Asymmetric drug-drug interaction prediction based on diffusion-augmented graph attention network

The prediction of potential Drug-Drug Interactions (DDIs) is crucial to reduce the harmful side effects from the combined use of multiple drugs. However, existing methods cannot comprehensively capture the DDI asymmetric information, resulting in inaccurate predictions, which increases the risk of harmful side effects. To address this, inspired by the advantages of directed graphs in representing asymmetric relationships, we propose a novel framework called Diffusion Graph Attention DDI (DiffGAT-DDI), designed to effectively capture the DDI asymmetric information to improve DDI prediction accuracy. First, we generate Morgan Fingerprints from drug molecular structures and construct a directed DDI network to provide rich contextual information for interaction prediction. Second, an extended bidirectional graph attention network is employed to learn dual-view representations of drug interactions, while leveraging attention mechanisms to allocate weights for node embeddings, thereby enhancing the model's interpretability. Third, we design an asymmetry-aware diffusion model that introduces edge structural noise and leverages reverse diffusion processes to capture complex asymmetric interaction patterns. Experimental results demonstrate that DiffGAT-DDI significantly outperforms existing state-of-the-art models in both direction-specific and direction-agnostic tasks, achieving the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC) scores of 99.2% and 99.1%, respectively. Compared to the best baseline, this represents improvements of 2.1% in AUROC and 4.0% in AUPRC.