【摘要】 蛋白质二级结构预测是蛋白质组计划中的一个重要组成部分，为提高蛋白质二级结构的预测准确率，同时减小训练规模，提出了一种基于自注意力生成对抗网络（Self-Attention generative adversarial network,SA-GAN）的深度学习模型。该模型利用生成对抗网络（generative adversarial network,GAN）提取隐式特征，其次将提取的特征结果与蛋白质序列的位置特异性矩阵（position specific scoring matrix,PSSM）结合作为网络的输入，其中，自注意力模块与卷积增强的GAN共同作用，得到预测结果。在测试数据集CASP10,CASP11,CASP12,CASP13,CASP14和CB513上分别获得了83.93%,83.61%,84.13%,84.86%,84.02%和83.37%的Q3准确率。实验结果表明，SA-GAN模型对于提取生物序列特征、获取长程依赖全局信息和提高蛋白质二级结构预测准确率的作用十分显著，具有较强的表达能力和竞争力。更多还原

【Abstract】 Protein secondary structure prediction is an important part of the proteome project. In order to improve the accuracy of protein secondary structure prediction and reduce the scale of training, a deep learning model based on Self-Attention generative adversarial network(SA-GAN) was proposed. The model utilized a generative adversarial network(GAN) to extract implicit features. Then both the extracted feature results and the position specific scoring matrix(PSSM) of protein sequences were set as the input of the network, in which prediction results were obtained by the conjuction of Self-Attention module and convolutional generative adversarial neural networks. The Q3 accuracy of 83.93%, 83.61%, 84.13%,84.86%, 84.02% and 83.37% were obtained from the test datasets CASP10, CASP11, CASP12, CASP13,CASP14 and CB513, respectively. The experimental results show that the SA-GAN model plays a very significant role in extracting biological sequence features, obtaining long-range dependent global information and improving the accuracy of protein secondary structure prediction. It has strong expressiveness and competitiveness.更多还原