Statistical Methods For Microarray Data Analysis

Because of the microarray technology and deep sequencing technology, thefield of current biological research has been developed to molecular level since21thcentury. Diferent from the single individual study, it is now possible for scientiststo measure millions of genomic items,such as genes and promoters, simutaneouslyin a single experiment and provide extra high dimension data set. Questions givento statisticians is how to understand these data systematically and extract usefulbiological information from them. Unfortunately, because of the technology lim-itation, plenty of noise, system bias, even serious error could be constructed intothe original data. As a result, How to get reasonable conclusion from the datasetis a challenge problem.In this thesis, we propose two statistical methods which can be used to an-alyze the data generated from diferent microarray experiments design. One istime-course gene expression data, the other is ChIP-chip genome expression data.The former one is generated from stem cell diferentiation process in order to com-pare diferent kinds of diferential processes in which genes with diferent expressionpattern are detected. We present a method for co-expression network based com-parison of temporal expression patterns (NACEP). NACEP proposes a B-splinebased linear model to describe the processwise gene expression pattern, comparesthe temporal patterns of a gene between two experimental conditions, taking intoconsideration all of the possible co-expression modules that this gene may par-ticipate in and finally provide robust results. The relative parameter inferenceis implemented by a delicately constructed Gibbs sampler algorithm. ChIP-chipgenome expression data is used to evaluate the profile of histone H3lysine9mod-ifications in human mesenchymal stem cell osteogenic diferentiation. p-value iscomputed based on a finite mixture model, the number of components is estimatedby BIC method. Both theory and simulations are developed to evaluate the ef- fect of each method. Results show that our method can give reasonable inferenceresults while the relative measurement error can be controlled consequently.