GNN-Based Hierarchy-Aware Embeddings Predict Yeast Gene Deletion Effects
An innovative technique employs graph neural networks (GNNs) enhanced by semantic loss derived from ontologies to generate embeddings of knowledge graphs (KGs) that are aware of hierarchy. This method is utilized to forecast and analyze the impacts of gene deletions in the yeast Saccharomyces cerevisiae. A knowledge graph for yeast, built from community databases and ontology terms, is integrated with low-dimensional box embeddings and GNNs to estimate cell growth following double gene knockouts. Through over 10-fold cross-validation, the predictions yield a mean R² score of 0.360, significantly surpassing baseline results, indicating that high-level qualitative knowledge influences experimental results. Additionally, this method allows for box embeddings in KGs without a prediction task and lays the groundwork for assessing KG modifications.
Key facts
- Method uses GNNs with semantic loss from ontologies for hierarchy-aware KG embeddings.
- Applied to predict and interpret effects of gene deletions in Saccharomyces cerevisiae.
- Yeast KG built from community databases and ontology terms.
- Low-dimensional box embeddings combined with GNNs predict cell growth for double gene knockouts.
- 10-fold cross-validation yields mean R² score of 0.360.
- Predictions significantly higher than baseline comparisons.
- Box embeddings can be learned without a prediction task.
- Box embeddings serve as basis for evaluating KG revisions.
Entities
—