First, we curated from the literature a list of experimentally validated NRSE sites in the human genome 1819, including 38 sites with site lengths of 21 nucleotides (see supplementary table 1 in Additional data file 1). Based on the 38 known sites, we derived a profile (also called a position weight matrix) on the distribution of different nucleotides at each position of NRSE. The profile shows an uneven contribution to the binding of the REST protein from each of the 21 positions (Figure 1a). The positions 2 to 9 and 12 to 17 nucleotides, which will be referred as 'core positions' of NRSE, are much less variable than the remaining positions.

Next we examined the conservation properties of the known NRSE sites. To carry this out, we extracted orthologous regions of these sites in three other fully sequenced mammalian genomes (mouse, rat and dog) 31323334, and generated an alignment for each site in the four species (see supplementary table 1 in Additional data file 1). The alignment data show that the NRSE sites are highly conserved across the mammalian lineages: out of the 38 reference sites, only one cannot be detected in other mammals. We further examined the conservation of NRSE by counting the number of bases mutated in other species from the aligned human site at each of its 21 positions. Similar to the profile, conservation levels at different NRSE positions are highly non-uniform (Figure 1b). However, the conservation levels at different positions are remarkably well correlated with the NRSE profile: highly constrained positions show much stronger conservation in orthologous species than those with higher variability. The core positions are highly constrained and permit few mutations. Among the 37 aligned sites, all core positions contain fewer than two mutations and no insertions or deletions in any of the other species when compared with a human site. By contrast, in a random control, only 0.47 out of the 38 sites are expected to be called conserved with the same criteria. Therefore, the functional NRSE sites demonstrate a 78-fold increase of evolutionary conservation, suggesting the usefulness of evolutionary conservation as an efficient tool for detecting NRSE sites.

We then used the profile to search the entire human genome for sites that are better described by the profile than other background models. For each candidate 21-nucleotide window in the genome, we calculated a log-odds score quantifying how well the site fits to the NRSE profile (see Materials and methods). The overall distribution of the log-odds scores computed over the regulatory regions of all protein-coding genes in humans is shown in Figure 1c, which follows a normal distribution (mean = -37; standard deviation (SD) = 10). We were interested in sites with scores significantly higher than the bulk of the overall distribution: over the entire human genome, we identified 171,152 sites with log-odds scores above 5 (corresponding to 4.2 SDs away from the mean).

The next step was to examine orthologous sequences of these sites in other mammals and filter the list to 1,498 sites based on two criteria: (a) the log-odds scores at the orthologous sites of mouse, rat and dog are also greater than 5, and (b) the number of bases mutated from the corresponding human sequence at the core positions is fewer than two in any of the orthologous sites. The criterion (b) is based on the conservation properties of the known NRSE sites described above.

We then estimated the number of sites that could be discovered purely by chance. For this purpose, we generated a cohort of control profiles with the same base composition and the same information contents as those of the NRSE profile, and searched the instances of the control profiles using the same procedure. Only 328 sites were found for the control profiles, suggesting that approximately 78% of the 1,498 sites are likely to be bona fide NRSE sites. To balance the need for an even smaller rate of false positives, we further identified 895 sites with log-odds scores above 10 in all aligned species. Only 30 sites are expected by chance, suggesting a false positive rate of 3.4%. The distribution on the log-odds scores of these sites falls distinctly to the far right of the bulk of the background distribution (Figure 1c). These sites are distributed across all chromosomes of the human genome and include 37 out of the 38 known NRSE sites that we have curated.

Next we identified the nearest protein-coding genes located around each of the 895 candidate NRSE sites. Over 60% of these genes have NRSE sites within 20 kb of their transcriptional starts (Supplementary figure 1 in Additional data file 1), while a few NRSE sites are located more than 150 kb away from genes, suggesting the possibility of long-range interactions. To study the properties of these genes further, we generated a list of 566 genes that contain at least one NRSE site within 100 kb of their transcriptional start sites (see supplementary website 35). Interestingly, 75 (13.2%) of the genes contain more than one NRSE site in their regulatory regions. For instance, NSF (N-ethylmaleimide-sensitive factor) contains as many as four NRSE sites in its regulatory region in a segment of sequence of less than 100 base pairs; another gene NPAS4 (neuronal PAS domain protein 4) contains three NRSE sites spread over a region of 3 kb.

If the predicted genes are bona fide REST targets, we would expect that the expression of these genes should inversely correlate with the expression of REST. To test this, we examined the expression of these genes and REST across a battery of mouse tissues in a dataset generated previously 36. The tissue gene expression dataset contains 409 of the predicted target genes. It confirms that REST is expressed at low levels in brain-related tissues, and at much higher levels in non-neuronal tissues (Figure 2a). In contrast to the expression profile of REST, most of the predicted REST target genes are specifically expressed in brain-related tissues (Figure 2b). We calculated the correlation coefficient between REST and each of the predicted target genes: the mean correlation coefficient for the genes shown in Figure 2b is -0.21, which is much lower (P value = 2.2e^-16) than what is expected by chance (Figure 2c). Using a stringent threshold (See Materials and methods), we screened out 188 (46% of all 409 genes, 5.4-fold enrichment) genes that demonstrate specific expression in brain-related tissues. A list of these genes and their expression profiles across different tissues is shown in Additional data file 1, supplementary figure 2.

We then examined the functional annotation of all 566 predicted REST target genes. Specifically we were aiming to test if these target genes are enriched in any of the functional categories specified in gene ontology. Based on an annotation provided in 37, we found that the gene set is highly enriched with genes implicated in nervous system development and function (Figure 3). For example, 51 genes (5.2-fold enrichment, P value = 1.3e^-22) encode ion channel activity, and 28 genes (7.3-fold enrichment, P value = 6.6e^-17) are involved in synaptic functions. Interestingly, the list also contains a large number of genes (60, 4.4-fold enrichment and P value = 2.1e^-22) implicated in nervous system development; 15 genes are involved in neuronal differentiation, which include a set of important transcription factors such as NeuroD1, NeuroD2, NeuroD4, LMX1A, SOX2 and DLX6.

However, we also observed some genes that do not seem to encode obvious neural-specific functions. This is consistent with what we observed when examining gene expression patterns for these genes (Figure 2b): a significant portion of them show specific expression in non-neuronal tissues such as brown fat, pancreas, spleen and thyroid (Figure 2b). Interestingly, in most of the tissues the expression of REST is also low (Figure 2a), consistent with the role of REST as a transcriptional repressor. The extent to which REST contributes to the function of other cell types is unclear. A recent study identified REST as a tumor suppressor gene in epithelia cells 38. Together with our findings, this may suggest that REST could potentially regulate a set of genes not necessarily specific to neuronal functions. Alternatively, the observed expression of some REST target genes in non-neuronal tissues might be due to other confounding factors, such as the heterogeneous cell population in these tissues, added levels of regulation caused by transcriptional regulators which themselves are targeted by REST, and the potential regulation by miRNAs, which we will discuss in more detail later.

Thus, using a profile constructed from 38 known NRSE sites and requiring evolutionary conservation in other mammalian species, we have identified 895 sites in the human genome with an estimated false positive rate of 3.4%. We have identified protein-coding genes near these elements, and found that most of these genes are expressed specifically in neuronal tissues.

We noticed that there is a set of miRNAs that are located in close proximity to the predicted 895 NRSE sites in the human genome (Table 1). This includes 10 miRNA genes that are located within 25 kb of at least one NRSE site, where no protein-coding genes can be found nearby. Three of the miRNAs, miR-124a, miR-9 and miR-132, have further experimental support for targeting by REST, as demonstrated in a chromatin immunoprecipitation analysis by Conaco et al. 39. Additionally, we discovered that miR-29a, miR-29b and miR-135b are also located in the vicinity of the NRSE sites. All these 10 miRNA genes are located in intergenic regions, and are transcribed with their own promoters. We also found that there is a set of miRNA genes likely regulated by REST indirectly through the promoters of protein-coding genes that host these miRNAs. These miRNA genes are located in the introns of protein-coding genes, which themselves are predicted REST targets. It is known that miRNAs located inside protein-coding genes are often cotranscribed with the host, and spliced out only after transcription. The set of miRNAs include miR-153 within PTPRN, miR-346 within glutamate receptor GRID1, and miR-218 within SLIT3.

Overall, we identified 16 miRNA genes that are potentially regulated by REST (Table 1) directly or indirectly through their protein-coding hosts. Interestingly, most of these miRNAs are expressed in the brain, and some of them show brain-specific/enriched expression patterns. In a recent survey of several miRNA expression-profiling studies, Cao et al. generated a list of 34 miRNAs that demonstrate brain-specific/enriched expression in at least one study 14. The 16 miRNA genes we identified correspond to 13 unique miRNA mature products. Out of the 13 miRNAs, eight (62%) are contained in the list of 34 brain-specific/enriched miRNAs summarized by Cao et al., which is about sixfold enrichment when compared with what is expected by chance (34 out of 319 all miRNAs, 10.6%). Among the six miRNAs not included in the list of 34 brain-related miRNAs, mir-29 has been demonstrated to show dynamic expression patterns during brain development, and is strongly expressed in glial cells during neural cell specification 1440; mir-346, mir-95 and mir-455 are contained in the introns of (and share the same strand as) their protein-coding hosts, which themselves are specifically expressed in brain-related tissues (supplementary figure 5 in Additional data file 1). It is unclear how these miRNAs and their host genes appear to demonstrate different expression patterns.

In summary, this suggests that similar to neuronal genes, a set of brain-related miRNAs are likely under the control of REST as well. REST might play an important role in repressing the expression of these miRNAs in cells outside the nervous system.

MiRNAs have been suggested to regulate the expression of thousands of genes. Our next step was to seek to identify genes that are targeted by the set of brain-related miRNAs mentioned above. We used an approach similar to previous analyses 2127, and identified candidate targets by searching for conserved matches of the miRNA seeds (2 to 7 nucleotides of the miRNA) in the 3'UTRs of the protein-coding genes. To reduce the rate of false positives, we required the seed to be conserved not only in eutherian mammals as used in the previous analysis, but also in marsupials. For this purpose, we first generated an aligned 3'UTR database in the orthologous regions of the human, mouse, rat, dog and opossum genomes (HMRDO). Then we searched the aligned 3'UTRs for conserved 7-nucleotide sequences that could form a perfect Watson-Crick pairing to each of the miRNA seeds. This effort lead to hundreds of predicted targets for the brain-related miRNAs, including 315 targets for miR-124a, 273 targets for miR-9, and 80 targets for miR-132. The complete list of predicted target genes for each of the brain-related miRNAs can be viewed at the supplementary website 35.

We examined the expression of the predicted target genes in different mouse tissues. The expression profile of the predicted target genes for each of the miRNAs across different tissues is shown in the supplementary website 35. Interestingly, we noticed that the brain-related miRNAs target many genes that are highly transcribed in neural tissues (supplementary figure 3 in Additional data file 1). For instance, among 191 genes targeted by mir-124a that have been profiled across different tissues, 45 (23.6%) are specifically expressed in brain-related tissues, which is 2.8-fold enrichment of that which would be expected by chance (8.54%). The enrichment also holds true for mir-9 in that 25.8% of its target genes show brain-specific expression (threefold enrichment). The coexistence of the predicted target genes and the miRNAs in the same tissues suggests that the brain-related miRNAs are likely involved in extensive regulation of a large number of neuronal genes.

Interestingly, the miRNA target list includes several proteins forming the core REST complex, such as MeCP2 and CoREST. For example, MeCP2 is targeted by numerous brain-specific miRNAs including miR-132, miR-212, miR-9*, miR-218, and miR-124a. Similarly, corepressor CoREST is targeted by miR-124a, miR-218, miR-135b, and miR-153 (Figure 4).

As to the REST itself, our initial analysis did not identify any miRNA that could bind to its 3'UTR. However, a closer examination indicates that gene REST harbors a much longer 3'UTR transcript, not annotated by any gene prediction programs (Additional data file 1, supplementary figure 4). This longer 3'UTR is supported by three pieces of evidence: 1) multiple ESTs detected in this region; 2) high levels of conservation across all mammalian species, and even chicken; and 3) a perfectly conserved poly-adenylation site (AATAAA) in all mammals at the end of the new transcript.

Based on the new 3'UTR transcript, we performed the target prediction again and discovered that REST itself is also targeted by several brain-related miRNAs including miR-9, miR-29a, and miR-153. Together with the discovery of regulation by REST on these miRNAs, this suggests the existence of an extensive double feedback loops between the REST complex and the brain-related miRNAs.

We notice that the 3'UTR of the REST also harbors predicted target sites for several miRNAs that do not seem to have obvious neuronal-specific functions. Out of the seven unique target sites (conserved in HMRDO), three sites are not contained in the list of 34 brain-specific/enriched miRNAs curated by Cao et al. 14, including one site targeted by mir-93 family, one site targeted by mir-25 family, and one site targeted by mir-377. Both mir-93 and mir-25 are enriched in non-neuronal tissues such as spleen and thymus 41. This seems to reinforce the observation of expression patterns for the predicted protein-coding targets of REST, where we also noticed a set of target genes specifically expressed in non-neuronal tissues (Figure 2). We speculate that REST might be involved in the regulation of genes outside the nervous systems.

Next we sought to understand the regulatory machinery controlling the expression of the set of brain-related miRNAs. Besides the negative regulation by REST, we are particularly interested in factors that positively regulate the expression of these miRNAs. Given the scarcity of data on the regulation of miRNA in general, we decided to take an unbiased approach to look for short sequence motifs enriched in the regulatory regions of these miRNAs.

Since few primary transcripts of the miRNA genes are available, we decided to examine a relatively big region (from upstream 10 kb to downstream 5 kb) around each of the miRNAs. On the other hand, however, using big regions significantly increases the difficulty of detecting any enriched motifs. We therefore resorted to comparative sequence analysis again, by searching only for sequence motifs present in aligned regions of the four mammals. For this purpose, we generated a list of all 7-nucleotide motifs, and for each motif we counted the number of conserved and total instances in those regions, and computed a score quantifying the enrichment of the conserved instances (see Materials and methods section. The analysis yielded 35 motifs that are significantly enriched in these regions with a P value less than 10^-6 (Table 2). The top motif is GACGTCA, which is a consensus cAMP response element (CRE) recognized by CREB, a basic leucine zipper transcription factor. We repeated the motif discovery using 6-mer and 8-mer motifs, and consistently identified the CRE element as the most significant motif. For the ten miRNA genes (Table 1) predicted to be directly regulated by REST, we found nine containing a conserved CRE site nearby. This set of miRNAs includes miR-124a, miR-9, miR-29a/29b, and miR-132 (Table 3, Figure 4). Although this association is purely computational, a recent study demonstrated experimentally that one of these miRNAs, miR-132, is regulated by CREB and is involved in regulating neuronal morphogenesis 42.

In addition to CREB, we also identified several other potential regulators such as E47, SMAD3, POU3F2, and MYOD. For instance, besides REST and CREB, miR-9-3 is predicted to be regulated by SMAD3, OCT1, and POU3F2 (Figure 5a), and miR-132 is predicted to be regulated by MYOD and MEF2 (Figure 5b). Interestingly, a recent study shows that MEF2 and MYOD control the expression of another miRNA, miR-1, and play an important role in regulating cardiomyocyte differentiation 11. As well as being expressed in muscle tissues, MEF2 is also highly expressed in brain, where it plays an important role in controlling postsynaptic differentiation and in suppressing excitatory synapse number 43. It would be interesting to examine whether miRNAs are involved in such processes via the regulation by MEF2.

Thus, we have identified several transcription factors that potentially regulate the expression of the brain-related miRNAs with CREB being the top candidate. It is likely that the expression of the brain-related miRNAs is under rigorous control of these regulators during different developmental stages and in different cell types.

We used the whole-genome mammalian alignments generated by the UCSC genome browser 47. From the whole-genome alignment, we then extracted regions of interest. For instance, we generated the aligned NRSE sequences based on genome coordinates of NRSE sites in human. Similarly, we constructed the aligned 3'UTR database using the coordinates of 3'UTRs of all protein-coding genes. For 3'UTRs, we used five-way alignments (human, mouse, rat, dog and opossum). The annotation of genes and their 3'UTRs are from the collection of known genes deposited in the UCSC genome browser.

The NRSE profile was constructed from 38 known NRSE sites each with a site length of 21 nucleotides. We used the 38 sites to compute the frequency of different nucleotides at each position, and generated a position weight matrix representation P of the profile, where p_ij represents the probability of nucleotide j at position i. The information content of a profile is defined as IC_i = 2+Σ_j p_ij*log₂(p_ij) for position i. For any candidate 21-nucleotide sequence, we then calculated a log-odds score to evaluate how well the sequence matched to the NRSE profile. The log-odds score is defined as LO = Σ_i log₂(p_{i, j(i)}/b_j(i)) where j(i) is the nucleotide at position i of the sequence, and b_j represents the probability of observing nucleotide j in a background model. The log-odds score computes the log ratio of two likelihoods, one that the site is generated by the NRSE profile, and the other that the site is generated by a neutral background model. In the neutral background model, we assume each nucleotide is generated independently according to a given nucleotide composition. We estimated the nucleotide composition based on sequences extracted from regulatory regions (5 kb upstream) of all known genes for each of the species separately.

We used the microarray gene expression data published previously by Su et al. 36, which profiled expression patterns of genes across 61 mouse tissues. We postprocessed the dataset and removed any probe with a mean expression level across different tissues of less than 100, and an SD less than 50. For genes containing multiple probes in the array, we used values averaged over different probes to represent the expression level for that gene. In total, 13,743 genes were used for further analysis. For each of the genes, we then normalized their expression values across different tissues such that the mean expression across different tissues was zero and the SD was 1. Based on the normalized values, we then screened out genes with expression values higher than 0.35 in at least one of the brain-related tissues. A total number of 1,174 genes was identified, and we refer to the gene set as the brain-related genes.

First we generated a multiple sequence alignment between human, mouse, rat and dog for the region from 10 kb upstream to 5 kb downstream for each miRNA. We then searched the occurrence of all 7-mers in the aligned regions. For each 7-mer, we counted the number of total instances (N) in human, and the number of instances (K) perfectly conserved in the aligned regions of mouse, rat and dog. We then calculated a Z-score defined as (K-Np₀)/[Np₀(1-p₀)]^1/2, where p₀ is the background conservation rate. The Z-score measures the number of standard deviations on the number of conserved instances away from what is expected by chance by assuming a binomial model on whether a site is conserved. The Z-score quantifies the enrichment of conserved motifs in the aligned regions. To achieve a significant Z-score, a 7-mer must be highly conserved and occur in high frequencies.