Evaluating disease-disease associations based on network representation learning

Evaluating disease-disease associations is critical for understanding the complex mechanism of less well-studied disease and facilitating disease diagnosis and therapy. Therefore, people present some methods to explore the potential connection between disease. However, they mostly rely on overlaps between disease-related gene sets or distances within ontological levels, where the diseases are described by discrete or sparse feature vectors failing to capture the latent semantic information of disease. In this study, inspired by the successfully application of network representation in bioinformatics tasks, we propose a novel method to evaluate disease-disease associations by combining network representation learning with best-match average similarity strategy. Our method can capture more semantically rich disease feature vectors and make better use of prior biological knowledge. The experimental results on a public benchmark dataset showed that our method can improve the predictive performance, and the visualization analysis further demonstrate the effectiveness of disease latent features learned by network representation learning.