iSnoDi-LSGT: identifying snoRNA-disease associations based on local similarity constraints and global topological constraints

Abstract

Growing evidence proves that small nucleolar RNAs (snoRNAs) have important functions in various biological processes, the malfunction of which leads to the emergence and development of complex diseases. However, identifying snoRNA-disease associations is an ongoing challenging task due to the considerable time- and money-consuming biological experiments. Therefore, it is urgent to design efficient and economical methods for the identification of snoRNA-disease associations. In this regard, we propose a computational method named iSnoDi-LSGT, which utilizes snoRNA sequence similarity and disease similarity as local similarity constraints. The iSnoDi-LSGT predictor further employs network embedding technology to extract topological features of snoRNAs and diseases, based on which snoRNA topological similarity and disease topological similarity are calculated as global topological constraints. To the best of our knowledge, the iSnoDi-LSGT is the first computational method for snoRNA-disease association identification. The experimental results indicate that the iSnoDi-LSGT predictor can effectively predict unknown snoRNA-disease associations. The web server of the iSnoDi-LSGT predictor is freely available at http://bliulab.net/iSnoDi-LSGT.

Keywords: global topological constraint; local similarity constraint; network embedding technology; snoRNA-disease association identification.

FIGURE 1.

The influence of parameters on the performance of iSnoDi-LSGT predictor via 10-fold cross-validation on $\mathbb{S}$_benchmark. Different bar charts denote different AUC values based on all predicted results via 10-fold cross-validation on $\mathbb{S}$_benchmark. X-axis, y-axis, and different color bars represent the values of γ, AUC, and β, respectively. Error bars represent deviation across the different 10-fold cross-validation on $\mathbb{S}$_benchmark. The black dotted line denotes a horizontal line at the level of the minimum value.

FIGURE 2.

The framework of iSnoDi-LSGT predictor. The four main steps are as follows: (i) Local similarity constraint. The snoRNA sequence similarity network, snoRNA-disease association network and disease semantic similarity network are constructed, in which snoRNA sequence similarity network and disease semantic similarity network are used as local similarity constraint. (ii) Global topological constraint. Disease topological similarity and snoRNA topological similarity are calculated as global topological constraint based on network embedding technology and heterogeneous network constructed by snoRNAs, diseases and their associations. (iii) Nonnegative matrix factorization. Nonnegative matrix factorization with local similarity constraint and global topological constraint is employed to identify potential snoRNA-disease associations. (iv) Candidate disease-associated snoRNA ranking. The candidate disease-associated snoRNAs are ranked according to the scores calculated by nonnegative matrix factorization.

Wenxiang Zhang