Computational modeling of DNA methylation-mediated gene regulation

摘要 大量完整的甲基组正在通过临床测序项目获得，例如 通过癌症基因组图谱、蓝图表观基因组计划和国际癌症基因组 财团。此外，第三代纳米孔测序仪，它检测DNA甲基化和基因 单个实验中的变异，几乎准备好进行常规临床测序，并将提供完整的 适用于所有需要进行全基因组测序的患者。然而，仅限当前分析工具 执行初步的甲基组处理和编目差异甲基化区域(DMRS)。为了 将甲基组分析转化为临床有用的诊断/预后测试，我们需要开发预测性 用于解释已识别的甲基化变化的功能和病理后果的工具。朝向这个方向 Goal，我们已经发表了一系列的论文，证明了基于机器学习的模型利用高效率的 启动子周围所有甲基化变化的分辨率签名远远优于传统的DMR 方法：研究方法。我们的模型准确地预测了可能受甲基化和 揭示便于机械性解释的预测性甲基化特征。尽管如此，有几个 在我们能够实现将全基因组甲基化数据转换为常规数据的目标之前，仍然存在挑战 临床应用：(1)据我们所知，目前还没有模型整合远端增强子，其激活受以下因素的影响 DNA甲基化。这种综合分析对于理解甲基化变化的后果是必要的 在癌症中，其基因组经常经历广泛的甲基化变化。另外，这样的造型 对于理解5-羟甲基胞嘧啶(5HmC)的作用是至关重要的，它可能同时发挥抑制和 激活神经元的作用取决于它是在启动子还是在增强子中发现的。(2)我们目前的模式 (和传统方法)表示独立于DNA序列的甲基化数据，尽管是机械的 研究表明，甲基化的变化可以有不同的功能影响，这取决于 序列会根据当地调控语法的上下文而变化。在这个提案中，我们将相遇 这些挑战是通过首先开发一个包含5-甲基胞嘧啶和5-HmC的预测模型来实现的 推动者和增强者，以确定这些标记如何协同起作用。特别是，我们将研究 假设5HmC在皮质神经元中作为启动子的抑制者和增强子的激活者的双重作用。 然后，我们将使用自然语言处理的新进展来模拟DNA序列和甲基化，以 预测表达式状态。我们的结果将揭示哪些调控元件和转录因子结合位点 受到DNA甲基化的影响，以及不同位点的变化如何协作影响表达变化。 我们将使用报告分析和CRISPR相结合的方法来实验验证我们的计算机预测- 基于表观基因组的编辑工具。因此，我们开发的软件工具将形成分析的重要工具包 以及在实验室和临床上对全基因组甲基化研究的机械性解释。

Abstract Large numbers of complete methylomes are being acquired through clinical sequencing projects, such as through The Cancer Genome Atlas, Blueprint Epigenome Project, and International Cancer Genome Consortium. Furthermore, third-generation nanopore sequencers, which detect DNA methylation and genetic variation in a single experiment, are nearly ready for routine clinical sequencing and will provide complete methylomes for all patients where whole-genome sequencing is indicated. Current analysis tools however only perform preliminary methylome processing and catalogue differentially methylated regions (DMRs). In order to transform methylome analysis into a clinically useful diagnostic/prognostic test, we need to develop predictive tools to interpret the functional and pathological consequences of identified methylation changes. Towards this goal, we have published a series of papers demonstrating that machine-learning based models utilizing high- resolution signatures of all methylation changes around a promoter vastly outperform conventional DMR methods. Our models accurately predict expression states at genes potentially regulated by methylation and reveal predictive methylation signatures that facilitate mechanistic interpretation. Nonetheless, several challenges remain before we can achieve our goals of translating genome-wide methylation data for routine clinical use: (1) To our knowledge, no current models integrate distal enhancers, whose activation is affected by DNA methylation. Such integrative analysis is necessary to understand consequences of methylation changes in cancers, whose genomes frequently undergo wide-spread methylation changes. In addition, such modelling will be essential to understand the role of 5-hydroxymethylcytosine (5hmC), which may play both repressive and activating roles in neurons depending on whether it is found at promoters or enhancers. (2) Our current models (and conventional approaches) represent methylation data independent of DNA sequence despite mechanistic studies demonstrating that methylation changes can have different functional effects depending on which sequences change and depending on the context of the local regulatory grammar. In this proposal, we will meet these challenges by first developing a predictive model that incorporates 5-methylcytosine and 5hmC at promoters and enhancers to determine how these marks act in concert. In particular, we will examine the hypothesized dual role of 5hmC as a repressor at promoters and as an activator at enhancers in cortical neurons. We will then use new advances in natural language processing to model DNA sequence and methylation to predict expression states. Our results will reveal which regulatory elements and transcription factors binding sites are affected by DNA methylation and how changes at different sites collaborate to affect expression changes. We will experimentally validate our in silico predictions using a combination of reporter assays and CRISPR- based epigenome-editing tools. Thus, the software tools we develop will form an important toolkit for the analysis and mechanistic interpretation of whole-genome methylation studies, both in the laboratory and clinic.