RENET2

RENET2 extracts gene-disease associations from full-text biomedical articles using deep learning-based relation extraction to enable comprehensive full-document analysis.

Key Features:

• Deep learning relation extraction: Uses deep learning models to perform relation extraction across entire full-text biomedical articles.
• Section Filtering: Implements Section Filtering to focus extraction on informative sections of articles.
• Ambiguous relations modeling: Models ambiguous relations to capture uncertain or partial gene-disease links.
• Iterative training data expansion: Employs an iterative training data expansion strategy to augment and refine labeled full-text training data.
• Manual curation and evaluation: Evaluated on a manually curated full-text dataset, achieving an F1-score of 72.13%.
• Benchmarking: Outperformed BeFree, DTMiner, BioBERT, and RENET by margins of 23.87%–30.30% in F1-score on the curated dataset.
• Large-scale PMC application: Applied to approximately 1.89 million PubMed Central full-text articles.
• Extraction scale: Identified approximately 3.72 million gene-disease associations from the PMC corpus.
• LitCovid analysis: Applied to LitCovid articles to rank top proteins associated with COVID-19.

Scientific Applications:

• Large-scale relation extraction: Extraction of gene-disease associations from ~1.89 million PubMed Central full-text articles, yielding ~3.72 million associations.
• COVID-19 protein ranking: Analysis of LitCovid literature to rank proteins associated with COVID-19.
• Training data generation and benchmarking: Generation and augmentation of labeled full-text datasets for relation extraction and benchmarking against BeFree, DTMiner, BioBERT, and RENET.

Methodology:

Applies deep learning-based relation extraction to full texts, implements Section Filtering and ambiguous relations modeling, and uses iterative training data expansion; evaluated by F1-score on a manually curated full-text dataset.