It has been reported that clusters may contain unrelated miRNAs 2122. But the frequency of such heterogeneity has not been comprehensively evaluated. Prior research 10 has identified the clustering of 326 registered human miRNA genes, using the strategy proposed by Altuvia et al. 23. According to miRBase 24, the number of registered miRNA genes has doubled. Still more homologs of the known miRNAs are expected in human genome, especially when more and more miRNA genes derived from TEs (transcriptional elements) have recently been identified 252627.

In order to analyze the genomic clustering of miRNAs as systematically and exhaustively as possible, we conducted a cross-species Blast search of known pre-miRNA sequences in the entire human, mouse and rat genomes. In rat genome we found 152 novel miRNA loci that are homologues of previously annotated human or mouse miRNAs, whereas the corresponding numbers of loci in the mouse and human genomes are 33 and 19, respectively (see Additional File 1). We clustered miRNA genes within 50 kb of each other according to Bartel et al. 9. After filtering with RNA polymerase II promoter analysis software and expression profiles (for details, see Material and Methods), a total of 104 miRNA clusters representing 67 unique miRNA combinations have been identified in human genome. Clusters composed of only hsa-mir-566 were not included, since hsa-mir-566 is thought to be derived from repeat sequence Alu, which is pervasive in the human genome 28. Similarly, 91 clusters representing 64 unique miRNA combinations were found in mouse (mmu-mir-680 excluded), and 56 miRNA clusters, all unique, were identified in rat.

In recent releases of miRBase, similar miRNA precursors are assembled into miRNA families (miFams) based on computational analysis and manual inspection. miRNAs in a family have been shown to be phylogenetically related 2930. We further mapped the miRNAs in our data to their families according to miRBase's assignment. As results, 55 distinct miRNA clusters were identified at family level in human genome, along with 51 in mouse, and 42 in rat (see Additional File 2). After mapping to families, 31 clusters have been found to be conserved among the three species, which take up 56% in human, 61% in mouse and 74% in rat (see Additional File 3).

Approximately half of the identified clusters are composed of multiple miRNA families and are denoted as miRNA hetero-cluster throughout this paper. On average, a hetero-cluster contains three miRNA families, and around half of the miRNA genes in a hetero-cluster do not belong to the main family (Table 1). Additionally, even miRNAs in the same family show sequence divergence and possibly target different proteins. This cluster heterogeneity is unlikely to merely derive from random mutation and retention. The fact that these clusters have been fixed in several modern animal genomes implies some evolutionary advantages of such a miRNA gene organization, which may provide an efficient internal mechanism for them to function in coordination 81231.

Coordination among regulators from the transcriptional and miRNA regulatory layers has recently drawn great interests 3233. To reveal why non-homologous miRNAs are found in the same cluster, we analyzed their potential coordination in the context of PPI networks. Generally, the coordination of miRNAs in a cluster should be analyzed directly through justifying the relative proximity between their targets in the network. However, because of the incompleteness of credible PPI networks (e.g. Human Protein Reference Database, HPRD) and the low reliability of other integrated databases, we primarily chose to reverse the procedure to analyze whether proteins in close proximity incline to be regulated by miRNAs from the same cluster (henceforth referred to as "sc-miRNAs").

We first investigated directly interacting protein pairs. It has recently been reported that interacting protein pairs have a certain propensity to be regulated by the same miRNA 34. There may exist a similar tendency that two interacting proteins are also regulated by sc-miRNAs. Sc-miRNAs can fulfil the same function through a complementary approach – by assembling two non-related miRNA loci in the same transcriptional unit. Before analysis, we prepare the data as follows:

We used HPRD as the golden standard positive (GSP, interacting protein pairs) dataset, and the dataset generated by Rhodes et al. 35 according to protein subcellular localization as the golden standard negative (GSN, non-interacting protein pairs) dataset. Comparisons of various miRNA target prediction programs suggest that TargetScan could achieve both high sensitivity and specificity 3637. We therefore employed TargetScan predictions for clustered miRNAs. For targets of a miRNA family, common targets shared by all its miRNA members were used. The number of miRNA sites is 9.7 ± 8.9 per transcript for GSP, and 10.1 ± 9.4 for GSN. To specifically evaluate the coordination among different miRNA families in a cluster, we constructed a family-represented version of hetero-clusters by mapping miRNAs in a cluster to their families, and then having redundancies removed (We will later denote these clusters as family-represented hetero-clusters and clusters before collapsing as original clusters). As a result, the 53 hetero-clusters have been collapsed to 37 family-represented hetero-clusters. To curb possible bias in the target prediction and isolate the effects of clustering, a control was generated by randomly reshuffling miRNAs (or families, when studying family-represented hetero-clusters) among the clusters.

Two measures were employed to assess sc-miRNAs' coordination: one measure is the number of clusters that regulate the two proteins, which reflects the strength of coordination; the other is the percentage of protein pairs that are regulated by sc-miRNAs, which reflects the range of coordination. We calculated the prevalence of miRNAs' coordination on the GSP and GSN datasets for real clusters and 1,000 random cluster sets. Interacting proteins are found to have a significantly higher tendency to be co-regulated by miRNAs from real than from randomly organized clusters (Figure 1). This tendency towards coordinated regulation was not observed for the GSN dataset, which justifies our study of sc-miRNAs' coordination in the context of PPI network. The large coordination measurements detected in random cluster set might be due to: 1) considerable noise exists in current protein-protein interaction data and miRNA target prediction; 2) when reshuffling 95 miRNAs families (contains redundancies, when a family appears in more than one cluster) in 37 hetero-clusters, there is a chance that two families in a real cluster are assigned to a random cluster again; meanwhile, miRNAs from the same family, which usually locate in one or two clusters, were dispersed into more clusters in randomization, and hence substantially increased the number of clusters they coincided.

Since the interacting pairs demonstrate stronger propensity to be regulated by sc-miRNAs than non-interacting pairs, we next asked whether there are some correlation between the propensity and the distance between two proteins in the network. The shortest paths were calculated between all the proteins in GSP dataset. We found a significant negative correlation between the number of clusters regulating the two proteins and the length of the shortest path between them in the PPI network (Spearman's rank correlation, correlation coefficient Rs = -0.24, P < 2.2e-16). This result suggests that distance in the PPI network is an important factor with respect to the coordinated regulation from sc-miRNAs, and lend further support to our previous finding.

We have also noticed the great variance in the number of clusters regulating an interacting protein pair. Still 50% of the directly interacting protein pairs are not co-regulated by miRNAs from same cluster, and the numbers of clusters that do regulate them vary from 1 to 14. Despite the noises and data incompleteness, directly interacting pairs are still likely to enjoy different levels of tendency. Therefore, we studied another factor, connectivity, that might influence this propensity.

Connectivity, the number of neighbours, is one of the most important properties of a protein node in a PPI network. In general, it quantitatively measures functional complexity and importance of a protein node in the PPI network. We counted all proteins' connectivity in HPRD network. The connectivity of proteins in GSP is 15.7 ± 19.3, and in GSN 11.3 ± 17.6. We added up the connectivity of the two proteins in each pair and discovered a positive correlation between connectivity sum and the number of miRNA clusters targeting them in GSP (Spearman's rank correlation, correlation coefficient Rs = 0.11, P = 3.136e-8, as detailed in Table 2). More interestingly, this positive correlation is two times stronger in GSP than GSN (Rs is only 0.06 in GSN-HC, which is close to a random control). Considering the major difference between GSP and GSN is that the shortest path is longer than 3 between GSN proteins in HPRD network, this result indicates that the sc-miRNAs preferentially act on protein pairs within close proximity, rather than those with greater connectivity (importance) in PPI network. It may explain why coordination has mainly been reported in regulating pathways.

To see whether these observations are robust, the same analysis was performed with PicTar prediction set, the overlap of TargetScan and PicTar prediction set as recommended by Sethupathy 36, as well as miRanda prediction set with alignment score greater than 155 (see Additional File 4). They all gave very consistent results.

A module in networks is a local structure characterized by more internal than external links. Since our findings suggest that sc-miRNAs favour coordinated regulation of local regions in PPI network, does sc-miRNAs' target gene set demonstrate some modularity? Xu and Wong 38 have found three miRNA clusters involved in regulating 15 signaling pathways by filtering mouse Biocarta pathway data. Pathway has been recognized as a common form of modularity in network analysis. Here, we explored another form of modularity, network community, in sc-miRNAs' targets with respect to HPRD network. Community structure in networks means the appearance of densely connected groups of vertices, with only sparser connections between groups 39. We have found 32 network communities of size no less than 10 from HPRD network by joining cliques with size 3 to 7 together, using CFinder 2.0 40. We selected those network communities with at least 50% proteins targeted by a miRNA cluster and at most 40% regulated by any member of the cluster to guarantee that they are co-regulated by the cluster rather than any single member. For the 8 community-cluster pairs we obtained, the P-value was calculated based on the overlap of each community with the targets of one thousand simulated miRNA clusters of the same size. Finally, we obtained 5 community-cluster pairs with P < 0.03 (FDR < 0.2, Table 3). Among them, mir-17 cluster's role in transcription regulation through this protein community is well-known 4142. Xie et al. recently reported mir-512 cluster's involvement in histone acetylation in embryonic stem cells 43. Another interesting one is mir-379 cluster. It targets a network community of ten proteins that are involved in circadian rhythm (Figure 2). Although there is no experimental report on it, circadian rhythm is modulated by neural system while mir-379 cluster is brain-specific and is required for dendritogenesis 19.

Although our work mainly focus on protein-protein interaction data, for it is more complete and more suitable for global analysis, we re-ran our analysis with KEGG human pathway data, and presented the results in Additional File 4.

Cluster provides a neat mechanism to transcribe many cooperative miRNAs simultaneously. It could hardly be imaged that this exquisite coordination has been generated in a single event; rather, it should have been consolidated through a long period of evolutionary process. To investigate the relationship between coordination and conservation, we have to calculate the target interactions between miRNAs in a cluster.

A comparison between human, mouse and rat miRNA clusters enables us to classify miRNAs in hetero-clusters into two categories: one is conserved miRNAs, which are observed in human, mouse and rat clusters; the other is non-conserved miRNAs, which exist in human clusters but are unobtainable in mouse or rat clusters. We counted the target interactions irredundantly within and between non-conserved miRNAs and conserved miRNAs. miRNA pairs with targets overlapping more than 100 within HPRD-recorded proteins were excluded because their mature sequences are so similar as to be suspected of having identical function. We calculated the empirical P-value of each miRNA pair's number of target interactions by 10,000 times randomization of miRNA-target protein association. The negative logarithm of P-value (-lgP) was then used to assess the coordination. The greater -lgP is, the stronger the coordination is. As results, 33.3% of conserved miRNAs pairs have been found to cooperate significantly in regulating PPI network (-lgP > 1.5, FDR < 0.1), which comports with our previous finding of intra-cluster coordination. Coordination within non-conserved miRNAs, however, is much weaker than that within conserved ones in a cluster (Figure 3, one-sided two-sample Kolmogorov-Smirnov test: D = 0.34, P = 1.38e-11). In other words, conserved miRNAs are generally more cooperative with each other than non-conserved ones in a cluster. The result is consistent on miRanda target set.

Although results in previous section indicate that miRNAs in a cluster may co-regulate proteins in close proximity of PPI network, they provide no information on each miRNA cluster's coordination. Here, those highly coordinated miRNA pairs (-lgP > 1.5), conserved and non-conserved, come from 17 (30.1%) hetero-clusters. We have found literature support for the function of 15 of them, all listed in Additional File 2. Thus it can be estimated that at least 30% of the miRNA hetero-clusters we can find evidence in current HPRD network for their intra-cluster coordination.