基于深度学习的核糖核酸二级结构预测方法  

Deep learning-based method for ribonucleic acid secondary structure prediction

作　　者：刘元宁[1,2] 臧子楠 张浩 刘震[1,3] LIU Yuan-ning;ZANG Zi-nan;ZHANG Hao;LIU Zhen(College of Computer Science and Technology,Jilin University,Changchun 130012,China;Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education,Jilin University,Changchun 130012,China;Graduate School of Engineering,Nagasaki Institute of Applied Science,Nagasaki 851-0193,Japan)

机构地区：[1]吉林大学计算机科学与技术学院,长春130012 [2]吉林大学符号计算与知识工程教育部重点实验室,长春130012 [3]长崎综合科学大学研究生院工学研究科,长崎851-0193

出　　处：《吉林大学学报(工学版)》2025年第1期297-306,共10页Journal of Jilin University:Engineering and Technology Edition

基　　金：吉林省自然科学基金项目(YDZJ202101ZYTS144)。

摘　　要：本文提出了一种基于深度学习的方法 UCEfold,用于预测核糖核酸(Ribonucleic acid,RNA)二级结构。UCEfold是一种同时采用“序列”和“图像”作为深度学习模型输入提取隐藏特征的全新方法,并在模型中加入一定的先验知识提高预测精度。在RNAStralign和ArchiveⅡ两个数据集上测试UCEfold模型,结果表明UCEfold性能显著优于传统方法,能够更准确地预测带假结的RNA序列,并具有较强的泛化能力,有效解决了传统算法复杂度高、效率低下且无法预测假结的瓶颈。A new method based on deep learning,UCEfold,is proposed for predicting ribonucleic acid(RNA) secondary structure using both“sequence”and“image”as input to the deep learning model to extract hidden features.It also added some prior knowledge to the model to improve the prediction accuracy.There have tested the UCEfold model on both RNAStralign and ArchiveⅡdatasets,and the results show that UCEfold outperforms the traditional method significantly,and can predict the RNA sequences with pseudoknots more accurately and has strong generalization ability,which effectively solves the bottleneck of the traditional algorithm with high complexity,low efficiency and inability to predict pseudoknots.

关 键 词：计算机应用 深度学习 核糖核酸二级结构预测 假结 注意力机制 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]