Introduction: Adverse drug reactions (ADRs) contribute to up to 400,000 deaths annually in the EU, with some being caused by drug-drug interactions (DDIs). Most existing tools for DDI prediction rely on predefined rules. In this study, we propose an alternative, fully data-driven approach to analyzing ADRs that can be caused by drug-drug interactions.
Methods: We utilized the US FDA Adverse Event Reporting System (FAERS) as the data source for ADRs potentially caused by DDIs. After curating the original database, we obtained approximately 4 million records. Two modeling approaches were applied: i) a Transformer-based model, and ii) finetuned Large Language Models (LLaMA 3 and LLaMA 3.1). For the Transformer-based model, we incorporated additional drug information, including text embeddings for drug descriptions, molecular descriptors, and predicted drug-target interactions from ChEMBL. For the LLMs, we used only standardized drug names as input. All outputs were standardized using MedDRA terminology.
Results: At the System Organ Class level in MedDRA, the Transformer-based model achieved an F1-score of 0.44, while the finetuned LLaMA 3.1 model achieved an F1-score of 0.41. However, manual analysis of the results revealed that the F1-score systematically underestimates predictive performance, highlighting the need for the development of an alternative metric for ADRs.
Conclusions: We demonstrated that both Transformer-based models and finetuned Large Language Models can predict ADRs resulting from potential DDIs. The performance of the LLMs was lower than that of the Transformer-based models, which can leverage more available data. A key advantage of our approach is the explainability of predictions. By using cross-attention scores, we can uncover the internal relationships between drug combinations and their corresponding ADRs.

David Mirk

Appearances