Onto2Vec

This example corresponds to the paper Onto2Vec: joint vector-based representation of biological entities and their ontology-based annotations.

This method is an approach to learn numerical representations (embeddings) of (biomedical) ontologies by representing ontology axioms as text sequences and applying an unsupervised learning algorithm such as Word2Vec. Onto2Vec uses an ontology reasoner to infer new axioms as a preprocessing step. The algorithm is tested on the protein-protein interaction task.

For this algorithm, we need three components:

• The reasoner

• The corpus generator

• The Word2Vec model

import mowl
mowl.init_jvm("20g")

from mowl.datasets.builtin import PPIYeastSlimDataset
from mowl.corpus import extract_and_save_axiom_corpus
from mowl.owlapi import OWLAPIAdapter
from mowl.reasoning import MOWLReasoner

from org.semanticweb.elk.owlapi import ElkReasonerFactory
from java.util import HashSet

from gensim.models.word2vec import LineSentence
from gensim.models import Word2Vec

import os

Inferring new axioms

Onto2Vec uses an ontology reasoner to infer new axioms as a preprocessing step. In the original paper, the authors used the HermiT reasoner. For this example, we use the ELK reasoner.

dataset = PPIYeastSlimDataset()

reasoner_factory = ElkReasonerFactory()
reasoner = reasoner_factory.createReasoner(dataset.ontology)
mowl_reasoner = MOWLReasoner(reasoner)

We wrap the reasoner into the MOWLReasoner class in order to use some shortcuts the mOWL provides such as:

• inferring subclass axioms

• inferring equivalent class axioms

• inferring disjoint axioms (not applicable for this example since we use ELK reasoner)

classes = dataset.ontology.getClassesInSignature()
subclass_axioms = mowl_reasoner.infer_subclass_axioms(classes)
equivalent_class_axioms = mowl_reasoner.infer_equivalent_class_axioms(classes)

We can now add the inferred axioms to the ontology.

adapter = OWLAPIAdapter()
manager = adapter.owl_manager

axioms = HashSet()
axioms.addAll(subclass_axioms)
axioms.addAll(equivalent_class_axioms)

manager.addAxioms(dataset.ontology, axioms)

<java object 'org.semanticweb.owlapi.model.parameters.ChangeApplied'>

Generating the corpus and training the model

Now that we have an extended ontology, we can generate the corpus out of it. After that, we can train the Word2Vec model.

extract_and_save_axiom_corpus(dataset.ontology, "onto2vec_corpus.txt")

sentences = LineSentence("onto2vec_corpus.txt")
model = Word2Vec(sentences, vector_size=5, window=2, min_count=1, workers=4)

Cleaning up memory

os.remove("onto2vec_corpus.txt")

Estimated memory usage: 165 MB