Selection of PS-specific motifs

The motif-prediction process per set of promoters identified 142 flower-specific motifs, 183 seed-specific motifs, 171 root-specific motifs, 142 shoot-specific motifs and 141 whole plant-specific motifs, respectively (see table 1). To remove redundant motifs, each position frequency matrix was converted to a k- mer frequency vector that was then used to build a distance matrix by the Pearson Correlation distance. This matrix was used to cluster each group of PS-specific motifs by average-linkage hierarchical method. The optimal number of clusters per PS was 6, 3, 5, 4 and 2 for flower, seed, root, shoot and whole plant, respectively (see table 1). Hereafter the whole plant will be referred to as a PS for simplicity. The group specificity score (measure of how well a motif targets the promoter regions where it was found) [15] of each motif was computed and motifs with the smallest score per cluster were chosen for further analysis. The selected motifs were further compared with plant cis- acting regulatory elements in the PLACE database [13]. Motifs with p- values less than 0.001 were regarded as known motifs, otherwise, novel ones. In order to restrict as much as possible our motif comparison, we chose a strict p- value equal to that successfully used to validate the motif comparison algorithm TOMTOM (see additional data file in [16]). As a result, motif Rt_1 (see Figure 1) matched to ACIIPVPAL2 (motif known for playing a key role in vascular tissue whose primary component "xylem" is usually located close to the interior of roots), motif Sd_1 (see Figure 2) matched to ACGTSEED3 ("ACGT motif" related to seed expression) and motif Plt_1 (see Figure 3) matched to INTRONLOWER (motif involved in "3' intron-exon splice junctions" in the plant). On the contrary, flower-specific motifs Flw_1, Flw_2, Flw_3 and Flw_5 (see Figure 4), root-specific motifs Rt_2 and Rt_4 (see Figure 1), seed-specific motif Sd_2 (see Figure 2) and shoot-specific motif Sht_2 (see Figure 5) did not match significantly to any known cis- acting regulatory element in the PLACE database, thus representing potentially new regulatory motifs in plants. We also compared our predicted motifs with others previously reported in A. thaliana [12]. As a result, motif Plt_2 (see Figure 3) matched to Motif_8 (see Figure 1 in [12]), motif Rt_3 (see Figure 1) matched to Motif_3 (see Figure 1 in [12]) and motif Sd_1 (see Figure 2) matched to Motif_11 (see Figure 1 in [12]) with p- values less than 0.001. In addition, we compared our 8 novel motifs to those stored in JASPAR database [17] and found that all the compared plant motifs matched significantly to motifs in other organisms (see table 2).

Models	Gene Sets		Motif Predictions		Accuracy♣ (%)	Genes predicted genome-wide♠
	motif-prediction†	model-build‡	overall*	over-represented§
Flower	55	83	142	6	75.8	49/ 63
Seed	59	88	183	3	69.0	134/ 165
Root	64	95	171	5	65.2	34/ 48
Shoot	62	92	142	4	60.2	51/ 95
Whole Plant	58	87	141	2	64.1	76/ 86

Notes

^† and ^‡ columns indicate the number of genes in the "motif-prediction" and "model-build" sets

^* column shows the number of motifs predicted by the three motif-discovery programs

^§ column indicates the amount of over-represented motifs per model

^♣ column shows the accuracy achieved by the corresponding SVMs

^♠ column depicts the amount of genes predicted genome-wide

Classification of PS-specific promoters

Each group of training promoter regions was scanned for TFBSs of its PS-specific motifs on both strands and matrices (referred to as TRAINING MATRIX) composed of 12-component, 6-component, 10-component, 8-component and 4-component vectors characterizing the promoter regions in flower, seed, root, shoot and the whole plant were created. For each "TRAINING MATRIX", another matrix composed of background promoter regions, which were not included in both the "model-build set" and the "motif-prediction set" (see "Final promoter sets" in Methods section) of a PS was further formed. After training the support vector machine (SVM) with the corresponding matrices, each remaining single-promoter set was used for evaluating its performance (see table 1 and Additional File 1). Considering a leave one-out cross-validation approach, the SVM of the flower model achieved the highest accuracy of 75.8%. In addition, the SVMs of seed, root and plant models reached similar accuracies of 69.0%, 65.2% and 64.1%, whereas that of the shoot model achieved the lowest accuracy of 60.2%.

Creation of novel motif patterns

Flower-pattern

The pattern for promoters of genes expressed in flower comprises four motifs (see Figure 6a and Additional File 2). It was observed that motif Flw_5 has a strong tendency to be present throughout the promoter region on both strands, whereas motifs Flw_3 and Flw_4 have a tendency to be found on both strands at the region 0 to -100 near the translation start site (TLS). The presence of Flw_3 and Flw_4 at the core promoter region on both strands could possibly facilitate a stronger binding of the transcriptional machinery. On the other hand, motif Flw_2 has a tendency to be at the region -100 to -150 on both strands. Figure 6a (see Additional File 2) shows the promoter region of genes expressed in petal differentiation identified by our method. In addition, motifs Flw_1 and Flw_6 were present in less than 47% of promoters, so it is possible that their TFs do not act independently at specific distances from the TLS but their role in transcription might be related to the presence of other TFs with which they act in cooperation. Motif Flw_2 is, on the other hand, present on minus strand at the region 0 to -100 in 57.4% of promoters, whereas on both strands at the region -150 to -200 in 44.4% of promoters.

Seed-pattern

The pattern for promoters of genes expressed in seed combines all the motifs over-represented in this PS (see Figure 6b and Additional File 3). Motif Sd_2 shows a tendency to appear on plus strand at the region -50 to -100, but on both strands at the region 0 to -50. The presence of motif Sd_3 is restricted to the region -50 to -100 on both strands, whereas motif Sd_1 tends to appear on minus strand at the region 0 to -100. Figure 6b (see Additional File 3) shows the promoter region of genes expressed in synergid cells. In addition, motif Sd_1 is sparsely present (less than 40% of promoters) on both strands at the region -150 to -200. Motif Sd_2, on the other hand, is also poorly represented (less than 35% of promoters) on both strands at the region -150 to -200.

Root-pattern

The pattern for promoters of genes expressed in root combines the presence of four motifs (see Figure 6c and Additional File 4). Motif Rt_5 shows a strong tendency to be on both strands throughout the promoter region. Motifs Rt_3 and Rt_4 tend to appear at the region -100 to -150 on both strands and motif Rt_3 that significantly matched to TATABOX4 has a tendency to be bound at about the same distance reported for a TATA box. Figure 6c (see Additional File 4) shows the promoter region of genes expressed in root hair. Since motifs Rt_1 and Rt_2 are poorly present (less than 40% of promoters) at the region 0 to -50 on both strands, the TFs of both motifs might be somehow linked. The TF binding to motif Rt_5 seems to have an important role within the core promoter region, whereas the TFs of motifs Rt_3 and Rt_5 could be cooperating at specific distances from each other on both strands at the region -50 to -150.

Shoot-pattern

The pattern for promoters of genes expressed in shoot combines the presence of three motifs (see Figure 6d and Additional File 5). Motif Sht_3 tends to appear throughout the promoter region on both strands, whereas motifs Sht_1 and Sht_4 show a tendency to be found at the region -50 to -200 on minus strand. Figure 6d (see Additional File 5) shows the promoter region of genes expressed in trichome. The fact that motifs Sht_1 and Sht_4 tend to be on the same strand at specific distances from the TLS may suggest not only a presence of their TFs at these specific positions but also at precise distances between them.

Whole plant-pattern

The pattern for promoters of housekeeping genes expressed in the whole plant comprises the presence of only two motifs (see Figure 6e and Additional File 6). Motif Plt_1 tends to appear throughout the promoter region on both strands, whereas motif Plt_2 has a tendency to be found at the region 0 to -100 on minus strand. Figure 6e (see Additional File 6) shows the promoter region of plant housekeeping genes. Surprisingly, motif Plt_2 is poorly present (less than 8% of promoters) at the region -150 to -200 on plus strand, while its presence is more clearly visible near the core promoter region. It is possible that more than two TFs might be involved in the transcription of genes expressed in the whole plant, but our method of obtaining over-represented motifs might have ruled out their motifs.

Genome-wide prediction of co-expressed genes

We used the above described motif patterns: "flower-pattern", "seed-pattern", "root-pattern", "shoot-pattern" and "whole plant-pattern" to search for genes expressed in each PS within the entire A. thaliana genome. We ruled out all the A. thaliana's genes whose promoter regions were more than 60% similar. As a result, the initial set of 22 591 genes was reduced to 19 212 genes. We thus identified 63, 165, 48, 95 and 86 genes whose promoters satisfied the motif patterns. In order to illustrate the validity of our predictions we checked the plant ontology terms (POTs) per group of predicted genes. We found 49 (77.8%) out of 63 genes expressed in petal differentiation and expansion stage, 134 (81.2%) out of 165 genes expressed in synergid cells, 34 (70.8%) out of 48 genes expressed in root hair, 51 (53.7%) out of 95 genes expressed in trichome and 76 (88.4%) out of 86 genes with housekeeping function (see table 1). The poor accuracy of trichome could be due to similar promoter structures between genes expressed in "shoot" and those expressed in the other four PSs. Such similarity could impede the SVM from correctly differentiating the promoters of trichome-expressed genes.

Initial database

We worked with the version TAIR10 (ftp://ftp.arabidopsis.org/home/tair/Sequences/whole_chromosomes) of the A. thaliana genome. In addition, an A. thaliana trans-factor and cis- element prediction database (ATTED-II) [14] comprising information of co-expressed genes deduced from microarray data was used. ATTED-II contains the expression of 22 591 genes in different experimental series.

Since we are interested in PS-specific genes, five distinct datasets composed of the normalized expression of 22 591 genes from 81, 27, 21, 27 and 9 microarrays based on annotation of flower, seed, root, shoot and whole plant were formed. Each gene set was further used to identify co-expressed genes and thus form reduced gene groups per plant structure (PS). The standard deviation of a gene's average expression values through all the PS-specific sets was computed. It was multiplied by a threshold (number manually chosen to get sets of over a hundred genes) and the resulting product was compared with the difference between the two greatest average expressions. As a result, the target gene is assigned to the PS-specific set in which its average expression was greatest. Thus, sets of 138, 147, 159, 154 and 145 genes expressed in flower, seed, root, shoot and in the whole plant were obtained with thresholds of 2.05, 2.35, 2.36, 0.80 and 0.75, respectively.

Final promoter sets

Each reduced set composed of genes co-expressed in flower, seed, root, shoot and whole plant was split into two distinct subsets: (1) model-build set and (2) motif-prediction set which are randomly composed of 60% and 40% of genes in the corresponding original set. The model-build set was used to differentiate promoters containing a precise combination of motifs from background genomic promoters and thus build novel patterns of sets of motifs based on the correctly classified PS-specific promoters. The motif-prediction set was, on the other hand, employed to search for de novo motifs. The promoter regions stretching 50bp, 100bp, 150bp and 200bp upstream from the translation start site (TLS) [18] were considered and promoters of genes in the motif-prediction set were grouped using these distances. As a result, four different sets composed of promoters 50bp, 100bp, 150bp and 200bp long were created. For each PS-specific promoter set, an additional one (non-PS specific) composed of background genomic promoters other than those of genes in the subsets: motif-prediction set and model-build set was formed.

Identification and comparison of motifs

We used three different motif-discovery programs: Seeder [19], Weeder [20] and MEME [21]. For Seeder [19], motifs 6bp, 8bp, 10bp and 12bp long with a seed length of 7 were predicted. Several runs were done regarding both strands. Weeder [20] was also run on both strands and the following motif lengths: 6bp with 1 mutation, 8bp with 2 and 3 mutations, 10bp with 3 and 4 mutations and 12bp with 4 mutations were searched. For MEME [21], motifs whose length was between 6bp and 12bp, and with any number of repetitions on both strands, were predicted. In order to remove redundant motifs, the position frequency matrix of each motif was converted into a k- mer frequency vector [22]. A distance matrix was then built by using the Pearson Correlation distance and PS-specific motifs were clustered according to their similarity by average-linkage hierarchical method. The optimal number of clusters was 6, 3, 5, 4 and 2. Additionally, each motif's group specificity score (measure of how well a motif targets the promoter regions where it was found) [15] per set of predicted motifs in flower, seed, root, shoot and whole plant was computed and the motif with the smallest score of every cluster was considered for further analysis. The chosen motifs were compared with plant cis- acting regulatory elements stored in the PLACE database [13] and motifs matching with p- values higher than 0.001 were regarded novel motifs.

Characterization of promoter regions

where x represents the distance from the TLS of a TFBS and n stands for the number of TFBSs on the same strand.

The promoter regions were characterized by different-size vectors depending on how many motifs are over-represented in the model under analysis. For instance, 6 motifs were chosen in "flower" (6 average distances in each strand for 6 motifs) hence the promoters of genes co-expressed in "flower" will be characterized by a 12-component vector. Furthermore, the distances were divided by the promoter's length (200bp) for normalization and the average distance of an absent motif on a specific strand was regarded to be zero. As a result, a matrix referred to as "TRAINING MATRIX" that characterizes the training promoter regions was prepared per promoter set.

Training of a support vector machine

A support vector machine (SVM) [23] is a supervised-learning algorithm able to predict the class of a new instance (unknown category) once a set of objects that belong to two possible categories is given. This algorithm seeks a hyperplane that optimally separates instances of either class with a maximal distance (margen) from them. Each model-build set composed of PS-specific promoters was randomly split into single-promoter groups. By following the leave one-out cross-validation technique, each single-promoter set was employed for testing the model's performance, whereas the remaining groups were used for training the SVM. Accordingly, the number of feature vectors in the "TRAINING MATRIX" varies depending on how many promoters are used for training. The SVM's performance was thus evaluated using each remaining single-promoter set. Moreover, a testing promoter region is also characterized as previously explained in the section "Characterization of promoter regions".

Creation of motif patterns using true positives

To decipher a similar promoter architecture for genes co-expressed in each PS and in the whole plant, we picked out the promoters (labeled as true positives) of genes expressed in each specific PS because they seem to share a similar architecture. The incorrectly classified promoters, on the other hand, do not seem to contain an alike architecture of interest to us. We thus attained five separate groups composed of 56, 55, 70, 60 and 63 promoters of genes expressed in flower, seed, root, shoot and whole plant, respectively. We regarded four distinct regions: [0, -50], [-50, -100], [-100, -150] and [-150, -200] which cover the entire promoter region and calculated each PS-specific motif's distribution through both strands per promoter group. Those motifs present in more than 60% of promoters in flower, seed, root and shoot sets as well as 50% of promoters in whole plant set were regarded for further analysis. As a result, we formed five specific patterns of sets of motifs to decipher the promoter architecture of those co-expressed genes.

Genome-wide prediction of genes co-expressed in four PSs and the entire plant

The promoter regions with more than 60% of similarity were removed from the initial set of 22 591 genes, thus resulting a final group of 19 212 genes. Each identified motif pattern was used to predict genes expressed within one of the four PSs and the entire plant. Although genes whose promoters were employed to train each model were not ruled out, no overlapping was detected after the genome-wide prediction was carried out. Predictions were tested for accuracy using their plant ontology annotations for cellular location.