TP-DDI: Transformer-based pipeline for the extraction of Drug-Drug Interactions

Area of Science:

• Biomedical Informatics
• Computational Linguistics
• Pharmacology

Background:

• Drug-Drug Interaction (DDI) extraction from literature is crucial for drug discovery but remains costly and time-consuming.
• Existing machine learning methods for Drug Named Entity Recognition (DNER) and DDI classification require improvement in prediction accuracy.
• Automated DDI extraction can significantly reduce the laborious efforts in the drug development cycle.

Purpose of the Study:

• To develop a novel end-to-end approach for DDI extraction.
• To improve the accuracy of both DNER and DDI classification tasks.
• To leverage Transformer models and biomedical domain pre-trained weights for enhanced DDI extraction.

Main Methods:

• A pipelined approach integrating DNER and DDI classification using the Transformer model architecture.
• Utilizing BioBERT pre-trained weights to incorporate prior biomedical knowledge.
• Successive execution of DNER and DDI classification for overall DDI extraction.

Main Results:

• The proposed TP-DDI approach demonstrated improved performance in both DNER and overall DDI extraction.
• The method achieved state-of-the-art results on the DDI Extraction 2013 corpus.
• Integration of pre-trained weights significantly enhanced the model's predictive capabilities.

Conclusions:

• The TP-DDI approach offers a more accurate and efficient method for DDI extraction.
• This pipeline method effectively addresses the challenges in identifying drug entities and their interactions.
• The study highlights the potential of Transformer models with domain-specific pre-training for biomedical text mining.