基于多尺度卷积和循环神经网络的蛋白质二级结构预测  

Protein Secondary Structure Prediction Based on Multiscale Convolution and Recurrent Neural Networks

作　　者：包晨 董洪伟[1] 钱军浩[1] Bao Chen;Dong Hongwei;Qian Junhao(College of The Internet of Things Engineering,Jiangnan University,Wuxi,214000)

机构地区：[1]江南大学物联网工程学院,无锡214000

出　　处：《基因组学与应用生物学》2020年第7期3025-3030,共6页Genomics and Applied Biology

摘　　要：蛋白质的二级结构预测是生物信息学中的一个重要问题。本研究针对传统的蛋白质结构预测模型精度不够高的问题,提出一种端到端的融合多次多尺度卷积和多层次双向长短期记忆网络的模型,充分提取了氨基酸序列的局部和长程的特征信息,进而用于蛋白质的二级结构预测。首先模型分别对氨基酸残基独热序列信息和氨基酸进化结构信息进行多次多尺度卷积提取特征信息,将提取的特征信息与原始序列信息进行融合构成残差模块送入多层双向长短期记忆网络进行局部和长程相互作用,最后送入全连接层进行8类蛋白质二级结构预测。实验结果表明,本研究提出的模型相较于基准方法,提高了8类蛋白质二级结构预测的精度。Protein secondary structure prediction is an important issue in bioinformatics.To address the problem that traditional prediction accuracy is not high enough,this paper proposes a model of end-to-end fusion multi-scale convolution and multi-level bidirectional long short-term memory networks,which fully extracts the local and long-range characteristic information of amino acid sequences,and then used for protein secondary structure prediction.Firstly,the model performs multiple multi-scale convolution on amino acid residue one-hot sequence information and evolution information partly,and fuses the extracted information with the original data to form a residual module to be sent to a multi-layer bidirectional long short-term memory network to interact.Finally,which is sent to the fully connected layer for the prediction of 8 classes of protein secondary structure.The experimental result showed that the proposed model improves the accuracy which is compared with the benchmark method.

关 键 词：蛋白质二级结构 多尺度卷积 循环神经网络 

分 类 号：Q518.1[生物学—生物化学]