa = self.scorer.best_match_weighted_average(
df, weights_a, weights_b)
self.assertAlmostEqual(score_bmwa, 0.3741, 4)
# set all weights to 1.0, result should be the same as BMA without weights
weights_a = {'disease_frequency': [1.] * len(terms_a)}
weights_b = {'disease_frequency': [1.] * len(terms_b)}
score_bmwa = self.scorer.best_match_weighted_average(
df, weights_a, weights_b)
self.assertAlmostEqual(score_bmwa, 0.2985, 4)
# test term not in network
terms_a = ['HP:Not_a_term']
weights_a = calculate_age_weights(terms_a, age_b, self.hpo_network)
self.assertEqual(weights_a, [1.0])
# term in network no age
terms_a = ['HP:0000001']
weights_a = calculate_age_weights(terms_a, age_b, self.hpo_network)
self.assertEqual(weights_a, [1.0])
def score(self, record_a, record_b):
"""
Scores the comparison of terms in list A to terms in list B.
:param record_a: record A.
:param record_b: record B.
:return: `float` (comparison score)
"""
if self.summarization_method not in ['BMA', 'BMWA', 'maximum']:
raise ValueError(
'Unsupported summarization method, please choose from BMA, BMWA, or maximum.'
)
# if either set is empty return 0.0
terms_a = record_a['terms']
terms_b = record_b['terms']
if not terms_a or not terms_b:
return 0.0
# calculate weights for record_a and record_b
weights_a = record_a['weights'].copy(
) if record_a['weights'] is not None else []
weights_b = record_b['weights'].copy(
) if record_b['weights'] is not None else []
# set weights
# if we have age of record_a use it to set age weights for record_b
if 'age' in record_a:
weights_b['age'] = calculate_age_weights(record_b['terms'],
record_a['age'],
self.hpo_network)
# if we have age of record_b use it to set age weights for record_a
if 'age' in record_b:
weights_a['age'] = calculate_age_weights(record_a['terms'],
record_b['age'],
self.hpo_network)
term_pairs = itertools.product(terms_a, terms_b)
df = pd.DataFrame(
[(pair[0], pair[1], self.score_hpo_pair_hrss(pair[0], pair[1]))
for pair in term_pairs],
columns=['a', 'b', 'score']).set_index(['a', 'b']).unstack()
if self.summarization_method == 'maximum':
return self.maximum(df)
elif self.summarization_method == 'BMWA' and any(
[weights_a, weights_b]):
return self.best_match_weighted_average(df,
weights_a=weights_a,
weights_b=weights_b)
else:
return self.best_match_average(df)
def score(self, record_a, record_b):
"""
Scores the comparison of terms listed in record A to terms listed in record B.
:param record_a: record A.
:param record_b: record B.
:return: record_a record id, record_b record id, `float` (comparison score)
:rtype: tuple
"""
if self.summarization_method not in ['BMA', 'BMWA', 'maximum']:
raise ValueError('Unsupported summarization method, please choose from BMA, BMWA, or maximum.')
# if either set is empty return 0.0
terms_a = record_a['terms']
terms_b = record_b['terms']
if not terms_a or not terms_b:
return record_a['record_id'], record_b['record_id'], 0.0
if self.scoring_method == 'Jaccard':
intersection = len(list(set(terms_a).intersection(terms_b)))
union = (len(terms_a) + len(terms_b)) - intersection
return record_a['record_id'], record_b['record_id'], float(intersection) / union
elif self.scoring_method == 'word2vec':
in_vocab_terms_a = [x for x in terms_a if x in self.word_vectors.vocab]
in_vocab_terms_b = [x for x in terms_b if x in self.word_vectors.vocab]
if in_vocab_terms_a and in_vocab_terms_b:
return self.word_vectors.n_similarity(in_vocab_terms_a, in_vocab_terms_b)
else:
return record_a['record_id'], record_b['record_id'], 0.0
# calculate weights for record_a and record_b
weights_a = record_a['weights'].copy() if record_a['weights'] is not None else []
weights_b = record_b['weights'].copy() if record_b['weights'] is not None else []
# set weights
# if we have age of record_a use it to set age weights for record_b
if 'age' in record_a:
weights_b['age'] = calculate_age_weights(record_b['terms'], record_a['age'], self.hpo_network)
# if we have age of record_b use it to set age weights for record_a
if 'age' in record_b:
weights_a['age'] = calculate_age_weights(record_a['terms'], record_b['age'], self.hpo_network)
term_pairs = itertools.product(terms_a, terms_b)
df = pd.DataFrame(
[(pair[0], pair[1], self.score_hpo_pair_hrss(pair[0], pair[1]))
for pair in term_pairs],
columns=['a', 'b', 'score']
).set_index(
['a', 'b']
).unstack()
if self.summarization_method == 'maximum':
return record_a['record_id'], record_b['record_id'], self.maximum(df)
elif self.summarization_method == 'BMWA' and any([weights_a, weights_b]):
return record_a['record_id'], record_b['record_id'], self.best_match_weighted_average(df, weights_a=weights_a, weights_b=weights_b)
else:
return record_a['record_id'], record_b['record_id'], self.best_match_average(df)