Abstract
Pancreatic ductal carcinoma (PDC) remains one of the most intractable human malignancies, mainly because of the lack of sensitive detection methods. Although gene expression profiling by DNA microarray analysis is a promising tool for the development of such detection systems, a simple comparison of pancreatic tissues may yield misleading data that reflect only differences in cellular composition. To directly compare PDC cells with normal pancreatic ductal cells, we purified MUC1-positive epithelial cells from the pancreatic juices of 25 individuals with a normal pancreas and 24 patients with PDC. The gene expression profiles of these 49 specimens were determined with DNA microarrays containing >44 000 probe sets. Application of both Welch's analysis of variance and effect size-based selection to the expression data resulted in the identification of 21 probe sets corresponding to 20 genes whose expression was highly associated with clinical diagnosis. Furthermore, correspondence analysis and 3-D projection with these probe sets resulted in separation of the transcriptomes of pancreatic ductal cells into distinct but overlapping spaces corresponding to the two clinical classes. To establish an accurate transcriptome-based diagnosis system for PDC, we applied supervised class prediction algorithms to our large data set. With the expression profiles of only five predictor genes, the weighted vote method diagnosed the class of samples with an accuracy of 81.6%. Microarray analysis with purified pancreatic ductal cells has thus provided a basis for the development of a sensitive method for the detection of PDC.