BBL-GAT: A Novel Method for Drug-Drug Interaction Extraction From Biomedical Literature

Abstract

The identification of Drug-Drug Interactions (DDIs) is crucial for optimizing patient treatment and avoiding adverse reactions. With the rapid growth of biomedical literature, manual screening for DDIs has become impractical. Hence, the demand for automated DDI extraction methods is continuously increasing. Currently, although many methods have been proposed, feature supplementation and the latest GCN-based methods still face the problem of being unable to effectively extract key information. In this paper, we propose BBL-GAT, a novel method combining BioBERT-BiLSTM and Graph Attention Network (GAT), to extract DDIs from biomedical literature. BioBERT is employed for its ability to capture the semantic relationships between complex medical terms and drug names. BiLSTM is utilized to handle bidirectional contextual information, which is essential for understanding the context of drug-disease relationships. GAT dynamically learns the significance of drug nodes in different interactions through attention mechanisms, enhancing the precision of relationship extraction. We evaluated BBL-GAT on the DDI Extraction 2013 dataset and compared it with other popular DDI extraction methods. The experimental results demonstrate that BBL-GAT achieves an precision of 81.76%, a recall of 84.38%, and an F1-score of 82.47%, illustrating its effectiveness and superiority in DDI relationship extraction tasks.

Keywords

• Drug-drug interactions
• deep learning
• BioBERT
• BiLSTM
• GAT