In the last decade, the advent of high throughput technologies led to the emergence of the systems biology era and prompted the research community to systematically define the expression levels of mRNAs and micro-RNA (miRNAs) through thousands of cell and tissues under physiological and pathological conditions 1. Now, one of the crucial issues is to define the biological mechanisms that drives genes expression with the ultimate goal of reverse-engineering gene regulatory networks (GRN) as a whole in order to predict the system outcome under molecular perturbations.

One current limit for biologists interested in mining regulatory information or for bioinformaticians interested in creating regulatory maps for modeling, is that this information is scattered over the Internet under various formats making it difficult to handle. Thus one needs to create a unified database that would list known and predicted molecular interactions. This information can be obtained from different sources: (i) from the literature, (ii) from large-scale experimental methods that allow genome-wide profiling of transcription factors (TFs) binding sites to DNA or (iii) from DNA sequence analysis, by searching 3'UTR regions for miRNA specific motifs or by scanning gene promoters with transcription factor specific position weight matrices (PWMs). In the latter case, the use of comparative genomics is known to greatly improve predictions of functional TF binding sites by limiting the number of false positives (though increasing false negative rate) 23. Another limit of GRN analysis is the intrinsic complexity of the data. In this regard, several graph-based tools have been developed to draw a global picture of the putative interactions taking place in the biological context of interest (for a review, see reference 4). In these, genes or proteins appear as nodes in a graph, and functional relations (physical/regulatory interactions) are represented as edges connecting the corresponding entities. The topology of the subsequent network can later be analyzed using advanced tools such as Cytoscape 5. However, as data integration is a challenge that requires to map various types of evidence onto a set of stable gene ids, most applications are oriented toward a single data type (mostly regulatory or physical interactions, see table 1 for an overview) 678910 Moreover, another challenge is the development of graph-based tools producing clear, meaningful and integrated visualizations from which users can draw new hypotheses without being overwhelmed by the density of the presented graphic information. In this regard, the Cytoscape plug-in "Cerebral" proposes an intuitive visualization method through a "cell compartment-based layout" that shows interacting proteins on a layout resembling "traditional" signalling pathway/system diagrams 11.

Here, we sought to create a compendium of predicted and validated molecular interactions in human and mouse. First, we used a large collection of PWMs obtained from TRANSFAC (n = 523), JASPAR (n = 303) and UNIPROBE (n = 387) to search, in gene promoter regions, for candidate transcription factor binding sites (TFBSs) conserved over human, mouse, rat and dog genomes 121314. Overall, our analysis of these PWMs corresponding to 347 human and 475 mouse transcription factors (TFs) provides a systematic overview of gene regulation in the human and mouse. Data generated in this study were next integrated with a large set of molecular interactions from various sources including (i) potential protein/DNA interactions derived from ChIP-seq experiments (ChIP-X database), (ii) curated regulatory interactions obtained from the literature (OregAnno, LymphTF-DB), (iii) predicted miRNA/targets interactions (TargetScan) (iv) protein kinase-substrate interactions derived from multiple online sources (KEA) and (v) physical protein-protein interactions obtained from HPRD, Reactome and various databases of the IMEx consortium 15161718192021222324252627282930. Informations related to these interactions were stored as MySQL tables that were integrated in the back-end database of TranscriptomeBrowser, our previously published microarray datamining software 31. Finally, we developed InteractomeBrowser (IBrowser) as a plugin for TranscriptomeBrowser. IBrowser was developed using the prefuse Java library and can be used to translate any gene list into a meaningful graph. The specificity of the IBrowser plugin relies on a new "cell compartments-based layout" that makes use of a subset of the Gene Ontology to map gene products onto relevant cell compartments. This layout is particularly powerful for visual integration of heterogeneous biological information. Moreover, IBrowser is integrated into the TranscriptomeBrowser suite, which allows an easy communication with other tools, for instance to retrieve lists of genes that are frequently coexpressed in given conditions, thus creating context-specific views of the interactome and regulome.

IBrowser is intended both for biologists and bioinformaticians. On one hand, it is a graph-based knowledge browser, that is intended to provide new insight into any user-defined gene list. On the other hand it is also intended to fill the gap between heterogeneous genomic data and gene regulatory network analysis. In this regard, graphs produced inside IBrowser may be exported into Cytoscape and GINsim, a dynamic modeling software 32. In the following sections we provide several examples underlying the benefits of this visualization tool for large gene set analysis.

We first used phylogenetic footprinting to predict regulatory elements in the human and mouse genomes. A dataset of 1,213 PWMs corresponding to mouse or human transcription factors was obtained from various sources (TRANSFAC 10.2, JASPAR 2010, UNIPROBE). The multiz28way (with hg18 as a reference) and the multiz30way (with mm9 as a reference) cross-species multiple alignments were obtained from UCSC 33. We retained for analysis alignments flanking transcription start sites on both sides (-3000, 3000) of any RefSeq transcript and devoid of coding sequences. Sequences were scored following the commonly used formula 34:

S C O R E p , c = ∑ w = 0 W - 1 l o g 2 P s e e i n g S p + w a t p o s i t i o n w | P W M P s e e i n g S p + w a t p o s i t i o n w | B a c k g r o u n d m o d e l

where SCORE_{p, c} represents the PWM score for a PWM of length W in the DNA sequence of a species c between positions p and p+W-1 and S_p+w represents the nucleotide observed at position p+w. The probability of observing each nucleotide under the background distribution was assumed to be 0.25. For each PWM m, a score threshold t_m with p-value below 5.10^-5 was computed using matrix-distrib from RSAT ensuring high stringency of sequence scoring 35. A sequence in the reference genome was considered as a putative TFBS if its score for PWM m at position p in the alignment was found above t_m in human, mouse rat and dog. Each PWM was then linked to its corresponding transcription factors and putative targets. Information was stored in a MySQL relational database.

We also integrated several informations obtained from popular databases. Protein/DNA interactions (n = 174,168) derived from various genome wide analysis (e.g.; ChIP-on-chip, ChIP-seq and ChIP-PET) and encompassing interactions corresponding to 38 human TFs and 55 mouse TFs were obtained from the ChIP-X database. TFBS predictions were obtained from the present work (see below) and TFBSConserved UCSC track (n = 367,829 and n = 686,936 respectively). A set of regulatory interactions curated from the literature were obtained from LymphTF-DB (392 directed interactions) and OregAnno (1,991 interactions). Protein-protein interaction datasets were obtained from HPRD (n = 78,325), Reactome (n = 166,001) and IMEx (n = 110,578). Protein kinase-substrate relationships were retrieved from KEA (n = 14,084). Finally, miRNA/target relationships were obtained from TargetScan database predictions (n = 260,068). For all datasets, all identifiers were mapped onto Entrez Gene ids. This compendium of molecular interactions is available as flat files at: ftp://tagc.univ-mrs.fr/public/TranscriptomeBrowser/DB_Tables/.

InteractomeBrowser was developed using the Prefuse Java library which was modified according to our needs. InteractomeBrowser requires Java 1.6.

InteractomeBrowser and its underlying approach can be compared to the Cerebral (Cell Region-Based Rendering And Layout) plugin of Cytoscape that also combines molecular interactions with a cell-compartment based layout 11.

But there are qualitative differences in the conception of Cerebral and InteractomeBrowser, which make the latest an interesting alternative for exploring networks.

On one hand, Cerebral uses a layered representation of the cell to create a "pathway-like" view of the network of interacting proteins. This layout thus provides a linear organisation of the network. On the other hand, the layout of InteractomeBrowser is based on a schematic view of the entire cell and displays the hierarchical structure of the underlying Gene Ontology subset as nested zones. First, this helps visually separating different parts of the network corresponding to different cellular localisations, as in Cerebral. But this is a more generic visualisation method, in the sense it does not restrict the visual message to an 'input-intermediates-output' mechanism such as in linear pathway diagrams. As a consequence it is suited for a more general study of various types of networks. Moreover, since visual zones correspond to Gene Ontology terms, this layout handles different levels of accuracy in the localisation of proteins: for instance a precisely-annotated protein might be placed in the zone corresponding to "endoplasmic reticulum", while a less well-annotated can be placed in the more generic, higher level zone "intracellular".

In Cerebral, each gene product is represented by one instance whose cell compartment may be defined by the user. In contrast, InteractomeBrowser displays, by default, several instances of a given gene product that may be placed in several cell-compartments according to informations provided by the GO Cellular-component ontology. Although this may lead to a more complex graph, it provides a more exhaustive presentation of current knowledge and may draw the attention of users to unexpected locations of gene products in the cells. The user may choose to delete some of these instances hence selecting a posteriori the most representative one.

The main benefit of InteractomeBrowser resides in its direct interaction with the database described in this report. Indeed, it provides a ready-to-use web-based service that requires only few manipulations to retrieve a network of interactions (see video tutorial provided as additional file). Notably, in addition to physical interactions it offers a unified access to miRNA targets and results from ChIP-Seq experiments derived from CHEA.

Presently, the data sources associated with the InteractomeBrowser plug-in are restricted to human and mouse. Indeed, one of the main objectives of InteractomeBrowser is to help users in creating regulatory maps to study human gene regulatory networks in physiological and pathological conditions. The choice of mouse as an additional organism supported by our database is a natural choice as it is a widely used model of human physiopathology. However, we are already planning to add new organisms in the near future.

As more and more experimentally validated interactions are available, we hope that this tool will prove very useful for researchers.

InteractomeBrowser comes as a plugin for TranscriptomeBrowser and is available at: http://tagc.univ-mrs.fr/tbrowser/. Our database is updated on a regular basis. See additional files for a video tutorial.

• Project name: InteractomeBrowser

• Project home page: http://tagc.univ-mrs.fr/tbrowser/

• Operating system(s): Platform independent (Java)

• Programming language: Java

• Other requirements: Java > 1.6.X

• License: no license required

• Any restrictions to use by non-academics: none

PWM: Position Weight Matrices; GRN: gene regulatory network; GO: Gene Ontology; micro RNA: miRNA; TF: transcription factors; TFBS: transcription factor binding site; TBMC: TranscriptomeBrowser Motif Conservation.

The authors declare that they have no competing interests.

CL, AB, FL, CN, JI and DP conceived the project. CL, AB and FL developed the Java application. AB, CL and NBP developed the database. DP performed the TFBS analysis. DP, CN and JI supervised the project. DP wrote the manuscript. All authors read and approved the final manuscript.