Multiscale Resolution and Deep Network Approaches for Deconvolving Different Cell Types in Bulk Tumor using Single-cell Sequencing Data (scDEC)

摘要 脑组织由异质细胞群组成。了解脑细胞的变化 类型或状态组成的神经退行性变的未来治疗有重要意义 阿尔茨海默病（AD）。例如，小胶质细胞 发展 （scRNA-Seq） 个脑袋 包括 差 型特异性基因表达变化发生在 的AD。在过去的几年里，单细胞RNA测序的发展和应用 已经彻底改变了大脑研究，从而使我们能够研究细胞的异质性， 随着scRNA-Seq的出现，我们现在可以识别健康和患病的脑细胞类型或状态 罕见的细胞类型或状态群体，并鉴定这些细胞群内的转录改变。 除了AD的细胞复杂性之外，该疾病的分子复杂性也仍然存在。 明白 在 到目前为止，研究已经确定了许多与生殖细胞相关的生殖细胞基因组变异， 对AD的易感性然而，在这方面， 从 水平 体细胞鉴定 举例来说， 然而，在这方面， 大脑 健康和患病scRNA-Seq数据 DNA的改变是不同的 胚系，术语称为"脑体细胞基因组镶嵌"。大脑体细胞变异发生在低- 等位基因频率，这只能使用单细胞DNA测序检测。C 变体 在散发性AD中，APP基因拷贝数在单个神经元细胞中呈镶嵌性增加。 表征这些体细胞变体是具有挑战性的， 单细胞DNA测序数据。因此，具有很大的利用价值 研究单个脑细胞中的体细胞嵌合现象， 脑细胞，尤其是神经元， 引起呕吐的 将解释体细胞嵌合体对AD的贡献。 在 越来越多的 识别 与AD易感性相关的基因组变异。 在这个提议中，我们描述了一种新的深度网络方法，用于对不同的细胞类型进行去卷积， 使用单细胞RNA测序数据，在散装AD样品中检测到的状态。我们的方法不仅可以估算 细胞类型或状态的比率，而且还使用scRNA-Seq数据测定AD样品中体细胞克隆嵌合现象的比率。我们 将体细胞克隆嵌合现象定义为含有体细胞基因组变体的细胞群，即克隆 等 asCNVs、SNP或插入缺失。体细胞基因组变异 一种基于多尺度分辨率信号处理的新算法CaSpER。 将从scRNA-Seq数据中识别 然后我们将提取 使用我们 细胞或 使用生成式深度学习方法，从scRNA-Seq数据中克隆类型基因签名， 对抗网络（GANs）。 我们还将使用 适应放射基因组学方法，我们将图像特征与细胞类型比率相关联。我们提出的 该方法将导致临床护理的重大改进，以指导AD的治疗和预后。 从scRNA-Seq数据鉴定的这些细胞类型基因签名将在稍后进行。 使用卷积神经网络（CNN）推断大块AD组织中的细胞类型的分数。

Abstract Brain tissue is composed of heterogeneous cell populations. Understanding the changes in brain cell type or state composition during neurodegeneration have important implications for future treatment of Alzheimer's disease (AD). For example, microglia cell development (scRNA-Seq) brain. including poorly type specific gene expression changes occur early in the of AD. Over the past few years, the development and application of single-cell RNA sequencing have revolutionized brain research thus enabling us to study the cellular heterogeneity of the With the advent of scRNA-Seq, we can now identify healthy and diseased brain cell types or states rare cell type or state populations and identify transcriptional alterations within these cell groups. addition to the cellular complexity of the AD, the molecular complexity of the disease also remains understood. In Until now, studies have identified numerous germline genomic variants associated with susceptibility to AD. However, from level identification of somatic example, However, brain healthy and diseased scRNA-Seq data DNA alterations that are distinct the germline, term referred as `brain somatic genomic mosaicism'. Brain somatic variants occur at a low- allele frequency, which could only be detected using single-cell DNA sequencing. C variants For sporadic AD the copy number of APP gene is mosaically increased in single neuron cells. it is challenging to characterize these somatic variants since there is a lack of AD brain or healthy single-cell DNA sequencing data. Therefore, there is a great value in utilizing to investigate somatic mosaicism in single brain cells and brain cells, especially neurons, harbor diverse omprehensive in brain cells will explain the contribution of somatic mosaicism to AD. in the growing number of identify genomic variants that are associated with susceptibility to AD. In this proposal, we describe a novel deep network approach for deconvolving different cell types or states in bulk AD sample using single-cell RNA sequencing data. Our approach will estimate not only the ratio of cell types or states but also the ratio of somatic clonal mosaicism in AD samples using scRNA-Seq data. We define somatic clonal mosaicism as the groups of cells, i.e. clones, harboring somatic genomic variants such asCNVs, SNPs, or indels. Thesesomatic genomic variants novel multiscale resolution signal processing based algorithm named CaSpER. will be identified from scRNA-Seq data We will then extract using our cell or clone type gene signatures from scRNA-Seq data using a generative deep learning approach called General Adversarial Networks (GANs). used We will also adapt radiogenomics approaches where we correlate image features with cell type ratios. Our proposed approach will lead to major improvements in clinical care to guide the treatment and prognosis of AD. These cell type gene signatures identified from scRNA-Seq data will be later to infer fractions of cell type in bulk AD tissue using convolutional neural networks (CNNs).