Scalable tools for consistent identification of neuronal cell types in mouse and human

项目摘要 拟议的工作将通过以下方式解决我们对神经元表型和细胞类型理解上的一个关键差距 开发用于多通道集成的机器学习算法和基于云的软件 描述老鼠和人的大脑皮层神经元的大量和不断增长的数据集。通过优化和 创新地使用可能不完整的数据，并强调自动形态表征， 建议的工具将使更丰富和一致的特征神经元从转录，解剖，或 电生理侧写。 虽然由BICCN资助的大规模细胞类型研究项目依赖于独特神经元身份的概念 它决定了细胞在不同观察模式下的表型，总体上 生理、解剖和分子特征仍然难以捉摸。尽管这些大规模的项目 成功生成广泛的多通道数据集，缺乏原则性、准确性和广泛性 可用的计算比对和推理工具是总体成功的障碍 程序。第二个问题是解剖学特征，尽管是经典的方法 对于细胞类型的理解，在吞吐量方面明显落后于分子和生理方法。 本文提出的研究旨在通过建立在耦合自动编码器的基础上来解决对准问题 方法，它提供了一个高效的优化框架，以神经元身份的无处不在为中心。 重要的是，建议的软件可以利用不完全特征化的数据点，这在 实践中，产生抽象神经元同一性的统一可视化和分析。这个工具既灵活又灵活 (例如，特征集可以改变)并且是可扩展的(例如，可以为关节添加更多观察模式 对齐)。对齐的表示使神经元群体能够跨 不同的观察方式，这是现代神经科学中一个紧迫的问题。 我们建议使用端到端的计算管道来解决解剖吞吐量问题，从原始的 局部神经元乔木的图像到解剖描述符，可以很容易地对齐和解释 耦合自动编码器软件。通过利用我们广泛的黄金标准手动重建，我们将训练 有监督的深层人工神经网络在稀疏标记场景中分割神经元树枝。富豪们 训练实例，加上算法创新，将赋予这种自动化的卓越的可推广性 分割工具，加速基于光学显微镜的研究的科学。

Project Summary The proposed work will address a critical gap in our understanding of neuronal phenotypes and cell types by developing machine learning algorithms and cloud-based software for the integration of multiple modality characterizations large and growing datasets of cortical neurons in mouse and human. Through optimal and innovative use of potentially incomplete data and emphasis on automated morphological characterization, the proposed tools will enable richer and consistent characterizations of neurons from transcriptomic, anatomical, or electrophysiological profiling. While large-scale, BICCN-funded cell type research programs rely on the notion of unique neuronal identity which determines the cell’s phenotype across different observation modalities, overarching agreements across physiological, anatomical and molecular characterizations remain elusive. Although these large-scale programs succeeded in generating extensive multiple modality datasets, the lack of principled, accurate and widely available computational alignment and inference tools presents a roadblock to the success of the overall program. A second issue is that anatomical characterization, despite being the classical approach to understanding cell types, lags significantly behind molecular and physiological methods in terms of throughput. The research proposed here aims to address the alignment problem by building on the coupled autoencoder approach, which presents an efficient optimization framework centered on the ubiquity of neuronal identity. Importantly, the proposed software can utilize incompletely characterized data points, which is common in practice, to produce unified visualization and analysis of abstract neuronal identity. This tool will be both flexible (e.g., the feature set can be changed) and extensible (e.g., more observation modalities can be added for joint alignment). The aligned representations enable consistent clustering of the neuronal population across the different observation modalities, which is a pressing problem in modern neuroscience. We propose to address the anatomical throughput issue with an end-to-end computational pipeline, from the raw image of local neuronal arbors to the anatomical descriptor that can be readily aligned and interpreted by the coupled autoencoder software. By utilizing our extensive gold-standard manual reconstructions, we will train supervised deep artificial neural networks to segment neuronal arbors in sparse labeling scenarios. The rich set of training examples, together with algorithmic innovations, will endow superior generalizability of this automated segmentation tool, accelerating science for light microscopy-based studies.