def main():
seed = 1234
data_name = "cpf"
kg_dp_path = "../data/"
print("Importing dataset files ... ")
train_data_raw = load_kg_file(os.path.join(kg_dp_path, "train.tsv"))
test_data_raw = load_kg_file(os.path.join(kg_dp_path, "test.tsv"))
all_data = np.array(
[[s, p, o]
for s, p, o in np.concatenate([train_data_raw, test_data_raw])])
chemicals_list = set(list(all_data[:, 0]))
effects_list = set(list(all_data[:, 1]))
proteins_list = set(list(all_data[:, 2]))
dataset = KgDataset(name=data_name)
dataset.load_triples(train_data_raw, tag="train")
dataset.load_triples(test_data_raw, tag="test")
nb_rels = dataset.get_rels_count()
nb_ents = dataset.get_rels_count()
train_data = dataset.data["train"]
test_data = dataset.data["test"]
all_data_indices = np.concatenate([train_data, test_data])
fn_known_facts = {k: set() for k in range(nb_rels)}
for s, p, o in all_data_indices:
fn_known_facts[p].add((s, o))
fn_test_dict = {k: [] for k in np.unique(test_data[:, 1])}
for s, p, o in test_data:
fn_test_dict[p].append([s, p, o])
print("Initializing the knowledge graph embedding model... ")
# model pipeline definition
model = ComplEx(seed=seed, verbose=2)
pipe_model = Pipeline([('kge_model', model)])
# set model parameters
model_params = {
'kge_model__em_size': 100,
'kge_model__lr': 0.01,
'kge_model__optimiser': "AMSgrad",
'kge_model__log_interval': 10,
'kge_model__nb_epochs': 100,
'kge_model__batch_size': 5000,
'kge_model__initialiser': 'xavier_uniform',
'kge_model__nb_ents': nb_ents,
'kge_model__nb_rels': nb_rels
}
# add parameters to the model then call fit method
pipe_model.set_params(**model_params)
print("Training ... ")
pipe_model.fit(X=train_data, y=None)
def generate_fn_negatives(fn_idx, neg_data_size):
"""
:param fn_idx:
:param neg_data_size:
:return:
"""
candidate_neg_size = int(neg_data_size * 1.2)
candidate_subs = np.random.randint(0, nb_rels, [candidate_neg_size, 1])
candidate_rel = np.ones([candidate_neg_size, 1]) * fn_idx
candidate_objs = np.random.randint(0, nb_rels, [candidate_neg_size, 1])
candidate_negs = np.concatenate(
[candidate_subs, candidate_rel, candidate_objs], axis=1)
true_negs = []
for s, p, o in candidate_negs:
if (s, o) not in fn_known_facts[rel_idx]:
true_negs.append([s, p, o])
true_negs = np.array(true_negs)
return true_negs[:neg_data_size, :]
rel_results = []
for idx, rel_idx in enumerate(fn_test_dict):
rel_name = dataset.get_rel_labels([rel_idx])[0]
rel_test_data_pos = np.array(fn_test_dict[rel_idx])
rel_test_size = len(rel_test_data_pos)
rel_pos_scores = pipe_model.predict(rel_test_data_pos)
res = {
1: {
"auroc": 0.0,
"aupr": 0.0
},
10: {
"auroc": 0.0,
"aupr": 0.0
},
50: {
"auroc": 0.0,
"aupr": 0.0
}
}
for neg_ratio in [1, 10, 50]:
rel_test_data_neg = generate_fn_negatives(
rel_idx, rel_test_size * neg_ratio)
neg_scores = []
for rel_test_data_neg_batch in generate_batches(rel_test_data_neg,
batch_size=10000):
batch_scores = pipe_model.predict(rel_test_data_neg_batch)
neg_scores.extend(batch_scores)
rel_neg_scores = np.array(neg_scores)
rel_all_scores = np.concatenate([rel_pos_scores, rel_neg_scores])
rel_all_labels = np.concatenate([
np.ones([
len(rel_pos_scores),
]),
np.zeros([
len(rel_neg_scores),
])
])
rel_aupr = auc_pr(rel_all_labels, rel_all_scores)
rel_auroc = auc_roc(rel_all_labels, rel_all_scores)
res[neg_ratio]["aupr"] = rel_aupr
res[neg_ratio]["auroc"] = rel_auroc
print(
"[%d] N1:AUROC %1.4f - N1:AUPR %1.4f\tN10:AUROC %1.4f - N10:AUPR %1.4f\tN50:AUROC %1.4f - N50:AUPR %1.4f"
"\t[Count: %d]\tREL:%s" %
(idx + 1, res[1]["auroc"], res[1]["aupr"], res[10]["auroc"],
res[10]["aupr"], res[50]["auroc"], res[50]["aupr"], rel_test_size,
rel_name))
rel_results.append([
res[1]["auroc"], res[1]["aupr"], res[10]["auroc"], res[10]["aupr"],
res[50]["auroc"], res[50]["aupr"]
])
rel_results = np.array(rel_results)
n1_au_roc = np.mean(rel_results[:, 0])
n1_au_pr = np.mean(rel_results[:, 1])
n3_au_roc = np.mean(rel_results[:, 2])
n3_au_pr = np.mean(rel_results[:, 3])
n10_au_roc = np.mean(rel_results[:, 4])
n10_au_pr = np.mean(rel_results[:, 5])
print(
"-----------------------------------------------------------------------------------------------------------"
)
print(
"N1:AUROC %1.4f - N1:AUPR %1.4f\tN10:AUROC %1.4f - N10:AUPR %1.4f\tN50:AUROC %1.4f - N50:AUPR %1.4f = [AVERAGE]"
% (n1_au_roc, n1_au_pr, n3_au_roc, n3_au_pr, n10_au_roc, n10_au_pr))
print(
"-----------------------------------------------------------------------------------------------------------"
)
def main():
seed = 1234
nb_epochs_then_check = None
data_name = "pse"
kg_dp_path = "../data/"
se_map_raw = [
l.strip().split("\t")
for l in open(os.path.join(kg_dp_path, "se_maps.txt")).readlines()
]
se_mapping = {k: v for k, v in se_map_raw}
print("Importing dataset files ... ")
benchmark_train_fd = gzip.open(
os.path.join(kg_dp_path, "ploypharmacy_facts_train.txt.gz"), "rt")
benchmark_valid_fd = gzip.open(
os.path.join(kg_dp_path, "ploypharmacy_facts_valid.txt.gz"), "rt")
benchmark_test_fd = gzip.open(
os.path.join(kg_dp_path, "ploypharmacy_facts_test.txt.gz"), "rt")
benchmark_train = np.array(
[l.strip().split() for l in benchmark_train_fd.readlines()])
benchmark_valid = np.array(
[l.strip().split() for l in benchmark_valid_fd.readlines()])
benchmark_test = np.array(
[l.strip().split() for l in benchmark_test_fd.readlines()])
benchmark_triples = np.array(
[[d1, se, d2] for d1, se, d2 in np.concatenate(
[benchmark_train, benchmark_valid, benchmark_test])])
pse_drugs = list(
set(
list(
np.concatenate(
[benchmark_triples[:, 0], benchmark_triples[:, 2]]))))
pse_list = set(list(benchmark_triples[:, 1]))
rel_dict = dict()
for s, p, o in benchmark_triples:
if p not in rel_dict:
rel_dict[p] = 1
else:
rel_dict[p] += 1
pair_dict = dict()
for s, p, o in benchmark_triples:
if s > o:
pair = (s, o)
else:
pair = (o, s)
if pair not in rel_dict:
pair_dict[pair] = 1
else:
pair_dict[pair] += 1
drug_combinations = np.array(
[[d1, d2] for d1, d2 in list(itertools.product(pse_drugs, pse_drugs))
if d1 != d2])
print("Processing dataset files to generate a knowledge graph ... ")
# delete raw polypharmacy data
del benchmark_triples
dataset = KgDataset(name=data_name)
dataset.load_triples(benchmark_train, tag="bench_train")
dataset.load_triples(benchmark_valid, tag="bench_valid")
dataset.load_triples(benchmark_test, tag="bench_test")
del benchmark_train
del benchmark_valid
del benchmark_test
nb_entities = dataset.get_ents_count()
nb_relations = dataset.get_rels_count()
pse_indices = dataset.get_rel_indices(list(pse_list))
d1 = np.array(dataset.get_ent_indices(list(
drug_combinations[:, 0]))).reshape([-1, 1])
d2 = np.array(dataset.get_ent_indices(list(
drug_combinations[:, 1]))).reshape([-1, 1])
drug_combinations = np.concatenate([d1, d2], axis=1)
del d1
del d2
# grouping side effect information by the side effect type
train_data = dataset.data["bench_train"]
valid_data = dataset.data["bench_valid"]
test_data = dataset.data["bench_test"]
bench_idx_data = np.concatenate([train_data, valid_data, test_data])
se_facts_full_dict = {se: set() for se in pse_indices}
for s, p, o in bench_idx_data:
se_facts_full_dict[p].add((s, p, o))
print("Initializing the knowledge graph embedding model... ")
# model pipeline definition
model = TransE(seed=seed, verbose=2)
pipe_model = Pipeline([('kge_model', model)])
# set model parameters
model_params = {
'kge_model__em_size': 100,
'kge_model__lr': 0.01,
'kge_model__optimiser': "AMSgrad",
'kge_model__log_interval': 10,
'kge_model__nb_epochs': 100,
'kge_model__nb_negs': 6,
'kge_model__batch_size': 5000,
'kge_model__initialiser': 'xavier_uniform',
'kge_model__nb_ents': nb_entities,
'kge_model__nb_rels': nb_relations
}
# add parameters to the model then call fit method
pipe_model.set_params(**model_params)
print("Training ... ")
pipe_model.fit(X=train_data, y=None)
metrics_per_se = {
se_idx: {
"ap": .0,
"auc-roc": .0,
"auc-pr": .0,
"p@50": .0
}
for se_idx in pse_indices
}
se_ap_list = []
se_auc_roc_list = []
se_auc_pr_list = []
se_p50_list = []
print(
"================================================================================"
)
for se in tqdm(pse_indices,
desc="Evaluating test data for each side-effect"):
se_name = dataset.get_rel_labels([se])[0]
se_all_facts_set = se_facts_full_dict[se]
se_test_facts_pos = np.array([[s, p, o] for s, p, o in test_data
if p == se])
se_test_facts_pos_size = len(se_test_facts_pos)
se_test_facts_neg = np.array(
[[d1, se, d2] for d1, d2 in drug_combinations
if (d1, se,
d2) not in se_all_facts_set and (d2, se,
d1) not in se_all_facts_set])
# shuffle and keep negatives with size equal to positive instances so positive to negative ratio is 1:1
np.random.shuffle(se_test_facts_neg)
se_test_facts_neg = se_test_facts_neg[:se_test_facts_pos_size, :]
set_test_facts_all = np.concatenate(
[se_test_facts_pos, se_test_facts_neg])
se_test_facts_labels = np.concatenate([
np.ones([len(se_test_facts_pos)]),
np.zeros([len(se_test_facts_neg)])
])
se_test_facts_scores = model.predict(set_test_facts_all)
se_ap = average_precision(se_test_facts_labels, se_test_facts_scores)
se_p50 = precision_at_k(se_test_facts_labels,
se_test_facts_scores,
k=50)
se_auc_pr = auc_pr(se_test_facts_labels, se_test_facts_scores)
se_auc_roc = auc_roc(se_test_facts_labels, se_test_facts_scores)
se_ap_list.append(se_ap)
se_auc_roc_list.append(se_auc_roc)
se_auc_pr_list.append(se_auc_pr)
se_p50_list.append(se_p50)
se_code = se_name.replace("SE:", "")
metrics_per_se[se] = {
"ap": se_ap,
"auc-roc": se_auc_roc,
"auc-pr": se_auc_pr,
"p@50": se_p50
}
print(
"AP: %1.4f - AUC-ROC: %1.4f - AUC-PR: %1.4f - P@50: %1.4f > %s: %s"
% (se_ap, se_auc_roc, se_auc_pr, se_p50, se_code,
se_mapping[se_code]),
flush=True)
se_ap_list_avg = np.average(se_ap_list)
se_auc_roc_list_avg = np.average(se_auc_roc_list)
se_auc_pr_list_avg = np.average(se_auc_pr_list)
se_p50_list_avg = np.average(se_p50_list)
print(
"================================================================================"
)
print(
"[AVERAGE] AP: %1.4f - AUC-ROC: %1.4f - AUC-PR: %1.4f - P@50: %1.4f" %
(se_ap_list_avg, se_auc_roc_list_avg, se_auc_pr_list_avg,
se_p50_list_avg),
flush=True)
print(
"================================================================================"
)
def main():
seed = 1234
nb_epochs_then_check = None
data_name = "TS-PROTEIN-GO"
dataset_dir = "../data/dataset/"
# loading dataset
train_fp = os.path.join(dataset_dir, "train.txt.gz")
train_facts_labeled = [
l.strip().split("\t") for l in gzip.open(train_fp, "rt").readlines()
]
train_facts = np.array([[s, p, o] for s, p, o, f in train_facts_labeled
if f == "1"])
# train_facts_neg = np.array([[s, p, o] for s, p, o, f in train_facts_labeled if f == "0"])
test_fp = os.path.join(dataset_dir, "test.txt.gz")
test_facts_labeled = [
l.strip().split("\t") for l in gzip.open(test_fp, "rt").readlines()
]
test_facts = np.array([[s, p, o] for s, p, o, f in test_facts_labeled
if f == "1"])
test_facts_neg = np.array([[s, p, o] for s, p, o, f in test_facts_labeled
if f == "0"])
tissue_list = list(set(list(test_facts[:, 1])))
dataset = KgDataset(name=data_name)
dataset.load_triples(train_facts, "train")
dataset.load_triples(test_facts, "test")
dataset.load_triples(test_facts_neg, "test_neg")
del train_facts
del test_facts
del test_facts_neg
train_data = dataset.data["train"]
test_data = dataset.data["test"]
test_data_neg = dataset.data["test_neg"]
tissue_list = dataset.get_rel_indices(tissue_list)
# model pipeline definition
model = TriModel(seed=seed, loss="pt_log", verbose=2)
pipe_model = Pipeline([('kge_model', model)])
# set model parameters
model_params = {
'kge_model__em_size': 30,
'kge_model__lr': 0.01,
'kge_model__nb_negs': 2,
'kge_model__nb_epochs': 200,
'kge_model__batch_size': 4000,
'kge_model__nb_ents': dataset.get_ents_count(),
'kge_model__nb_rels': dataset.get_rels_count()
}
# add parameters to the model then call fit method
pipe_model.set_params(**model_params)
pipe_model.fit(X=train_data)
ts_auc_roc_list = []
ts_auc_pr_list = []
print("============================================================")
print("= Tissue-specific evaluation =")
print("============================================================")
for tissue_idx in tissue_list:
tissue_name = dataset.get_rel_labels([tissue_idx])[0]
ts_test_facts_pos = np.array([[s, p, o] for s, p, o in test_data
if p == tissue_idx])
ts_test_facts_neg = np.array([[s, p, o] for s, p, o in test_data_neg
if p == tissue_idx])
set_test_facts_all = np.concatenate(
[ts_test_facts_pos, ts_test_facts_neg])
se_test_facts_labels = np.concatenate([
np.ones([len(ts_test_facts_pos)]),
np.zeros([len(ts_test_facts_neg)])
])
se_test_facts_scores = model.predict(set_test_facts_all)
se_auc_pr = average_precision_score(se_test_facts_labels,
se_test_facts_scores)
se_auc_roc = roc_auc_score(se_test_facts_labels, se_test_facts_scores)
ts_auc_roc_list.append(se_auc_roc)
ts_auc_pr_list.append(se_auc_pr)
print("= AUC-ROC: %1.4f - AUC-PR: %1.4f > %s" %
(se_auc_roc, se_auc_pr, tissue_name),
flush=True)
se_auc_roc_list_avg = np.average(ts_auc_roc_list)
se_auc_pr_list_avg = np.average(ts_auc_pr_list)
print("============================================================")
print("= AUC-ROC: %1.4f - AUC-PR: %1.4f > [AVERAGE]" %
(se_auc_roc_list_avg, se_auc_pr_list_avg),
flush=True)
print("============================================================")