Source-linked AI summary
OPA2Vec: combining formal and informal content of biomedical ontologies to improve similarity-based prediction
Fatima Zohra Smaili, Xin Gao, Robert Hoehndorf
TL;DR
OPA2Vec addresses the limited use of ontology meta-data by combining formal logical content with annotation axioms to represent biomedical entities as vectors. Across protein-interaction and gene-disease-association prediction, it significantly improves predictive performance over Onto2Vec and traditional semantic similarity measures.
Problem
Biomedical ontologies are logical theories whose annotation meta-data are valuable but generally underused by ontology-based analysis methods.
Method
OPA2Vec combines asserted and inferred logical axioms with annotation axioms to generate dense vector representations for named ontology entities.
Results
OPA2Vec significantly outperforms Onto2Vec and traditional semantic similarity measures in predicting protein-protein interactions and gene-disease associations.
Takeaways & Limitations
OPA2Vec can encode ontology information comprehensively and support applications using biomedical ontologies, with potential to include or exclude different annotation properties.
Abstract
from arXiv · showhide
Motivation: Ontologies are widely used in biology for data annotation, integration, and analysis. In addition to formally structured axioms, ontologies contain meta-data in the form of annotation axioms which provide valuable pieces of information that characterize ontology classes. Annotations commonly used in ontologies include class labels, descriptions, or synonyms. Despite being a rich source of semantic information, the ontology meta-data are generally unexploited by ontology-based analysis methods such as semantic similarity measures. Results: We propose a novel method, OPA2Vec, to generate vector representations of biological entities in ontologies by combining formal ontology axioms and annotation axioms from the ontology meta-data. We apply a Word2Vec model that has been pre-trained on PubMed abstracts to produce feature vectors from our collected data. We validate our method in two different ways: first, we use the obtained vector representations of proteins as a similarity measure to predict protein-protein interaction (PPI) on two different datasets. Second, we evaluate our method on predicting gene-disease associations based on phenotype similarity by generating vector representations of genes and diseases using a phenotype ontology, and applying the obtained vectors to predict gene-disease associations. These two experiments are just an illustration of the possible applications of our method. OPA2Vec can be used to produce vector representations of any biomedical entity given any type of biomedical ontology. Availability: https://github.com/bio-ontology-research-group/opa2vec Contact: robert.hoehndorf@kaust.edu.sa and xin.gao@kaust.edu.sa.
1 Introduction
Biomedical ontologies formally represent biological knowledge and support integration and analysis, but also contain human-oriented metadata such as labels, descriptions, and synonyms. OPA2Vec extends ontology-based vector representations by combining logical content with this metadata and is evaluated for protein interactions and gene-disease associations.
- Biomedical ontologies provide explicit, shared specifications that integrate biological data across disparate sources.
- Ontologies encode biological knowledge through logical axioms and are widely used to capture information extracted from literature.
- Ontology metadata include labels, descriptions, and synonyms intended primarily for human use.
- Existing ontology applications include semantic similarity measures and knowledge graph embeddings for biomedical analysis.
- OPA2Vec extends Onto2Vec by jointly representing biomedical entities from logical axioms and ontology metadata.
- The method is evaluated with Gene Ontology for protein interactions and PhenomeNET for gene-disease associations, outperforming Onto2Vec and traditional semantic similarity measures.
2 Results
OPA2Vec extends ontology-based vectorization by combining formal logical axioms with annotation metadata and biomedical language semantics. Across protein-interaction and gene-disease prediction tasks, these representations support improved similarity-based prediction.
- 2.1 Encoding ontologies plus annotations as vectors: OPA2Vec combines asserted and inferred logical axioms with annotation axioms to generate dense vectors for ontology entities.It extends Onto2Vec by incorporating metadata such as labels, descriptions, synonyms, and creators.
- 2.1 Encoding ontologies plus annotations as vectors: Annotation axioms are converted into sentences and combined with logical-axiom sentences before applying a Word2Vec skipgram model.A biomedical Word2Vec model pretrained on PubMed supplies semantics for natural-language annotation values.
- 2.1 Encoding ontologies plus annotations as vectors: The OPA2Vec pipeline accepts an OWL ontology and entity-class associations, producing vectors that encode both logical axioms and metadata.The method is illustrated as a feature-vector generation workflow for entities in the ontology and associated instances.
- 2.2 OPA2Vec performance in predicting interactions between proteins: Protein-interaction prediction uses GO-based protein vectors, cosine similarity, and optionally a neural network trained to predict interactions.The evaluation compares OPA2Vec with Onto2Vec and Resnik-based semantic similarity using protein interactions characterized in STRING.
- 2.2 OPA2Vec performance in predicting interactions between proteins: Natural-language descriptions and class labels produce the largest performance improvements among the evaluated annotation properties.Creator metadata adds a minor improvement, whereas creation date and namespace do not improve prediction.
- 2.3 Evaluating performance in predicting gene-disease associations: In the PhenomeNET use case, gene and disease vectors are compared by cosine similarity and used with a supervised neural network for gene-disease prediction.The knowledge base includes mouse genes and human diseases associated with phenotypes; the evaluation uses two gene-disease datasets.
3 Conclusion
OPA2Vec combines formal logical content with ontology metadata and natural-language descriptions to generate entity representations. Its applications show improved performance for semantic-similarity-based protein interaction and gene-disease association prediction, while supporting flexible ontology analysis.
- OPA2Vec combines formal logical content, metadata, and natural-language descriptions to produce vector representations of biological entities.
- The method was applied to Gene Ontology and PhenomeNET for semantic-similarity-based prediction tasks.
- Transfer learning with a Word2Vec model pretrained on biomedical literature significantly improved performance in protein-protein interaction and gene-disease association prediction.
- OPA2Vec can include or exclude any annotation property and is based on OWL standards, supporting applications that use biomedical ontologies.
4.1 Ontology and annotation resources
The study assembled Gene Ontology and PhenomeNET resources with associated biological annotations for evaluation. Ontology processing used OWL API and the ELK OWL reasoner.
- Gene Ontology was downloaded in OWL format, and protein annotations were obtained from UniProt-GOA.Annotations with evidence codes IEA and ND were removed; STRING supplied human and yeast protein-interaction data for validation.
- PhenomeNET was downloaded from AberOWL, while mouse phenotype annotations were obtained from the Mouse Genome Informatics database.The study reports 302,013 unique mouse phenotype annotations.
- The Gene Ontology and PhenomeNET resources were processed with OWL API 4.2.6 and the ELK OWL reasoner.
4.2 PubMed
The study collected a large PubMed abstract corpus for biomedical text processing and used figures to report annotation-property and gene-disease prediction evaluations.
- The PubMed corpus contained 28,189,045 abstracts collected in MEDLINE format.Publication metadata were removed, retaining each article’s title and abstract text.
- Figure 4 reports the contribution of ontology annotation properties to protein-protein interaction prediction for human and yeast.
- Figure 5 reports ROC curves and AUC values for gene-disease association prediction across methods in human and mouse.
4.3 Word2Vec
OPA2Vec processes ontology and biomedical-text corpora with Word2Vec to generate entity vectors. The implementation uses skip-gram training, PubMed pretraining, and subsequent retraining on ontology data.
- The study used ontologies, entity annotations, and PubMed abstracts as text corpora for Word2Vec processing.
- The skip-gram model was selected because it produces better vectors for infrequent words than CBOW.This was useful because the biological entities represented in the corpora may occur infrequently.
- Word2Vec was pretrained on PubMed abstracts and then retrained on the Gene Ontology and PhenomeNET data.Grid search was used to optimize skip-gram parameters, with min_count changed from 25 for PubMed training to 1 before ontology training.
- The Word2Vec training parameters are reported in Table 1.
4.4 Similarity
The paper compares biological entities by combining similarities across their ontology annotations. It uses cosine similarity for vector representations and ontology-based semantic similarity, aggregated across multiple annotations with Best Match Average.
- Cosine Similarity: Cosine similarity measures the cosine angle between two Word2Vec vectors using their dot product and magnitudes.The vectors are A and B, and A · B denotes their dot product.
- Semantic similarity: Resnik similarity quantifies the information content of the most informative common ancestor of two ontology concepts.Information content is defined as −log p(c), where p(c) is the probability of encountering concept c.
- Semantic similarity: Resnik similarity avoids assuming that all ontology relations represent uniform distances because edge values can vary.The measure remains conceptually simple while accounting for differing relational information.
- Semantic similarity: Best Match Average aggregates similarities between entities whose annotations contain multiple ontology concepts.It averages the best matching terms between the two entities’ annotation sets, using Resnik or another semantic similarity measure.
4.5 Supervised Learning
The authors train neural-network models to improve PPI and gene-disease association prediction from entity representations. Training and testing use separate protein or association sets, with positives drawn from reference databases and negatives sampled or defined as non-associations.
- Prediction models: Neural networks were trained to improve prediction performance for protein-protein interactions and gene-disease associations.The PPI models used separate yeast and human protein training and testing sets.
- Prediction models: For gene-disease prediction, mouse and human associations were divided into separate training and testing sets.The positive pairs came from MGI_DO.rpt, while all other associations were treated as negative.
4.6 Evaluation using ROC curve and AUC
The evaluation uses ROC curves to assess classification performance and AUC to support quantitative comparisons between methods. ROC curves plot sensitivity against false-positive rate, with better models closer to the ideal curve.
- ROC evaluation: ROC curves plot true-positive rate against false-positive rate to evaluate PPI and gene-disease prediction models.TPR is sensitivity, while FPR is 1−specificity.
- ROC evaluation: For PPI evaluation, true positives are STRING interactions predicted as interacting, while false positives and true negatives are defined from STRING membership and predictions.The definitions use protein pairs’ presence or absence in the STRING network.
- ROC evaluation: A perfect ROC curve connects (0, 0), (0, 1), and (1, 1), and curves closer to it indicate better classification.Visual comparison alone can be insufficient when methods’ ROC curves overlap.
- AUC evaluation: AUC summarizes a ROC curve by integrating it over the entire false-positive-rate axis.It provides a quantitative measure for formal comparison between methods.
Funding
The research was supported by multiple awards from the King Abdullah University of Science and Technology Office of Sponsored Research.
- Funding: The study received funding from the KAUST Office of Sponsored Research through multiple listed award numbers.The cited awards include FCC/1/1976-04, FCC/1/1976-06, URF/1/2602-01, URF/1/3007-01, URF/1/3412-01, URF/1/3450-01, and URF/1/3454-01.