The Integrated Stress Response in Human Islets During Early T1D

摘要 该项目，在人类胰岛综合应激反应在早期1型糖尿病（T1 D），假设， 整合应激反应的激活和应激颗粒的形成是早期细胞反应 在T1 D中启动β细胞应激，决定细胞存活，并可以在T1 D前期和早期个体中进行监测 以最小的侵入性。一个多学科的团队科学方法正在采取测试这一假设， 收集大量的异质性数据，如mRNA、脂质组学、蛋白质组学和免疫学 测量.机器学习被用于提取多生物标志物面板，以帮助对压力进行分层， 人类胰岛，并将这些发现转化为T1 D和新发T1 D风险个体。虽然我们 在父授权内格式化用于该特定机器学习任务的多组学数据，该数据被 生成的数据，以及我们从以前的合作中收集的数据，通常不适合AI/ML。 方法的应用。然而，它们是用作AI/ML“旗舰”数据集的优秀候选者 准备就绪，既可以测试新的AI/ML方法来解决数据预处理挑战，也可以提取分子 T1 D的签名分析中的这两个差距是两个目标的中心主题。第一个目标集中在 生成AI/ML就绪的组学数据集，这些数据集经过适当注释，以解决稀疏性和偏差方面的挑战， 例如插补和批量校正。第二个目标关注AI/ML就绪的多组学数据集， 使用机器学习从数据中提取生物标志物和途径水平分子特征的新研究 专注于标准AI/ML方法，以及那些专门用于小样本量的方法。数据集机器 学习模型卡将用于更好地使AI/ML研究社区能够利用这些数据集， 高效的方式。对于这两个目标，都有一个关键的重点是生成可重用的软件方法，以生成 数据包，可以直接导入到最常见的AI/ML包中，并发布到AI/ML 社区通过各种资源，使反馈不断改进和完善AI/ML 准备软件开发计划。

ABSTRACT The project, Integrated Stress Response in Human Islets During Early Type 1 Diabetes (T1D), hypothesizes that the activation of the integrated stress response and formation of stress granules is an early cellular response initiating β cell stress in T1D that determines cell survival and can be monitored in pre- and early-T1D individuals with minimal invasiveness. A multidisciplinary Team science approach is being taken to test this hypothesis, collecting a large suite of heterogenous data, such as mRNA, lipidomics, proteomics and immunologic measurements. Machine learning is being used to extract a multi-biomarker panel to aid in stratifying stress in human islets and translating these findings to individuals at-risk for T1D and new-onset T1D. Although we are formatting the multi-omics data for this specific machine learning task within the parent grant, the data being generated, as well as our data collected from prior collaborations, are not generally AI/ML-ready for general application of methods. They are however excellent candidates to be used as “flagship” datasets for AI/ML readiness, both to test novel AI/ML approaches to tackle data pre-processing challenges and to extract molecular signatures of T1D. These two gaps in analyses are the central themes of two aims. The first aim focuses on the generation of AI/ML ready omics datasets that are properly annotated to address challenges in sparsity and bias, such as imputation and batch correction. The second aim focuses AI/ML ready multi-omic datasets to enable new studies in using machine learning to elicit biomarkers and pathway-level molecular signatures from the data focused on standard AI/ML methods, as well as those specialized for small sample size. Dataset machine learning model cards will be utilized to better enable to AI/ML research communities to utilize these datasets in an efficient manner. For both aims there is a key focus on generating reusable software approaches to generate data packages that can be directly imported into the most common AI/ML packages and released to the AI/ML community through a variety of resources that enable feedback to continually improve and refine the AI/ML readiness software development plan.