Operon information improves gene expression estimation for cDNA microarrays

Abstract

Background: In prokaryotic genomes, genes are organized in operons, and the genes within an operon tend to have similar levels of expression. Because of co-transcription of genes within an operon, borrowing information from other genes within the same operon can improve the estimation of relative transcript levels; the estimation of relative levels of transcript abundances is one of the most challenging tasks in experimental genomics due to the high noise level in microarray data. Therefore, techniques that can improve such estimations, and moreover are based on sound biological premises, are expected to benefit the field of microarray data analysis

Results: In this paper, we propose a hierarchical Bayesian model, which relies on borrowing information from other genes within the same operon, to improve the estimation of gene expression levels and, hence, the detection of differentially expressed genes. The simulation studies and the analysis of experiential data demonstrated that the proposed method outperformed other techniques that are routinely used to estimate transcript levels and detect differentially expressed genes, including the sample mean and SAM t statistics. The improvement became more significant as the noise level in microarray data increases.

Conclusion: By borrowing information about transcriptional activity of genes within classified operons, we improved the estimation of gene expression levels and the detection of differentially expressed genes.

Figure 1

ROC curves of simulation settings. The Figure (A), (B), and (C) are the ROC for simulations 1,2 and 3, respectively. It shows that the sample mean and the SAM t statistic have similar performance in detecting DE genes, and our hierarchical model outperformed both of them.

Figure 2

Histogram of the correlation coefficients. Histogram of pairwise correlation coefficients for (A)the genes within operons and (B) random gene pairs. The correlations are calculated across experimental conditions. The correlation of genes organized in operons is much higher than that of random genes, strongly indicating the co-expression of genes within operons.

Figure 3

Number of false positives vs. Number of total positives. For the E. coli motility data, the genes are ranked by using the proposed method, SAM t statistic and sample mean. The number of false positives is plotted against the number of total positives for each ranking criterion. It shows that ranking genes by proposed method has less false positives than ranking genes by SAM t or sample mean.

Figure 4

FDR estimation. Estimate FDR by using posterior probability and the functional annotation from Macnab et al. The solid line is for the FDR estimate using existing functional annotation while the dashed line for using posterior probability. It indicates that estimate the FDR by using posterior probability yields reasonable result.

Figure 5

Within-operon standard deviation of the estimated gene expressions. Distribution of the standard deviation of the expression estimates of the genes from the same operon. The solid line is for the sample mean method while the dashed line for our proposed Bayesian model. Comparing to the sample mean method, our proposed method yields smaller within-operon standard deviations of the gene expressions.