Web server implementation
The cGRNB (http://www.scbit.org/cgrnb) is released as a freely accessible tool for modelling combinatorial gene regulatory networks from built-in regulation libraries and users-uploaded gene expression datasets. cGRNB is designed under a PHP and R framework (Figure 2). The PHP modules control the data flow and the R modules perform the calculations.
Two main functional R modules, Mod_MPGE and Mod_Parallel, are deployed in cGRNB. Mod_MPGE works on MPGE datasets to build combinatorial gene regulatory networks covering miRNA-gene and TF-gene regulation. Mod_Parallel works on miRNA/mRNA expression datasets to build combinatorial regulatory networks covering miRNA-gene, TF-gene and TF-miRNA regulations.
The interface is web-based and users without R programming expertise can freely utilize the calculation modules. After the expression datasets are uploaded, users can set the required parameters on the web graphic interface (Figure 3). When the calculation is finished, users can view or download the HTML formatted reports through a URL sent to the pre-designated email addresses.
Data libraries
There are three data libraries (TF2gene, TF2miR and miR2gene) deposited in cGRNB as built-in components that will be called at every calculation process. The TF2gene and TF2miR libraries are comprised of forward-engineered putative "TF to gene" and "TF to miRNA" regulation relationships respectively. These two libraries were extracted mainly from the source file 'tfbsConsSites.txt' and 'tfbsConsFactors.txt' obtained from UCSC hg19, where the two source files were the results of scanning the human genome for human/mouse/rat conserved TF binding sites (http://genome.ucsc.edu/cgi-bin/hgTables). The miR2gene library, extracted from original dataset of starBase(http://starbase.sysu.edu.cn/) [13], includes putative "miRNA to gene" regulation relationships mapped from CLIP-Seq and Degradome-Seq data. We processed the original data files so that miRNA transcript names are consolidate into their root forms since they are indexed according to their genome coordinates. For example, 'hsa-let-7a-1', 'hsa-let-7a-2' and 'hsa-let-7a-3' are renamed to be 'hsa-let-7a'. TF2miR and miR2gene libraries are also subject to this rule, and we strongly recommend users to process their expression datasets in the same manner. A Perl script tailored to this goal can be found in the download page of cGRNB.
Detailed information about how to process and access these data libraries can also be found at the help page of the web server.
Mod_MPGE
In an MPGE experiment, a miRNA is first transfected into a certain cell line. After a time period (usually 12h or 24h), the mRNA levels in the miRNA-transfected and pre-transfected cells are both measured and compared. A MPGE dataset can be utilized for building a miRNA-driven two-layer combinatorial gene regulatory network.
Based on the MPGE dataset and two data libraries (miR2gene and TF2gene), Mod_MPGE is aimed at three mutually related goals and in the end arrives at a two-layer regulatory networks centring on the perturbing miRNA and its downstream regulating TFs (Figure 4). The three goals are as follows: 1) to evaluate the significance of miRNA degradating mRNAs in human cells; 2) to refine miRNA's degraded targets from forward-predicted putative targets; 3) to identify mediating TFs that transfer miRNA's regulation effect to downstream secondary targets. Goal one and goal two are achieved through non-parametric statistical tests that compare the mRNA level(s) of miRNA's putative targets against those of the non-targets (the complement set to the putative targets). For goal three, we first perform a pre-filtering uni-variate linear regression to screen out highly plausible regulator-target relations one by one, and then apply a multi-variate linear regression to further refine the combinatorial regulators of each target. Been taken together, the miRNA2gene links and TF2gene links output from goal two and goal three make up a two-layer regulatory network centring on the perturbed miRNA and its downstream mediating TFs. The overall diagram of the Mod_MPGE algorithm is illustrated in Figure 4 and the full mathematics details can be found in our previous related algorithm paper [11].
A case study of Mod_MPGE
We tested Mod_MPGE with an MPGE dataset related with "hsa-miR-1" miRNA. This MPGE dataset GDS1858 was obtained from Gene Expression Omnibus (http://www.ncbi.nlm.nih.gov/sites/GDSbrowser?acc=GDS1858). It contains the expression log ratios of a total of 20,127 protein-coding genes, which are obtained by comparing the expression profiles of HeLa cells before and after the transfection of hsa-miR-1. We recommend the tabular data of the MPGE dataset to be stored in a CSV (Comma-Separated Values) file. Spreadsheet software like Microsoft Excel will enable CSV extension conversion. In the CSV file, the first line contains a description of the columns; all other lines must contain a gene symbol and an expression log ratio in the first and second column separated with a comma. The only procedure to begin a calculation is to upload the MPGE data file to the server, set the appropriate parameters, provide a valid e-mail address and click the run button. The job runs in the web server and an e-mail including the URL of the report page is immediately sent to the user. In this example, it took about 5 minutes to finish the calculation and get the result report.
The result include three sections: 1) the targets of the particular miRNA (hsa-miR-1 in this case) and a PP-plot chart with detailed information on the miRNA's target analysis; 2) a list of the TF mediators; 3) a two-layer combinatorial gene regulatory network with the TFs mediating and the miRNA-initiating regulatory effects. Most of the results are shown as tab-delimited text tables with the gene identities and relevant statistics (http://www.scbit.org/cgrnb/doR_target_result.php?jobID=821338391035). As R is not good at displaying a dynamic graphic object on the web interface, we suggest users to download the original CVS file 'network.edge.txt' from the report page and reload it to an external graphical tool, for example CytoScape [14].
Mod_Parallel
A parallel miRNA and mRNA expression dataset includes two data matrices of the same set of column headers (experimental conditions) but different sets of row headers (biological molecules) - one set for miRNAs and the other for mRNAs. A parallel miRNA and mRNA expression dataset can be utilized for building a combinatorial gene regulatory network encompassing three types of gene regulations ("TF to gene", "TF to miRNA" and "miRNA to gene").
Based on the parallel miRNA/mRNA expression dataset and three data libraries (TF2gene, miR2gene and TF2miR), Mod_Parallel sets out to map a comprehensive TF-and-miRNA-involving combinatorial gene regulatory network and this also goes further to analyze its various topological properties (Figure 5). Similar to Mod_MPGE, here the multi-variate linear regression model is adopted to infer plausible regulation relationships. In this module, because heterogeneous expression data types (miRNA expression and mRNA expression) are available, we build up two multi-variate linear equations to model the expression of mRNAs and miRNAs separately. TF-gene and miRNA-gene regulations are output of the mRNA-targeted equation, and TF-miRNA regulations are output of the miRNA-targeted equation. Been taken together, the three types of regulations made up a comprehensive combinatorial gene regulatory network correlating to the particular experimental conditions.
Mod_Parallel then conducts topological investigation of the resultant combinatorial regulatory network and identifies the important vertices/edges, regulator pairs and three-vertex regulating motifs. We first pinpoint the crucial vertices and edges of the network according to degree rank and betweenness rank. Then the 'co-regulating regulator pairs' in the network is marked. To this end, we check all regulator pairs by testing the significance of their potential co-regulating targets. Finally we carry out the triple-vertex motif analysis. Theoretically, there are eighteen triple-vertex regulatory motifs involving at least one miRNA and one TF. These motifs are defined as closed triple-vertex regulatory circuits, and can be classified into 'feed-forward loops' (FFLs) and 'feed-backward-loops' (FBLs) according to the ways of the directional regulations being connected. We count the occurrences of all possible triple-vertex motifs in the resulting network and estimate the corresponding p-values by comparing the real occurrence against the counterpart occurrences in the randomly shuffled networks.
While the overall diagram of the Mod_Parallel algorithm is illustrated in Figure 5, more details of the algorithm can be found in our previous algorithm paper [12].
A case study of Mod_Parallel
The parallel cancer gene expression datasets were downloaded from CellMiner (http://discover.nci.nih.gov/cellminer/loadDownload.do). The two datasets, one for miRNAs and the other for mRNAs, were designed to study a total of 60 types of human cancer cell lines. The experiments were carried out on the 41,000-probe Agilent Whole Human Genome Oligo Microarray and the 15,000-feature Agilent Human microRNA Microarray V2.
The miRNA expression dataset included 365 human miRNAs with detectable expression levels [13]. After miRNA names and their corresponding data were pre-processed, this dataset covered only 266 miRNAs. The mRNA expression dataset had more than 41,000 data rows, while 40,155 rows have Entrez GeneID available. After removing data rows without GeneID and combining rows for identical genes, we finally obtained the expression data for 21,319 protein-coding genes. These two parallel expression datasets are taken as sample datasets for Mod_Parallel, which can be obtained from the cGRNB's "Download" page. The calculation takes around 60 minutes on our web server.
The results of Mod_Parallel include four parts: 1) a CSV file that indicates the edges of the combinatorial gene regulatory network; 2) the vertices/edges ranked by topological features; 3) significantly co-regulating regulator pairs (p < 0.01, one-sided Fisher's exact test); 4) significance of the recurrence of the 18 triple-vertex motifs and all instances of the existing motifs. The results are provided as tab-delimited tables and can be downloaded as CSV formatted files from the example report (http://www.scbit.org/cgrnb/doR_network_result.php?jobID=931338180093).