To inspect the influence of miRNAs on gene expression fluctuation, we first conducted a comprehensive analysis of microarray data to retrieve the SE genes and FL genes. We collected the expression profiles under various environmental conditions based on the HGU133plus2.0 platform. To minimize variation caused by different experimental platform, we only investigated expression data generated from this platform. For each gene, a fluctuant score (FL score) was calculated by meta-analysis to quantify the expression sensitivity in response to environmental perturbations. The top and bottom 5% of genes in the list were defined as SE genes and FL genes respectively. To evaluate the validity of this categorization, we performed Gene Ontology (GO) enrichment analysis on these genes 13 . From the resultant GO graph, we observed a distinct function distribution for these two groups of genes (see Additional File 1 and Additional File 2). Specifically, for "molecular function", the SE genes were enriched in terms of some basic activities, such as RNA binding, protein binding, NADH dehydrogenase activity, constituent of the ribosome etc, whereas FL genes are involved in environmental factor response, such as receptor binding, cytokine activity, growth factor receptor binding, peptide hormone binding and dopamine binding. For "biological processes", the SE genes were enriched in translation, gene expression, metabolic processes, and biosynthetic processes, whereas FL genes were enriched in signaling pathways, defence response, regulation of immune system process and mediation by a chemical signal etc. Similar results were also obtained when the top and bottom 10% of genes were defined as SE genes and FL genes. This suggests that our classification of SR and FL genes are biologically meaningful and these genes occupy distinct positions in the cell.

We evaluated the propensity of miRNA regulation based on the SE gene and FL gene classification scheme. The predicted targets of human miRNAs were retrieved from TargetScan 6 , PicTar 14 , PITA 15 and miRanda 16 17 , which to our knowledge are regarded in the community as having higher prediction accuracy. A more stringent prediction result derived from intersection of TargetScan and PicTar provided by the miRGen database was also used 18 . In addition, another set of experimentally validated miRNA targets integrated from miRTarBase 19 , miRrecords 20 , miRWalk 21 and miR2Disease 22 was also included in this analysis. Based on these data sets, we observed that miRNA targets were significantly enriched in SE genes. As shown in Figure 1A, miRNA targets comprised 42.5% of SE genes, but only 28% of FL genes as predicted by PicTar (Fisher exact test p-value = 5.8e-07). We observed similar results when using the data sets from other algorithms and experimentally validated miRNA targets (Figure 1B-F). As a control, we randomly selected the same number of genes from the list to analyze this trend, no obvious propensity of miRNA regulation was found in the control data sets (see Additional File 3). The propensity of miRNA regulation was also observed when we selected the top and bottom 10% of the genes as SE genes and FL genes respectively (see Additional File 4). Furthermore, to exclude the interference of datasets from cancer tissue or cell lines, we selected 69 microarray datasets that were derived only from normal tissues to screen the SE genes and FL genes. The propensity analysis of miRNA regulation gave similar results (see Additional File 5).

To avoid potential bias derived from sampling, we next divided the total genes into two groups and calculated the average FL score in each group. The first group contained all the predicted miRNA target genes whereas the second group contained the other genes. We found that the expression fluctuation of miRNA target genes was significantly lower than that of the non-miRNA-target genes for the four data sets (Table 1). For example with PicTar, the average FL score of miRNA target genes was 5154.0, significantly lower than the non-miRNA targets (average FL score = 5717.3, Wilcoxon rank sum test, p-value: 3.53e-58). For a more detailed analysis, we subgrouped the total genes according to their FL scores, and calculated the average FL score and miRNA target proportion in each group. As shown in Figure 2, there was a negative correlation between expression fluctuation and miRNA target proportion, and miRNA target proportion declined dramatically with increasing FL score. Taken together, these results indicated that miRNA target genes are significantly enriched in SE genes, which suggests that miRNAs have a negative effect on whole genome expression fluctuation.

Several studies have demonstrated that a single miRNA can regulate hundreds of mRNAs and that a single mRNA can be regulated by multiple miRNAs. This complex interaction makes the synergistic effect of miRNA regulation in biological networks and pathways possible 23 24 . The synergistic effect of different miRNAs on the expression level of a single gene has been reported 25 . However, whether this effect exists on the genome-wide level is largely unknown. We therefore analyzed the correlation between number of regulatory miRNAs in the 3'-UTR and gene expression fluctuation. In the following analysis, we only use the predicted miRNA targets from PicTar, TargetScan and PITA in that too large or too small data sets may introduce interference. Predicted miRNA target genes were subgrouped according to the number of regulatory miRNAs within 3'-UTR and then the average FL score in each group was calculated. We did not observe any significant correlation between the number of regulatory miRNAs and the expression fluctuation (see Additional File 6). This result is somewhat in disagreement with the previously reported positive correlation between gene expression variability and miRNA seed number 11 . To account for such disagreement, we propose that following explanation. On one hand, a gene that is regulated by multiple microRNAs may be an indication of its functional importance, which requires complex post-transcriptional control by miRNAs. Such functional importance suggests that the expression of such genes are tightly controlled and has less variations. On the other hand, such sophisticated regulation by multiple miRNAs may render it prone to fluctuations and accumulation of noise. We believe that these two factors may be both in play for the majority of the miRNA target genes, and for any given gene it is uncertain which factor is more dominant. As a result, we do not expect any straightforward and overwhelming correlation between the gene expression fluctuation level and the number of miRNA seeds.

Most of the miRNA target sites are located in the 3'-UTR of mRNAs, whereas the lengths of 3'-UTR of protein coding genes vary substantially, and it has been shown that miRNA regulation has an effect on 3'-UTR evolution. It is also known that genes with different 3'-UTR lengths have distinct expression patterns 26 27 . Along this line, we performed a correlation analysis between 3'-UTR length and gene expression fluctuation. Predicted miRNA targets were subgrouped by length in 300 nt intervals and the average FL score within each group was calculated. As shown in Figure 3, a positive correlation between expression fluctuation and 3'-UTR length was observed. The average FL score increased with the 3'-UTR length, for example among the PicTar prediction results (Figure 3A), r = 0.85, p value = 1.69e-05. Similar results were obtained when using miRNA targets predicted by TargetScan (Figure 3B) and PITA (Figure 3C). This result was confirmed when we directly compared the 3'-UTR length of miRNA targets in both SE genes and FL genes. We found that the 3'-UTR length of miRNA targets in SE genes was shorter as compared to that of the FL genes (Figure 4A), which suggested that miRNA targets with longer 3'-UTR length were more likely to have higher expression fluctuation, thus other confounding factors may interfere with the gene expression.

To investigate whether the expression intensity of miRNAs have an effect on target expression fluctuation, we obtained the miRNA expression data from micorRNA.org database 17 and calculated the average expression level in 31 normal human tissues. We compared the average expression intensity of miRNAs that regulate SE genes and FL genes. We did not find any significant difference between these groups (Figure 4B), indicating that miRNA expression level is not a decisive factor for target expression fluctuation.

For identification of the SE genes and FL genes in human genome, firstly we collected gene expression data sets based on the standard and widely-used Affymetrix HGU133plus2.0 platform from the Gene Expression Omnibus database 45 . We collected expression profiles that consist of samples under a variety of environmental factors, including hypoxia, hyperthermia, smoking, alcohol, medicine, strong magnetic field, metal ion, small-sized compounds, chemotherapy, UV, etc. Only data sets with more than six arrays were retained. Finally, a total of 149 data sets were obtained. These data sets were classified as from normal tissues, cancer tissue or cell lines and other disease (see Additional File 8). For each data set, the expression values were logarithmically transformed (base 2) if it was above 0, otherwise turned to 0. Only the maximum expression value was selected if there were multiple probes for a given gene in each sample.

Identification of SE genes and FL genes was performed according to the method previously described by Hao et al. with minor modifications 46 . Briefly, the coefficient of variance (CV = standard deviation/mean value) of the expression for each gene in every data set was calculated. Due to the heterogeneity of the data sets, the CVs of specific genes from different data set could not be compared directly. Thus the CVs in each data set were rank ordered in ascending order, to generate a ranked CV matrix. For each gene, the FL score was defined as the average rank order of the CV in the matrix, and was used as the indication of expression fluctuation. For a specific gene, a relatively high CV was expected if it was more vulnerable to the perturbation of environmental factors. Its confidence was deemed higher if this trend was observed in multiple data sets, thus relative high FL score were expected, and vice versa. Based on this hypothesis, the genes occupying the top or bottom of the genes list were taken as the SE genes and the FL genes respectively (presented as Additional File 9). To validate this classification, Gene Ontology enrichment analysis was used to investigate the functional difference between SE genes and FL genes, performed using the hypergeometric test from web based software GOEAST 47 . In addition, the embeded tool of Multi-GOEAST was used to compare the difference of the GO terms that were enriched in these two sets of genes.

Pre-compiled predicted miRNA targets were retrieved from previously constructed databases including TargetScan (http://www.targetscan.org/, release 5.1: April 2009), PicTar (from UCSC table browser, http://genome.ucsc.edu/) and miRanda (http://www.microrna.org/, August 2010). These algorithms are considered as having high accuracy for miRNA target prediction 48 49 . The intersection dataset generated by both TargetScan and PicTar were retrieved from miRGen database. We also included another dataset generated from PITA software (from the Weizman Institute website, http://genie.weizmann.ac.il/pubs/mir07/mir07_data.html, no flank, TOP catalog), which makes predictions based only on sequence features and target site accessibility. Experimentally validated miRNA targets were integrated from miRTarBase http://mirtarbase.mbc.nctu.edu.tw/, miRrecords http://mirecords.umn.edu/miRecords, miRWalk http://mirwalk.uni-hd.de/ and miR2Disease http://www.miR2Disease.org.

Three different methods were used to analyze the influence of miRNAs on gene expression fluctuation. Firstly, we calculated the proportion of predicted miRNA targets among SE genes and FL genes at different level of significance. As a control, the same numbers of genes were randomly selected from the gene list and the proportion of miRNA targets among these genes was calculated. Secondly, to avoid potential sampling bias, we divided the total genes into two distinct groups. The first group contained the union of the predicted miRNA targets (predicted to be a target by at least one method), whereas the second group contained all of the non-miRNA targets, i.e., the genes that were not predicted to be a target by any of these prediction tools. The average FL score from different groups was calculated to compare the differences. Lastly, we used a sliding window method to calculate the correlation between the average FL score and the proportion of miRNA targets. Specifically, genes were rank ordered according to their FL scores, the average FL score and the miRNA target proportion was calculated for the top 2,000 genes ( = window size) in the first group, then the window was shifted by 50 genes ( = step size) to perform the same calculation on the next group until the end. Pearson's correlation coefficient was calculated between the average FL score and the miRNA target proportion from different groups.