Computational Modeling of Heterogeneous Gene Expression in Single Cells

 描述（由申请人提供）：单细胞RNA-seq方案的最新发展使得能够在细胞水平上对生物体系统进行全基因组研究，为研究开辟了许多新的生物学问题。单细胞分辨率允许表征罕见或未知的细胞类型，能够解剖分化过程，并有助于解码负责细胞健康和患病状态的调控网络。然而，目前的单细胞RNA-seq研究受到现有计算方法的关键差距的限制。 我们已经设计了策略来解决当前单细胞RNA-seq分析方法的三个关键限制：（1）缺乏异构体特异性表达的模型，（2）无法将基因表达差异与细胞功能的可测量变化联系起来，以及（3）缺乏研究基因表达变化顺序进展的方法。为了解决第一个缺点，我们开发了SingleSplice，这是一种用于识别基因的算法，这些基因的异构体比例在一组单细胞中的变化超过了预期（Aim 1）。我们还开发了一种新的微筏平台，可以进行培养、功能表征、分离， 和随后的单细胞测序。利用该平台生成的数据，我们将进行监督机器学习，以识别与细胞间功能差异相关的基因（目标2）。为了解决第三个限制，我们将使用局部线性嵌入（一种非线性降维技术）来识别细胞通过顺序过程（如发育和对刺激的反应）进行的“轨迹”（目标3）。我们将把我们的方法应用于我们自己从微筏实验中产生的数据，以及来自发育中的肺组织和对免疫刺激做出反应的免疫细胞的公开可用的单细胞RNA-seq数据。 使用来自实验的数据，其中以恒定的已知量向细胞中添加spike-in转录物以模拟选择性剪接变化，我们发现SingleSplice以高灵敏度（73%）和特异性（93%）检测同种型转换。我们使用microafts对胰腺癌细胞系的单细胞进行测序，发现这种方法产生的高质量数据与Fluidigm C1相当。microraft技术还使我们能够在吉西他滨治疗后对胰腺癌细胞的RNA进行测序，并测量细胞的增殖，识别分裂的“耐药”细胞和不增殖的细胞，提供具有匹配功能和转录组测量的数据集。树突状细胞用细菌脂多糖（LPS）刺激的数据集的初步研究表明，局部线性包埋（LLE）可以根据细胞暴露于LPS的时间长度对细胞进行排序。

 DESCRIPTION (provided by applicant): The recent development of single cell RNA-seq protocols enabled genomewide investigation of organismal systems at the cellular level, opening many new biological questions for study. Single cell resolution allows characterization of rare or unknown cell types, enables dissection of differentiation processes, and aids in decoding regulatory networks responsible for healthy and diseased states of cells. However, current single cell RNA-seq studies are limited by crucial gaps in existing computational methods. We have devised strategies to address three key limitations of current single cell RNA-seq analysis methods: (1) lack of models for isoform-specific expression, (2) inability to link gene expression differences with measurable changes in cell function, and (3) lack of methods for studying sequential progression of gene expression changes. To address the first shortcoming, we developed SingleSplice, an algorithm for identifying genes whose isoform ratios vary more than expected by chance across a set of single cells (Aim 1). We have also developed a novel microraft platform that allows culturing, functional characterization, isolation, and subsequent sequencing of single cells. Using data generated from this platform, we will perform supervised machine learning to identify genes linked to functional differences among cells (Aim 2). To address the third limitation, we will use locally linear embedding, a nonlinear dimensionality reduction technique, to identify "trajectories" of cells proceeding through sequential processes such as development and response to stimuli (Aim 3). We will apply our methods to our own data generated from microraft experiments, as well as publicly available single cell RNA-seq data from developing lung tissue and immune cells responding to immune stimulation. Using data from experiments in which spike-in transcripts are added at constant, known amounts to cells to mimic an alternative splicing change, we found that SingleSplice detects isoform switching with high sensitivity (73%) and specificity (93%). We used microrafts to sequence single cells from a pancreatic cancer cell line and found that this approach produced high-quality data comparable to that from the Fluidigm C1. The microraft technology also enabled us to sequence RNA from pancreatic cancer cells after gemcitabine treatment and measure the proliferation of the cells, identifying both "drug resistant" cells that divide and cels that do not proliferate, giving a dataset with matched functional and transcriptomic measurements. Preliminary investigation of a dataset in which dendritic cells were stimulated with bacterial lipopolysaccharide (LPS) shows that locally linear embedding (LLE) can order cells according to the length of time they have been exposed to LPS.