| 本文已被：浏览 372次 下载 7次 |
码上扫一扫！ |
| 结合太赫兹光谱与密度泛函理论的黄曲霉毒素B1指纹特征验证 |
|
张思怡1,2,3, 杨硕1,2,3, 梁晓颖1,2,3, 姚志凤1,2,3, 陈煦1,2,3, 宋怀波1,2,3
|
|
1.西北农林科技大学 机械与电子工程学院，陕西 杨凌 712100;2.农业农村部农业物联网重点试验室，陕西 杨凌 712100;3.陕西省农业信息感知与智能服务重点试验室，陕西 杨凌 712100
|
|
| 摘要: |
| 【 目的】融合密度泛函理论（DFT）与太赫兹时域光谱（THz-TDS）技术，为黄曲霉毒素 B1（AFB1）的 高精度定量分析与指纹特征验证提供参考。【 方法】基于DFT对AFB1分子进行结构优化与振动频率计算，获取其在0.1~5.0 THz频段内的理论吸收谱，并从分子振动模式角度对吸收峰进行解释；系统比较S-G平滑、移动平滑、归一 化、标准正态变量变换（SNV）、多元散射校正（MSC）、一阶导数与二阶导数等 7 种光谱预处理方法对AFB1太赫兹光 谱偏最小二乘回归（PLSR）建模效果的影响；最后针对传统特征提取方法（SPA/MCUVE/CARS/SFLA）对低浓度信号表征不足的问题，构建基于一维卷积神经网络（1D-CNN）和残差网络（ResNet）的黄曲霉毒素B1定量预测模型，通 过卷积与池化操作，自适应地对原始光谱进行局部感知，并利用深度学习捕捉邻近波段的隐性特征，提升指纹峰的提 取能力。【 结果】在0.5~3.0 THz时，试验吸收峰与理论预测高度一致，验证了C-O/C-O振动模式作为AFB1特异性指纹峰的可靠性。在AFB1质量浓度为0.25~10.00 μg/mL 时，1D-CNN 模型预测集的决定系数为0.92，均方根误 差为0.77。相较于传统方法，1D-CNN得到的特征频率与理论吸收峰具有良好的对应关系，提取的特征频率几乎完 全覆盖2.114 THz处的理论峰值，而在1.108 THz处仅存在0.05 THz的偏移。【 结论】1.108，1.767和2.114 THz处 振动模式与试验光谱的匹配度较高。1D-CNN模型对黄曲霉毒素B1具有最佳预测能力，所建立的AFB1太赫兹指纹 峰的特征验证方法为复杂基质中痕量毒素检测提供了新范式，为食品安全领域现场快检技术的建立提供了参考。 |
| 关键词: 黄曲霉毒素B1 太赫兹时域光谱 密度泛函理论 指纹峰分析 机器学习 |
| DOI：10. 13207/j. jnwafu. 2026. 09. 008 |
| 分类号: |
| 基金项目:国家自然科学基金项目（32201662）；陕西省重点研发计划项目（2024NC-YBXM-218） |
|
| Verification of the fingerprint characteristics of aflatoxin B1 using terahertz spectroscopy combined with density functional theory |
|
ZHANG Siyi1,2,3, YANG Shuo1,2,3, LIANG Xiaoying1,2,3, YAO Zhifeng1,2,3, CHEN Xu1,2,3, SONG Huaibo1,2,3
|
|
1.College of Mechanical and Electronic Engineering，Northwest A&F University，Yangling，Shaanxi 712100，China;2.Key Laboratory of Agricultural Internet of Things，Ministry of Agriculture and Rural Affairs，Yangling，Shaanxi 712100，China;3.Shaanxi Key Laboratory of Agricultural Information Perception and Intelligent Service，Yangling，Shaanxi 712100，China
|
| Abstract: |
| 【 Objective】This study integrates Density Functional Theory（DFT） and Terahertz Time-Domain Spectroscopy（THz-TDS），to provide a reference for high-precision quantitative analysis and fingerprint verification of Aflatoxin B1（AFB1）.【Method】Based on DFT，the AFB1 molecule was structurally optimized and its vibrational frequencies were calculated to obtain the theoretical absorption spectrum in the 0.1-5.0 THz range，with absorption peaks interpreted from the perspective of molecular vibrational modes. Seven spectral preprocessing methods，including S-G smoothing，movement smoothing，normalization，standard normal variate（SNV），multiplicative scatter correction（MSC），first derivative，and second derivative，were systemati-cally compared to determine their effects on the PLSR modeling of AFB1 terahertz spectra. To address the inade-quacy of traditional feature extraction methods（SPA，MCUVE，CARS，SFLA） in characterizing lowconcen-tration signals，quantitative prediction models for AFB1 based on one-dimensional convolutional neural network（1D-CNN） and residual metwork（ResNet） were constructed. Through convolution and pooling operations，the models adaptively performed local perception on raw spectra，and leveraged deep learning to capture implicit features from adjacent spectral bands，thereby enhancing the extraction efficiency of fingerprint peaks.【Result】In the 0.5-3.0 THz range，the experimental absorption peaks were highly consistent with theoretical predictions，validating the reliability of the C-O/C-O vibrational modes to obtain characteristic fingerprint peaks for AFB1. In the concentration range of 0.25-10.00 μg/mL，the prediction set of 1D-CNN model achieved a determination coefficient（R） of 0.92，with a root mean square error of 0.77. Compared to traditional methods，the characteristic frequencies extracted by the 1D-CNN model corresponded well with the theoretical absorption peaks：the extracted frequencies almost fully covered the theoretical peak at 2.114 THz，while only a 0.05 THz deviation was observed at 1.108 THz.【Conclusion】The vibrational modes at 1.108，1.767，and 2.114 THz show a high degree of agreement with the experimental spectrum. The 1D-CNN model demonstrates optimal predictive capability for aflatoxin B1. The established method for verifying the characteristic terahertz fingerprint peaks of AFB1 provides a new paradigm for detecting trace toxins in complex matrices and serves as a valuable reference for developing on-site rapid detection technologies in the field of food safety. |
| Key words: aflatoxin B1 terahertz time-domain spectroscopy density functional theory fingerprint peak analysis machine learning |
