Penalization methods for identifying gene envrionment interactions and applications to melanoma and other cancer types

描述（由申请人提供）：在癌症GWAS研究中，已经投入了相当大的努力来开发用于识别G*E相互作用的统计方法。现有的方法有严重的局限性。首先，它们中的大多数采用基于模型的方法。模型假设在数据分析中难以验证，存在模型不规范的高风险，从而导致错误的标记识别。现有的鲁棒方法适用性有限。其次，现有的方法采用了无效的统计技术。最近，我们和其他人引入了有效的惩罚技术来识别重要的G*E相互作用，并表明它们明显优于现有的技术。但是，现有的处罚方法也存在局限性。它们采用基于估计的标记识别策略，该策略对调优参数选择敏感，缺乏稳定性，并且没有直接的错误发现率控制。此外，它们还会产生过高的计算成本。上述限制可能掩盖重要影响的识别，导致研究结果不一致，并导致次优预测模型。在这项研究中，我们将开发新的方法来检测G*E相互作用，以分析癌症病因、预后和生物标志物数据。该方法具有基于模型的方法所不具有的鲁棒性。他们将采用新的惩罚技术，并通过采用和直接比较多种标记识别策略，在现有惩罚方法的基础上进行改进。他们将能够进行边际分析和联合分析，以及个体标记和途径水平的分析。通过采用一种渐进的方法，他们将在全基因组数据的计算上负担得起。具体而言，我们将（目标1）开发强大的惩罚方法，以识别与癌症风险、生存和生物标志物相关的重要环境、遗传和G*E风险因素。我们将开发有效的计算算法，并严格证明鲁棒性和一致性。将进行广泛的模拟和比较。（目标2）开发用户友好的软件和项目网站。我们将使软件和其他研究成果易于访问。（目标3）分析黑色素瘤和其他癌症类型的数据，确定重要的G*E相互作用。我们将对鉴定的标记物进行综合评价，并与现有方法得到的结果进行比较。本研究将提供一套新颖的方法，这些方法将具有优越的统计和数值特性，并识别现有方法遗漏的重要标记。它们将广泛适用于大量的癌症类型和多种类型的遗传、基因组和表观遗传测量。在数据分析中，确定的标记物将为黑色素瘤和其他癌症的生物学机制提供重要见解，并为未来的验证研究和临床实践提供基础。

DESCRIPTION (provided by applicant): Considerable effort has been devoted to developing statistical methods for identifying G*E interactions in cancer GWAS studies. The existing methods suffer serious limitations. First, most of them take a model-based approach. The model assumptions are difficult to verify in data analysis, and there is a high risk of model mis- specification, which leads to false marker identification. The existing robust methods have limited applicability. Second, the existing methods adopt ineffective statistical techniques. Recently, we and others introduced effective penalization techniques for identifying important G*E interactions and showed that they significantly outperform the existing techniques. However, the existing penalization methods also have limitations. They adopt an estimation-based marker identification strategy, which is sensitive to tuning parameter selection, lacks stability, and does not have a direct false discovery rate control. In addition, they incur prohibitively high computational cost. The aforementioned limitations can mask the identification of important effects, lead to inconsistent findings across studies, and result in suboptimal predictive models. In this study, we will develop novel methods for detecting G*E interactions in the analysis of cancer etiology, prognosis, and biomarker data. The proposed methods will have the robustness property not shared by the model-based approach. They will adopt novel penalization techniques and advance from the existing penalization methods by adopting and directly comparing multiple marker identification strategies. They will be able to conduct both marginal and joint analyses and both individual marker- and pathway-level analyses. By adopting a progressive approach, they will be computationally affordable with whole-genome data. Specifically, we will (Aim 1) Develop robust penalization methods for identifying important environmental, genetic, and G*E risk factors associated with cancer risk, survival, and biomarker. We will develop effective computational algorithms and rigorously prove the robustness and consistency properties. Extensive simulations and comparisons will be conducted. (Aim 2) Develop user-friendly software and a project website. We will make the software and other research results easily accessible. (Aim 3) Analyze data on melanoma and other cancer types and identify important G*E interactions. We will comprehensively evaluate the identified markers and compare with the results obtained using existing methods. This study will deliver a set of novel methods which will have superior statistical and numerical properties and identify important markers missed by existing methods. They will be broadly applicable to a large number of cancer types and to multiple types of genetic, genomic, and epigenetic measurements. In data analysis, the identified markers will provide important insights into the biological mechanisms underlying melanoma and other cancers and serve as basis for future validation studies and clinical practice.