LLMs For drug-Drug interaction prediction using textual drug descriptors

As treatment plans involve more medications, anticipating and preventing drug-drug interactions (DDIs) becomes increasingly important. Such interactions can result in harmful side effects and may reduce therapy effectiveness. Currently, most computational approaches for DDI prediction rely heavily on complex feature engineering and require chemical information to be structured in specific formats to enable accurate detection of potential interactions. This study presents the first investigation of the application of Large Language Models (LLMs) for DDI prediction using drug characteristics expressed solely in free-text form. Specifically, we use SMILES notations, target organisms, and gene associations as inputs in purpose-designed prompts, allowing LLMs to learn the underlying relationships among these descriptors and accordingly predict possible DDIs. We evaluated the performance of 18 distinct LLMs under zero-shot, few-shot, and fine-tuning settings on the DrugBank dataset (version 5.1.12) to identify the most effective paradigm. We then assessed the generalizability of the fine-tuned models on 13 external DDI datasets against well-known machine learning baselines. The results demonstrated that, while zero-shot and few-shot paradigms showed only modest utility, fine-tuned models achieved superior sensitivity while maintaining competitive accuracy and F1-score compared to baselines. Notably, despite its small size, the Phi-3.5 2.7B model attained a sensitivity of 0.978 and an accuracy of 0.919. These findings suggest that computational efficiency and task-specific adaptation are more important than model size in order to capture the complex patterns inherent in drug interactions, and outline a more accessible paradigm for DDI prediction that can be integrated into clinical decision support systems.