Survival Bump Hunting for Finding Informative Subgroups in High Dimensional Data.

描述（由申请人提供）：基于高维基因组数据的亚群发现可能为疾病过程提供新的见解。这通常是通过各种形式的聚类分析（有监督的和无监督的）完成的。极端亚群被定义为那些在本质上是同质的，但呈现极端价值结果的群体。在这个项目中，特别感兴趣的是开发方法来识别这些与生存结果有关的极端亚群（例如，那些在癌症治疗中表现异常良好的个体，可以根据高维基因组预测因子进行描绘）。如果这些亚群是真实的并且被发现，其意义将包括提高对疾病病因的理解，发现具有潜在治疗靶点的新生物标志物，并允许早期和个性化的治疗干预。从统计学上讲，这个问题可以用稀疏生存碰撞搜索框架来描述。我们汇集了一个生物统计学家团队，他们率先开发了连续反应的稀疏肿块搜索模型，以及两个国际公认的实验室作为合作者，他们分别致力于儿科髓母细胞瘤和非小细胞肺癌的多平台基因组分析。因此，我们提出以下具体目标：1)开发稀疏凹凸狩猎的新模型，允许具有连续和名义预测因子（例如基因表达和snp）的生存结果；2)开发一种稀疏生存碰撞狩猎方法，使我们能够通过三种不同的方法（稀疏指导、碰撞表型和稀疏中介分析）整合SNP和基因表达谱数据；3)为稀疏生存凹凸搜索模型的渐近性能提供详细的理论；一种新的基于栅格的模型验证研究方法并在详细的模拟和合作实验室提供的数据集上进行实证研究和比较；4)用R语言开发一个基于java的用户友好界面和一个命令行终端用户CRAN包，实现我们所有的方法及其扩展。

DESCRIPTION (provided by applicant): Subgroup discovery based on high dimensional genomic data can potentially provide novel insights into a disease process. Typically this has been done with various forms of cluster analysis (both supervised and unsupervised). Extreme subgroups are defined as those which are homogeneous in nature but which present extreme valued outcomes. Of particular interest in this project is to develop methodology to identify such subgroups which are extreme with respect to survival outcomes (e.g. those individuals that do unusually well on a cancer treatment and can be delineated based on high dimensional genomic predictors). If such subgroups are real and are uncovered, implications would include improved understanding of the disease etiology, discovery of new biomarkers with potential therapeutic targets, and allow early and personalized therapeutic interventions. Statistically, thi problem can be framed within a sparse survival bump hunting framework. We have brought together a team of biostatisticians who have pioneered the first sparse bump hunting models for continuous responses, as well as two internationally recognized laboratories as collaborators, who work on multi-platform genomic profiling for pediatric medulloblastoma and non-small cell lung cancer respectively. We thus propose the following specific aims: 1) To develop new models for sparse bump hunting that allow survival outcomes with both continuous and nominal predictors (e.g. gene expression and SNPs).; 2) To develop a sparse survival bump hunting approach that will allow us to integrate SNP and gene expression profile data by three different approaches - sparse coaching, bump phenotyping and sparse mediation analysis; 3) To develop detailed theory for asymptotic performance of these sparse survival bump hunting models; theory for a new fence-based methodology for studying model validation; and to empirically study and compare the performance in detailed simulations as well as on the datasets provided by our collaborator laboratories; 4) To develop a Java-based user-friendly interface and a command line end-user CRAN package in the R language that will implement all of our methodologies and its extensions.