SoyDB contains the annotations of 5,671 putative transcription factors. These proteins were classified into 64 families (for details see the section "Transcription Factor Family Prediction Using SAM Hidden Markov Models"). Figure 1 illustrates the architecture of the SoyDB website. Users can access the main components from the home page: full-text search, PSI-BLAST sequence search, family classification by hidden Markov model, family browsing, protein browsing, family information, protein information, and FTP site.

The soybean genome sequences and gene model predictions used in this study were acquired from the publicly available database Phytozome 5 . These sequences were generated by the preliminary GenomeScan 40 , FgenesH 41 , and PASA 42 gene annotations based on the Gm1.01 version of soybean assembly data 4 .

We used the standalone versions of InterProScan 21 to search all the soybean protein sequences against 11 databases integrated in InerPro 38 . These databases and their corresponding scanning methods include: PROSITE (pfscan) 39 , PRINTS (FingerPRINTScan) 43 , Pfam (HMMPfam) 27 , ProDom (ProDomBlast3i) 44 , SMART (HMMSmart) 34 , TIGRFAMs (HMMTigr) 45 , PIR SuperFamily (HMMPIR) 46 , SUPERFAMILY (superfamily) 47 , Gene3D (gene3d) 48 , PANTHER (HMMPanther) 49 , and HAMAP (pfscan) 50 . InterProScan systematically searches each of these databases using their corresponding scanning methods to find domains. The proteins predicted to contain TF related domain(s) were considered as putative transcription factors. Using the Plant Transcription Factor Database (PlnTFDB) 51 and the classification of Medicago truncatula TF genes 52 as references, we manually curated the list of putative transcription factors and eliminated any mistakenly identified sequences. In this way, we identified 5,671 putative TF sequences.

The transcription factors of Arabidopsis thaliana have been well studied and classified into 64 families 33 . This provides a model for us to classify soybean transcription factors. We used MUSCLE 53 to generate a multiple sequence alignment for each Arabidopsis thaliana TF family. The multiple sequence alignment was then input into SAM 3.5 41 to build a hidden Markov model (HMM) for each family. Every soybean TF sequence was aligned with each of the 64 HMMs, which outputs an e-value. This e-value can be considered as a fitness score between a TF sequence and a hidden Markov model: lower e-value indicates better fitness. Finally, a transcription factor was classified into the family whose HMM yields the lowest e-value.

The standalone versions of several bioinformatics tools were locally installed and executed to generate annotations for soybean transcription factors. An accurate protein structure prediction tool MULTICOM 54 was used to predict the tertiary structure of each transcription factor when homologous template proteins could be found in PDB. According to the official evaluations of the 8^th community-wide Critical Assessment of Techniques for Protein Structure Prediction (CASP8) 55 , MULTICOM was able to predict high-accuracy tertiary structures with an average GDT-TS score 56 0.87 if suitable templates can be found. GDT-TS score ranges from 0 to 1 measuring the similarities between the predicted and experimental structures, whereas 1 indicates completely the same and 0 completely different. Figure 2 illustrates the predicted tertiary structure of a transcription factor in SoyDB with ID GM00002, and the electrostatic polarization of the predicted structure. The blue area in the electrostatic polarization shows residues positively charged, which is found to be highly identical to the green area in Figure 2(b), which is the putative DNA-binding sites identified by a pair-wise alignment between GM00002 and its template protein 1WID (Figure 2(d)). Since it has been studied and found that the DNA-binding area is positively charged if analyzed by electrostatic potentials 57 , the highly identical area in Figure 2(a) and 2(b) strongly confirms that the predicted structure has the electrostatic properties of a transcription factor. This further confirms the qualities of MULTICOM predictions and the correctness of the predicted binding sites derived from the homology alignment. In SoyDB, a predicted tertiary structure is visualized by Jmol 58 . In order to clearly visualize the tertiary structure of the DNA-binding region, only the segments containing homologous DNA binding domains are visualized by Jmol. Users can view a TF structure from various perspectives in a three-dimensional way and perform many operations including selecting and highlighting interested regions, changing view styles and colors, and measuring atom distances and angles by right clicking on the Jmol console and selecting corresponding menus. Detailed instructions about Jmol menus can be found at Jmol website 58 . During the structure prediction process, MULTICOM generates the sequence alignments between the transcription factor and its homologous templates using PSI-BLAST. These sequence alignments can be used to predict the binding sites of a transcription factor based on the experimentally determined binding sites of its template as shown in Figure 2.

A predicted structure was parsed into domains by Protein Domain Parser (PDP) 59 . Since some transcription factors did not have homologous templates found in PDB, DOMAC 60 , an accurate ab initio domain prediction tool, was also used to predict domains for each transcription factor.

The protein sequences in the same family were aligned by MUSCLE 53 and visualized by WebLogo 61 . A consensus sequence was derived from the multiple sequence alignment by selecting the most frequently appeared amino acid at each position. The multiple sequence alignments were also used to identify the conserved signatures (likely the DNA binding domains) for each family.

All of the bioinformatics tools incorporated to construct SoyDB can be used to automatically annotate other species in the future.

In order to annotate the functions of soybean transcription factors, each TF protein sequence was searched against the soybean TF database PlantTFDB, NCBI known EST sequences, and other general protein databases by PSI-BLAST or TBLASTN. The external protein databases include Swiss-Prot 32 , TAIR 33 , RefSeq 24 , SMART 34 , Pfam 27 , KEGG 35 , SPRINTS 36 , EMBL 37 , InterPro 38 , PROSITE 39 , and Gene Ontology 20 . Web links to these databases were created when the same transcription factor or its homologous proteins were found in them; and for each database or EST dataset only the PSI-BLAST or PBLASTN hit with the smallest e-values was listed in SoyDB. To search the known EST sequences, PSI-BLAST was first used to build a position-specific score matrix for each transcription factor. TBLASTN was then used to search each protein sequence against three known EST datasets: EST human, EST mouse, and EST others. GenBank 62 web page of each EST hit was linked to SoyDB website. The gene expression of a subset of TF genes (about 1,000 TF genes) was recently published 63 . Transcription profile of all soybean TF genes in various conditions is under investigation.

These external links greatly expand the annotation scope of SoyDB providing related knowledge from various perspectives. SoyDB provides a systematic view of a transcription factor -- from the features of the protein itself, to the biological pathway it locates in. The links to the external databases and datasets can be updated by a re-run of PSIBLAST and TBLASTN. Currently, these links are scheduled to be updated once every six months. This time interval can be changed if necessary.

The programs used to automatically annotate proteins were written in PERL. The relational database was built on MySQL with database schemas automatically generated by programs written in PERL. The website was implemented in PHP. The database and web site were automatically constructed by computer programs with little human intervention.

This component contains the complete annotations for each transcription factor, including protein ID, protein name and description, tools used for TF identification, family ID, family name and description, amino acid sequence, homology domain prediction, ab initio domain prediction, PDB homologous templates, and predicted tertiary structure. This component can be reached by clicking the sequence ID, such as GM00001, or the Phytozome protein name, such as Glyma01g11670.1, at the "Protein Browsing" webpage (for details see the following "Protein Browsing" section). Figure 3 illustrates the protein information page. The sequence ID and family ID, such as GM00001 and GMF0001, are internal indices used by the SoyDB, and the sequence name is the standard soybean TF name used by the soybean genome database Phytozome 5 . We noticed the trend of unifying annotation formats within the soybean community. Therefore, the commonly used TF ID format, such as PTGm00009.1, is also compatible in SoyDB. Details are described in the "Full-Text Search" section below.

This component contains the complete annotations for each TF family, including family ID, family name and description, number of sequences within the family, consensus sequence, consensus signatures (likely the DNA binding regions), web logo of the signature profile, and multiple sequence alignment of the protein sequences within the family. Figure 4 demonstrates a family information web page. This component can be reached from the "Family Browsing" web page.

The transcription factors within a family are listed in the order of sequence IDs. The list contains the thumbnail of tertiary structure, protein ID and name, family ID, and family name of each transcription factor (Figure 5). Users can further view the complete annotations by clicking its sequence ID or the Phytozome protein name. This component can be reached by clicking the number of sequences in the "Family Browsing" or the "Family Information" web page.

A user can browse SoyDB from TF family perspective. The 64 TF families are listed in the order of family IDs. The family ID, family name, and the number of transcription factors within each family are listed. By clicking the family ID or name, users can view the complete annotations for a family, or further browse the sequences within a family by clicking the number of sequences. This component can be reached by clicking "Browse Database" in both the top and bottom menu bars from any SoyDB web pages. Additional file 1 (Figure S1) illustrates the web page showing a TF family list.

This component allows users to search the entire SoyDB database by a query text, such as protein name or family name. Given input keywords, SoyDB searches all the fields in the database and returns matched transcription factors with links to their annotations. Users can also search SoyDB by the TF IDs used in PlantTFDB. The search component will return the homologous soybean TFs found in SoyDB.

This component allows users to search a query sequence against every TF sequence stored in SoyDB. Users can submit a query sequence and adjust PSI-BLAST parameters from a web page as shown in Additional file 2 (Figure S2). After a PSI-BLAST search is performed, the significant hits, with links to their annotation web pages, are ranked based on the e-values generated by PSI-BLAST. Additional file 3 (Figure S3) illustrates a PSI-BLAST result web page.

This component classifies a query protein sequence into one of the 64 TF families. Additional file 4 (Figure S4) illustrates the web page for family classification. A submitted query sequence is aligned with each of the 64 hidden Markov models built based on the 64 Arabidopsis thaliana TF families. The query sequence is classified into a family whose hidden Markov model outputs the lowest e-value (correspondingly the highest alignment score or fitness score). More details about family classification can be found at the "Transcription Factor Family Prediction Using SAM Hidden Markov Models" section under "Construction and Content".

All of the information in SoyDB is available for users to download from an FTP site. For example, users can download all of the TF protein sequences in the FASTA format and the multiple sequence alignments for each family in plain text. This makes it possible for other websites to link to SoyDB by performing PSI-BLAST searches on SoyDB sequences, similarly as SoyDB links with other external databases.

In total, SoyDB has 5,671 transcription factors - 4,306 of them (75.9%) have hits found in PlantTFDB identified by PSI-BLAST with an e-value threshold 10^-3. PlantTFDB has 1,891 soybean transcription factors (based on the FASTA file downloaded from PlantTFDB FTP site), and 1,805 of them (95.5%) have hits found from SoyDB based on a PSI-BLAST search with an e-value threshold 10^-3.

The collection and analyses in SoyDB allows us to perform comparisons of TF family distributions across the plant kingdom. The large number of soybean TF genes (5,671) described in this study is likely due to the two soybean whole genome duplication events; one estimated to have occurred 40-50 million years ago (mya) and the most recent one approximately 10-15 million years ago 64 65 . By comparing the total number of genes in different organisms, it was found that the increase of plant gene number is related to multicellularity and ploidy. For example, compared to the unicellular eukaryote Chlamydomonas reinhardtii where 15,143 genes were predicted 66 , larger numbers of protein-encoding genes were reported in multicellular plant organisms, e.g., Physcomitrella patens (35,938; 67 ), Arabidopsis thaliana (32,944; TAIR 33 ), and the tetraploid Glycine max (66,153, Phytozome). We hypothesize that TF gene number also follows the same trend as land plants, which have a larger number of TF genes compared to algae. To perform comparisons of plant TF genes and their distributions across TF gene families, we mined the last updated DBD database 19 for 11 plant species (C. reinhardtii, P. patens, Oryza sativa, Zea mays, Sorghum bicolor, Lotus japonicum, Medicago truncatula, A. thaliana, Vinis vinifera, Ricinus communis, and Populus trichocarpa). These species were then compared with the soybean TF genes stored in our SoyDB database.

Our analysis showed that the unicellular C. reinhardtii has the lowest number of TF genes compared to multicellular land plants (the exceptions are L. japonicus and M. truncatula where only a partial genome sequence is available). This trend also reflects the differences of total gene number in the organisms shown in Figure 6. For example, it is interesting to note that homeobox, MYB, NAC, and WRKY TF genes in C. reinhardtii lack or have very low representations compared to the 11 other plant models (Table 1). Previous studies defined a role for homeobox 68 and WRKY genes 13 in plant development. Therefore, the occurrence of these genes only in multicellular plants may reflect their special roles in development. In addition, a close relationship between TF gene number and total gene number 69 is observed when comparing the TF gene numbers of G. max and A. thaliana with their total gene numbers (i.e., G. max encodes 66,153 protein-coding genes including 5,683 TF genes; A. thaliana encodes 32,944 protein-coding genes and 1,738 TF genes). Thus, the family distribution of soybean TF genes is similar to other land plant species, except for P. patens (e.g., AP2 represents 7% of total TF genes in soybean vs. 8-12% for other land plants; bZIP: 3% vs. 3-7%; bHLH: 7% vs. 8-11%; homeobox: 6% vs. 4-7%; MYB: 14% vs. 7-14%; NAC: 4% vs. 4-9%; WRKY: 3% vs. 4-7%; ZF-C2H2: 7% vs. 5-9%) (Figure 6 and Table 1).

Collectively, these results suggest that soybean TF genes were not lost following soybean genome duplication, and may have evolved for specialized functions in plant development or response to the environment.

In the future, we plan to link to more soybean database, such as SoyBase, and add a human expert discussion section for each transcription factor where biologists can register, log in, and make comments on any annotation items. Also, we plan to link the protein name, such as Glyma01g11670.1, listed in each protein information page to its entry in Phytozome. By doing this, SoyDB can be linked with other soybean genome annotations. Furthermore, we may identify the binding regions on the soybean DNA sequences, which can further help biologists target the regulated regions on soybean genome.