FGSGT-DDI: An LLM-Enhanced Functional Group Semantic Graph Transformer for Drug-Drug Interaction Prediction

Area of Science:

• Computational Chemistry
• Pharmacology
• Artificial Intelligence

Background:

• Drug-drug interaction (DDI) prediction is crucial for drug discovery and clinical safety.
• Current DDI prediction models primarily use molecular graph structures, often neglecting valuable functional group semantic information.
• This limitation hinders the accurate identification of complex DDI patterns.

Purpose of the Study:

• To introduce FGSGT-DDI, a framework that integrates LLM-enhanced functional group semantics with Graph Transformer-based structural learning for improved DDI prediction.
• To address the limitations of existing methods by incorporating rich semantic information of molecular functional groups.

Main Methods:

• Extraction of molecular functional groups using SMARTS patterns.
• Generation of semantic embeddings for functional groups, SMARTS patterns, and descriptive information via a large language model (LLM).
• Construction of parallel structural and semantic channels, utilizing cross-attention for deep fusion of molecular graph and functional group semantic features.

Main Results:

• FGSGT-DDI demonstrated competitive performance on three public datasets (Deng, Ryu, MUDI).
• Ablation studies confirmed the efficacy of functional group semantic modeling, graph structure modeling, and the cross-channel interaction module.
• Explainability analysis revealed the model's ability to identify key chemical substructures associated with DDIs.

Conclusions:

• FGSGT-DDI offers an effective approach for DDI prediction by jointly modeling multigranular functional group semantics and molecular structures.
• The integration of LLM-enhanced semantic information significantly contributes to downstream DDI prediction tasks.
• The framework provides insights into the chemical basis of drug-drug interactions.