DSNet: Predicting drug-side effect frequencies via Dual-Graph Ensemble and Similarity Learning

Drug safety remains a critical concern in healthcare, making the accurate prediction of drug-side effect frequencies essential for risk-benefit assessments. Recent advancements in graph neural network-based methods for predicting drug-side effect frequencies have shown significant promise. However, the inherent complexity of drug molecular structures, often characterized by multi-ring and long-chain substructures, poses a challenge. Mainstream graph-based approaches are limited in expressive power and suffer from low information transmission efficiency, which hampers the ability to capture deep structural features. Additionally, the sparsity of drug-side effect interaction networks restricts the effective utilization of similarity information between drugs and side effects, substantially degrading prediction quality.

To address these challenges, we propose a novel framework for predicting drug-side effect frequencies, termed DSNet. By integrating multi-source heterogeneous features to construct embedding representations, and designing a Dual-Graph Ensemble Network with residual connections, DSNet enhances the capture of local, subtle features of drug molecules while preserving global structural consistency. To mitigate the sparsity limitations of drug-side effect interaction networks, we introduce a Structural Consistency Preservation Loss, which ensures that critical information is retained in the low-dimensional space. Additionally, we propose a Temperature-Adaptive Similarity Loss to dynamically adjust the sharpness of the similarity distribution between drugs and side effects. Experimental results on the SIDER dataset demonstrate that DSNet significantly improves prediction performance in both warm-start and cold-start scenarios. Furthermore, molecular docking experiments targeting tigecycline further validate the effectiveness of DSNet in predicting drug-side effect frequencies.