NIEE: Modeling Edge Embeddings for Drug-Disease Association Prediction via Neighborhood Interactions

Abstract

Using computational methods to search for potential drugs for diseases can speed up the drug development process. The majority of current research focuses on obtaining node embedding representations for link prediction using deep learning techniques. They use a simple inner product to simulate the association between drug and disease nodes, which is insufficient, thus we propose an edge embedding model, which named NIEE, based on the interaction between drug neighborhood and disease neighborhood for performing link prediction tasks. The core idea of NIEE is to simulate the embedding of edges between source and target nodes using the interaction between their neighborhoods. The model first samples the neighborhoods of nodes on the heterogeneous network in accordance with the specially designed meta-paths, and then uses the interaction module to simulate the interaction between the neighborhoods. We de-signed a hierarchical attention mechanism to aggregate heterogeneous nodes within meta-paths and perform semantic-level aggregation between meta-paths. Finally, use the MLP to predict whether the edge exists. We compared our model with four GNN models, and the experiments show that our model outperforms other models in all indicators, confirming the effectiveness of NIEE.