A multi-layer hypergraph framework for drug-drug interaction prediction based on transformer and hypergraph convolution.

Drug-drug interactions represent a key problem for drug research, development, and clinical practice. It is crucial to accurately predict interactions when drugs combine to improve treatment safety and optimize medication regimens. However, the exponential increase in potential drug combinations, along with the limitations of conventional graph and multi-layer network prediction models- which primarily capture only binary relationships between drugs and struggle to represent multi-element synergistic interactions-limits prediction performance. To overcome these challenges, this paper proposes a Multi-Layer Hypergraph framework for drug interaction prediction using Transformer and Hypergraph Convolution (MLHTHC). This framework first constructs a multi-layer similarity hypergraph of drugs based on four attribute types: chemical structure, ATC code, drug category, and corresponding targets. Using drug-drug interaction data from KEGG database as a benchmark, the spectral Hamming similarity method is adopted to calculate the structural similarity between the constructed hypergraph and the benchmark hypergraph, enabling the determination of the importance weight for each hypergraph layer. Subsequently, a hypergraph convolutional neural network performs network embedding on each layer of drug nodes; the Transformer model is used to weight and fuse the multi-layer features; and finally, The multi-layer perceptron (MLP) is used to predict drug-drug interactions (DDIs).Experimental results demonstrate that this model outperforms existing methods such as DPSP and DANN, with the integration of Transformer and hypergraph convolution significantly enhancing prediction accuracy. This approach provides an effective tool for drug-drug interaction prediction.