Novel and simple transformation algorithm for combining microarray data sets
1 Oral Cancer Research Institute, Yonsei University College of Dentistry, Seoul, 120-752, Korea
2 Cancer Metastasis Research Center, Yonsei University College of Medicine, Seoul, 120-752, Korea
3 National Biochip Research Center, Yonsei University College of Medicine, Seoul, 120-752, Korea
4 Brain Korea 21 Project for Medical Science, Yonsei University College of Medicine, Seoul, 120-752, Korea
5 Yonsei Cancer Center, Yonsei University College of Medicine, Seoul, 120-752, Korea
6 Department of Internal Medicine, Yonsei University College of Medicine, Seoul, 120-752, Korea
BMC Bioinformatics 2007, 8:218 doi:10.1186/1471-2105-8-218Published: 25 June 2007
With microarray technology, variability in experimental environments such as RNA sources, microarray production, or the use of different platforms, can cause bias. Such systematic differences present a substantial obstacle to the analysis of microarray data, resulting in inconsistent and unreliable information. Therefore, one of the most pressing challenges in the field of microarray technology is how to integrate results from different microarray experiments or combine data sets prior to the specific analysis.
Two microarray data sets based on a 17k cDNA microarray system were used, consisting of 82 normal colon mucosa and 72 colorectal cancer tissues. Each data set was prepared from either total RNA or amplified mRNA, and the difference of RNA source between these two data sets was detected by ANOVA (Analysis of variance) model. A simple integration method was introduced which was based on the distributions of gene expression ratios among different microarray data sets. The method transformed gene expression ratios into the form of a reference data set on a gene by gene basis. Hierarchical clustering analysis, density and box plots, and mixture scores with correlation coefficients revealed that the two data sets were well intermingled, indicating that the proposed method minimized the experimental bias. In addition, any RNA source effect was not detected by the proposed transformation method. In the mixed data set, two previously identified subgroups of normal and tumor were well separated, and the efficiency of integration was more prominent in tumor groups than normal groups. The transformation method was slightly more effective when a data set with strong homogeneity in the same experimental group was used as a reference data set.
Proposed method is simple but useful to combine several data sets from different experimental conditions. With this method, biologically useful information can be detectable by applying various analytic methods to the combined data set with increased sample size.