BRLA-DDI: A novel framework for drug-drug interaction extraction

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Drug-drug interaction (DDI) extraction is crucial for identifying adverse drug reactions (ADRs).
• Existing models face challenges with complex sentences and implicit drug relationships.
• There is a need for more accurate and generalizable DDI extraction methods.

Purpose of the Study:

• To present BRLA-DDI, a novel model for enhanced DDI extraction.
• To improve the accuracy and generalization capabilities of DDI extraction models.
• To address limitations in handling complex and implicit drug interactions.

Main Methods:

• The BRLA-DDI model integrates BioBERT-LSTM for feature extraction and Relational Graph Convolutional Network (R-GCN) with multihead attention.
• An innovative loss-attention mechanism combines cross-entropy loss with attention-based regularization.
• A dynamic negative sampling strategy is employed to mitigate the zero-loss issue and enhance robustness.

Main Results:

• BRLA-DDI achieved high performance on the DDI Extraction 2013 dataset with 87.68% precision, 88.06% recall, and 87.87% F1 Score.
• The model demonstrated superior and robust performance on the external TAC 2018 dataset, indicating strong generalizability.
• The proposed methods significantly outperformed existing DDI extraction approaches.

Conclusions:

• The BRLA-DDI model offers a significant advancement in drug-drug interaction extraction.
• The synergistic integration of BioBERT-LSTM, R-GCN, and loss-attention enhances model performance and generalizability.
• The publicly released code and data facilitate further research in biomedical information processing.