Bayesian Differential Causal Network and Clustering Methods for Single-Cell Data

项目描述 DMS/NIGMS 2：单细胞数据的贝叶斯差分因果网络和聚类方法 重要性 A.1要解决的问题的重要性 单细胞RNA测序（scRNA-seq）技术促进了新的生物学发现， 批量RNA-seq是不可能的，例如在单细胞水平上发现新的基因调控活性， 细胞类型。然而，为了翻译scRNA所推进的基础生物学知识， 为了改进疾病的诊断、治疗和预防，需要新的方法， 研究正常和病理细胞/组织之间的分子差异，以及对照和 病例/治疗组。虽然两个样本组中差异表达基因的鉴定 已被广泛研究，迄今为止，绝大多数现有的方法用于鉴定基因regu- 到目前为止，实验网络（GRNs）和细胞类型的研究主要集中在单一的scRNA-seq数据下产生的scRNA-seq数据上。 实验条件原则上，这些方法可以一次应用于一个实验条件， 在此基础上，可以进行事后比较，以发现实验引起的差异 干预措施。然而，与联合建模方法相比，这种两步过程被认为较少 由于缺乏适当的不确定性传播， 第一步到第二步。此外，大多数scRNA-seq网络模型在本质上是相关的，并且不依赖于它们。 推断因果基因调控关系。因此，迫切需要开发新的模式， 确定实验干预对因果基因调控和细胞组成的影响， 跨实验组对scRNA-seq数据进行建模。在没有这些工具的情况下，机械地不- 了解基因调控和细胞分化，充分认识scRNA-seq的翻译价值 研究可能仍然很困难。 A.2先前研究的严谨性 目标1.许多现有的scRNA-seq网络方法将标准关联度量调整为零关联。 膨胀的scRNA-seq数据，例如Pearson相关性[1]和互信息[2]。一个共同的局限性 这些方法的一个缺点是，它们只量化了边际依赖性，这很容易受到虚假的间接影响。 协会[3]。处理条件关联的图形模型是 边际关联测度许多方法已被提出用于一般目的[4，5]。 包括非高斯数据的发展[6-9]。特别是对于scRNA-seq数据，两个非定向的 最近提出了基于邻域选择的图形模型，包括Co-I Cai的工作[10，11]， 然而，这并不意味着因果基因调控。为了确定因果关系， 方法[12，13]。然而，这些方法或者忽略scRNA-seq的计数性质， 数据，需要已知的伪时间（这在真实的scRNA-seq数据中很少已知），或者理论上不存在。 调查横截面观察的因果可识别性。对于差分网络，许多方法 [14-18]包括PI的先前工作[19]已经开发了大量RNA-seq数据，这些数据显示出很大的差异。 联合分析优于独立分析。然而，存在更少的差分网络， scRNA-seq数据的工作方法，例如，[20][21][22][23][24] PT的常见局限性， scdNet是他们只考虑边际依赖（因此容易受到错误发现的影响）， 发现因果关系。我们的初步结果（§C.1）表明，所提出的贝叶斯 网络模型能够识别横截面scRNA-seq中的因果基因调控关系 数据，并且通常优于最先进的替代方法。 目标二。很少有方法可用于构建特定于细胞的网络，因为它很难估计 网络的样本量基本上是1。最近，开发了一种假设检验方法[22] 来估计特定于细胞的网络。该方法对每个单元进行近似网络推理， 对它的邻居。然而，它只考虑对称（无向）边际依赖，因此 不能推断出因果调节关系，容易产生虚假的关联。PI之前的工作 [23]解决了批量RNA-seq数据中的“样本量一”问题，假设因果网络是 其他协变量的平滑函数。然而，该方法在没有协变量的情况下不适用， 不允许反馈回路，这是GRN中的常见主题。现有的工作[24，25]，包括PI的[19]， 1

Project Description DMS/NIGMS 2: Bayesian Differential Causal Network and Clustering Methods for Single-Cell Data A Significance A.1 Importance of the Problem to Be Addressed Single-cell RNA-sequencing (scRNA-seq) technologies have facilitated new biological discoveries that were impossible with bulk RNA-seq, such as discovering at the single-cell level new gene regulatory activities and cell types. However, in order to translate the fundamental biological knowledge advanced by the scRNA- seq to improved disease diagnosis, treatment, and prevention, new methods are required to comparatively study the molecular differences between normal and pathological cells/tissues, and between control and case/treatment groups. Although identification of differentially expressed genes across two sample groups has been extensively studied, to date, the vast majority of the existing methods for identifying gene regu- latory networks (GRNs) and cell types have, so far, focused on scRNA-seq data generated under a single experimental condition. In principle, these methods can be applied to one experimental condition at a time, based on which post hoc comparisons can be made in order to find the differences caused by experimental interventions. However, compared to joint modeling approaches, this two-step procedure is deemed less efficient and more susceptible to false discoveries due to lack of proper uncertainty propagation from the first step to the second. Moreover, most scRNA-seq network models are correlative in nature and do not infer causal gene regulatory relationships. There is, therefore, a critical need to develop new models for identifying the effects of experimental interventions on causal gene regulation and cell composition by jointly modeling scRNA-seq data across experimental groups. In the absence of such tools, mechanistically un- derstanding gene regulation and cell differentiation, and fully realizing the translational values of scRNA-seq studies will likely remain difficult. A.2 Rigor of Prior Research Aim 1. Many existing scRNA-seq network approaches adapt standard association measures to zero- inflated scRNA-seq data, e.g. Pearson correlation [1] and mutual information [2]. A common limitation of these methods is that they only quantify marginal dependencies, which is susceptible to spurious indirect associations [3]. Graphical models which deal with conditional associations are powerful alternatives to the marginal association measures. Numerous methods have been proposed for general purposes [4, 5] including the development on non-Gaussian data [6–9]. Specifically for scRNA-seq data, two undirected graphical models including Co-I Cai's work [10, 11] were recently proposed based on neighborhood selec- tion which, however, do not infer causal gene regulation. To identify causal relationships, several alternative methods [12, 13] were developed. However, these methods either ignore the count nature of scRNA-seq data, require a known pseudotime (which is rarely known in real scRNA-seq data), or do not theoretically in- vestigate causal identifiability for cross-sectional observations. For differential networks, many approaches [14–18] including the PI's prior work [19] have been developed for bulk RNA-seq data which showed great advantages of joint analyses over independent analyses. However, there exist much fewer differential net- work methods for scRNA-seq data, e.g., PT [20] and scdNet [21] . The common limitation of PT and scdNet is that they only consider marginal dependence (hence susceptible to false discoveries) and do not discover causality. Results from our preliminary results (§C.1) demonstrate that the proposed Bayesian network model is capable of identifying causal gene regulatory relationships in cross-sectional scRNA-seq data and often outperforms the state-of-the-art alternative methods. Aim 2. Very few methods are available to construct cell-specific networks because it is difficult to estimate networks with, in essence, sample size one. Recently, a hypothesis testing approach [22] was developed to estimate cell-specific networks. The method makes approximate network inference of each cell based on its neighbors. However, it only considers symmetric (undirected) marginal dependence, and therefore cannot infer causal regulatory relationships and is susceptible to spurious associations. The PI's prior work [23] addressed the "sample-size-one" problem in bulk RNA-seq data assuming the causal networks are smooth functions of additional covariates. However, the method is not applicable without covariates and does not allow feedback loops, a common motif in GRN. Existing work [24, 25] including the PI's [19] has 1