BO-LSTM

BO-LSTM applies recurrent neural networks with long short-term memory units and ontology-based entity representations to detect and classify biomedical relations from text, improving performance when labeled data are limited.

Key Features:

• Integration of Domain-Specific Ontologies: Uses biomedical ontologies such as Chemical Entities of Biological Interest (ChEBI), Human Phenotype, and Gene Ontology, representing each entity as a sequence of its ontology ancestors to provide structured background knowledge.
• Recurrent Neural Network Architecture: Implements recurrent neural networks equipped with long short-term memory (LSTM) units to process sequences and capture long-range dependencies in text.
• Ontology-Based Representation Improves Performance: Combines ontological information with traditional word embeddings and WordNet to improve relation detection and classification F1-scores, particularly on datasets with sparse annotations.
• Application to Drug-Drug Interaction (DDI) Detection: Evaluated on an international challenge corpus comprising 792 drug descriptions and 233 scientific abstracts for DDI detection, demonstrating superior performance relative to existing models.
• Adaptability to Various Relation Types: Adapted to other relation types such as gene–phenotype interactions and evaluated on a newly developed corpus of 228 abstracts annotated with these relations.

Scientific Applications:

• Drug-Drug Interaction Extraction: Extracts and classifies DDIs from drug descriptions and scientific abstracts to support drug-related research.
• Gene–Phenotype Relation Extraction: Identifies gene–phenotype interactions in biomedical abstracts using ontology-enriched representations.
• Relation Extraction with Limited Annotations: Enhances relation extraction accuracy in domains with scarce labeled data by leveraging domain-specific ontologies.
• Broad Life Sciences Research: Supports uncovering complex biological interactions across life sciences through improved text-mining of biomedical entities.

Methodology:

Entities are encoded as sequences derived from their hierarchical positions within ontologies; these ontology-derived sequences are combined with traditional word embeddings and WordNet information and fed into recurrent neural networks with LSTM units for relation detection and classification.