Sequence data for A. thaliana were downloaded from TAIR 1617, annotation release version 6.0, for P. trichocarpa they were downloaded from JGI/DOE 18, annotation release version 1.1, for O. sativa from TIGR 19, annotation release version 4.0, for C. reinhardtii from JGI/DOE 13, annotation release version 3.1, and for O. tauri from the University of Ghent 20, annotation release version August 2006.

Transcription factors can be identified and grouped into different families according to their domain architecture, mainly taking into account their DNA-binding domains, as described by Riechmann et al. 21 for A. thaliana. We have extended this approach by including new TF families and applied it in a systematic manner to other plant species.

Therefore, in a first step, we identified – using current literature – the list of all domains, which are known to occur in TFs and that are generally employed to classify proteins as transcriptional regulators. The list was established from available PFAM profile Hidden Markov Models (HMMs) (v20.0, 22), additionally we generated new models for further TF families, as indicated below.

To group TF proteins into families, we identified – based on previously published data – those domains, or in some cases domain combinations, that were specific for each family ('Literature survey' in Fig. 1). Then, we established a set of rules for each TF family. The rules can be depicted as a bipartite graph with two types of nodes and two types of edges (Fig. 2).

One set of nodes (blue squares) represents protein families (i.e. transcription factors, solid color, or other transcriptional regulators, shaded) and the other set of nodes (yellow circles) represents protein domains. The edges indicate the connections between protein domains and families. A continuous edge represents a required relationship, i.e. the indicated domain must be present in a protein to be assigned to the respective TF family. A discontinuous edge represents a forbidden relationship, i.e. the definition of such a family excludes the presence of the given domain. Rules were implemented in a PERL script as "IF . . . THEN" statements ('Classifier' in Fig. 1).

The general pipeline we have developed for the identification and classification of TFs is shown in Fig. 1. Typically, the process starts with retrieving the complete set of predicted proteins for a given species, followed by a profile-HMM search with all available PFAM HMMs (v20.0, 22) and the models that we have generated for further TF families. The search is carried out using the software package HMMER (v2.3.2, 23). All significant HMM hits are kept. For the PFAM models, only those hits with a bit-score larger than the gathering score reported for the HMM were considered significant. For our own HMMs, hits with an e-value smaller than 10^-3 and a bit-score threshold that differed for each HMM were considered significant. From this set of significant HMM hits, we discarded all proteins that contained domains having DNA-related activity but not generally regarded as being parts of transcriptional regulators (such as e.g. transposase-related domains). Thereby, we eliminated potential false positives right at the beginning. Finally, we applied the PERL script implementing the set of established rules for the identification and classification of TFs on the remaining set of proteins ('Classifier' in Fig. 1). The script produces as output a list of proteins that belong to the different classes of transcriptional regulators and their classification into the identified families.

For 31 out of 68 families the presence of a single domain was sufficient to assign membership (two out of the 31 families belong to the category of other transcriptional regulators). The remaining families were characterized by combinations of different domains. In this way we were able to classify transcription factors into 58 families plus 10 families for other types of transcriptional regulators, such as chromatin remodeling factors.

Table 1 summarizes the total number of TFs per species identified through the procedure outlined above. We detected 7597 different proteins classified as transcription factors or other transcriptional regulators in the five species analyzed. It is not surprising that the number of TFs generally increases with the number of genes in the genome (e.g. 24). On average there are 4.2 ± 2.5 TFs per 100 genes. The INPARANOID software implements a variation of the best-reciprocal-BLAST-hits method to search for orthologs between pairs of species 25. In finding functionally equivalent orthologous proteins INPARANOID has been shown to be the best ortholog identification method 26. We used INPARANOID to detect orthologs between the analyzed species in a pairwise manner, starting from the complete sets of predicted proteins in each species. The predicted orthologous relationships were used to create cross-references between the species-centered databases.

For the families Alfin-like, CCAAT-Dr1, CCAAT-HAP3, CCAAT-HAP5, DBP, G2-like, GRF, HRT, LUG, NOZZLE, SAP, Trihelix, ULT and VOZ no appropriated models were found in the PFAM (v20.0) database. Consequently we created our own profile-HMMs based on either published multiple sequence alignments, or on alignments we created based on outputs of PSI-BLAST searches run against the NCBI protein database. The alignments used to build the HMMs are available through our web interfaces.

Data of the different TF families are stored in five MySQL relational databases, one for each species, and in a further, global database for PlantTFDB. To uniformly structure the databases two different schemes were implemented (Fig. 3). The first scheme (Fig. 3A) was applied for each of the five independent species-specific databases. The second scheme (Fig. 3B) was implemented for PlantTFDB, which was generated as an entry site to allow access to the species-specific databases.

The basic information in each species-specific database is structured in two sets of tables. One set (right side of the TF table) contains in several tables the information about the TF family: literature references, family description and domains relevant for their classification. The field relating the information in these tables is the family_id. The second set (left side of TF table) contains five tables with the information related to the TFs themselves: sequences, domains present, domain alignments, expressed sequence tags (ESTs), orthologs. The main field here is the cds_id that unequivocally identifies every TF. One additional table, the TF table relates the two sets of tables. This table has both keys, i.e., cds_id and family_id, and contains the information about the classification of the transcription factors into families. The PlantTFDB consists of a single table with the following fields: coding sequence identifier, locus identifier, transcription factor family, md5sum of the protein sequence, description of the protein sequence, species name and TF family. The field md5sum_pep contains the md5sum of the protein sequence, which is a sequence of 32 hexadecimal digits that identifies unequivocally each protein sequence in the database.

A web resource with a uniform look-and-feel was developed in PHP (i) for each of the species studied, and (ii) for the PlantTFDB. We have taken care to follow W3 standards regarding HTML v4.01 and CSS v2.1 to assure browser interoperability as much as possible. Data can be downloaded from the databases as plain text files (Fig. 4).

The information provided in the species-specific web databases is linked through the gene identifiers or domain names to different external resources, when available and appropriate: TAIR 17, TIGR's rice genome annotation 19, JGI/DOE's poplar genome 18, and C. reinhardtii genome annotation 13, University of Ghent's O. tauri genome annotation 20, AthaMap 27, PlantGDB 28, Gramene 29, INPARANOID 30, SIMAP 31, and PFAM 22. Additional external links to other databases and computational tools will continually be included.

To evaluate the confidence in our lists of putatively complete sets of transcription factors, we decided to compare our predictions to published data sets on detailed phylogenetic single-family analyses in A. thaliana. In this way the published analyses were taken as the gold standard. We measured the sensitivity and the positive predicive value (PPV) of our approach- in a similar fashion as done by Iida et al. 6 (The terminus 'specificity' used by Iida et al. 6 is in fact the PPV, see 3233).

The sensitivity is defined as:

S e n s i t i v i t y = T P T P + F N , MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqWGtbWucqWGLbqzcqWGUbGBcqWGZbWCcqWGPbqAcqWG0baDcqWGPbqAcqWG2bGDcqWGPbqAcqWG0baDcqWG5bqEcqGH9aqpdaWcaaqaaiabdsfaujabdcfaqbqaaiabdsfaujabdcfaqjabgUcaRiabdAeagjabd6eaobaacqGGSaalaaa@45AB@

where, TP is the number of true positives, i.e. the number of TFs listed in our database that are also found in the gold standard, and TP + FN, is the number of true positives plus the number of false negatives, i.e. TP + FN is equivalent to the total number of TFs in the gold standard.

The PPV is defined as:

P P V = T P T P + F P , MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqWGqbaucqWGqbaucqWGwbGvcqGH9aqpdaWcaaqaaiabdsfaujabdcfaqbqaaiabdsfaujabdcfaqjabgUcaRiabdAeagjabdcfaqbaacqGGSaalaaa@39FD@

with the same notation as before, and FP being the number of false positives. Thus, TP + FP is equivalent to the total number of TFs listed in our database.

According to these definitions, the sensitivity gives an idea of the probability not to miss a true TF: a high sensitivity implies a low number of false negatives. The PPV, in contrast, gives an idea of the goodness of our method at only reporting true TFs: a high PPV implies a low number of false positives. The results of this evaluation are shown in Table 2. For 10 out of 12 tested TF families we obtained sensitivity and PPV values larger than 0.90 for both measurements (bold face in Table 2). Therefore the numbers of false negatives and false positives, respectively, are very low. Thus, the agreement with published results is still acceptable. For the remaining two families the agreement is still reasonable since both values are larger than 0.80, however at least one of them is smaller than 0.90.

The computational identification and classification of TFs is a very dynamic process that relies on the available computational models and tools, which in turn rely on the accumulated biological knowledge. This fact is reflected by the calculated Sensitivity and PPV values. As more experimental data become available over time, further improvements in HMMs are expected helping to minimize further the existing gaps between the gold standards and the reported data in the database.

The number of sequenced and annotated plant genomes is rapidly increasing. The computational pipeline described in this article will be applied to new plant genomes as soon as they become available and the new information will be added to future releases of PlnTFDB@uni-potsdam.de. Upcoming versions of the database will also include additional structural data about the domains employed for the identification and classification of TFs, and detailed information about the hierarchical family classification of DNA-binding domains 43738.

We are currently extending the TF discovery pipeline towards large EST collections. The next release of PlnTFDB@uni-potsdam.de will include such information and will classify TFs from plant species whose genomes have not yet been sequenced but for which large EST collections are available.