Transcriptional and structural impact of TATA-initiation site spacing in mammalian core promoters
© Ponjavic et al.; licensee BioMed Central Ltd. 2006
Received: 3 May 2006
Accepted: 17 August 2006
Published: 17 August 2006
The TATA box, one of the most well studied core promoter elements, is associated with induced, context-specific expression. The lack of precise transcription start site (TSS) locations linked with expression information has impeded genome-wide characterization of the interaction between TATA and the pre-initiation complex.
Using a comprehensive set of 5.66 × 106 sequenced 5' cDNA ends from diverse tissues mapped to the mouse genome, we found that the TATA-TSS distance is correlated with the tissue specificity of the downstream transcript. To achieve tissue-specific regulation, the TATA box position relative to the TSS is constrained to a narrow window (-32 to -29), where positions -31 and -30 are the optimal positions for achieving high tissue specificity. Slightly larger spacings can be accommodated only when there is no optimally spaced initiation signal; in contrast, the TATA box like motifs found downstream of position -28 are generally nonfunctional. The strength of the TATA binding protein-DNA interaction plays a subordinate role to spacing in terms of tissue specificity. Furthermore, promoters with different TATA-TSS spacings have distinct features in terms of consensus sequence around the initiation site and distribution of alternative TSSs. Unexpectedly, promoters that have two dominant, consecutive TSSs are TATA depleted and have a novel GGG initiation site consensus.
In this report we present the most comprehensive characterization of TATA-TSS spacing and functionality to date. The coupling of spacing to tissue specificity at the transcriptome level provides important clues as to the function of core promoters and the choice of TSS by the pre-initiation complex.
Elucidation of the mechanisms that govern the regulation of genes at the transcriptional level remains one of the most important challenges in biology. Transcriptional regulation is achieved by a combination of cellular events, including binding of cis-regulatory elements to transcription factor binding sites (TFBSs), chromatin structure modification, and the assembly of the pre-initiation complex (PIC) at transcription start sites (TSSs) .
Presently, we have a reasonable understanding of components used in the transcription initiation process but only limited insight into the mechanisms of the cognate elements [2–6]. The generally accepted model for transcriptional initiation by core promoter elements is centered on the complexes formed by TATA box binding protein (TBP) with RNA polymerases and associated factors . It is a common textbook-inflicted misconception that 'typical' RNA polymerase II eukaryotic core promoters have a TATA box guiding the PIC. Recent evidence [7, 8] provided genome-wide confirmation of the existence of at least two distinct modes of transcription initiation: CpG-island based, TATA independent initiation with multiple TSSs; and TATA dependent initiation, in which TSSs are concentrated on one or few consecutive genome positions (called single peak [SP] promoters). SP promoters and, by association, TATA-driven promoters are strongly associated with genes with tissue-specific and/or context-specific expression . This is in agreement with recent large-scale statistical studies that confirmed the previously anecdotal correlation between CpG island promoters and housekeeping genes on one hand, and TATA box promoter and tissue-specific genes on the other . The fact that TATA box promoters evolve more slowly than other types of promoters  implies that changes in such promoters are less tolerated and that this type of mechanism is more ancient than the more plastic promoters with many TSSs , in which evolutionary events can include evolutionary turnover .
In TATA driven promoters, the primary role of the TATA box is to anchor the PIC. In higher eukaryotes, this process sterically constrains the selection of transcription initiation sites, but TATA-TSS distance can vary slightly. The exact mechanism of start site selection, and therefore the TATA-TSS distance, remains unknown [3, 12].
Because TATA boxes are highly overrepresented in promoters where the TSSs are concentrated in one or few consecutive genome positions, the TATA box location relative to the TSS is likely to have an impact on the efficiency of inducible expression. The unavailability of precise TSS locations has limited the study of the TATA-TSS spacing to a handful of promoters [13–18]. These studies indicated that the TATA box is functionally linked to the determination of the initiation site, and that TATA-TSS spacing affects the efficiency of transcriptional initiation.
It is evident that inducible expression is not solely orchestrated by events at the core promoter, but is also subject to long-range cis-regulatory element interactions  as well as cellular events on a larger scale, including epigenetic control of chromatin superstructure . Nonetheless, core promoter elements have been confirmed as important determinants for transcriptional specificity , and our goal in this work is to determine the constraints imposed on such determinants.
We recently showed that the FANTOM cap analysis of gene expression (CAGE) data allow us, for the first time, to analyze simultaneously the precise locations of TSSs and the spatiotemporal expression patterns of the corresponding transcripts . This permits detailed analysis of constraints imposed on TATA driven promoters for regulating inducible expression. Here, we show that, in TATA-driven promoters, the TATA-TSS spacing affects the transcriptional specificity of the downstream transcript. We then proceed to show that different TATA-TSS spacings affect a number of core promoter features, including the consensus sequence of the -3 to +1 region and the distribution of alternative TSS. Finally, we show that the overall TSS distribution within SP class promoters is indicative of tissue specificity as well as TATA box and initiation site properties.
CAGE data and promoter classifications
CAGE [20, 21] enables genome-wide localization of TSSs by rapid large-scale sequencing of 5' ends of mRNAs. The data structure and content of the CAGE data repository were described by Carninci and coworkers . CAGE tags consist of sequenced 20-21 base pair (bp) long, 5' ends of full-length cDNAs that have been mapped to the corresponding (mouse or human) genome. Protocols for CAGE were described by Kodzius and colleagues . Overlapping tags on the same strand form a tag cluster (TC) . A TC and its surrounding genomic sequence can be considered a core promoter and is the basic unit used in this work.
A wide variety of RNA libraries (209) and tissues (23) was used for CAGE sequencing in mouse. Because all CAGE tags originate from defined RNA libraries isolated from specific tissues, for each TSS detected by CAGE the distribution of source libraries and tissues is also available. There are multiple lines of evidence for the high reliability and nucleotide-level resolution of CAGE tags, as discussed in detail in the supplementary material presented by Carninci and coworkers .
As discussed above, we previously discovered that promoters where the vast majority of TSSs are constrained to one to four consecutive nucleotides are enriched for TATA boxes and are associated with tissue-specific expression . In the present study, in order to avoid ambiguous estimation of TATA-TSS distances, we analyzed promoters that have a single dominant peak located at a single nucleotide position (see Materials and methods, below). We shall refer to this type of promoters as 'single-TSS' promoters. For clarity, they are a subset of the SP promoter class, as defined by Carninci and coworkers .
Measuring tissue specificity using CAGE expression data
To assess the specificity of the expression of the downstream gene, we compared the tissue distribution of the CAGE tags within the TC with the tissue distribution of all CAGE tags, by computing the relative entropy (the Kullback-Leibler distance) [22, 23] between the two distributions (see Materials and methods, below).
The concept of relative entropy has been applied to diverse computational biology problems, including sequence conservation , single nucleotide polymorphism selection , binding site predictions , and gene expression analysis [9, 23, 27–31]. Yan and coworkers  recently showed that relative entropy can distinguish differentially expressed genes better than other popular methods, such as t-tests, whereas Kasturi and coworkers  showed that clustering of gene expression using relative entropy was superior to Pearson correlation. In particular, Shannon entropy has been used in a number of studies to analyze transcriptional specificity based on cDNA and expressed sequence tag (EST) libraries [9, 30]. Stekel and coworkers  presented a detailed study of statistical properties of related metrics in this context, whereas Schug and colleagues  showed that entropy-based metrics are useful for classifying expression profiles in GNF Gene Expression Atlas  and EST libraries as source datasets.
To demonstrate that relative entropy in combination with the CAGE data correlates with tissue-specific expression, we collected three sets of genes expected to be ubiquitously expressed: a set of 263 housekeeping genes from the HuGEIndex database (identified from microarray experiments) ; 14 genes of the citric acid cycle; and 23 genes of the ubiquitin-mediated proteolysis pathway, as annotated in the KEGG database . We then collected six gene sets identified as tissue-specific using diverse approaches: 17 whole-brain specific genes (based on microarray expression profiles) [35, 36]; 10 heart-specific genes (based on statistical over-representation in EST libraries) ; nine testis-specific genes (based on microarray expression profiles) [35, 38]; 66 liver-specific genes; 12 lung-specific genes; and 20 cerebellum-specific genes, all from the GNF Gene Expression Atlas . We then calculated the tissue specificity for each gene in the sets using relative entropy based on CAGE tags as well as on an independent dataset of EST cluster expression profiles within UniGene  (see Materials and methods, below).
Correlation of tissue specificities measured by relative entropy in CAGE and UniGene EST clusters
EST versus CAGE: Spearman rank correlation coefficient
Spearman rank correlation P value
Number of genes
Whole brain specific
1.10 × 10-3
9.68 × 10-2
1.48 × 10-2
1.32 × 10-6
1.81 × 10-2
<2.20 × 10-16
Citric acid cycle
2.90 × 10-1
Ubiquitin-mediated proteolysis pathway
5.94 × 10-3
8.54 × 10-6
All sets combined
<2.20 × 10-16
High relative entropy signifies great discrepancy between the TC tissue distribution and the background tissue distribution, and therefore temporally or spatially constrained expression of the corresponding gene, whereas two identical distributions will have a relative entropy value of zero. In this report we refer to the relative entropy measurement between the sample and expected distribution as the 'tissue specificity' or 'transcriptional specificity'.
TATA-TSS spacing is associated with transcriptional specificity in vertebrates
A previous, basic descriptive analysis of the distribution of TATA-TSS spacing established that the most common spacings are 30 and 31 bp and that the great majority of TATA-driven promoters have a distance of 27-34 bp between TATA and the TSS  in mouse. Because our goal in this work was to elucidate whether there is a link between transcriptional specificity and TATA-TSS spacing, we sought both to increase the number of promoters analyzed and to focus on cases in which the TATA-TSS distance is unambiguous. Therefore, we applied a more conservative detection procedure to a larger amount of core promoters where the absolute majority of TSSs were concentrated on a single nucleotide position (the single-TSS class of TCs [see Materials and methods, below]). Only promoters with at least one predicted TATA box with a score greater than 75% within the -40 to -19 bp region relative to the dominant start site were used for subsequent analyses. This resulted in 784 single-TSS promoters used for the subsequent analysis.
Initially, we focused on the most prominent TATA box found in each single-TSS promoter (the highest scoring predicted) [6, 40] TATA box location. We then measured the spacing between the first T in the TATA box (as defined by Bucher ) and the highest CAGE tag peak found in the TC (for simplicity, we refer to this position as 'TSS'). The findings we present below are not dependent on these specific cutoffs; changes in score cutoff and/or application of cross-species filtering of the promoter sets give similar results (data not shown).
We assessed the impact of TATA-TSS spacing on overall tissue specificity by measuring the relative expression entropy of the TCs grouped by TATA-TSS distance, as described in Materials and methods (below). When discussing positions within the promoter, we use the word 'upstream' to mean in the 5' direction of a given location in the promoter, with respect to the strand of the produced transcript (in all relevant figures, this is equivalent to the left-hand side). Similarly, we use the word 'downstream' for locations 3' of a given position (right-hand side in figures).
Correlation between TATA location and initiation signal
Although small differences in TATA box consensus exist between spacing classes, the most important difference is in the properties of the sequence motif near the TSS; the initiation site consensus as well as the distribution of alternative TSSs are dependent on TATA-TSS spacing.
Given the findings above, we argue that the additional signal strength (Figures 6 and 7) found around the initiation site in promoters with extended TATA-TSS spacings is not due to the existence of shared PIC binding site motifs in these promoters, but is due to the absence of a PyPu transcription initiation site at a more favorable spacing (Figure 5). Because information content is a measure of constraints in selection of symbols (in this case nucleotides), negative selection against a subset of symbols will increase the information content.
Consistent with the previous initiation site analysis (Figure 4), promoters where the TATA box is located at -28 have a weaker initiation site with an SR consensus at [-1, +1] (Figure 6a). The atypical promoter structure together with the low tissue specificity suggests that the mechanism for TATA-driven transcription is different in promoters with this spacing type.
We checked for the possibility that the TATA boxes at -28 could actually represent bona fide TATA boxes at -30, which would render the TATA box at position -28 redundant. However, the logo summarizing TATA boxes detected at -28 shows no support for this explanation. Additionally, the promoter structure and differential use of initiation site sequences between the promoters with TATA boxes at -30 and -28 makes the proposition unlikely. If a majority of TATA boxes located at -28 had a functional (and preferentially used) TATA box at -30, then we would expect the initiation site distributions to be similar for both spacing classes. On the other hand, the logo representing the TATA boxes located at position -34 (Figure 6g) has a TATATAA consensus instead of the TATAAA seen in the other spacings. This clearly shows the potential for parallel usage of TATA boxes at positions -34 and -32 (using the first and second T in the TATATAA).
The CAGE tag distribution around the dominant start position also reflects the spacing classes (Figure 6). As expected, positions -31 and -30 have the smallest CAGE tag distribution skew (the number of CAGE tags at each side of the dominant start site is approximately equal). Interestingly, the CAGE tag distribution in promoters with TATA boxes at position -31 is close to perfectly symmetrical, whereas there is a small skew toward the larger spacings at promoters where the TATA box is located at -30 (Figure 6c,d). Promoters where the TATA box is located elsewhere exhibit a considerable skew, which is fully consistent with the location of the sites, because they are skewed in the direction of more favorable spacings; promoters where the TATA box is located at positions -28 and -29 have alternative TSSs located downstream of the main TSS. Conversely, promoters with the TATA box at -32, -33, and -34 have alternative TSSs upstream of the main TSS (Figure 6). In both cases, the effect of choosing the indicated alternative TSSs would be a TATA-TSS spacing of 30 or 31 bp. In the case of promoters where the TATA-box is located at -34, there is potential for usage of both alternative TSSs and alternative TATA-boxes, as discussed above.
TBP binding strength has minor effects on transcriptional specificity compared to spacing
Next, we investigated whether the existence of several, possibly overlapping bona fide TATA boxes in a single core promoter can influence the expression specificity, by analyzing the correlation between the sum of scores for all predicted TATA boxes, exceeding a 75% score threshold along the promoter (see Materials and methods, below), and their transcriptional specificity (Figure 9b). As above, we found no correlation (R2 = 6.1 × 10-3). Finally, we repeated the same analysis with no score constraints in order to investigate whether the total binding potential for TBP along the promoter might have a significant influence (data not shown), but we found no correlation (R2 = 8.1 × 10-3).
Taken together, these results imply that there exists a certain operational range of dissociation constant values for TBP-DNA interaction that is required for efficient TATA box guided transcription, but that there is no preferred strength of interaction within that range.
Promoter shape modulates TATA-driven expression
Apart from the TATA-TSS distance, we found that the overall shape of the promoters within the SP class is indicative of transcriptional specificity and/or other promoter characteristics. We have focused on two 'borderline' subtypes of promoters found within the SP class set defined by Carninci and coworkers .
The first subtype includes promoters with a single peak in combination with a uniform distribution of CAGE tags stretching over 50 bp. We refer to these as 'shallow-TSS' promoters. This set is a subset of the single-TSS set analyzed above for TATA spacing properties.
The second subtype includes promoters with two dominant peaks with a spacing of 0-3 bp. We refer to these as 'twin-TSS' promoters. This set is disjoint from the single-TSS set.
Representative examples of tag clusters of the shallow-TSS and twin-TSS promoter subclasses are shown in Figure 1.
Shallow-TSS promoters are less effective for driving context-specific expression
We previously showed that promoters where the CAGE tags are distributed shallowly (the broad class [BR]) are associated with ubiquitously expressed genes and have high over-representation of CpG islands . Therefore, it is not unreasonable that SP promoters with BR-like characteristics would be less suitable for directing specific expression. As described above, we tested the subset of 76 shallow-TSS promoters harboring TATA boxes against the remaining set of 708 single-TSS promoters harboring TATA boxes. The overall transcriptional selectivity of shallow-TSS promoter subset is lower (P = 4.0 × 10-2; one-tail Wilcoxon test), although the P value is marginally significant. Interestingly, this is also true if we only consider the dominant peak of the promoters in both sets (we ignore the flanking tags; P = 4.1 × 10-2; one-tail Wilcoxon test). Within a shallow-TSS promoter, the dominant peak generally has a higher transcriptional specificity than the flanking tags (P = 1.32 × 10-4; one-tail paired Wilcoxon test). Unexpectedly, the transcriptional specificity of the dominant peaks are highly correlated with that of the flanking tags (P < 2.2 × 10-16; two-sided Spearman rank correlation test), suggesting that the shape of these promoters cannot be explained by two overlaid tag distributions with different levels of tissue specificity.
Spacing between TSSs in twin-TSS promoters affects promoter structure
We found that promoters with a genomic spacing of 1-3 bp between the peaks have an unmistakable TATA consensus starting at around -30 and exhibit PyPu consensus initiation sites (Figure 10b-d). Conversely, promoters with two adjacent peaks (no spacing) have a significant under-representation of TATA boxes compared with the other twin-TSS promoters (P = 5.6 × 10-6; two-tailed Fisher's exact test [see Materials and methods, below]). These promoters also have a radically different signal near the initiation site: a GGG consensus, where the last G is located at position +1 (Figure 10a). Although the consensus is similar to the initiation site motif found previously in transcripts starting in 3' untranslated regions of protein encoding genes , we can at present only speculate on whether the mechanisms governing these types of promoters are similar.
We also investigated whether the transcriptional specificity of TATA-driven twin-TSS promoters is significantly different from that of single-TSS promoters. Intriguingly, the twin-TSS promoters might have a greater transcriptional specificity than the single-TSS promoters (P = 4.5 × 10-2; one-tail Wilcoxon test). However, because relative entropy values for the twin-TSS set are dominated by a few extreme outliers, it is unclear whether this observation holds in general. This implies that there are highly tissue-specific promoters that use two closely located TSSs, but it is unclear whether these are guided by two overlapping TATA boxes or by a mechanism in which the PIC chooses between the two comparably favorable TSSs (see Discussion, below).
Determination of optimal TATA-TSS spacing
We have found that the spacing of the TATA-TSS is associated with tissue-specific expression (Figure 3). In particular, positions -31 and -30 are most strongly associated with context-specific transcription.
In comparison with the TATA-TSS spacing, the strength of TBP-TATA interaction does not appear to be correlated with the tissue specificity, only requiring that the interaction strength between TBP and a potential TATA box exceeds some threshold level.
The effects of TATA-TSS spacing on transcriptional specificity have been studied in depth within a few plant promoters. Zhu and coworkers  showed that, in Oryza sativa, the phenylalanine-lyase promoter activity in vitro was eliminated when a 6 bp element was either deleted from positions -21 to -16 or inserted between positions -18 and -19. This is entirely consistent with our more comprehensive study, because transferring the TATA box 6 bp upstream or downstream would take its starting locations outside the range of acceptable TSSs, as defined above. In a more detailed study of the developmentally important β-phaseolin gene promoter , multiple insertions and deletions were used to dissect the promoter function. Insertions between the TATA boxes and the initiation sites conferred either a significant decrease in transcription or creation of new TSS with a more favorable spacing (30 or 31 bp) relative to the TATA box, which is consistent with our analysis. Similarly, O'Shea-Greenfield and coworkers  showed that maximal expression in an in vitro system using human cell nuclear extracts was achieved when the TATA-TSS distance was 30 bp, and that when extending the distance from 30 to 35 or 40 nucleotides the start site was dislocated to a position 30 bp downstream of the TATA box.
Although our study shows the functional importance of the distance separating the TATA box and the TSS, the underlying mechanism that determines the start site selection is not fully understood, despite high-resolution X-ray structure determinations of the PIC and the polymerase II complex . In TATA-driven promoters in higher eukaryotes, the TATA box functions as an anchor for the rest of the PIC, thus sterically focusing the selection of initiation sites to a limited range of positions. It is important to note that at present it is not fully understood whether the TATA-TSS spacing in itself contributes to changes in transcriptional specificity, or whether the observed spacings are consequences of other events, such as the mechanistic constraints imposed by the PIC and other trans-acting regulatory proteins.
A recent genome-wide survey of Arabidopsis thaliana core promoters  indicates that plant TATA box driven promoters probably share the spatial constraints presented herein. The authors estimated that the ideal TATA-TSS spacing in A. thaliana is 32 bp, but this analysis lacked the depth and resolution of TSS data that now are available for mouse [7, 8]. The allowed TATA-box position distribution is similar to that of mouse, in which TATA boxes at positions closer than -29 are rarely observed, and larger distances are tolerated more often. The results in A. thaliana clearly show an immediate application for the insights we have presented in this study; the precise rules established here are valid across many eukaryotes, and can be applied for annotation of TATA-driven TSSs of those genomes in which the TSS data are not available or not precise enough.
Promoter shape and initiation site consensus
As discussed, our results indicate that the TATA box must lie within a narrow 4 bp region (-32 to -29) in order to achieve high transcriptional specificity. When the TATA box is located within this region, the initiation site at [-1, +1] is dominated by a PyPu dinucleotide consensus. In the case of TATA motifs located upstream of -32, the PyPu consensus is retained but extended for TATA boxes at positions -33 and -34 (Figure 6f,g and Figure 7), which is due to an absence of PyPu initiation sites at more favorable distances upstream of the actual TSS (Figure 5). In these promoters there is also an evident skew in the CAGE tag distribution, indicating that if alternative minor start sites exist then they preferentially use more favorable spacings (closer to positions -30 and -31).
Our interpretation of these extended spacing classes can be divided into two different but not mutually exclusive hypotheses. First, because the TATA motif is more expanded and variable at these positions, there is a possibility that a weaker TATA box 2 bp downstream is used instead of the site indicated in our analysis. However, a consistent use of the downstream TATA box would not explain the high scoring TATA boxes at position -34 or -33, the skew of usage of minor initiation sites towards canonical spacing, or the depletion of PyPu dinucelotides at positions [-2, -3] (Figure 5), which is not present at other positions. A more likely explanation is the parallel use of both TATA boxes in the promoter. Only experimental follow up can resolve which of the putative sites is preferentially used.
A second, alternative explanation is that there is an intrinsic 'stretching' potential in the PIC anchored to the TATA box, resulting in the possibility of selecting TSS located further downstream when no suitable initiation site is present at the canonical distance. Promoters with a TATA box located at position -28 have a significantly different initiation site distribution in terms of PyPu (Figure 4). Because the PyPu initiation site is ambiguous in these promoters, it is reasonable to believe that the PIC stretching potential suggested above can accommodate extended but not decreased TATA-TSS distances. As in the case of more extended spacings, there is a skew in the CAGE distribution toward more canonical spacings.
These results suggest that the mechanism for TATA-TSS interaction by the PIC is comparable for promoters where the TATA box is located at positions -34 to -29. Conversely, the combination of atypical initiation sites and radically decreased transcriptional specificity for promoters where the TATA-box is located at position -28 suggests that this type of interaction is governed by at least partially different mechanisms.
Correlation between CAGE distribution and TATA occurrence
In the concluding part of our analysis we looked at two related classes or promoters that depart from the 'ideal' single-peaked distributions. Our data indicate that the shallow-TSS class promoter might have a lower transcriptional specificity than the remaining single-peak class; this seems to be true also for the dominant TSS position.
The twin-TSS class promoters have a TATA box pattern when the spacing of the two dominant TSS peaks ranges from 1 to 3 bp. However, if the two TSS peaks have no spacing, then the promoters are TATA depleted and have a novel initiation site sequence motif (Figure 10).
The presented results demonstrate that the interdependence of TATA motifs and the associated TSSs reflect underlying promoter architecture and mechanisms.
The underlying features of the CAGE data used in this study have enabled the discovery that TATA-TSS spacing is associated with the transcriptional specificity of the downstream transcript, the TSS distribution of the promoters, and initiation site motifs. Although our understanding of the functional mechanism that governs core promoters in general and the TATA-TSS interaction in particular is still limited, the results presented here will provide fertile ground for more detailed studies of core promoters. Our findings can be also used to resolve a substantial subset of ambiguities that arise from unreliable determination of TSSs, and will be an asset when annotating putative TATA boxes in uncharacterized promoters. The rules inferred for TATA boxes are directly applicable to the design of expression vectors in vertebrate systems, and suggest further directions in experimental investigation of transcriptional initiation from TATA-dependent promoters.
The combination of accurate, high-throughput TSS determination, systematic detection of cis-acting elements (for instance, ChIP2-chip [52, 53]), and computational analysis offers a breadth of targets with a sufficient data depth to explore genome-wide principles. CAGE tag distributions reveal patterns of TSS usage in core promoters that will greatly advance our understanding of core promoter function and help to guide future promoter annotation and characterization experiments, both individual and genome wide.
Materials and methods
Experimental data sources
We used the FANTOM3 CAGE collection [7, 8] for assessing TSSs in mouse (Mus musculus). The experimental procedure for production and mapping of CAGE tags to the genome Is described elsewhere [8, 21]. The full set of 7,151,511 mapped CAGE tags was derived from 209 different RNA libraries and 23 tissues.
In our analysis we used a restricted set based on the 5,655,682 mapped tags originating from the 15 tissues, each containing at least 10,000 mapped tags. We removed tags from whole-body libraries, as well as macrophage libraries, because macrophages are present in almost all tissues and macrophage-specific genes have purine-rich proximal promoters that are not TATA associated . The CAGE data are described by Kawaji and coworkers  and publicly available on the internet . We consistently used the G-correction algorithm, as presented and used by Carninci and coworkers , for TSS locations.
Promoter sets used in analyses
As in our previous study, CAGE TCs were used to define core promoter locations. Briefly, a CAGE TC consists of CAGE tags overlapping by at least 1 bp on the same strand . Mouse TCs, containing at least 50 tags from the CAGE set defined above, were assigned a single peak shape (SP) if the distance between the 25 and 75 tag density percentile was less than 4 bp. A total of 2863 core promoters fulfilled this classification criterion and formed the initial set for selecting TATA-driven promoters; this is the same definition as was used by Carninci and coworkers , although that study used TCs with at least 100 tags. The reason for using the same initial definition was to make it possible to reflect our findings here with those made previously . From this initial set we analyzed three subsets.
A promoter was classed as twin-TSS if it fulfilled the following criteria: one of the neighboring TSSs (± 4 bp) relative to the highest TSS peak must contain at least 25% of CAGE tags of the highest TSS peak; and these two start site positions must contain more than 75% of the total CAGE tags within the TC. In total, 465 promoters were classed as twin-TSS, in which the lowest observed tag count for any of the two peaks was nine.
This is equivalent to the initial set of SP core promoters, excluding the twin-TSS promoters defined above. In total, 2398 promoters were classed as single-TSS, in which the lowest observed tag count of the main peak was 16.
This is a subset of the single-TSS promoters that fulfilled the following criteria: the TC must consist of at least 30 start site positions spanning a region greater than 50 bp; the sum of all tags within the TC, excluding those contained in the highest TSS peak, must be at least 100; and except for the dominant peak, each distinct start site must contain 20% or fewer of CAGE tags of the most dominant peak. In total, 185 promoters were classed as shallow-TSS.
Determination of TATA box locations
We determined the occurrence of TATA boxes upstream 40 to 19 bp of the most dominant tag peak of these promoters by using the TATA model constructed by Bucher  deposited in the JASPAR database  and the TFBS Perl programming module  for predicting potential TBP binding sites. For clarity, the start of the TATA box was annotated as the first T of the TATA motif and the second position of Bucher's model.
For selecting likely TBP-binding sites we only accepted site predictions on the same strand as the transcript and exceeding a relative score threshold of 75%. For the different types of analysis in this work, we distinguished between three cases: we considered the best scoring TATA box prediction using the thresholds as above, the sum of all predicted sites with each site scoring greater than the defined threshold, or the sum of all predicted sites in the specified region without any relative score threshold constraint.
The vast majority of the analyses were made on the single-TSS promoter set. In total, 784 TCs in this set were assigned a predicted best scoring TATA box, whereas 2114 TCs in this set were considered when applying no relative score threshold criteria.
Measuring transcriptional specificity using CAGE
Transcriptional specificity was measured by the relative entropy (the Kullback-Leibler distance) [22, 59] of the tissue distribution of a sample TC with respect to the tissue distribution of all 5,655,682 CAGE tags:
where k is the number of different tissues (n = 15), p is the discrete probability distribution of tissues in the sample tag cluster, and q is the discrete probability distribution of tissues for all tags. The distance cannot be negative, and if p = q then the distance d will be 0.
Measuring transcriptional specificity using ESTs
In a comparative study, we measured the transcriptional specificity based on EST clusters from the UniGene database; more specifically, the Mm.profiles file from the UniGene ftp repository  was used as a data source. It summarizes the expression profile of ESTs in each cluster from libraries with curated and controlled vocabulary tissue annotation, in which each cluster has at least 10 tags. Relative entropy was calculated using the equation given above, where k is the number of tissues, p is the discrete probability distribution of tissues in the sample EST cluster, and q is the discrete probability distribution of tissues for all ESTs.
Extraction of tissue-specific genes from literature and Internet sources
The gene sets were taken from the supplementary material of each publication, except for the GNF-derived data, which were retrieved using the SymAtlas web tool ; we selected all mouse genes from the Mouse GeneAtlas U74A  set, which had an expression fold over 30 of the median using the web retrieval tool for liver and adipose separately. The same procedure was repeated for lung using a 25-fold threshold.
For all gene sets, we only included genes that were covered both by CAGE and EST data. To be able to compare EST clusters and CAGE TCs, we used the official mouse gene symbol names for linking purposes. In cases in which several alternative promoters existed in the CAGE database, we selected the TC with the largest number of tags. Within this analysis, we did not exclude macrophage and whole body libraries, because it was unreasonable to treat CAGE and EST sets differently. We also used a CAGE tag count threshold of 30 tags for TCs included in the analysis in order to be closer to the 10 EST threshold used in the UniGene cluster database.
Analysis of differential initiation site distribution for promoters with the TATA box located at -28
We applied the χ2 test for the frequency distribution, as implemented by Ihaka and Gentleman , of the four different dinucleotide classes (PyPu, PyPy, PuPy, and PuPu) at the initiation site [-1, +1] in order to determine whether the initiation site distribution from promoters in which the TATA box is located at -28 can be considered significantly different from the initiation site distribution from all promoters with the TATA box situated at position -34 to -29.
Analysis of PyPu dinucleotide usage in the vicinity of observed TSS
We wished to assess the occurrence of PyPu dinucleotides immediately upstream and downstream of the TSS [-1, +1] in the different TATA-TSS spacing classes (-34 to -28). Using a 2 bp sliding window, we counted the PyPu dinucleotides in the region ± 5 bp of the TSS for each TATA-dependent promoter sequence, normalized by the number of promoters in each spacing class.
Specific TATA-TSS sequence logos and corresponding CAGE tag distributions
We classified each TC in terms of the spacing between the best scoring TATA box prediction that fulfilled the selection criteria listed above, and the initiation site (the highest CAGE peak within the TC, for convenience referred to as 'TSS' and located at +1). We only considered TATA boxes in a restricted spacing interval from position -34 to -28, because the absolute majority of functional TATA boxes reside in this region (Figure 3b). We then extracted the -40 to +25 sequence region relative to the TSS and created a sequence logo  for each TATA-TSS spacing class using the TFBS programming modules . Small sample correction was applied as described Schneider and coworkers [43, 44] and implemented by Lenhard and Wasserman . We measured the signal strength of the region surrounding the initiation site by calculating the total information content  of the -5 to +5 sequence region for each given spacing class.
In order to visualize the CAGE tag distributions in that region, the frequency distribution of CAGE tags was obtained for each TC (one bin per bp) and then normalized by its total number of CAGE tags within the TC. For each position in the logo, we calculated the median tag density from the array of vectors defined above.
Hidden Markov model simulation for exploring signal strength effects of PyPu depletion
To investigate how the signal strength immediately upstream of the observed TSS is affected by depletion of PyPu dinucleotides, we constructed a simple HMM that generates sequences according to a set of rules. The rules are a simplification of the biologic reality, because the goal is just to explore the principal effects of PyPu depletion. The HMM generates a sequence corresponding to the regions surrounding the TSS (-7 to +3) from left to right using the following rule set: there must be a PyPu at [-1, +1]; when selecting a nucleotide at position i in the region immediately upstream of the [-1, +1] region, nucleotides [i-1, i] must not form a PyPu dinucleotide; and aside from these constraints, all nucleotides are considered equally likely for selection.
Unless otherwise indicated, we used the R language  for statistical analysis and graph visualization. We used the TFBS  library for promoter pattern analysis and sequence logo drawing, and the AT libraries (Engström P, Andersen M, Sandelin A, Fredman D, Lenhard B, unpublished data) for sequence handling and genome informatics.
We thank Ann Karlsson at the Genome Exploration Research Group (Genome Network Project Core Group), RIKEN Genomic Sciences Center (GSC), RIKEN Yokohama Institute, and Chris Ponting at the MRC Functional Genetics Unit, University of Oxford for help with editing of the manuscript. JP gratefully acknowledges a joint research grant from the Studienstiftung des deutschen Volkes and the RIKEN Institute, a graduate Clarendon Award, Oxford Balliol College Domus Award, and a graduate scholarship by the Studienstiftung des deutschen Volkes. BL was supported by Pharmacia Corporation (now Pfizer), the Swedish Research Council, and The National Programme for Research in Functional Genomics in Norway (FUGE) of the Research Council of Norway. This work is supported by a Research Grant for the RIKEN Genome Exploration Research Project from the Ministry of Education, Culture, Sports, Science and Technology of the Japanese Government to YH; a grant of the Genome Network Project from the Ministry of Education, Culture, Sports, Science and Technology, Japan to YH; and a grant for the Strategic Programs for R&D of RIKEN to YH. We thank two anonymous referees for constructive criticism.
- Kadonaga JT: Regulation of RNA polymerase II transcription by sequence-specific DNA binding factors. Cell. 2004, 116: 247-257. 10.1016/S0092-8674(03)01078-X.PubMedView ArticleGoogle Scholar
- Butler JE, Kadonaga JT: The RNA polymerase II core promoter: a key component in the regulation of gene expression. Genes Dev. 2002, 16: 2583-2592. 10.1101/gad.1026202.PubMedView ArticleGoogle Scholar
- Smale ST, Kadonaga JT: The RNA polymerase II core promoter. Annu Rev Biochem. 2003, 72: 449-479. 10.1146/annurev.biochem.72.121801.161520.PubMedView ArticleGoogle Scholar
- Hampsey M: Molecular genetics of the RNA polymerase II general transcriptional machinery. Microbiol Mol Biol Rev. 1998, 62: 465-503.PubMedPubMed CentralGoogle Scholar
- Hahn S: Structure and mechanism of the RNA polymerase II transcription machinery. Nat Struct Mol Biol. 2004, 11: 394-403. 10.1038/nsmb763.PubMedPubMed CentralView ArticleGoogle Scholar
- Wasserman WW, Sandelin A: Applied bioinformatics for the identification of regulatory elements. Nat Rev Genet. 2004, 5: 276-287. 10.1038/nrg1315.PubMedView ArticleGoogle Scholar
- Carninci P, Kasukawa T, Katayama S, Gough J, Frith MC, Maeda N, Oyama R, Ravasi T, Lenhard B, Wells C, et al: The transcriptional landscape of the mammalian genome. Science. 2005, 309: 1559-1563. 10.1126/science.1112014.PubMedView ArticleGoogle Scholar
- Carninci P, Sandelin A, Lenhard B, Katayama S, Shimokawa K, Ponjavic J, Semple CA, Taylor MS, Engstrom PG, Frith MC, et al: Genome-wide analysis of mammalian promoter architecture and evolution. Nat Genet. 2006, 38: 626-635. 10.1038/ng1789.PubMedView ArticleGoogle Scholar
- Schug J, Schuller WP, Kappen C, Salbaum JM, Bucan M, Stoeckert CJ: Promoter features related to tissue specificity as measured by Shannon entropy. Genome Biol. 2005, 6: R33-10.1186/gb-2005-6-4-r33.PubMedPubMed CentralView ArticleGoogle Scholar
- Taylor MS, Kai C, Kawai J, Carninci P, Hayashizaki Y, Semple CA: Heterotachy in mammalian promoter evolution. PLoS Genet. 2006, 2: e30-10.1371/journal.pgen.0020030.PubMedPubMed CentralView ArticleGoogle Scholar
- Frith MC, Ponjavic J, Fredman D, Kai C, Kawai J, Carninci P, Hayshizaki Y, Sandelin A: Evolutionary turnover of mammalian transcription start sites. Genome Res. 2006, 16: 713-722. 10.1101/gr.5031006.PubMedPubMed CentralView ArticleGoogle Scholar
- Bushnell DA, Westover KD, Davis RE, Kornberg RD: Structural basis of transcription: an RNA polymerase II-TFIIB cocrystal at 4.5 Angstroms. Science. 2004, 303: 983-988. 10.1126/science.1090838.PubMedView ArticleGoogle Scholar
- Grosveld GC, Shewmaker CK, Jat P, Flavell RA: Localization of DNA sequences necessary for transcription of the rabbit beta-globin gene in vitro. Cell. 1981, 25: 215-226. 10.1016/0092-8674(81)90246-4.PubMedView ArticleGoogle Scholar
- Kovacs BJ, Butterworth PH: The effect of changing the distance between the TATA-box and cap site by up to three base pairs on the selection of the transcriptional start site of a cloned eukaryotic gene in vitro and in vivo. Nucleic Acids Res. 1986, 14: 2429-2442.PubMedPubMed CentralView ArticleGoogle Scholar
- O'Shea-Greenfield A, Smale ST: Roles of TATA and initiator elements in determining the start site location and direction of RNA polymerase II transcription. J Biol Chem. 1992, 267: 1391-1402.PubMedGoogle Scholar
- Zhu Q, Dabi T, Lamb C: TATA box and initiator functions in the accurate transcription of a plant minimal promoter in vitro. Plant Cell. 1995, 7: 1681-1689. 10.1105/tpc.7.10.1681.PubMedPubMed CentralView ArticleGoogle Scholar
- Grace ML, Chandrasekharan MB, Hall TC, Crowe AJ: Sequence and spacing of TATA box elements are critical for accurate initiation from the beta-phaseolin promoter. J Biol Chem. 2004, 279: 8102-8110. 10.1074/jbc.M309376200.PubMedView ArticleGoogle Scholar
- Ren D, Nedialkov YA, Li F, Xu D, Reimers S, Finkelstein A, Burton ZF: Spacing requirements for simultaneous recognition of the adenovirus major late promoter TATAAAAG box and initiator element. Arch Biochem Biophys. 2005, 435: 347-362. 10.1016/j.abb.2004.12.028.PubMedView ArticleGoogle Scholar
- Wolffe AP, Guschin D: Review: chromatin structural features and targets that regulate transcription. J Struct Biol. 2000, 129: 102-122. 10.1006/jsbi.2000.4217.PubMedView ArticleGoogle Scholar
- Shiraki T, Kondo S, Katayama S, Waki K, Kasukawa T, Kawaji H, Kodzius R, Watahiki A, Nakamura M, Arakawa T, et al: Cap analysis gene expression for high-throughput analysis of transcriptional starting point and identification of promoter usage. Proc Natl Acad Sci USA. 2003, 100: 15776-15781. 10.1073/pnas.2136655100.PubMedPubMed CentralView ArticleGoogle Scholar
- Kodzius R, Kojima M, Nishiyori H, Nakamura M, Fukuda S, Tagami M, Sasaki D, Imamura K, Kai C, Harbers M, et al: CAGE: cap analysis of gene expression. Nat Methods. 2006, 3: 211-222. 10.1038/nmeth0306-211.PubMedView ArticleGoogle Scholar
- Kullback S, Leibler R: On information and sufficiency. Ann Math Stat. 1951, 22: 79-86.View ArticleGoogle Scholar
- Steuer R, Kurths J, Daub CO, Weise J, Selbig J: The mutual information: detecting and evaluating dependencies between variables. Bioinformatics. 2002, 8 (Suppl 2): S231-S240.View ArticleGoogle Scholar
- Tseng YY, Liang J: Are residues in a protein folding nucleus evolutionarily conserved?. J Mol Biol. 2004, 335: 869-880. 10.1016/j.jmb.2003.11.007.PubMedView ArticleGoogle Scholar
- Liu Z, Lin S: Multilocus LD measure and tagging SNP selection with generalized mutual information. Genet Epidemiol. 2005, 29: 353-364. 10.1002/gepi.20092.PubMedPubMed CentralView ArticleGoogle Scholar
- Prakash P, Yellaboina S, Ranjan A, Hasnain SE: Computational prediction and experimental verification of novel IdeR binding sites in the upstream sequences of Mycobacterium tuberculosis open reading frames. Bioinformatics. 2005, 21: 2161-2166. 10.1093/bioinformatics/bti375.PubMedView ArticleGoogle Scholar
- Fuhrman S, Cunningham MJ, Wen X, Zweiger G, Seilhamer JJ, Somogyi R: The application of shannon entropy in the identification of putative drug targets. Biosystems. 2000, 55: 5-14. 10.1016/S0303-2647(99)00077-5.PubMedView ArticleGoogle Scholar
- Kasturi J, Acharya R, Ramanathan M: An information theoretic approach for analyzing temporal patterns of gene expression. Bioinformatics. 2003, 19: 449-458. 10.1093/bioinformatics/btg020.PubMedView ArticleGoogle Scholar
- Pollet N, Schmidt HA, Gawantka V, Niehrs C, Vingron M: In silico analysis of gene expression patterns during early development of Xenopus laevis. Pac Symp Biocomput. 2000, 443-454.Google Scholar
- Stekel DJ, Git Y, Falciani F: The comparison of gene expression from multiple cDNA libraries. Genome Res. 2000, 10: 2055-2061. 10.1101/gr.GR-1325RR.PubMedPubMed CentralView ArticleGoogle Scholar
- Yan X, Deng M, Fung WK, Qian M: Detecting differentially expressed genes by relative entropy. J Theor Biol. 2005, 234: 395-402. 10.1016/j.jtbi.2004.11.039.PubMedView ArticleGoogle Scholar
- Su AI, Cooke MP, Ching KA, Hakak Y, Walker JR, Wiltshire T, Orth AP, Vega RG, Sapinoso LM, Moqrich A, et al: Large-scale analysis of the human and mouse transcriptomes. Proc Natl Acad Sci USA. 2002, 99: 4465-4470. 10.1073/pnas.012025199.PubMedPubMed CentralView ArticleGoogle Scholar
- Haverty PM, Weng Z, Best NL, Auerbach KR, Hsiao LL, Jensen RV, Gullans SR: HugeIndex: a database with visualization tools for high-density oligonucleotide array data from normal human tissues. Nucleic Acids Res. 2002, 30: 214-217. 10.1093/nar/30.1.214.PubMedPubMed CentralView ArticleGoogle Scholar
- Kanehisa M, Goto S, Hattori M, Aoki-Kinoshita KF, Itoh M, Kawashima S, Katayama T, Araki M, Hirakawa M: From genomics to chemical genomics: new developments in KEGG. Nucleic Acids Res. 2006, 34 (Database issue): D354-D357. 10.1093/nar/gkj102.PubMedPubMed CentralView ArticleGoogle Scholar
- Plotkin JB, Robins H, Levine AJ: Tissue-specific codon usage and the expression of human genes. Proc Natl Acad Sci USA. 2004, 101: 12588-12591. 10.1073/pnas.0404957101.PubMedPubMed CentralView ArticleGoogle Scholar
- Hsiao LL, Dangond F, Yoshida T, Hong R, Jensen RV, Misra J, Dillon W, Lee KF, Clark KE, Haverty P, et al: A compendium of gene expression in normal human tissues. Physiol Genomics. 2001, 7: 97-104.PubMedView ArticleGoogle Scholar
- Megy K, Audic S, Claverie JM: Heart-specific genes revealed by expressed sequence tag (EST) sampling. Genome Biol. 2002, 3: RESEARCH0074-PubMedPubMed CentralGoogle Scholar
- Warrington JA, Nair A, Mahadevappa M, Tsyganskaya M: Comparison of human adult and fetal expression and identification of 535 housekeeping/maintenance genes. Physiol Genomics. 2000, 2: 143-147.PubMedGoogle Scholar
- Wheeler DL, Barrett T, Benson DA, Bryant SH, Canese K, Chetvernin V, Church DM, DiCuccio M, Edgar R, Federhen S, et al: Database resources of the National Center for Biotechnology Information. Nucleic Acids Res. 2006, 34 (Database issue): D173-D180. 10.1093/nar/gkj158.PubMedPubMed CentralView ArticleGoogle Scholar
- Stormo GD: DNA binding sites: representation and discovery. Bioinformatics. 2000, 16: 16-23. 10.1093/bioinformatics/16.1.16.PubMedView ArticleGoogle Scholar
- Bucher P: Weight matrix descriptions of four eukaryotic RNA polymerase II promoter elements derived from 502 unrelated promoter sequences. J Mol Biol. 1990, 212: 563-578. 10.1016/0022-2836(90)90223-9.PubMedView ArticleGoogle Scholar
- Burke TW, Kadonaga JT: The downstream core promoter element, DPE, is conserved from Drosophila to humans and is recognized by TAFII60 of Drosophila. Genes Dev. 1997, 11: 3020-3031.PubMedPubMed CentralView ArticleGoogle Scholar
- Schneider TD, Stephens RM: Sequence logos: a new way to display consensus sequences. Nucleic Acids Res. 1990, 18: 6097-6100.PubMedPubMed CentralView ArticleGoogle Scholar
- Schneider TD, Stormo GD, Gold L, Ehrenfeucht A: Information content of binding sites on nucleotide sequences. J Mol Biol. 1986, 188: 415-431. 10.1016/0022-2836(86)90165-8.PubMedView ArticleGoogle Scholar
- Schneider TD: Evolution of biological information. Nucleic Acids Res. 2000, 28: 2794-2799. 10.1093/nar/28.14.2794.PubMedPubMed CentralView ArticleGoogle Scholar
- Shannon CE: A mathematical theory of communication. Bell Syst Tech J. 1948, 27: 379-423.View ArticleGoogle Scholar
- Eddy SR: Profile hidden Markov models. Bioinformatics. 1998, 14: 755-763. 10.1093/bioinformatics/14.9.755.PubMedView ArticleGoogle Scholar
- Durbin R, Eddy S, Krogh H, Mitchison G: Biological Sequence Analysis. 1999, Cambridge: Cambridge University PressGoogle Scholar
- Berg OG, von Hippel PH: Selection of DNA binding sites by regulatory proteins. Statistical-mechanical theory and application to operators and promoters. J Mol Biol. 1987, 193: 723-750. 10.1016/0022-2836(87)90354-8.PubMedView ArticleGoogle Scholar
- Sandelin A, Wasserman WW, Lenhard B: ConSite: web-based prediction of regulatory elements using cross-species comparison. Nucleic Acids Res. 2004, 32 (Web Server issue): W249-W252.PubMedPubMed CentralView ArticleGoogle Scholar
- Molina C, Grotewold E: Genome wide analysis of Arabidopsis core promoters. BMC Genomics. 2005, 6: 25-10.1186/1471-2164-6-25.PubMedPubMed CentralView ArticleGoogle Scholar
- Martone R, Euskirchen G, Bertone P, Hartman S, Royce TE, Luscombe NM, Rinn JL, Nelson FK, Miller P, Gerstein M, et al: Distribution of NF-kappaB-binding sites across human chromosome 22. Proc Natl Acad Sci USA. 2003, 100: 12247-12252. 10.1073/pnas.2135255100.PubMedPubMed CentralView ArticleGoogle Scholar
- Mukherjee S, Berger MF, Jona G, Wang XS, Muzzey D, Snyder M, Young RA, Bulyk ML: Rapid analysis of the DNA-binding specificities of transcription factors with DNA microarrays. Nat Genet. 2004, 36: 1331-1339. 10.1038/ng1473.PubMedPubMed CentralView ArticleGoogle Scholar
- O'Reilly D, Addley M, Quinn C, MacFarlane AJ, Gordon S, McKnight AJ, Greaves DR: Functional analysis of the murine Emr1 promoter identifies a novel purine-rich regulatory motif required for high-level gene expression in macrophages. Genomics. 2004, 84: 1030-1040. 10.1016/j.ygeno.2004.08.016.PubMedView ArticleGoogle Scholar
- Kawaji H, Kasukawa T, Fukuda S, Katayama S, Kai C, Kawai J, Carninci P, Hayashizaki Y: CAGE Basic/Analysis Databases: the CAGE resource for comprehensive promoter analysis. Nucleic Acids Res. 2006, 34 (Database issue): D632-D636. 10.1093/nar/gkj034.PubMedPubMed CentralView ArticleGoogle Scholar
- The CAGE basic viewer. [http://fantom31p.gsc.riken.jp/cage/mm5/]
- Vlieghe D, Sandelin A, De Bleser PJ, Vleminckx K, Wasserman WW, van Roy F, Lenhard B: A new generation of JASPAR, the open-access repository for transcription factor binding site profiles. Nucleic Acids Res. 2006, 34 (Database issue): D95-D97. 10.1093/nar/gkj115.PubMedPubMed CentralView ArticleGoogle Scholar
- Lenhard B, Wasserman WW: TFBS: computational framework for transcription factor binding site analysis. Bioinformatics. 2002, 18: 1135-1136. 10.1093/bioinformatics/18.8.1135.PubMedView ArticleGoogle Scholar
- Cover TM, Thomas JA: Elements of Information Theory. 1991, New York: WileyView ArticleGoogle Scholar
- UniGene ftp repository. [ftp://ftp.ncbi.nih.gov/repository/UniGene/]
- GNF SymAtlas. [http://symatlas.gnf.org/SymAtlas/]
- Ihaka R, Gentleman R: R: a language for data analysis and graphics. J Comput Graph Stat. 1996, 5: 299-314. 10.2307/1390807.Google Scholar
This article is published under license to BioMed Central Ltd. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.