AUTHOR=Fan Chunyan , Lei Xiujuan , Pan Yi 

TITLE=Prioritizing CircRNA–Disease Associations With Convolutional Neural Network Based on Multiple Similarity Feature Fusion

JOURNAL=Frontiers in Genetics

VOLUME=Volume 11 - 2020

YEAR=2020

URL=https://www.frontiersin.org/journals/genetics/articles/10.3389/fgene.2020.540751

DOI=10.3389/fgene.2020.540751

ISSN=1664-8021

ABSTRACT=Increasing evidences have shown that circular RNAs (circRNAs) play significant roles in the occurrence and development of human health and diseases. Biological researchers have identified diseases related circRNAs by wet-lab experiments, and circRNAs could be considered as potential biomarkers for clinical diagnosis, prognosis and treatment. However, identification of circRNA-disease associations with traditional biological experiments is still expensive and time-consuming. In this study, we proposed a novel method named MSFCNN for the task of circRNA-disease prediction, which applied two layer convolutional neural networks on the feature matrix that fused multiple similarity kernels and interaction features among circRNAs, miRNAs and diseases. First, four circRNA similarity kernels and seven disease similarity kernels are constructed based on the biological or topological properties of circRNAs and diseases. Subsequently, similarity kernel fusion (SKF) method is used to integrate the similarity kernels into one circRNA similarity kernel and one disease similarity kernel, respectively. Then, feature matrix of each circRNA-disease pair is constructed by integrating fused circRNA similarity kernel, fused disease similarity kernel, interactions or features among circRNAs, miRNAs and diseases, in which interaction features of circRNA-miRNA and disease-miRNA interactions are selected with principal component analysis (PCA) method. Finally, taking the constructed feature matrix as input, we employed a two layer convolutional neural networks (CNN) to predict the circRNA-disease association labels and mine the potential novel associations. The five-fold cross validation (5-fold CV) evaluation shows that our proposed model outperforms conventional machine learning methods including support vector machine (SVM), random forest (RF) and multilayer perception (MPL). Furthermore, case studies of predicted circRNAs for specific disease and top predicted circRNA-disease associations are analyzed. The above results show that MSFCNN model could be an effective tool for mining potential circRNA-disease associations.