We obtained two HCC-related networks corresponding to without- or with- metastasis status. The types of nodes included TF, miRNA, and non-TF-gene; the types of edges included TF-TF, TF-miRNA, TF-gene, miRNA-TF, miRNA-gene. The statistics about nodes and edges are shown in Table 1 in which the regulators included all the nominal TFs or miRNAs in the network, with targets or not. From a global view, the metastatic network was larger and more complex as there are more nodes and edges. We could also see that overall the amount of TFs as nodes and TF-involved edges are both larger than that of miRNAs, no matter in the non-metastatic or metastatic network, which substantiate the critical role of the TF in gene regulation.

To verify whether our networks are correlated to HCC, we performed one-sided Fisher's exact test respectively on the genes from the two networks and the collected HCC-related genes and HCC-metastasis-related genes resorted from a series of a priori databases and literatures. It turned out that genes from the non-metastatic network were significantly overlapped with HCC-related genes (p = 8.35e-8) but not to HCC-metastasis-related genes (p = 0.094), and that genes from the metastatic network were not only significantly overlapped with HCC-related genes (p = 3.81e-9) but also with HCC-metastasis-related genes (p = 0.031). Such results gave us confidence that our constructed networks reasonably lie in the context of HCC and HCC metastasis.

In order to explain the difference of the two networks, we categorized all the nodes and edges into three groups: a) NM-specific nodes or edges that appeared in the non-metastatic network only, b) common nodes or edges that existed in both non-metastatic and metastatic networks, and c) M-specific nodes or edges in metastatic network only. The percentage of each type of nodes and edges in different groups are shown in Figure 2A, B. MiRNAs as nodes appear most in the M-specific group compared with TFs and non-TF genes which appear nearly equally in different groups. MiRNA-involved edges, including TF-miRNA, miRNA-TF, miRNA-gene, also count for an overwhelming proportion in the M-specific group. When considering the increased rate of average number of targets of a regulator in the metastatic network versus the non-metastatic(Figure 2C, Additional File 3), on average a TF regulates more miRNAs but less TFs, and a miRNA regulates both more TFs and genes. These discoveries suggest that miRNAs might participate more actively in tumors with metastasis, which supports the important role of miRNAs in tumor progression 17.

The basic standards on the defining of our key regulatory modules from the combinatorial networks are as follows: i) the selected module should possess clear biological structure to decipher its regulatory pattern. ii) the selected module should contain nodes and edges discriminative of the metastasis status. With such standards, we obtained 71 ranked differential regulatory modules from the two networks in total, each including one specific regulator and all of its first-layer targets, of which 26 were from the non-metastatic network(NM modules) and 45 from the metastatic network(M modules). Based on these differential regulatory modules a series of classification analyses were performed to further identify predictive modules.

Firstly each single module was tested for classification efficiency. Each of the top 20 modules (involving 5 NM- and 15 M- modules) from the ranked differential list was sequentially taken as the single-module classifier and tested in the recursive partitioning classification model 18. The performance of these single-modules was evaluated by leave-one out cross validation (LOOCV), the best of which achieved accuracy (ACC) of about 82%, and Matthew Correlation coefficient (MCC)of 64%. And there was no significant difference between the performance of modules from the non-metastatic network and the metastatic one(Additional File 3).

Then cumulative modules were examined for their predictive ability of metastasis status. The classification procedure was repeated by adding one more candidate module at a time from the top down the previously prepared ranked list. Our results showed that when the top 17 modules were chosen as cumulative-module classifier, the LOOCV accuracy overrode 90% and MCC overrode 80% (Figure 3A, Additional File 3), which was an explicitly great improvement than the single-module classifiers. From then on when more modules were added to the classifier, more genes were brought into, while the performance did not improve significantly. So these 17 modules were considered as the key regulatory modules predictive of HCC metastasis in our work, which altogether involved 139 unique genes and miRNAs in 5 NM and 12 M modules. The full list of these modules is displayed in Table 2.

Some previous reports demonstrated the advantage of subnetwork classification over single gene-lists, probably because of functional relevance in the classifier 11. In order to check whether our key regulatory modules possess such advantage, we performed gene-list-based classification procedure in a counterpart way to our module-based classification. Signature genes were the selected differentially expressed genes in HCC metastatic vs. non-metastatic samples, with the Student t-test Benjamini-Hochberg adjusted p value < 0.001, and further ranked by the method of minimum redundancy and maximum relevance(MRMR) 19(Additional File 1), which resulted in a list of 349 ranked candidate genes. Same number of genes as in the cumulative-module classifier were picked with priority from the ranked gene list to compose the single-module classifier or perform metastasis classification and the performance was evaluated by LOOCV. Our results showed, there existed no significant difference of ACC or MCC between the top 20 single-module-classifiers and counterpart gene-list-classifiers (two sided t test p value > 0.5) (Additional File 3). However, when combining modules (even just two) the cumulative-module classifier achieved consistently better performance than the classifying models of corresponding number of signature genes (Additional Figure 2, Additional File 3).

According to the approach of identifying key regulatory modules predictive of metastasis sub-statuses, these modules possess regulatory patterns which were disturbed in tumors with metastasis(either disappeared or appeared). To investigate the disturbed pathways of these NM and M modules, we conducted enrichment analysis on genes from each of the key regulatory modules with all non-metabolic KEGG resource containing gene regulatory and signaling pathways. It turned out that only 6 modules (headed by 6 regulators) out of the 17 were significantly enriched in 28 non-metabolic pathways (Table 3). NM key modules(FOXO3_NM, TP53_NM) were enriched mostly in various cancer pathways and several cancer related processes such as cell cycle and apoptosis; M key modules (hsa_miR_16_M, has_let_7e_M, has_miR_30a_M, STAT1_M) were mostly enriched in signaling pathways related to tumor progression. In order to illustrate a full functional regulatory landscape in the context of HCC metastasis, we merged the topology structure of the 17-module classifier and the enriched 28 non-metabolic KEGG pathways (Figure 4A). The graph structure of KEGG pathways embodied by gene(protein)-gene(protein) interactions was retrieved by the R/Bioconductor package KEGGgraph. In this way, individual key regulators and its module can be zoomed in to see in detail the regulatory pattern transition between non-metastatic status to metastatic status.

Out of 17 key classifying modules predictive of metastasis, six were enriched in KEGG non-metabolic pathways: FOXO3_NM, TP53_NM, STAT1_M, hsa_miR_16_M, has_let_7e_M, has_miR_30a_M. Based on the hypothesis that miRNAs might actively participate in the tumor progression and metastasis process and act as key roles, we further focused on the regulatory patterns of the three modules headed by miRNAs, which were zoomed in from the regulatory landscape constructed above.

MiR-30a. The module led by miR-30a in the metastatic network shows inextricable links to various cancer-related pathways and some other important regulators (Figure 4B). EP300 and CREBBP, regulated by way of miR-30a and CREB1, are highly related transcriptional co-activators possessing histone acetyltransferase activity and were known to be involved in the survival and invasion pathways of prostate cancer 19. Functions of TP53 and STAT1 might be modulated through acetylation by CREBBP/EP300. Meanwhile, NEDD4, another target of miR-30a, by further targeting EGFR, might interfere with key cellular signaling pathways. According to a previous report, miR-30a was reported to inhibit epithelial-to-mesenchymal transition in non-small cell lung cancer 20. The exact role miR-30a might play in HCC metastasis requires more exploration.

MiR-16. MiR-16 targets human nuclear co-repressor 2(NCOR2) in the metastatic network. NCOR2 is a transcriptional co-repressor linked to Notch (Figure 4C). According to previous reports, Notch signaling cascade was regarded as anti-proliferative rather than oncogenic in hepatocellular carcinoma 2122, so if miR-16 repressed Notch it might result in more aggressive HCC that would lead to metastasis.

Let-7e and miR-204. In tumor with later metastasis, let-7e targets nerve growth factor (NGF), whose deprivation was supposed to induce apoptosis 23. Upstream let-7e is ATF2 and miR-204 (Figure 4D). Because we only performed the first-step targets enrichment in KEGG pathways, miR-204 was not among the six key regulators whose targets showed pathway enrichment, yet it was one of the heading regulators of 17 key regulatory modules. Besides, referring to the topology of our HCC metastatic network, miR-204 is a bottleneck with the 7th highest betweenness, and two of miR-204-involved edges rank into the top 10 list of edge betweenness(Additional File 3). Therefore we may hypothesize that the important role of let-7e regulation was driven by its upstream regulator miR-204. MiR-204 represses the expression of its target ATF2, blocking its activation to downstream target let-7e. The lack of let-7e may release NGF deprivation and therefore inhibit apoptosis, leading to tumor aggression and HCC metastasis. Indeed miR-204 was previously reported to regulate mesenchymal progenitor cell differentiation 24and to be related to head and neck tumor metastasis 25. Therefore we list both let-7e and its upstream miR-204 to be key regulatory miRNAs that might relate to HCC metastasis.

With the 17 key regulatory modules, all the HCC patients could be classified into two groups: without or with venous metastasis. To investigate the survival outcomes of these two groups, Kaplan-Meier analysis was performed, and the survival difference between the two groups was evaluated by log-rank test (p = 0.01) (Figure 3B). It turned out that the predicted NM group correlated significantly with a longer overall survival, whereas the M group correlated significantly with a shorter one. The difference of clinical characteristics between predicted groups of venous metastasis were also investigated (Table 4). BCLC stage, AFP level, TNM stage, each showed significant difference between these two groups. As we know, AFP is a common critical index in HCC progression, and BCLC and TNM stages reflect the progression stage of hepatocellular carcinoma. These results suggested that the predicted HCC subgroups of venous metastasis might represent distinct prognosis and clinical stages, which on the other hand subordinate the rational performance of the identified 17 key regulatory modules.

Next we scanned the association of each module to clinical features using GlobalAncova (Figure 5). Among the clinical parameters, AFP related with most modules, followed by another clinical parameter, ALT, and then BCLC stage, TNM stage and Child-Pugh class. To be noted, the three modules led by the key regulatory miRNAs identified above (miR-16, miR-30a, miR-204/let-7e) were all closely associated with AFP and ALT. Moreover, hsa_miR_16_M was simultaneously related with the most number of clinical characteristics (AFP, ALT, and 5 cancer staging indexes), which further suggested that miR-16 and its targets might correlate with invasive tumor cell behavior.

mRNA and miRNA expression microarray data on the same cohort of HBV-infected HCC patients who underwent radical resection in Zhongshan Hospital were used for integration in this study. Both datasets (GSE5975 and GSE6857) were downloaded from the Gene Expression Omnibus (GEO) database http://www.ncbi.nlm.nih.gov/geo/. The mRNA signal intensities were retrieved from GSE5975, which was generated using the NCI/ATC Hs-OperonV2 array. The miRNA expression levels were obtained from GSE6857, which was generated using OSU-CCC MicroRNA Microarray Version 2.0. Status of venous metastasis of patients were collected from GSE6857. Other clinical pathologic characteristics and survival time of patients were provided by Zhongshan Hospital.

Microarray data preprocessing was conducted on each dataset separately, and then both mRNA profiles and miRNA profiles were combined to prepare the combined expression profiles among the 198 patients, 150 non-metastatic and 48 metastatic. After quantile normalization across arrays on the combined expression profiles, irrelevant genes and mature-miRs within the 5% smallest standard deviations of tumor/nontumor profiles between metastasis and non-metastasis samples were filtered. Finally, the combined mRNA and miRNA expression profiles of 198 patients included 12434 genes and 132 mature-miRs altogether. More detailed information for data processing is available in Additional File 1.

In the following data selection, a gene list of 1318 previously defined TFs 34 from a previous report were regarded as TFs, while others as non-TF genes.

MiRNA-gene. Candidate miRNA-target relationships were downloaded from miRBase Target Version 5.0, TargetScanHuman Version 5.1, and miRDB Version 3.0, each was based on the predicting algorithm miRanda 35, TargetScan 36, and miRTarget2 37, respectively. The predicted miRNA-gene relationships with accordance in at least two algorithms were retained in our study.

TF-gene. A set of predicted TF-gene relationships were compiled with methods mainly described in our previous work 16, where TF-TFBS(TF binding sites) relationships and TFBS-gene relationships were first calculated, based on which TF-gene relationships were linked. The difference between the method in this work and our previous work was that the promoter region of each gene in our work was defined as 1k bp up- and down- stream (instead of 1 kb upstream to 0.5 kb downstream of the transcription start site (TSS) according to the ENCODE project 38.

TF-miRNA. TFBSs mapped to the regions upstream of miRNA primary transcript TSSs were downloaded from miRGen 2.0 39. Precursor-miRs were mapped to mature-miRs according to miRBase database. Then the candidate TF-miRNA relationships were generated based on the above TF-TFBS relationships and TFBS-miRNA relationships.

The statistics of the final set of 327711 regulatory relationships based on sequence-matched in human between TFs, miRNAs, and genes were displayed in Additional File 3.

HCC-related genes were collected from HCCdb 40, EHCO-II 41, and HCCNet 42. The union set of 5088 genes from these three databases was taken as the HCC-related genes. HCC-metastasis-related genes were collected using the text-mining tool, SciMiner 43. ("carcinoma, hepatocellular"[MeSH Terms] OR hepatocellular carcinoma[Text Word]) AND ("liver neoplasms"[MeSH Terms] OR liver cancer[Text Word]) AND ("neoplasm metastasis"[MeSH Terms] OR metastasis[Text Word]) AND metastatic[Text Word]) was set as the query string for full text mining. The resultant 322 genes each cited by at least 2 papers were regarded as the HCC-metastasis-related genes in this study. All the collected HCC-related genes and HCC-metastasis-related genes are listed in the Additional File 4.

One sided Fisher's Exact Test was performed to examine whether genes in our constructed HCC non-metastatic and metastatic networks were significantly overlapped with the collected HCC- or HCC-metastasis- related genes from databases and literatures. All the 12434 genes in the combined expression profiles were used as the set of universe genes in the test.

The combined expression profile was divided into two sub-profiles by sample labels, namely profile of non-metastasis and profile of metastasis, so as to construct gene regulatory network of HCC without and with metastasis respectively.

We assumed that sequence-matched pairs were more possible to be real interaction pairs than sequence-unmatched pairs, and that real interaction pairs were more possible to be correlated in expression than random pairs. In order to construct the network as credible as possible, we filtered out untrustworthy pairs before expression-based network inference. The candidate 327711 sequence-matched relationships genome-wide were first reduced to 78310 non-self-looping pairs whose both nodes were genes and miRNAs with expression in the combined profiles. Then the absolute spearman correlation of the expression was calculated between each of these 78310 sequence-matched pairs, and the mean absolute spearman correlations of the expression were also calculated between randomly sampled 78310 pairs from the combined expression profile for 100 random times. Pairs with the absolute spearman correlation higher than 95% of random pairs were retained as candidate pairs, which were processed to infer the transcriptional interactions.

Based on the two sub-profiles respectively, based on all nodes from the above remaining pairs, Context Likelihood of Relatedness (CLR) 30 was then applied as the network inference algorithm to identify transcriptional interactions using an R/Bioconductor package minet with default parameters. The CLR algorithm returned a non-negative matrix which was the weighted adjacency matrix of the network whose values represented the edge weights of the network. We set the cutoff for edge weights as 1, and edges whose edge weight below 1 were thus removed, since edges with little weight were considered as marginal relationships and might be noise.

CLR uses mutual information as a metric of similarity between the expression profiles of two genes. Formally, the mutual information for two discrete random variables X and Y is defined as:

I ( X ; Y ) = ∑ i , j p ( x i , y j ) log p ( x i , y j ) p ( x i ) p ( y j )

where p(x_i, y_j) is the joint probability distribution function of X and Y, and p(x_i) and p(y_j) are the marginal probability distribution functions of X and Y respectively. In the case of continuous random variables, the summations over X and Y are replaced by integrals. For genes, X and Y represent a transcription factor and its potential target gene, and x_i and y_i represent particular expression levels (Further description in Additional File 1).

The composition of our 'modules' was defined as one specific regulator and all of its first-layer targets (more than one), and was named as Regulator_Status. Regulator was the name of the regulator, i.e. a TF or a miRNA. Status represented the source network of the module; it could be from the non-metastatic or metastatic network. All the modules in our work included targets only one step down from the regulator such that the regulatory attributes of each module was explicit to read.

Differential modules were first selected before identifying predictive classifying modules of metastasis sub-statuses. As to edges, the non-discriminative edges were excluded from the networks. For all the edges appearing in any of the two networks, we calculated the absolute value of the edge weight difference (The edge weights were directly carried on from the CLR results. The edge weight of a non-existing edge was regarded as zero.) between the two sub-statuses. The edges whose absolute edge weight difference were within the lowest 25% among all the edges were regarded as non-discriminative ones and were filtered out. As to nodes, GlobalAncova 44 test was performed on each module to measure the discriminance of nodes in that module between the metastasis statuses, which was implemented using R/Bioconductor package GlobalAncova (Additional File 1). Significant differential modules with Benjamini-Hochberg adjusted p < 0.001 were taken as candidate predictive modules, which were sorted by their corresponding p values from smallest to the largest. Finally, these ranked differential modules were proceeded to classification.

A multivariate algorithm, recursive partitioning, was chosen as the classification model 45. It creates a decision tree that strives to correctly classify members of the patients based on several dichotomous dependent variables, which is simple and intuitive (Further description in Additional File 1). Recursive partitioning has been successfully applied in other cancer biology context to identify multi-gene biomarkers or signatures 464748. In this study, the classification procedure was performed using R/Bioconductor package rpart with default setting of parameters. The predicted group and the prediction possibility for each individual were returned at each performance using this program. For cumulative modules as one classifier (a list of modules), the final predicted label for each individual was determined as the label with the larger overall predicted probability by modules in the classifier; for single-module as one classifier (a list of genes), the final predicted label for each individual was determined as the label with the larger predicted probability. Leave-one-out cross-validation (LOOCV) was used to evaluate the classification performance.

The survival analysis was performed to compare patient overall survival. Kaplan-Meier estimation was calculated to plot the survival curve. Log-rank test was used to compare two survival distributions and generate the p value. Comparison between clinical pathological indicators was conducted using chi-square test for discrete variable and Wilcoxon test for continuous variables. The association between clinical pathologic characteristics and classifying modules was examined using GlobalAnova test by R/Bioconductor package GlobalAncova (Additional File 1).