We first compiled the genes regulated by CNV-miRNAs using the targets from TargetScan5.1 32 , which predicts miRNA targets based on sequence complementarities, sequence context information and binding energy. Because of its high confidence, TargetScan5.1 has been widely used in a variety of “omics” studies (see Methods) 32 33 34 . From among the miRNA-Target associations that were obtained, the representative miRNA for each family with the lowest total context score was presented, but all other miRNAs from the same family were considered to target the same gene at the same target sites 34 . To study the non-redundant miRNA binding sites directly, we replaced the miRNAs by their miRNA-family ID. Finally, 63,428 regulatory relationships were constructed comprising 541 miRNA-families and 9,174 targets (see Additional file 1).

According to the study by Marcinkowska et al. 31 , a total of 209 miRNAs were found to locate in the human CNV-regions. These miRNAs belong to 172 families (see Additional file 2); the remaining 369 miRNA-families had no members in the CNV-regions. In the following analysis, these two types were referred to as CNV-miRNAs and non-CNV-miRNAs, respectively.

We investigated target genes of the non-CNV-miRNAs and CNV-miRNAs and classified them into three groups (see Additional file 3). The first group contains a total of 1,134 target genes that are regulated exclusively by CNV-miRNAs, 823 of the genes are regulated by one CNV-miRNA, 211 by two CNV-miRNAs, 67 by three CNV-miRNAs, 22 by four CNV-miRNAs, and 11 by ≥ 5 CNV-miRNAs. The second group contains a total of 5,710 target genes that are regulated by non-CNV-miRNAs and at least one CNV-miRNA. The third group consists of 2,330 target genes that are regulated exclusively by non-CNV-miRNAs, 1,408 of the genes are regulated by one non-CNV-miRNA, 515 by two non-CNV-miRNAs, 207 by three non-CNV-miRNAs, 95 by four non-CNV-miRNAs and 105 by ≥ 5 non-CNV-miRNAs.

To explore the target-recognition preference of CNV-miRNAs and non-CNV-miRNAs, we devised a sampling method to investigate whether the observed number of target genes for each regulatory type could be expected from random sampling. The simulation analysis involved two steps: (a) 172 miRNAs were selected randomly from the 541 miRNAs, and assumed to be pseudo-CNV-miRNAs; (b) in the miRNA-target regulatory network (see Additional file 1), the edges connecting genes and pseudo-CNV-miRNAs, and the edges connecting genes and pseudo-non-CNV-miRNAs were marked, respectively; the number of target genes (k) was recorded for each type. The steps (a) and (b) were repeated 1,000 times, and resulted in normal distributions of target genes for each type of miRNA regulation. The Z-scores and their transformed p-values (calculated by NORMDIST function in Microsoft Excel) were then used to assess the statistical significance of whether the observed number deviated significantly from random expectation. The simulations provide clues to the regulatory patterns of CNV-miRNAs. As shown in Table 1, the number of genes regulated exclusively by one CNV-miRNA (823 genes were regulated by 137 CNV-miRNAs, approximately 6 target genes per CNV-miRNA) was significantly higher than the number expected from random simulations (p~0.05). In contrast, the number of genes regulated exclusively by one non-CNV-miRNA (1,408 genes were regulated by 280 non-CNV-miRNAs, approximately 5 target genes per non-CNV-miRNA) was significantly lower than the number expected from random simulations (p~0.05). Thus CNV-miRNAs tend to target a higher average number of genes compared with non-CNV-miRNAs. Besides, two and more CNV-miRNAs tend to synergistically regulate the same genes; that is, these genes are preferentially targeted by a combination of CNV-miRNAs in which directional selection may be involved in increasing the frequency of CNV-miRNAs in the human genome 35 36 37 . Obviously, the copy number variation of miRNAs is not independent of copy number variation of the other miRNAs if their binding sites are co-located in the same untranslated regions (UTRs) and regulate the same genes. As shown in Figure 1A for this type of co-regulation, miRNA-α<−>miRNA-β, the copy number alteration of miRNA-α could influence copy number alteration of miRNA-β, or vice versa. Theoretically, it is required that dosage of miRNA-α and miRNA-β should be balanced in synergistically regulating the same genes, which may promote the simultaneous retention of concurrent CNV-miRNAs and finally increase reciprocally the number of genes regulated by CNV-miRNAs. To verify this speculation, we analyzed 211 target genes that were regulated exclusively by two CNV-miRNAs, this dataset contained 422 interactions among 211 genes and 113 CNV-miRNAs (see Additional file 3). If CNV-miRNAs were retained or occurred independently, the number of target genes should follow a normal distribution of N(134,22) (see Table 1 and Figure 1B). Therefore, the number of genes affected by non-independent CNV-miRNAs can be estimated as 211-N(134,22) = N(77,22) (see Figure 1C). To investigate how many of the CNV-miRNAs were caused by the dosage-balance in co-regulation of the same genes, we (a) removed the information of CNV-miRNAs and then drew a number (m) from a normal distribution N(77, 22), (b) randomly assigned m genes to the miRNA-target regulatory network (see Additional file 1), miRNAs which targeted the selected genes were marked, and their number (f) was recorded. The two steps, (a) and (b), were repeated 1,000 times. f followed a normal distribution as N(74, 14) and was then divided by 2 to give N(37, 7). Thus, the miRNA-target recognition retained about 37 CNV- miRNAs with the standard error of 7 (see Figure 1D); at least one-third (calculated by 37/113) of the CNV-miRNAs were attributable to the requirement of dosage-balance for synergistic regulation.

The p-values were obtained from the Z-score, which was calculated as (observed number – mean number of 1000 simulations) / standard deviation of 1000 simulations.

Intuitively, CNVs of miRNA genes can dramatically change their dosage, and this would then affect the expression levels of the target genes in the corresponding individuals 5 15 . Recently, a series of genome-wide gene expression profiles have been measured in four HapMap ethnic populations, CEU (U.S. residents with Northern and Western European ancestry), YRI (Yoruba people of Ibadan, Nigeria), CHB (Chinese Han in Beijing) and JPT (Japanese from Tokyo). We calculated the coefficient of variation (CV) for each protein-coding gene across individuals in the four populations to quantify the within-population expression variability of each of the genes (see Methods). Briefly, the CV is the ratio of the standard deviation of gene’s expression to its mean intensity, which is considered to be an unbiased and comprehensive metric to measure the regulation diversity at the expression level among individuals 38 (see Additional file 4).

As shown in Figure 2A for the YRI population, the mean CV was 0.0251 for target genes regulated exclusively by non-CNV-miRNAs and increased to 0.0258 for target genes regulated by both CNV-miRNAs and non-CNV-miRNAs (p=0.0110, Mann–Whitney U, two-tail test), the mean CV was further increased to 0.0274 for target genes regulated exclusively by CNV-miRNAs (p=0.0072, Mann–Whitney U, two-tail test). Using the CVs calculated in CEU (Figure 2B), CHB (Figure 2C) and JPT (Figure 2D) populations, we obtained similar results.

The associated sequence variants, such as causative bi-allelic SNPs, could also lead to the different expression variability 12 13 14 39 , we explored whether the minor allele frequencies (MAFs) of SNPs in the target genes of the CNV-miRNAs were significantly different from target genes of non-CNV-miRNAs. The 5^′UTR and 3^′UTR sequences of human Ensembl genes were downloaded using BioMart 40 , and then the HapMap Phase III SNPs (retrieved from http://hgdownload.cse.ucsc.edu/goldenPath/hg19/database/) 41 were mapped onto the sequences (see Methods and Additional file 5). As shown in Figure 3, genes regulated exclusively by either non-CNV-miRNAs or CNV-miRNAs have similar proportions of genes that have SNPs in 5^′UTRs and 3^′UTRs; furthermore, the SNPs in the 5^′UTRs and 3^′UTRs have similar MAFs in each of four HapMap populations (p-values range from 0.13 to 0.97, two-tailed t-test). Because genome-wide association and regression analyses have mainly used the MAFs to infer statistical correlations of SNPs with a trait; similar MAFs often indicate that the corresponding SNPs have similar probability to be detected. Therefore, the cis-elements of 5^′UTRs and 3^′UTRs may contain less information than trans-elements in explaining gene expression variations, it is possible that the regulation of some CNV-miRNAs adds a more diversifying control and promotes the differential expression of their target genes among individuals.

A good study has demonstrated that the within-population expression variability of genes can influence the propensity of their differential expression levels between populations 42 . Here, some CNV-miRNAs may live in different populations; thus, the genes targeted by these CNV-miRNAs are likely to be differentially expressed among individuals within a population and also between different populations.

To verify this prediction, we identified the genes that were differentially expressed between any two of the four populations. Taking the CEU and YRI populations as example, we first (a) regress average gene expression intensity, M

yri

Using the method described above, we identified genes that were differentially expressed in at least one of the four ethnic populations (see Additional file 6). As shown in Figure 4B, a similar number of genes were differentially expressed among six population pairs selected from the four ethnic populations. We then investigated whether genes targeted by CNV-miRNAs were over-represented in these differentially expressed genes. As shown in Figure 4C, the proportion of differentially expressed genes was 15.7% for targets regulated exclusively by non-CNV-miRNAs, 17.4% for targets regulated by both CNV-miRNAs and non-CNV-miRNAs (p=0.060, Chi-square, two-tail test), the proportion increased further to 21.7% for targets regulated exclusively by CNV-miRNAs (p=0.001, Chi-square, two-tail test).

For miRNAs that are specifically expressed in a particular tissue or at a particular developmental stage, the copy number duplication or deletion of miRNAs may lead to either weaker or stronger expression of their target genes in the corresponding tissue and developmental stage. For each human gene, we obtained its Differential Expression Ratio (DER) from the FitSNPs 43 . This DER value was a measure of the frequency of differential expression of the gene in multiple microarray studies across thousands of samples (see Methods). Because the DER is derived from all available human microarray datasets deposited in NCBI’s GEO database (http://www.ncbi.nlm.nih.gov/geo/), it provides a comprehensive metric to measure the regulation diversity of genes at the expression level 44 . As shown in Figure 5, the mean DER was 0.506 for 9,784 genes that are not regulated by miRNAs, 0.514 for 2,249 target genes regulated exclusively by non-CNV-miRNAs (p=1.81E-7, Mann–Whitney U, two-tail test), and increased further to 0.535 for 6,730 target genes of CNV-miRNAs (p=2.36E-36, Mann–Whitney U, two-tail test), which include 5,626 targets regulated by non-CNV-miRNAs and CNV-miRNAs, and 1,104 targets regulated exclusively by CNV-miRNAs (see Additional file 7). Therefore, CNV-miRNAs indeed add a more diversifying and complex regulation control to their targets and contribute to an increased likelihood of differential expression among different tissues, cell types, developmental and disease stages.

The Gene Ontology annotation system 45 contained 190,525 associations among 14,117 human genes and 412 GO terms. This data was downloaded and intersected with the 9,174 miRNA target genes that were identified using TargetScan5.1. We obtained GO terms for 6,952 miRNA targets and sought to determine whether the genes that were regulated exclusively by CNV-miRNAs encode proteins that have specific molecular functions or that are involved in particular biological processes (see Methods). As shown in Figure 6B and 6D, targets regulated exclusively by non-CNV-miRNAs were significantly enriched for fundamental biological processes such as maintenance of chromatin, organelle and biogenesis, chromosome segregation, extracellular transport and nucleic acid metabolic process. These processes are known to be essential and dosage-sensitive and their radical fluctuation usually reduces an organism’s fitness. In contrast, targets regulated exclusively by CNV-miRNAs are enriched for processes responsible for stimulus response, immune response, amino acid glycosylation and the MAPKKK cascade (Figure 6A and 6C). These processes were environment-oriented and transduce a large variety of external signals, leading to a wide range of cellular responses such as growth, differentiation, inflammation and apoptosis. The flexible regulation for these processes is required and generally provides positive selectiveness to an organism’s survival.

The miRNAs and their predicted targets were taken from TargetScan (http://www.targetscan.org version 5.1) 32 33 . Targets with a total context score of −0.3 or lower were ignored, where the score quantitatively measure the overall target efficacy 58 . A total of 9,174 targets with at least one conserved 7-mer or 8-mer were selected as reliable miRNA targets 59 (see Additional file 1).

The microarray-based gene expression profiles were derived from lymphoblastic cell lines of 270 HapMap individuals (http://www.sanger.ac.uk/humgen/genevar, GSE6536), including 90 samples of YRI (Yoruba people of Ibadan, Nigeria), 90 samples of CEU (U.S. residents with northern and western European ancestry), 45 samples of CHB (Chinese Han in Beijing) and 45 samples of JPT (Japanese from Tokyo) 60 61 . The annotation table was retrieved from http://www.ncbi.nlm.nih.gov/projects/geo/query/acc.cgi?acc=GPL2507. The RefSeq identifiers were transformed to Ensembl Gene ID through BioMart 40 . Finally, the expression profiles of 16,686 human genes (including 8,636 miRNA targets) across four HapMap populations were complied.

The following formulas were adopted to calculate the coefficient of variation (CV) of gene i in each ethnic population.

The mean intensity M _i calculated by M i = ∑ j = 1 n S ij n

The standard deviation σ _i calculated by σ i = S ij − M i 2 n − 1 ,

The coefficient of variation CV _i calculated by C V i = σ i M i

Where j=1 to n, n represents the number of samples in a population, S _ij represents the expression signal of gene i in sample j. Greater CV implies higher expression variability of a gene across individuals within the corresponding population (see Additional file 4).

Minor allele frequency (MAF) refers to the frequency at which the less common allele occurs in a given population. SNPs with a minor allele frequency of 5% or greater were targeted by the HapMap project and have been widely employed in Genome Wide Association Studies for complex traits (GWAS) 62 63 .

For a SNP A/a, the minor allele frequency was calculated by the following formula

MAF = min 2 N AA + N Aa , 2 N aa + N Aa 2 N AA + 2 N aa + N Aa

Where N _aa represents the count of individuals who are homozygous for allele1, N _Aa represents the count of individuals who are heterozygous, N _aa represents the count of individuals who are homozygous for allele2.

The differential expression ratios (DER) of human genes were obtained from the study by Chen et al. (FitSNPs, http://fitsnps.stanford.edu/download.php) 43 . Briefly, the authors downloaded 476 human GEO datasets from the NCBI Gene Expression Omnibus and categorized each GEO dataset into 24 types of comparisons, such as disease state, cell type, metabolism and so on. A total of 4,877 subset-versus-subset comparisons were performed to identify differentially expressed genes with a cutoff of q value ≤ 0.05 by SAM package 44 . For each human gene, the count of GEO datasets in which it was differentially expressed was divided by the count of its measured GEO.

The gene symbols and EntrezGene IDs were transformed to their Ensembl gene IDs using the BioMart program 40 .The Ensembl genes with available DERs were then intersected with the genes that were used for TargetScan5.1 prediction. Finally, the DER values of 9,784 genes that are not regulated by miRNAs and 8,979 target genes of miRNAs were obtained.

The Gene Ontology (GO) has developed three structured controlled vocabularies to describe gene products in terms of their associated biological processes, cellular components and molecular functions 45 . The human gene association file was downloaded from http://www.geneontology.org/gene-associations/. For each GO term, the proportion of annotated genes was compared between the genes regulated exclusively by CNV-miRNAs and the genes regulated exclusively by non-CNV-miRNAs. The p-value was estimated by Fisher’s exact two-tailed test, and a cutoff of p ≤ 0.05 was used to identify the over-represented or under-represented GO terms among the genes that are regulated exclusively by CNV-miRNAs.

The project was started and completed in Dalian Institute of chemical Physics. Computations were performed on a Linux cluster with 50 nodes (Intel 5130, 2.0 GHz CPU, 4G memory, Laboratory of Molecular Modeling and Design, Dalian Institute of Chemical Physics, Chinese Academy of Sciences). Perl (http://perl.org) and R (http://www.r-project.org/) scripts were used for analysis, and can be obtained on request.