BO-LSTM: classifying relations via long short-term memory networks along biomedical ontologies

Area of Science:

• Biomedical informatics
• Computational biology
• Natural Language Processing

Background:

• Deep learning, specifically recurrent neural networks, shows promise for biomedical text mining.
• Current deep learning methods often overlook valuable domain-specific resources like ontologies.
• Biomedical ontologies formalize knowledge about entities (genes, chemicals, disorders) and offer supplementary data.

Purpose of the Study:

• To propose a novel deep learning model, BO-LSTM, that integrates domain-specific ontologies for improved relation detection and classification in biomedical texts.
• To enhance the performance of biomedical text mining, particularly in data-scarce environments.

Main Methods:

• Developed BO-LSTM, a recurrent neural network utilizing Long Short-Term Memory (LSTM) units.
• Represented entities as sequences of their ontology ancestors.
• Utilized open biomedical ontologies: Chemical Entities of Biological Interest (ChEBI), Human Phenotype Ontology, and Gene Ontology.
• Integrated word embeddings and WordNet with ontology information.

Main Results:

• BO-LSTM significantly improved the F1-score for detecting and classifying drug-drug interactions (DDIs) in a benchmark corpus.
• Performance gains were most notable in document sets with limited annotations.
• An adapted existing DDI extraction model using the ontology-based method outperformed the original model.
• A new corpus of gene-phenotype relations was created and demonstrated BO-LSTM's applicability to other relation types.

Conclusions:

• Domain-specific ontologies are valuable for enhancing deep learning models in biomedical text mining.
• Ontologies effectively mitigate challenges posed by limited labeled data.
• The BO-LSTM model offers a robust approach for relation extraction in the life and health sciences.