Multi-View Disease-Symptom Association Prediction Based on Disease Knowledge Graph Embedding

Disease symptoms are significant in disease treatment and diagnosis, as well as in developing a hierarchical model from disease phenotypes to genes to improve our understanding of complex disease causes. While previous studies have focused on genetics and proteomics, this study proposes a novel compu-tational model, MGAEMF, based on disease knowledge graph embedding for disease-symptom association (DSA) prediction (i.e., disease nodes have connections with multiple other node types). First, the disease knowledge graph is decomposed into multiple disease-related bipartite networks, form which multiple disease similarity networks and symptom similarity networks can be extracted, and the similarity networks about diseases and symptoms are separately integrated into a single similarity network. Then, the integrated disease similarity network and symptom similarity network are combined with bipartite networks built of known disease-symptom connections to generate nonlinear feature representations of diseases and symptoms using GAE. Meanwhile, the disease-symptom association matrix is calculated using a non-negative matrix factorization to provide linear feature representations of diseases and symptoms. Finally, a fully connected neural network incorporating the linear and nonlinear representations of diseases and symptoms calculates the final prediction scores for all disease-symptom pairs. During the 10-fold cross-validation experiment, the MGAEMF model outperforms other comparable methods with a superior average AUC value of 0.8372.