生物活性肽作为21世纪人类健康的新宠儿, 研究证明其对人体生命活动有着很好的作用, 其检测方法也是备受关注, 太赫兹时域光谱技术因为其独特的性质在检测生物活性肽中有着不可比拟的优势。 选用牛骨肽、 海参肽和牛肽这三种生物活性肽, 通过透射式太赫兹时域光谱系统得到其在0.5～2 THz的吸收系数曲线。 从太赫兹吸收系数曲线来看, 鱼肽吸收系数大于海参肽和牛骨肽。 因为生物活性肽的氨基酸种类和肽键的相互作用, 导致其在太赫兹频段内没有明显的吸收峰, 为了更好的对其进行检测区分, 建立分类判别模型, 寻找出最适合这类物质的方法。 在对太赫兹原始吸收系数数据进行S-G平滑处理, 归一化预处理之后, 随机选取四分之三预处理好的数据划分为训练集, 其余为预测集, 导入分类判别模型。 模型包括分类器和最优参数选取两部分, 分类器选取支持向量机, 随机森林和极限学习机等有监督的分类方法, 使用遗传算法、 粒子群算法和网格搜索等智能优化算法选取支持向量机最优参数。 为了减少原始光谱数据维数并提高模型的运算速度, 使用主成分分析进行预处理, 将降维之后的结果导入分类模型。 综合考虑其准确率和运行时间等因素, 虽然基于粒子群算法的支持向量机具有最高的准确率98.3%, 但是运行时间较长为180 s; 使用极限学习机能够有着最短的运行时间0.2 s, 但是准确率为73.3%。 基于网格搜索的支持向量机准确率为95%, 运行时间为11 s, 能够在准确率较高的情况下使用较短的时间, 证明基于网格搜索的支持向量机对生物活性肽太赫兹吸收光谱具有快速, 准确的分类结果。 研究结果表明, 利用太赫兹时域光谱技术结合机器学习算法能够实现快速、 无损检测生物活性肽, 为生物活性肽的检测提供了一种新思路, 同时也为THz-TDS结合机器学习对吸收峰不明显的多肽之间的鉴别提供参考。

Bioactive peptides, as the new darling of human health in the 21st century, have been proved that they have a good effect on human life activities, and their detection methods are also of great concern. Terahertz time-domain spectroscopy technology has incomparable advantage in detecting bioactive peptides because of its unique properties. In this paper, three bioactive peptides, bovine bone peptide, sea cucumber peptide and fish peptide, were used to obtain the absorption coefficient curve of 0.5~2 THz by the transmission terahertz time domain spectroscopy system. From the terahertz absorption coefficient curve, the absorption coefficient of the fish peptide is higher than that of sea cucumber peptide and fish bone peptide. Because of the interaction between the amino acid species of bioactive peptides and peptide bonds, there is no obvious absorption peak in the terahertz frequency band. In order to better detect and distinguish them, a classification discriminant model is established to find the most suitable for such substances. After the S-G smoothing and normalization preprocess performed on the terahertz original absorption coefficient data, two-thirds of the pre-processed data are randomly selected into training sets, and the rest are prediction set. The classification discriminant model is introduced. The model includes two parts: the classifier and the optimal parameter selection. The classifier selects the supervised classification method such as support vector machine, random forest and extreme learning machine, and uses the intelligent optimization algorithm such as genetic algorithm, particle swarm optimization and grid search to select the support vector machine optimal parameters. In order to reduce the original spectral data dimension and improve the computational speed of the model, Principal Component Analysis is used for preprocessing, and the results after dimensionality reduction are imported into the classification model. Considering the factors such as accuracy and running time, although the support vector machine based on particle swarm optimization has the highest accuracy rate of 98.3%, the running time is longer than 180 seconds; the ultimate learning machine can have the shortest running time of 0.2 seconds. However, the accuracy rate is 73.3%. The support vector machine based on grid search has an accuracy rate of 96% and a running time of 11 seconds. It can use a shorter time in the case of higher accuracy, and proves that the support vector machine based on grid search is better for detecting bioactive peptide. The results show that the use of terahertz time-domain spectroscopy combined with machine learning algorithms can achieve rapid and non-destructive detection of bioactive peptides, providing a new idea for the detection of bioactive peptides. It also demonstrates that THz-TDS combined with machine learning is a way better way for the identification of inconspicuous peptides.