GGAECDA: Predicting circRNA-disease associations using graph autoencoder based on graph representation learning

Guanghui Li; Yawei Lin; Jiawei Luo; Qiu Xiao; Cheng Liang

doi:10.1016/j.compbiolchem.2022.107722

GGAECDA: Predicting circRNA-disease associations using graph autoencoder based on graph representation learning

Comput Biol Chem. 2022 Aug:99:107722. doi: 10.1016/j.compbiolchem.2022.107722. Epub 2022 Jul 4.

Authors

Guanghui Li¹, Yawei Lin², Jiawei Luo³, Qiu Xiao⁴, Cheng Liang⁵

Affiliations

¹ School of Information Engineering, East China Jiaotong University, Nanchang, China. Electronic address: ghli16@hnu.edu.cn.
² School of Information Engineering, East China Jiaotong University, Nanchang, China.
³ College of Computer Science and Electronic Engineering, Hunan University, Changsha, China. Electronic address: luojiawei@hnu.edu.cn.
⁴ College of Information Science and Engineering, Hunan Normal University, Changsha, China.
⁵ School of Information Science and Engineering, Shandong Normal University, Jinan, China.

PMID: 35810557
DOI: 10.1016/j.compbiolchem.2022.107722

Abstract

Numerous studies have shown that circular RNAs (circRNAs) can serve as ideal disease markers as they are involved in most cellular activities of organisms and are key regulators in various pathological processes. Therefore, the association analysis of circRNAs and diseases can explore the role of circRNAs in diseases and provide help for practical medical research. However, the traditional biotechnology are not convenient for identifying unconfirmed interactions between circRNAs and diseases, which need too many resources and long experimental period. In this work, a new deep learning model is advanced, which is based on graph autoencoder (GAE) constructed with graph attention network (GAT) and random walk with restart (RWR) for predicting circRNA-disease associations (GGAECDA). In detail, GAT is designed to learn the hidden representations of circRNAs and diseases through using low-order neighbor information from circRNA similarity network and disease similarity network respectively, while RWR is employed to learn the latent features of circRNAs and diseases via using high-order neighbor information from the same two networks respectively. After that, these two parts of features of circRNAs and diseases are combined to form new feature representations of circRNAs and diseases respectively. Finally, two GAEs are constructed for co-training to fully integrate information from circRNA space and disease space and calculate potential association prediction scores. Unlike previous models, GGAECDA deeply mines low-dimensional representations from node similarity network through using GAT and RWR. The average AUC value obtained from GGAECDA with a five-fold cross-validation result is 0.9359. Furthermore, case studies demonstrate the ability of GGAECDA to detect potential candidate circRNAs for human diseases. The above results show that the GGAECDA model can be used as a reliable tool to guide subsequent studies on the regulatory functions of circRNAs.

Keywords: CircRNA-disease associations; Graph attention network; Graph autoencoder; Random walk with restart.

MeSH terms

Computational Biology* / methods
Humans
RNA, Circular* / genetics

Substances

RNA, Circular