Development of plasmon-enhanced biosensing for multiplexed profiling of extracellular vesicles

母公司R01项目旨在推进纳米等离子体传感技术，以实现健壮的多路传输 胞外囊泡(EV)分析和良好的重复性。开发的技术在油井上得到了验证。 建立临床前和临床样本，以证明新技术的可行性和潜力 用于临床应用。在项目过程中，我们生成了大量的成像数据 作为肿瘤和非肿瘤模型来源的细胞和EV，以及人类临床样本。这 为开发分析高维和高变量的深度学习模型提供了新的机会 为机器派生的分类提供成像数据，并发现新的见解。对于健壮的深度学习模型， 然而，除了数据的数量外，训练数据的质量也是至关重要的。数据不平衡或嵌入 技术混杂因素往往导致深度学习或机器学习模型基于非 相关或任意参数。另一个问题是当新的输入数据不同时进行错误分类 来自用于训练的数据，这被称为非分布样本。这些问题严重阻碍了 深度学习模型的稳健性与结果和精度的变化，导致失望和 降低了使用人工智能模型的热情。父R01项目的这一行政补充的目标是 是开发一个基于深度学习的数据管理软件，它可以消化大量的细胞和分子 成像数据，并产生平衡、无混杂的数据集，为新的深度学习或机器学习做好准备 任务。具体地说，我们将为不一定了解以下内容的普通用户设计算法 深度学习框架。我们将对生成的多通道电动汽车成像数据进行软件应用和测试 来自父R01项目以及来自过去NIH项目的其他蜂窝和电动汽车成像数据以及其他 实验室。最终的软件包和人工智能就绪数据将与其他研究人员公开分享，以及 我们将通过我们的IT核心团队不断提供反馈和更新。我们设想该软件将 为研究人员提供独特的机会，以随时创建高质量的培训数据，降低 研究人员，并促进将他们的数据用于深度学习或机器学习模型。

The parent R01 project aims to advance nanoplasmonic sensing technology for robust multiplexed extracellular vesicle (EV) analysis and good reproducibility. The developed technology is validated using well- established preclinical and clinical samples to demonstrate the feasibility and potential of the new technology for clinical applications. In the course of the project, we have generated a large amount of imaging data from cells and EVs that are originating tumor and non-tumor models, as well as human clinical samples. This provides a new opportunity to develop a deep learning model to analyze high-dimensional and high-variate imaging data for machine-derived classification and uncover new insights. For robust deep-learning models, however, the quality of training data, besides the quantity of the data, is critical. Unbalanced data or embedded technical confounding factors often lead to the deep- or machine-learning models' decisions based on non- related or arbitrary parameters. Another problem is making false classifications when new input data is different from the data used for training, which is called out-of-distribution samples. These issues significantly hampered the deep-learning models' robustness with variable results and accuracies, resulting in disappointment and reduced enthusiasm for using AI models. The goal of this Administrative Supplement to the Parent R01 project is to develop a deep-learning-based data management software that digests massive cellular and molecular imaging data and produces balanced, confounder-free data sets ready for new deep- or machine-learning tasks. Specifically, we will design the algorithm for general users who do not necessarily have knowledge of the deep-learning framework. We will apply and test the software for multi-channel EV imaging data generated from the Parent R01 project and other cellular and EV imaging data from the past NIH project and other laboratories. The final software package and AI-ready data will be publicly shared with other researchers, and we will continuously provide feedback and updates through our IT core team. We envision that the software will offer a unique opportunity for researchers to create quality training data ready, reduce the technical barrier for researchers, and promote the use of their data for deep- or machine-learning models.