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

Drug-drug interaction (DDI) prediction is of great importance for drug discovery and safe clinical medication use. Existing methods mainly rely on molecular graph structure modeling but make insufficient use of functional group semantic information, which limits their ability to identify complex DDI patterns. To address this issue, we propose FGSGT-DDI, a DDI prediction framework that integrates large language model (LLM)-enhanced functional group semantic information with Graph Transformer-based structural learning. The method first extracts molecular functional groups based on SMARTS patterns and then uses an LLM to generate semantic embeddings of functional group names, SMARTS patterns, and descriptive information. It subsequently constructs a structural channel and a semantic channel and employs cross-attention to achieve deep fusion of molecular graph features and functional group semantic features, thereby enabling multiclass DDI prediction. Experiments on three public data sets, namely, Deng, Ryu, and MUDI, show that FGSGT-DDI achieves competitive performance. Ablation studies verify the effectiveness of functional group semantic modeling, graph structure modeling, and the cross-channel interaction module. Comparisons among different LLM-based semantic pipelines further demonstrate the positive contribution of LLM-enhanced functional group semantic information to downstream DDI prediction. Explainability analysis shows that the proposed model can identify key local chemical substructures related to DDIs. Overall, FGSGT-DDI provides an effective solution for jointly modeling of multigranular functional group semantic information and molecular structures in DDI prediction.