RegPredict: an integrated system for regulon inference in prokaryotes by comparative genomics approach

INTRODUCTION

Accurate reconstruction of transcriptional regulatory networks and annotation of regulatory elements is one of critical aspects of microbial genomics in the era of large-scale genome sequencing. Despite the growing number of genome-wide gene expression studies, our abilities to convert the results of these studies into accurate regulatory annotations (such as ‘gene A is transcriptionally regulated by regulator B’) and to project such annotations from model organisms to related genomes are extremely limited. Likewise, despite the abundance of gene annotation resources and tools, very few of them support regulatory annotations beyond a handful of model species. In our vision, the genomics-driven reconstruction of regulatory interactions in a well-populated taxonomic group should include two main steps: (i) accurate reconstruction of a reference set of regulons in a representative set of genomes from the taxonomic group and (ii) automated propagation of the inferred regulons to all closely related genomes in this group.

Genes and operons directly co-regulated by the same transcription factor (TF) or RNA structure (e.g. a metabolite-sensing riboswitch) are considered to be a part of a regulon. Activation or repression of gene expression is achieved via sequence-specific binding of TFs to regulatory sites (or TF-binding sites, TFBSs) located in promoter regions of genes. The TFBSs often possess an intrinsic symmetry, e.g. palindromic, due to mostly dimeric conformation of bacterial TFs in the DNA-bound state. The size of a single TFBS usually varies between ∼12 and 30 nt, with the most common length being 16–20 nt. Significant variations in binding sites of particular TFs co-regulating multiple genes pose a challenge to the identification of all TFBSs that conform to a consensus motif in the same genome. Nucleotide frequency positional weight matrices (PWMs) are used by many computational algorithms as a more sensitive and more precise way for TFBS recognition in microbial genomes (1).

Comparative genomic approaches are becoming widely used for genome-scale inference of cis-acting regulatory elements (e.g. TFBSs, promoters and RNA regulatory sites) and reconstruction of transcriptional regulatory networks in numerous groups of bacteria (2). The consistency check and phylogenetic footprinting techniques are based on the assumption that regulons have a tendency to be conserved between the genomes that contain orthologous TFs (3–7). The presence of similar candidate TFBSs upstream of orthologous genes is an indication that it is a true regulatory site, whereas candidate TFBSs scattered at random in the genome are considered false positives. Simultaneous analysis of multiple genomes from the same taxonomic group allows one to make reliable predictions of TFBSs even with weak recognition rules. In the general approach, after validation of TFBSs by comparative genomics, the improved PWM is constructed and used to scan the genomes for the presence of additional instances of the TFBS motif, yielding regulon expansion beyond the initial training sets. This leads to another cycle (or several cycles) of iterative refinement aimed at maximizing coverage and consistency of the reconstructed regulons.

In our previous work, we provided multiple examples of the power of the comparative genomic approach applied to a large number of different TF- and riboswitch-driven regulons reconstructed for a broad spectrum of diverse bacteria (2). In particular, this approach was applied to: (i) propagate previously known regulons from model organisms to many others; (ii) ab initio predict and reconstruct novel regulons either for a metabolic pathway of interest or for a particular family of TFs; and (iii) perform wide-ranging reconstruction of multiple regulons in a group of closely related species (8–13). These, and other previous studies, helped us to determine major workflows for genomic reconstruction of regulons and develop the RegPrecise database for collection and visualization of the accumulated regulatory reconstructions (14).

In this paper, we present the RegPredict web server designed to bring our extensive experience to the scientific community and facilitate the procedure of regulon inference in prokaryotic genomes from diverse taxonomic groups by providing a highly integrated set of computational modules for semi-automated and accurate analysis of regulons. RegPredict is designed to approach the goal of fast accumulation of reference sets of regulons in diverse groups of microorganisms.

CLUSTERS OF CO-REGULATED ORTHOLOGOUS OPERONS

In a typical scenario of comparative genomic analysis of regulation, candidate TFBSs are identified in the upstream regions of orthologous genes and subsequent consistency check procedure is carried out (3–7). Two potentially challenging issues that could hinder automatic implementation of this procedure are (i) high conservation rate in closely related genomes (source of false positives), and (ii) frequent operon rearrangements and appearance of additional TFBSs in distantly related genomes (false negatives). To address these issues and facilitate comparative assessment of putative regulon members in a group of closely related genomes, we developed an automatic procedure to split the entire regulon into a set of candidate subregulons that are defined as Clusters of co-Regulated Orthologous operoNs (CRONs). For each analyzed regulon, the set of constructed CRONs is prioritized based on the level of conservation of regulatory interactions, allowing focusing on the most prominent regulon members. At the next step, the functional and genomic context analysis of each CRON is conducted using the advanced web interface (see below), facilitating the decision on CRON inclusion in the final regulon model.

The procedure of CRON construction is outlined in Figure 1. First, the upstream regions of all putative operons are scanned for candidate TFBSs. Second, all orthologous operons with candidate TFBSs are linked together in clusters, allowing one to evaluate the level of conservation of the candidate regulatory interactions. Two operons are considered orthologous if they share at least one orthologous gene. Finally, the initial clusters are extended by adding orthologous operons lacking candidate TFBSs to build a candidate CRON. Each constructed CRON can then be analyzed separately as an independent unit. Combining all accepted CRONs for a given TFBS motif yields the reconstructed TF regulon for a group of target genomes. The CRON-based splitting of a regulon into subregulons is especially important for the analysis of large regulons for global TFs such as CRP in Proteobacteria and CcpA in Firmicutes.

The procedure of CRON construction in RegPredict is based on the genomic data available in the MicrobesOnline database including: (i) high-quality orthologs based on the analysis of phylogenetic trees of protein domains and (ii) predicted operons (15,16).

WORKFLOWS FOR REGULON INFERENCE

The RegPredict web server was designed to provide a set of integrated tools for the complete pipeline of regulon inference in a set of taxonomically related prokaryotic genomes. The server implements two well-established workflows that are widely used for the comparative genomic analysis of regulation: (i) regulon reconstruction for known regulatory motifs with available PWMs and (ii) de novo inference of regulons for previously unknown regulatory motifs (Figure 2).

‘Regulon inference based on known PWM workflow’ utilizes our collection of PWMs for known TFBS motifs. All genomes in the analyzed group of species are scanned by a particular PWM and candidate TFBSs are identified. Then, CRONs are automatically computed, ranked by the number of genomes with candidate TFBSs and by site scores, and finally provided as an output for manual curation in the interactive RegPredict web interface (see below). The PWM-based regulon inference is straightforward when the PWM training set belongs to genomes from the same taxonomic group as target genomes. In a more general case, the applicability of a particular PWM, initially built based on a set of known TFBS from one group of genomes, to another group of genomes can be determined by preliminary checking for the presence of orthologous TFs in the analyzed genomes.

The current version of RegPredict provides a comprehensive collection of PWMs from three publicly available web resources. The RegPrecise database (http://regprecise.lbl.gov), recently developed by our group (14), captures and visualizes predicted TF regulons that were reconstructed by the comparative genomic approach in a wide variety of prokaryotic genomes (∼11 500 TFBSs for ∼400 orthologous groups of TFs from over 350 prokaryotic genomes). The majority of these data are represented by genome-wide reconstructions in several diverse taxonomic groups of bacteria (including Shewanellaceae, Streptococcaceae, Staphylococcaceae, Desulfovibrionales and Thermotogales) and a large-scale prediction of regulons for the LacI family of TFs. The literature-based database RegTransBase (http://regtransbase.lbl.gov) (17) captures experimental knowledge on regulatory sequences and interactions published for a variety of microorganisms using a controlled vocabulary. The RegTransBase database provided data for ∼140 PWMs based on published TFBSs. The RegulonDB database (http://regulondb.ccg.unam.mx) (18) contains models of the TF regulons that were experimentally characterized in Escherichia coli (∼70 TFs with PWMs).

The current list of microbial genomes available for regulon reconstruction in RegPredict includes representative set of species from major taxonomic groups of Bacteria available in the MicrobesOnline database. Currently, RegPredict allows to analyze up to 15 genomes simultaneously.

‘De novo regulon inference workflow’ is intended for prediction of regulons when there is no available PWM. The procedure starts with a set of potentially co-regulated genes from one or multiple related genomes from a particular taxonomic group. The input training set might be composed based on many sources including (i) genes that make up a functional pathway; (ii) genes homologous to regulons from a well-studied species (e.g. E. coli and Bacillus subtilis); (iii) genes derived from conserved chromosomal loci or operons containing orthologous TF genes; and (iv) genes with similar expression profile as determined by microarray experiments (Figure 2). Subsequent important steps are identification of candidate TFBS motifs within a set of DNA upstream regions for genes from the training set and construction of corresponding PWM profiles. The common approach is to find profiles of different types, such as palindromes of different length, or direct repeats, using the MEME-like iterative algorithm previously implemented by us in the SignalX software (4,19) and massively applied for microbial regulon inference (2,20).

The list of candidate motif profiles ranked by information content is provided as an output of the ‘Discover Profiles’ procedure in RegPredict. The selection of a correct motif is a critical point in the de novo regulon inference workflow, and it can be usually obtained by application of the selected candidate motifs to the target genomes and subsequent analysis of the resulting regulon content. RegPredict provides an interface to expedite the evaluation of candidate motifs, as described for the PWM-based regulon reconstruction workflow. Once the optimal motif is identified, its application to the target genomes may result in identification of new candidate regulon members that can be iteratively added to the initial training set. The iterative procedure of PWM construction is simplified by the ‘Gene cart’ for the collection of candidate target genes identified during the de novo regulon inference analysis.

WEB INTERFACE

The RegPredict web server is publicly accessible through a web interface at http://regpredict.lbl.gov. To start the session of regulon inference, the user selects a set of genomes from one or several taxonomic groups of interest. Organisms in the ‘Select Genome’ dialog are now grouped by taxonomy. The selected genome set remains constant for the entire working session and is used in all regulon reconstruction procedures. Then the user is provided with a wizard to start from one of the two main strategies of regulon inference.

EXAMPLE USAGE ON RECONSTRUCTION OF HRCA REGULON IN STAPHYLOCOCCI

Typical scenario of the RegPredict usage is the identification of TFBSs and regulon reconstruction in a group of genomes for a TF with a known motif in a model organism. This scenario can be demonstrated by the inference of the heat shock regulon HrcA, previously described in Streptococcus spp., in a novel taxonomic group, for instance, the Staphylococcaceae (currently includes SEVEN nonredundant genomes).

First, the potential HrcA regulon in the Staphylococcaceae can be estimated by preliminary ‘Run Profile’ scan using the Streptococcus-specific PWM available in the RegPrecise-based collection of profiles and the default profile scan parameters (threshold = 6.3). Three CRONs highly conserved in all Staphylococcaceae will be obtained including two heat shock response operons, groES-groEL and hrcA-grpE-dnaK-dnaJ-SA1407-SA1406-SA1405, and the hypothetical gene SA1838, which shares the same TFBS with the divergently transcribed groES gene. Then additional putative HrcA regulon members can be checked by repeating ‘Run Profile’ with the reduced score threshold.

Second, the ‘Discover profiles’ procedure can be applied to build the Staphylococcaceae-specific HrcA motif. The training set of genes initially found with candidate sites during the previous step in all Staphylococcaceae genomes can be easily collected using ‘Add gene to cart’ function in ‘Gene properties’ panel (three genes in seven genomes). Upstream regions of 21 genes from ‘Gene cart’ are extracted in the ‘Discover profiles’ dialog and the resulting training set is used to build the profile with the parameters specified for the HrcA motif (palindrome; length = 27, 27, any; min palindromic positions = 4; min GC pairs = 1; training set size = 100%). The best resulting palindromic profile (information content = 1.04) can be immediately applied to automatically calculate set of CRONs for HrcA regulon in Staphylococcaceae (profile parameters: score threshold = 7). Sites with scores slightly below the threshold could be easily checked using ‘Show minor sites’ option. For instance, the third CRON including SA1838 has a minor site with score = 6.98 upstream of the orthologous gene in Macrococcus caseolyticus.

It should be noted that the final regulon reconstruction in a particular case could be achieved by manual assessment of each CRON taking into consideration individual scores and overall conservation of TFBSs across the genomes, as well as the genomic context and functional annotations of regulated genes.

SUMMARY AND PERSPECTIVES

RegPredict is a highly interactive web server specifically designed for fast and accurate comparative analysis of microbial regulons and identification of TFBSs in multiple taxonomically related genomes in semi-automatic way. The server has user-friendly interface and multiple functionality including: (i) scanning of microbial genomes with annotated DNA motifs and collections of aligned binding sites, (ii) collecting upstream DNA regions for any gene set and identification of novel TFBS motifs and (iii) comparative analysis of the genomic context of candidate TFBSs in the group of genomes. Among few other web tools available for genomic analysis of TFBS, RSAT suite of tools (22) allows for sequence retrieval, pattern discovery and pattern matching. Key features of RegPredict include: (i) simultaneous analysis of TFBSs in multiple genomes and (ii) distribution of the analysis of large regulons using CRONs. RegPredict is an extremely powerful tool for fast accumulation of reference sets of regulons in diverse groups of microorganisms.

We are planning further development and extension of the RegPredict web server in three main directions. First, we will achieve tight integration of the RegPredict server with the RegPrecise database (14), including direct deposition of the inferred regulons and PWMs to the database using the individual sign-on capability. Such two-way integration will facilitate accumulation of the reference sets of regulons that will further be used for large-scale automated propagation of regulatory annotations to closely related genomes. Second, we will add new types of input gene sets for the de novo regulon inference, including: (i) genes with similar expression profiles using microarray experiments deposited in the MicrobesOnline database (15); (ii) genes derived from conserved gene clusters and operons containing a putative TF gene; (iii) genes that are homologous to previously inferred regulons in other lineages described in the RegPrecise database; and (iv) experimentally known sets of co-regulated genes from the RegTransBase (17), DBTBS (23), and CoryneRegNet (24) databases. Finally, we are planning to add a module for the analysis of RNA regulatory elements described in the RFAM database (25).

FUNDING

The US Department of Energy, Office of Science, Office of Biological and Environmental Research, Genomics Program: GTL through contract DE-AC02-05CH11231 between Lawrence Berkeley National Laboratory and the US Department of Energy; National Science Foundation (award DBI-0850546 to D.A.R.); Howard Hughes Medical Institute (55005610 to M.S.G.); Russian Foundation for Basic Research (08-04-01000 to A.E.K., 09-04-92745 to M.S.G., 10-04-01768 to D.A.R., 09-04-92742 to A.A.M.); Russian Academy of Sciences (program ‘Molecular and Cellular Biology’ to D.A.R and M.S.G.); Federal Agency on Education (P2581 to E.D.S.); Russian Science Agency (contract 2.740.11.0101 to M.S.G.); and Russian President’s grant for young scientists (MK-422.2009.4 to D.A.R.). Funding for open access charge: The US Department of Energy, Office of Science, Office of Biological and Environmental Research, Genomics Program.

Conflict of interest statement. None declared.