Abstract
BACKGROUND: Microarray technology shows great potential but previous studies were limited by small number of samples in the colorectal cancer (CRC) research. The aims of this study are to investigate gene expression profile of CRCs by pooling cDNA microarrays using PAM, ANN, and decision trees (CART and C5.0). METHODS: Pooled 16 datasets contained 88 normal mucosal tissues and 1186 CRCs. PAM was performed to identify significant expressed genes in CRCs and models of PAM, ANN, CART, and C5.0 were constructed for screening candidate genes via ranking gene order of significances. RESULTS: The first screening identified 55 genes. The test accuracy of each model was over 0.97 averagely. Less than eight genes achieve excellent classification accuracy. Combining the results of four models, we found the top eight differential genes in CRCs; suppressor genes, CA7, SPIB, GUCA2B, AQP8, IL6R and CWH43; oncogenes, SPP1 and TCN1. Genes of higher significances showed lower variation in rank ordering by different methods. CONCLUSION: We adopted a two-tier genetic screen, which not only reduced the number of candidate genes but also yielded good accuracy (nearly 100%). This method can be applied to future studies. Among the top eight genes, CA7, TCN1, and CWH43 have not been reported to be related to CRC.