Sequence-based Machine Learning for Inference of Dynamic Cell State Gene Network Models

大多数疾病相关的GWAS变体对报告基因或转录因子的表达具有相对适度的影响。 CRISPR扰动测定。此外，增强子在体内的破坏通常具有令人惊讶的弱的表达。 表型的后果我们假设，一个关键的缺失因素是我们缺乏定量的 多个TF如何在增强子处相互作用，以及多个增强子如何在基因座处相互作用， 通过改变基因网络活动以非线性方式对扰动作出反应。预测影响 因此，基因组变异的研究需要对一个变异的影响如何依赖于其他变异进行定量建模 变异通过它们对改变的细胞调节状态的组合作用。其核心目标是 建议是开发计算方法来推断这些组合的定量模型， 通过对基因活性、增强子活性和 早期人类细胞状态转换过程中核心细胞命运调节转录因子（TF）活性 发展我们的初步研究表明，虽然启动子敲除对靶基因有很强的影响， 基因表达，单个增强子敲除通常较弱，并影响时间转换 动力学，但不是最终的稳定状态。我们证明了基于序列的基因网络模型 机器学习与这些观察是一致的。我们对我们的序列提出了改进 基于模型开发动力学速率方程和随机模拟基因网络模型预测 增强子扰动的可变且经常是暂时的影响。我们将生成高时间分辨率 ATAC、H3 K27 ac和scRNA-seq数据来训练这些模型，并验证基因网络预测 在天然基因组背景下CRISPRi的网络响应。我们先关注我们的胚胎- 干细胞到定形内胚层（ESC-DE）系统，然后我们将开发方法来推广 将这些重点模型应用于更大的ENCODE监管数据集。我们的工作将使 定量了解调节元件的活性改变如何影响稳定性， 该元件在其中运作的基因调控网络的动态，以及它们如何在 控制发育重要的和疾病相关的细胞状态转变。

Most disease associated GWAS variants have relatively modest effects on expression in reporter or CRISPR perturbation assays. In addition, enhancer disruption in vivo often has surprisingly weak phenotypic consequences. We hypothesize that a critical missing element is our lack of quantitative models of how multiple TFs interact at an enhancer, and how multiple enhancers interact at a locus to respond to perturbations in a nonlinear way through altered gene network activity. Predicting the impact of genomic variation thus requires quantitative modeling of how one variant's impact depends on other variants through their combined effect on altered cellular regulatory state. The central aim of this proposal is to develop computational methods to infer quantitative models of these combinatorial interactions by training on temporally-resolved measurements of gene activity, enhancer activity, and core cell fate-regulating transcription factor (TF) activity across cell state transitions in early human development. Our preliminary studies show that while promoter knockdown has robust effects on target gene expression, individual enhancer knockdown is often weaker and affects temporal transition dynamics, but not the final steady state. We show that gene network models based on sequence-based machine learning are consistent with these observations. We propose improvements to our sequence based models to develop kinetic rate equation and stochastic simulation gene network models to predict the variable and often temporal effects of enhancer perturbation. We will generate high time resolution ATAC, H3K27ac, and scRNA-seq data to train these models, and validate the gene network predictions of network response with CRISPRi in a native genomic context. We will first focus on our embryonic- stem-cell to definitive-endoderm (ESC-DE) system, and we will then develop methods to generalize application of these focused models to larger ENCODE regulatory datasets. Our work will enable a quantitative understanding of how the altered activity of regulatory elements affects the stability and dynamics of the gene regulatory networks within which the element operates, and how they play a role in controlling developmentally important and disease relevant cell state transitions.