Collaborative Research: Analysis of longitudinal multiscale data in immunological bioinformatics - Feature selection, graphical models, and structure identification

该项目旨在开发一套统计分析工具系统，以应对复杂生物信息学数据分析中的几个重要挑战，这些数据涉及各种响应变量和数万个独立变量。人们的兴趣往往在于识别与响应变量相关的关键自变量，并了解这种关联以及自变量之间的相互作用。生物信息学数据的极端规模和复杂性给数据分析带来了严重的挑战。为了克服这些挑战，我们建议(I)在我们可以应用我们的新的建模和分析方法之前，系统地和适当地整合多尺度数据，因为我们探索的数据是由大量来自表型、细胞、蛋白质和遗传水平的独立研究收集的，这些信息来自非常不同的时间和维度尺度；(2)结合双变量纵向回归模型中的相关检验和Benjamini-Hochberg-Yekutieli程序，为混合类型的纵向数据制定特征筛选标准，(3)开发图形模型，允许变量是连续和离散纵向变量的混合，节点代表变量，每条边表明两个相关变量的相关性，条件是其他变量；以及(Iv)通过在具有连续或离散响应和预测器的高维向量的混合效应回归模型的框架下借助于数据本身来调查每个预测器的功能形式，所产生的过程允许用户同时确定每个预测器效应的形式为零、线性或非线性。

This project aims to develop a system of statistical analysis tools to tackle several important challenges in analysis of complex bioinformatics data, which involves a variety of response variables and tens of thousands independent variables. The interest often lies in identifying the key independent variables associated with the response variables, and understanding such associations as well as the interactions among the independent variables.The extreme magnitude and complexity of bioinformatics data have posed serious challenges for data analysis. To overcome these challenges, we propose (i) to systematically and properly integrate multi-scale data before we can apply our novel modeling and analysis methods since the data we explore are collected by numerous independent studies at phenotypic, cellular, protein, and genetic levels with information from very different time and dimension scales; (ii) to develop feature screening criteria for a mixed type of longitudinal data using the combination of correlation tests in bivariate longitudinal regression models and the Benjamini-Hochberg-Yekutieli procedure, (iii) to develop graphical models that allow the variables being a mix of continuous and discrete longitudinal variables, with the nodes representing variables and each edge indicating the dependence of the two relevant variables conditional on the other variables; and (iv) to investigate the functioning form of each predictor by resorting to the data themselves under the framework of a mixed effects regression model with a continuous or discrete response and a high dimensional vector of predictors, with the resulting procedure allowing a user to simultaneously determine the form of each predictor effect to be zero, linear or nonlinear.