| 郭莎,郭翔,高洁,许东泽,梅玉婷,王翠,夏喜杰,李灵坤,贺鹏飞,吴宏宇,吴昊,王兰.机器学习在抗体制剂不溶性微粒分类识别中的应用[J].中国药事,2024,38(1):71-81 |
| 机器学习在抗体制剂不溶性微粒分类识别中的应用 |
| Application of Machine Learning in Classifi cation and Recognition of InsolubleParticles in Antibody Formulations |
| 投稿时间：2023-08-17 |
| DOI：10.16153/j.1002-7777.2024.01.010 |
| 中文关键词: 机器学习 蛋白质制剂 卷积神经网络 不溶性微粒 |
| 英文关键词: machine learning protein formulation Convolutional Neural Network insoluble particle |
| 基金项目:国家药典标准提高课题（编号 2020S11）；中国食品药品检定研究院关键技术研究基金（编号 GJJS-2022-1-4）；辽宁省博士启动基金计划项目（编号 2021-BS-131） |
|
| 摘要点击次数: 529 |
| 全文下载次数: 195 |
| 中文摘要: |
| 目的：对微流成像技术收集的抗体注射剂的不溶性微粒图片进行分类建模，以建立不溶性微粒的分类和溯源分析方法。方法：本研究首先制备并用微流成像系统获得气泡、硅油液滴、玻璃颗粒、反复冻融产生的蛋白颗粒4种不同类型颗粒的图片，并将这些图片分为训练集和测试集两部分。采用3种卷积神经网络（Convolutional Neural Network，CNN）模型，即ResNet50、DenseNet201和ShuffleNetV2，对训练集中的图片进行训练和识别，建立数据集，并用于对测试集图片的识别。此外，通过在实际样品中的应用，将CNN模型识别和人眼分类的准确性和速度进行对比。结果与结论：各模型对于测试集的识别准确率都达到96%以上，且DenseNet201模型具有更优的稳定性；与人眼识别相比，准确率没有显著差异，识别速度更快且具有显著差异。本研究表明CNN模型能够对蛋白质制剂中不溶性微粒图片进行分类和溯源分析，以便采取有针对性的措施，降低药品的潜在风险和安全隐患。 |
| 英文摘要: |
| Objective: To classify and modeling of insoluble particles in antibody injections collected bymicrofluidic imaging technology, in order to establish a classification and traceability analysis method forinsoluble particles. Methods: In this study, four diff erent types of images including bubble, silicone oil droplets,glass particles and protein particles generated by repeated freezing and thawing were obtained by the microfl owimaging system. These images were divided into training set and test set. The characteristics of images in training set were studied by and trained in three Convolutional Neural Network(CNN) models, namely ResNet50,DenseNet201 and Shuffl eNetV2, in order to establish data model. The recognition and classifi cation performanceof the model were tested by the images in the test set. Besides, the classification accuracy and speed of thethree CNN models were compared with that of human eyes through application in an actual case. Results andConclusion: The recognition accuracy of each model for the test set is above 96%, and the DenseNet201 modelhas better stability. Compared with human eye recognition, there is no signifi cant diff erence in accuracy, but asignifi cant acceleration in recognition speed. This study proves that CNN model can be applied in classifi cationand traceability analysis of insoluble particles in protein preparations, so that targeted measures can be taken toreduce the potential risks and safety hazards of drugs. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|