Ochratoxin A (OTA), which is prone to contaminating food products, poses a significant threat to human health. To accurately detect and provide timely early warnings for OTA contamination (COTA) in food, a novel four-modal nano-biosensor detection system based on MXenes was developed in this work. MXenes is a new family of two-dimensional materials including vanadium carbide (V2C) MXenes, which has great potential in the field of biosensing and detection due to its large specific surface area, rich surface functional groups, and good chemical stability. By in-situ chemical etching, metal ion intercalation, and synergistic physical exfoliation techniques, V2C nano-materials (V2C-NMS) with excellent fluorescence quenching characteristics were synthesized, and V2C-NMS@ssDNA with excellent peroxidase like activity was constructed by functionalizing V2C-NMS with adapter. Based on this, a fluorescence/colorimetric biosensor for OTA was constructed, and the fluorescence/colorimetric signal output by the sensor had good linear relationship with COTA, with the detection limit (LOD) as low as 6.77 pg mL−1. Furthermore, a fully connected artificial neural network (FCANN) was developed based on a series of RGB values obtained from the fluorescence/colorimetric mode of the biosensor, and the fluorescence/colorimetry channel of the FCANN could accurately predict the COTA of the sample on-site or remotely in just a few seconds, with the LOD as low as 7.10 pg mL−1. Importantly, the four-mode method performed well in real sample detection, with recovery rates ranging from 95.33% to 105.79%, and the detection results of the four modalities could be mutually verified.
翻译：
赭曲霉毒素A (OTA) 易于污染食品，对人类健康构成重大威胁。为了准确检测并及时预警食品中的OTA污染 (COTA)，本研究开发了一种基于MXenes的新型四模态纳米生物传感器检测系统。MXenes是一类新型二维材料，包括碳化钒 (V2C) MXenes，由于其比表面积大、表面官能团丰富、化学稳定性好，在生物传感和检测领域具有巨大潜力。通过原位化学蚀刻、金属离子插层和协同物理剥离技术，合成了具有优异荧光猝灭特性的V2C纳米材料 (V2C-NMS)，并通过适配体功能化V2C-NMS构建了具有优异类过氧化物酶活性的V2C-NMS@ssDNA。在此基础上，构建了用于OTA的荧光/比色生物传感器，传感器输出的荧光/比色信号与COTA具有良好的线性关系，检测限 (LOD) 低至6.77 pg mL−1。此外，基于从生物传感器荧光/比色模式获得的一系列RGB值，开发了一个全连接人工神经网络 (FCANN)，FCANN的荧光/比色通道可以在几秒钟内准确预测样品的现场或远程COTA，LOD低至7.10 pg mL−1。重要的是，该四模式方法在实际样品检测中表现良好，回收率在95.33%到105.79%之间，并且四种模式的检测结果可以相互验证。

Aflatoxins, which are both potent and highly carcinogenic, commonly contaminate peanuts and related products, presenting considerable risks to food safety. This research introduces an innovative approach for predicting the content of aflatoxin B1 (AFB1) by fine-tuning the Babbage-002 large language model (LLM). This approach can streamline food safety monitoring processes, allowing for more reliable assessments of contamination levels. The relative aflatoxin content is labeled through sub-pixel decomposition, and the spectral sequences are converted into a Json file for fine-tuning the LLM. The dataset includes 7000 spectral sequences for training and 3000 for test. After 1400 optimization steps, the model achieved a training loss of 0.5177 and a test loss of 0.5304. The fine-tuned model has a root mean square error (RMSE) of 0.0588, a coefficient of determination (R²) of 0.8898, and a ratio of performance to deviation (RPD) of 2.7408 on the test set. Experimental results show that the LLM-based fine-tuning model's ability to predict aflatoxin content exceeds the one-dimensional and two-dimensional convolutional neural network models. In addition, this study also established an online AFB1 spectral detection platform based on a fine-tuned model, which facilitates the AFB1 content analysis visualization. The fine-tuned LLM model also demonstrated excellent ability to continuously learn and adapt to new data.
翻译：
黄曲霉毒素效力强且具有高致癌性，常见于花生及相关产品中，对食品安全构成重大风险。本研究提出了一种创新方法，通过微调Babbage-002大语言模型（LLM）来预测黄曲霉毒素B1（AFB1）的含量。该方法可以简化食品安全监测流程，实现对污染水平更可靠的评估。通过子像素分解对黄曲霉毒素相对含量进行标注，并将光谱序列转换为Json文件以用于LLM微调。数据集包含7000条用于训练的光谱序列和3000条用于测试的序列。经过1400步优化后，模型达到了0.5177的训练损失和0.5304的测试损失。微调后的模型在测试集上的均方根误差（RMSE）为0.0588，决定系数（R²）为0.8898，性能与偏差之比（RPD）为2.7408。实验结果表明，基于LLM微调模型预测黄曲霉毒素含量的能力超过了一维和二维卷积神经网络模型。此外，本研究还建立了一个基于微调模型的在线AFB1光谱检测平台，方便了AFB1含量分析的可视化。微调后的LLM模型还表现出了出色的持续学习和适应新数据的能力。

This study evaluated the quality and freshness of pork patties stored at 4°C over varying durations. Using an electronic nose, the flavor profiles of the samples were analyzed, focusing on sensor responses to volatile compounds. Machine learning models, including partial least squares discriminant analysis (PLS-DA), support vector machines (SVM), and back-propagation artificial neural networks (BP-ANN), were applied to predict freshness. Total volatile basic nitrogen (TVB-N) was used as the key indicator in these models. Results showed a decline in the quality of pork patties with extended storage. The electronic nose sensors, particularly S4, S6, and S14, demonstrated strong sensitivity to volatile substances, highlighting their relevance in assessing freshness. Among the predictive models, BP-ANN achieved the highest accuracy (94%), followed by SVM (91%) and PLS-DA (89%). Based on the highest-performing BP-ANN model, the test results confirmed that the model’s predicted TVB-N values closely align with the actual values, demonstrating excellent predictive accuracy. This further highlights the practical applicability of the present study. These findings illustrate that machine learning models can reliably predict the freshness of pork patties, offering a non-destructive and efficient alternative to traditional methods, which are often labor-intensive and time-consuming.
翻译：
本研究评估了猪肉饼在4°C下不同贮藏时间内的品质和新鲜度。使用电子鼻分析了样品的风味图谱，重点关注传感器对挥发性化合物的响应。应用了包括偏最小二乘判别分析（PLS-DA）、支持向量机（SVM）和反向传播人工神经网络（BP-ANN）在内的机器学习模型来预测新鲜度。挥发性盐基氮（TVB-N）被用作这些模型中的关键指标。结果表明，随着贮藏时间的延长，猪肉饼的品质下降。电子鼻传感器，特别是S4、S6和S14，对挥发性物质表现出很强的敏感性，突出了它们在评估新鲜度方面的相关性。在预测模型中，BP-ANN达到了最高的准确率（94%），其次是SVM（91%）和PLS-DA（89%）。基于性能最佳的BP-ANN模型，测试结果证实该模型预测的TVB-N值与实际值高度吻合，表现出优异的预测准确性。这进一步突显了本研究的实际应用性。这些发现表明，机器学习模型能够可靠地预测猪肉饼的新鲜度，为传统方法提供了一种非破坏性且高效的替代方案，而传统方法通常劳动密集且耗时。

食品操作员的卫生习惯和行为对确保生产过程中的食品安全至关重要。良好卫生规范（GHP）要求食品操作员遵循严格的卫生标准，包括正确的洗手程序和使用个人防护装备（PPE），如发网、口罩、手套、防护服、裤子和工作鞋，以防止食品工厂中的交叉污染。然而，在大规模操作中，人工检查通常效率低下且不切实际。为此，本研究提出了一种基于深度学习的系统，利用计算机视觉和卷积神经网络（CNN）在工人进入生产区域前自动检测手部卫生和PPE合规性。这确保了GHP标准的遵守，增强了食品安全，提高了检查效率并降低了成本。使用一个包含20，222个条目的数据集（涵盖8个角度的12个洗手动作和彩色编码的PPE状态）来训练和评估10个模型：YOLOv8、YOLOv7、YOLOv6、YOLOv5、ResNet、DenseNet、MobileNetv2、EfficientNetv2、VGG和Vision Transformer（ViT）。YOLOv6和YOLOv8在洗手识别方面达到了最高准确率（0.999），而DenseNet在PPE检测方面达到了最高准确率（0.956）。该系统为监测卫生习惯提供了一种高效、自动化的解决方案，有助于防止交叉污染并确保工厂环境内的食品安全。

Secondary metabolites of Alternaria alternata in juice extracted from fruit of pomegranate ( Punica granatum ) ‘Acco’ and ‘Wonderful’ affected by heart rot were identified using the UHPLC-Q-TOF-MS analytical method. Juice from healthy and infected fruit was analyzed comparatively with particular focus on mycotoxin content. Overall, 22 metabolites were detected, including the mycotoxins tenuazonic acid, alternariol, alternariol monomethyl ether (AME) and altenuene (ALT). Multivariate analyses, including Principal Component Analysis and Random Forest models, clearly separated the metabolic profiles of healthy and infected fruit. Symptomatic fruits, especially those of ‘Wonderful’, exhibited markedly higher levels of mycotoxins, such as alternarlactone A and altertoxin, suggesting that these metabolites are potential biomarkers for early detection of heart rot. In contrast, they were characterized by a substantial lower level of punicic acid, indicating heart rot infection affected negatively the nutraceutical value of juice due to the depletion of the content in this omega 5 fatty acid. This study provides better insights into the effects of A. alternata infection on the quality of pomegranate juice as well as new evidences in favour of a regulation of the mycotoxin level in pomegranate juice.
翻译：利用UHPLC-Q-TOF-MS分析方法，对受“心腐病”影响的‘Acco’和‘Wonderful’石榴果实榨取的果汁中链格孢菌的次级代谢物进行了鉴定。对健康和感染果实的果汁进行了比较分析，特别关注霉菌毒素含量。共检测到22种代谢物，包括霉菌毒素细交链孢菌酮酸、交链孢酚、交链孢酚单甲醚和细交链孢菌烯酸。多元分析，包括主成分分析和随机森林模型，清晰地区分了健康和感染果实的代谢谱。有症状的果实，尤其是‘Wonderful’品种，表现出显著更高水平的霉菌毒素，如交链孢内酯A和交链孢毒素，表明这些代谢物是早期检测心腐病的潜在生物标志物。相反，它们的石榴酸（punica acid）水平显著降低，表明心腐病感染由于消耗了这种Omega-5脂肪酸的含量，对果汁的营养价值产生了负面影响。本研究为更好地理解链格孢菌感染对石榴汁品质的影响提供了见解，并为石榴汁中霉菌毒素水平的监管提供了新的证据。