Leveraging Textual Drug Information for Effective Drug-Drug Interaction Identification

Identification of drug-drug interaction (DDI) seeks to determine whether two drugs influence each other's mechanisms within the human body. This study focuses on evaluating textual fields for the purpose of identifying DDIs. Specifically, twelve distinct textual fields that describe drug characteristics were utilized, with data obtained from DrugBank. The textual fields that demonstrated a significant individual impact on DDI identification were identified. Subsequently, the combined use of these informative textual fields was assessed through three different models: PI-DDI, PII-DDI, and SI-DDI. These models share a similar underlying architecture, with the primary difference being the stage at which the textual data of the drugs is concatenated. The experiments conducted measured the impact of varying the textual data, revealing that the fields for description, indication, mechanism, and pharmacodynamics yielded the highest F1 scores, with performances of 89%, 89%, 88%, and 88%, respectively. It was observed that these four textual fields are effective in determining whether drugs interact. The PI-DDI and PII-DDI models, which process the textual data of drugs in parallel, achieved satisfactory performance scores. However, the SI-DDI model, which leverages the textual data as a single input, improved model performance.