NIRG: Taming the skew - empowering Mendelian randomization trials by controlling stratification bias

背景孟德尔随机化（MR）是一种使用常见遗传变异作为感兴趣的暴露的替代物来统计地估计暴露对疾病的因果效应的工具。MR是随机对照试验的一种强大、快速和廉价的替代方案，随机对照试验可能无法执行。然而，MR假设遗传多态性（这是暴露的代表）和疾病结局之间没有混杂。群体分层偏倚是由于研究个体的祖先差异而发生的，并且它可以使全基因组关联研究（GWAS）和MR试验产生偏倚，因为它违反了遗传同质性的假设。类似的问题也存在于双样本MR中，它使用来自两组独立研究的汇总统计量（一组研究SNP暴露相关性，另一组研究SNP结果相关性）来确定总体效应估计。在这种情况下，不太可能充分控制人群分层，因此将出现偏倚。为了继续利用大数据资源，如英国生物银行（50万人的全基因组数据），我们必须识别和纠正MR的人群分层偏见。目标本项目将开发工具和统计方法，以测试，识别，测量，解决，并提出建议，如何处理人口分层偏倚，特别是在MR分析。然而，为此，我们首先需要为所有人类群体开发一个统一的祖先模型。然后，我们将开发工具来估计祖先，并将我们的工具与现有的工具进行比较。接下来，我们将开发工具来优化基于血统的病例对照匹配，从而克服在血统方面病例和对照不匹配的问题，这可能会使结果产生偏差。我们还将使用一系列模拟和真实的数据库将我们的工具与现有的工具进行比较。根据结果，我们将提出建议，以指导分析何时使用适当的方法。我们将为MR分析量身定制特定的解决方案，开发一种测量偏差的统计数据，并测量在MR中使用GWAS汇总统计数据的偏差。最后，我们将应用我们的方法研究科学中的两个开放问题，其中人口分层可能会导致工作偏差，即：维生素D水平影响身高，牛奶摄入量（使用LCT基因）和糖尿病之间是否存在关系？计算机代码、附带的指导说明和培训材料将免费提供，以确保分析人员能够轻松实施我们的方法。

BackgroundMendelian randomization (MR) is a tool which uses common genetic variants as surrogates for an exposure of interest to statistically estimate the causal effect of an exposure on a disease. MR is a powerful, rapid, and cheap alternative to randomised control trials, which may be unfeasible to execute. However, MR assumes that there is no confounding between the genetic polymorphism (which is a proxy for the exposure) and the disease outcome. Population Stratification bias occurs due to ancestral differences in the studied individuals, and it can bias Genome-wide Association Studies (GWAS) and MR trials as it violates the assumption of genetic homogeneity. A similar problem exists in two sample MR, which uses summary statistics from two separate sets of studies (one set of studies for the SNP-exposure association and another for the SNP-outcome association) to determine the overall effect estimate. In this scenario it is unlikely that population stratification will be adequately controlled for and thus bias will arise. To continue utilizing Big Data resources, like the UK BioBank (whole genome data for 500,000 people), for MR we must identify and correct the population stratification bias. ObjectivesThis project will develop tools and statistical methods to test, identify, measure, address, and make recommendations about how to handle population stratification bias particularly in MR analyses. However, for that, we first require to develop a unified ancestry model for all human populations. We will then develop tools to estimate ancestry and compare our tools with existing ones. Next, we will develop tools to optimise case-control matches based on ancestry, thus overcoming problems of mismatching case and controls in terms of ancestry, which may bias the results. We will also compare our tools with existing ones using a series of simulated and real databases. Based on the results, we will make recommendation to guide analyses when to use the appropriate methods. We will tailor specific solutions to MR analyses in developing a statistic that measures the bias and measuring the bias in using GWAS summary statistics in MR. Finally, we will apply our methods to study two open questions in science where population stratification may bias efforts, namely: do vitamin D levels affect height and is there a relationships between milk intake (using the LCT gene) and diabetes?Computer code, accompanying guidance notes, and training materials will be made freely available to ensure analysts can easily implement our methods.