Bayesian causal learning: A novel framework for drug-target discovery using Mendelian randomization on single-cell transcriptomics

新型治疗药物的开发是一项昂贵而耗时的努力，在大规模的II/III期临床试验中进行测试时往往会失败，这些试验需要花费数千万美元。近年来，人们越来越有兴趣利用遗传学证据来支持选择哪些药物靶点。在这个项目中，我们将使用帝国理工学院开发的一种独特的单细胞转录组学资源，该资源测量了人脑海马和前额叶皮层两个区域数十万个细胞的基因表达，这两个区域对记忆和行为至关重要。这个数据集的独特优势是能够放大单个细胞，从而找出哪些基因和哪种类型的细胞导致疾病。该数据集与遗传信息相关联，以定义单细胞表达的遗传调控，也称为表达数量性状位点（eQTL）。然而，分析工具落后于我们生成数据的能力，因此这些数据集目前没有得到充分利用。为此，我们将结合两个因果推理概念。第一个概念被称为孟德尔随机化（MR），它利用了一个人的DNA序列在受孕时被随机分配的事实。核磁共振利用这种“自然随机化”来推断暴露（即细胞中基因的表达）对结果（神经系统疾病）的因果关系。第二个概念是因果网络图和结构学习算法，它考虑一组基因，并推断哪些基因是连接的，以及效果的方向性，也就是说，如果基因a导致基因B，反之亦然。结合单细胞转录组学测量的细胞类型特异性的附加信息将使我们了解分子效应如何通过不同的细胞类型传播。研究遗传数据可以整合全基因组关联研究中公开可用的遗传关联数据。在本提案中，我们考虑了20多种神经表型和脑部疾病作为结果，包括例如阿尔茨海默病，帕金森病和癫痫。大型全基因组关联研究表明，我们的基因组中存在与一种以上疾病相关的变异。从这些数据集，我们可以了解是否存在导致多种疾病的共同分子机制。在这里，我们提出了一种新的计算工具包，称为单细胞MR (scMR)，用于分析单细胞转录组学eQTL数据，并结合疾病结果的全基因组关联研究。ScMR将实现以下三个目标的计算方法：1)确定哪些基因在哪些细胞中起作用并导致疾病。2)了解基因之间以及不同细胞类型之间的关系，以及分子过程如何通过不同的分子层进行传播。3)将多种反应整合到模型中，以识别导致多种疾病的基因。我们的发现不仅限于需要改变哪些基因来治疗神经系统疾病，还包括需要针对哪些特定的细胞类型。此外，我们希望了解不同细胞类型的下游分子过程，这将有助于更好地设计治疗干预措施。最后，整合许多相关神经系统疾病结果的数据将使我们能够定义影响多种疾病结果的共享分子机制。这些信息可用于现有药物的重新利用，确定具有降低一种以上疾病风险的协同效益的药物靶点，或突出潜在副作用的风险。

The development of novel therapeutic drugs is a costly and time-consuming endeavour that often fails when tested in large-scale Phase II/III clinical trials which cost tens of millions of dollars. In recent years there has been increasing interest in using evidence from genetics to support the choice of which drug targets to take forward.In this project, we will use a unique resource of single-cell transcriptomics developed at Imperial College which has measured the expression of genes in hundreds of thousands of cells in two regions of the human brain, the hippocampus and pre-frontal cortex, which are critical for memory and behaviour. The unique advantage of this dataset is the ability to zoom in on individual cells, and thereby work out which genes and which types of cells cause disease. This dataset has been linked with genetic information to define genetic regulation of single-cell expression, also known as expression quantitative trait loci (eQTL). However, analytical tools lag behind our ability to generate data and so these datasets are currently not being utilized to their full capacity.To this end, we will combine two causal inference concepts. The first concept called Mendelian randomization (MR) uses the fact that variations in a person's DNA sequence are randomly assigned at conception. MR uses this "natural randomization" to infer the causal effect of an exposure (i.e., the expression of a gene in a cell) on an outcome (neurological disease). The second concept is causal network graphs and structural learning algorithms which consider a group of genes and infer which genes are connected and the directionality of the effect, that is if gene A causes gene B or vice versa. Incorporating the additional information on cell-type specificity from the single-cell transcriptomics measurements will allow us to understand how molecular effects propagate through different cell-types. Working on genetic data allows the integration of publicly available genetic association data from genome-wide association studies. In this proposal, we consider a wide range of more than twenty neurological phenotypes and brain diseases as outcomes, including for example Alzheimer's disease, Parkinson's disease, and epilepsy. Large genome-wide association studies have shown that there is variation in our genome that is associated with more than one disease. From these datasets, we can learn if there are shared molecular mechanisms that cause more than one disease. Here, we propose a novel computational toolkit, called single-cell MR (scMR) for the analysis of single-cell transcriptomics eQTL data combined with genome-wide association studies on the disease outcomes. ScMR will implement computational methodology with the following three aims:1) To identify which genes act in which cells and cause disease. 2) To understand how genes relate with each other and across different cell-types and how molecular processes propagate through different molecular layers.3) To integrate many responses into the model to identify genes that cause more than one disease.Our findings will not be limited to which genes need to be altered to treat neurological disease, but also which specific cell-types need to be targeted. Moreover, we want to learn downstream molecular processes in different cells-types which will help to better design therapeutic interventions. Finally, integrating data on many related neurological disease outcomes will allow us to define shared molecular mechanisms affecting more than one disease outcome. This information can be used for the repurposing of existing drugs, the identification of drug-targets with co-benefits that reduce risk of more than one disease, or highlight the risk of potential side-effects.