K-RET

K-RET extracts biomedical relations from unstructured text by injecting external biomedical ontologies into a relation extraction framework to improve prediction accuracy and explainability.

Key Features:

• Knowledge Injection: Leverages external biomedical ontologies to augment relation extraction predictions and improve explainability compared with methods relying solely on bidirectional encoders like BERT.
• Handling Diverse Associations: Manages multiple types of biomedical associations and integrates multiple knowledge sources while determining optimal points for knowledge application within the text-mining process.
• Multi-Token Entity Recognition: Processes multi-token biomedical entities to accurately identify complex relationships between biological concepts.
• Corpus-level Validation: Was evaluated on open-access corpora DDI, BC5CDR, and PGR using four different biomedical ontologies to assess performance across relation types.

Scientific Applications:

• Drug–Drug Interaction extraction: Applied to the DDI corpus to extract and characterize drug–drug interactions from text.
• Chemical–Disease and entity relation extraction: Applied to BC5CDR to extract chemical and disease associations and related relations.
• Protein–Gene relation extraction: Applied to PGR to identify relations between proteins and genes from biomedical literature.
• Discovery of novel or missing associations: Supports identification of previously unreported associations to inform studies of biological interactions and disease mechanisms.

Methodology:

Integrates external biomedical ontologies into a relation extraction framework via knowledge injection, processes multi-token entities, integrates multiple knowledge sources and determines optimal insertion points for that knowledge, and was evaluated on the DDI, BC5CDR, and PGR corpora using four biomedical ontologies.