Image Tools for Computational Cellular Barcoding and Automated Annotation

项目总结 随着高通量单细胞成像和筛查技术的突破，我们可以精确地监测 自然细胞对不同刺激的反应行为。进一步提高分辨率和新的检测能力 丰富每个单元格的录制内容。许多图像处理步骤，帮助提取完整的 记录可以自动化到与成像本身相当的吞吐量。然而，因为生物 样本可以是复杂的和非均匀的，在寻址时指定像元子集是很有价值的 下游的生物学问题。使用现在可以从高吞吐量生成的数据量 测量时，这一子集选择步骤是获得关于 生物样本。目前，可靠地筛选单元格数据的黄金标准是通过 技术员。这种方法既昂贵又耗时，给回答重要的问题造成了严重的瓶颈 生物问题。为了克服这一瓶颈，我们将开发工具来使用三个唯一的 方法：化学注释法、注释法和细胞条码法。化学注解将 通过额外的生物标记物提供计算机可读的细胞标签。注释放大将使用小的、 精选数据集以生成大型数据集。蜂窝条形码将识别基于像素的签名以唯一 识别单个细胞。一旦对注记进行了计算处理，就可以对相关像元进行内联分类 随着收购的进行。然后，我们可以生成一个大型的带注释的数据集。无论是计算工具还是数据 将与科学界共享存储库，作为新模型的验证集并作为 可以在研究荧光成像细胞的研究小组之间开发的算法。目标是 这项工作的重点是生成技术并定义实验计算方法，使 通过湿实验室和机器学习技术之间的强烈相互作用，高度手动的细胞管理步骤。 我们提出的技术适用于大范围的高通量测量应用，因为 它可以通过计算而不是实验来管理样本。

PROJECT SUMMARY With technological breakthroughs in high-throughput single-cell imaging and screening, we can precisely monitor native cell behavior in response to diverse stimuli. Improvements in resolution and new detection capacity further enrich the recording from each cell. Many image-processing steps that help to extract the full breadth of the recording can be automated to a throughput comparable to the imaging itself. However, because biological samples can be complex and nonhomogeneous, it is valuable to specify subsets of cells when addressing downstream biological questions. With the amount of data that can now be generated from high-throughput measurements, this subset-selection step is a significant bottleneck to obtaining a quantitative result about the biological sample. Currently, the gold standard to reliably filter through cell data is manual annotation by a technician. This approach is costly and time-consuming, creating a significant bottleneck to answering important biological questions. To overcome this bottleneck, we will develop tools to automate annotation with three unique approaches: chemical annotation, annotation amplification, and cellular barcoding. Chemical annotation will deliver a computer-readable cell label via an additional biomarker. Annotation amplification will use small, curated datasets to generate large ones. Cellular barcoding will identify pixel-based signatures to uniquely identify individual cells. Once annotation is addressed computationally, relevant cells can be classified in-line with the acquisition. We can then produce a large annotated dataset. Both the computational tools and data repository will be shared with the scientific community as a validation set for new models and as a foundation for algorithms that could be developed across research groups studying cells with fluorescence imaging. The goal of this work is to generate the technology and define the experimental-computational methods that automate the highly manual steps of cell curation through a strong interplay between wet-lab and machine-learning techniques. The technology we propose is relevant to a broad scope of high-throughput measurement applications, because it enables curating samples computationally rather than experimentally.