Related Experiment Videos
HyHO-DDI: A hypergraph neural network for high-order Drug-Drug interaction prediction
Naeem Shirmohammady1, Zaynab Mousavian2, Somayeh Kafaie1
1Department of Mathematics and Computing Science, Saint Mary's University, Halifax, NS, Canada.
Objective:
Drug-drug interactions (DDIs) are a major cause of adverse drug reactions (ADRs), particularly in patients receiving multiple medications. Existing computational methods predominantly focus on pairwise interactions, limiting their ability to capture the complex effects of high-order (multi-drug) therapies. This study proposes a computational framework for predicting myopathy-associated high-order DDIs.
Methods:
We propose HyHO-DDIs, a hypergraph neural network model for predicting myopathy-associated high-order drug combinations using a dataset curated from the FDA Adverse Event Reporting System (FAERS). In our model, each drug is represented by molecular fingerprint features and a protein-target profile, and we construct a hypergraph in which nodes are drugs and hyperedges correspond to multi-drug combinations co-reported with myopathy. The model integrates Graph Convolutional Network (GCN) layers with hypergraph-based attention to learn high-order interaction representations, and a decoder outputs the predicted probability of a myopathy association for an input drug combination. We evaluated HyHO-DDI with 5-fold cross-validation and compared it against an MLP with the same inputs, as well as established high-order baselines (HODDI, D3I), and DeepPSE on the pairwise subset.
Results:
HyHO-DDI achieved strong overall performance across combination sizes and remained competitive with DeepPSE on the pairwise subset, achieving the same F1-score with higher recall. For larger combinations, the proposed model reduced false-positive predictions while maintaining strong performance on previously unseen drug combinations, demonstrating robust generalization.
Conclusion:
HyHO-DDI provides an effective framework for predicting myopathy-associated high-order DDIs by explicitly modeling multi-drug interactions with hypergraphs. The proposed approach has the potential to improve computational pharmacovigilance by prioritizing potentially hazardous drug combinations and supporting medication safety.