DDI-MuG: Multi-aspect graphs for drug-drug interaction extraction.

Area of Science:

• Biomedical Natural Language Processing
• Computational Pharmacology
• Bioinformatics

Background:

• Drug-drug interactions (DDIs) pose significant risks, necessitating accurate extraction from biomedical literature.
• Existing DDI extraction methods often overlook valuable corpus-level information, limiting their scope.

Purpose of the Study:

• To develop a novel Multi-aspect Graph-based DDI extraction model (DDI-MuG).
• To improve the accuracy and interpretability of DDI extraction by incorporating multi-aspect graph information.

Main Methods:

• Utilized a bio-specific pre-trained language model for contextualized representations.
• Employed two graphs: one for instance-level syntax and another for corpus-wide word co-occurrence.
• Combined drug entity and verb token representations for classification.

Main Results:

• Achieved superior performance on the DDIExtraction-2013 and TAC 2018 datasets.
• Outperformed all twelve compared state-of-the-art models in DDI extraction tasks.

Conclusions:

• DDI-MuG offers a more interpretable approach compared to black-box models by visualizing word relationships.
• This work pioneers the use of multi-aspect graphs for DDI extraction, setting a foundation for future research.