DataSheet1_GCNCMI: A Graph Convolutional Neural Network Approach for Predicting circRNA-miRNA Interactions.zip

The interactions between circular RNAs (circRNAs) and microRNAs (miRNAs) have been shown to alter gene expression and regulate genes on diseases. Since traditional experimental methods are time-consuming and labor-intensive, most circRNA-miRNA interactions remain largely unknown. Developing computational approaches to large-scale explore the interactions between circRNAs and miRNAs can help bridge this gap. In this paper, we proposed a graph convolutional neural network-based approach named GCNCMI to predict the potential interactions between circRNAs and miRNAs. GCNCMI first mines the potential interactions of adjacent nodes in the graph convolutional neural network and then recursively propagates interaction information on the graph convolutional layers. Finally, it unites the embedded representations generated by each layer to make the final prediction. In the five-fold cross-validation, GCNCMI achieved the highest AUC of 0.9312 and the highest AUPR of 0.9412. In addition, the case studies of two miRNAs, hsa-miR-622 and hsa-miR-149-5p, showed that our model has a good effect on predicting circRNA-miRNA interactions. The code and data are available at https://github.com/csuhjhjhj/GCNCMI.

环状RNA（circular RNAs, circRNAs）与微小RNA（microRNAs, miRNAs）之间的相互作用已被证实可调控基因表达，并参与疾病相关的基因调控机制。由于传统实验方法耗时耗力，目前绝大多数环状RNA与微小RNA的相互作用仍未被探明。开发可用于大规模挖掘环状RNA与微小RNA相互作用的计算方法，有助于填补这一研究空白。本文提出了一种基于图卷积神经网络（graph convolutional neural network）的计算方法GCNCMI，用于预测环状RNA与微小RNA之间的潜在相互作用。GCNCMI首先在图卷积神经网络中挖掘相邻节点的潜在相互作用，随后在图卷积层上以递归方式传播相互作用信息。最终，该模型会融合各层生成的嵌入表征，以完成最终的预测任务。在五折交叉验证实验中，GCNCMI取得了0.9312的最高受试者工作特征曲线下面积（Area Under Curve, AUC）与0.9412的最高精确召回曲线下面积（Area Under Precision-Recall Curve, AUPR）。此外，针对hsa-miR-622与hsa-miR-149-5p两种微小RNA的案例研究证实，本模型在预测环状RNA与微小RNA相互作用方面表现优异。本研究的代码与数据可通过https://github.com/csuhjhjhj/GCNCMI获取。