PromptSE: drug side effect prediction with LLM-derived pharmacological representations

Abstract Predicting drug-side effect associations is vital for drug discovery and patient safety. Accurate prediction requires high-quality representations of both drugs and side effects. While drug representations have advanced due to rich structured data, side effect information is mostly unstructured, symptom-oriented and heterogeneous texts, making it hard to capture underlying pharmacological mechanisms. We propose prompt-based side effect prediction (PromptSE), a hybrid framework that combines the reasoning power of large language models with the predictive capability of deep learning. PromptSE employs stepwise prompting tailored to the characteristics of side effect texts, moving beyond simple encoding or conventional prompting, to generate pharmacologically relevant representations. These representations are then fed into a deep learning module for drug-side effect prediction. Building on this framework, PromptSE+ extends the prediction module by integrating multi-modal drug information. Rare entity representations are further refined via a custom graph neural network module. Experiments show that PromptSE outperforms non-drug-informed baselines by 9.26% in AUPR, confirming the effectiveness of our representations. PromptSE+ enhances state-of-the-art drug-side effect prediction methods across all metrics, including a 1.81% AUPR gain, demonstrating its compatibility with advanced methods and potential to support safer drug development and more reliable pharmacological research.