一种基于多模态语义对齐的药物-药物相互作用预测方法
By integrating text, molecular structure, and knowledge graph modalities through the MultiMod-DDI framework and utilizing adaptive positional interaction vectors and the ProbSparse self-attention mechanism, the problems of insufficient semantic alignment and long-distance dependency modeling ability in drug-drug interaction prediction are solved, achieving more accurate DDI prediction and clinical decision support.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GUANGDONG PHARMA UNIV
- Filing Date
- 2025-09-16
- Publication Date
- 2026-07-17
AI Technical Summary
Existing methods suffer from insufficient semantic alignment when fusing molecular structure, biomedical entity knowledge graphs, and biomedical entity text semantics. This makes it difficult to fully explore the deep mechanisms by which drug molecular structure affects efficacy through biomedical entities, and the methods also have limited ability to model long-distance semantic dependencies in text, resulting in inaccurate identification of drug-drug interaction relationships.
The MultiMod-DDI framework is adopted to integrate three data modalities: text, molecular structure, and knowledge graph. Text is encoded using the BioBERT model, and adaptive positional interaction vectors are introduced to construct a molecular graph neural network PS-AEGNN with a ProbSparse self-attention mechanism. Combined with a multi-stage adaptive fusion module, the synergistic effect of multimodal data is realized.
It significantly improves the accuracy and interpretability of drug-drug interaction prediction, captures deep relationships through cross-modal information fusion, solves the problem of insufficient long-distance dependency modeling capabilities, provides more accurate DDI prediction results, and supports clinical decision-making.
Smart Images
Figure CN121215085B_ABST