Single Cell Data Analysis Algorithms

单细胞RNA-seq数据使我们能够解决组织样本中的细胞亚群和基因表达异质性。然而，scRNA-seq揭示单个细胞基因表达的能力也使得其比大量RNA-seq数据更难以分析，因为细胞发育、分化和复制可能混淆从差异表达分析获得的结果。scRNA-seq比批量RNA-seq测量更容易产生噪声，批量RNA-seq测量是在许多细胞上平均的。我们在这里提出了一种方法，用于将单细胞RNA-seq数据模拟为基于基因-基因相互作用网络的动态系统，该网络模拟每个细胞随时间的发展。它有能力通过模拟基础动力学系统背景下的差异基因调控来产生不同的细胞类型。 生成模型通常学习数据和模型的联合概率分布，而判别模型学习条件概率分布。虽然判别模型更适合分类任务，但生成模型允许我们对单个类的分布进行建模，从而可以表示数据中的关系。由于单细胞数据可能具有时间非平稳性以及发育细胞和分化细胞类型之间的相关软边界，因此生成模型可能能够更好地从该数据中提取生物功能。 我们尝试在我们的生成模型的背景下从scRNA-seq数据推断基因调控网络。我们研究了如何从数据中逆向工程这样的网络，并量化了随着细胞数量的增加而增加的准确性。我们将我们的方法应用于真实的数据，并将我们的模型与已发表的方法的性能进行比较。

Single-cell RNA-seq data allows us to address cell subpopulations within a tissue sample and gene expression heterogeneity. However, the ability of scRNA-seq to reveal the gene expression of individual cells also makes it more difficult to analyze than bulk RNA-seq data, as cell development, differentiation and replication can confound the results obtained from differential expression analyses. scRNA-seq is prone to more noise than bulk RNA-seq measurement, which are averaged over many cells. We present here a method for simulating single-cell RNA-seq data as a dynamical system based on a gene-gene interaction network that models the development of each cell over time. It has the ability to generate different cell types through simulation of differential gene regulation in the context of the underlying dynamical system. Generative models in general learn the joint probability distribution of the data and the model, while discriminative models learn conditional probability distributions. While discriminative models are better suited to classification tasks, generative models allow us to model the distributions of individual classes, allowing representations of the relations in the data. As single-cell data may have temporal non-stationarity and related soft boundaries between developing cells and differentiating cell types, generative models may be able to better extract biological function from this data. We attempt inference of gene regulatory networks from scRNA-seq data in the context of our generative model. We study how well such networks can be reverse-engineered from the data, and quantify the increase in accuracy with increasing numbers of cells. We apply our method to real data and compare the performance of our model with published methodologies.