21150 Background: The aim of the study was to analyze the gene expression profile of pancreatic cancer by multivariate methods of class prediction. Methods: The snap-frozen or RNA-later preserved samples of 18 pancreatic adenocarcinomas, 9 chronic pancreatitis cases and 6 specimens collected from microscopically unchanged pancreas (N/CP) were analyzed by HG-U133 Plus 2.0 oligonucleotide microarrays (Affymetrix). The obtained dataset was pre-processed using GC-RMA method, gene selection was carried out both by class comparison methods (Welch test with Benjamini-Hochberg correction, False Discovery Rate FDR<5%) and by our own algorithms of class prediction, based on Support Vector Machines technique (Recurrent Feature Replacement and Bootstrap-Based Feature Ranking). Real- time quantitative PCR (Q-PCR) was carried out on Applied Biosystems 7900 HT machine, with Universal Probe Library (Roche) fluorescent probes and normalization by three reference genes index (geNorm, Vandesompele et al.). Results: We compared gene expression profiles between pancreatic cancer samples and N/CP specimens. 23850 probesets significantly differentiated between these three classes (FDR<5%). No ideal discrimination between cancer and N/CP samples was possible by any of single markers. We selected the optimal multi- gene classifier by Support Vector Machines, using Bootstrap-Based Feature Ranking method. The smallest classifier resulting in 100% accuracy consisted of three genes, 45 genes were included in more than half of the diagnostic genesets obtained during bootstrapping process. 14 genes were selected for Q-PCR validation, again none of them ideally discriminated between cancer and normal specimens, with the area under the receiver-operating-characteristic curve ranging from 0.82–0.93. Three-gene combinations allowed for proper classification of all samples. Conclusions: The multi-gene classifier, derived both by microarray technique and Q-PCR analysis, is properly discriminating between pancreatic cancer and chronic pancreatitis/normal pancreas. At least three genes must be included in the classifier to obtain satisfying accuracy. MO and MJ equally contributed to the study. No significant financial relationships to disclose.