BMBQA:融合MCNN和BiGRU的蛋白质模型质量评估算法  

BMBQA:Protein Model Quality Assessment Algorithm Fused with MCNN and BiGRU

作　　者：聂良鹏 权丽君[1,2] 吴庭芳 孙晓雨 何如吉 吕强 NIE Liang-peng;QUAN Li-jun;WU Ting-fang;SUN Xiao-yu;HE Ru-ji;LV Qiang(School of Computer Science and Technology,Soochow University,Suzhou 215006,China;Jiangsu Province Key Lab for Information Processing Technologies,Suzhou 215006,China)

机构地区：[1]苏州大学计算机科学与技术学院,江苏苏州215006 [2]江苏省计算机信息处理技术重点实验室,江苏苏州215006

出　　处：《小型微型计算机系统》2022年第7期1419-1425,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(31801108)资助;国家自然科学基金青年基金项目(62002251)资助;江苏省自然科学基金青年基金项目(BK20200856)资助;江苏高校优势学科建设工程资助项目资助.

摘　　要：蛋白质模型质量评估是指对计算手段预测出的蛋白质结构进行评分,以选择尽可能接近天然结构的蛋白质模型,这对在蛋白质结构预测中挑选最优的蛋白质模型和在生物医学研究中恰当使用它们起到至关重要的作用.随着3D蛋白质数据的不断增长,基于深度学习的蛋白质模型质量评估方法得到了快速发展,但该领域只探索了深度学习方向较浅层的应用.本文为了探索更精确的质量评估方法,提出了一个基于多尺度卷积(MCNN)和双向门控循环神经网络(BiGRU)的深度模型,预测蛋白质模型的GDT_TS(Global Distance Test_Total Score)分数,并将这一方法命名为BMBQA(Quality Assessment Base on MCNN-BiGRU).其中,多尺度卷积神经网络用来提取蛋白质模型中浅层的细节信息以及深层的抽象信息,双向门控循环神经网络用来提取每个残基的长程相互作用信息,通过数据增强来提高深度模型在目标蛋白质中挑选最优蛋白质模型的性能.本文利用CASP13中的数据集与现有的先进方法进行比较,实验结果表明本文方法在4个经典的评价指标中均具有很强的竞争力.Protein model quality assessment refers to scoring the protein structure predicted by computational methods to select protein models that approach to the natural structure mostly.This plays a vital role in selecting the best protein model for the protein structure prediction and application in biomedical research.With the continuous growth of 3 D protein data, protein model quality assessment methods based on deep learning have great developed.But the deep learning frameworks they used sonly involve shallow networks.In order to explore a more accurate quality assessment method, we propose a deep model based on multiscale convolutional neural network(MCNN)and bidirectional gated recurrent neural network(BiGRU)to predict the GDT_TS(Global Distance Test_Total Score)of the protein model, which is named BMBQA(Quality Assessment Base on MCNN-BiGRU).Here, the multiscale convolutional neural network is used to extract detailed information on shallow layer and abstract information on deep layer in the protein model, and the bidirectional gated recurrent neural network is used to extract the long-range interaction information of each residue.Besides, data enhancement is introduced to improve the performance of the deep model in selecting the best protein model from the target proteins.Finally, compared with state-of-the-art methods in CASP13 data set, our method has strong competitiveness according to four classic evaluation indicators.

关 键 词：蛋白质模型质量评估 多尺度卷积神经网络 双向门控循环神经网络 GDT_TS 数据增强 CASP13 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]