Ontologies provide structured vocabularies and relationships essential for organizing biological knowledge, yet their symbolic nature limits integration with modern machine learning methods. Leveraging recent advances in graph neural networks (GNNs) and transformer-based language models, we present on2vec, a toolkit developed during the DBCLS BioHackathon 2025 for generating vector embeddings from OWL ontologies. on2vec integrates structural information from ontology hierarchies with semantic features from textual annotations using HuggingFace Sentence Transformers, producing domain-aware embeddings suitable for downstream biomedical applications and ontology-based reasoning tasks.

Figures

Performance heatmap of ontology-enhanced models across MTEB biomedical clustering tasks
Performance heatmap of ontology-enhanced models across MTEB biomedical clustering tasks
Average model performance across biomedical clustering tasks with standard deviation
Average model performance across biomedical clustering tasks with standard deviation
Comparison of model performance across different architectures and fusion strategies
Comparison of model performance across different architectures and fusion strategies
Hyperparameter importance analysis across experimental configurations
Hyperparameter importance analysis across experimental configurations
Performance breakdown by source ontology showing domain-specific patterns
Performance breakdown by source ontology showing domain-specific patterns

These figures come from the source repository of the preprint; their captions are also in the PDF.