优化多核SVM的蛋白质二级结构预测
2020-08-03刘斌温雪岩
刘斌 温雪岩



摘 要: 蛋白质序列的不同特征提取方式对蛋白质结构分类有很大的影响。为更好地表达蛋白质结构信息,基于特征融合思想构建特征向量,并使用一种基于多核支持向量机的方法,以多个核函数的线性加权代替传统的单一核函数,在对多类特征进行整合后构造SimpleMKL分类模型;利用梯度下降法迭代求解核函数的权值系数,并校准核函数参数和不同特征表达的融合效果。实验结果表明,该方法提高了蛋白质二级结构分类精度,在分类精度方面有明显优势,有助于准确预测蛋白质的二级结构。
关键词: 蛋白质; 二级结构预测; 多核支持向量机; 特征提取; 特征融合; 線性加权
中图分类号: TN911?34 文献标识码: A 文章编号: 1004?373X(2020)08?0139?04
Protein secondary structure prediction based on optimized multi?kernel SVM
LIU Bin, WEN Xueyan
(College of Electronic Information and Artificial Intelligence, Shaanxi University of Science and Technology, Xian 710021, China)
Abstract: The different feature extraction methods of protein sequence have great influence on protein structure classification. For better expression of protein structure information, the feature vectors are constructed based on feature fusion idea, and the traditional single kernel function is replaced by the linear weighting of multiple kernel functions by means of the method based on multi?kernel SVM (support vector machine); the SimpleMKL classification model is constructed after integrating the multi?class features, the weight coefficient of kernel function is solved iteratively by means of the gradient descent method, and the fusion effects of kernel function parameters and different feature expressions are calibrated. The experimental results show that the proposed method improves the classification accuracy of protein secondary structure, and has obvious advantages in classification accuracy, which is helpful to accurately predict the secondary structure of protein.
Keywords: protein; secondary structure prediction; multi?kernel support vector machine; feature extraction; feature fusion; linear weighting
通过模式识别和机器学习方法来获得蛋白质的结构信息是生物信息学的基本任务。SVM已被广泛用于蛋白质二级结构的预测[1],为扩展SVM分类器性能,有研究利用多层SVM方法增强了单个SVM方案的预测[1]。有研究基于改进模糊支持向量机的预测蛋白质二级结构的新方法[2],通过将位置特定信息和非位置特定信息与更好的核函数相结合,改进了基于SVM的预测[3]。该研究中提出的SVM?PHGS规范SVM核函数的参数,动态校准了不同核函数的融合结果,提高了分类精度,通过多次调整权重来求解最优核参数,其中核权重是线性加权的。但当数据维度过高时,已有文献研究成果的SVM求解速度会受到严重影响。
针对上述问题,本文使用一种自适应的L2范数正则化方法来考虑多核学习问题,即SimpleMKL算法,该算法将核矩阵定义为多个内核的线性组合,可以解决多特征分类问题[4]。……
