def predict_links(
input_path: str,
num_walks: int,
walk_length: int,
dimension: int,
window_size: int,
) -> Tuple[float, float]:
g2v = Gat2Vec(input_path,
input_path,
label=False,
tr=gat2vec_config.training_ratio)
model = g2v.train_gat2vec(
num_walks,
walk_length,
dimension,
window_size,
output=True,
)
df = get_embeddingDF(model)
labels = pd.read_csv(gat2vec_paths.get_labels_path(input_path),
sep='\t',
header=None,
names=['disease', 'gene', 'label'])
disease_idx = labels.iloc[0, 0]
labels['prob'] = pd.Series([
calculate_prob(df[disease_idx], df[gene_idx])
for gene_idx in labels.loc[:, 'gene']
])
auc = roc_auc_score(labels['label'], labels['prob'])
aps = average_precision_score(labels['label'], labels['prob'])
return auc, aps
def get_prediction_probs_for_entire_set(self, model):
embedding = parsers.get_embeddingDF(model)
embedding = embedding[self.label_ind, :]
log_reg = linear_model.LogisticRegression(solver='lbfgs')
clf = OneVsRestClassifier(log_reg)
clf.fit(embedding, self.labels) # for multi-class classification
probs = clf.predict_proba(embedding)
logger.debug('ROC: %.2f', roc_auc_score(self.labels, probs[:, 1]))
return probs
def evaluate(self, model, label=False, evaluation_scheme="tr"):
embedding = 0
clf = self.get_classifier()
if not label:
embedding = parsers.get_embeddingDF(model)
if evaluation_scheme == "cv":
results = self.evaluate_cv(clf, embedding, 5)
elif evaluation_scheme == "tr" or label:
results = defaultdict(list)
for tr in self.TR:
logger.debug("TR ... %s", tr)
if label:
model = paths.get_embedding_path_wl(
self.dataset_dir, self.output_dir, tr)
if isinstance(model, str):
embedding = parsers.get_embeddingDF(model)
results.update(self.evaluate_tr(clf, embedding, tr))
logger.debug("Training Finished")
df = pd.DataFrame(results)
return df.groupby(axis=0, by="TR").mean()
def predict_links_cv(network, dataset, use_dge, num_walks, walk_length,
dimension, window_size) -> pd.DataFrame:
results = defaultdict(list)
disease_abv = dataset_to_disease_abv(dataset)
do_id = disease_identifiers[disease_abv]
for i in range(10):
for cv_labels in network.write_gat2vec_cv_split(
home_dir=g2v_path,
disease_id=do_id,
filter_pleiotropic_targets=True,
use_dge_data=use_dge):
g2v = Gat2Vec(input_dir=g2v_path,
output_dir=g2v_path,
label=False,
tr=gat2vec_config.training_ratio)
model = g2v.train_gat2vec(
num_walks,
walk_length,
dimension,
window_size,
output=True,
)
model_df = get_embeddingDF(model)
labels = pd.DataFrame(data=[(x, y) for x, y in cv_labels.items()],
columns=['gene', 'label'])
disease_idx = network.get_index_for_disease(do_id)
labels['prob'] = pd.Series([
calculate_prob(model_df[disease_idx], model_df[gene_idx])
for gene_idx in labels.loc[:, 'gene']
])
auc = roc_auc_score(labels['label'], labels['prob'])
aps = average_precision_score(labels['label'], labels['prob'])
results['TR'].append(i)
results['dge'].append(dataset)
results['eval'].append(use_dge)
results['auc'].append(auc)
results['aps'].append(aps)
logger.debug(results)
df = pd.DataFrame(results)
return df.groupby(by=['TR', 'dge', 'eval']).mean().reset_index()