ENHANCED INTER-STUDY PREDICTION AND BIOMARKER DETECTION IN MICROARRAY WITH APPLICATION TO CANCER STUDIES

Although microarray technology has been widely applied to the analysis of many malignancies, integrative analyses across multiple studies are rarely investigated, especially for studies of different platforms or studies of different diseases. Difficulties with the technology include issues such as different experimental designs between studies, gene matching, inter-study normalization and disease heterogeneity. This dissertation is motivated by these issues and investigates two aspects of inter-study analysis.First, we aimed to enhance the inter-study prediction of microarray data from different platforms. Normalization is a critical step for direct inter-study prediction because it applies a prediction model established in one study to data in another study. We found that gene-specific discrepancies in the expression intensity levels across studies often exist even after proper sample-wise normalization, which cause major difficulties in direct inter-study prediction. We proposed a sample-wise normalization followed by a ratio-adjusted gene-wise normalization (SN+rGN) method to solve this issue. Taking into account both binary classification and survival risk predictions, simulation results, as well as applications to three lung cancer data sets and two prostate cancer data sets, showed a significant and robust improvement in our method.Second, we performed an integrative analysis on the expression profiles of four published studies to detect the common biomarkers among them. The identified predictive biomarkers achieved high predictive accuracy similar to using whole genome sequence in the within-cancer-type prediction. They also performed superior to the method using whole genome sequences in inter-cancer-type prediction. The results suggest that the compact lists of predictive biomarkers are important in cancer development and represent common signatures of malignancies of multiple cancer types. Pathway analysis revealed important tumorogenesis functional categories.Our research improved predictions across clinical centers and across diseases and is a necessary step for clinical translation research