Effect of missing values in detecting differentially expressed genes in a cDNA microarray experiment

The aim of this paper is to discuss the effect of missing values in detecting differentially expressed genes in a cDNA microarray experiment in the context of a one sample problem. We conducted a cDNA micro array experiment to detect differentially expressed genes for the metastasis of colorectal cancer based on twenty patients who underwent liver resection due to liver metastasis from colorectal cancer. Total RNAs from metastatic liver tumor and adjacent normal liver tissue from a single patient were labeled with cy5 and cy3, respectively, and competitively hybridized to a cDNA microarray with 7775 human genes. We used $M=log_2(R/G)$ for the signal evaluation, where Rand G denoted the fluorescent intensities of Cy5 and Cy3 dyes, respectively. The statistical problem comprises a one sample test of testing E(M)=0 for each gene and involves multiple tests. The twenty cDNA microarray data would comprise a matrix of dimension 7775 by 20, if there were no missing values. However, missing values occur for various reasons. For each gene, the no missing proportion (NMP) was defined to be the proportion of non-missing values out of twenty. In detecting differentially expressed (DE) genes, we used the genes whose NMP is greater than or equal to 0.4 and then sequentially increased NMP by 0.1 for investigating its effect on the detection of DE genes. For each fixed NMP, we imputed the missing values with K-nearest neighbor method (K=10) and applied the nonparametric t-test of Dudoit et al. (2002), SAM by Tusher et al. (2001) and empirical Bayes procedure by $L\ddot{o}nnstedt$ and Speed (2002) to find out the effect of missing values in the final outcome. These three procedures yielded substantially agreeable result in detecting DE genes. Of these three procedures we used SAM for exploring the acceptable NMP level. The result showed that the optimum no missing proportion (NMP) found in this data set turned out to be 80%. It is more desirable to find the optimum level of NMP for each data set by applying the method described in this note, when the plot of (NMP, Number of overlapping genes) shows a turning point.