Prediction Method of RNA Contact Map Based on Attention Mechanism

Objective Long non-coding RNA play an important role in genetics, metabolism and gene expression regulation. But it is time-consuming and costly to analyze the RNA structure by experimental approaches. However, prediction software based on co-evolutionary algorithm has not made breakthrough progress in prediction accuracy in recent ten years. Therefore, it is necessary to propose a new prediction algorithm to accurately predict the tertiary structure of RNA. So, this paper develops prediction method of base contact map of RNA that can be used to improve the accuracy of tertiary structure prediction. Methods To utilize the physical and chemical characteristics of RNA, we propose a deep learning algorithm based on multi-layer convolutional neural network and long short-term memory networks to predict the contact map between base pair. In addition, we employ attention mechanism to deal with complex global spatial independence features in RNA sequences. Results By combining multilayer neural networks with the attention mechanism, our method can effectively obtain local and global information in RNA features, which improves the robustness and generalization ability of the model. The computations show that the proposed model achieves 0.84, 0.82, 0.82 and 0.75 prediction accuracies for the base contact map of 4 criteria (L/10, L/5, L/2, L) of sequence length L. Conclusion Prediction method based on attention method is better than traditional computational methods and common deep learning algorithms, respectively.