Onto2Vec: joint vector-based representation of biological entities and their ontology-based annotations

1 Introduction

• Ontologies formally structure classes and relations, while annotations associate biological entities with ontology classes and may include evidence and source metadata.Ontologies are used across biological databases, web services, and file formats as semantic metadata.Source: §1 · pp. 1–2
• Ontology-based semantic similarity measures compare ontology classes, sets of classes, or entities annotated with ontology classes.Most measures treat ontologies as graphs, although not all axioms can naturally be represented that way.Source: §1 · p. 2
• Existing binary annotation vectors and semantic-similarity features do not retain explicit ontology structure and dependencies for machine learning.Semantic-similarity features reduce ontology information to a single similarity value, while binary vectors no longer retain the ontology structure.Source: §1 · p. 2
• Semantic-similarity measures are expert-designed, and different measures perform well on some datasets and tasks but worse on others, without clear superiority across multiple tasks.The introduction notes that data-driven training may help define a common measure when labelled outcomes are available.Source: §1 · p. 2
• Onto2Vec jointly represents biological entities, their ontology-based annotations, and the ontology structure for applications including protein-interaction prediction, interaction-type classification, and clustering.The introduction also says its implementation is freely available.Source: §1 · pp. 2–3

2.1 Onto2Vec

• Reasoning over OWL axioms supplies a corpus for skip-gram class vectors, which can represent annotated entities through joint annotation axioms or combinations of class vectors.The GO application included proteins with manual UniProt annotations.Source: §2.1 · p. 3

2.2 Similarity-based prediction of biological relations

• Resnik outperformed all evaluated unsupervised methods for PPI prediction, while joint Onto2Vec representations ranked second.The comparison covered human and yeast datasets.Source: Table 1 · p. 4, §2.2 · p. 5
• Supervised Onto2Vec models outperformed predefined similarity measures; SVMs and neural networks exceeded them, while trained scores correlated more strongly with STRING confidence scores.Logistic regression performed comparably to Resnik.Source: §2.2 · p. 6, Table 2 · p. 7
• Multiclass SVMs and neural networks using Onto2Vec vectors distinguished STRING interaction types, whereas fixed similarity measures could not; Onto2Vec also outperformed sparse binary vectors.The interaction types were reaction, activation, binding, and catalysis.Source: Table 3 · p. 7, §2.2 · p. 7

3.1 Ontologies as graphs and axioms

• Onto2Vec learns from arbitrary OWL axioms and deductive closure, including complex class expressions, disjointness, and property restrictions that graph-only methods mostly ignore.The authors describe this as the first approach to apply feature learning to arbitrary OWL axioms in biomedical ontologies while incorporating deductive closure.Source: §3.1 · p. 9

3.2 Towards “trainable” semantic similarity measures

• Cosine similarity between Onto2Vec protein vectors almost matches an established semantic similarity measure.The representations can also support trainable, problem- and dataset-specific similarity measures.Source: §3.2 · p. 9
• Trainable similarity measures can identify the importance of classes and class combinations for a predictive task and significantly improve predictive performance.Source: §3.2 · p. 9
• Supervised similarity measures over Onto2Vec vectors outperform binary representations and can learn how ontology structure contributes, potentially differently across ontology branches.This data-driven approach combines ontology-based annotations and ontology structure in a single representation.Source: §3.2 · p. 10

• Onto2Vec combines neural and symbolic methods in biology and demonstrates significant improvement over state-of-the-art methods.Source: §4 · p. 10
• Onto2Vec adds a method for semantic analysis of ontology-annotated datasets to biology’s limited knowledge-driven analysis toolbox.Source: §4 · p. 10
• The demonstrated application is improved protein–protein interaction prediction; experiments with other ontologies may identify further application areas.The paper anticipates that future neural-symbolic research will extend these results and enable more comprehensive analysis of symbolic representations in biology and biomedicine.Source: §4 · p. 10

5.1 Data set

• GO and UniProt-GOA data were collected, automatically assigned IEA and ND annotations were filtered, and yeast and human STRING interaction datasets were assembled.The study also obtained EC-number annotations for 10,000 proteins from Expasy.Source: §5.1 · p. 10

5.2 Automated reasoning

• HermiT inferred and materialized 80,133 new subsumption, equivalence, and disjointness axioms from GO.The processed GO contained 577,454 logical axioms and 43,828 classes.Source: §5.2 · p. 10

5.3 Representation learning using Word2Vec

• Onto2Vec processes ontology axioms as sentences and uses skip-gram Word2Vec to learn vectors, selected to represent infrequent biological entities.The reported settings include 200-dimensional vectors and five iterations; parameters were selected through limited grid search.Source: §5.3 · p. 11

5.4 Similarity

• Onto2Vec vector similarity uses cosine, while the Resnik baseline scores ontology classes through their most informative common ancestor and BMA compares sets of annotated classes.The BMA comparison can use Resnik or another semantic similarity measure.Source: §5.4 · pp. 11–12

5.5 Supervised Learning

• Logistic regression, linear-kernel SVMs, and ANNs use protein-pair feature vectors to predict interactions, with multi-class SVMs and ANNs also trained to predict interaction types.The interaction datasets were split 70%/30% for training and testing, with negative pairs randomly sampled to match positive-set sizes.Source: §5.5 · p. 12

5.6 Evaluation

• Protein-interaction predictions were evaluated with ROC curves and ROCAUC, using STRING-reported pairs as positives and absent pairs as negatives.The passage defines TP, FP, and TN for this evaluation.Source: §5.6 · pp. 12–13

5.7 Clustering and Visualization

• Protein vectors were visualized with two-dimensional t-SNE and clustered by k-means, with cluster purity measured against six first-level EC families using six clusters.The t-SNE analysis used perplexity 30.Source: §5.7 · p. 13

Funding

• The research was supported by KAUST Office of Sponsored Research under four listed award numbers.The awards are URF/1/1976-04, URF/1/1976-06, URF/1/3450-01-01, and URF/1/3454-01-01.Source: Funding · p. 13