.png){kind=icon}
.png){kind=icon}
Return
Multiview feature fusion-based graph representation model for drug-drug interaction prediction
.png){kind=icon}
PRE
AI
M
D
E
S
G
F
DOI：10.1007/s10489-026-07397-6.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In drug combination therapy, Drug-Drug interaction (DDI) prediction can effectively improve efficacy or reduce adverse reactions. However, existing artificial intelligence (AI) approaches have limitations in fully considering and effectively extracting rich drug features. To overcome the limitations of single-view feature extraction, this study presents a novel multiview feature fusion-based graph representation model (MFF-GRM) for predicting DDI. This architecture integrates drug molecular graphs, SMILES sequences, DDI information networks, and drug biological features (target, enzyme, transport) to learn drug features more comprehensively. First, we use a Graph Convolutional Network (GCN) to update atomic features and extract drug molecular graph features. Secondly, we propose a cascade graph representation model combined with a multi-head self-attention mechanism and graph convolution to capture the topological features of the DDI network. This design allows the model to capture the heterogeneity of neighborhood interactions and further fuse and constrain the representation through graph structure propagation. After encoding the SMILES (Simplified Molecular Input Line Entry System) sequences, SMILES representations are obtained using the Gated Recurrent Unit (GRU) model. Furthermore, we employ a hybrid similarity calculation strategy to reconstruct varying biological feature matrices for drug representations. Finally, a multi-view feature fusion and prediction module is proposed to fuse multiple drug representations for DDI prediction. We conducted comparative experiments on the ZhangDDI and DrugBank datasets. Additionally, parameter sensitivity tests, ablation parameters, and imbalance data testing were performed. The experimental results demonstrate that MFF-GRM can effectively learn drug representations and outperform the baselines for DDI prediction, achieving improvements across all evaluation metrics, with gains of $$0.07-11.3\%$$ and $$0.92-17.9\%$$ on the two benchmark datasets. The MFF-GRM model shows great application potential in assisting drug repositioning and clinical medication decision-making.
Keywords:
Drug-drug interaction
Multiview feature fusion
Cascade graph representation
Natural language processing
Hybrid similarity calculation strategy
Journal
IF:
3.5
Papers:
7.5K
Citations:
1.7W
Organization
Cited Papers