基于注意力机制的RNA碱基关联图预测方法  

Prediction Method of RNA Contact Map Based on Attention Mechanism

作　　者：曹一航 黄强[1] CAO Yi-Hang;HUANG Qiang(School of Life Sciences,Fudan University,Shanghai 200438,China)

机构地区：[1]复旦大学生命科学学院,上海200438

出　　处：《生物化学与生物物理进展》2023年第3期657-667,共11页Progress In Biochemistry and Biophysics

基　　金：国家重大科技专项“重大新药开发”课题(2018ZX09J18112);国家自然科学基金(31971377)资助项目。

摘　　要：目的长链非编码RNA在遗传、代谢和基因表达调控等方面发挥着重要作用。然而,传统的实验方法解析RNA的三级结构耗时长、费用高且操作要求高。此外,通过计算方法来预测RNA的三级结构在近十年来无突破性进展。因此,需要提出新的预测算法来准确的预测RNA的三级结构。所以,本文发展可以用于提高RNA三级结构预测准确性的碱基关联图预测方法。方法为了利用RNA理化特征信息,本文应用多层全卷积神经网络和循环神经网络的深度学习算法来预测RNA碱基间的接触概率,并通过注意力机制处理RNA序列中碱基间相互依赖的特征。结果通过多层神经网络与注意力机制结合,本文方法能够有效得到RNA特征值中局部和全局的信息,提高了模型的鲁棒性和泛化能力。检验计算表明,所提出模型对序列长度L的4种标准(L/10、L/5、L/2、L)碱基关联图的预测准确率分别达到0.84、0.82、0.82和0.75。结论基于注意力机制的深度学习预测算法能够提高RNA碱基关联图预测的准确率,从而帮助RNA三级结构的预测。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 multilayer 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.

关 键 词：深度学习 RNA碱基关联图 结构预测 注意力机制 

分 类 号：Q812[生物学—生物工程]