RDense: A Protein-RNA Binding Prediction Model Based on Bidirectional Recurrent Neural Network and Densely Connected Convolutional Networks

Abstract

Complexes formed by proteins binding to RNAs are essential in biological processes, and can also be useful for identifying causal disease variants, gene expression regulation and translation. Protein-RNA interactions identified in vivo can be affected by experimental condition, noise, and some bias, while in vitro experiments yield clearer signals. Therefore, accurately inferring RNA-protein binding models from in vitro data, to predict bound and unbound RNA transcripts in vivo, has become a key challenge. We constructed RDense, a novel deep neural network model. Using existing RNA sequences and secondary structure information, we introduced the pairwise probability feature extracted from the RNA secondary structure as the input. The bidirectional long and short memory neural network (Bi-LSTM) and densely connected convolutional neural networks (DenseNet) were then combined to learn protein-RNA binding preferences. We found that our prediction of in vitro binding was better than all current methods with a significant improvement in model accuracy. In addition, there was also some improvement when in vitro data-trained RNA binding models were used to predict in vivo binding. In summary, we have introduced new pairwise probability feature of RNA to improve the robustness of the model. By comparing the Deepbind and DLPRB methods based on CNN, our method combines the strength of Bi-LSTM and DenseNet, with better performance in accuracy and scalability of prediction.