Selection and Integration of -Omics Data for Biomarkers Discovery

各种组学技术的最新进展允许同时定量数百到数千个分子（例如，基因），彻底改变了科学家寻找分子生物标志物来测量致病过程或对治疗反应的方式。尽管可用于生物标志物研究的技术资源的数量和质量得到认可，但需要新的统计和计算方法来询问和理解这些技术产生的丰富信息。我的研究计划提出了创新的统计方法来选择和整合相关的分子变量，而传统的方法可能无法检测到它们与感兴趣的表型的关联。特别是，我将结合工具变量估计、稳健性和惩罚估计的元素，以解释测量误差、混杂因素、离群观察和复杂的数据结构，这些在组学数据集中很常见，可能会危及临床有用生物标志物的发现。工具变量估计器类似于经典回归估计器，但它们借用补充变量（仪器）的力量来解释测量误差和混杂因素。例如，遗传或基因组数据可以用作蛋白质组学生物标志物发现研究的工具，以评估蛋白质和临床表型之间的因果关系。由于组学研究需要分析大量的候选解释变量（如蛋白质）和潜在的工具（如基因），其中只有少数是相关的（稀疏模型），因此不能使用经典的IV估计器。此外，组学数据集通常包含与技术问题或具有罕见分子谱的患者相关的外围观察结果。因此，开发对异常值和数据中的杠杆点具有鲁棒性的估计器是至关重要的。在文献中提出了惩罚回归估计器来估计从复杂数据集（例如LASSO）中选择最重要的解释变量的稀疏模型。尽管在我过去的工作和文献中对惩罚IV估计器和鲁棒惩罚估计器有一些初步结果，但没有一个提议的估计器集成了所有三个组成部分：惩罚、鲁棒性和工具变量。通过探索与基因、蛋白质和疾病状态相关的合理生物机制，有一个融合这些概念的统一框架对于促进蛋白质组学生物标志物的发现至关重要。虽然我的大部分研究集中在统计蛋白质组学上，但所提出的分析技术与数据科学中常见的复杂高维数据的分析相关，为更广泛的社区带来价值。

Recent advances in various -omics technologies allow the simultaneous quantitation of hundreds to thousands of molecules simultaneously (e.g., genes), revolutionizing the way that scientists search for molecular biomarkers to measure pathogenic processes or responses to therapies. Despite the recognized number and quality of the technical resources available for biomarker studies, new statistical and computational methods are needed to interrogate and understand the rich information generated by these technologies. My research program proposes innovative statistical methods to select and integrate relevant molecular variables, where classical approaches may fail to detect their association with the phenotype of interest. In particular, I will combine elements from instrumental variables estimation, robustness, and penalized estimation to account for measurement errors, confounding factors, outlying observations, and complex data structures, which are common in -omics dataset and can jeopardize the discovery of clinically useful biomarkers. Instrumental variables estimators are analogous to classical regression estimators but they borrow strength from supplemental variables (the instruments) to account for measurement errors and confounding factors. For example, genetic or genomic data can be used as instruments in a proteomic biomarkers discovery studies to assess causal effects between proteins and clinical phenotypes. Since -omics studies require the analysis of a large number of candidate explanatory variables (e.g., proteins) and potential instruments (e.g., genes), of which only a few would be relevant (sparse model), classical IV estimators cannot be used. Furthermore, -omics datasets usually contain outlying observations associated, for example, with technical problems or patients with rare molecular profiles. Thus, the development of estimators that are robust to outliers and leverage points in the data is of fundamental importance. Penalized regression estimators have been proposed in the literature to estimate sparse models selecting the most important explanatory variables from complex datasets (e.g., LASSO). Despite some initial results on penalized IV estimators and robust penalized estimators in my past work and the literature, none of the proposed estimators integrates all three components: penalization, robustness, and instrumental variables. Having a unifying framework that blends these concepts is essential to boost proteomic biomarker discoveries by exploiting the plausible biological mechanisms that relate genes, proteins, and disease state. Although most of my research is focused in Statistical Proteomics, the analytical technics proposed are relevant for the analysis of complex high-dimensional data commonly found in Data Science bringing value to a broader community.