基于Wasserstein生成对抗网络和残差网络的8类蛋白质二级结构预测  

Protein 8-State Secondary Structure Prediction Based on Wasserstein Generative Adversarial Network and Residual Network

作　　者：李舜 马玉明 刘毅慧[1] 

机构地区：[1]齐鲁工业大学(山东省科学院)计算机科学与技术学院,山东 济南

出　　处：《计算生物学》2023年第1期1-9,共9页Hans Journal of Computational Biology

摘　　要：蛋白质二级结构包含充分的蛋白质信息,而且蛋白质二级结构是研究蛋白质三级结构和药物设计的基础,因此,本文提出了一种基于Wasserstein生成对抗网络(WGAN)和残差网络(ResNet)的蛋白质8态二级结构预测的方法。该方法首先通过Wasserstein生成对抗网络(WGAN)提取蛋白质特征,将其与PSSM结合成新的特征集合,然后将新的特征集合输入到残差网络(ResNet)预测并得到最终结果。经过实验,该方法在测试集CASP10-14和CB513中的Q8预测准确率分别为73.21%,72.43%,71.67%,69.83%,70.17%和73.89%。通过实验表明,Wasserstein生成对抗网络(WGAN)具有出色的特征提取能力,ResNet能够有效地训练深层网络结构,从而提高蛋白质二级结构的预测精度。Protein secondary structure is the basis for studying protein tertiary structure and drug design, because the 8-state protein secondary structure can provide sufficient protein information for this. Therefore, this paper proposes a method for predicting the 8-state secondary structure of proteins based on Wasserstein generative adversarial network (WGAN) and residual network (ResNet). This method first extracts protein features by Wasserstein generative adversarial network (WGAN), combines them with PSSM to form a new feature set, and then inputs the new feature set to the residual network (ResNet) prediction and obtains the final result. After experiments, the Q8 prediction accuracy of this method in the test set CASP10-14 and CB513 was 73.21%, 72.43%, 71.67%, 69.83%, 70.17% and 73.89%, respectively. Experiments show that the Wasserstein generative adversarial network (WGAN) has excellent feature extraction ability, and ResNet can effectively train the deep network structure, thereby improving the prediction accuracy of protein secondary structure.

关 键 词：生成对抗网络 残差网络 蛋白质二级结构 特征提取 深层网络 二级结构预测 预测准确率 药物设计 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]