Typically, DNA elements recognized by a protein (complex) is within the range of 5 to 15 bp long 34. Within this range, we decided to analyze localization patterns of hexamer and octamer sequences. Our results suggest that sequences longer than 9 bps would not provide enough number of appearance to survive statistical analysis.

For each hexamer sequence, a distribution profile in relation to distance from the TSS was analyzed for Arabidopsis thaliana. Looking through all the distribution profiles, we noticed that there are quite a few patterns. Most sequences have a flat distribution profile with no special tendency (Fig. 1, GAAGAG). Sometimes the base line has a slight slope with a higher frequency toward the TSS. There are also groups with peaks, and they can be classified according to the peak position. We refer these sequences as LDSS-positive.

One example of a LDSS-positive sequence, (Fig. 1, CTCTTC) has a peak of appearance at the TSS. Its complementary sequence (Fig. 1, GAAGAG) has a distinct distribution profile, showing that its appearance is sensitive to the direction of transcription. Although hexamers with this type of distribution profile tend to have only C and T in the sequence (see later), there seems to be weak sequence preference, and not all the sequences filled with C and T show a peak-positive distribution (Fig. 1, CCTTTT is a peak-negative example).

A second example (Fig. 1, CTATAA) is a TATA box-related sequence. This has a peak around -35 bp, and the peak is very sharp. The complementary sequence showed a different pattern with no peak (Fig. 1, TTATAG).

A third example (Fig. 1, TGGGCC) has a relatively wide and low peak. Complementary sequence of this sequence shows the same peak (Fig. 1, GGCCCA). Peak position and direction-insensitivity suggest that sequences with this type of distribution profile are so-called cis-regulatory sequences involved in transcriptional regulation 34. In fact, TGGGCC in Figure 1 is reported to be necessary for meristematic expression in Arabidopsis, and mutation to TG

AA

A

As controls, a set of random genomic sequences of 1 kb length was used for the distribution analysis instead of the promoter database. When sequences with distribution patterns of peak-positive sequences were applied to this analysis, they were found to have no peaks in the random genome fragments (Fig. 1, CTCTTC/random genome, CTATAA/random genome, TGGGCC/random genome).

Beside LDSS-positive elements, there are many LDSS-negative sequences. Among them, frequently observed sequences beyond the theoretical occurrence rate (0.24 per a 1 kb region) are rich in AT and might promote promoter context, and rare sequences are rich in GC and they might disturb promoter function when located within the promoter region. Therefore, it might be possible to utilize these LDSS-negative sequences as well for evaluation of promoter context.

Figure 2A shows a close-up of a typical distribution profile of the regulator type. In order to detect peak-positive sequence, we calculated several parameters. Curve fitting with Gaussian did not give good results (data not shown), because the peak shape is not symmetrical, as seen in the figure. Through analysis of distribution profiles of all the hexamers, we noticed that all of the observed peaks were located downstream of -200 bp. This enabled a base line to be established (Base in the figure) as an average of occurrence between -1,000 and -500 bp. Then we calculated the Relative Peak Height (RPH), and Relative Peak Area (RPA) for evaluation of peak strength. Fluctuation around the base line between -1,000 and -500 bp was also evaluated (see figure legend).

Figure 2B shows the relationship between peak position and a parameter of peak strength. As shown, all the strong peaks locate downstream of -200 bp while weak peaks are scattered throughout the promoters. One important point of the figure is the continuous distribution of hexamers across the vertical axis. The continuous nature was also observed when RPH or RPA was represented in the graph on the vertical axis (data not shown). These results mean that there is no clear way to separate peaky and flat groups. In this study, we took a strategy to list sequences with strong peaks, leaving out a flat group and a group with ambiguous peaks.

Considering peak height, peak area, and fluctuation from the base line, we selected 247 sequences from all the hexamers as peak positive (Fig. 2B, black dots, Table S1 [see Additional file 1]).

The LDSS-positive hexamers identified were then classified into three major groups as mentioned above. The first group, including CTCTTC, in Figure 1, localize from -100 to -13 bp. They typically have a peak at the most downstream region of the promoter (position -13, Table 1), but peak positions distribute from -13 to -60. Most of their sequences are composed of only C and T, we refer to this group as Y Patch (Y for pyrimidine). As shown in the table, Y Patch sequences are found in the majority of Arabidopsis promoters.

The second group contains TATA box-related sequences. An example is shown as CTATAA in Figure 1. The characteristics of this group are high peak height, narrow peak width, and stringent peak position (Table 1, TATA Box). Similar to Y Patch, the TATA box group sequences are also found in the majority of Arabidopsis promoters, although promoters with the TATA Box within the peak are is about 1,000 or less for each sequence.

The third group, including TGGGCC in Figure 1, is referred to as REG, for Regulatory Element Group, in this study. The peak positions of this group locate around -80 bp, and they have a wide peak width in comparison with that of the TATA box group (Table 2). Another feature of the group is high coverage of Peak Area against total area. This means high specificity of localization within a promoter. As shown in Relative Peak Area (RPA) of the table, around 50% to 30% of a REG sequence is found in the peak area. These ratios are much higher than those of the Y Patch (25 to 10%) or TATA box (11 to 5%) groups. Compared to these, the number of promoters containing a REG sequence is smaller, consistent with the idea that each REG is not a component of the general core promoter but a specific regulator of gene expression. In fact, Table 2 contains several known cis-regulatory elements, including Element II of Arabidopsis PCNA-2 (GGCCCA, TGGGCC, and AGCCCA) 35 and G-box/ABRE (CACGTG, CGTGGC, CCACGT, and GCCACG) 36.

In addition to these three groups, there is also small number of exceptional hexamers with peak positions in the core promoter (-13 to -60). They might constitute a minor type(s) within the core promoter (Table S1, "others" [see Additional file 1]. See also Table S2 and S3 for these elements). The complete list of the extracted sequences is shown in Table S1. The table shows 103 Y Patch, 39 TATA-related, 38 REG, and 22 unclassified hexamer sequences.

Subsequently, we examined if the orientation of the hexamers is critical. The identified hexamers were tested to determine if their complementary sequences were also included or not. If the complementary sequence was also found in this positive group, the original sequence is considered as direction-insensitive, and if not, direction-sensitive. As shown in Figure 3, the downstream region from -50, that is known to be the core promoter region 1 and includes the Y Patch and TATA box groups, is occupied with direction-sensitive sequences ("uniq" in the figure), while the upstream region, containing the REG group, is rich in direction-insensitive sequences ("comp" in the figure). These findings are consistent with the established idea that the core promoter determines position and direction of transcription, and cis-elements are direction insensitive. These findings further support the idea that the Y Patch and TATA box sequences are core promoter elements and REG sequences are the cis-elements 34.

Subsequently, we analyzed the distribution of octamer sequences. The average of octamer appearance rates is 15.7-fold less than the one of hexamers, consistent with a mathematical expectation of 16-fold difference (data not shown). Because rare sequences tend to show more fluctuations by chance, statistical evaluation was more critical for octamer analysis. We prepared random distribution populations and used them for statistical evaluation of each octamer (Figure S1 [see Additional file 2]). In this study, we have set a p value of 1 × 10^-5 as a threshold. In addition, data of the complementary sequences was merged only for REG detection to increase total count of an octamer in the database. Through the octamer analyses, we have identified 350 and 418 LDSS-positive core elements (Table S2 [see Additional file 3] and S3 [see Additional file 4]), and 308 and 242 REG sequences from Arabidopsis and rice, respectively (Table S4 [see Additional file 5] and S5 [see Additional file 6]). Sum of the p values for all the extracted octamers of individual species were around 1 × 10^-3 each, so false-positive sequences by pure random distribution are not likely to be included in the lists.

For comparison of Arabidopsis and rice elements, Relative Peak Height (RPH) values of all the positive octamers in either of the two promoter databases were represented (Fig. 4A). If a sequence has the same RPH value in the Arabidopsis and rice databases, a dot appears on the diagonal line. As shown in the figure, we found that RPH values are moderately conserved between Arabidopsis and rice (Fig. 4B). The figure also indicates that a considerable number of the sequences have a large difference in the parameter between Arabidopsis and rice. Figure 4B shows Venn diagram of the number of positive octamers in Arabidopsis and rice. As shown in the figure, approximately 30 to 50% of the octamers are conserved between Arabidopsis and rice for both core groups of Y Patch and TATA box, and the REG group. Presence of all the three categories in Arabidopsis and rice, and sequential conservation as shown in the figure indicate that promoter architecture of these plant species is essentially conserved. On the other hand, divergence of the positive sequences might reflect differentiation of the corresponding trans-factors between these species.

All the LDSS-positive sequences from Arabidopsis were subjected to clustering analysis according to their distribution profiles. As expected from previous hexamer analyses, major clusters are REGs, TATA box, and Y Patch (Fig. 5). As shown in the figure, distribution profiles within each group (clusters) are quite similar, suggesting functional conservation within each group. The observed clear classification of the LDSS-positive sequences, represented in Figure 5, suggest that the local distribution is a quite useful feature in extraction of putative functional elements in the promoter.

Subsequently, we did classification of 308 Arabidopsis REGs with the aid of the promoter database. For each promoter, number of appearance for each REG was scored, and two-dimensional REG-promoter clustering was performed. This REG-promoter association has revealed that 10,334 out of 12,951 Arabidopsis promoters have at least one REG at the region of -400 to -40 bp. This high coverage (80%) is due to the long list of REG sequences.

This 2D clustering puts co-localized REGs proximal, and promoters with similar REG compositions also come close. Two promoter clusters are shown in Figure 6A and 6B. One cluster of promoters (A) are rich in GCCCA-containing REGs, and another cluster (B) have ACGT-containing REGs. GCCCA-containing REGs is the same kind as TGGGCC (Figure 1) and known to show cell cycle-dependent expression and meristematic expression (Group 1, Table 3). Interestingly, this promoter group is rich in ribosomal proteins. As shown in Figure 6A, as high as 38% (6 out of 16) of the annotated promoters are for ribosomal proteins (Fig. 6A, blue). In contrast, ribosomal promoters are not included in the ACGT-containing promoter clusters (Fig. 6B). Instead, the latter cluster is rich in photosynthesis-related genes and stress-responsive genes, both of which would show environmental responses. In fact, as many as 34 out of 38 genes in this cluster with expression data are responsive to light (Fig. 6B, green) or abiotic stress including salt, drought, and cold (Fig. 6B, red and orange), according to public microarray data 3738. Although this clustering is not so accurate as to distinguish between light and stress responses, it has been proved to classify genes with respect to gene expression with a certain range of accuracy. The results are reasonable because cis-elements for light response (G-box: CACGTG, 36) and stress response (ABRE: ACGTGTC, 39) are related sequences both of which belong to the ACGT motif for environmental responses (Group 2, Table 3). Therefore, clustering of promoters appears reasonable, although the accuracy may not be enough for pinpoint speculation of gene function.

Clustering of REGs turned out to be reliable as well, and thus useful for REG classification. According to this 2D method, overlapping REGs (e.g., C

ACGTGGA

ACGTGGA

ACGTGGA

ACGTGGA

One group has a GGCCCA core sequence that is known as Site IIa or Element II (Group 1, Table 3). Element II is necessary for cell cycle-related expression and for meristematic expression 35. Many sequences containing GGCCCA in the center of an octamer were found in REG group of both Arabidopsis and rice (Table 4). As seen in the table, this group is a good indicator of conservation.

Another group shown in the table has the bZIP protein-binding motif containing ACGT core sequence. This group mediates various environmental signals 36. Both species have this group in common, but Arabidopsis has wider variations than rice (Table 4).

Classification of Arabidopsis and rice REGs are shown in Table 3. The largest group is the Group 1, which includes Element II of the Arabidopsis PCNA-2 involved in cell cycle-related expression, as mentioned above. As shown in the table, this group is well conserved between Arabidopsis and rice and has many members for both species. There are several other REG groups, some of which are rich in only Arabidopsis and some are found from both (several examples in Table 4 and summarized in Table 3). Comparison between Arabidopsis and rice suggests both conserved and differentiated types of REGs.

The identified Arabidopsis REG sequences were referred to the PLACE database that is a collection of reported plant cis-regulatory elements 40. The comparison revealed that 155 out of 308 Arabidopsis REGs show 100 % match with at least one of the Arabidopsis PLACE entries, giving an estimation that 50 % of the REGs are of established cis-regulatory elements (Table S6 [see Additional file 7]). These results again provide strong evidence for biologically meaningful extraction of sequences by the LDSS method. From another point of view, 21 out of 48 Arabidopsis PLACE entries have been found in the REG list (Table 5). Comparison with another cis-element database, AGRIS 41, resulted in lower match than PLACE (27%) among Arabidopsis motif entries shorter than 9 bps (data not shown). These results suggest that not all of the cis-regulatory elements are detected by the LDSS strategy. One of the valuable finding of this analysis is the identification of a large number of novel REGs.

We then analyzed sequence characteristics around the TSS. In this region, the Initiator motif (Inr: YY

A

A

A

R

Our simple LDSS analysis has successfully revealed three distinct groups consisting of hundreds of short sequences. Figure 9A illustrates the architecture of plant promoters based on these findings.

Tight positioning of the TATA boxes relative to the TSS fits with the general idea that the TATA boxes determine the position of the TSS. In addition, the YR Rule of Arabidopsis would be another important determinant as well. The Y Patches locate between the TATA boxes and the TSS, but they can be upstream of the TATA boxes, considering the wide distribution profiles (Figure 5). The role of the Y Patch is not known. The above three elements are orientation-sensitive, and constituents of a core promoter. REGs appear upstream of the TATA box, and they exist in an orientation-insensitive manner. Rice promoters share the above characteristics, showing architectural conservation between dicots and monocots.

An example of an Arabidopsis promoter that has the Y Patch and TATA box is shown in Figure 9B. Octamer analysis of the promoter revealed one cluster of Group 2 REGs (Table 3), one cluster of Y Patches, one cluster of TATA box, and YR Rule. An interesting feature of the figure is the multiple hits of a locus, detecting a longer element. This demonstrates that octamer analysis can detect long functional units as clusters of octamers.

In this study, we have identified hundreds of novel sequences solely based on local distribution in the promoter region of Arabidopsis and rice. Biological information, such as microarray data, was not used at all for sequence extraction, and it becomes useful only during interpretation of the extracted sequence. This method is equally sensitive in detection of major and minor motifs in a promoter population as demonstrated by simultaneous detection of major TATA elements and minor REG elements. This feature is an advantage of the LDSS method over other methods of detection of consensus sequences among promoter populations, such as Gibbs Sampling method. We successfully applied the LDSS method to Arabidopsis and rice promoters, and of course, it is applicable to bacterial and mammalian research as well.

The observed localized distribution is a direct result of the selection pressure. While the localization is an indication of a beneficial role for the organism, the relationship between local distribution of a sequence and its functionality is indirect. Therefore, the question arises if all regulatory elements can be picked up by the LDSS strategy.

When we compared REG sequences with established cis-elements in the PLACE database, it was found that 27 out of 48 Arabidopsis PLACE entries are absent in the extracted REGs (Table 5). These results indicate that not all of the functional elements are LDSS-positive, and thus some would not be detected by this method. There are two possibilities for the presence of cis-elements that do not show local distribution. One possibility is that these elements are relatively "new" so there has not been selection pressure for a long enough period. Another possibility is that there has not been any selection pressure because of functional differences from the LDSS-positive elements. The latter idea suggests localization-insensitive classes of regulatory elements that are distinct from REGs. So called long range-regulators 4344 might be one of the classes.

Generally, any functional sequences in the genome are recognized by trans-acting factors that are DNA-binding proteins. Promoter elements and their trans-factors have a relationship of co-evolution. Therefore, differentiation of REGs in the two species would reflect a different status of the corresponding trans-factors. Functional comparison of DNA-binding proteins of Arabidopsis and rice is expected to give some answers as to why these two species have differentiated REG sequences. As for the conserved REGs, it is reasonable that cell cycle-related elements (Group 1, Table 3) comprise the most conserved group, because the cell cycle is one of the most conserved activities in organisms.

REG sequences can be extracted form mammalian promoters as well. However, our preliminary analyses suggest that the LDSS method can detect much less REGs than of plants (YYY and JO, unpublished results). This may be reflected by different promoter architecture between plants and animals.

The discovery that the Y Patch is conserved in monocots and dicots is one of the major achievements of this study. A related motif is reported by Molina and Grotewold from Arabidopsis core promoter analysis using the Gibbs-sampling method (Motif 1 with a typical sequence, TTCTTCTTC, 29). The biochemical role of Y Patch is not known, but its position, direction sensitivity, and its abundant nature strongly suggest that it is a general component of the core promoter. Our LDSS analyses suggest that human and mouse do not share this element with plants and thus this is a plant-specific core element (YYY and JO, unpublished results).

At the TSS, the Initiator (Inr) motif (Y Y

A

This rule is not an artifact by the Cap-Trapper method that is the basis of TSS mapping of this study and mammalian studies mentioned above 45, because it is applicable to human TSS determined by another method (Oligo-Cap method, 46) as well (YYY and JO, unpublished results).

A plant consensus around TSS (A/T n T/a C/t

A/c

The TFIIB-Recognition Element (BRE) is another core promoter element of animal genes. It is located just upstream of the TATA box and has a GC-rich sequence, (G/C)(G/C)(G/C)CGCC 148. Our analysis did not detect the BRE as a LDSS-positive element, although CC is preferred at the neighboring sequence of the TATA box at the upstream side in both Arabidopsis and rice promoters (Table S2 [see Additional file 3] and S3 [see Additional file 4]).

The hundreds of octamer sequences identified by the LDSS analysis can be used for promoter prediction. The presence of the TATA box is an important feature of a promoter, but there are many false-positives in the genome. For example, a TATA octamer sequence with the highest specific localization is found within the peak area 30% of times in the promoter region, meaning that 70% are found outside of the peak area. This is essentially consistent with a previous study, where more than 200,000 putative TBP-binding sites were detected from the Arabidopsis genome 27. Utilization of preferential sequence around the TATA box, and coexistence with the Y Patch and REG are expected to elevate accuracy of prediction. Although such a combinational approach is incorporated into several promoter prediction programs 13, motifs to be detected have been limited so far. Our long list of the LDSS-positive octamers is expected to serve as a thick dictionary for precise interpretation of plant genomes.

Cap-Trapper 49 is one of the most reliable methods for identification of the 5' end of mRNA and thus suitable for determination of TSS. So-called full-length (fl) cDNAs of Arabidopsis and rice were made by the Cap-Trapper method, and around ten to twenty thousand of non-redundant fl-cDNA clones for each species have been completely sequenced 5051. Therefore, we decided to use the information from the fl-cDNAs for positioning of promoters. Genome sequences of promoter regions from -1,000 to -1 bp were prepared with the aid of information of the 5' ends of fl-cDNAs of Arabidopsis 5052 and rice 51. The established Arabidopsis promoter database 5053 and a rice database with 11,370 promoters, prepared in this study, were utilized for our analysis.

Positions of rice fl-cDNA clones of rice 51 were mapped on to corresponding BAC clones according to description of "MappingData.txt" obtained from the KOME web site 54, and promoter regions from -1 kb to +200 bp relative to the TSS, that are 1.2 kbp long, were collected. BAC and fl-cDNA sequences were obtained from DDBJ. Special care was taken for 5' end of fl-cDNA sequences, and ones with less than 2 bp mismatch with the corresponding genomic sequences were used for the promoter mapping. Sequences of non-redundant 11,370 rice promoters have been prepared. For analyses of the TSS region, as shown in Figure 6, rice fl-cDNA sequences with no mismatch to the 5' end (6,209 promoters) were used. Establishment of the Arabidopsis promoter database is described elsewhere 5053. Earlier analyses with Arabidopsis hexamers have been done using the distributed database containing 15,607 promoters. This database is based on distinct TSS and allows multiple promoters belonging to a single gene. A smaller set of 12,951 promoters was re-selected from the 15,607-version so as to pick-up one promoter from one gene, and used for octamer analyses. For preparation of random genomic fragments, non-overlapping Arabidopsis BAC clones were selected by consulting a TAIR web site 55, they were successively cut into 1 kb pieces and serial numbers were given to the fragments. Sequences corresponding to 3,000 randomly chosen numbers based on the Mersenne Twister method 56 were used as random genomic fragments of 1 kb length.

The programs used in this study will be freely provided upon request for non-profit purposes. A searchable web site to obtain results in this work will be released.

Random distribution samples were generated with respect to Total Area, that is indication of total count in a promoter database. For each Total Area, 1,000 samples were prepared, and their RPA values were subjected to statistical analysis. Average and standard deviation are functions of Total Area (Figure S1 [Additional File 2]) and affected by a smoothing window. Model RPA populations of random distribution were calculated as the following equations:

REG detection (smoothing with a 21-bin (width of window), and Total Area < 2,000): log₁₀(average) = -0.1861Ln(Total Area) – 0.5329, SD = 0.17 CORE detection (smoothing with a 3-bin, and Total Area < 10,000): log₁₀(average) = -0.1784Ln(Total Area) – 0.8026, SD = 0.13

These models were utilized for estimation of p value for each octamer distribution.

Sequence analysis was achieved by a combination of home-made Perl and C⁺⁺ programs and also Excel software (Microsoft Japan, Tokyo). The first step of the analysis was the preparation of index files for each promoter with all the possible 4,096 hexamer and 65,536 octamer sequences. Information of the index files was then rearranged for each hexamer and octamer sequence, and the occurrence of the short sequences was summarized according to the promoter position. Summarized distribution data of each hexamer was then subjected to smoothing with a bin of 15 bp. Generally, smoothing with a wide bin lowers the peak height of a sharp peak, and with a narrow bin capturing a wide and low peak is not always possible. Considering these tendencies, a bin of 21 bp was used for identification of octamer REGs, and a bin of 3 bp was used for octamer core elements. Octamer REGs were extracted after merging the distribution data of the complementary sequence to increase the count of occurrence. As for extraction of octamer Core elements that is orientation-sensitive, merging was avoided. Positions of octamers and hexamers were counted from the first base of the sequence. For example, the position of a hexamer sequence that locates from -6 to -1 is expressed as -6. Positions of average values for line smoothing are indicated at the centre of the region. Therefore, positions closest to TSS vary depending on the bin length as well.

Thresholds for distribution of peaks are as follows:

Hexamer: (peak height/Base Line > 3) & (peak height/SD > 5) & (Peak Area/basal fluctuation) > 5),

Octamer Core: (p value < 10^-4) & (peak height/Base Line > 5) & (peak height/SD > 10) & (Peak Area/basal fluctuation > 6) & (peak position > -51),

Octamer REG: (p values < 10^-4) & (peak height/Base Line > 3) & (Peak Area/total area > 0.1) & (peak height/SD > 5) & (Peak Area/basal fluctuation > 6) & (peak position <-50).

Fitting the distribution data with the Gaussian curve was achieved using Igor Pro (Hulinks, Tokyo). All the LDSS-positive octamers together with above parameters can be viewed at our web site (57).

Clustering analyses were achieved with Cluster 58 and visualized with TreeView 59. For clustering of LDSS-positive elements based on distribution profiles, peak value of each profile was adjusted to 5.0. For REG-promoter clustering, number of each REG appeared at a region between -400 to -40 bp was scored for each promoter and a REG-promoter table was prepared. Among the Cluster options, the hierarchical clustering method (centroid linkage) gave the most natural results over the k-means and SOM methods.

Among the PLACE database 40, 48 entries with definition sequences of 8 bases or less and also with description containing "Arabidopsis" were subjected to REG survey.