Statistical Methods for Ultrahigh-dimensional Biomedical Data

描述（由申请人提供）：本提案开发了新的统计方法，从癌症研究的高通量数据（如微阵列和蛋白质组学数据）中选择一小群分子。该研究的挑战是这些研究中遗传的超高维度，特别是当引入基因-基因相互作用时。超高维对统计计算、方法发展和理论研究都有很大的影响。本文提出的独立筛选方法将解决这一问题，该方法还解决了超高维统计推断中的计算需求和稳定性以及随机误差积累问题。引入了一种迭代独立筛选方法来发现隐藏的特征基因，这些基因对临床结果不太重要，但对临床结果非常重要。它还使我们能够消除冗余的分子，这些分子与临床结果的关联程度很高，但联合起来却很弱。随着特征数量减少到可管理的水平，将引入惩罚伪似然方法来进一步选择相关基因。此外，还介绍了寻找分子协同基团的方法。独立筛选的思想及其迭代版本将应用于各种统计问题，从高通量数据的分析，从超高维回归和分类到生存时间的分析，全基因方差的估计，以及微阵列的规范化。所提出的方法的有效性将通过渐近理论和模拟研究进行评估。来自正在进行的癌症生物医学研究的数据集，如乳腺癌、多发性骨髓瘤、神经母细胞瘤、肺癌和肝癌，将使用新开发的统计和生物信息学工具进行严格分析。

DESCRIPTION (provided by applicant): This proposal develops novel statistical methods to select a small group of molecules from high-throughput data such as microarray and proteomic data from cancer research. The challenge of the study is the ultrahigh dimensionality inherited in these studies, particular when gene-gene interactions are introduced. The ultrahigh dimensionality has large impact on statistical computation, methodological developments, and theoretical studies. The challenge will be dealt by using the proposed novel independence screening methods, which also addresses the computational demand and stability, and the issues of stochastic error accumulation in ultra-high dimensional statistical inferences. An iterative independence screening method is introduced to find hidden signature genes that are marginally unimportant but jointly extremely important to the clinical outcomes. It also enables us to eliminate redundant molecules that are marginally highly but jointly weakly associated with clinical outcomes. With number of features surely reduced to a manageable level, penalized pseudo-likelihood methods will be introduced to further select relevant genes. In addition, methods for finding synergetic groups of molecules are introduced. The idea of independence screening and its iterated version will be applied to various statistical problems from the analysis of high throughput data, ranging from ultrahigh dimensional regression and classification to the analysis of survival time, estimation of genewide variance, and normalization of microarrays. The efficacy of the proposed methods will be evaluated via asymptotic theory and simulation studies. Data sets from on-going biomedical studies on cancer such as breast cancer, multiple myeloma, neuroblastoma, lung tumor, and liver carcigogen will be critically analyzed using the newly developed statistical and bioinformatic tools.