Enhancing drug-drug interaction classification by leveraging textual drug arguments.

Background: The accurate identification and classification of drug-drug interactions (DDIs) are critical for ensuring patient safety and optimizing treatment outcomes in modern healthcare. Traditional methods for DDI classification primarily focus on analyzing the chemical properties and pharmacological data of drugs.

Methods: This study introduces a novel methodology that incorporates textual arguments from DrugBank in addition to the conventional reliance on chemical properties. It explored the impact of integrating text embeddings and text-based similarity matrices alongside the existing chemical properties. To achieve this, types and concepts extracted from the Unified Medical Language System (UMLS), as well as entities, were utilized to create similarity matrices as new input features. The impact of these features was evaluated through a series of experiments conducted across various implementation scenarios by utilizing a deep neural network.

Results: The results highlighted the most discriminative feature types and demonstrated the overall contribution of the proposed approach to the existing literature. The code and associated resources for this study are available for download at the following link: https://github.com/kivancbayraktar/enhancing-ddi-classification-via-textual-arguments.