A methodology for analyzing SAGE libraries for cancer profiling

Serial Analysis of Gene Expression (SAGE) has proven to be an important alternative to microarray techniques for global profiling of mRNA populations. We have developed preprocessing methodologies to address problems in analyzing SAGE data due to noise caused by sequencing error, normalization methodologies to account for libraries sampled at different depths, and missing tag imputation methodologies to aid in the analysis of poorly sampled SAGE libraries. We have also used subspace selection using the Wilcoxon rank sum test to exclude tags that have similar expression levels regardless of source. Using these methodologies we have clustered, using the OPTICS algorithm, 88 SAGE libraries derived from cancerous and normal tissues as well as cell line material. Our results produced eight dense clusters representing ovarian cancer cell line, brain cancer cell line, brain cancer bulk tissue, prostate tissue, pancreatic cancer, breast cancer cell line, normal brain, and normal breast bulk tissue. The ovarian cancer and brain cancer cell lines clustered closely together, leading to a further investigation on possible associations between these two cancer types. We also investigated the utility of gene expression data in the classification between normal and cancerous tissues. Our results indicate that brain and breast cancer libraries have strong identities allowing robust discrimination from their normal counterparts. However, the SAGE expression data provide poor predictive accuracy in discriminating between prostate and ovarian cancers and their respective normal tissues.