PMCCPMCCPMCC

Search tips
Search criteria 

Advanced

 
Logo of plosonePLoS OneView this ArticleSubmit to PLoSGet E-mail AlertsContact UsPublic Library of Science (PLoS)
 
PLoS One. 2010; 5(8): e12274.
Published online 2010 August 18. doi:  10.1371/journal.pone.0012274
PMCID: PMC2923625

Tissue-Specific and Ubiquitous Expression Patterns from Alternative Promoters of Human Genes

Chad Creighton, Editor

Abstract

Background

Transcriptome diversity provides the key to cellular identity. One important contribution to expression diversity is the use of alternative promoters, which creates mRNA isoforms by expanding the choice of transcription initiation sites of a gene. The proximity of the basal promoter to the transcription initiation site enables prediction of a promoter's location based on the gene annotations. We show that annotation of alternative promoters regulating expression of transcripts with distinct first exons enables a novel methodology to quantify expression levels and tissue specificity of mRNA isoforms.

Principal Findings

The use of distinct alternative first exons in 3,296 genes was examined using exon-microarray data from 11 human tissues. Comparing two transcripts from each gene we found that the activity of alternative promoters (i.e., P1 and P2) was not correlated through tissue specificity or level of expression. Furthermore neither P1 nor P2 conferred any bias for tissue-specific or ubiquitous expression. Genes associated with specific diseases produced transcripts whose limited expression patterns were consistent with the tissue affected in disease. Notably, genes that were historically designated as tissue-specific or housekeeping had alternative isoforms that showed differential expression. Furthermore, only a small number of alternative promoters showed expression exclusive to a single tissue indicating that “tissue preference” provides a better description of promoter activity than tissue specificity. When compared to gene expression data in public databases, as few as 22% of the genes had detailed information for more than one isoform, whereas the remainder collapsed the expression patterns from individual transcripts into one profile.

Conclusions

We describe a computational pipeline that uses microarray data to assess the level of expression and breadth of tissue profiles for transcripts with distinct first exons regulated by alternative promoters. We conclude that alternative promoters provide individualized regulation that is confirmed through expression levels, tissue preference and chromatin modifications. Although the selective use of alternative promoters often goes uncharacterized in gene expression analyses, transcripts produced in this manner make unique contributions to the cell that requires further exploration.

Introduction

The identity of a cell is largely a consequence of the genes transcribed within it. Collectively, these units define the transcriptome, which includes RNA generated from protein-coding and noncoding genes. As a primary determinant of transcriptional activation, promoters of genes position the RNA Pol II molecule at their transcription start sites (TSSs). The position of any TSS is defined in the genome by the position of the first nucleotide in the RNA transcript, also known as the +1 position. This genomic coordinate supplies additional information regarding the position of the core promoter, which lies both adjacent to and overlapping the TSS (i.e., typically from positions −50 to +50) [1]. Promoters confer activation in “cis”, by acting linearly along the genomic sequence, and therefore regulate the transcription of exons that are located in downstream positions. Over half of human genes contain more than one promoter [2][5], which are collectively described as alternative promoters. Each promoter flanks the TSS of a transcript and regulates transcription of exons falling downstream of its genomic position, which may be within an intron of the locus.

Alternative promoters provide transcript diversity and confer dimensional complexity to a cell [6]. For example, individual promoters can be selectively utilized to provide tissue-specific expression, to incorporate distinct 5′ exons into transcripts, or to change the length or content of the open reading frame [7], as illustrated in the downstream transcript of the CD5 gene, which excludes a membrane-binding domain [8]. Notably, a substantial number of alternative promoters have no sequence conservation because they evolved independently among species [9].

Although promoter activity cannot be quantitatively addressed by DNA sequence alone, transcript expression levels serve as a proxy for evaluating promoter function. One technique for measuring transcript expression levels is CAGE (cap-analysis of gene expression), providing information about TSS location and transcript abundance. The CAGE technique has been used to assess a collection of 5′ initiation sites in 17 normal tissues [3]; however, like any resource for TSSs, CAGE data must be combined with annotated gene structures to fully elucidate the relationship between regulatory elements and transcriptional products.

Traditional cDNA or oligonucleotide microarrays represent a second extensively used method to measure transcript expression levels. However, these platforms do not address the alternative promoter activity because many isoforms share the 3′ exons that hybridize to the array probes [10]. Advances in microarray technology have enabled exon-specific microarray platforms, in which all known and predicted exons in the genome have hybridization probes. These platforms are most often used to assess alternative splicing among transcripts [11]. Using the exon microarray platform, we describe a novel approach to study alternative promoter activity using hybridization probes that anneal to first exons of genes, which enables us to evaluate the exclusive activity of the alternative promoters located upstream of these exons. Using a computational assessment of exon-array expression data, we identified and characterized a sizeable set of alternative promoters and discriminated the expression levels and tissue profiles of their transcript isoforms. Although this approach does not identify the precise boundaries of alternative promoters, the measurement of the hybridization signal associated with the first exon enables direct interpretation of the promoter-associated activity in multiple tissues. We quantified and visualized expression data for the subset of human genes containing these distinct alternative first exons (DAlFEs) to elucidate details of the “promoterome” regulating human gene expression and distinguish among alternative promoters whose mutation may contribute to disease phenotypes.

Results

Genome-wide analysis of alternative promoters and alternative first exons

To identify genes with distinct alternative 5′ exons (Figure 1A), we used the UCSC annotations for human genes (NCBI36) [12], in which one or more start sites are annotated for each gene. In the initial set of 26,304 genes, 12,767 genes had more than one mRNA isoform (Figure 1B; see Methods). Alternative promoters existed in 7,708 genes, of which 57% transcribed a first exon that was not exclusive to the isoform and 42% specifically incorporated a DAlFE into the isoform (Figure 1C). We chose two DAlFE isoforms from every gene, whose first exons do not overlap in their genomic coordinates and do not overlap with any exons of the other isoform. Although genes can have more than two alternative promoters of this type, we found that only 19% of the genes showed that characteristic, indicating that genes with three or more distinct, alternative first exons were relatively uncommon in the genome. To test the gene models underlying our predictions we compared the DAlFEs to experimental CAGE evidence in K562 cells and found that 45% of our set corresponded to CAGE data at the same position, in proximity to TSSs of genes (see Methods).

Figure 1
Illustration of alternative promoters and their prevalence in the human genome.

In all cases, the downstream (P2) promoter was fully embedded in an intron of the P1 transcript (Figure 1A). The median distance between the alternative P1 and P2 transcription start sites (TSS) of the isoforms was 4.8 kb; however, inter-TSS distances fell into two main groups, with median distances of 700 or 30,000 intervening bp (Figure S1A). As first exons of transcripts, DAlFEs were predicted to have no splicing to upstream exons and no participation as alternatively spliced exons of larger transcripts. To test this hypothesis, we scored DAlFEs for the presence of 3′ splice junctions (splice acceptor sites) at their leading edge. Maximum entropy (MaxENT) calculations [13] from DAlFEs were compared to a control set of spliced, internal exons from the same transcripts. The maximum scores of 3′ splice-sites in P1 exons (i.e., upstream DAlFEs) were significantly smaller than those of the internal exons. Median scores were 0.5 versus 9.3 for DAlFEs and internal exons, respectively (P-value<2.2×10−16, Wilcoxon test) (Figure S1B). Moreover, the distribution of 3′ splice-site scores for P2 exons (i.e., downstream DAlFEs) was virtually identical to that of upstream DAlFEs (Figure S1B).

We postulated that one role of DAlFEs was to contribute regulatory diversity through the role of noncoding secondary structure selectively affecting the rate of translational initiation [7]. Consistent with this idea, 79.1% of DAlFE exons contained only UTR sequences. In contrast, 79.9% of constitutive first exons (CFEs) from single promoter genes had coding sequence in the first exon (a 3.5-fold higher presence; P-val<10−16, Fisher's exact test). DAlFEs also showed a smaller average size than CFEs by 52 bp (159 vs. 211 bps respectively, P-val = 3.3×10−95, Student's t-test).

A common regulatory component of promoter sequences is the CpG-island. We examined the DAlFE genes and found that 68% lacked a CpG-island promoter for one of the TSSs. In contrast, CFE genes lacked a CpG-island in only 35% of the promoters (consistent with previous publications) [12]. The large percentage of non-CpG-island promoters in DAlFE genes suggested a different mode of evolution from CpG-island promoters, which could resemble a more ancestral form, such as found in fly, yeast and bacterial promoters, or a more recently acquired, lineage-specific form. The possibility of lineage specific acquisition was addressed through the existence of transposable elements or transposon-derived repetitive elements, which are known to introduce sequences that can be exapted for regulatory roles in the genome [9]; [14][16]. To test this hypothesis, we mapped repetitive elements in the first exons and found a 4.3-fold higher occurrence in DAlFEs than CFEs (22.6% and 5.3% respectively; see Methods). To account for the bias of coding sequences against repetitive element insertion, we repeated the analysis using only noncoding CFEs and found that the difference remained significant (2.2-fold higher in DAlFEs, P-val = 8.3×10−22, Fisher's exact test), indicating that repetitive elements were strongly associated with the presence of alternative first exons. Furthermore, we found that the repetitive elements frequently supplied the TSS used for transcription (2.3-fold higher in DAlFEs, P-val = 3.4×10−16) (Figure S1C).

Although the presence of repetitive elements within the first exons was informative about alternative promoter evolution, our subsequent analyses dealt with microarray data generated from DAlFE transcripts. We therefore excluded all gene loci that contained a repetitive element in at least one of the first exons (750 out of 3296, 22.8%) to reduce the possibility of artifacts caused by cross hybridization in the assay.

Global P1 and P2 comparisons

Using the DAlFE dataset, we developed a novel computational method to discriminate the activity of alternative promoters at gene loci by utilizing hybridization probes complementary to the first exons of the transcripts. From the initial set of 3,296 genes whose isoforms used distinct first exons, 1,447 were chosen because they had exon array probes for their first exons that lacked repetitive elements (Figure 1C). We downloaded a public Affymetrix Exon 1.0 array data from 11 human tissues (breast, cerebellum, heart, kidney, liver, muscle, pancreas, prostate, spleen, testes and thyroid), with three biological replicates per tissue, normalized the data using the Affymetrix PLIER software, assessed the median signal of the replicates, and assigned the expression values to the DAlFE probes (see Methods).

Expression analyses

The range of expression data for DAlFE isoforms varied over 3 orders of magnitude from 0 to 4,700 (Figure 2). Comparison of the activity of P1 and P2 promoters across 11 tissues revealed a diversity of patterns, ranging from overlapping expression patterns to mutually exclusive and tissue-specific. Unsupervised hierarchical clustering of the expression intensities over all 11 tissues separated the upstream DAlFEs from their downstream counterparts, indicating that the activities of P1 and P2 promoters were more dissimilar than similar (Figure 2). The median expression intensity for all P2 DAlFEs was slightly lower than the P1 counterparts (22.3 versus 27.1 respectively; see Table 1 for a summary by tissue), but was not significantly different. The distributions of expression intensities conferred by P1 and P2 promoters are similar (Figure S2). All expression values are reported in Table S1.

Figure 2
Expression data illustrating the activity of P1 or P2 promoters.
Table 1
Mean and median expression values for each tissue and the Pearson correlation for DAlFE expression levels.

Since differential expression levels were identified, we assessed expression intensity values in liver by sorting data relative to the P1 promoter, keeping P1 and P2 pairs together (Figure 3A). In contrast to the ascending values of P1 hybridization signals, P2 transcripts generated a random pattern, indicating no global expression correlation between P1 and P2 promoters. This conclusion held true for the other tissues as well. We also found no correlation among the expression levels of P1 and P2 promoters (Pearson correlation; r = 0.00; Figure 3B). These results verified that the vast majority of alternative promoters were differentially activated. Moreover, the data confirmed that the transcription level of any one isoform should not represent all isoforms from the gene locus.

Figure 3
Assessment of coordinated expression patterns among alternative promoters.

Entropy as a quantitative measure for expression uniformity

To objectively quantify tissue-specificity, we used the concept of entropy (HT) to measure the deviation in expression uniformity for a given transcript across multiple tissues [17]. For these purposes, ubiquitous expression was defined as the uniformity of expression across tissues regardless of the magnitude of expression whereas tissue-specificity was interpreted as a significant positive deviation from uniformity in one of multiple tissues. Overall, entropy scores ranged from 0.13 (i.e., transcripts expressed predominantly in one tissue) to larger than 10, with a median HT score of 4.9. To gain an impression of the proportion of the data defined as tissue-specific, heterogeneous, or uniformly expressed, a threshold approach was used to bin the transcripts into each category. With HT levels less than 3, tissue-specific expression patterns were found for 14% of the transcripts (Figure 4). Heterogeneous expression occurred for the majority of alternative promoters (79%) (HT values between 3 and 7). Uniform expression identified 7.0% of the transcripts (HT values greater than or equal to 7). Thus 86% of DAlFE transcripts examined in this analysis were expressed in multiple tissues without tissue-specific patterns. All entropy scores are reported in Table S1.

Figure 4
Entropy scores (HT) for all isoforms.

Entropy scores revealed that the position of a promoter within a gene locus (as P1 or P2) did not favor tissue-specific or ubiquitous expression (Figure 4A). Entropy scores also enabled an analysis of the influence of CpG-islands on expression patterns, whose transcripts were hypothesized to have broad rather than tissue-restricted expression patterns. The median entropy scores for CpG-island promoters were higher than non-CpG island promoters (HT = 5.19 versus 4.59; Student's t-test P-val = 3.1×10−35) (Table S1), indicating that CpG-island promoters had more uniform transcript expression levels. Nevertheless, the CpG-island promoter set was not uniformly active and the non-CpG-island promoter set was not strongly tissue-specific.

The majority of DAlFE transcripts had entropy scores of 3 to 7 (Figure 4A and Table S1). Expression plots for individual genes confirmed the expected patterns by showing increasing uniformity corresponding to increasing HT values (Figure 4B–I). One transcript with a tissue-specific entropy score displayed 100-fold higher expression in cerebellum (i.e., the P2 transcript of CNTN1; HT = 1.7) (Figure 4B). The smallest entropy score was 0.13, for the P1 transcript of PRSS2, a pancreas-specific trypsinogen enzyme whose expression value in this analysis was 2,805 in pancreas and 0.1 to 7.38 in all other tissues. In contrast, examples of transcripts with uniform expression levels had high entropy scores (HT = 7.88 for P2 of RNPS1), which is a part of a surveillance complex against premature stop codons that acts to trigger nonsense-mediated mRNA decay (Figure 4I).

Assessment of genes with ubiquitous function

We performed a case study on a known set of 130 DNA-repair genes [18]. The ubiquitous function of DNA repair genes suggested that single promoters might be sufficient to direct transcription without a need for alternative expression patterns. However, eight genes from the DNA repair set had evidence for a second promoter in our DAlFE annotations. The expression profiles generated by the P1 or P2 promoters of these genes varied across tissues, indicating that DNA repair genes had specialized regulation beyond ubiquitous expression (Figure 5A). For instance, the P1 and P2 isoforms of CHEK1 (checkpoint homolog, encoding a cell-cycle checkpoint gene) had differential regulation, yet similar entropy scores (HT = 4.35 and 4.39, respectively) and showed the highest expression intensity values in testes (71 or 25, respectively). A literature search of published experimental results revealed northern blot data with multiple bands corroborating high expression of CHEK1 in testes [19], without examining the contribution of alternative promoters. In contrast to the relatively low level of expression displayed by isoforms of CHEK1, isoforms of NEIL1 and MPG showed strong expression for one transcript. A 32-fold difference in expression levels occurred between the P2 and P1 transcripts of MPG. FANCD2 showed strong expression signals for both transcripts.

Figure 5
The expression levels and tissue specificities conferred by alternative promoters in a set of DNA-repair genes (A) and a set of housekeeping genes (B) are listed on the x-axes.

We subsequently assessed a set of 532 independently characterized housekeeping genes [20] to address the hypothesis that alternative promoters would not contribute significantly to the expression diversity of housekeeping genes. However, 31 genes refuted that expectation by containing alternative promoters with variable expression levels (Figure 5B). For example, FBXO7 (F-box only protein 7) showed a heterogeneous, moderate-level expression pattern (HT = 3.62) with a seven-fold range from the P1 promoter (47.56 to 371.87) and a high level, ubiquitous expression pattern (HT = 7.0) with 2-fold range from the P2 promoter (612.65 to 1,178.63). Additionally, CNTN1, was expressed in all 11 tissues, but showed entropy scores of 1.18 and 1.75 (P1 and P2, respectively). The highest expression was found in cerebellum for P1 and P2 (858.64 and 106.15, respectively), whereas lowest expression was in heart (P1, 14.22) and pancreas (P2, 1.24). Although widely expressed, the increased cerebellar expression of CNTN1 suggested a more targeted role than a housekeeping function. Consistent with this interpretation, CNTN1 has a published role as a glycosylphosphatidylinositol (GPI)-anchored neuronal membrane protein that functions as a cell adhesion molecule and exhibits high levels of transcription in brain [21].

Alternative promoters and tissue preferences

Similarly to the expression data, we analyzed the correlation between entropy scores for P1 and P2 promoters. The Pearson correlation coefficient was 0.15, indicating no overall correlation for tissue-specific activity, although the weakly positive signal suggested that isoforms from some loci might each have tissue-specific expression. To address non-random co-expression of isoforms, we lowered our requirement from tissue-specificity at both promoters in favor of a measurement of “tissue preference” at both promoters, which identified the most favored tissue from a profile allowing heterogeneous expression. The preferred expression levels were identified by a signal that was two-fold higher than the median value for the probe across all tissues (Table S1 and Table 2). Consistent with previously documented expression patterns reported in cerebellum and testes [22], these two tissues held the highest counts of P1 or P2 transcripts with a tissue preference, including 27% of all transcripts expressed principally in cerebellum and 19% in testes (with some overlap). Spleen and liver showed the next highest occurrences of favored expression.

Table 2
The breakdown of tissue preferences by count of DAlFE promoters.

We examined how often a P1 and P2 promoter pair had the same tissue preference (Table S1). In total, 237 alternative promoter pairs (16%) were predominantly active in the same tissue, almost twice as many as the expected number of 123.8 (see Methods). A chi-square goodness-of-fit test confirmed that this result was highly significant (P-val = 2.1×10−26) compared to the random expectation. The status of most-preferred tissue was shared for P1 and P2 promoters in cerebellum or testes more than any other tissues. This finding was consistent with abundant transcription in these tissues, but novel regarding the coordinated regulation of alternative promoters in these tissues.

Tissue preference distinguishes alternative promoter activity

We found that most often only one promoter of a gene showed a tissue preference. For example, P1 of PEMT (phosphatidylethanolamine N-methyltransferase) had an entropy score of 5.7 versus 0.6 for P2 (liver-specific) (Figure 6). These data were consistent with previously published work that identified P2 of PEMT as liver-specific [23]. Additional isoforms with liver-specific expression were equivalently divided between P1 and P2 activity. One third of those liver specific genes showed co-expression of the P1 and P2 promoters in a liver-specific manner (Figure 6). Only one gene, UGT1A8, had dual promoter expression that was specific to liver and kidney, which contradicts the published literature reporting only extra-hepatic expression, but is consistent with the recent publication of Li et al. [24], which documents expression in the liver.

Figure 6
Liver-specific expression attributed to alternative promoters.

To assess whether the expression patterns had biological relevance, a gene ontology analysis (using GOSTAT) was performed on all genes showing a preference for liver expression from either promoter. The results confirmed that the functional roles of the genes were consistent with liver-related functions, albeit from several types of liver cells. For example, the GO terms for carboxylic, organic acid and lipid metabolic processes were likely contributed by genes that were active in liver mitochondria, which heavily populate hepatic cells (Table 3). In contrast, the term humoral immune response was likely contributed by resident Kupffer cells, which are specialized macrophages located in the liver [25].

Table 3
Best GO terms for molecular function in the liver-preferred dataset.

Additional examples of a specific expression patterns conferred predominantly from one promoter include PDZD4 (P1; cerebellum), C6 (P2; liver), ASB5 (P1; muscle), PPP1R12B (P2; heart), SLC5A12 (P2; kidney), APOM (P2; liver), ADPRHL1 (P1; heart), and TP73L (P1; muscle) (Table S1). Our data matched published literature for the expression patterns attributed to these gene loci [26][33]; however, the contribution of alternative promoters was rarely addressed in those studies. The relevance of alternative promoters in mechanisms of disease is apparent in the tissue-preferences identified for these isoforms. For example, the PDZD4 gene was reported to be upregulated in 13 of 13 synovial sarcomas (SS) [26], whereas siRNA against PDZD4 resulted in suppression of SS tumor cell growth. Alternative splicing of PDZD4 coding exons was reported in Nagayama et al. [26], but the presence of an alternative promoter was not. APOM, known for cholesterol regulation [31], confers risk for the development of type 2 diabetes through SNP T-778C in the proximal promoter region among Han Chinese [34]. This SNP (rs805295) maps to a region of the P2 promoter, consistent with our data showing that P2 has liver-specific expression (HT = 0.8; 626.49 expression value) (Figure 7). TP73L, known for monogenic malformation syndromes such as cleft palate [33] (suggesting a tissue-specific role), has strong sequence similarity to the tumor suppressor TP53. Its isoforms, which are created by alternative promoters and alternative splicing, have divergent abilities to transactivate p53 reporter genes and induce apoptosis [33] (suggesting a more generalized role). The P1 transcript showed strongest expression in muscle (HT = 1.2), whereas P2 showed heterogeneous expression patterns (HT = 5.66). C6, traditionally known as a liver-specific gene [27], showed preferential P2 activity in liver (HT = 0.87; expression value 636.36) as well as in heart (expression value 141.21) (Table S1). These results are consistent with published data stating that the human heart expresses all mRNAs and proteins for the classical complement pathway; furthermore that expression of these genes becomes upregulated in areas of myocardial infarct, causing myocardial damage that may become chronic [35]. The expression level of transcript P1 was 0.54 in heart, strongly implicating the P2 transcript in disease processes of the heart.

Figure 7
P1 and P2 of APOM.

Case Study – HNF4A

Mutations in the HNF4A gene are documented dominantly inherited maturity onset diabetes of the young (MODY) [36]. Two HNF4A promoters have been identified, one with kidney and liver expression (downstream, but called P1 in the literature) and one with pancreatic-specific expression (upstream, but called P2 in the literature) (Figure S3). Thus far, three MODY-associated mutations have been described in the P2 promoter coinciding with transcription factor binding sites for IPF1, HNF1a, or HNF1a/b [36]. Data generated in this analysis confirmed the preferred liver and kidney expression of the downstream promoter (HT = 1.25) (expression levels of 540.15 in liver and 92.45 in kidney). In contrast, the upstream promoter had heterogeneous expression (HT = 5.63) with its highest expression in pancreas (signal intensity 32.15), prostate (signal intensity 39.7) and cerebellum (signal intensity 36.36) (Table S1 and Table S2). Consistent with previous reports, pancreatic expression levels of the upstream transcript were higher than the downstream transcript [37], [38] and liver expression levels were higher for the downstream transcript than the upstream transcript [39]. As with our results, low-level expression has been documented for tissues besides liver and kidney from the downstream promoter [38], [40].

Comparison to public repositories

Comparison of the alternative transcript annotations and expression data to the cDNA microarray data from the Genomics Institute of the Novartis Research Foundation (GNF) [41], showed that the majority of DAlFE isoforms were not assessed in the GNF expression platform. Specifically, GNF expression data quantified expression from both of our isoforms in 313 out of the 1,447 genes examined (21%), whereas the rest had a single expression profile assigned to their locus. Assessment of our collection of transcripts in the UNIPROT database [42] indicated that only 18% of all transcripts (534 of 2,894 isoforms) had a citation for tissue-specific expression, whereas 57% (1,671 of 2,894) showed a preference for a tissue in our analysis, illustrating that distinct expression profiles of isoforms are far more complex than indicated in most repositories.

Epigenetic analyses

The independence of the P1 and P2 promoters suggests that regulatory characteristics should appear independently at each site. We examined epigenetic data associated with the activation of promoters. Specifically, the tri-methylation of the fourth lysine on histone H3 (H3K4me3) is known as a mark of gene activation associated with promoters [43]. Using available chromatin-immunoprecipitation sequencing (ChIP-seq) data (for CD4+ T-cells) [43], we found a significant enrichment of H3K4me3 around the TSSs of P1 and P2 promoters (Figure 8A), with similar peak heights. Similarly [43], the H2A.Z histone variant (Figure 8B), which is associated with active transcription in mammalian sequences [44] had similar peak heights for P1 and P2 promoters, although lower than H3K4me3.

Figure 8
Characterization of the histone modification domains for P1 and P2 alternative promoters (in CD4+ T Cells).

We plotted tag counts of H3K4me3 and H3K9me1 at ABL1 (Abelson murine leukemia viral oncogene) and PEMT as examples. ABL1 was chosen since it was ubiquitously expressed from both promoters at high levels across the 11 tissues, and was likely to show a similar trend in CD4+ T-cells. This hypothesis was confirmed by the histone methylation data from CD4+ cells (Figure 9A). In contrast, our analysis showed that PEMT was only ubiquitously expressed from the P1 promoter, whereas the P2 promoter activity was restricted to the liver. Consistent with this data, the P1 promoter of PEMT showed the activation marks of histone methylation in CD4+ T-cells, whereas the P2 promoter did not (Figure 9B). These results illustrate the presence of distinct epigenetic marks at active and inactive promoters of a gene. Overall, these methylation marks indicate that active, P1 or P2 promoters could be identified by their epigenetic features, regardless of the status of other promoters in the locus. The signals were identical at the P1 and P2 promoters when transcription was active at both sites, whereas the signals were opposite when only one promoter was active. This analysis also helped in confirming that gene annotations correctly identified promoter regions of ABL1 and PEMT by their distinct chromatin features [44].

Figure 9
Histone modifications in alternative promoters.

Extending the analysis of chromatin marks to the full set of DAlFE genes showed that 34.4% have H3K4me3 at one or the other promoter, but not both, as indicated by their overlap with “tag islands” defined for this histone modification (see Methods). The modifications also identified 52% of genes that share signal for H3K4me3 at P1 and P2, and 13% that have no marks at either promoter in CD4+ T-cells. The median distance between alternative promoters was relatively short for genes with simultaneous H3K4me3 modifications (800 bp) (Figure 10), suggesting that promoters with coordinated modifications or those coordinately lacking modifications should share this inter-TSS characteristic. However, we found that promoters coordinately lacking the modifications had the opposite profile, in which the median inter-TSS distance was 14.9 kb. Median distances greater than 12 kb were measured for those sets of promoters with mutually exclusive modifications as well. Combined with other data in this analysis, we concluded that a slight majority of alternative promoters contain epigenetic marks indicative of concurrent activation, likely due to their close spatial proximity. However, the absence of strong correlations for transcriptional output levels or tissue-preferences indicates that alternative promoters confer targeted and specialized functions to their transcripts and do not represent a mechanism for redundant activity.

Figure 10
Frequency of H3K4me3 modification versus inter-TSS distance.

Discussion

Alternative promoters are recognized for creating transcript diversity in a cell, with mostly uncharacterized consequences. Of the different types of alternative promoters, our analysis focused on the subset that utilized distinct first exons. In a novel method to address expression diversity in the cell, we analyzed microarray data to measure transcription levels and tissue preferences. Consistent with the conclusion of Faulkner and Carninci, 2009 [16] we found that repetitive elements are associated with alternative 5′ ends and upstream regions of some transcripts. These data indicate that sequences of repetitive elements contribute to the function of alternative promoters and imply that repetitive elements selectively contribute to the formation of lineage-specific alternative start sites.

In this analysis the expression patterns of distinct first exons served as a proxy for the activity of the promoter and all associated distal regulatory elements. Although the analysis utilized existing annotations of 5′ ends of transcripts (i.e. in UCSC Known Genes), the interpretation of promoter activity was associated directly with the expression data and was not dependent on the precise endpoints of the exon annotations. Therefore, each distinct first exon corresponded to an expression profile that was uniquely its own, revealing the specificity and transcriptional activation of the associated promoter. Individual regulatory signals contributed by distal elements were not evaluated in this analysis. Consistent with a role in transcriptome diversification, the expression levels and tissue preferences of promoters P1 and P2 in each gene were markedly different. The diversity indicated that the conventional descriptions of genes, which routinely include only a single transcript or a mixture of all transcripts, omit important details about the biological functions contributed by individual isoforms. Supporting this conclusion, genes classified as ubiquitous or tissue-specific were found to contain alternative isoforms with different expression profiles. The presence of such examples indicates that genes can undergo sub-functionalization through the appearance of distinct isoforms, comparable to sub-functionalization that occurs through the creation of paralogous genes. Our results also show that alternative promoters maintain their independent expression patterns through the presence of distinct epigenetic marks. These data confirm that the regulation of alternative promoters is a targeted rather than passive feature. Recent methylation analyses [45] corroborate this conclusion by showing that intragenic CpG-islands (putative promoters) are often silenced compared to 5′ CpG-island promoters (in brain), indicating that downstream promoters remain inactive despite the presence of active transcription throughout the locus. Moreover, different cellular environments activate alternative promoters to different extents. For example, we found that cerebellum has more transcripts expressed preferentially and coordinately than other tissues. This result suggests a more permissive environment in cerebellum, however we recognize the fact that cerebellum is a complex tissue that contains a collection of many heterogeneous cell types whose unique transcription patterns could be mixed together.

Altogether this analysis indicates that the further characterization of alternative promoter activity will help elucidate the biological complexities of every cell type. Addtional studies will also extend the functional classification of genomic DNA sequences, which remain largely under-characterized. As sequence polymorphisms are likely to be identified in these functional regions, they represent candidates for altering gene expression and causing complex diseases [46]. Noncoding SNPs affecting gene expression are being identified in genome-wide association testing [47] and experimental validation of the role of these sequences is a necessity for biological interpretation. Further characterization of alternative promoters and their products will extend to discovery of co-regulatory mechanisms of gene expression, since related transcription factor binding sites should control expression in the same tissues. Ultimately, transcript-isoform information and the use of alternative promoters should be considered in all genomic analyses. A small number of databases currently provide information on alternative transcripts, such as FANTOM4 EdgeExpressDB [48]. As independently regulated mRNA isoforms gain more appreciation, the definition of a gene [49] may expand to address different cellular functions for each isoform.

Despite the emphasis on promoters as major regulatory elements of transcription, the act of generating mRNA in a cell is a complex problem. Promoters work collectively with enhancers to confer cell-type appropriate levels of transcription. Moreover, microRNA can selectively down-regulate the accumulation of transcripts in a cell or interfere with the process of translation. Without an integrated picture of the regulatory components of a cell, the basic information necessary for understanding regulatory networks and therapeutic interventions to disease may remain incomplete.

Methods

We used the definition of a gene as a set of transcripts that share sequences [49], which corresponded to the “clusterId” field in the “knownIsoforms” Table of the UCSC Genome Browser (hg18), where each cluster represents a gene (http://genome-test.cse.ucsc.edu/goldenPath/gbdDescriptions.html# KnownIsoforms). This analysis was run on the human genome assembly hg18. The gene clusters represent summary-level annotations integrating information from all transcripts in the same genomic locus. Our pipeline records this information as the total number of clusters. Custom software was developed for these analyses.

Identification of DAlFE Gene Set

We identified all isoform clusters for which there are at least 2 transcripts with a unique first exon. Given this definition, we were able to use the strand and exon position columns from the “KnownGene” Table of the Genome Browser for human genome assembly 18 to identify genes whose 5′ exon of an alternative downstream (or P2) transcript was completely contained within an intron of an upstream (or P1) transcript, as shown in Figure 1. Only the first two promoters encountered along the DNA sequence were considered for any given gene. All genes that contain non-standard exonic features were removed. This includes first exons that had repetitive elements (via REPEATMASKER), or those that were too short for adequate detection in a microarray analysis (60 bp or less). To examine the occurrence of coding sequence in DAlFEs, we constructed a control set of 9,929 CFEs from loci with at least three exons and a first exon that is shared by all isoforms. MaxENT splice-site scores were generated using software from the Burge lab at MIT, http://genes.mit.edu/burgelab/maxent/download/fordownload.tar.gz.

Assignment of probe accession numbers

Once filtered, the DAlFE dataset was compared to the exons represented on the Affymetrix exon array. This step was done through a table join at Galaxy2 database (www.bx.psu.edu). The join function returned probe accession numbers for each exon. These numbers served as the input for the Affymetrix software, ExACT, which provided access to microarray data collected from 3 replicates each of 11 tissues (including breast, cerebellum, heart, kidney, liver, muscle, pancreas, prostate, spleen, testes and thyroid).

Expression Data

The expression data were generated from Human Exon 1.0 ST microarrays by Affymetrix (Santa Clara, CA). The data is publicly available from Affymetrix (http://www.affymetrix.com/support/technical/sample_data/exon_array_data.affx), in GeneChip Operating Software flat file format. The GEO platform platform ID is GPL5175. Each microarray contained 5 million probes, with approximately four probes per exon. Probes were at least 25 bp long with a median length of 123 bp. Approximately 1000 anti-genomic probes were used for background subtraction, categorized by GC count. In the expression data, all exons had three intensity values corresponding to their expression levels in triplicate assays for 11 human tissues: breast, cerebellum, heart, kidney, liver, muscle, pancreas, prostate, spleen, testes and thyroid. Data were analyzed with the Affymetrix Expression Console tool using the Probe Logarithmic Error Intensity Estimate summary method and sketch quantile normalization (http://www.affymetrix.com/analysis/index.affx). The normalized expression values varied from an intensity of near 0 to just above 10,000, with median values of 27.1 and 22.3 from the P1 and P2 promoters, respectively.

Median expression values for the three intensity values were calculated, disregarding values exceeding 3.5-fold difference from the median. Using this criterion 5.6% of the probe values were excluded by being too far above and 2.8% too far below the median. This method created a continuum of probe values, for which only neighboring categories showed overlap (Figure S4). Therefore, when comparing probe values we only interpreted changes >2-fold to be significantly different. Consequently, in almost all comparisons the median probe values differ by orders of magnitude and derive from raw probe values with no overlap.

These data were consistent with results produced by Affymetrix (Figure S5), despite slightly different normalization procedures (within tissues, versus across all tissues). The Affymetrix data had a maximum intensity of just over 10,000 as well, but the minimum value was two, rather than zero. The median values in the Affymetrix data were 29.1 and 23.8 for the P1 and P2 promoters, respectively.

CAGE data

The following UCSC Genome Browser datasets were used for “− strand” transcripts:

http://hgdownload.cse.ucsc.edu/goldenPath/hg18/database/wgEncodeRikenCageMinusClustersK562NucleusLongpolya.txt.gz or for “+ strand” transcripts:

http://hgdownload.cse.ucsc.edu/goldenPath/hg18/database/wgEncodeRikenCagePlusClustersK562NucleusLongpolya.txt.gz. The 3292 tags on the “+ strand” and 3300 on the “− strand” were examined for overlap with the DAlFE isoforms within distances of 100 bps (using a 201-nt interval centered on the TSS) or within 50 bp (using a 101-nt interval centered on the TSS).

Repetitive Elements

Repetitive elements were obtained from the UCSC Genome Browser hg18 RepeatMasker annotation track. We excluded repeats from “Simple_repeat”, “Satellite” and “Low_complexity” categories.

Entropy

As a measure of tissue-specificity, we applied Shannon entropy (H), similar to Schug et al. [17]. We used a transformed version (HT) to normalize the distribution of values. HT is shown in Equation 1, where T is the sum of all expression values, et is the expression in tissue T, and n is the number of tissues (11, for our data set).

equation image

We added a small arbitrary number (5) to each intensity value (et) before calculating entropy in order to avoid misleading readings for low expression values. Otherwise, an intensity value of 5 compared to a value of 0.1 would look like a 50-fold change, even though noise alone could account for the difference.

Shared tissue preferences

To calculate the expected value, if 23.4% of the promoters showed a preference for cerebellum, the expectation is that 5.5% (or 0.2342) of the promoter pairs should have a preference for activity in cerebellum (0.2342×1447 = 79.2). These values are shown in the Expected Pairs column of Table 2. However, we observed 111 promoter pairs with a preference for cerebellum (see the Observed Pairs column). Overall, 123.8 promoter pairs were expected to share a tissue preference, while almost twice as many (237) were observed. A chi-square goodness-of-fit test confirmed that this result was highly significant (p = 2.1×10−26).

Histone Modifications

Tag islands for H4K4me3 histone modifications were computed following the Barski et al. [43] algorithm. Each of the P1 and P2 TSSs for 3,296 genes were overlapped with the tag islands to determine their inclusion status.

Supporting Information

Figure S1

(A) The distribution of the distances between transcription-start sites of pairs of upstream and downstream promoters. The full set of DAlFEs containing 3,296 genes and the restricted set used in the expression analyses (1,447 genes) are plotted with solid and dashed lines, respectively. (B) MaxENT scores of 3′ splice-sites at DAlFEs or second exons in the same transcript. (C) The profile of repetitive element occurrence (%) in a 300-bp window centered on the TSS of DAlFE and CFE transcripts. For comparison we used only non-coding exons including 3,001 DAlEs (blue) and 1,200 CFEs (red).

(0.07 MB PDF)

Figure S2

The log10 expression data for transcripts in all tissues, grouped as P1 or P2 transcript isoforms.

(0.13 MB PDF)

Figure S3

(A) Illustration of the HNF4A locus with alternative promoters P2 and P1. Terminal exons from the P1 promoter are shaded white, whereas the 5-prime P2 exon has a dashed border. Exons shared by both forms are gray with solid borders. (B) In the absence of alternative splicing, the coding sequences generated from the P2 and P1 promoters differ at their amino termini.

(0.95 MB PDF)

Figure S4

Measurement of variation among replicate probe intensities. The variation of raw probe intensities including three replicates for every probe is shown on the y-axis. The range of mean values is shown on the x-axis. Each vertical bar represents probe values whose mean was near the x value; for example, the bar at 10 represents values between 8.91 (100.95) and 11.22 (101.05). Each box represents 50% of the data points and all remaining points are within the dotted lines. The data illustrate that signal intensities varying by orders of magnitude represented significantly different values, i.e. the mean probe intensities at 10 do not overlap those at 100 or 1000.

(0.16 MB PDF)

Figure S5

A scatter plot of the log10 expression values calculated for DAlFE exons compared to the data set analyzed by Affymetrix (available on the UCSC Genome Browser), showing a strong correlation. The Affymetrix set had a minimum value of two, whereas the DAlFE expression values started at zero.

(0.19 MB PDF)

Table S1

All expression values and entropy scores.

(0.42 MB XLSX)

Table S2

HNF4 data.

(0.05 MB DOC)

Acknowledgments

We thank members of the NHGRI faculty for their helpful comments, as well as the NIH Fellows Editorial Board for suggestions in revising this manuscript.

Footnotes

Competing Interests: The authors have declared that no competing interests exist.

Funding: LE is supported by the Intramural Research Program of the National Human Genome Research Institute, National Institutes of Health. IO is supported by the Intramural Research Program of the National Center for Biotechnology Information, National Institutes of Health. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

References

1. Smale ST, Kadonaga JT. The RNA polymerase II core promoter. Annu Rev Biochem. 2003;72:449–479. [PubMed]
2. Baek D, Davis C, Ewing B, Gordon D, Green P. Characterization and predictive discovery of evolutionarily conserved mammalian alternative promoters. Genome Res. 2007;17:145–155. [PubMed]
3. Carninci P, Sandelin A, Lenhard B, Katayama S, Shimokawa K, et al. Genome-wide analysis of mammalian promoter architecture and evolution. Nat Genet. 2006;38:626–635. [PubMed]
4. Kimura K, Wakamatsu A, Suzuki Y, Ota T, Nishikawa T, et al. Diversification of transcriptional modulation: large-scale identification and characterization of putative alternative promoters of human genes. Genome Res. 2006;16:55–65. [PubMed]
5. Tsuritani K, Irie T, Yamashita R, Sakakibara Y, Wakaguri H, et al. Distinct class of putative “non-conserved” promoters in humans: comparative studies of alternative promoters of human and mouse genes. Genome Res. 2007;17:1005–1014. [PubMed]
6. Landry JR, Mager DL, Wilhelm BT. Complex controls: the role of alternative promoters in mammalian genomes. Trends Genet. 2003;19:640–648. [PubMed]
7. Davuluri RV, Suzuki Y, Sugano S, Plass C, Huang TH. The functional consequences of alternative promoter use in mammalian genomes. Trends Genet. 2008;24:167–177. [PubMed]
8. Garaud S, Le Dantec C, Berthou C, Lydyard PM, Youinou P, et al. Selection of the alternative exon 1 from the cd5 gene down-regulates membrane level of the protein in B lymphocytes. J Immunol. 2008;181:2010–2018. [PubMed]
9. Jordan IK, Rogozin IB, Glazko GV, Koonin EV. Origin of a substantial fraction of human regulatory sequences from transposable elements. Trends Genet. 2003;19:68–72. [PubMed]
10. Luzzi V, Mahadevappa M, Raja R, Warrington JA, Watson MA. Accurate and reproducible gene expression profiles from laser capture microdissection, transcript amplification, and high density oligonucleotide microarray analysis. J Mol Diagn. 2003;5:9–14. [PubMed]
11. Blencowe BJ. Alternative splicing: new insights from global analyses. Cell. 2006;126:37–47. [PubMed]
12. Antequera F. Structure, function and evolution of CpG island promoters. Cell Mol Life Sci. 2003;60:1647–1658. [PubMed]
13. Yeo G, Burge CB. Maximum entropy modeling of short sequence motifs with applications to RNA splicing signals. J Comput Biol. 2004;11:377–394. [PubMed]
14. Lowe CB, Bejerano G, Haussler D. Thousands of human mobile element fragments undergo strong purifying selection near developmental genes. Proc Natl Acad Sci U S A. 2007;104:8005–8010. [PubMed]
15. Huh JW, Kim YH, Lee SR, Kim H, Kim DS, et al. Gain of new exons and promoters by lineage-specific transposable elements-integration and conservation event on CHRM3 gene. Mol Cells. 2009;28:111–117. [PubMed]
16. Faulkner GJ, Carninci P. Altruistic functions for selfish DNA. Cell Cycle. 2009;8:2895–2900. [PubMed]
17. Schug J, Schuller WP, Kappen C, Salbaum JM, Bucan M, et al. Promoter features related to tissue specificity as measured by Shannon entropy. Genome Biol. 2005;6:R33. [PMC free article] [PubMed]
18. Wood RD, Mitchell M, Sgouros J, Lindahl T. Human DNA repair genes. Science. 2001;291:1284–1289. [PubMed]
19. Sanchez Y, Wong C, Thoma RS, Richman R, Wu Z, et al. Conservation of the Chk1 checkpoint pathway in mammals: linkage of DNA damage to Cdk regulation through Cdc25. Science. 1997;277:1497–1501. [PubMed]
20. Eisenberg E, Levanon EY. Human housekeeping genes are compact. Trends Genet. 2003;19:362–365. [PubMed]
21. Reid RA, Bronson DD, Young KM, Hemperly JJ. Identification and characterization of the human cell adhesion molecule contactin. Brain Res Mol Brain Res. 1994;21:1–8. [PubMed]
22. Jongeneel CV, Delorenzi M, Iseli C, Zhou D, Haudenschild CD, et al. An atlas of human gene expression from massively parallel signature sequencing (MPSS). Genome Res. 2005;15:1007–1014. [PubMed]
23. Shields DJ, Agellon LB, Vance DE. Structure, expression profile and alternative processing of the human phosphatidylethanolamine N-methyltransferase (PEMT) gene. Biochim Biophys Acta. 2001;1532:105–114. [PubMed]
24. Li X, Bratton S, Radominska-Pandya A. Human UGT1A8 and UGT1A10 mRNA are expressed in primary human hepatocytes. Drug Metab Pharmacokinet. 2007;22:152–161. [PMC free article] [PubMed]
25. Nemeth E, Baird AW, O'Farrelly C. Microanatomy of the liver immune system. Semin Immunopathol. 2009;31:333–343. [PubMed]
26. Nagayama S, Iiizumi M, Katagiri T, Toguchida J, Nakamura Y. Identification of PDZK4, a novel human gene with PDZ domains, that is upregulated in synovial sarcomas. Oncogene. 2004;23:5551–5557. [PubMed]
27. Hobart MJ, Lachmann PJ, Calne RY. C6: synthesis by the liver in vivo. J Exp Med. 1977;146:629–630. [PMC free article] [PubMed]
28. Ota T, Suzuki Y, Nishikawa T, Otsuki T, Sugiyama T, et al. Complete sequencing and characterization of 21,243 full-length human cDNAs. Nat Genet. 2004;36:40–45. [PubMed]
29. Okamoto R, Kato T, Mizoguchi A, Takahashi N, Nakakuki T, et al. Characterization and function of MYPT2, a target subunit of myosin phosphatase in heart. Cell Signal. 2006;18:1408–1416. [PubMed]
30. Gopal E, Umapathy NS, Martin PM, Ananth S, Gnana-Prakasam JP, et al. Cloning and functional characterization of human SMCT2 (SLC5A12) and expression pattern of the transporter in kidney. Biochim Biophys Acta. 2007;1768:2690–2697. [PMC free article] [PubMed]
31. Jiang J, Zhang X, Wu C, Qin X, Luo G, et al. Increased plasma apoM levels in the patients suffered from hepatocellular carcinoma and other chronic liver diseases. Lipids Health Dis. 2008;7:25. [PMC free article] [PubMed]
32. Waye MMY, Cheung HKY, Lam WY, Law PTW, Lo ASY, Lui VWY, Luk SCW, Tsui SKW, Tung CKC, Yam NYH, Liew CC, Lee CY. Gene expression of adult human heart as revealed by random sequencing of cDNA library. Miami Winter BioTechnol Symp Proc. 1995;6
33. Yang A, Kaghad M, Wang Y, Gillett E, Fleming MD, et al. p63, a p53 homolog at 3q27–29, encodes multiple products with transactivating, death-inducing, and dominant-negative activities. Mol Cell. 1998;2:305–316. [PubMed]
34. Niu N, Zhu X, Liu Y, Du T, Wang X, et al. Single nucleotide polymorphisms in the proximal promoter region of apolipoprotein M gene (apoM) confer the susceptibility to development of type 2 diabetes in Han Chinese. Diabetes Metab Res Rev. 2007;23:21–25. [PubMed]
35. Yasojima K, Schwab C, McGeer EG, McGeer PL. Human heart generates complement proteins that are upregulated and activated after myocardial infarction. Circ Res. 1998;83:860–869. [PubMed]
36. Murphy RaC, Tim, Ellard, Sian DWP6-4 Novel hepatocyte nuclear factor 4-alpha (HNF4A) promoter mutation in a New Zealand family. Diabetes Research and Clinical Practice. 2008;79:S30–S30.
37. Eeckhoute J, Moerman E, Bouckenooghe T, Lukoviak B, Pattou F, et al. Hepatocyte nuclear factor 4 alpha isoforms originated from the P1 promoter are expressed in human pancreatic beta-cells and exhibit stronger transcriptional potentials than P2 promoter-driven isoforms. Endocrinology. 2003;144:1686–1694. [PubMed]
38. Briancon N, Weiss MC. In vivo role of the HNF4alpha AF-1 activation domain revealed by exon swapping. Embo J. 2006;25:1253–1262. [PubMed]
39. Briancon N, Bailly A, Clotman F, Jacquemin P, Lemaigre FP, et al. Expression of the alpha7 isoform of hepatocyte nuclear factor (HNF) 4 is activated by HNF6/OC-2 and HNF1 and repressed by HNF4alpha1 in the liver. J Biol Chem. 2004;279:33398–33408. [PubMed]
40. Harries LW, Locke JM, Shields B, Hanley NA, Hanley KP, et al. The diabetic phenotype in HNF4A mutation carriers is moderated by the expression of HNF4A isoforms from the P1 promoter during fetal development. Diabetes. 2008;57:1745–1752. [PubMed]
41. Su AI, Wiltshire T, Batalov S, Lapp H, Ching KA, et al. A gene atlas of the mouse and human protein-encoding transcriptomes. Proc Natl Acad Sci U S A. 2004;101:6062–6067. [PubMed]
42. Riggins GJ, Strausberg RL. Genome and genetic resources from the Cancer Genome Anatomy Project. Hum Mol Genet. 2001;10:663–667. [PubMed]
43. Barski A, Cuddapah S, Cui K, Roh TY, Schones DE, et al. High-resolution profiling of histone methylations in the human genome. Cell. 2007;129:823–837. [PubMed]
44. Li B, Carey M, Workman JL. The role of chromatin during transcription. Cell. 2007;128:707–719. [PubMed]
45. Maunakea AK, Nagarajan RP, Bilenky M, Ballinger TJ, D'Souza C, et al. Conserved role of intragenic DNA methylation in regulating alternative promoters. Nature. 2010;466:253–257. [PubMed]
46. Davuluri RV, Suzuki Y, Sugano S, Plass C, Huang TH. The functional consequences of alternative promoter use in mammalian genomes. Trends Genet. 2008;24:167–177. [PubMed]
47. Stranger BE, Nica AC, Forrest MS, Dimas A, Bird CP, et al. Population genomics of human gene expression. Nat Genet. 2007;39:1217–1224. [PMC free article] [PubMed]
48. Severin J, Waterhouse AM, Kawaji H, Lassmann T, van Nimwegen E, et al. FANTOM4 EdgeExpressDB: an integrated database of promoters, genes, microRNAs, expression dynamics and regulatory interactions. Genome Biol. 2009;10:R39. [PMC free article] [PubMed]
49. Gerstein MB, Bruce C, Rozowsky JS, Zheng D, Du J, et al. What is a gene, post-ENCODE? History and updated definition. Genome Res. 2007;17:669–681. [PubMed]

Articles from PLoS ONE are provided here courtesy of Public Library of Science