Regulatory feedback loops bridge the human gene regulatory network and regulate carcinogenesis

Regulatory feedback loops bridge the human gene regulatory network and regulate carcinogenesis Abstract The development of disease involves a systematic disturbance inside cells and is associated with changes in the interactions or regulations among genes forming biological networks. The bridges inside a network are critical in shortening the distances between nodes. We observed that, inside the human gene regulatory network, one strongly connected core bridged the whole network. Other regulations outside the core formed a weakly connected component surrounding the core like a peripheral structure. Furthermore, the regulatory feedback loops (FBLs) inside the core compose an interface-like structure between the core and periphery. We then denoted the regulatory FBLs as the interface core. Notably, both the cancer-associated and essential biomolecules and regulations were significantly overrepresented in the interface core. These results implied that the interface core is not only critical for the network structure but central in cellular systems. Furthermore, the enrichment of the cancer-associated and essential regulations in the interface core might be attributed to its bridgeness in the network. More importantly, we identified one regulatory FBL between HNF4A and NR2F2 that possesses the highest bridgeness in the interface core. Further investigation suggested that the disturbance of the HNF4A-NR2F2 FBL might protect tumor cells from apoptotic processes. Our results emphasize the relevance of the regulatory network properties to cellular systems and might reveal a critical role of the interface core in cancer. gene regulatory network, regulatory bridge, carcinogenesis, feedback loop Introduction The gene regulatory network (GRN) is responsible for the control of biological processes in cells [1, 2]. In metazoans, the GRN consists of two major families of gene regulators, transcription factors (TFs) and microRNAs (miRNAs) [3]. TFs are proteins that bind to specific nucleic acid sequences to regulate target gene expressions at the transcriptional level [4]; miRNAs are small (∼21–22 nucleotides), noncoding RNAs that regulate gene expression at the posttranscriptional level in eukaryotic cells [5]. TFs are regarded as the primary regulators of gene expression and can be either activators or repressors [6, 7]. On the other hand, miRNAs are considered critical in regulatory systems through fine-tuning the expression of target genes and usually suppress target mRNA expression [8, 9]. Previous studies have observed that cooperation between TFs and miRNAs is prevalent within living cells [3]. Accordingly, the various regulatory abilities and the widespread cooperation of TFs and/or miRNAs make the GRN subtle and difficult to investigate. Like other biological networks, the GRN is scale-free: a few nodes in the networks are highly connected [10]. Moreover, it is hierarchical [11, 12]. There are so-called master regulators, which control most of the regulatory information flow in the GRN and have the maximum impact in affecting gene expression [13, 14]. However, a previous study found that the nonmaster regulators in the GRN are more relevant to cell viability [11]. This study also suggested these middle-level regulators, which are the regulators with more than one in-degree, as control bottlenecks, i.e. regulatory bridges, in the hierarchical structure of the GRN. Moreover, the bridgeness of regulations in the GRN has been demonstrated to be influential in cancers [15, 16]. However, how the GRN hierarchy is involved in the cellular system and the roles of these regulatory bridges in regulating biological processes has not been well studied. To address this knowledge gap, we reconstructed the human GRN that constitutes TF and miRNA regulations. We first separated the human GRN into three layers: the constituent core, periphery, and interface core in between. Our analyses found that the interface core could be central and bridge the regulatory information flow in the human GRN, and could be critical to cell viability and cancer. Finally, we identified one feedback loop (FBL) between HNF4A and NR2F2 that possesses the highest bridgeness in the interface core. Furthermore, literature-based evidence supported our network-based results and indicated that the disequilibrium of the HNF4A-NR2F2 FBL might protect tumor cells from intrinsic apoptotic processes. In summary, our study clarified the influence of regulation bridgeness in the human GRN, as well as unveiled a critical role of bridging regulations in cancer. Results Structure of the human GRN The human GRN constructed in this study covers TF and miRNA regulations and is scale-free-like (Supplementary Figure S1). That is, only a few targets are highly regulated, and a few regulators govern a large number of targets. In the network, <10% of biomolecules are regulators and only 0.2% of regulations are regulatory FBLs (Supplementary Table S1). The FBLs are bidirectional regulations formed by two regulators, which regulated each other. In addition, the GRN contains a single strongly connected component (SCC) that covers 1175 (5.71%) TFs/miRNAs and 41 102 (5.44%) regulations (Figure 1A, left panel). To examine if the human GRN was generated randomly, we built two models with 1000 random networks for each. In the first model, we assumed that the regulators randomly recruited the targets with equal probability (Figure 1A, middle panel). We also kept the numbers of regulators and targets the same as those in the constructed GRN. In the second model, the in- and out-degree distributions of random networks were controlled to be identical to the in- and out-degree distributions of the GRN (Figure 1A, right panel). We observed that the probability of forming more than one single SCC is <0.001 and 0.018 in the first and second model, respectively. This observation suggested that the single SCC property might be structural [17]. Even so, the SCC in the constructed GRN is distinguishable from those in the random networks by size. In the first model, all the sizes of the SCCs were 1494, which are significantly larger than the SCC in the constructed GRN (Figure 1B, left table and middle panel). Interestingly, the number of regulated regulators (regulators with in-degree ≥1) in the GRN was 1494. Notably, all the nodes in a SCC are required to be regulated regulators. Therefore, the significantly larger sizes of the SCCs in the random networks from the first model could be an artifact of the number of the regulated regulators in the constructed SCC. On the other hand, the SCC in the GRN was significantly larger than the random ones in the second model (Figure 1B, left table and right panel). This observation showed that the constructed GRN possessed more regulated regulators, i.e. regulators in the middle level, which has been demonstrated to be relevant to cell viability [14]. The above results indicate that the random network structure is incomparable with the constructed GRN. Additionally, we observed that older genes tended to possess higher in- and out-degrees (Figure 1D). The phylostratum data were obtained from Neme et al. [18]. This observation and the scale-free-like characteristic suggested that the evolution of the human GRN could be a preferential attachment process [19]. In other words, older genes could be favored to have a new target or regulator. Accordingly, we could conclude that the human GRN might have evolved, and the regulators could selectively recruit targets or vice versa. As the SCC in the GRN comprised regulators that are naturally enriched in the nucleus, the non-SCC was observed to be enriched in the cytoskeleton (Supplementary Table S2). We then denoted the SCC as the core layer and non-SCC region as the periphery layer in the GRN. Of note, as the FBLs are naturally SCCs, the periphery layer (non-SCC region) contains no FBLs. Figure 1. View largeDownload slide Network characteristics of the human GRN. (A) The schematic diagram of the core structure in the human GRN and random network model. The network could be separated into three layers: the interface core, constituent core and periphery. The interface core (yellow) was formed by all the regulatory FBLs in the networks. The constituent core (red) is the SCC of the network containing regulators only and excluding all the regulatory FBLs. All the other regulations, i.e. not in the interface or constituent core, were categorized to the periphery region (blue, green and gray arrows). Of note, nodes in the periphery form no regulatory loops with either the node in the periphery or the core. Additionally, in the first random model, as no regulated regulators (regulators with in-degree ≥1) were in the periphery, the depth of the periphery is only one. (B) The distribution of the SCC size and (C) the number of FBLs in the random networks. The left table shows the statistical significance of the human GRN compared with the random network models. The P-values in the table were calculated using the Wilcoxon rank-sum test. The red line in each panel indicates the value of the corresponding network feature in the GRN. (D) The phylogenetic association of regulatory degree in the human GRN. The oldest phylogenetic origins of genes are listed at the x-axis from older to younger (left to right). (E) The connectivity and (F) closeness centrality distribution of the nodes participating in the core, FBL and periphery. Figure 1. View largeDownload slide Network characteristics of the human GRN. (A) The schematic diagram of the core structure in the human GRN and random network model. The network could be separated into three layers: the interface core, constituent core and periphery. The interface core (yellow) was formed by all the regulatory FBLs in the networks. The constituent core (red) is the SCC of the network containing regulators only and excluding all the regulatory FBLs. All the other regulations, i.e. not in the interface or constituent core, were categorized to the periphery region (blue, green and gray arrows). Of note, nodes in the periphery form no regulatory loops with either the node in the periphery or the core. Additionally, in the first random model, as no regulated regulators (regulators with in-degree ≥1) were in the periphery, the depth of the periphery is only one. (B) The distribution of the SCC size and (C) the number of FBLs in the random networks. The left table shows the statistical significance of the human GRN compared with the random network models. The P-values in the table were calculated using the Wilcoxon rank-sum test. The red line in each panel indicates the value of the corresponding network feature in the GRN. (D) The phylogenetic association of regulatory degree in the human GRN. The oldest phylogenetic origins of genes are listed at the x-axis from older to younger (left to right). (E) The connectivity and (F) closeness centrality distribution of the nodes participating in the core, FBL and periphery. It is worth noting that when comparing the two random models, the human GRN possessed significantly more FBLs (Figure 1C). This result and the significantly larger SCC suggested that the regulators in the human GRN might tend to regulate each other directly or indirectly (regulatory path length ≥1). We then applied a connectivity measurement to investigate the roles of the FBLs in the core structure. This connectivity was calculated by dividing the degree of a gene/miRNA by the number of connections between the core and the periphery layer. Interestingly, TFs/miRNAs forming FBLs possessed significantly higher connectivity than other genes/miRNAs in the human GRN (Figure 1E). That is, the TFs/miRNAs forming FBLs play pivotal roles in connecting the core and the periphery. Additionally, the FBL TFs/miRNAs possess significantly higher closeness centrality (Figure 1F). This result implies that the FBL TFs/miRNAs might also be the center of the GRN. Briefly, the FBLs could compose an interface-like structure between core and periphery and act as the regulatory center of the GRN. Therefore, we further separated the core layer into two sublayers: the interface core formed by the FBLs and the constituent core formed by other regulations in the core. In summary, we separated the human GRN into three layers: the interface core, constituent core and periphery. Furthermore, the structure of the human GRN could be evolved, and thus, its structural characteristics and the embedded biological features were worth discovering. The interface core bridges the human GRN In the network, the regulator–target edges are dead ends, while the regulator–regulator edges are the only connections that can be passed through. Thus, the fullness of regulator–regulator edges could lead the core to bridge the GRN. In graph theory, a bridge is an edge whose removal breaks the graph into several connected components. Thus, the removal of a bridge decreases the connectivity of a network. In a network, edges with higher edge betweenness centrality (eBC) control a larger number of the shortest paths and therefore could act as bridges. Indeed, the removal of the edges following a descending order of directed eBC rapidly disrupted the GRN (Supplementary Figure S2). Therefore, we used directed eBC to measure the bridgeness of an edge (regulation) in this study. As expected, the regulations in the core, including the interface and constituent cores, possessed significantly higher eBC than those in the periphery (Figure 2A and Supplementary Table S3). The same phenomenon was observed in the random networks with controlled in- and out-degree distribution (Supplementary Figure S3). Accordingly, the high bridgeness of the core could be attributed to its structural influence: the fullness of regulator–regulator edges. Therefore, the core could bridge a graph naturally. Furthermore, the bridging role might lead the core to control the major part of the regulatory information flow in the human GRN. Figure 2. View largeDownload slide Bridging roles of the constituent and interface cores in the human GRN. (A) The eBC distribution of regulations in the human GRN. The regulations in the periphery region were further categorized into two types: RR, formed by two regulators; RT, formed by one regulator and one target. The proportion of disconnected shortest paths after removing the constituent and interface cores and periphery (RR) in the human GRN compared with (B) random sampling and (C) the random network. The histograms show the distribution of the proportion of disconnected shortest paths as doing (B) random sampling and (C) in random networks. The vertical lines represent the value of the human GRN. Figure 2. View largeDownload slide Bridging roles of the constituent and interface cores in the human GRN. (A) The eBC distribution of regulations in the human GRN. The regulations in the periphery region were further categorized into two types: RR, formed by two regulators; RT, formed by one regulator and one target. The proportion of disconnected shortest paths after removing the constituent and interface cores and periphery (RR) in the human GRN compared with (B) random sampling and (C) the random network. The histograms show the distribution of the proportion of disconnected shortest paths as doing (B) random sampling and (C) in random networks. The vertical lines represent the value of the human GRN. To further discover the bridging function of the core, we removed the core and then recalculated the shortest paths between all the genes and miRNAs. Indeed, after removing the constituent or interface core, a significantly larger proportion of paths were disconnected, i.e. no path could be found between genes/miRNAs, compared with randomly removing the same amount of regulator–regulator edges in the constituent or interface core (Figure 2B, Supplementary Table S4). Of note, the removal of regulator–regulator edges in the periphery also disconnected a significantly larger proportion of paths (Figure 2B, Supplementary Table S4). Interestingly, when comparing with the random networks that controlled the in- and out-degree distribution, only the interface core remained significant in disconnecting regulatory paths (Figure 2C, Supplementary Table S4). These observations imply that the high bridgeness of the constituent core might be a result of its structural importance, i.e. the regulator–regulator edges are the connections that can be passed through. Moreover, this result demonstrated that the bridgeness of the interface core may not only be attributed to its structural influence but also to other pivotal roles in the GRN. Moreover, it recapitulates the roles of the interface core in bridging between the core and the periphery of the human GRN. However, removal of the interface core only disconnected 5.72% of biomolecule pairs in the GRN. Thus, there might be alternative regulatory paths that could keep the network structure unbroken when the interface core was attacked. A possible scenario is that the interface-controlled regulatory paths might be critical, and these alternative paths might be required to increase the robustness of the human GRN. Previous studies have demonstrated that FBLs could play essential roles in critical biological processes, such as differentiation and cell cycle transitions [20, 21]. Accordingly, these studies supported the proposed scenario for the existing alternatives of the interface-controlled paths and may emphasize the importance of the interface core in the human GRN. Nevertheless, the biological role of the interface core still needs to be studied. The interface core is influential in cell viability and cancer Even though the FBLs in the human GRN are distinguishable from those in the random networks, the FBLs are natural bridges in the graph. Therefore, to investigate the relevance of the interface core to human cells, we collected a set of cancer-associated biomolecules, including genes and miRNAs, and a set of essential human genes. It is worth noting that the proportion of overlapping genes between the essential genes and the cancer-associated genes was only 8% (calculated using the Jaccard index). This observation suggested that the two data sets are nearly independent. Then, we compiled a set of cancer-associated regulations linked by two cancer-associated biomolecules and a set of essential regulations formed by at least one essential gene. We found that the cancer-associated and essential regulations possessed significantly higher eBC than the non-cancer-associated and nonessential regulations, respectively (Supplementary Table S5). Moreover, both of the cancer-associated and essential regulation proportions increased with eBC in ascending order (cancer: P-value = 1.33 × 10−6, χ2 test, Figure 3A; essential: P-value < 2.2 × 10−16, χ2 test, Figure 3B). These results suggested that the bridges could be more cancer-associated and essential compared with the other regulations, and the eBC might be capable of revealing the important regulations involved in cancer and cell viability. Figure 3. View largeDownload slide Cancer association and essentiality of eBC and the core in the human GRN. The association between eBC and the proportion of the (A) cancer-associated and (B) essential regulation. The colored circles represent the significance derived from Fisher’s exact test. The red and green circles show significantly overrepresented and underrepresented, respectively, and the gray show nonsignificant. (C) The enrichment of cancer-associated and essential biomolecules and regulations in the four regions. Asterisks show the significance of enrichment. For each bar, two asterisks represent the significance when the whole human GRN (left) and only regulators (right) were used as a reference. Red asterisks are significantly overrepresented, green underrepresented and gray insignificant. The periphery (RT) is marked by only the left asterisk, i.e. using the whole human GRN as the reference. Figure 3. View largeDownload slide Cancer association and essentiality of eBC and the core in the human GRN. The association between eBC and the proportion of the (A) cancer-associated and (B) essential regulation. The colored circles represent the significance derived from Fisher’s exact test. The red and green circles show significantly overrepresented and underrepresented, respectively, and the gray show nonsignificant. (C) The enrichment of cancer-associated and essential biomolecules and regulations in the four regions. Asterisks show the significance of enrichment. For each bar, two asterisks represent the significance when the whole human GRN (left) and only regulators (right) were used as a reference. Red asterisks are significantly overrepresented, green underrepresented and gray insignificant. The periphery (RT) is marked by only the left asterisk, i.e. using the whole human GRN as the reference. Notably, the core possessed significantly higher eBC in the GRN. Thus, the constituent and interface cores might be highly associated with cancer and cell viability. Indeed, compared with the whole GRN, the cancer-associated biomolecules were significantly enriched in the interface core as well as in the constituent core (Figure 3C and Supplementary Table S6). Consistently, the essential genes were significantly enriched in the interface and constituent cores (Figure 3C and Supplementary Table S6). These observations suggested that both the constituent and interface cores might be influential in cancer and cell viability. However, the cancer-associated biomolecules and essential genes were significantly enriched in regulators (Supplementary Figure S4). Thus, the enrichment of cancer-associated biomolecules and essential genes in the core might be biased by its fullness of regulators. Indeed, compared with all the regulators, the cancer-associated biomolecules become significantly underrepresented, and enrichment of the essential genes became insignificant in the constituent core (Figure 3C and Supplementary Table S6); only the interface core was still significantly enriched with the cancer-associated biomolecules and essential genes (Figure 3C and Supplementary Table S6). On the other hand, compared with the whole human GRN, the interface and constituent cores were significantly enriched with the cancer-associated and essential regulations (Figure 3C and Supplementary Table S6). Likely, the cancer-associated and essential regulations were significantly overrepresented in the interface and constituent cores by using the regulator–regulator edges as the reference (Figure 3C and Supplementary Table S6). Notably, the interface core contains a higher proportion of cancer-associated and essential biomolecules and regulations than the constituent core. Additionally, the enrichment of the interface core is not biased by the fullness of regulators. Furthermore, the cancer-associated biomolecules were significantly overrepresented in the interface core when using the core as the reference (P-value = 4.21 × 10−25, Fisher’s exact test). The essential genes were also enriched in the interface with moderate significance (P-value = 0.10, Fisher’s exact test) when using the core as the reference. The enrichment of cancer-associated and essential biomolecules and regulations emphasized that the interface core could be critical in human cells. These results further imply that a disturbance to the interface core might result in more damage to the human cells compared with the constituent core. Moreover, these observations recapitulate the scenario explaining the existence of the alternative regulatory paths accompanying the interface-controlled paths. In summary, the interface core is not only pivotal to the structure of the human GRN but might also be influential to cancer and cell viability. These results further emphasized the pertinence of the network properties to the human GRN and the potentially critical role of the core in cancer and normal cells. Regulatory role of the interface core in cancers To embody the regulatory function of the interface core in cancer, we investigated the FBL between HNF4A and NR2F2, which possesses the highest eBC in the interface core. HNF4A has been reported to be downregulated and act as a critical transcriptional regulator in hepatocellular carcinoma (HCC) [22, 23], while the upregulation of HNF4A is reportedly a key event in gastric cancer [24]. Previous studies also found that dysregulation of HNF4A is influential to various cancers [25–28]. NR2F2 has also been found to be important in various cancer types, including prostate, colorectal and breast cancer [29–31]. These results suggested that HNF4A and NR2F2 could play critical roles in carcinogenesis. However, the effects of dysregulating the HNF4A-NR2F2 FBL in cancer have not been well studied. In The Cancer Genome Atlas (TCGA) data set, HNF4A was significantly upregulated in five cancer types and downregulated in four (Figure 4A); NR2F2 is upregulated in breast cancer and downregulated in eight cancer types (Figure 4A). This observation suggested that NR2F2 might be turned off, and HNF4A might dominate and disequilibrate this FBL in the five cancer types. Roughly, two regulators forming an FBL must tightly control each other to keep the equilibrium status of an FBL. The dysregulation of either one of the two regulators could lead to a FBL disequilibrium. For example, the disequilibrium effects of the double-negative FBL between OCT4 and miR-145 have been observed, i.e. OCT4 and miR-145 inhibited each other [32, 33]. When miR-145 dominated the FBL, differentiation of stem cells was initiated, while the domination of OCT4 could inhibit miR-145 and then promote tumorigenesis. Herein, the downstream pathways regulated by NR2F2 might be turned off because of the downregulation of NR2F2 in cancers. The upregulation of HNF4A could dominate the HNF4A-NR2F2 FBL and thus turn on the corresponding downstream pathways in the five cancer types. Therefore, we focused on the downstream pathways regulated by HNF4A to reveal the consequence of disturbing this FBL via investigating the HNF4A-regulated functional modules in the five cancer types. We found that most of the functional modules were apoptosis-related (Supplementary Table S7). However, we did not find any evidence that HNF4A could directly regulate apoptosis. HNF4A might regulate apoptosis through its targeting network. Indeed, HNF4A could promote apoptosis through the regulation of its target miRNA (miR-548p) in HCC [34]. A mouse model study has suggested that HNF4A might be associated with HCC development via the targeting of the Perp, the p53/p63 apoptosis effector gene [35]. The evidence suggested that HNF4A might be involved in the development of the five cancer types through regulating apoptosis. Figure 4. View largeDownload slide HNF4A-NR2F2 FBL regulation in cancers. (A) The log2 fold ratio of HNF4A and NR2F2 expression in cancers. The bold red and green numbers represent the significantly overexpressed and underexpressed, respectively, with the gray denoting nonsignificant expression. (B) The differential expression patterns of the HNF4A-regulated apoptotic module. We colored HNF4A target genes with red and framed the Bcl2 family with a green rectangle. (C) The hypothesis of the downstream effect caused by the loss of equilibrium mediated by the HNF4A-NR2F2 FBL. The potential molecular mechanism describes how the domination of HNF4A in the regulatory FBL between HNF4A and NR2F2 protect cancer from the apoptotic process. The mitochondrial apoptosis-induced channel (MAC) and cytochrome c are potentially inhibited in this hypothesis. Figure 4. View largeDownload slide HNF4A-NR2F2 FBL regulation in cancers. (A) The log2 fold ratio of HNF4A and NR2F2 expression in cancers. The bold red and green numbers represent the significantly overexpressed and underexpressed, respectively, with the gray denoting nonsignificant expression. (B) The differential expression patterns of the HNF4A-regulated apoptotic module. We colored HNF4A target genes with red and framed the Bcl2 family with a green rectangle. (C) The hypothesis of the downstream effect caused by the loss of equilibrium mediated by the HNF4A-NR2F2 FBL. The potential molecular mechanism describes how the domination of HNF4A in the regulatory FBL between HNF4A and NR2F2 protect cancer from the apoptotic process. The mitochondrial apoptosis-induced channel (MAC) and cytochrome c are potentially inhibited in this hypothesis. To confirm this implication, we investigated the genes involved in the identified HNF4A-regulated apoptotic functional module in the five cancer types (Figure 4B). In the module, 8 of 21 members were genes encoding BCL2 family proteins, i.e. apoptosis regulator Bcl-2 [36]. Among these eight genes, two pro-apoptotic genes, BAK1 and BID, were upregulated by HNF4A. These observations suggested that human cells might initiate apoptosis to respond to carcinogenesis. However, carcinogenesis is known to be because of excessive cell proliferation and decreased cell death, i.e. inappropriate inhibition of apoptosis [37]. Interestingly, there are four genes, VIL1, HSPD1, SRC and TRAF2, upregulated by HNF4A. VIL1 encodes the vilin protein and has been reported to inhibit apoptosis through maintaining mitochondrial integrity [38]. BAK1 could initiate apoptosis via increasing mitochondrial outer membrane permeability to release cytochrome c and other proapoptotic factors [39]. Therefore, the over-expression of VIL1 could inhibit the BAK1-initiated apoptosis process. Additionally, the previous study has demonstrated that SRC could be capable of inhibiting the Bax/Bak activation cascade through increasing the activation threshold of Bax [40]. Furthermore, HSPD1 has been found to interact with Bax and Bak in vitro, and this interaction may block the proapoptotic ability of the Bax/Bak complex [41]. Accordingly, VIL1, SRC and HSPD1 might cooperate to promote carcinogenesis via inhibiting Bax/Bak-initiated apoptotic processes. Moreover, TRAF2 has been reported to protect cancer cells from ER stress-induced apoptosis [42], to suppress apoptotic processes through regulating tumor necrosis factor and nuclear factor kappa-light-chain-enhancer of activated B cells (NF-κB) signaling [43], and to inhibit TRAIL- and CD95L-induced apoptosis [44]. Briefly, HNF4A may activate both antiapoptotic and proapoptotic processes simultaneously during cancer development. That is, HNF4A might possess a dual role, i.e. antiapoptotic and proapoptotic, in carcinogenesis of the five cancer types. Based on the results above, we proposed a hypothesis on how the overexpressed HNF4A promoted carcinogenesis of the five cancer types (Figure 4C). During the development of the five cancer types, the protective mechanism of human cells might highly express HNF4A to activate proapoptotic genes BAK1 and BID for the initiation of intrinsic apoptosis. However, the upregulation of HNF4A may also upregulate the expression of VIL1, SRC and HSPD1, which could repress the initiation of apoptosis either by maintaining mitochondrial membrane integrity [38] or obstructing the formation of the Bax/Bak complex [40, 41]. Moreover, HNF4A also activated TRAF2, which could inhibit apoptosis upstream of the apoptotic pathway [44]. Finally, the tumor cells were able to survive Bax/Bak-triggered intrinsic apoptosis. Collectively, our findings suggest that the domination of HNF4A might disequilibrate the FBL between HNF4A and NR2F2 to allow tumor cells to survive apoptosis and even to promote carcinogenesis of the five cancer types. However, more experiments are needed to validate our hypothesis. Discussion In this study, we investigated the structure of the human GRN and discovered the critical regulations of cancer through analyzing eBC of regulations. We observed that the human GRN structure is distinguishable from the random ones and thus could be evolved rather than randomly generated. However, the human GRN structure was strongly affected by the nature of the gene regulatory system—the proportion of regulators is small, but they dominate all (thousands) of the regulations. In other words, the averaged out-degree of regulators becomes much higher than randomly expected (408 versus 37). For example, this nature could make the network, with high probability, form a single SCC with regulators as critical nodes inside the network, such as hubs and bottlenecks. Accordingly, the biological significance of structurally critical nodes inside the human GRN becomes a paradox—it is hard to answer if the significance of structurally critical nodes was attributed to biology or graph structure. Nevertheless, as the regulators were significantly enriched with the essential and cancer-associated biomolecules (Supplementary Figure S4), the SCC could be influential in a cellular system because of the fact that all the nodes in the SCC are regulators. Moreover, the single SCC characteristic could be derived from the nature of the regulatory system, the high average out-degree of the regulators [17]. Accordingly, the single SCC characteristic might be biologically relevant. However, further studies are needed to justify if the single SCC characteristic is biologically derived or not. On the other hand, we found that the human GRN could be naturally bridged by the SCC, including the constituent and interface cores. To further clarify the bridge role of the SCC, we investigated the loss rate of the shortest paths after removing the SCC. Interestingly, only the removal of the interface core caused a significantly higher loss rate of the shortest paths than the random networks with controlled degree distribution. In other words, the interface core in the human GRN occupied significantly more shortest paths than the random interface cores did. Of note, in the random networks, the interface core also possessed the highest eBC and could bridge the random networks. This result implied that the interface core could be global, whereas the random interface core could be local [45]. The high betweenness centrality of local bridges is contributed by their high degree that makes local bridges control lots of paths between neighbors [45]. Indeed, the random interface core possessed higher degree than the interface core in the constructed human GRN (Supplementary Figure S5). These observations demonstrated that the random interface cores could bridge the networks because of their high degree caused by the nature of the gene regulatory system. In contrast, the interface core in the human GRN could further act as a global bridge by occupying the longer shortest paths. In addition, our results suggested that the interface core could be influential in cell viability and cancer without biasing from the fullness of regulators. Briefly, we observed that (1) eBC is positively correlated with the proportion of essential and cancer-associated regulations, (2) the interface core possessed the highest eBC and (3) the interface core was significantly enriched with essential and cancer-associated biomolecules and regulations. Accordingly, we could conclude that the biological significance of the interface core might be attributed to its highest eBC. However, we did not observe that the essential and cancer-associated biomolecules and regulations were significantly underrepresented in the regulations with low eBC in the interface core. This result might be attributed to the relatively high eBC of the interface core in the human GRN (Figure 2A). In other words, there is no regulation in the interface core possessing relatively low eBC. Consequently, other analyses are required to robustly demonstrate this negative association that the biological significance of the interface core is attributed to its bridgeness. Conclusions In summary, we observed that the regulatory FBLs form an interface-like structure between the constituent core and periphery region in the human GRN. We then denoted the FBLs as the interface core in the human GRN. Further investigation suggested that the interface core could not only be influential to the structure of the human GRN but also crucial to cancer and cell viability. Moreover, we discovered the regulatory HNF4A-NR2F2 FBL possesses the highest eBC in the interface core. Advanced analysis showed that HNF4A might possess a dual role of antiapoptosis and proapoptosis during carcinogenesis and suggested that the disturbance of this FBL might protect tumor cells from apoptotic elimination. Methods For a detailed description of the constructed GRN, cancer-associated biomolecules, essential genes and mRNA and miRNA expression profiles from TCGA, see Supplementary SI Materials and Methods. Identification of regulatory bridges in the human GRN In this study, we applied eBC to evaluate the influence of one regulation on the structure of GRN, the sensitivity of the GRN structure to the removal of the given regulation. The eBC of an edge, e, is the sum of the proportion of all-pairs of shortest paths passing through e and can be calculated as below: BCe=∑i,j∈VSeS, (1) where BCe is the eBC of edge, e; V is the set of nodes in the GRN; S is the number of all-pairs shortest paths; and Se is the number of shortest paths passing through e. All network properties, such as SCCs and closeness centrality, in this study were calculated by python package NetworkX. Discovery of HNF4A-regulated functional modules To discover the HNF4A-regulated downstream functional modules, we collected HNF4A target genes that are (1) significantly positively coexpressed with HNF4A and (2) significantly upregulated in the cancer types in which HNF4A was also upregulated. Through these two conditions, we obtained those genes potentially activated by HNF4A in cancer. The detailed definition of these two conditions is specified in the Supplementary SI Materials and Methods. To discover HNF4A regulation in more detail, we recruited the first neighbors of the selected HNF4A target genes in the human protein interaction network (PIN). The protein–protein interactions (PPI) were obtained from the PIN Analysis v2 [46]. We then denoted the network constructed by the selected HNF4A target genes, their protein interacting partners and PPIs among them as the HNF4A-regulated network. Next, we performed enrichment analysis to determine the functions involved in the identified HNF4A-regulated network. Of note, we conducted the functional enrichment analysis using two methods, conventional and network-wise. With the conventional method, the overrepresentation of selected HNF4A target genes and their PPI partners defines the significance of HNF4A-regulated functions. On the other hand, the network-wise enrichment analysis evaluates the significance of HNF4A-regulated functions through the overrepresentation of functional PPIs among selected HNF4A targets. The procedure of functional enrichment analysis is described in the Supplementary SI Materials and Methods in detail. To preserve the HNF4A regulation in the identified functions, we calculated the enrichment significance of the selected HNF4A targets in each functional module using Fisher’s exact test. Finally, we considered those functions passing all three tests (P-value < 0.05) as potential HNF4A-regulated functional modules. Key Points The regulatory FBLs inside the core form an interface-like structure between the constituent core and the periphery region, which also act as the regulatory center in the human GRN. The regulatory FBLs cover a significantly higher proportion of the cancer-associated and essential regulation and biomolecules than the constituent core and periphery region. The regulatory FBLs are not only influential for the network structure but are also critical in cellular systems. The disturbance of the regulatory FBL between HNF4A and NR2F2 might protect tumor cells from intrinsic apoptotic elimination. Supplementary Data Supplementary data are available online at https://academic.oup.com/bib. Funding This work was supported by the Ministry of Science and Technology (grant number MOST 105-2628-E-010-001-MY3 and MOST 104-2320-B-010-037-). Yun-Ru Chen is a PhD student in the Institute of Biomedical Informatics, National Yang-Ming University. Hsuan-Cheng Huang is a professor in the Institute of Biomedical Informatics, National Yang-Ming University. Chen-Ching Lin is an assistant professor in the Institute of Biomedical Informatics, National Yang-Ming University. References 1 Peter IS , Davidson EH. Evolution of gene regulatory networks controlling body plan development . Cell 2011 ; 144 ( 6 ): 970 – 85 . http://dx.doi.org/10.1016/j.cell.2011.02.017 Google Scholar CrossRef Search ADS PubMed 2 Lander AD. How cells know where they are . Science 2013 ; 339 ( 6122 ): 923 – 7 . http://dx.doi.org/10.1126/science.1224186 Google Scholar CrossRef Search ADS PubMed 3 Shalgi R , Lieber D , Oren M , et al. Global and local architecture of the mammalian microRNA-transcription factor regulatory network . PLoS Comput Biol 2007 ; 3 ( 7 ): e131 . Google Scholar CrossRef Search ADS PubMed 4 Latchman DS. Transcription factors: an overview . Int J Biochem Cell Biol 1997 ; 29 ( 12 ): 1305 – 12 . http://dx.doi.org/10.1016/S1357-2725(97)00085-X Google Scholar CrossRef Search ADS PubMed 5 Guo H , Ingolia NT , Weissman JS , et al. Mammalian microRNAs predominantly act to decrease target mRNA levels . Nature 2010 ; 466 ( 7308 ): 835 – 40 . http://dx.doi.org/10.1038/nature09267 Google Scholar CrossRef Search ADS PubMed 6 Lee TI , Young RA. Transcription of eukaryotic protein-coding genes . Annu Rev Genet 2000 ; 34 : 77 – 137 . http://dx.doi.org/10.1146/annurev.genet.34.1.77 Google Scholar CrossRef Search ADS PubMed 7 Shen-Orr SS , Milo R , Mangan S , et al. Network motifs in the transcriptional regulation network of Escherichia coli . Nat Genet 2002 ; 31 ( 1 ): 64 – 8 . http://dx.doi.org/10.1038/ng881 Google Scholar CrossRef Search ADS PubMed 8 Bartel DP. MicroRNAs: genomics, biogenesis, mechanism, and function . Cell 2004 ; 116 ( 2 ): 281 – 97 . http://dx.doi.org/10.1016/S0092-8674(04)00045-5 Google Scholar CrossRef Search ADS PubMed 9 Harfe BD. MicroRNAs in vertebrate development . Curr Opin Genet Dev 2005 ; 15 ( 4 ): 410 – 15 . http://dx.doi.org/10.1016/j.gde.2005.06.012 Google Scholar CrossRef Search ADS PubMed 10 Babu MM , Luscombe NM , Aravind L , et al. Structure and evolution of transcriptional regulatory networks . Curr Opin Struct Biol 2004 ; 14 ( 3 ): 283 – 91 . http://dx.doi.org/10.1016/j.sbi.2004.05.004 Google Scholar CrossRef Search ADS PubMed 11 Yu H , Gerstein M. Genomic analysis of the hierarchical structure of regulatory networks . Proc Natl Acad Sci USA 2006 ; 103 ( 40 ): 14724 – 31 . http://dx.doi.org/10.1073/pnas.0508637103 Google Scholar CrossRef Search ADS PubMed 12 Bhardwaj N , Yan KK , Gerstein MB. Analysis of diverse regulatory networks in a hierarchical context shows consistent tendencies for collaboration in the middle levels . Proc Natl Acad Sci USA 2010 ; 107 ( 15 ): 6841 – 6 . http://dx.doi.org/10.1073/pnas.0910867107 Google Scholar CrossRef Search ADS PubMed 13 Ihmels J , Levy R , Barkai N. Principles of transcriptional control in the metabolic network of Saccharomyces cerevisiae . Nat Biotechnol 2004 ; 22 ( 1 ): 86 – 92 . http://dx.doi.org/10.1038/nbt918 Google Scholar CrossRef Search ADS PubMed 14 Yu H , Luscombe NM , Qian J , et al. Genomic analysis of gene expression relationships in transcriptional regulatory networks . Trends Genet 2003 ; 19 ( 8 ): 422 – 7 . http://dx.doi.org/10.1016/S0168-9525(03)00175-6 Google Scholar CrossRef Search ADS PubMed 15 Kotlyar M , Fortney K , Jurisica I. Network-based characterization of drug-regulated genes, drug targets, and toxicity . Methods 2012 ; 57 ( 4 ): 499 – 507 . http://dx.doi.org/10.1016/j.ymeth.2012.06.003 Google Scholar CrossRef Search ADS PubMed 16 Azevedo H , Moreira-Filho CA. Topological robustness analysis of protein interaction networks reveals key targets for overcoming chemotherapy resistance in glioma . Sci Rep 2015 ; 5 ( 1 ): 16830 . http://dx.doi.org/10.1038/srep16830 Google Scholar CrossRef Search ADS PubMed 17 Bollobás B , Frieze AM. On matchings and Hamiltonian cycles in random graphs . North Holland Math Stud 1985 ; 118 : 23 – 46 . Google Scholar CrossRef Search ADS 18 Neme R , Tautz D. Phylogenetic patterns of emergence of new genes support a model of frequent de novo evolution . BMC Genomics 2013 ; 14 : 117 . http://dx.doi.org/10.1186/1471-2164-14-117 Google Scholar CrossRef Search ADS PubMed 19 Barabasi AL , Albert R. Emergence of scaling in random networks . Science 1999 ; 286 ( 5439 ): 509 – 12 . http://dx.doi.org/10.1126/science.286.5439.509 Google Scholar CrossRef Search ADS PubMed 20 Wang F , Zhu Y , Guo L , et al. A regulatory circuit comprising GATA1/2 switch and microRNA-27a/24 promotes erythropoiesis . Nucleic Acids Res 2014 ; 42 ( 1 ): 442 – 57 . http://dx.doi.org/10.1093/nar/gkt848 Google Scholar CrossRef Search ADS PubMed 21 Ma Y , Wang B , Jiang F , et al. A feedback loop consisting of microRNA 23a/27a and the beta-like globin suppressors KLF3 and SP1 regulates globin gene expression . Mol Cell Biol 2013 ; 33 ( 20 ): 3994 – 4007 . http://dx.doi.org/10.1128/MCB.00623-13 Google Scholar CrossRef Search ADS PubMed 22 Lazarevich NL , Cheremnova OA , Varga EV , et al. Progression of HCC in mice is associated with a downregulation in the expression of hepatocyte nuclear factors . Hepatology 2004 ; 39 ( 4 ): 1038 – 47 . http://dx.doi.org/10.1002/hep.20155 Google Scholar CrossRef Search ADS PubMed 23 Ning BF , Ding J , Yin C , et al. Hepatocyte nuclear factor 4 alpha suppresses the development of hepatocellular carcinoma . Cancer Res 2010 ; 70 ( 19 ): 7640 – 51 . http://dx.doi.org/10.1158/0008-5472.CAN-10-0824 Google Scholar CrossRef Search ADS PubMed 24 Chang HR , Nam S , Kook MC , et al. HNF4alpha is a therapeutic target that links AMPK to WNT signalling in early-stage gastric cancer . Gut 2016 ; 65 ( 1 ): 19 – 32 . http://dx.doi.org/10.1136/gutjnl-2014-307918 Google Scholar CrossRef Search ADS PubMed 25 Vuong LM , Chellappa K , Dhahbi JM , et al. Differential effects of hepatocyte nuclear factor 4alpha isoforms on tumor growth and T-Cell factor 4/AP-1 interactions in human colorectal cancer cells . Mol Cell Biol 2015 ; 35 ( 20 ): 3471 – 90 . http://dx.doi.org/10.1128/MCB.00030-15 Google Scholar CrossRef Search ADS PubMed 26 Koizume S , Yokota N , Miyagi E , et al. Hepatocyte nuclear factor-4-independent synthesis of coagulation factor VII in breast cancer cells and its inhibition by targeting selective histone acetyltransferases . Mol Cancer Res 2009 ; 7 ( 12 ): 1928 – 36 . http://dx.doi.org/10.1158/1541-7786.MCR-09-0372 Google Scholar CrossRef Search ADS PubMed 27 Lucas B , Grigo K , Erdmann S , et al. HNF4alpha reduces proliferation of kidney cells and affects genes deregulated in renal cell carcinoma . Oncogene 2005 ; 24 ( 42 ): 6418 – 31 . http://dx.doi.org/10.1038/sj.onc.1208794 Google Scholar CrossRef Search ADS PubMed 28 Snyder EL , Watanabe H , Magendantz M , et al. Nkx2-1 represses a latent gastric differentiation program in lung adenocarcinoma . Mol Cell 2013 ; 50 ( 2 ): 185 – 99 . http://dx.doi.org/10.1016/j.molcel.2013.02.018 Google Scholar CrossRef Search ADS PubMed 29 Qin J , Wu SP , Creighton CJ , et al. COUP-TFII inhibits TGF-beta-induced growth barrier to promote prostate tumorigenesis . Nature 2013 ; 493 ( 7431 ): 236 – 40 . Google Scholar CrossRef Search ADS PubMed 30 Wang C , Zhou Y , Ruan R , et al. High expression of COUP-TF II cooperated with negative Smad4 expression predicts poor prognosis in patients with colorectal cancer . Int J Clin Exp Pathol 2015 ; 8 ( 6 ): 7112 – 21 . Google Scholar PubMed 31 Al-Rayyan N , Litchfield LM , Ivanova MM , et al. 5-Aza-2-deoxycytidine and trichostatin A increase COUP-TFII expression in antiestrogen-resistant breast cancer cell lines . Cancer Lett 2014 ; 347 ( 1 ): 139 – 50 . http://dx.doi.org/10.1016/j.canlet.2014.02.001 Google Scholar CrossRef Search ADS PubMed 32 Chivukula RR , Mendell JT. Abate and switch: miR-145 in stem cell differentiation . Cell 2009 ; 137 ( 4 ): 606 – 8 . http://dx.doi.org/10.1016/j.cell.2009.04.059 Google Scholar CrossRef Search ADS PubMed 33 Xu N , Papagiannakopoulos T , Pan G , et al. MicroRNA-145 regulates OCT4, SOX2, and KLF4 and represses pluripotency in human embryonic stem cells . Cell 2009 ; 137 ( 4 ): 647 – 58 . http://dx.doi.org/10.1016/j.cell.2009.02.038 Google Scholar CrossRef Search ADS PubMed 34 Hu XM , Yan XH , Hu YW , et al. miRNA-548p suppresses hepatitis B virus X protein associated hepatocellular carcinoma by downregulating oncoprotein hepatitis B x-interacting protein . Hepatol Res 2016 ; 46 ( 8 ): 804 – 15 . http://dx.doi.org/10.1111/hepr.12618 Google Scholar CrossRef Search ADS PubMed 35 Bonzo JA , Ferry CH , Matsubara T , et al. Suppression of hepatocyte proliferation by hepatocyte nuclear factor 4alpha in adult mice . J Biol Chem 2012 ; 287 ( 10 ): 7345 – 56 . http://dx.doi.org/10.1074/jbc.M111.334599 Google Scholar CrossRef Search ADS PubMed 36 Chao DT , Korsmeyer SJ. BCL-2 family: regulators of cell death . Annu Rev Immunol 1998 ; 16 : 395 – 419 . http://dx.doi.org/10.1146/annurev.immunol.16.1.395 Google Scholar CrossRef Search ADS PubMed 37 Fernald K , Kurokawa M. Evading apoptosis in cancer . Trends Cell Biol 2013 ; 23 ( 12 ): 620 – 33 . http://dx.doi.org/10.1016/j.tcb.2013.07.006 Google Scholar CrossRef Search ADS PubMed 38 Wang Y , Srinivasan K , Siddiqui MR , et al. A novel role for villin in intestinal epithelial cell survival and homeostasis . J Biol Chem 2008 ; 283 ( 14 ): 9454 – 64 . http://dx.doi.org/10.1074/jbc.M707962200 Google Scholar CrossRef Search ADS PubMed 39 Wei MC , Zong WX , Cheng EH , et al. Proapoptotic BAX and BAK: a requisite gateway to mitochondrial dysfunction and death . Science 2001 ; 292 ( 5517 ): 727 – 30 . http://dx.doi.org/10.1126/science.1059108 Google Scholar CrossRef Search ADS PubMed 40 Lopez J , Hesling C , Prudent J , et al. Src tyrosine kinase inhibits apoptosis through the Erk1/2- dependent degradation of the death accelerator Bik . Cell Death Differ 2012 ; 19 ( 9 ): 1459 – 69 . http://dx.doi.org/10.1038/cdd.2012.21 Google Scholar CrossRef Search ADS PubMed 41 Kirchhoff SR , Gupta S , Knowlton AA. Cytosolic heat shock protein 60, apoptosis, and myocardial injury . Circulation 2002 ; 105 ( 24 ): 2899 – 904 . http://dx.doi.org/10.1161/01.CIR.0000019403.35847.23 Google Scholar CrossRef Search ADS PubMed 42 Habelhah H , Zhang L , Xialikaer A , et al. Abstract 3496: TRAF2 protects mammary epithelial and cancer cells from endoplasmic reticulum stress-induced apoptosis . Cancer Res 2016 ; 76(Suppl 14) : 3496 . Google Scholar CrossRef Search ADS 43 Etemadi N , Chopin M , Anderton H , et al. TRAF2 regulates TNF and NF-kappaB signalling to suppress apoptosis and skin inflammation independently of Sphingosine kinase 1 . Elife 2015 ; 4 : e10592 . Google Scholar CrossRef Search ADS PubMed 44 Karl I , Jossberger-Werner M , Schmidt N , et al. TRAF2 inhibits TRAIL- and CD95L-induced apoptosis and necroptosis . Cell Death Dis 2014 ; 5 : e1444 . Google Scholar CrossRef Search ADS PubMed 45 Jensen P , Morini M , Karsai M , et al. Detecting global bridges in networks . J Complex Netw 2016 ; 4 ( 3 ): 319 – 29 . http://dx.doi.org/10.1093/comnet/cnv022 Google Scholar CrossRef Search ADS 46 Cowley MJ , Pinese M , Kassahn KS , et al. PINA v2.0: mining interactome modules . Nucleic Acids Res 2012 ; 40 : D862 – 5 . Google Scholar CrossRef Search ADS PubMed © The Author 2017. Published by Oxford University Press. All rights reserved. For Permissions, please email: journals.permissions@oup.com http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Briefings in Bioinformatics Oxford University Press

Regulatory feedback loops bridge the human gene regulatory network and regulate carcinogenesis

Loading next page...
 
/lp/ou_press/regulatory-feedback-loops-bridge-the-human-gene-regulatory-network-and-62pRK2s8eu
Publisher
Oxford University Press
Copyright
© The Author 2017. Published by Oxford University Press. All rights reserved. For Permissions, please email: journals.permissions@oup.com
ISSN
1467-5463
eISSN
1477-4054
D.O.I.
10.1093/bib/bbx166
Publisher site
See Article on Publisher Site

Abstract

Abstract The development of disease involves a systematic disturbance inside cells and is associated with changes in the interactions or regulations among genes forming biological networks. The bridges inside a network are critical in shortening the distances between nodes. We observed that, inside the human gene regulatory network, one strongly connected core bridged the whole network. Other regulations outside the core formed a weakly connected component surrounding the core like a peripheral structure. Furthermore, the regulatory feedback loops (FBLs) inside the core compose an interface-like structure between the core and periphery. We then denoted the regulatory FBLs as the interface core. Notably, both the cancer-associated and essential biomolecules and regulations were significantly overrepresented in the interface core. These results implied that the interface core is not only critical for the network structure but central in cellular systems. Furthermore, the enrichment of the cancer-associated and essential regulations in the interface core might be attributed to its bridgeness in the network. More importantly, we identified one regulatory FBL between HNF4A and NR2F2 that possesses the highest bridgeness in the interface core. Further investigation suggested that the disturbance of the HNF4A-NR2F2 FBL might protect tumor cells from apoptotic processes. Our results emphasize the relevance of the regulatory network properties to cellular systems and might reveal a critical role of the interface core in cancer. gene regulatory network, regulatory bridge, carcinogenesis, feedback loop Introduction The gene regulatory network (GRN) is responsible for the control of biological processes in cells [1, 2]. In metazoans, the GRN consists of two major families of gene regulators, transcription factors (TFs) and microRNAs (miRNAs) [3]. TFs are proteins that bind to specific nucleic acid sequences to regulate target gene expressions at the transcriptional level [4]; miRNAs are small (∼21–22 nucleotides), noncoding RNAs that regulate gene expression at the posttranscriptional level in eukaryotic cells [5]. TFs are regarded as the primary regulators of gene expression and can be either activators or repressors [6, 7]. On the other hand, miRNAs are considered critical in regulatory systems through fine-tuning the expression of target genes and usually suppress target mRNA expression [8, 9]. Previous studies have observed that cooperation between TFs and miRNAs is prevalent within living cells [3]. Accordingly, the various regulatory abilities and the widespread cooperation of TFs and/or miRNAs make the GRN subtle and difficult to investigate. Like other biological networks, the GRN is scale-free: a few nodes in the networks are highly connected [10]. Moreover, it is hierarchical [11, 12]. There are so-called master regulators, which control most of the regulatory information flow in the GRN and have the maximum impact in affecting gene expression [13, 14]. However, a previous study found that the nonmaster regulators in the GRN are more relevant to cell viability [11]. This study also suggested these middle-level regulators, which are the regulators with more than one in-degree, as control bottlenecks, i.e. regulatory bridges, in the hierarchical structure of the GRN. Moreover, the bridgeness of regulations in the GRN has been demonstrated to be influential in cancers [15, 16]. However, how the GRN hierarchy is involved in the cellular system and the roles of these regulatory bridges in regulating biological processes has not been well studied. To address this knowledge gap, we reconstructed the human GRN that constitutes TF and miRNA regulations. We first separated the human GRN into three layers: the constituent core, periphery, and interface core in between. Our analyses found that the interface core could be central and bridge the regulatory information flow in the human GRN, and could be critical to cell viability and cancer. Finally, we identified one feedback loop (FBL) between HNF4A and NR2F2 that possesses the highest bridgeness in the interface core. Furthermore, literature-based evidence supported our network-based results and indicated that the disequilibrium of the HNF4A-NR2F2 FBL might protect tumor cells from intrinsic apoptotic processes. In summary, our study clarified the influence of regulation bridgeness in the human GRN, as well as unveiled a critical role of bridging regulations in cancer. Results Structure of the human GRN The human GRN constructed in this study covers TF and miRNA regulations and is scale-free-like (Supplementary Figure S1). That is, only a few targets are highly regulated, and a few regulators govern a large number of targets. In the network, <10% of biomolecules are regulators and only 0.2% of regulations are regulatory FBLs (Supplementary Table S1). The FBLs are bidirectional regulations formed by two regulators, which regulated each other. In addition, the GRN contains a single strongly connected component (SCC) that covers 1175 (5.71%) TFs/miRNAs and 41 102 (5.44%) regulations (Figure 1A, left panel). To examine if the human GRN was generated randomly, we built two models with 1000 random networks for each. In the first model, we assumed that the regulators randomly recruited the targets with equal probability (Figure 1A, middle panel). We also kept the numbers of regulators and targets the same as those in the constructed GRN. In the second model, the in- and out-degree distributions of random networks were controlled to be identical to the in- and out-degree distributions of the GRN (Figure 1A, right panel). We observed that the probability of forming more than one single SCC is <0.001 and 0.018 in the first and second model, respectively. This observation suggested that the single SCC property might be structural [17]. Even so, the SCC in the constructed GRN is distinguishable from those in the random networks by size. In the first model, all the sizes of the SCCs were 1494, which are significantly larger than the SCC in the constructed GRN (Figure 1B, left table and middle panel). Interestingly, the number of regulated regulators (regulators with in-degree ≥1) in the GRN was 1494. Notably, all the nodes in a SCC are required to be regulated regulators. Therefore, the significantly larger sizes of the SCCs in the random networks from the first model could be an artifact of the number of the regulated regulators in the constructed SCC. On the other hand, the SCC in the GRN was significantly larger than the random ones in the second model (Figure 1B, left table and right panel). This observation showed that the constructed GRN possessed more regulated regulators, i.e. regulators in the middle level, which has been demonstrated to be relevant to cell viability [14]. The above results indicate that the random network structure is incomparable with the constructed GRN. Additionally, we observed that older genes tended to possess higher in- and out-degrees (Figure 1D). The phylostratum data were obtained from Neme et al. [18]. This observation and the scale-free-like characteristic suggested that the evolution of the human GRN could be a preferential attachment process [19]. In other words, older genes could be favored to have a new target or regulator. Accordingly, we could conclude that the human GRN might have evolved, and the regulators could selectively recruit targets or vice versa. As the SCC in the GRN comprised regulators that are naturally enriched in the nucleus, the non-SCC was observed to be enriched in the cytoskeleton (Supplementary Table S2). We then denoted the SCC as the core layer and non-SCC region as the periphery layer in the GRN. Of note, as the FBLs are naturally SCCs, the periphery layer (non-SCC region) contains no FBLs. Figure 1. View largeDownload slide Network characteristics of the human GRN. (A) The schematic diagram of the core structure in the human GRN and random network model. The network could be separated into three layers: the interface core, constituent core and periphery. The interface core (yellow) was formed by all the regulatory FBLs in the networks. The constituent core (red) is the SCC of the network containing regulators only and excluding all the regulatory FBLs. All the other regulations, i.e. not in the interface or constituent core, were categorized to the periphery region (blue, green and gray arrows). Of note, nodes in the periphery form no regulatory loops with either the node in the periphery or the core. Additionally, in the first random model, as no regulated regulators (regulators with in-degree ≥1) were in the periphery, the depth of the periphery is only one. (B) The distribution of the SCC size and (C) the number of FBLs in the random networks. The left table shows the statistical significance of the human GRN compared with the random network models. The P-values in the table were calculated using the Wilcoxon rank-sum test. The red line in each panel indicates the value of the corresponding network feature in the GRN. (D) The phylogenetic association of regulatory degree in the human GRN. The oldest phylogenetic origins of genes are listed at the x-axis from older to younger (left to right). (E) The connectivity and (F) closeness centrality distribution of the nodes participating in the core, FBL and periphery. Figure 1. View largeDownload slide Network characteristics of the human GRN. (A) The schematic diagram of the core structure in the human GRN and random network model. The network could be separated into three layers: the interface core, constituent core and periphery. The interface core (yellow) was formed by all the regulatory FBLs in the networks. The constituent core (red) is the SCC of the network containing regulators only and excluding all the regulatory FBLs. All the other regulations, i.e. not in the interface or constituent core, were categorized to the periphery region (blue, green and gray arrows). Of note, nodes in the periphery form no regulatory loops with either the node in the periphery or the core. Additionally, in the first random model, as no regulated regulators (regulators with in-degree ≥1) were in the periphery, the depth of the periphery is only one. (B) The distribution of the SCC size and (C) the number of FBLs in the random networks. The left table shows the statistical significance of the human GRN compared with the random network models. The P-values in the table were calculated using the Wilcoxon rank-sum test. The red line in each panel indicates the value of the corresponding network feature in the GRN. (D) The phylogenetic association of regulatory degree in the human GRN. The oldest phylogenetic origins of genes are listed at the x-axis from older to younger (left to right). (E) The connectivity and (F) closeness centrality distribution of the nodes participating in the core, FBL and periphery. It is worth noting that when comparing the two random models, the human GRN possessed significantly more FBLs (Figure 1C). This result and the significantly larger SCC suggested that the regulators in the human GRN might tend to regulate each other directly or indirectly (regulatory path length ≥1). We then applied a connectivity measurement to investigate the roles of the FBLs in the core structure. This connectivity was calculated by dividing the degree of a gene/miRNA by the number of connections between the core and the periphery layer. Interestingly, TFs/miRNAs forming FBLs possessed significantly higher connectivity than other genes/miRNAs in the human GRN (Figure 1E). That is, the TFs/miRNAs forming FBLs play pivotal roles in connecting the core and the periphery. Additionally, the FBL TFs/miRNAs possess significantly higher closeness centrality (Figure 1F). This result implies that the FBL TFs/miRNAs might also be the center of the GRN. Briefly, the FBLs could compose an interface-like structure between core and periphery and act as the regulatory center of the GRN. Therefore, we further separated the core layer into two sublayers: the interface core formed by the FBLs and the constituent core formed by other regulations in the core. In summary, we separated the human GRN into three layers: the interface core, constituent core and periphery. Furthermore, the structure of the human GRN could be evolved, and thus, its structural characteristics and the embedded biological features were worth discovering. The interface core bridges the human GRN In the network, the regulator–target edges are dead ends, while the regulator–regulator edges are the only connections that can be passed through. Thus, the fullness of regulator–regulator edges could lead the core to bridge the GRN. In graph theory, a bridge is an edge whose removal breaks the graph into several connected components. Thus, the removal of a bridge decreases the connectivity of a network. In a network, edges with higher edge betweenness centrality (eBC) control a larger number of the shortest paths and therefore could act as bridges. Indeed, the removal of the edges following a descending order of directed eBC rapidly disrupted the GRN (Supplementary Figure S2). Therefore, we used directed eBC to measure the bridgeness of an edge (regulation) in this study. As expected, the regulations in the core, including the interface and constituent cores, possessed significantly higher eBC than those in the periphery (Figure 2A and Supplementary Table S3). The same phenomenon was observed in the random networks with controlled in- and out-degree distribution (Supplementary Figure S3). Accordingly, the high bridgeness of the core could be attributed to its structural influence: the fullness of regulator–regulator edges. Therefore, the core could bridge a graph naturally. Furthermore, the bridging role might lead the core to control the major part of the regulatory information flow in the human GRN. Figure 2. View largeDownload slide Bridging roles of the constituent and interface cores in the human GRN. (A) The eBC distribution of regulations in the human GRN. The regulations in the periphery region were further categorized into two types: RR, formed by two regulators; RT, formed by one regulator and one target. The proportion of disconnected shortest paths after removing the constituent and interface cores and periphery (RR) in the human GRN compared with (B) random sampling and (C) the random network. The histograms show the distribution of the proportion of disconnected shortest paths as doing (B) random sampling and (C) in random networks. The vertical lines represent the value of the human GRN. Figure 2. View largeDownload slide Bridging roles of the constituent and interface cores in the human GRN. (A) The eBC distribution of regulations in the human GRN. The regulations in the periphery region were further categorized into two types: RR, formed by two regulators; RT, formed by one regulator and one target. The proportion of disconnected shortest paths after removing the constituent and interface cores and periphery (RR) in the human GRN compared with (B) random sampling and (C) the random network. The histograms show the distribution of the proportion of disconnected shortest paths as doing (B) random sampling and (C) in random networks. The vertical lines represent the value of the human GRN. To further discover the bridging function of the core, we removed the core and then recalculated the shortest paths between all the genes and miRNAs. Indeed, after removing the constituent or interface core, a significantly larger proportion of paths were disconnected, i.e. no path could be found between genes/miRNAs, compared with randomly removing the same amount of regulator–regulator edges in the constituent or interface core (Figure 2B, Supplementary Table S4). Of note, the removal of regulator–regulator edges in the periphery also disconnected a significantly larger proportion of paths (Figure 2B, Supplementary Table S4). Interestingly, when comparing with the random networks that controlled the in- and out-degree distribution, only the interface core remained significant in disconnecting regulatory paths (Figure 2C, Supplementary Table S4). These observations imply that the high bridgeness of the constituent core might be a result of its structural importance, i.e. the regulator–regulator edges are the connections that can be passed through. Moreover, this result demonstrated that the bridgeness of the interface core may not only be attributed to its structural influence but also to other pivotal roles in the GRN. Moreover, it recapitulates the roles of the interface core in bridging between the core and the periphery of the human GRN. However, removal of the interface core only disconnected 5.72% of biomolecule pairs in the GRN. Thus, there might be alternative regulatory paths that could keep the network structure unbroken when the interface core was attacked. A possible scenario is that the interface-controlled regulatory paths might be critical, and these alternative paths might be required to increase the robustness of the human GRN. Previous studies have demonstrated that FBLs could play essential roles in critical biological processes, such as differentiation and cell cycle transitions [20, 21]. Accordingly, these studies supported the proposed scenario for the existing alternatives of the interface-controlled paths and may emphasize the importance of the interface core in the human GRN. Nevertheless, the biological role of the interface core still needs to be studied. The interface core is influential in cell viability and cancer Even though the FBLs in the human GRN are distinguishable from those in the random networks, the FBLs are natural bridges in the graph. Therefore, to investigate the relevance of the interface core to human cells, we collected a set of cancer-associated biomolecules, including genes and miRNAs, and a set of essential human genes. It is worth noting that the proportion of overlapping genes between the essential genes and the cancer-associated genes was only 8% (calculated using the Jaccard index). This observation suggested that the two data sets are nearly independent. Then, we compiled a set of cancer-associated regulations linked by two cancer-associated biomolecules and a set of essential regulations formed by at least one essential gene. We found that the cancer-associated and essential regulations possessed significantly higher eBC than the non-cancer-associated and nonessential regulations, respectively (Supplementary Table S5). Moreover, both of the cancer-associated and essential regulation proportions increased with eBC in ascending order (cancer: P-value = 1.33 × 10−6, χ2 test, Figure 3A; essential: P-value < 2.2 × 10−16, χ2 test, Figure 3B). These results suggested that the bridges could be more cancer-associated and essential compared with the other regulations, and the eBC might be capable of revealing the important regulations involved in cancer and cell viability. Figure 3. View largeDownload slide Cancer association and essentiality of eBC and the core in the human GRN. The association between eBC and the proportion of the (A) cancer-associated and (B) essential regulation. The colored circles represent the significance derived from Fisher’s exact test. The red and green circles show significantly overrepresented and underrepresented, respectively, and the gray show nonsignificant. (C) The enrichment of cancer-associated and essential biomolecules and regulations in the four regions. Asterisks show the significance of enrichment. For each bar, two asterisks represent the significance when the whole human GRN (left) and only regulators (right) were used as a reference. Red asterisks are significantly overrepresented, green underrepresented and gray insignificant. The periphery (RT) is marked by only the left asterisk, i.e. using the whole human GRN as the reference. Figure 3. View largeDownload slide Cancer association and essentiality of eBC and the core in the human GRN. The association between eBC and the proportion of the (A) cancer-associated and (B) essential regulation. The colored circles represent the significance derived from Fisher’s exact test. The red and green circles show significantly overrepresented and underrepresented, respectively, and the gray show nonsignificant. (C) The enrichment of cancer-associated and essential biomolecules and regulations in the four regions. Asterisks show the significance of enrichment. For each bar, two asterisks represent the significance when the whole human GRN (left) and only regulators (right) were used as a reference. Red asterisks are significantly overrepresented, green underrepresented and gray insignificant. The periphery (RT) is marked by only the left asterisk, i.e. using the whole human GRN as the reference. Notably, the core possessed significantly higher eBC in the GRN. Thus, the constituent and interface cores might be highly associated with cancer and cell viability. Indeed, compared with the whole GRN, the cancer-associated biomolecules were significantly enriched in the interface core as well as in the constituent core (Figure 3C and Supplementary Table S6). Consistently, the essential genes were significantly enriched in the interface and constituent cores (Figure 3C and Supplementary Table S6). These observations suggested that both the constituent and interface cores might be influential in cancer and cell viability. However, the cancer-associated biomolecules and essential genes were significantly enriched in regulators (Supplementary Figure S4). Thus, the enrichment of cancer-associated biomolecules and essential genes in the core might be biased by its fullness of regulators. Indeed, compared with all the regulators, the cancer-associated biomolecules become significantly underrepresented, and enrichment of the essential genes became insignificant in the constituent core (Figure 3C and Supplementary Table S6); only the interface core was still significantly enriched with the cancer-associated biomolecules and essential genes (Figure 3C and Supplementary Table S6). On the other hand, compared with the whole human GRN, the interface and constituent cores were significantly enriched with the cancer-associated and essential regulations (Figure 3C and Supplementary Table S6). Likely, the cancer-associated and essential regulations were significantly overrepresented in the interface and constituent cores by using the regulator–regulator edges as the reference (Figure 3C and Supplementary Table S6). Notably, the interface core contains a higher proportion of cancer-associated and essential biomolecules and regulations than the constituent core. Additionally, the enrichment of the interface core is not biased by the fullness of regulators. Furthermore, the cancer-associated biomolecules were significantly overrepresented in the interface core when using the core as the reference (P-value = 4.21 × 10−25, Fisher’s exact test). The essential genes were also enriched in the interface with moderate significance (P-value = 0.10, Fisher’s exact test) when using the core as the reference. The enrichment of cancer-associated and essential biomolecules and regulations emphasized that the interface core could be critical in human cells. These results further imply that a disturbance to the interface core might result in more damage to the human cells compared with the constituent core. Moreover, these observations recapitulate the scenario explaining the existence of the alternative regulatory paths accompanying the interface-controlled paths. In summary, the interface core is not only pivotal to the structure of the human GRN but might also be influential to cancer and cell viability. These results further emphasized the pertinence of the network properties to the human GRN and the potentially critical role of the core in cancer and normal cells. Regulatory role of the interface core in cancers To embody the regulatory function of the interface core in cancer, we investigated the FBL between HNF4A and NR2F2, which possesses the highest eBC in the interface core. HNF4A has been reported to be downregulated and act as a critical transcriptional regulator in hepatocellular carcinoma (HCC) [22, 23], while the upregulation of HNF4A is reportedly a key event in gastric cancer [24]. Previous studies also found that dysregulation of HNF4A is influential to various cancers [25–28]. NR2F2 has also been found to be important in various cancer types, including prostate, colorectal and breast cancer [29–31]. These results suggested that HNF4A and NR2F2 could play critical roles in carcinogenesis. However, the effects of dysregulating the HNF4A-NR2F2 FBL in cancer have not been well studied. In The Cancer Genome Atlas (TCGA) data set, HNF4A was significantly upregulated in five cancer types and downregulated in four (Figure 4A); NR2F2 is upregulated in breast cancer and downregulated in eight cancer types (Figure 4A). This observation suggested that NR2F2 might be turned off, and HNF4A might dominate and disequilibrate this FBL in the five cancer types. Roughly, two regulators forming an FBL must tightly control each other to keep the equilibrium status of an FBL. The dysregulation of either one of the two regulators could lead to a FBL disequilibrium. For example, the disequilibrium effects of the double-negative FBL between OCT4 and miR-145 have been observed, i.e. OCT4 and miR-145 inhibited each other [32, 33]. When miR-145 dominated the FBL, differentiation of stem cells was initiated, while the domination of OCT4 could inhibit miR-145 and then promote tumorigenesis. Herein, the downstream pathways regulated by NR2F2 might be turned off because of the downregulation of NR2F2 in cancers. The upregulation of HNF4A could dominate the HNF4A-NR2F2 FBL and thus turn on the corresponding downstream pathways in the five cancer types. Therefore, we focused on the downstream pathways regulated by HNF4A to reveal the consequence of disturbing this FBL via investigating the HNF4A-regulated functional modules in the five cancer types. We found that most of the functional modules were apoptosis-related (Supplementary Table S7). However, we did not find any evidence that HNF4A could directly regulate apoptosis. HNF4A might regulate apoptosis through its targeting network. Indeed, HNF4A could promote apoptosis through the regulation of its target miRNA (miR-548p) in HCC [34]. A mouse model study has suggested that HNF4A might be associated with HCC development via the targeting of the Perp, the p53/p63 apoptosis effector gene [35]. The evidence suggested that HNF4A might be involved in the development of the five cancer types through regulating apoptosis. Figure 4. View largeDownload slide HNF4A-NR2F2 FBL regulation in cancers. (A) The log2 fold ratio of HNF4A and NR2F2 expression in cancers. The bold red and green numbers represent the significantly overexpressed and underexpressed, respectively, with the gray denoting nonsignificant expression. (B) The differential expression patterns of the HNF4A-regulated apoptotic module. We colored HNF4A target genes with red and framed the Bcl2 family with a green rectangle. (C) The hypothesis of the downstream effect caused by the loss of equilibrium mediated by the HNF4A-NR2F2 FBL. The potential molecular mechanism describes how the domination of HNF4A in the regulatory FBL between HNF4A and NR2F2 protect cancer from the apoptotic process. The mitochondrial apoptosis-induced channel (MAC) and cytochrome c are potentially inhibited in this hypothesis. Figure 4. View largeDownload slide HNF4A-NR2F2 FBL regulation in cancers. (A) The log2 fold ratio of HNF4A and NR2F2 expression in cancers. The bold red and green numbers represent the significantly overexpressed and underexpressed, respectively, with the gray denoting nonsignificant expression. (B) The differential expression patterns of the HNF4A-regulated apoptotic module. We colored HNF4A target genes with red and framed the Bcl2 family with a green rectangle. (C) The hypothesis of the downstream effect caused by the loss of equilibrium mediated by the HNF4A-NR2F2 FBL. The potential molecular mechanism describes how the domination of HNF4A in the regulatory FBL between HNF4A and NR2F2 protect cancer from the apoptotic process. The mitochondrial apoptosis-induced channel (MAC) and cytochrome c are potentially inhibited in this hypothesis. To confirm this implication, we investigated the genes involved in the identified HNF4A-regulated apoptotic functional module in the five cancer types (Figure 4B). In the module, 8 of 21 members were genes encoding BCL2 family proteins, i.e. apoptosis regulator Bcl-2 [36]. Among these eight genes, two pro-apoptotic genes, BAK1 and BID, were upregulated by HNF4A. These observations suggested that human cells might initiate apoptosis to respond to carcinogenesis. However, carcinogenesis is known to be because of excessive cell proliferation and decreased cell death, i.e. inappropriate inhibition of apoptosis [37]. Interestingly, there are four genes, VIL1, HSPD1, SRC and TRAF2, upregulated by HNF4A. VIL1 encodes the vilin protein and has been reported to inhibit apoptosis through maintaining mitochondrial integrity [38]. BAK1 could initiate apoptosis via increasing mitochondrial outer membrane permeability to release cytochrome c and other proapoptotic factors [39]. Therefore, the over-expression of VIL1 could inhibit the BAK1-initiated apoptosis process. Additionally, the previous study has demonstrated that SRC could be capable of inhibiting the Bax/Bak activation cascade through increasing the activation threshold of Bax [40]. Furthermore, HSPD1 has been found to interact with Bax and Bak in vitro, and this interaction may block the proapoptotic ability of the Bax/Bak complex [41]. Accordingly, VIL1, SRC and HSPD1 might cooperate to promote carcinogenesis via inhibiting Bax/Bak-initiated apoptotic processes. Moreover, TRAF2 has been reported to protect cancer cells from ER stress-induced apoptosis [42], to suppress apoptotic processes through regulating tumor necrosis factor and nuclear factor kappa-light-chain-enhancer of activated B cells (NF-κB) signaling [43], and to inhibit TRAIL- and CD95L-induced apoptosis [44]. Briefly, HNF4A may activate both antiapoptotic and proapoptotic processes simultaneously during cancer development. That is, HNF4A might possess a dual role, i.e. antiapoptotic and proapoptotic, in carcinogenesis of the five cancer types. Based on the results above, we proposed a hypothesis on how the overexpressed HNF4A promoted carcinogenesis of the five cancer types (Figure 4C). During the development of the five cancer types, the protective mechanism of human cells might highly express HNF4A to activate proapoptotic genes BAK1 and BID for the initiation of intrinsic apoptosis. However, the upregulation of HNF4A may also upregulate the expression of VIL1, SRC and HSPD1, which could repress the initiation of apoptosis either by maintaining mitochondrial membrane integrity [38] or obstructing the formation of the Bax/Bak complex [40, 41]. Moreover, HNF4A also activated TRAF2, which could inhibit apoptosis upstream of the apoptotic pathway [44]. Finally, the tumor cells were able to survive Bax/Bak-triggered intrinsic apoptosis. Collectively, our findings suggest that the domination of HNF4A might disequilibrate the FBL between HNF4A and NR2F2 to allow tumor cells to survive apoptosis and even to promote carcinogenesis of the five cancer types. However, more experiments are needed to validate our hypothesis. Discussion In this study, we investigated the structure of the human GRN and discovered the critical regulations of cancer through analyzing eBC of regulations. We observed that the human GRN structure is distinguishable from the random ones and thus could be evolved rather than randomly generated. However, the human GRN structure was strongly affected by the nature of the gene regulatory system—the proportion of regulators is small, but they dominate all (thousands) of the regulations. In other words, the averaged out-degree of regulators becomes much higher than randomly expected (408 versus 37). For example, this nature could make the network, with high probability, form a single SCC with regulators as critical nodes inside the network, such as hubs and bottlenecks. Accordingly, the biological significance of structurally critical nodes inside the human GRN becomes a paradox—it is hard to answer if the significance of structurally critical nodes was attributed to biology or graph structure. Nevertheless, as the regulators were significantly enriched with the essential and cancer-associated biomolecules (Supplementary Figure S4), the SCC could be influential in a cellular system because of the fact that all the nodes in the SCC are regulators. Moreover, the single SCC characteristic could be derived from the nature of the regulatory system, the high average out-degree of the regulators [17]. Accordingly, the single SCC characteristic might be biologically relevant. However, further studies are needed to justify if the single SCC characteristic is biologically derived or not. On the other hand, we found that the human GRN could be naturally bridged by the SCC, including the constituent and interface cores. To further clarify the bridge role of the SCC, we investigated the loss rate of the shortest paths after removing the SCC. Interestingly, only the removal of the interface core caused a significantly higher loss rate of the shortest paths than the random networks with controlled degree distribution. In other words, the interface core in the human GRN occupied significantly more shortest paths than the random interface cores did. Of note, in the random networks, the interface core also possessed the highest eBC and could bridge the random networks. This result implied that the interface core could be global, whereas the random interface core could be local [45]. The high betweenness centrality of local bridges is contributed by their high degree that makes local bridges control lots of paths between neighbors [45]. Indeed, the random interface core possessed higher degree than the interface core in the constructed human GRN (Supplementary Figure S5). These observations demonstrated that the random interface cores could bridge the networks because of their high degree caused by the nature of the gene regulatory system. In contrast, the interface core in the human GRN could further act as a global bridge by occupying the longer shortest paths. In addition, our results suggested that the interface core could be influential in cell viability and cancer without biasing from the fullness of regulators. Briefly, we observed that (1) eBC is positively correlated with the proportion of essential and cancer-associated regulations, (2) the interface core possessed the highest eBC and (3) the interface core was significantly enriched with essential and cancer-associated biomolecules and regulations. Accordingly, we could conclude that the biological significance of the interface core might be attributed to its highest eBC. However, we did not observe that the essential and cancer-associated biomolecules and regulations were significantly underrepresented in the regulations with low eBC in the interface core. This result might be attributed to the relatively high eBC of the interface core in the human GRN (Figure 2A). In other words, there is no regulation in the interface core possessing relatively low eBC. Consequently, other analyses are required to robustly demonstrate this negative association that the biological significance of the interface core is attributed to its bridgeness. Conclusions In summary, we observed that the regulatory FBLs form an interface-like structure between the constituent core and periphery region in the human GRN. We then denoted the FBLs as the interface core in the human GRN. Further investigation suggested that the interface core could not only be influential to the structure of the human GRN but also crucial to cancer and cell viability. Moreover, we discovered the regulatory HNF4A-NR2F2 FBL possesses the highest eBC in the interface core. Advanced analysis showed that HNF4A might possess a dual role of antiapoptosis and proapoptosis during carcinogenesis and suggested that the disturbance of this FBL might protect tumor cells from apoptotic elimination. Methods For a detailed description of the constructed GRN, cancer-associated biomolecules, essential genes and mRNA and miRNA expression profiles from TCGA, see Supplementary SI Materials and Methods. Identification of regulatory bridges in the human GRN In this study, we applied eBC to evaluate the influence of one regulation on the structure of GRN, the sensitivity of the GRN structure to the removal of the given regulation. The eBC of an edge, e, is the sum of the proportion of all-pairs of shortest paths passing through e and can be calculated as below: BCe=∑i,j∈VSeS, (1) where BCe is the eBC of edge, e; V is the set of nodes in the GRN; S is the number of all-pairs shortest paths; and Se is the number of shortest paths passing through e. All network properties, such as SCCs and closeness centrality, in this study were calculated by python package NetworkX. Discovery of HNF4A-regulated functional modules To discover the HNF4A-regulated downstream functional modules, we collected HNF4A target genes that are (1) significantly positively coexpressed with HNF4A and (2) significantly upregulated in the cancer types in which HNF4A was also upregulated. Through these two conditions, we obtained those genes potentially activated by HNF4A in cancer. The detailed definition of these two conditions is specified in the Supplementary SI Materials and Methods. To discover HNF4A regulation in more detail, we recruited the first neighbors of the selected HNF4A target genes in the human protein interaction network (PIN). The protein–protein interactions (PPI) were obtained from the PIN Analysis v2 [46]. We then denoted the network constructed by the selected HNF4A target genes, their protein interacting partners and PPIs among them as the HNF4A-regulated network. Next, we performed enrichment analysis to determine the functions involved in the identified HNF4A-regulated network. Of note, we conducted the functional enrichment analysis using two methods, conventional and network-wise. With the conventional method, the overrepresentation of selected HNF4A target genes and their PPI partners defines the significance of HNF4A-regulated functions. On the other hand, the network-wise enrichment analysis evaluates the significance of HNF4A-regulated functions through the overrepresentation of functional PPIs among selected HNF4A targets. The procedure of functional enrichment analysis is described in the Supplementary SI Materials and Methods in detail. To preserve the HNF4A regulation in the identified functions, we calculated the enrichment significance of the selected HNF4A targets in each functional module using Fisher’s exact test. Finally, we considered those functions passing all three tests (P-value < 0.05) as potential HNF4A-regulated functional modules. Key Points The regulatory FBLs inside the core form an interface-like structure between the constituent core and the periphery region, which also act as the regulatory center in the human GRN. The regulatory FBLs cover a significantly higher proportion of the cancer-associated and essential regulation and biomolecules than the constituent core and periphery region. The regulatory FBLs are not only influential for the network structure but are also critical in cellular systems. The disturbance of the regulatory FBL between HNF4A and NR2F2 might protect tumor cells from intrinsic apoptotic elimination. Supplementary Data Supplementary data are available online at https://academic.oup.com/bib. Funding This work was supported by the Ministry of Science and Technology (grant number MOST 105-2628-E-010-001-MY3 and MOST 104-2320-B-010-037-). Yun-Ru Chen is a PhD student in the Institute of Biomedical Informatics, National Yang-Ming University. Hsuan-Cheng Huang is a professor in the Institute of Biomedical Informatics, National Yang-Ming University. Chen-Ching Lin is an assistant professor in the Institute of Biomedical Informatics, National Yang-Ming University. References 1 Peter IS , Davidson EH. Evolution of gene regulatory networks controlling body plan development . Cell 2011 ; 144 ( 6 ): 970 – 85 . http://dx.doi.org/10.1016/j.cell.2011.02.017 Google Scholar CrossRef Search ADS PubMed 2 Lander AD. How cells know where they are . Science 2013 ; 339 ( 6122 ): 923 – 7 . http://dx.doi.org/10.1126/science.1224186 Google Scholar CrossRef Search ADS PubMed 3 Shalgi R , Lieber D , Oren M , et al. Global and local architecture of the mammalian microRNA-transcription factor regulatory network . PLoS Comput Biol 2007 ; 3 ( 7 ): e131 . Google Scholar CrossRef Search ADS PubMed 4 Latchman DS. Transcription factors: an overview . Int J Biochem Cell Biol 1997 ; 29 ( 12 ): 1305 – 12 . http://dx.doi.org/10.1016/S1357-2725(97)00085-X Google Scholar CrossRef Search ADS PubMed 5 Guo H , Ingolia NT , Weissman JS , et al. Mammalian microRNAs predominantly act to decrease target mRNA levels . Nature 2010 ; 466 ( 7308 ): 835 – 40 . http://dx.doi.org/10.1038/nature09267 Google Scholar CrossRef Search ADS PubMed 6 Lee TI , Young RA. Transcription of eukaryotic protein-coding genes . Annu Rev Genet 2000 ; 34 : 77 – 137 . http://dx.doi.org/10.1146/annurev.genet.34.1.77 Google Scholar CrossRef Search ADS PubMed 7 Shen-Orr SS , Milo R , Mangan S , et al. Network motifs in the transcriptional regulation network of Escherichia coli . Nat Genet 2002 ; 31 ( 1 ): 64 – 8 . http://dx.doi.org/10.1038/ng881 Google Scholar CrossRef Search ADS PubMed 8 Bartel DP. MicroRNAs: genomics, biogenesis, mechanism, and function . Cell 2004 ; 116 ( 2 ): 281 – 97 . http://dx.doi.org/10.1016/S0092-8674(04)00045-5 Google Scholar CrossRef Search ADS PubMed 9 Harfe BD. MicroRNAs in vertebrate development . Curr Opin Genet Dev 2005 ; 15 ( 4 ): 410 – 15 . http://dx.doi.org/10.1016/j.gde.2005.06.012 Google Scholar CrossRef Search ADS PubMed 10 Babu MM , Luscombe NM , Aravind L , et al. Structure and evolution of transcriptional regulatory networks . Curr Opin Struct Biol 2004 ; 14 ( 3 ): 283 – 91 . http://dx.doi.org/10.1016/j.sbi.2004.05.004 Google Scholar CrossRef Search ADS PubMed 11 Yu H , Gerstein M. Genomic analysis of the hierarchical structure of regulatory networks . Proc Natl Acad Sci USA 2006 ; 103 ( 40 ): 14724 – 31 . http://dx.doi.org/10.1073/pnas.0508637103 Google Scholar CrossRef Search ADS PubMed 12 Bhardwaj N , Yan KK , Gerstein MB. Analysis of diverse regulatory networks in a hierarchical context shows consistent tendencies for collaboration in the middle levels . Proc Natl Acad Sci USA 2010 ; 107 ( 15 ): 6841 – 6 . http://dx.doi.org/10.1073/pnas.0910867107 Google Scholar CrossRef Search ADS PubMed 13 Ihmels J , Levy R , Barkai N. Principles of transcriptional control in the metabolic network of Saccharomyces cerevisiae . Nat Biotechnol 2004 ; 22 ( 1 ): 86 – 92 . http://dx.doi.org/10.1038/nbt918 Google Scholar CrossRef Search ADS PubMed 14 Yu H , Luscombe NM , Qian J , et al. Genomic analysis of gene expression relationships in transcriptional regulatory networks . Trends Genet 2003 ; 19 ( 8 ): 422 – 7 . http://dx.doi.org/10.1016/S0168-9525(03)00175-6 Google Scholar CrossRef Search ADS PubMed 15 Kotlyar M , Fortney K , Jurisica I. Network-based characterization of drug-regulated genes, drug targets, and toxicity . Methods 2012 ; 57 ( 4 ): 499 – 507 . http://dx.doi.org/10.1016/j.ymeth.2012.06.003 Google Scholar CrossRef Search ADS PubMed 16 Azevedo H , Moreira-Filho CA. Topological robustness analysis of protein interaction networks reveals key targets for overcoming chemotherapy resistance in glioma . Sci Rep 2015 ; 5 ( 1 ): 16830 . http://dx.doi.org/10.1038/srep16830 Google Scholar CrossRef Search ADS PubMed 17 Bollobás B , Frieze AM. On matchings and Hamiltonian cycles in random graphs . North Holland Math Stud 1985 ; 118 : 23 – 46 . Google Scholar CrossRef Search ADS 18 Neme R , Tautz D. Phylogenetic patterns of emergence of new genes support a model of frequent de novo evolution . BMC Genomics 2013 ; 14 : 117 . http://dx.doi.org/10.1186/1471-2164-14-117 Google Scholar CrossRef Search ADS PubMed 19 Barabasi AL , Albert R. Emergence of scaling in random networks . Science 1999 ; 286 ( 5439 ): 509 – 12 . http://dx.doi.org/10.1126/science.286.5439.509 Google Scholar CrossRef Search ADS PubMed 20 Wang F , Zhu Y , Guo L , et al. A regulatory circuit comprising GATA1/2 switch and microRNA-27a/24 promotes erythropoiesis . Nucleic Acids Res 2014 ; 42 ( 1 ): 442 – 57 . http://dx.doi.org/10.1093/nar/gkt848 Google Scholar CrossRef Search ADS PubMed 21 Ma Y , Wang B , Jiang F , et al. A feedback loop consisting of microRNA 23a/27a and the beta-like globin suppressors KLF3 and SP1 regulates globin gene expression . Mol Cell Biol 2013 ; 33 ( 20 ): 3994 – 4007 . http://dx.doi.org/10.1128/MCB.00623-13 Google Scholar CrossRef Search ADS PubMed 22 Lazarevich NL , Cheremnova OA , Varga EV , et al. Progression of HCC in mice is associated with a downregulation in the expression of hepatocyte nuclear factors . Hepatology 2004 ; 39 ( 4 ): 1038 – 47 . http://dx.doi.org/10.1002/hep.20155 Google Scholar CrossRef Search ADS PubMed 23 Ning BF , Ding J , Yin C , et al. Hepatocyte nuclear factor 4 alpha suppresses the development of hepatocellular carcinoma . Cancer Res 2010 ; 70 ( 19 ): 7640 – 51 . http://dx.doi.org/10.1158/0008-5472.CAN-10-0824 Google Scholar CrossRef Search ADS PubMed 24 Chang HR , Nam S , Kook MC , et al. HNF4alpha is a therapeutic target that links AMPK to WNT signalling in early-stage gastric cancer . Gut 2016 ; 65 ( 1 ): 19 – 32 . http://dx.doi.org/10.1136/gutjnl-2014-307918 Google Scholar CrossRef Search ADS PubMed 25 Vuong LM , Chellappa K , Dhahbi JM , et al. Differential effects of hepatocyte nuclear factor 4alpha isoforms on tumor growth and T-Cell factor 4/AP-1 interactions in human colorectal cancer cells . Mol Cell Biol 2015 ; 35 ( 20 ): 3471 – 90 . http://dx.doi.org/10.1128/MCB.00030-15 Google Scholar CrossRef Search ADS PubMed 26 Koizume S , Yokota N , Miyagi E , et al. Hepatocyte nuclear factor-4-independent synthesis of coagulation factor VII in breast cancer cells and its inhibition by targeting selective histone acetyltransferases . Mol Cancer Res 2009 ; 7 ( 12 ): 1928 – 36 . http://dx.doi.org/10.1158/1541-7786.MCR-09-0372 Google Scholar CrossRef Search ADS PubMed 27 Lucas B , Grigo K , Erdmann S , et al. HNF4alpha reduces proliferation of kidney cells and affects genes deregulated in renal cell carcinoma . Oncogene 2005 ; 24 ( 42 ): 6418 – 31 . http://dx.doi.org/10.1038/sj.onc.1208794 Google Scholar CrossRef Search ADS PubMed 28 Snyder EL , Watanabe H , Magendantz M , et al. Nkx2-1 represses a latent gastric differentiation program in lung adenocarcinoma . Mol Cell 2013 ; 50 ( 2 ): 185 – 99 . http://dx.doi.org/10.1016/j.molcel.2013.02.018 Google Scholar CrossRef Search ADS PubMed 29 Qin J , Wu SP , Creighton CJ , et al. COUP-TFII inhibits TGF-beta-induced growth barrier to promote prostate tumorigenesis . Nature 2013 ; 493 ( 7431 ): 236 – 40 . Google Scholar CrossRef Search ADS PubMed 30 Wang C , Zhou Y , Ruan R , et al. High expression of COUP-TF II cooperated with negative Smad4 expression predicts poor prognosis in patients with colorectal cancer . Int J Clin Exp Pathol 2015 ; 8 ( 6 ): 7112 – 21 . Google Scholar PubMed 31 Al-Rayyan N , Litchfield LM , Ivanova MM , et al. 5-Aza-2-deoxycytidine and trichostatin A increase COUP-TFII expression in antiestrogen-resistant breast cancer cell lines . Cancer Lett 2014 ; 347 ( 1 ): 139 – 50 . http://dx.doi.org/10.1016/j.canlet.2014.02.001 Google Scholar CrossRef Search ADS PubMed 32 Chivukula RR , Mendell JT. Abate and switch: miR-145 in stem cell differentiation . Cell 2009 ; 137 ( 4 ): 606 – 8 . http://dx.doi.org/10.1016/j.cell.2009.04.059 Google Scholar CrossRef Search ADS PubMed 33 Xu N , Papagiannakopoulos T , Pan G , et al. MicroRNA-145 regulates OCT4, SOX2, and KLF4 and represses pluripotency in human embryonic stem cells . Cell 2009 ; 137 ( 4 ): 647 – 58 . http://dx.doi.org/10.1016/j.cell.2009.02.038 Google Scholar CrossRef Search ADS PubMed 34 Hu XM , Yan XH , Hu YW , et al. miRNA-548p suppresses hepatitis B virus X protein associated hepatocellular carcinoma by downregulating oncoprotein hepatitis B x-interacting protein . Hepatol Res 2016 ; 46 ( 8 ): 804 – 15 . http://dx.doi.org/10.1111/hepr.12618 Google Scholar CrossRef Search ADS PubMed 35 Bonzo JA , Ferry CH , Matsubara T , et al. Suppression of hepatocyte proliferation by hepatocyte nuclear factor 4alpha in adult mice . J Biol Chem 2012 ; 287 ( 10 ): 7345 – 56 . http://dx.doi.org/10.1074/jbc.M111.334599 Google Scholar CrossRef Search ADS PubMed 36 Chao DT , Korsmeyer SJ. BCL-2 family: regulators of cell death . Annu Rev Immunol 1998 ; 16 : 395 – 419 . http://dx.doi.org/10.1146/annurev.immunol.16.1.395 Google Scholar CrossRef Search ADS PubMed 37 Fernald K , Kurokawa M. Evading apoptosis in cancer . Trends Cell Biol 2013 ; 23 ( 12 ): 620 – 33 . http://dx.doi.org/10.1016/j.tcb.2013.07.006 Google Scholar CrossRef Search ADS PubMed 38 Wang Y , Srinivasan K , Siddiqui MR , et al. A novel role for villin in intestinal epithelial cell survival and homeostasis . J Biol Chem 2008 ; 283 ( 14 ): 9454 – 64 . http://dx.doi.org/10.1074/jbc.M707962200 Google Scholar CrossRef Search ADS PubMed 39 Wei MC , Zong WX , Cheng EH , et al. Proapoptotic BAX and BAK: a requisite gateway to mitochondrial dysfunction and death . Science 2001 ; 292 ( 5517 ): 727 – 30 . http://dx.doi.org/10.1126/science.1059108 Google Scholar CrossRef Search ADS PubMed 40 Lopez J , Hesling C , Prudent J , et al. Src tyrosine kinase inhibits apoptosis through the Erk1/2- dependent degradation of the death accelerator Bik . Cell Death Differ 2012 ; 19 ( 9 ): 1459 – 69 . http://dx.doi.org/10.1038/cdd.2012.21 Google Scholar CrossRef Search ADS PubMed 41 Kirchhoff SR , Gupta S , Knowlton AA. Cytosolic heat shock protein 60, apoptosis, and myocardial injury . Circulation 2002 ; 105 ( 24 ): 2899 – 904 . http://dx.doi.org/10.1161/01.CIR.0000019403.35847.23 Google Scholar CrossRef Search ADS PubMed 42 Habelhah H , Zhang L , Xialikaer A , et al. Abstract 3496: TRAF2 protects mammary epithelial and cancer cells from endoplasmic reticulum stress-induced apoptosis . Cancer Res 2016 ; 76(Suppl 14) : 3496 . Google Scholar CrossRef Search ADS 43 Etemadi N , Chopin M , Anderton H , et al. TRAF2 regulates TNF and NF-kappaB signalling to suppress apoptosis and skin inflammation independently of Sphingosine kinase 1 . Elife 2015 ; 4 : e10592 . Google Scholar CrossRef Search ADS PubMed 44 Karl I , Jossberger-Werner M , Schmidt N , et al. TRAF2 inhibits TRAIL- and CD95L-induced apoptosis and necroptosis . Cell Death Dis 2014 ; 5 : e1444 . Google Scholar CrossRef Search ADS PubMed 45 Jensen P , Morini M , Karsai M , et al. Detecting global bridges in networks . J Complex Netw 2016 ; 4 ( 3 ): 319 – 29 . http://dx.doi.org/10.1093/comnet/cnv022 Google Scholar CrossRef Search ADS 46 Cowley MJ , Pinese M , Kassahn KS , et al. PINA v2.0: mining interactome modules . Nucleic Acids Res 2012 ; 40 : D862 – 5 . Google Scholar CrossRef Search ADS PubMed © The Author 2017. Published by Oxford University Press. All rights reserved. For Permissions, please email: journals.permissions@oup.com

Journal

Briefings in BioinformaticsOxford University Press

Published: Nov 29, 2017

There are no references for this article.

You’re reading a free preview. Subscribe to read the entire article.


DeepDyve is your
personal research library

It’s your single place to instantly
discover and read the research
that matters to you.

Enjoy affordable access to
over 18 million articles from more than
15,000 peer-reviewed journals.

All for just $49/month

Explore the DeepDyve Library

Search

Query the DeepDyve database, plus search all of PubMed and Google Scholar seamlessly

Organize

Save any article or search result from DeepDyve, PubMed, and Google Scholar... all in one place.

Access

Get unlimited, online access to over 18 million full-text articles from more than 15,000 scientific journals.

Your journals are on DeepDyve

Read from thousands of the leading scholarly journals from SpringerNature, Elsevier, Wiley-Blackwell, Oxford University Press and more.

All the latest content is available, no embargo periods.

See the journals in your area

DeepDyve

Freelancer

DeepDyve

Pro

Price

FREE

$49/month
$360/year

Save searches from
Google Scholar,
PubMed

Create lists to
organize your research

Export lists, citations

Read DeepDyve articles

Abstract access only

Unlimited access to over
18 million full-text articles

Print

20 pages / month

PDF Discount

20% off