Differential expression analysis for proteomics data by two-dimensional gel electrophoresis

The objective of differential expression analysis of proteomics data generated by 2D gel electrophoresis is to identify the differentially expressed proteins over various experimental conditions. In this dissertation, we present the procedures of differential expression analysis and propose solutions of the missing data problem in the 2D gel electrophoresis experiment.;Our differential expression analysis consists of data normalization, variance stabilization, analysis of variance(ANOVA), estimation of protein-specific variance, adjustment for multiple comparison and Chi-square tests. A case study of the impact of low dose arsenic and low dose ionizing radiation on human cells is presented. Several statistical methods are compared. It shows that ANOVA ajusted with the empirical Bayes estimate of protein-specific variance combined with pFDR adjustment gives the highest yield of differential protein expression while minimizing the false discovery rate.;The proteomics data generated by 2D gel electrophoresis has proven difficult to analyze due to its large proportion of missing data. The missing values can be either generated by random experimental variations or as a result of being below the limit of detection. It is not always possible to say which one is the cause of missing data. We develop a novel statistical technique that uses the customized Expectation-Maximization (EM) algorithm and the adjusted F-test to identify the differential expression in the presence of missing data. Simulation studies reveal that our EM-based approach has better performance in handling missing data compared with the other two traditional methods, discarding missing data and filling with the detection limit.;Furthermore, the empirical Bayes correction to the variance estimate in the EM algorithm is proposed to lessen the variance heterogeneity problem and increase the power of the F test. Simulation studies show that the empirical Bayes correction improves the performance of our EM-based approach and provides less biased parameter estimates.