Enhancing lncRNA-disease associations predictions by optimizing similarity on multiple heterogeneous networks.

Focusing on the problem about predicting long non-coding RNA(lncRNA)-disease associations, this article introduces an improved random walk with restart(RWR) model by optimizing similarity on multiple heterogeneous networks, referred to as OS-LDA. Most existing models for constructing heterogeneous networks only consider two types of networks(disease and lncRNA), neglecting the valuable information from other networks. Furthermore the construction of an accurate and reasonable similarity network is crucial to the effectiveness of model. This paper considers four types of networks and focuses on improving and optimizing the similarity framework of network to enhance the prediction capabilities of the model. Firstly, a new approach for measuring disease semantic similarity which combines the advantages of two conventional methods is introduced. Secondly, to overcome the sparsity of disease semantic similarity matrix, this paper proposes an improved measure based on the penalty factor, thereby making it more suitable to measuring similarity between different diseases. In addition, multiple interaction profiles are taken into account for the computation of Gaussian similarity. Finally, this paper constructs precise multilayer heterogeneous similarity networks (lncRNA-gene-miRNA-disease) and the random walk with restart method is implemented on heterogeneous networks to predict disease-related lncRNAs. The average AUC of OS-LDA reaches 0.9876 by ten-fold cross validation, outperforming several other models and indicating the effectiveness of the algorithm.