RGDA-DDI: Residual graph attention network and dual-attention based framework for drug-drug interaction prediction
Changjian Zhou, Xin Zhang, Jiafeng Li, Jia Song, Wensheng Xiang
TL;DR
DDI prediction is challenged by the need to integrate diverse drug features and DDP information. The authors propose RGDA-DDI, a framework that combines residual-GAT blocks for multi-scale feature learning with a dual-attention fusion module to capture local joint interactions, guided by biomedical knowledge graphs embedded via OpenKE. The approach constructs BKGs from DrugBank and KEGG, learns rich representations, and predicts interactions with a sigmoid on concatenated drug embeddings, achieving state-of-the-art results on two public datasets. This work advances drug safety assessment and development by enabling robust, multi-scale feature integration and interpretable joint representations for DDI prediction.
Abstract
Recent studies suggest that drug-drug interaction (DDI) prediction via computational approaches has significant importance for understanding the functions and co-prescriptions of multiple drugs. However, the existing silico DDI prediction methods either ignore the potential interactions among drug-drug pairs (DDPs), or fail to explicitly model and fuse the multi-scale drug feature representations for better prediction. In this study, we propose RGDA-DDI, a residual graph attention network (residual-GAT) and dual-attention based framework for drug-drug interaction prediction. A residual-GAT module is introduced to simultaneously learn multi-scale feature representations from drugs and DDPs. In addition, a dual-attention based feature fusion block is constructed to learn local joint interaction representations. A series of evaluation metrics demonstrate that the RGDA-DDI significantly improved DDI prediction performance on two public benchmark datasets, which provides a new insight into drug development.