MicroRNAs (miRNA) are approximately 22-nucleotide, non-coding RNA sequences important in the control of gene expression. They are involved in a variety of cellular processes, including development, cell differentiation, signaling, and tumorigenesis[1], and are believed to represent 1% of the predicted genes in mammalian and nematode genomes[2, 3]. Mammals in general (and primates in particular) appear to have a large number of miRNAs not found in other animal orders[2], suggesting that many functional miRNAs may have emerged during recent evolutionary periods. According to current functional and predictive models, each miRNA regulates multiple genes during differentiation and/or development at the transcription, translation, and posttranslational levels[1, 4, 5]. However, few of these targets and regulatory pathways have been experimentally validated, and the number of authentic (as opposed to predicted) miRNAs that exist in the mammalian genome as well as the actual number of their targets are not yet known.

Considerable effort has been directed toward understanding which mRNAs within the human genome are subject to regulation by miRNA-mediated repression. The miRGen Targets interface allows users to search either for targets or particular miRNA(s) that influence a particular gene. DIANA-microT[6], MiRanda[7], TargetScanS[5], and PicTar[4] are four genome-wide prediction algorithms whose results are available through miRGEN http://www.diana.pcbi.upenn.edu/miRGen.html[8–11], an integrated database of (i) positional relationships between animal miRNAs and genomic annotation sets, and (ii) animal miRNA targets according to combinations of widely used prediction programs. These algorithms can provide quite a variable picture of miRNA behavior, and it is difficult to assess which in silico predictive method is best for identifying true miRNA targets[12]. It is probable that the use of multiple programs combined with mRNA expression profiling will be necessary to address this question. As a result, we considered four different algorithms (PicTar, TargetScanS, miRanda(microrna.org), and miRanda(miRBase)) in this report when assessing the relationship between mRNA and miRNA expression.

Previous studies have evaluated the influence of a particular miRNA on potential mRNA targets and found a high degree of correlation in specific tissues [13–15]. In general, these reports have examined the regulatory relationships between miRNAs and the genome wide transcriptome, with a focus on pathological conditions (such as cancer) rather than acute perturbations such as hypoxia, although hypoxia in particular has been shown to regulate discrete miRNAs [16–23]. Hypoxia results in a change in expression of a significant portion of the human transcriptome. After oxygen restriction, we observed down-regulation of hundreds of transcripts, including the cystic fibrosis transmembrane conductance regulator (CFTR), in which we could not identify a consensus hypoxia regulatory motif (HRE, hypoxia regulatory element[24]; an indicator of transcriptional regulation by hypoxia inducible factor (HIF)). In the present study, we hypothesized that many of these transcripts may be down-regulated by miRNAs.

Kulshreshtha et al[22] recently demonstrated a functional link between hypoxia and microRNA expression, although the relationship to mRNA expression was not evaluated. We therefore investigated the effects of hypoxia on a model epithelia and found that 3125 unique genes were significantly altered. Of these, approximately 53% were down-regulated, presenting 1649 unique possible targets for miRNA-mediated repression. Expression data from a miRNA Bioarray (see Methods) was compared with the mRNA expression profile, and the strength of correlation against predicted targets with differentially expressed miRNAs was analyzed using computational techniques we developed specifically for this purpose. We found no compelling evidence that miRNA-mediated repression plays a major role in down-regulation of CFTR, and present evidence that the individual miRNA levels do not correlate well with their algorithm-predicted target mRNAs. However, the groups of miRNAs predicted to regulate the same mRNA target were found to be co-regulated, indicating that a level of combinatorial control may exist.

Cellular responses to hypoxia can occur through stabilization of HIF, a well-established transcriptional activator, and result in enhanced expression of a variety of hypoxia-related genes. Much less is known regarding hypoxia-dependent transcriptional repression. Our mRNA array data indicate that a large number of transcripts are robustly downregulated following oxygen restriction in human epithelial cells. One goal of the present study was to investigate the extent to which changes in miRNA could account for these variations in the cellular transcriptome. We hypothesized that miRNAs would play a role suppressing certain genes during hypoxia, and tested this by comparing expression data from miRNA and mRNA expression profiles investigated in parallel.

Recent studies describe hypoxia specific miRNA signatures in a variety of cell types [16–23]. A functional link between hypoxia and miRNA expression has therefore been observed by others, although the relationship between mRNA and miRNA from a genome-wide perspective has not been investigated previously. Kulshreshtha and colleagues emphasized that a spectrum of miRNAs can be induced during hypoxia, and at least some of these occur via a HIF-dependent mechanism. Ten miRNAs reported previously as hypoxia-responsive were also identified in our experiments (e.g. miR-23a, -23b, -27b, -30d, -191, -210, -213, -155, -200a, -181b) using a different method of oxygen deprivation (Tables 4 and 5). Interestingly, three miRNAs (miR-155, -200a, -181b) reported to be upregulated in other cell systems (reviewed in [16–23]) were noted to be repressed in colonic epithelia. These differences most likely relate to the various cell-types, growth conditions, or procedural aspects used in earlier studies.

The epithelial model of hypoxia described here represents a well-defined in vitro system for studying subacute (including transcriptional) effects of oxygen restriction [25–29]. We used the model to evaluate miRNA regulation of gene expression. We found changes in the epithelial transcriptome resulting from low oxygen, as well as further evidence for a potential signature of miRNAs induced by hypoxia [16–23]. However, in contrast to several extant models [13–15], we did not observe a significant correlation between mRNA expression levels and miRNAs on a genome-wide scale. Earlier studies have relied primarily on particular tissue types and developmental stages from a variety of organisms, suggesting results most relevant to embryologic gene regulation. The present investigation of mRNA:miRNA association applied a novel analytical approach to widely available data visualization tools, and monitored miRNA and mRNA expression on a genome-wide basis, including the potential role of environmental stressors (found commonly in pathologic conditions) on miRNA-mediated regulation.

Our analysis incorporated four miRNA target predictions programs (MiRanda, PicTar, miRBase, TargetScan). When miRNA targets were compared to mRNA output, the data sets failed to indicate a significant relationship between expression arrays. In a very recent study, Baek et al. [40] reported that the top third TargetScan predictions (ranked by 'total context score') may correlate best with protein downregulation. In the present experiments, applying this stringent threshold and strict site conservation (after [40]) did not result in a stronger association. This included use of the most recent TargetScan algorithm (release 4.2; http://www.targetscan.org/) and restriction of targets to a context score of 85% or higher. Our results therefore indicate limitations of the currently available target prediction algorithms. While high stringency methods can be valuable for an individual miRNA [40], TargetScan/PicTar modifications do not appear to enhance the available algorithms in a broader, genomic context.

The lack of a significant and robust correspondence between mRNA levels and miRNA expression could represent a lack of specificity and/or accuracy of miRanda or other target prediction algorithms. The observed magnitude of miRNA expression changes (Figures 8b–d) in the present experiments is lower than observed for mRNA (Figure 1). In addition, the relatively small sample sizes used in this study could contribute to a lack of information, making it difficult to test the assumptions underlying the statistical method (such as normality), in a fashion that could impact results. MiRanda typically produces more potential targets than other programs, and a large number of false targets would seriously limit the computational methods described here. We also note that the available programs have only partially overlapping predicted targets for the same miRNA and produce smaller data sets than miRanda. Due to the differences among databases and because there are no clearly superior methods, future studies of mRNA and miRNA regulation should consider analysis of multiple predictive algorithms rather than use of a single data analysis tool.

Although miRNAs can act to promote cleavage and subsequent degradation of their mRNA targets, this may not be the only (or even primary) mechanism of miRNA action in mammalian epithelia. A strong consensus is not yet available regarding the predominant pathway(s) that underly miRNA gene repression [41–44]. One explanation for our findings could relate to translational repression as a major action of miRNA in human cells. It has been shown that certain miRNAs bind their targets and prevent adequate translation. However, mRNA levels are not always affected by this process. A quantitative, proteomic approach to evaluate hypoxic protein expression in epithelia followed by in silico statistical correlation would be necessary to investigate this possibility. On the other hand, miRNA levels are also governed by DNA promoter elements, stability of miRNA, degradative pathways related to differential RNA editing, transport into the cytoplasm, and/or deficient processing by Drosha. Alternative transcript splicing and polyadenylation can eliminate miRNA regulatory sites from a message, and miRNA directed repression can be blocked by certain RNA binding proteins. It seems less likely that common promoter element(s) or a single pathway (by itself) could explain the very large number of up- or down-regulated miRNAs noted as a result of oxygen restriction (Figure 1). Moreover, translation of miRNA targets leads to secondary transcriptional and post-transcriptional regulation that contributes to the observed mRNA profile. The diversity of potential regulatory sequences, difficulty predicting biologic regulation based solely on a consensus miRNA binding site, and the increasingly apparent need for confirmation in living cells indicate that additional, cell-based studies should be used in the future to address transcriptome regulation by miRNA.

miRNA expression arrays represent a relatively new technology, and potential issues exist with regard to data acquisition. The correlation of biological replicates in our studies was >0.99, which indicates the technology is precise, although accuracy is undefined. In addition, we randomized the order of miRNA and mRNA extraction to minimize non-biological, confounding variables. The goal of identifing a method to predict levels of mRNAs based on miRNA profiling, regardless of the underlying regulatory mechanism, was strengthened by correlation against predicted mRNA targets across the entire transcriptome. While previous studies have evaluated effects of a single miRNA after high level recombinant overexpression, the present experiments allowed us to study the dynamics of miRNA and mRNA regulation in parallel with a common physiologic insult (oxygen deprivation). This approach avoided potential variables introduced by overexpression of foreign DNA elements or otherwise manipulating the cellular genome.

The present findings suggest that correlation between miRNAs and their predicted targets based primarily on the number of consensus sites in the 3'UTR may be overly simplistic. Combinatorial analysis reveals much more significant agreement between specific genes and their predicted miRNA regulators as a group; however, this too may reflect a one-dimensional view of miRNA activity. Based on evidence presented here that entire (GO) functional categories of mRNAs are regulated in parallel by hypoxia (Table 2), higher order miRNA groupings may exist along functional or developmental lines that respond as networks. In either case, the present experiments provide a means by which other predicted target lists – either currently available or under development – may be optimized to yield a better correlation between miRNA levels and gene expression.

The observation that a gene-specific group of miRNAs may work in concert to repress CFTR mRNA during hypoxia also points to a novel mechanism of regulation. Previous experiments have failed to establish a direct role for HIF during the pronounced inhibition of CFTR that occurs during oxygen deprivation. Moreover, very few gene products are believed to be down-regulated in a direct fashion by HIF. If a cohort of miRNAs target CFTR and coordinately suppress its message, this could represent an important and novel example of miRNA based repression following an environmental stress. The findings may also help explain in vivo suppression of CFTR mRNA during low oxygen exposure[45, 46], and suggest a role for miRNAs governing levels of hundreds of gene products following hypoxic insult (Figures 2 and 8).

Our results suggest that the expected inverse relationship between miRNA and target mRNA may be a rare event. Several previous studies [13–15] have indicated a clear correlation between a specific miRNA and suppression of a target mRNA. These earlier studies in some cases were based on marked overexpression of a particular miRNA, followed by expression studies of the mRNAs of interest. However, our experiments suggest that under physiological conditions in human epithelium, miRNA acts in a more subtle fashion distinct from that of marked overexpression. In addition, the physiologic impact of miRNAs on cellular transcription appears to result from a multifaceted network of miRNA and mRNA relationships, working together in an interconnected system and in context of hundreds of other RNA species. It may be that target prediction algorithms and expression profiling techniques do not yet adequately represent the complexity of miRNA-mediated gene repression, and new methods may be required to truly understand these systemic aspects.