PlantTFDB: a comprehensive plant transcription factor database

Data sources

Currently, PlantTFDB contains TFs identified in 22 species (Table 1). Genome sequences of Arabidopsis (A. thaliana), rice (Oryza sativa), poplar (Populus trichocarpa), green alga (Chlamydomonas reinhardtii) and moss (Physcomitrella patens) were downloaded from TAIR, TIGR and JGI. For the 17 species without available complete genome data, we downloaded the unique transcripts from the Plant Genome Database (PlantGDB, http://www.plantgdb.org/) (10). These plant unique transcripts (PUTs) were assembled by PlantGDB based on the mRNA and EST sequences. We applied the framefinder program in ESTate (Expressed Sequence Tag Analysis Tools Etc) package to predict the open reading frames and obtain protein sequences from these PUTs (http://www.ebi.ac.uk/~guy/estate/).

Plant TF HMM profiles

Transcription factors are always grouped as different families based on their DNA binding domains. Currently, 64 TF families have been characterized in plants (7). Among them, 48 families have hidden Markov Model (HMM) profiles in the Pfam database (v20.0) (11), while the remaining 16 families do not have available HMM profiles since they either were newly identified or only had a few members. To build the HMM profiles of these 16 families, we took their protein sequences from the previous TF databases of Arabidopsis, rice and poplar (5–7) and performed multiple sequence alignment. Then, we manually refined the alignment results and kept only the regions representing the conserved DNA binding domain. Finally, we used the hmmbuild program in the HMMER package (http://hmmer.janelia.org/, v2.3.2) to build the HMM profiles for these 16 TF families.

TF identification

We applied the hmmsearch program in HMMER to search against the protein sequences of each species to predict TFs. Based on our previous experience and manual inspection, we took E-value 0.01 as the cutoff, which was widely adopted for HMMER search.

Many TFs have more than one DNA binding domains (2). For example, the B3 domain (PF02362) was presented in either ABI3-VP1 family or RAV subfamily of the AP2 family. We assigned TFs into the ABI3-VP1 family if they only possessed the B3 domain, otherwise to the AP2 family if they had both B3 and AP2 domains (2). We developed a rules-driven program to handle such issues. Detailed rules to categorize the TFs can be found in the PlantTFDB help page (http://planttfdb.cbi.pku.edu.cn/help.php).

TF annotation

To provide comprehensive information for the identified TFs, we made extensive annotations on both the family and gene levels. For each TF family, a brief introduction and key reference were listed in the family page. BLAST search was performed against well-known public databases such as UniProt, RefSeq, EMBL and TRANSFAC. Putative functional domains were identified and annotated by InterProScan, and Gene Ontology annotations were further extracted. In addition, the expression profiles collected from UniGene EST/cDNA information were available for all putative TFs.

Putative ortholog annotation

To predict putative orthologous relationship of TFs among these species, we used the BLAST score ratio (BSR) method, which had been widely adopted by ENSEMBL (http://www.ensembl.org/) and other studies (12). An all-against-all BLASTP search with a strict cutoff E-value <1E−20 was performed, and the BSR value was calculated for each hit. After comparing results at different BSR value, we chose the BSR value ≥0.4 as the cutoff and we retrieved the top sequences in a species with the largest BSR value as the putative ortholog(s).

Protein sequences from PUTs

The PUT sequences assembled from transcripts may have insertion/deletion sites disrupting the open reading frames. We made TBLASTN against these PUTs using Arabidopsis proteins as query sequences and observed that more than half of them had frame shifts. Therefore, we used the framefinder program to obtain the protein sequences of the PUTs of 17 species.

Evaluation of self-built HMM profiles

Based on the multiple sequence alignment results of known members in Arabidopsis, rice and poplar, we built HMM profiles for DNA binding domains of 16 families, which did not have HMM profiles in the Pfam database (v20.0). We obtained the same hits from HMMER search against Arabidopsis proteins for CCAAT-HAP2 and MBF1 using the HMM profiles we built and the new HMM profiles added in Pfam (v22.0).

Evaluation of TF identification accuracy

To estimate the accuracy and reliability, we applied our pipeline to 10 well-annotated families in Arabidopsis. We measured the sensitivity and the specificity of our approach using the same approach described by Iida et al. (4) and Riano-Pachon et al. (8). Our results showed that the sensitivity and specificity of eight families were greater than 0.95, the sensitivity of two families were close to 0.90 and the specificity of one family was 0.935. This suggested that the approach we used had reasonable performance with acceptable accuracy.

Ortholog prediction

We used the widely-adopted BSR method to predict orthologs (12). To find an appropriate parameter, we made BLAST search against japonica rice to find orthologs of each Arabidopsis TF with six different BSR cutoffs (0.3, 0.33, 0.35, 0.4, 0.45 and 0.5) and compared the hit number, coverage and identity of the BLAST hits. Finally, we chose the BSR 0.4 as the cutoff of BLAST hits, which was relatively strict with an average coverage >80% and identity ∼60%. Based on these cutoffs, we chose the top sequences in a species with the largest BSR value as the putative ortholog(s).