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

Area of Science:

• Pharmacology
• Computational Biology
• Artificial Intelligence

Background:

• Accurate drug-side effect prediction is essential for drug discovery and patient safety.
• Existing methods struggle with unstructured, heterogeneous side effect data, hindering the capture of pharmacological mechanisms.
• High-quality drug and side effect representations are key to improving prediction accuracy.

Purpose of the Study:

• To develop a novel framework (PromptSE) for predicting drug-side effect associations.
• To generate pharmacologically relevant representations of side effects using stepwise prompting with large language models.
• To enhance prediction by integrating multi-modal drug information and refining rare entity representations (PromptSE+).

Main Methods:

• A hybrid framework combining large language models (LLMs) and deep learning (DL).
• Stepwise prompting tailored to side effect texts for generating enhanced representations.
• PromptSE+ integrates multi-modal drug data and uses graph neural networks (GNNs) for rare entities.

Main Results:

• PromptSE improved prediction performance over non-drug-informed baselines by 9.26% in AUPR.
• PromptSE+ further enhanced state-of-the-art methods, achieving an additional 1.81% AUPR gain.
• The proposed representations were validated as effective for drug-side effect prediction.

Conclusions:

• Prompt-based representations significantly improve drug-side effect prediction accuracy.
• PromptSE+ demonstrates compatibility with advanced methods and potential for safer drug development.
• The framework offers a promising approach for reliable pharmacological research and drug safety.