PEDL

PEDL predicts protein–protein associations (PPAs) and extracts supporting text spans from biomedical literature using deep language models and distant supervision to expand discovery of PPAs.

Key Features:

• Deep Language Models: Utilizes deep language models to process biomedical literature and identify PPAs that may be missed by methods reliant solely on manually annotated datasets.
• Distant Supervision: Incorporates distant supervision to expand the training data pool by an order of magnitude relative to conventional approaches, improving robustness in PPA prediction.
• Dataset Utilization: Introduces three curated datasets for PPA prediction and supports evaluation across two subtasks—predicting PPAs between protein pairs and identifying text spans that state these associations.
• Performance Evaluation: Demonstrates superior performance over existing models in both subtasks across all three datasets.
• Expert Validation: Expert evaluation shows the capability to identify PPAs that are absent from major pathway databases.

Scientific Applications:

• Pathway database curation: Enhances the completeness and accuracy of signalling pathway databases by extracting PPAs from biomedical literature.
• Literature-scale PPA extraction: Automates extraction of PPAs from large volumes of biomedical publications to help researchers maintain current and comprehensive PPA datasets.

Methodology:

Training deep language models on large-scale text corpora using distant supervision to predict protein–protein associations, with evaluation on two subtasks (protein-pair PPA prediction and text-span identification) across three curated datasets.