Pharmaguard: AI-Driven Prediction of Drug-Drug Interactions and Side Effects using LLMs

Abstract

Drug-Drug Interactions (DDIs) continue to pose sig- nificant challenges in pharmacological safety, particularly when multiple medicines are prescribed simultaneously. The growing scale of biomedical data has created new opportunities to identify such interactions computationally, yet existing approaches often lack interpretability and adaptability to unseen drug pairs. This study presents a customized hybrid framework that combines the biomedical understanding of BioBERT with the reasoning capabilities of GPT to predict and explain potential DDIs. The model was developed and evaluated using the Mendeley Medicine Information Dataset (MID), which provides both structured and descriptive clinical information. BioBERT was fine-tuned on this dataset to capture drug-specific contextual embeddings, while GPT was prompted to generate short, clinically meaningful explanations for each predicted interaction. The proposed system achieved improved accuracy across multiple evaluation metrics, with an observed ROC-AUC of 0.91 and precision-recall balance outperforming classical and transformer-based baselines. In ad- dition, the explanation component produced outputs that phar- macists rated as factually reliable and contextually useful. The findings suggest that integrating domain-tuned representation learning with generative reasoning can contribute meaningfully to safe prescription analysis and clinical decision support.