Prediction of circRNA-Binding Protein Site Based on Hybrid Neural Networks and Recurrent Forests Method

Circular RNAs (circRNAs) play an important role in the regulation of autoimmune diseases by binding to RNA–binding proteins (RBP). Therefore studying the binding sites of RBP on cyclic RNA is crucial for our understanding of the interactions between RBP and its RNA targets. In this paper, we propose the classification method CNBM-RRF based on hybrid neural networks and recurrent forests method for identifying circRNA-RBP interaction sites. In the CNBM-DRAF method, we use four coding methods to extract four features of the cyclic RNA sequences. The features include pseudo amino acid features, pseudo dipeptide features, pseudo secondary structure features, and pseudo word vector features. Then we feed the features into the hybrid neural network to obtain the common features of the cyclic RNA sequences. The hybrid neural network includes the convolutional neural network (CNN) and the bi-directional long short-term memory network (BiLSTM). In addition we use weighted generalized canonical correlation analysis (WGCCA) to extract the common features of the four features. Finally we input common features into recurrent forests for prediction of RBP binding sites on circular RNAs. The proposed recurrent forests method is inspired by Long Short Term Memory (LSTM). We test it on 10 circRNA datasets and compare it with 7 existing methods. The experimental results show that the prediction performance of CNBM-RRF method is improved compared with that of the existing 7 methods.