Enhancing Herbal Medicine-Drug Interaction Prediction Using Large Language Models.

Area of Science:

• Pharmacology
• Bioinformatics
• Artificial Intelligence

Background:

• Drug-herb interactions are crucial for optimizing treatments and advancing personalized medicine.
• Deep learning models for interaction prediction face challenges with data quality and distribution.
• Large language models (LLMs) offer robust solutions due to their extensive knowledge bases.

Purpose of the Study:

• To propose an integrated model for predicting herbal medicine-drug interactions (HDIs).
• To leverage LLMs, one-hot encoding, and variational graph autoencoders (VGAEs) for enhanced prediction accuracy and interpretability.
• To address data quality and distribution issues in existing interaction prediction methods.

Main Methods:

• Utilized LLMs to extract high-quality molecular features from drug SMILES strings.
• Applied one-hot encoding to represent multi-component herbal medicines, improving model interpretability.
• Employed VGAEs to reconstruct herbal medicine-drug interaction graphs and predict unknown interactions.
• Incorporated node degree differentiation to mitigate bias from high-degree nodes in VGAE message passing.

Main Results:

• The proposed model demonstrated significant performance in predicting herbal medicine-drug interactions.
• Experiments validated the effectiveness of individual components, including LLM feature extraction and VGAE graph reconstruction.
• The method successfully addressed challenges related to data quality and uneven distribution.
• Mitigation strategies for high-degree node dominance proved effective in enhancing prediction stability.

Conclusions:

• The integrated LLM-VGAE model offers a powerful new approach for predicting herbal medicine-drug interactions.
• This method holds significant potential for optimizing traditional Chinese medicine formulations and aiding new drug development.
• The findings support the advancement of precision medicine through more accurate interaction predictions.