Bayesian Models for Gene Expression with Microarray Data

描述（由申请人提供）：本项目涉及基因表达数据的参数化和半参数化建模。DNA微阵列和其他用于分析复杂核酸序列的高通量方法现在可以快速，有效和准确地测量生物样品中表达的许多基因的水平。微阵列数据分析的主要困难是，与问题的维度（基因数量）相比，样本量非常小。单个个体的基因数量通常在数千个，数据集中很少有个体。我们提出了几种新的参数贝叶斯建模方法，用于基因选择，肿瘤分类，贝叶斯网络，基因聚类和降维方法。大多数现有方法都不是基于模型的，因此无法解决有关不确定性的正式评估或特定模型拟合评估的具体问题。此外，基于模型的方法还提供了扩展到更复杂情况的可能性，例如，概率混合建模、处理缺失数据等。我们将为微阵列数据开发贝叶斯层次模型，该模型将灵活地适应不同层次的多个建模因素。在一些建模框架中，我们将保持模型空间的维度未知，以创建额外的灵活性。在这些灵活的模型类别中不可能得到解析答案，因此将使用基于模拟的马尔可夫链蒙特卡罗（MCMC）方法和维度跳跃算法来获得未知参数的估计（不确定性分布）。

DESCRIPTION (provided by applicant): This project is concerned with parametric and semiparametric modeling of gene expression data. DNA microarrays and other high-throughput methods for analyzing complex nucleic acid sequences now make it possible to rapidly, efficiently and accurately measure the levels of many genes expressed in a biological sample. The main difficulty with microarray data analysis is that the sample size is very small when compared to the dimension of the problem (the number of genes). The number of genes for a single individual is usually in the thousands and there are few individuals in the data set. We propose several novel parametric Bayesian modeling approaches for gene selection, tumor classification, Bayesian networks, gene clustering and dimension reduction methods. Most of the existing methods are not model-based and thus are unable to address specific questions regarding formal assessment of uncertainties or assessment of the fit of a specific model. Also model-based approaches offer the potential for extension to more complex situations, e.g., probabilistic mixture modeling, handling missing data, etc. We will develop Bayesian hierarchical models for microarray data, which will accommodate several modeling factors flexibly at different levels. In several of the modeling frameworks, we will keep the dimension of the model space unknown to create added flexibility. It is impossible to get analytical answers in these flexible classes of models so simulation based Markov Chain Monte Carlo (MCMC) methodology with dimensional jumping algorithms will be used to derive the estimates (uncertainty distributions) of the unknown parameters.