We studied the relationship between host and target genes in three different mouse developmental microarray datasets (see methods): embryonic stem cell development (SCD) 27 , somitogenesis (SG) 28 and neurite outgrowth (NO) 29 . We chose developmental datasets since regulatory effects of miRNAs are known to be strongly present in developmental processes 30 . During cell differentiation, groups of genes driving specific developmental processes are often commonly regulated, resulting in similar expression patterns in time course-data. A synergistic relationship between host and the miRNA target genes of differentiating cells is then indicated by positively correlated gene expression patterns. In reverse, antagonistic processes are expected to show anti-correlated or uncorrelated expression patterns between host and related target genes.

Since we argue that correlated expression indicates potentially common host gene functions, we initially tested for correlations between the expression patterns of known host genes. In order to generate statistically robust results (independent of data and prediction errors) we did not analyze single gene expression patterns but argue on groups of correlated genes. Therefore, for each dataset we identified all miRNA host genes and clustered their time-courses according to correlations above 0.8 (see methods).

Within all analyzed cell differentiation datasets, host genes tend to be co-expressed in clusters. As a result of our clustering we obtained seven host gene clusters with more than 5 host genes from all three datasets (see Table 1).

Intriguingly, some host genes appear to be clustered preferentially across the experiments. The genes H19, Igf2, Lpp, Plod3, and Rnf130 were clustered in the two clusters SCD I and NO I, and the genes Chm, Copz1, Dnm1, Nupl1, and Sf3a3 in the clusters SG I and NO II.

For each host gene cluster we identified the intronic miRNAs and all their expressed targets. Most prediction tools for miRNA target site prediction vary qualitatively and quantitatively. In order to get more confident predictions, we used a consensus model (CM) of several miRNA target prediction tools (see methods) 31 . Detailed lists of all analyzed miRNAs/clusters in this work including host genes and loci, are available as additional files 1 and 2.

We performed a hierarchical cluster analysis for the seven clusters based on the expression data of the target genes (see Figure 2). All resulting trees mainly split up in two subclusters: one subcluster of genes with positively correlated expression patterns and one with anti-correlated expressions compared to the host genes. Furthermore, within each dataset, the resulting trees of at least two target gene groups appeared to show completely flipped expression patterns of the main subclusters (SG I vs. SG II; NO I vs. NO II; SCD I vs. SCD III).

These results fit well to the observation that miRNAs dampen the output of preexisting mRNAs or optimize required protein output as it is proposed for metazoans 32 . Additionally, it was shown that genes preferentially expressed at the same time and place as a miRNA tend to avoid sites matching the miRNA 33 . In contrast, co-expression of transcripts with evolutionary conserved miRNA binding sites would then arise from a functional requirement.

The clear discrimination between the two expression patterns suggests a gradual order of differentiating cells, whereas miRNAs function as enhancers of robustness in gene regulation 34 35 . A plausible explanation would be that shortly after initiation of the differentiation process, genes that arrange the differentiating cell towards its new function are up-regulated. In this stage miRNAs are activated to inhibit processes required for self-renewal of stem cells but act perturbing during differentiation. After this reorganization the cell adopts its new functions. In this phase genes are up-regulated which now fulfill the cell's new responsibilities and simultaneously block activity that was only required for differentiation.

In order to confirm the above observed we statistically compared gene expression patterns of host genes with the expression patterns of predicted target genes and sets of randomly sampled genes. For each cluster, we determined the correlations between all hosts and predicted targets and all hosts and 500 sets of randomly sampled genes. Using Wilcoxon's rank sum test we tested the observed correlations to result from distributions with equal median (see methods).

To avoid bias in our CM, we further used three independent miRNA target prediction tools, namely Pictar (PT) 36 , TargetScan (TS) 37 and RNA22 (R) 38 . For each host gene cluster and each single host gene, expression patterns were compared to the expression of predicted targets. Only clusters with predicted and expressed targets in the respective dataset were used in the following analysis.

Concordant for all used methods and all analyzed datasets, we determined that 9 to 44% of the identified host gene clusters were significantly positively correlated or anti-correlated to their target gene expressions (see Figure 3). Thereby, the observed correlations between hosts and targets were independent from the analyzed dataset. The identification of significantly correlated target gene expressions based on the four used target prediction methods appears to be relatively similar between the datasets. The average amount of host gene clusters with significantly correlated target expression for the three datasets varies between 26% and 30%.

Comparing the four tools, PT performs strikingly weaker (15%) than TS, R and the CM with regard to the mean fraction of host gene clusters with significantly correlated predicted target expressions (37%, 25% and 34%). Since the number of targets predicted with PT for each host gene is on average considerably smaller compared to the two other methods, false positive predictions have a larger effect on the determined p-values.

Although, the CM graph is less dense than the other graphs as well as notably smaller than the TS and R graph, it performs best in this analysis with an equal fraction of significantly regulated clusters. However, our results are consistent over all datasets and all different miRNA target site prediction tools.

To further validate our results we used miRNA-mRNA interactions identified by Argonaute HITS-CLIP in the mouse brain 39 . From the 20 miRNAs tested in this study two (miR-153, miR-708) were identified as intronic. After mapping these to our expression data we found the miR-708 host gene Odz4 expressed in the NO and SG datasets. From the 374 targets, 209 targets were expressed in the NO dataset. After clustering these target genes, the resulting tree again splits up into two main subclusters, one with positively correlated and one with anti-correlated gene expression profiles (Additional file 3 - Figure A).

The histogram over the correlation coefficients also confirms an equal distribution of positive and negative correlation coefficients (Additional file 3 - Figure B). The considerable fraction of only weakly correlated target genes may result from (i) false positive experimental miRNA-mRNA associations, (ii) neutral regulatory effects, or (iii) experimental noise.

The results so far indicate a non-directed functional relation between host genes and intronic miRNAs, but do not provide any information on positive or negative correlations. To get robust results, we only used the two tools with the highest number of overall identified host-target gene clusters, which were TS and the CM for the following analysis.

To test whether one or both of the two proposed functional effects - synergistic or antagonistic - may be identified in our data, we calculated the distance between the medians, derived from the correlation distributions between host and predicted target genes and hosts and randomly sampled targets (Figure 4A and methods). The resulting distances Δ _m combined from all three datasets can be seen in Figure 4B and 4C. Both distance distributions show a bimodal distribution with a local minimum at Δ _m = 0, but no significant shift towards a negative or positive correlation. Hence, based on the assumption that highly positive or negative correlation of gene expression patterns indicates similar or opposite functions, we infer that the proposed synergistic and antagonistic effects appear to be equally represented in our data.

Since our investigation is only based on mRNA expression data and further information on protein levels is missing, the real impact on translation stays obscure in this analysis. However, it could be shown that most of miRNA-mRNA interactions function as fine-tuning adjustments to the proteome 40 . Considering the fact that our experimental analysis was based on mRNA expression data, only knock-down effects mediated by target cleavage are directly visible. However, in agreement with previous work 40 41 , the massive appearance of positively correlated miRNA and target expressions strongly indicates tuning effects of varying translational repression. Furthermore, as our findings so far were derived from developmental data, we have to state that the observed effects could be related to these datasets only.

The significantly correlated expression patterns between host genes and miRNA target genes support the notion that intronic miRNA regulation improves host-associated biological functions by either tuning or dampening the expression of target genes. We assume that this relation is also apparent via shared functional annotations. To test this hypothesis, we determined the commonly used functional similarity of gene products based on Gene Ontology (GO) 42 between a single or multiple host genes and their set of target genes. We then calculated the significance of the mean functional similarity by comparing the target set with randomly sampled sets of miRNA target genes (see methods).

We analyzed the previously defined clusters SCD I - NO II and calculated mean functional similarities between the host and all TS predicted target genes. According to the proposed synergistic and antagonistic effects, we divided the target genes upon their expression pattern correlations with the host genes into two groups of positive and negative correlated targets and calculated similarity scores. Results are shown in Table 1. All host gene clusters display a significantly higher functional similarity to their predicted target genes as compared to the null model of randomly chosen target genes (p < 0.01, after Bonferroni-correction). Comparing the similarity scores derived from the positively and negatively correlated target genes showed no significant differences between the two distributions. Independent from the sign of correlation, we determined a significantly higher similarity score than expected for all analyzed clusters.

To check whether a high functional similarity can be found for all host-target relations independent of expression patterns, we additionally calculated the functional similarity score for all single host genes and their predicted target gene sets. We expected the most robust results for the largest network of predicted miRNA target gene associations, since the score is given by the mean of all host gene - target gene pairs. In Figure 5A, we plotted the frequency distribution of similarity scores for TS. We found that the scores are well distributed within the range of 0 and 5. We compared each similarity score with a null model, where the same number of target genes is randomly selected from all miRNA target genes as provided by TS. For the host gene Copz1, for example, we found a significantly larger functional similarity to its targets as compared to 1000 randomly selected sets of miRNA targets (see Figure 5B).

For all annotated host genes with available annotations for the respective targets, we calculated p-values and Z-scores as measures of deviation from the null model. We found that surprisingly many host-target relations deviated from the null model, with high Z-scores as can be seen in Figure 5C. As many as 57 of all 75 host genes annotated in the ontology 'biological process' exhibited a greater similarity to their targets (z > 0) than expected by chance, 30 of them with a p-value < 0.05. For those pairs of host and target genes, a strong correlation in terms of their annotated 'biological process' existed. For the other prediction tools used on in this study, a similar trend to high Z-scores could be observed (Additional file 3 - Figure C). However, these predictions comprise less annotated host genes (48 and 45 for PT and CM, respectively) and also about 10 times less links, rendering significant deviations less possible (see methods for details).

With the use of GO annotations we could show that intronic miRNAs tend to target genes that are functionally more similar to the host genes than randomly chosen genes. The strong bias towards positive Z-scores and the absence of significant dissimilarities between host and target genes, independent from the sign of correlation, agrees with both former proposed regulatory principles (Figure 1A, B). Notably, GO terms are not classified with respect to antagonistic or synergistic effects but on biological relations. For instance, two pathways with conicting regulation on a cellular process like 'cell growth' are both children of the parental term and therefore close within the GO tree. Furthermore, two genes can have opposing regulatory effects on one pathway and would be still grouped together in the same term.

All analyzed datasets were taken from the GEO 43 database: (i) The stem cell development (SCD) dataset consists of three cell lines (R1, J1, V6.5) differentiated into embryoid bodies (EB) at 11 time points from t = 0 h until t = d 14. From each time point and each cell line 3 technical replicates were measured (combination of three cell line differentiations GSE2972, GSE3749, GSE3231). (ii) Within the somitogenesis dataset (SG) gene expression was measured from synchronized C2C12 myoblasts at 13 timepoints from t = 0 h until t = 6 h (GSE7012). (iii) The neurite outgrowth (NO) and regeneration dataset consists of transcriptional activity, measured from dorsal root ganglia during a time course of neurite outgrowth in vitro under two conditions: untreated and under potent inhibitory cue Semaphorin3A. Measurements were taken at 5 time points from t = 2 h until t = 40 h including two technical replicates (GSE9738).

Affymetrix raw data were preprocessed using Bioconductor's R package simpleaffy 44 . Data was normalized and detection calls were determined. Expression values were calculated using the RMA algorithm. Each dataset was filtered independently to remove all probesets with an absent flag in more than two third of all datapoints within the whole experimental setup.

Gene names and gene symbols for each probeset were derived from the Bioconductor Affymetrix Mouse Expression Set 430 annotation data package (moe430a.db). Gene symbols represented by more than one probeset were set to the median expression values.

Host gene clusters were defined upon a correlation-based adjacency matrix. For each microarray dataset we selected all known miRNA host genes and calculated a correlation matrix based on their expression profiles. Each entry representing a correlation coefficient above 0.8, was set to 1, all others to 0. This adjacency matrix now forms a graph of host genes. A host gene cluster was then defined as a maximal connected subgraph of this graph. This equals nearest neighbor method applied to hierarchical clustering algorithm with a defined cutoff of 0.8 of the dendrogramm. For each host gene cluster containing M host genes, the N corresponding target genes were determined upon the three miRNA target prediction tools. We calculated the cluster specific miRNA degree d _i = #T _i /#H _i where #T _i is the number of target genes and #H _i is the number of host genes of cluster i.

Depending on the respective expression profiles, we calculated the M × N cross-correlation coefficients between all hosts and all targets. As a null model we randomly sampled N targets 500 times. For each sample we calculated all M × N correlations. Statistically significant differences between the correlation distributions of our clusters and sampled data were estimated by determining p-values using Wilcoxon's rank sum test.

Distances between the medians of the correlation distributions were calculated as

with C _c being the correlation distribution between the host and the target genes of one cluster and C _s being the correlation distribution between host genes and sampled target genes. Hierarchical cluster analysis was performed using Matlab's Bioinformatics toolbox http://www.mathworks.com using average linkage with Euclidean distance metric.

A list of all murine intronic miRNAs and their host genes was downloaded from the miRBase website http://microrna.sanger.ac.uk. PT predictions were download from the UCSC genome browser http://genome.ucsc.edu. TS conserved miRNA target site predictions were downloaded from the TS website http://www.targetscan.org. In contrast to PT and TS, R is a prediction tool that does not rely on cross-species conservation. The data was downloaded from their website http://cbcsrv.watson.ibm.com/rna22.html. Redundant gene-to-miRNA relationships were removed from all datasets.

The CM prediction graph used in our analysis was built of five different miRNA target site prediction tools. Additionally to PT and TS we used predictions from PITA 45 , Miranda http://www.microrna.org 46 , and TargetSpy http://www.targetspy.org 47 . From all predictions based on RefSeq transcript IDs, we filtered out only miRNA-transcript relations that were predicted by a minimum of four different tools. Transcript mapping to gene symbols was done using a local copy of the RefSeq database (September 2008) 48 .

These genome-wide predictions can be represented by a bipartite graph, where the two different sets of nodes represent the miRNAs and the target genes, respectively, and the predicted interactions are formed by the edges. The three graphs vary primarily in their absolute sizes. PT with 242 miRNAs and 1335 overall predicted targets is very small compared to TS (382 miRNAs, 8879 targets), R (233 miRNAs, 9997 targets) and CM (219 miRNAs, 3249 targets). In Figure 6A relative densities for all graphs and in Figure 6B all degree distributions are shown. For each cluster a mean miRNA target recovery was calculated as the fraction of the number of all predicted and recovered target genes of one cluster to the number of clustered host genes. These distributions again are strikingly similar whereas the mean still varies strongly (Figure 6C, D).

The fraction of the cluster-specific miRNA degree compared to the complete graph miRNA degree of CM is very high (76%) compared to the other methods (TS: 50%, PT: 27%). Since TS predicts the highest number of targets per miRNA, one also expects a relatively large recovery of target genes within the dataset. The PT graph is the densest graph of all but also the smallest one, hence the weak recovery of targets. One reason for the high target recovery of the CM might be that the used prediction tools for the CM score are all trained upon validated data. Therefore, the resulting miRNA-target predictions contain more training data as the PT and TS, which results in the high recovery rate.

We assume that host genes confer regulatory control by translational inhibition of the respective intronic miRNA target genes in possibly related biological processes. To test this hypothesis for all hosts and target genes, we compare the similarity of their respective annotations. Functional gene annotations as provided by the GO 26 classify genes according to their function, associated biological processes or appearance within defined cellular components. They are organized hierarchically, typically in a directed acyclic graph. To each gene more than one classification term can be assigned.

The functional similarity between a host and a target was defined by Resnik's measure as described in 42 and calculated using the ProCope software suite 49 . This method scores relationships between genes by common appearance within one or more terms or, more abstract, by analyzing their distance within the GO graph. For genes with multiple term annotations the maximum scoring GO term pair was used. The functional similarity between a host and a set of targets was determined as the mean of all single host-target scores. For our study, we downloaded the most recent GO files and mouse gene annotation lists from the GO website (January, 2009).

In order to assign statistical significance of the functional host-target similarities in our network, we compared the average similarity of each host to all of its targets against 100.000 randomized networks. To evaluate the host-cluster to target relations, we compared the average host-target similarities in the real network against 100.000 networks with randomized target sets for each host cluster. We calculated a p-value as the relative number of random samples with scores exceeding the score from the data sample. The Z-score was calculated as the deviation of the real score s from the mean m of the sampled distribution, divided by its standard deviation σ, .