基于深度卷积神经网络的RNA三维结构打分函数  

An RNA tertiary structure scoring function based on deep neural network

作　　者：杜渊洋 邓成伟 张建[1,2] Du Yuanyang;Deng Chengwei;Zhang Jian(School of Physics,Nanjing University,Nanjing,210093,China;Collaborative Innovation Center of Advanced Microstructures,Nanjing,210093,China)

机构地区：[1]南京大学物理学院,南京210093 [2]人工微结构科学与技术协同创新中心,南京210093

出　　处：《南京大学学报（自然科学版）》2022年第3期369-376,共8页Journal of Nanjing University（Natural Science）

基　　金：国家自然科学基金(11774158,12090052,11334004)。

摘　　要：非编码RNA的三维结构对于人们理解和干预其生物功能具有重要的意义,从计算的角度发展RNA结构预测方法可以加速结构获取过程,对三维结构进行评分是进行结构预测的关键步骤.近年来,基于机器学习的方法,如AlphaFold2,已在分子结构预测领域取得了革命性的进展.基于深度卷积神经网络,建立了一个对RNA三维结构进行评估的方法 .为了训练这一网络,建立了一个非冗余的含有422个RNA以及126600个decoys结构的数据集.训练得到的模型在RNA-Puzzles数据集上进行了测试,结果表明,在28个RNA中,网络从众多decoys中挑选出实验结构的正确率约为71.4%,这一结果比之前有所提高.另外,还对网络的工作机制进行了分析,发现神经网络对结构评分的倾向性和已知的物理化学知识相一致.It is essential to acquire RNA tertiary structures for understanding and intervening their biological functions which accelerate the development of computational structure prediction. One of the key steps is to evaluate the quality of the structural candidates. Recently,approaches based on machine learning,such as AlphaFold2,have achieved revolutionary progress in protein structure prediction. In this study,we develop an RNA structure scoring function based on deep convolutional neural network. We also build a training dataset including 422 non-redundant RNAs and 126600 associated decoys. We test the trained model on the RNA-Puzzles dataset. The results show that,among 28 RNAs,the model correctly identify experimental structures out of decoys with a ratio of 71.4%,superior to our previous model. Furthermore,we analyze the underlying mechanism of the neural network, finding that the way it scores the structural elements is consistent with known physical-chemical principles.

关 键 词：RNA结构预测 打分函数 卷积神经网络 深度学习 机器学习 

分 类 号：Q615[生物学—生物物理学]