CS Bioinformatics Track 10 - 02 July 2010

BACKGROUD: Prostate cancer is the most common form of cancer in men with an incidence about 670,000 new cases annually worldwide (CDC, American and Canadian Cancer Societies; Boyle, 2004, ERSPC). It is the leading non-skin cancer in men above 65, and one out of six men will be affected by prostate cancer during his life. Screen for prostate-specific antigen (PSA) has led to the earlier detection of disease, but increased serum PSA can be present in non-malignant conditions such as benign prostatic hyperplasia as well. So the supplementary biomarkers are strongly needed to improve the diagnosis and prognosis accuracy. There are various biomarker techniques from genomics, proteomics, pharmacogenetics to integrative approaches. Immune response protein microarray as one of the proteomics techniques with the advantage of taking post-translational modification of protein into account is quickly emerging as a follow-up technology. METHOD: This work is based on a data set with 120 blood samples from five groups from control to advanced metastasis. Considering the difference between DNA microarrays and protein microarrays, the first step is to compare different well-established normalization methods such as global normalization, quantile normalization, VSN as well as robust linear model normalization on this protein microarray data. After the proper normalization, one-side Wilcoxon test is performed on biomarker selection. In order to overcome the problem of simultaneously multiple-test as well the sensitivity of the hit gene list highly depending on the training samples, the re-sampling tests instead of just selecting the most significant genes by one test is used. Besides, gene selection strategy in the process of classification such as random forest, shrunken centroid as well as recursive feature elimination are also used to derive different hit gene lists, which can be used to verify the biomarker selected in statistical test. All the hit genes from different methods are further checked by online annotation database. Except from biomarker discovery, the obtained gene lists from different methods are also used to clarify the underlying prostate cancer progression by enrichment pathway analysis, and the classification performance of these gene signatures will be evaluated by principle component analysis (PCA), and leave-one-out cross-validation error rate of different classifiers, including K-nearest neighbour (KNN) as well as support vector machine (SVM). CONCLUSION: Here we show that the quantile followed with robust linear model normalization strategy works better than other counterparts. At the same time, we developed a combinatorial strategy on gene selection, although the variation …