PhenoTagger

PhenoTagger performs phenotype concept recognition in unstructured biomedical text by combining Human Phenotype Ontology (HPO) dictionary matching with deep learning to identify ontology concepts.

Key Features:

• Hybrid approach: Integrates dictionary-based matching and deep learning classification to balance precision and recall.
• HPO dictionary construction: Builds a comprehensive dictionary using all HPO concepts and synonyms.
• Distant supervision: Automatically generates a distantly supervised training dataset from dictionary matches, removing the need for manual labeling.
• Candidate extraction: Extracts n-grams from input sentences as candidate phrases for concept assignment.
• Deep learning classification: Uses a deep learning model to classify candidate phrases into ontology concept labels.
• Ontology adaptability: Can be retrained on other ontologies such as MEDIC for disease concept recognition.
• Corpus validation: Evaluated on two HPO corpora and the NCBI disease corpus with performance comparable to supervised methods.

Scientific Applications:

• Phenotype concept recognition: Identification of HPO concepts in unstructured biomedical text.
• Ontology-driven annotation: Generation of large-scale distantly supervised annotated datasets for training models using HPO.
• Disease concept recognition: Adaptation to MEDIC for detecting disease concepts and evaluation on the NCBI disease corpus.
• Cross-ontology generalization: Assessment of model transferability across ontologies for biomedical concept recognition.

Methodology:

Constructs a dictionary from HPO concepts and synonyms, uses it to generate a distantly supervised training dataset, extracts n-gram candidate phrases from input sentences, and applies a deep learning model to classify candidates into concept labels; retraining was performed on MEDIC and evaluations used two HPO corpora and the NCBI disease corpus.