卷积神经网络在核小体定位识别中的应用  

Application of convolutional neural network based on Z-Curve theory in identifying nucleosome positioning

作　　者：崔颖[1,2] 施丹丹 徐泽龙 张兆功 李建中[2] CUI Ying;SHI Dandan;XU Zelong;ZHANG Zhaogong;LI Jianzhong(Harbin Medical University, College of Bioinformatics Science and Technology, Harbin 150086, China;Heilongjiang University, College of Computer Science and Technology, Harbin 150080, China)

机构地区：[1]哈尔滨医科大学生物信息科学与技术学院,黑龙江哈尔滨150086 [2]黑龙江大学计算机科学与技术学院,黑龙江哈尔滨150080

出　　处：《哈尔滨工程大学学报》2021年第5期751-758,共8页Journal of Harbin Engineering University

基　　金：国家自然科学基金重点项目(61832003).

摘　　要：为更准确识别核小体定位,本文提出一种基于Z曲线理论(Z-Curve)的卷积神经网络(CNN)方法,称为ZCN方法。ZCN方法以Z曲线三维坐标矩阵表示核小体序列特征,通过十倍交叉验证,进行卷积神经网络方法进行模型训练和验证,使用标准评估指标进行性能评价。结果表明:ZCN方法在酵母中具有良好的识别效能,敏感性Sn、准确性Sp、ROC曲线面积分别为92.4%、90.2%和0.9704,可推广到人类、线虫和果蝇的核小体定位识别中,其ROC曲线面积分别为0.796、0.940和0.772,与其他方法比较,进一步证实ZCN方法具有较好的识别效能和可推广性。在酵母全基因组进行核小体定位预测,发现16条染色体的预测准确率均值为78.83%,在基因GAL和GAL10中进行核小体定位预测,研究了降低假阳性的方法,给出了预测核小体定位的图谱。ZCN方法为研究核小体定位识别、预测及功能分析提供了有价值的方法和指导。In order to improve the recognition accuracy of nucleosome positioning,we applied a method,that convolution neural network method based on the Z-curve theory was used to identify nucleosome DNA sequence and called ZCN method,the z-curve three-dimensional coordinate matrix was used to represent the features of nucleosomes,the convolutional neural network was trained and verified by nucleosome sequences and linker sequences through 10-fold cross validation,and use standard evaluation indicators for performance evaluation.The results show that the convolutional neural network method has a good recognition performance in Saccharomyces cerevisiae,and Sensitivity Sn,accuracy Sp and ROC curve areas were 92.4%,90.2%and 0.9704,respectively.It can be applied to the localization of nucleosomes in H.sapiens,C.elegans and D.melanogaster,the ROC curve areas were 0.796,0.940 and 0.772,respectively,and compared with the other four methods,it is further confirmed that ZCN method has better identification efficiency and generalization.In the whole genome of S.cerevisiae,the prediction accuracy of nucleosome positioning was found to be 78.83%for the mean of 16 chromosomes.ZCN was used to predict nucleosome positioning on GALl and GAL10 genes,which the method of reducing false positive was studied and the map of predicting nucleosome was displayed.In conclusion,ZCN method provides valuable methods and guidance for the researching nucleosome positioning,recognition,prediction and function analysis.

关 键 词：计算生物学 卷积神经网络 Z曲线理论 核小体 DNA序列 连接区 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]