LLM-Enhanced Multimodal Framework for Drug-Drug Interaction Prediction.

Area of Science:

• Computational chemistry
• Bioinformatics
• Artificial intelligence in medicine

Background:

• Drug-drug interactions (DDIs) can alter drug efficacy and safety.
• Polypharmacy increases the risk of DDIs, especially in chronic disease patients.
• Accurate and scalable DDI prediction is essential for safe medication management.

Purpose of the Study:

• To develop a multimodal deep learning framework for enhanced DDI prediction.
• To integrate heterogeneous data modalities including chemical structure, biological networks, and pharmacological mechanisms.
• To address the challenge of data heterogeneity in unified DDI modeling.

Main Methods:

• Utilized a multimodal deep learning framework integrating chemical structure, BioBERT embeddings, and CTET proteins.
• Applied a random walk with restart (RWR) algorithm to incorporate indirect biological pathways.
• Employed BioBERT, a domain-specific large language model, for semantic embeddings.

Main Results:

• The fusion of structural features and BioBERT embeddings achieved the highest classification accuracy (0.9655).
• BioBERT embeddings effectively captured subtle pharmacological relationships between drugs.
• The model demonstrated superior performance in predicting potential DDIs.

Conclusions:

• The multimodal deep learning framework significantly improves DDI prediction accuracy.
• BioBERT embeddings are highly valuable for encoding pharmacological semantics and enhancing DDI prediction.
• The framework offers a practical tool for clinical decision-making in polypharmacy.