Discovering What Matters: Informative and Reproducible Variable Selection with Applications to Genomics

这个项目将开发统计方法，以发现在大量集合中，哪些变量与感兴趣的结果有意义地相关。这个问题的一个例子是，在我们测量的数百万个基因变异中，确定哪些基因变异会影响疾病风险。开发的方法将能够同时分析所有变量，说明它们的相互依存关系，并导致确定“可操作的”变量。提出的方法保证，平均而言，很大一部分已发现的特征确实会影响结果。正确识别重要变量的能力将增加许多领域的知识，并使专家能够制定干预措施。例如，了解记录在患者身上的哪些变量与他/她的治疗反应更相关，可以帮助开发具有更高成功率的个性化医疗干预。开发的方法将扩大统计学家和数据科学家可用的工具箱，因为他们试图从包含大量变量的数据集中提取有意义的信息。该方法建立在“仿制”框架的基础上，这是一种非常灵活和新颖的方法，不需要为感兴趣的结果和可能的协变量之间的关系指定一个模型。提供的推理保证针对选定变量，并控制错误发现率(FDR)，其中，如果选定变量与给定剩余协变量的结果无关，则认为发现是错误的。这为结果的重现性及其可解释性提供了保证。开发的方法将用于分析遗传学数据集，目的是获得更完整的DNA变异如何影响医学相关表型的模型。该项目得到了数学科学部和分子与细胞生物科学部的支持。

This project will develop statistical methods to discover which variables, in a large collection, are meaningfully related to an outcome of interest. An example of the problem is the identification of which genetic variants, among the millions we measure, influence disease risk. The methods developed will allow analysis of all variables at the same time, accounting for their interdependence, and leading to the identification of "actionable" ones. The approaches put forward come with the guarantee that, on average, a large fraction of the discovered features truly influence the outcome. The ability to correctly identify important variables will increase knowledge in many domains, and allow experts to devise interventions. For example, understanding which of the variables recorded on a patient are more relevant with respect to his/her response to therapy, can help develop personalized medical interventions with a higher success rate.The methods developed will enlarge the tool-box available to statisticians and data scientists as they attempt to extract meaningful information from datasets comprising a very large number of variables. The approach builds on the "knock-off" framework, a very flexible and novel approach that does not require specifying a model for the relation between an outcome of interest and possible co-variates. The inferential guarantees provided are on the selected variables, with control of the False Discovery Rate (FDR), where a discovery is considered false if a selected variable is independent of the outcome given the remaining covariates. This provides assurance on the reproducibility of results, as well as on their interpretability. The approaches developed will be used to analyze genetics datasets with the goal of obtaining more complete models of how DNA variation influences medically relevant phenotypes. This project is supported by the Division of Mathematical Sciences and the Division of Molecular and Cellular Biosciences.