In the post genome era we observe a continuously growing, vast amount of sequenced organisms spread over all domains of life. Besides the identification and annotation of functional sites within the emerging nucleic acid sequences, an important task in molecular genetics, biotechnology, and human medicine is to unravel the regulation of these sites. DNA-binding transcription factors (TFs) are the most important components of the cell's regulatory machinery [1]. They recognize specific operator sequences close-by the promoter regions of the controlled target genes, referred to as transcription factor binding sites (TFBSs), and thereby influence the amount of produced proteins. Although inevitable for the understanding of the cell's handling of changing environmental conditions, the wet-lab reconstruction of the resulting transcriptional regulatory networks is cost-intensive, time-consuming, and impossible to perform for any species separately [2, 3]. Even for prokaryotic model organisms, such as Escherichia coli or Corynebacterium glutamicum the monumental task of deciphering transcriptional regulatory networks for whole species is far from being complete. The current knowledge is brought together and stored in reference databases, such as RegulonDB [4] and CoryneRegNet [5]; see [6] for a more detailed analysis of such platforms.

The gathered information about substantial parts of the transcriptional regulatory apparatus is used to study conserved network structures, sensing mechanisms, and to uncover hidden architectures behind gene regulatory networks [7, 8]. In addition, specialized approaches, based on the evolutionary conservation of the responsible transcription factors and the controlled target genes, are used to transfer knowledge on gene regulatory networks between different organisms but aim to provide more general, qualitative conclusions (trends) across many species [9]. The main problem, however, is the neglect of the fact that orthologous regulators and target genes not necessarily are involved in conserved regulations. Another complicacy is the dependency on reliable homology detections. Other approaches utilize annotated transcription factor binding sites to compute mathematical models for further TFBS predictions; where the by far most widely used model for TFBSs are position weight matrices (PWMs) [10]. Here, the major intricacy lies in the comparatively low level of TFBS conservation between different organisms [11], even for essential factors such as the bacterial SOS response and DNA damage regulator LexA [12]. Hence, the consideration of PWM-based predictions apart from further evidence is not very meaningful. Moreover, there is a hidden problem with PWM calculations: The determination of the position to which a transcription factor binds is difficult and time-consuming. It is normally performed through electrophoretic mobility shift assays, DNAse footprinting, ChIP-to-chip assays, or mutations of putative TFBSs [13–15]. With all of these methods a precise identification that is accurate to one base pair is problematic. Furthermore, since TFs bind the double-stranded DNA it is a matter of interpretation which strand of the DNA sequence is annotated and stored in the database. This causes a practical problem when a motif from either strand based on approximate knowledge of its position is used for PWM construction.

In the past years, we extensively studied the transcriptional regulatory repertoire of the model organism C. glutamicum and other corynebacteria important in human medicine and biotechnology. We gathered all publicly available data, combined it with own wet-lab findings and developed the reference database and analysis platform CoryneRegNet [5, 11, 12, 16]. Here we introduce an integrative bioinformatics approach that aims for a reliable transfer of gene regulatory interactions, which combines both of the above introduced major approaches: homology detection and DNA binding site prediction. Instead of studying general trends and conserved network motifs across hundreds of organisms we are interested in high-quality predictions with just few or even no false-positives for C. diphtheriae, C. efficiens, and C. jeikeium based on evidenced observations from the model organism C. glutamicum.

The success of the approach relies on the optimal interplay of the used bioinformatics components. On a very general level, the workflow is depicted in figure 1. An integrated database system is used for data fusion of annotated nucleotide and amino acid sequences together with evidenced transcriptional regulatory relationships and corresponding sequence features. By having all required data at hand we can re-adjust inaccurately determined TF binding sequences by shifting some motifs by some positions and by assigning a strand annotation, if necessary. Sequence motif discovery tools [17] may be utilized for that purpose or publicly available special purpose tools, such as MoRAine [18]. After computing a PWM for each transcriptional regulator of the model organism from the re-annotated TFBSs, motif matching tools (such as PoSSuMsearch [19] or MATCH [20]) are used to predict binding sites in the target organism assuming that the DNA-binding motif of the regulator is sufficiently conserved. This is done for any pair of the conserved regulators and target genes. It is obvious that the TF itself has to be conserved in the target organism; but the second condition (homologous target genes) is very important as well to provide reliable, high-quality predictions since a high-scoring PWM-based sequence match alone is not very meaningful [10]. This is the only way to reduce the huge amount of false-positives without further background knowledge even for comparatively restrictive thresholds that miss most of the true-positives. An application example illustrating the problem for corynebacteria may be found in ref. [11]. The detection of conserved, orthologous proteins based on the given amino acid sequences alone was a long-standing challenge in computational biology. It emerged that clustering approaches utilizing BLAST[21]-based pairwise sequence similarity measures attack the problem comparatively well. Popular approaches are Markov clustering [22], spectral clustering [23] and graph cluster editing [24, 25].

To demonstrate our combined approach we decided to integrate MoRAine [18] for TFBS readjustment, PoSSuMsearch [19] for binding motif matching, and FORCE [25] for homology detection and applied it to corynebacterial data gained from CoryneRegNet [5] (see additional file 1). The goal was to use proven knowledge from the model organism C. glutamicum (CG) to predict reliable gene regulatory networks in C. diphtheriae (CD), which is a human pathogen, C. jeikeium (CJ) that is pathogen and extremely antibiotic resistant, and the sister organism of CG namely C. efficiens (CE), which is import for biotechnological production processes. First, the homology detection identified the common repertoire of conserved transcription factors between CG and CD, CE, and CJ (49, 77, and 31 regulators respectively). This observation of an average common TF set of ~70% fits well with the previously published study on the individual and common repertoire of corynebacteria [26]. For the next step in the pipeline the known transcription factor binding sites for each of the 69 characterized TFs of CG have been re-adjusted resulting in an average improvement of the mean information content of the underlying position frequency matrices of ~23% (see additional file 2). Subsequently, PWMs are computed and matched with all upstream sequences of orthologous target genes in CD, CE, and CJ to predict putative TFBSs. In the case of ambiguity of the homology prediction, the pipeline also considers the neighborhood of the respective genes in the target organism and chooses the candidate gene with the maximal number of homologous genes in the genetic surrounding. If the regulator, the target gene, and the binding site are conserved sufficiently, the regulation is considered to be conserved and added to the pipeline's output. Although these restrictions are very strict, we found 530 reliable novel gene regulations for CD, CE, and CJ and thereby increased the database content of the reference database CoryneRegNet considerably (factor 4.2). For the three organisms we identified reliable transcriptional gene regulations for ~40% of the common transcription factors, compared to ~5% for which knowledge was available before. Table 1 statistically summarizes the original and the transferred database content. Figure 2a exemplarily shows a network visualization of the PcaR regulon as known from CG compared to the transferred one of CE. It is obvious that all of the 11 target genes including the regulator itself are orthologous between CG and CE. Since the binding sites are conserved too, as indicated by the sequence logos in figure 2b/c, the pipeline considered all regulations as transferable from CG to CE. This observation fits well with the known PcaR regulon of C. efficiens [27].

The whole data analysis procedure is very time-efficient (< 1 min computing time) and we added the new datasets into the CoryneRegNet 5.0 database for public access. Although it is very likely that our predictions are highly reliable we separated the evidenced from the predicted datasets within the front-end and now provide two sub-versions of CoryneRegNet.

In the application described above, we combined the software packages MoRAine, FORCE, and PoSSuMsearch into a data fusion workflow that is responsible for the data transfer between model and target organisms. One could also think about a combination of other computational biology tools designed for the same purpose. For instance FORCE could be replaced by TribeMCL [22] and PoSSuMsearch by MATCH [20]. We decided to use PoSSuMsearch since it provides statistically sound p-value computations within reasonable response times [19]. FORCE has been included into the pipeline since it has been shown to outperform other clustering approaches on prokaryotic datasets [25].

With Regulogger [28], Alkema et al. presented a pipeline that mainly focuses on the reconstruction of conserved regulatory networks for Staphylococcus aureus. Regulogger mainly concentrates on the detection and characterization of conserved sequences in promoter regions (phylogenetic footprinting) of orthologous genes. For the determination of these orthologous genes, the COG database [29] is used, which is impracticable at least for our target species since no COG annotations are available for other corynebacteria than C. glutamicum. In the near future, novel ultra-fast sequencing technologies will provide much more data on organisms that are not included in COG. In [30] and a subsequent follow-up study [31], the TRACTOR_DB database [32] was used to identify conserved regulatory interactions between E. coli and thirty gamma-proteobacteria. Here, pairwise BLAST Bidirectional Best Hits (BBHs) are used directly to identify orthologous genes, which may result in suboptimal predictions [25]. However, the presented results support our strategy to combine orthology information with binding site detection. Besides, the conservation of gene regulatory networks between corynebacterial organisms has never been studied before and the integration of the presented results into the CoryneRegNet reference database provides a powerful tool for further network studies.

The current study is strongly limited by the level of phylogenetic conservation between the reference organism and the target species. Since remodeling of transcriptional gene regulation is a crucial strategy used by bacteria to evolve and regulate novel biochemical features, (1) specific regulatory pathways may have been altered and (2) unique transcriptional regulatory mechanisms may have been developed. The following three limitations with our approach for inter-species network transfer arouse: (1) Interactions that do not evolve in the reference species cannot be detected in a specific target organism. (2) Utilizing pure sequence-based similarities for in silico orthology detections neglect that proteins with comparably high overall amino acid sequence similarity may have different specific functions within the cell although they are predicted as putative homologs and vice versa. (3) In our approach, we assume that a conserved transcription factor in the target organism interacts with a binding site that is very similar to that in the reference organism. This is not necessarily the case and may result in both, false positive and false negative predictions. However, the rapidly increasing amount of fully sequenced organisms without much background knowledge about their gene regulatory repertoire strongly restricts our alternatives to computational methods that utilize sequence-based evolutionary conservation.

Besides the immediate advantages for the worldwide medical and biotechnological corynebacteria community, we anticipate the presented results to be a starting point for an integrated analysis of gene regulatory networks in the light of a combined analysis of orthologous genes conjointly with conserved DNA binding sites. Although we tested the presented strategy with corynebacteria, it is of general interest and can be applied to many other organisms as well. We conclude that trustworthy transfers of gene regulatory networks between organisms on a genome-scale are feasible but still limited by the level of evolutionary conservation.

Project name: CoryneRegNet 5.0

Project home page: http://coryneregnet.cebitec.uni-bielefeld.de/v5

Operating system(s): Platform independent

Programming language: PHP, Java 6

License: Academic Free

License (AFL)

Any restrictions to use by non-academics: No.

Here, we briefly introduce the integrated bioinformatics tools that have been utilized for this article.