Gene-disease associations identify a connectome with shared molecular pathways in human cholangiopathies

There is no author summary for this article yet. Authors can add summaries to their articles on ScienceOpen to make them more accessible to a non-specialist audience.

Abstract

<p class="first" id="P1">Cholangiopathies are a diverse group of progressive diseases whose primary cell targets are cholangiocytes. To identify shared pathogenesis and molecular connectivity among the three main human cholangiopathies (biliary atresia [BA], primary biliary cholangitis [PBC] and primary sclerosing cholangitis [PSC]), we built a comprehensive platform of published data on gene variants, gene expression and functional studies, and applied network-based analytics in search for shared molecular circuits. Mining the data platform with largest connected component and interactome analyses, we validated previously reported associations and identified essential- and hub-genes. In addition to disease-specific modules, we found a substantial overlap of disease neighborhoods, and uncovered a group of 34 core genes that are enriched for immune processes and abnormal intestine/hepatobiliary mouse phenotypes. Within this core, we identified a gene subcore containing <i>STAT3, IL6, TNF</i> and <i>FOXP3</i> prominently placed in a regulatory connectome of genes related to cellular immunity and fibrosis. We also found substantial gene enrichment in the AGE-RAGE pathway, and showed that RAGE activation induced cholangiocyte proliferation. </p><div class="section"> <a class="named-anchor" id="S1">  </a> <h5 class="section-title" id="d7177180e136">Conclusion</h5> <p id="P2">Human cholangiopathies share pathways enriched by immunity genes and a molecular connectome that links different pathogenic features of BA, PBC and PSC. </p> </div>

Related collections

Most cited references 37

Record: found
Abstract: found
Article: not found

Global and Local Architecture of the Mammalian microRNA–Transcription Factor Regulatory Network

Reut Shalgi, Daniel Lieber, Moshe Oren … (2007)

Introduction microRNAs (miRs) are short RNAs that post transcriptionally regulate messenger RNAs. Two main mechanisms for such effects are degradation of the target mRNA, and inhibition of its translation [1]. In recent years considerable progress within multiple genomes was obtained in the experimental identification of genes encoding for miRs [2–4], and in tools for the identification of target genes of miRs, based on miR sequences and the sequence of the targets' 3′ untranslated regions (UTRs) [5–11]. Compared with the regulation of transcription, the study of the regulatory networks spanned by miRs is only at its beginning. When it comes to transcriptional regulation, a lot is known about the main players and the interactions between them. Transcription factors (TFs) are well-characterized [12], and promoter binding motifs are available in a diversity of species [13]. The combinatorial interactions between TFs have been explored [14,15] as well as the global level properties of the transcription regulatory network [16]. In addition, the local structures of the network have been intensively investigated. It was found in several species that the transcription regulatory network may be decomposed into elementary building blocks, or network motifs, that recur in the network more than expected by chance, and that these motifs likely perform local “computations,” such as the detection of signal persistency or the coordinated gradual activation of a set of genes [17–20]. When it comes to posttranscriptional regulation, and in particular to the miR world, most of the parallel knowledge is lacking. While we do know about many miRs in multiple genomes [1], their targets are predicted with relatively limited accuracy [21]. Even more obvious is the lack of knowledge about the structure of the miR regulatory network, and about the potential interface between this network and the transcriptional one. In similarity to TFs, miRs are expected to work in combinations on their target genes [7]. The target specificity-determining site of the miRs is often short (seven to eight nucleotides [9]), hence some genes that contain a match to a single miR in their 3′ UTRs may represent false positive assignments. Thus, combinatorial interactions among the miRs are probably necessary to specify more precisely the set of affected targets of each miR. As in the realm of transcription regulators [14], combinatorics may also have the advantage of allowing multiple sources of information, each represented by a single miR, to be integrated into the regulation of individual transcripts. Since TFs regulate mRNA production, and miRs regulate transcript stability and its translation, an attractive possibility is that miRs and TFs cooperate in regulating shared target genes. This possibility is appealing since a gene that is regulated through multiple mechanisms may be tuned at a level of precision that is higher than what may be obtained by either mechanism alone. In addition, as with any other regulatory agent in cells, the question “what regulates the regulator” is of prime importance, as it may allow the exposure of multiple levels of hierarchies and their design within a control network. It is thus crucial to understand whether TFs and miRs collaborate in gene regulation, and also to characterize regulatory interactions that miRs and TFs may exert on each other. In similarity to the transcription network, local network motifs might exist which may also consist of miRs. One attractive role for such motifs has been suggested in a developmental context—to canalize “noise” in gene expression [22]. However, actual realization of such motifs remains to be explored. Here we report extensive combinatorial interactions among miRs and between miRs and TFs. We found hundreds of miRs target hubs—genes regulated by dozens of miRs—which are involved in a diversity of developmental processes and in transcription regulation. The miR–TF regulatory network features several motifs in which TF and miR partners that are suggested to regulate multiple target genes often exert regulation on one another. Results Connectivity Distributions in the miR–Gene Network We used two datasets of miRs and their predicted target genes: TargetScan [8,9] and PicTar [7]. The miRs used in this analysis are characterized by being evolutionarily conserved, and, in addition, their targets were defined based on conservation in orthologous genes in four species (human, mouse, rat, and dog). This evolutionary conservation criterion was assumed to constitute a good filter for false positive assignments of miRs to genes [9,23]. Yet, it must be emphasized that the accuracy of such assignments is still limited [21] (see “noise tolerance analysis” in Materials and Methods). Altogether we analyzed 8,672 and 9,152 human (RefSeq) genes in the TargetScan and PicTar datasets, respectively, that have at least one predicted miR binding site in their 3′ UTR, and a total of 138 miRs and 178 miRs in the respective datasets. We constructed a matrix whose rows are genes and columns are miRs, in which the ij-th element is “1” if gene i contains a predicted binding site for miR j in its 3′ UTR, and “0” otherwise. We created one such matrix for each of the two miR target prediction datasets. For the sake of clarity, from here on we will say interchangeably that “a miR targets a gene” or that “a gene contains in its 3′ UTR a predicted binding site for a miR.” We first characterized the matrix by the distribution of degree connectivity of each gene and of each miR. Figure 1A shows the distribution of the number of miRs assigned per gene, while Figure 1B shows the distribution of number of genes assigned to each miR. We compared each distribution with a set of distributions, each derived by randomization of the original matrix according to two alternative null models. Along with the distribution of number of miRs per gene (Figure 1A), we also plotted 100 distributions obtained after randomizing each of the columns in the matrix. In this randomization we preserved the number of genes per miR, yet assigned genes at random to each miR. The distributions obtained after the randomization differed markedly from the original distribution, both in terms of width and shape. While in the randomized distributions genes rarely have more than ten different miRs in their 3′ UTR, in the original distribution there are hundreds of genes subjected to extensive predicted miR regulation. In Figure 1B we also show the distribution of number of genes per miR. Along with it is shown a set of distributions obtained by randomizing each of the rows in the matrix, namely by randomly assigning miRs to each gene, preserving the real number of miRs predicted to target each gene, as in the original matrix. Here, too, the randomized distributions differed from the original one both in shape and width; the original data contains multiple miRs which appear to target more than 400 genes, significantly higher than the number that would be obtained by merely preserving the statistics of number of miR sites in genes UTRs. These observations lead us to highlight some special properties that seem to be unique to the miR regulatory network. Figure 1 miRs and Target Genes in the TargetScan Dataset (A) Distribution of the number of different miRs regulating each target gene in the TargetScan dataset. The thick red line represents the distribution in the original datasets, while each of the thin blue lines represents the distribution in one of the column-randomized matrices. The matrix contains only genes with at least one predicted site in their 3′ UTR. In each randomization, we shuffled the assignment of miRs to their targets, keeping constant the number of targets per miR. (B) Distribution of number of targets per miR in the TargetScan dataset. In the thick red line we depicted the original distribution, while each blue thin line represents the distribution in one of the 100 row-randomized matrices, which preserve the distribution of number of miRs targeting each gene. Target Hubs—Genes with Extensive miR Regulation The distribution of number of miRs regulating each target gene (Figure 1A) has a long right tail in contrast to the distributions in the randomized matrices that looked Gaussian (as befits a sum of independent random variables). We thus focused on the genes in that tail of the distribution (which are targeted by more than 15 miRs and 20 miRs in the TargetScan and PicTar datasets, respectively; see Materials and Methods for further details and cutoff justification). We named these genes target hubs following a recent definition of genes regulated by multiple TFs in yeast [24]. There are 470 such genes in the TargetScan dataset. We made similar observations with the PicTar dataset and identified 834 target hubs (see Figure S1)—the set of target hubs based on the TargetScan dataset has an 81% overlap with the target hubs defined by PicTar dataset. Inspecting the target hubs genes' annotations (using Gene Ontology, GO), we found that they are highly enriched for developmental processes, specifically for muscle development and nervous system development, as well as for TFs and transcription regulators (see Table 1 for enrichment statistics). Among the transcription regulators in the set of target hubs are included RUNX1, E2F-3, N-MYC, and SP3. Another very intriguing fact is that the Ago1 gene, one of the key components of the human RISC (RNAi induced silencing complex), is also a target hub, as in the dataset it appears to be potentially regulated by multiple miRs. Table 1 TargetScan Target Hubs GO Functional Enrichment We suspected, however, that the fact that target hubs host many miR binding sites may result from potentially longer 3′ UTRs [23]. Although we found that target hubs have a distribution of 3′ UTR lengths that is significantly longer than that of the rest of the genes in the current analysis (p-value = 4 × 10−85 and p-value = 3 × 10−101 for TargetScan and PicTar target hubs, respectively, using the Kolmogorov-Smirnov test), we still realized that many of them have relatively short 3′ UTRs (Figure S2A and S2B). To test whether the high number of miR binding sites in the target hubs is a simple reflection of their 3′ UTR lengths, we performed a randomization test, in which we sampled 100 times random gene sets from the entire dataset with the same or very similar length distributions as that of the target hubs (see Materials and Methods). We found that such gene sets always have a significantly lower average number of miR sites per gene compared with the target hubs (see Figure S3A). We further calculated the density of different miRs in the 3′ UTRs [23]. Density was defined as number of different miRs targeting a gene divided by 3′ UTR length. Remarkably, we found that the miR density in the target hubs is significantly higher than in the rest of the genes in the dataset (p-value = 2 × 10−85 and p-value = 6 × 10−124 for the TargetScan and PicTar target hubs, respectively, using the Kolmogorov-Smirnov test; the means are 2.84 and 1.80 times higher in the TargetScan and PicTar dataset means, respectively; see Figure 2 and Figure S2C for the entire distributions). We concluded that target hubs are rich in binding sites for different miRs to an extent that cannot be explained solely by their 3′ UTRs lengths. Figure 2 Distribution of the density of miRs in the 3′ UTRs of target hubs (thick red line) and all the genes (thin blue line) in the TargetScan dataset (all genes included in this figures have at least one miR site predicted in their 3′ UTR). The log10 densities were binned into bins of 0.1, and relative frequencies were plotted. Same analysis for the PicTar dataset is in Figure S2. Realizing that density of miR binding sites may be an important property by itself, we also used an alternative definition for target hubs—genes with particularly high density of miRs in their 3′ UTRs. We collected the genes in the top 85th percentile of the miR binding site density spectrum, then we performed a similar GO enrichment analysis to see whether particular functionalities were enriched among the genes with a high density of miR binding sites. Reassuringly, most of the functionalities that were enriched among the set of target hubs defined by number of differnet miRs were also significant in the set of high density target hubs (see Table 1). Moreover, we found that genes that were target hubs according to only one of the two definitions (i.e., genes that are not in the overlap of the two sets) were still significantly enriched for functionalities such as transcription regulator activity and development (unpublished data). A Combinatorial Network of miR Interactions Combinatorial interactions are a fundamental property of the transcription networks [25]. It may be anticipated that, similarly to TFs, miRs may work in combinations. One way to predict pairs of coregulating miRs is to ask which pairs show a high rate of co-occurrence in the same target genes' 3′ UTRs. A common statistical test in the field, previously used in the context of promoter motifs and TF binding site [26–28], is the cumulative hypergeometric statistic. According to this model, given the rate of occurrence of each of the regulators alone, and the total number of genes in the analysis, a p-value is computed on the size of the set of genes that are shared between the two regulators. The main assumption of this model, that assignment of a gene to the first regulator is independent of the assignment to the second one, is likely fulfilled in the context of fixed-length promoters. Yet when it comes to 3′ UTRs of varying length, the assumption does not hold anymore. Some genes, e.g., those with long 3′ UTRs, have a higher chance to contain predicted binding sites for miRs, hence a p-value calculated based on the hypergeometric model may overestimate the significance of the co-occurrence rate. We have thus devised an alternative, randomization-based test for identifying significantly co-occurring miR pairs. The model was designed such that it will capture the underlying distributions in Figure 1A and 1B, and test whether a given pair of miRs co-occurs at a higher rate, considering the above distributions as a background. For each pair of miRs, i and j, with their set of targets, Targets(i) and Targets(j), respectively, we calculated the “Meet/Min” score [29,30] defined in the present case as: namely, the size of the set of genes that contain sites for the two miRs together, divided by the smaller of the two sets of targets (we filtered from the calculation for each i,j pair, 3′ UTRs in which the sites for i and j are physically overlapping to avoid overestimation of significance of miR pairs with an overlapping or similar seed, see Materials and Methods for details). Yet this score is not a statistic, i.e., it lacks an estimate of the probability to obtain such score (or better) by chance given an appropriate null model. Following previous works [20], we used a null model that preserves for each gene the number miRs assigned to it, and for each miR the number of genes assigned to it in the input data. We generated 1,000 randomized matrices according to this null model. In each such matrix we randomized the original matrix in 100,000 steps, using an edge-swapping algorithm [20]. For each such randomized matrix we computed again the Meet/Min score for all pairs of miRs. The co-occurrence p-value for a pair of miRs was computed according to the pair's Meet/Min score and the population of 1,000 Meet/Min scores obtained for that same pair in each of the 1,000 edge-swapped matrices. The p-value for the pair is defined as the fraction of the 1,000 randomized matrices in which the Meet/Min score of that pair is greater than or equal to the Meet/Min score of the pair in the original matrix. In addition to calculating a score of co-occurrence, we also calculated, using the same formalism, a score that captures the tendency of every two miRs to avoid residing within shared 3′ UTRs. We will regard a pair of miRs that co-occur in the original matrix significantly less frequently than in the edge-swapped matrices as avoiding each other. Given the Meet/Min score of co-occurrence for a pair of miRs, and the Meet/Min scores obtained for that pair in the 1,000 edge-swapped matrices, we calculated the fraction of randomized scores that were lower than or equal to that obtained in the original matrix for that pair, as the avoidance p-value of a miR pair. In both cases of co-occurrence and avoidance, we used the false discovery rate (FDR) to control for the testing of multiple hypotheses. In the case of co-occurring miR pairs, using a restrictive FDR threshold (q-value = 0.05), we obtained 107 pairs with a significant p-value in the TargetScan dataset, and 199 pairs in the PicTar dataset (interestingly, the ratio between the number of interactions in the two datasets (∼0.54) is very close to the ratio expected based on the square of relative number of miRs in each dataset (∼0.6)). We created a combinatorial network based on the significant co-occurring miR pairs. The top miR pairs are given in Table 2 and are also depicted in Figures 3A and S4A. The full list of significant pairs is provided in Tables S1 and S2. This combinatorial network consists of several levels of hierarchy. At the top (Figure 3A) are a handful of miRs that interact with a relatively large number of miR partners, while at the bottom are “end-nodes” with very few miR partners each. Examination of the degree distribution in the miR combinatorial network revealed a power law with a slope of about −1.5 and R2 = ∼0.89 in TargetScan and R2 = 0.94 in PicTar (Figures 3B and S4B), indicating that the network of coregulating miRs is scale-free (alternative FDR cutoffs also resulted in scale-free networks with R2 always bigger than 0.72). Interestingly, expression data of the miRs provides some support for the predicted regulatory interactions between them. We found that coexpressed miRs tended to have relatively high co-occurrence scores, and significant co-occurrence p-values, while miR pairs with negatively correlated expression tended to avoid residing in shared 3′ UTRs (see below). Table 2 Top 20 Most Significant Pairs of Coregulating miRNAs in the TargetScan Network Figure 3 miR Co-Occurrence Network in the TargetScan Dataset (A) The TargetScan miR co-occurrence network, at FDR level of 0.05. A node represents a miR and an edge connects between pairs of miRs with significant rate of co-occurrence. The nodes in the figure are arranged from most highly connected on the top, to most lowly connected, on the bottom. For interactive viewing of the network, using Pajek (http://vlado.fmf.uni-lj.si/pub/networks/pajek/), see Datasets S1 and S2. (B) Degree distribution in the TargetScan miR combinatorial regulation network (co-occurring miR pairs that passed FDR of 0.05). Coordinated Regulation of Target Genes by miRs and TFs A potential regulatory design in the gene expression network is that genes belonging to the same regulon will be coregulated not only at the transcriptional level, but also posttranscriptionally [31]. One potential realization of this design may be that a particular miR and a particular TF would regulate common targets. A simple means to identify some of the cases of regulatory cooperation between a miR and a TF may be to find TF–miR pairs that co-occur in a large set of shared targets compared with the size expected by chance. Similar to the case of miRs sites in 3′ UTRs, we considered a TF to be present in a human gene's promoter only if its occurrence in the promoter is conserved in the promoters of orthologous genes from mouse and rat [32] (as taken from UCSC, see Materials and Methods). We then created a matrix whose rows are the genes and columns are TFs, with a “1” for the i-th gene and the j-th TF if the TF binding site (TFBS) occurs in the gene's promoter and “0” otherwise. To identify pairs of TFs and miRs that cooperate in regulating shared target genes, we looked for TF–miR pairs with a high rate of co-occurrence in the promoters and 3′ UTRs of the regulated genes. We tested the co-occurrence in shared genes of each of the 409 position specific scoring matrices (PSSMs) representing TF binding sites in TRANSFAC [13] with each of the 138 and 178 miRs in the TargetScan and PicTar databases, respectively. A PSSM and a miR are said to co-occur in the same gene if the PSSM has a conserved binding site in the promoter of the gene and the miR has a conserved predicted site in the gene's 3′ UTR. We used two statistical models to calculate the significance of rate of TF–miR co-occurrence, and ultimately considered TF–miR pairs that were found to be significant according to both tests. First, a hypergeometric p-value was calculated based on the number of genes that contain a TFBS in their promoter, the number of genes that contain a miR site in their 3′ UTR, and the number of genes that contain both the TF and the miR sites (see Materials and Methods for details). We computed such p-values on all TF–miRs pairs and set a threshold on the p-values obtained to account for the multiplicity of hypotheses, using FDR. Using an FDR q-value of 0.3, we obtained 111 miR-TF pairs with significant p-values using the TargetScan dataset and 1,263 miR-TF pairs with significant p-values using the PicTar dataset (see Materials and Methods for number of pairs with more stringent q-values). Reassuringly, there is a high overlap between the TargetScan and PicTar networks (68.7% of the TargetScan miR–TF network pairs were also found to be significant pairs in the PicTar network). The hypergeometric p-value has the advantage of being an analytical model with essentially unlimited resolution. Also, unlike the above situation of miR co-occurring pairs, which exhibited inherent dependency between the two regulators, the present case of TF–miR interaction does not present such limitation (and is in fact identical to the classical cases in which hypergeometric model is used [33]). Nevertheless, we decided to also back up the hypergeometric-based predictions with a randomization test, very similar to the one presented above for the case of miR co-occurrence, that preserves the distribution of number of regulators of each gene, the number of targets of each TF, and the number of targets of each miR in the input datasets. We calculated the co-occurrence rates and p-values of all TF–miR pairs, and used FDR as above to account for the multiplicity of hypotheses (see Materials and Methods for details). Reassuringly, 93% and 72% of the hypergeometric-based TF–miR interactions from the TargetScan and PicTar datasets, respectively, were also supported by this alternative model. The rest of the analyses were based on TF–miR pairs that passed the two statistical tests using FDR; there were 104 pairs in the TargetScan dataset and 916 pairs in the PicTar dataset. For simplicity we term a TF and a miR that significantly co-occur as partners. Table 3 lists the top TF–miR partners. The full networks of TF–miR partners can be downloaded as Tables S3–S5, and interactively viewed in Datasets S3–S5. Table 3 Top 20 Most Significant Pairs of Coregulating miRNAs and TFs in the TargetScan and PicTar Networks The Network of miR–TF Coregulation Reveals Recurring Local Architectures—Network Motifs Recently it was suggested that in circuits composed of a miR and a TF, in which these two regulators target the same genes, the TF may also exert a regulatory effect on the miR with which it coregulates the target genes [22]. It was suggested that such a feed-forward loop (FFL) [19,20], a well-known local feature of many biological networks, may have a beneficial function. An FLL consisting of a TF and a miR could act as a switch for developmental and other programs in cells, since it may acquire biological systems with robustness to noise by means of canalization of perturbations [22]. We wanted to check whether in any of the significant miR-TF partners discovered above, the miR and its partner TF may regulate each other. We determined how many of the miR–TF partner pairs (out of 104 pairs in the TargetScan dataset and 916 pairs in the PicTar dataset) had a conserved TF binding site of the partner TF in the putative upstream regulatory region of the partner miR (see Materials and Methods for definition of miRs' upstream putative regulatory regions). Interestingly, we found that ten of the TF–miR pairs in the TargetScan dataset (9.6% of the pairs), and 75 out of 916 pairs in the PicTar dataset (8.2%) fulfilled that additional requirement (see Figure 4). To establish whether this rate was significant, we carried out a randomization test (see Materials and Methods) in which we computed, in 10,000 randomized sets of TF–miR pairs, the rate of formation of a regulatory interaction between the TF and the miR. In the TargetScan network, we obtained a modest p-value of 0.024; however, in both PicTar networks we obtained the minimal possible p-value, 93%) in the hypergeometric derived set. The final set of significant pairs in the miR–TF network is presented in FDR q-value cutoffs of 0.1, 0.2, and 0.3. With q-value of 0.1 we obtained 20 TF–miR pairs with significant p-value using the TargetScan dataset, and 267 using the PicTar 10 kb dataset, and 70 using the PicTar 5 kb dataset. With a q-value of 0.2 we obtained 60 TF–miR pairs with significant p-value using the TargetScan dataset, and 555 using the PicTar 10 kb dataset, and 261 using the PicTar 5 kb dataset. With 0.3 we obtained 104 TF–miR pairs with significant p-value using the TargetScan dataset, and 916 using the PicTar 10 kb dataset, and 497 using the PicTar 5 kb dataset. miRs clusters and regulatory regions. As was shown in the past [41], miRs may be clustered on the genome, and are often transcribed as one unit. Therefore, to predict regulatory regions of miRs (i.e., proximal as well as potentially more distant promoters or enhancers) we had to first cluster miRs on the human genome. We mapped all 461 pre-miRs in miRBase (http://microrna.sanger.ac.uk, accessed June 2006) [47,48] onto the human genome and clustered them according to physical proximity (genomic locations of miRs were taken from UCSC hg17 and some miRs were mapped from hg18 back to hg17 using the UCSC “lift genome” web service). Two pre-miRs, that are consecutive on the genome, were considered belonging to the same cluster if the distance between them was shorter than a cutoff, provided that they are transcribed from the same strand. We kept adding miRs to clusters until we hit the first distance that was larger than the cutoff. To learn a meaningful cutoff from the data, we plotted the distribution of distances between all neighboring pre-miRs in the genome. Interestingly, we found the distribution to be bimodal—distances below and above 10 kb (on a log scale, Figure 6A) were highly represented in contrast to a lower representation at about 10 kb. This indicated that a reasonable cutoff on the distance between two adjacent miRs that still belong to the same cluster may be 10 kb. Using this clustering procedure we generated 301 clusters, the majority of which (∼82.39%) consists of a single miR; the cluster with the highest number of miRs contains 43 miRs (see Figure S7 for the distribution of number of miRs per cluster). In a previous study, which was based on 207 miRs (compared with the 461 used here), miRs were clustered using a different cutoff [49]. When we repeated our cluster analysis with the current set of miRs, with the previous cutoff, we got similar clustering, 94% of the present clusters are identical to the clusters generated with the alternative cutoff and average cluster lengths are very similar (unpublished data). Figure 6 Analysis of miR Clusters in the Human Genome (A) Distribution of distances between all neighboring pre-miR genes in the human genome. (B) Distribution of tissue expression correlations between pairs of miRs: all possible pairs in the data (thin blue line) and pairs of miRs which reside in shared clusters (thick red line). In the inset are shown tissue expression correlations between pairs of miRs in the same genomic clusters versus distances between them. (C) Distribution of number of conserved TFBS 30 kb upstream of the 5′ most nucleotide in each miR clusters. Conserved TFBSs were taken from UCSC hg17. Reassuringly, using expression data of miRs across tissues [34] we found that miRs that belong to the same cluster have a significant tendency to be coexpressed compared with miRs that do not map to shared clusters (Figure 6B). This tendency is preserved even in cases where miRs that belong to the same cluster are relatively far from each other on the genome (Figure 6B, inset). We have then defined, as a putative regulatory region of miRs, the sequence that lies 10 kb upstream of the 5′ most pre-miR in each miR cluster. The 10 kb promoter length was determined from the data as follows. A distribution of number of conserved TFBS upstream of clusters was generated (Figure 6C). We found that the number of conserved TFBS gradually declined as a function of the distance from the putative 5′ end of the cluster, with a plateau obtained at about 10 kb upstream. The distribution was rather noisy, probably due to the fact that primary-miR transcripts are much longer than the precursor miR we relate to (e.g., the primary transcript of the miR-17–92 cluster is C13orf25, which is 6,795 bp long [45]), and thus the transcription start site (TSS) taken here is only crudely defined. We considered the presence of a TFBS in a miR promoter only if such occurrence was conserved in mouse and rat, as taken from the UCSC hg17 conserved track in the relevant regions. Transcription factor binding sites. We used predicted binding sites for all human mouse and rat PSSMs from TRANSFAC [13] version 8.3, as they are defined by the UCSC hg17 genome assembly, in the tfbsConsSites (http://genome.ucsc.edu/) and tfbsConsFactors. All RefSeq genes genomic locations were taken from hg17. To determine the length of upstream regulatory regions, we measured the number of conserved TFBS upstream RefSeq genes as a function of distance from TSS (see Figure S6). The result shows that the signal decays and plateaus between 5 kb and 10 kb upstream of the TSS. We hence chose to work with two alternative cutoffs of promoter length, 5 kb and 10 kb. The regulatory regions thus defined probably consist of proximal promoters as well as distant enhancers. The recent Affymetrix (http://www.affymetrix.com) promoter chip for detection of ChIP experiments with TF binding in human promoters also consists of probes that span 10 kb of regulatory regions, and future experiments with this chip and as many TFs as possible will allow a better delineation of regulatory regions boundaries. Although we used regulatory regions which are longer than the common definition, our use of evolutionary conservation filter gives confidence in the present regulatory region definitions. Feed-forward loop statistics. FFL TF → miR: for all the significant pairs of coregulators (i.e., TF–miR partners that co-occur in a significantly high number of targets) we investigated whether the TF has a binding site in the putative promoter of the miR cluster from which the miR partner is transcribed. In some cases in which the mature miR sequence is transcribed from more than one genomic locus, all possible regulatory regions of the relevant miR clusters were examined. In addition, each PSSM may belong to a family of PSSMs, with similar binding sites, representing the same TF (a family was defined as several PSSMs representing the same TF, as determined from the UCSC hg17 tfbsConsFactors track). Thus, PSSM–miR pairs are treated as TF–miR pair, and given a pair of PSSM–miR partners, we say that the PSSM's TF regulates the miR if at least one of the PSSMs that corresponds to that TF has a match in the regulatory region of the miR partner (the same procedure was carried out in the randomizations described below). For testing the FFL miR → TF configuration, we had to connect first between TRANSFAC PSSMs and the genes encoding the TFs that bind these PSSMs. For that, PSSMs were mapped to the TF they represent which in turn was mapped to a SwissProt ID. These two mappings were done using the UCSC hg17 tfbsConsFactors track. These SwissProt IDs were then mapped to RefSeq IDs, for which the data on miR targets was maintained. This information served also in the process of indirect FFL search; for each of the TF–miR partners, we checked whether the miR is regulated by another mediator TF, which in turn is regulated by the partner TF. We note that not all TFs had a corresponding SwissProt ID in the UCSC hg17 tfbsConsFactors track, and therefore not all pairs served as candidates for the FFL miR → TF and the indirect FFL; only in 74 of the 104 (71%) TargetScan significant pairs, and in 680 of 916 (74%) of the PicTar pairs, could the PSSM be mapped to a RefSeq gene. The following procedure was used for the calculation of the significance of the FFLs and indirect FFL in the PicTar and TargetScan miR–TF networks. Since there were 104 and 916 pairs of miR–TF partners in the two respective networks, we have drawn 10,000 times the same number of random pairs of TFs and miRs out of all the possible pairs in each network. The number of each FFL and indirect FFL was recorded in each randomization and a p-value (and a corresponding z-score) on the hypothesis that a given network motif is over-represented in the network was taken to be the number of random sets with a greater or equal number of motifs in it. miR and mRNA tissue expression data. The expression profiles of 150 miRs across five healthy human tissues and organs (brain, liver, thymus, testes, and placenta) were previously measured using miR-dedicated microarrays [34]. miRs from the chips were mapped to PicTar and TargetScan; they cover 154 and 87 of the miRs in the two respective datasets. In addition, we used data from [35] for human mRNAs expression across the same set of tissues. Both sets of expression data were column centered (chip-wise centering: each chip's values were divided by the chip mean to account for differences in chip intensities) and then log2 transformed. Regarding mRNA expression chips, we particularly focused on genes coding for the TFs that participated in our analysis. Using the above mapping of PSSMs to their corresponding TF genes, we had a total of 127 TFs that could be matched to at least one probe set in the mRNA expression dataset [35]. We examined the tissue expression correlation of all significantly co-occurring miR and TF pairs for which we had an expression profile. When more than one gene was attributed to the same TF, we chose for each pair of TF and miR the one with the highest absolute value of correlation coefficient out of all options. We did that consistently both for the background statistics of all possible TF–miR pairs and for the predicted TF–miR partners. In total we calculated correlation coefficients for 361 such TF–miR partners out of 916 partners in PicTar, and for 30 out of 104 partners in TargetScan. The miR expression data [34] consisted of five healthy tissues, and HeLa cells, while the mRNA study that we focused on [35] overlapped with the miR data only in the five tissues. Therefore when we compared expression between miRs and TFs we only used the five healthy tissues, and when we compared expression of miR pairs we used all six samples. Noise-tolerance analysis. The assignments of miRs to targets are known to be of limited accuracy [21] . We thus wanted to assess the noise tolerance of our results. We adopted a procedure previously utilized for the case of network motifs in the bacterial transcription network [20]. We experimented with different percentages of the connections in the network that were randomly removed or added and the significance of the present FFL motifs was assessed for each case. Similarly to the findings in the E. coli network, we found that up to 20%–30% of the edges can be added or removed without appreciable effect on the FFL significance. Supporting Information Dataset S1 Pajek Input File for the miR Co-Occurrence Network, the TargetScan Dataset (Significant Co-Occurring miR Pairs with FDR q-Value 0.05) All networks in the Dataset files can be interactively viewed using the Pajek software, which can be freely downloaded from (http://vlado.fmf.uni-lj.si/pub/networks/pajek/). (12 KB TXT) Click here for additional data file. Dataset S2 Pajek Input File for the miR Co-Occurrence Network, the PicTar Dataset (Significant Co-Occurring miR Pairs in FDR q-Value 0.05) (20 KB TXT) Click here for additional data file. Dataset S3 Pajek Input File for the Network of miR–TF Coregulating Pairs This graph depicts all the significant miR–TF pairs in the TargetScan network, in addition to all the FFLs. A red node is a TF and a green node is a miR, and a blue edge is drawn if the TF and the miR are co-occurring partners. A yellow edge connects between a TF and a miR if, in addition to having a high rate of co-occurrence, they also form a FFL TF → miR; a pink edge represents the FFL miR → TF motif, while orange edge represents a FFL miR ← → TF (in all cases the set of target genes is not explicitly shown). (16 KB TXT) Click here for additional data file. Dataset S4 Pajek Input File for the Network of miR–TF Coregulating Pairs This graph depicts the 100 most significant pairs in the PicTar (10 kb) network, in addition to all the FFLs. (86 KB TXT) Click here for additional data file. Dataset S5 Pajek Input File for the Network of miR–TF Coregulating Pairs This graph depicts the 100 most significant pairs in the PicTar (5 kb) network, in addition to all the FFLs. (55 KB TXT) Click here for additional data file. Figure S1 Distribution of miRs to Target Gene Assignments in the PicTar Dataset (A) Distribution of the number of different miRs regulating each target gene in the PicTar dataset. The thick red line represents the distribution in the original datasets, while each of the thin blue lines represents the distribution in one of the column-randomized matrices. The matrix contains only genes with at least one predicted site in their 3′ UTR. In each randomization, we shuffle the assignment of miRs to their targets, keeping constant the number of targets per miR. (B) Distribution of number of targets per miR in the PicTar dataset. In the thick red line we depicted the original distribution, while each blue thin line represents the distribution in one of the 100 row-randomized matrices, which preserve the distribution of number of miRs targeting each gene. (1.6 MB EPS) Click here for additional data file. Figure S2 miR Binding Sites and 3′ UTR Length in the TargetScan and PicTar Datasets A dot plot depicting number of miRs targeting each gene and its 3′ UTR length of the target hubs, high miR number target hubs in green, high density target hubs in red, genes that are target hubs according to both criteria in magenta and the rest of the genes in blue for the (A) TargetScan dataset and (B) PicTar Dataset. (C) Distribution of the miR densities in the 3′ UTRs of target hubs (thick red line) and all the genes (thin blue line) in the PicTar dataset (all genes included in this figures have at least one miR site predicted in their 3′ UTR). The log10 densities were binned into bins of 0.1, and relative frequencies were plotted. (1.6 MB EPS) Click here for additional data file. Figure S3 Click here for additional data file. Figure S4 miR Pairs Interaction Network in the PicTar Dataset (A) The miR pairs interaction network in the PicTar database. (B) Degree distribution in the PicTar miR combinatorial regulation network (co-occurring miR pairs that passed FDR of 0.05) (1.6 MB EPS) Click here for additional data file. Figure S5 Positively Correlated miR Pairs Tend To Have Significant Co-Occurrence p-Values while Negatively Correlated Pairs Tend to Avoid Residing in the Same 3′ UTRs Highly expression correlated miR pairs tend to have significant co-occurrence or p-values, while negatively correlated pairs tend to have significant avoidance p-values. The figures depict the Kolmogorov-Smirnov p-values for the hypotheses that correlated miR pairs have lower co-occurrence p-values than the rest of the pairs. Correlated pairs were defined according to correlation cutoffs (depicted on the x-axis), with positively correlated pairs in blue, negatively correlated pairs in green. Positively correlated miR pairs tend to have significant co-occurrence p-values in both TargetScan (A) and PicTar (C). Negatively correlated pairs tend to have significant avoidance p-values in both TargetScan (B) and PicTar (D). (3.9 MB EPS) Click here for additional data file. Figure S6 Distribution of Number of Conserved TFBS 30 kb Upstream of TSS of RefSeq Protein-Coding Genes (1l KB EPS). Click here for additional data file. Figure S7 Distribution of Number of miRs per Cluster As seen, ∼82% of the 301 clusters contain a single miR. (12 KB EPS) Click here for additional data file. Table S1 Significant Co-Occurring miR Pairs in the TargetScan Dataset (30 KB XLS) Click here for additional data file. Table S2 Significant Co-Occurring miR Pairs in the PicTar Dataset (38 KB XLS) Click here for additional data file. Table S3 Significant Co-Occurring miR–TF Pairs in the TargetScan Network (32 KB XLS) Click here for additional data file. Table S4 Significant Co-Occurring miR–TF Pairs in the PicTar Network, Taking 10 kb Regulatory Regions for Protein Coding Genes (172 KB XLS) Click here for additional data file. Table S5 Significant Co-Occurring miR–TF Pairs in the PicTar Network, Taking 5 kb Regulatory Regions for Protein Coding Genes (103 KB XLS) Click here for additional data file. Table S6 Indirect FFLs in the TargetScan Dataset (22 KB XLS) Click here for additional data file. Table S7 Indirect FFLs in the PicTar Dataset Taking 10 kb Regulatory Regions for Protein Coding Genes (47 KB XLS) Click here for additional data file. Table S8 Indirect FFLs in the PicTar Dataset Taking 5 kb Regulatory Regions for Protein Coding Genes (29 KB XLS) Click here for additional data file.

0 comments Cited 208 times – based on 0 reviews      Review now

Bookmark

Record: found
Abstract: found
Article: not found

Pathogenesis of biliary atresia: defining biology to understand clinical phenotypes

Akihiro Asai, Alexander Miethke, Jorge A. Bezerra (2015)

Biliary atresia is a severe cholangiopathy of early infancy that destroys extrahepatic bile ducts and disrupts bile flow. With a poorly defined disease pathogenesis, treatment consists of the surgical removal of duct remnants followed by hepatoportoenterostomy. Although this approach can improve the short-term outcome, the liver disease progresses to end-stage cirrhosis in most children. Further improvement in outcome will require a greater understanding of the mechanisms of biliary injury and fibrosis. Here, we review progress in the field, which has been fuelled by collaborative studies in larger patient cohorts and the development of cell culture and animal model systems to directly test hypotheses. Advances include the identification of phenotypic subgroups and stages of disease based on clinical, pathological and molecular features. Stronger evidence exists for viruses, toxins and gene sequence variations in the aetiology of biliary atresia, triggering a proinflammatory response that injures the duct epithelium and produces a rapidly progressive cholangiopathy. The immune response also activates the expression of type 2 cytokines that promote epithelial cell proliferation and extracellular matrix production by nonparenchymal cells. These advances provide insight into phenotype variability and might be relevant to the design of personalized trials to block progression of liver disease.

0 comments Cited 75 times – based on 0 reviews      Review now

Bookmark

Record: found
Abstract: found
Article: not found

The Cholangiopathies.

Konstantinos Lazaridis, Nicholas F LaRusso (2015)

Cholangiocytes (ie, the epithelial cells that line the bile ducts) are an important subset of liver cells. They are actively involved in the modification of bile volume and composition, are activated by interactions with endogenous and exogenous stimuli (eg, microorganisms, drugs), and participate in liver injury and repair. The term cholangiopathies refers to a category of chronic liver diseases that share a central target: the cholangiocyte. The cholangiopathies account for substantial morbidity and mortality given their progressive nature, the challenges associated with clinical management, and the lack of effective medical therapies. Thus, cholangiopathies usually result in end-stage liver disease requiring liver transplant to extend survival. Approximately 16% of all liver transplants performed in the United States between 1988 and 2014 were for cholangiopathies. For all these reasons, cholangiopathies are an economic burden on patients, their families, and society. This review offers a concise summary of the biology of cholangiocytes and describes a conceptual framework for development of the cholangiopathies. We also present the recent progress made in understanding the pathogenesis of and how this knowledge has influenced therapies for the 6 common cholangiopathies-primary biliary cirrhosis, primary sclerosing cholangitis, cystic fibrosis involving the liver, biliary atresia, polycystic liver disease, and cholangiocarcinoma-because the latest scientific progress in the field concerns these conditions. We performed a search of the literature in PubMed for published papers using the following terms: cholangiocytes, biliary epithelia, cholestasis, cholangiopathy, and biliary disease. Studies had to be published in the past 5 years (from June 1, 2009, through May 31, 2014), and non-English studies were excluded.

0 comments Cited 68 times – based on 0 reviews      Review now

Bookmark

All references

Author and article information

Journal

Title: Hepatology

Abbreviated Title: Hepatology

Publisher: Wiley

ISSN: 02709139

Publication date Created: February 2018

Publication date (Print): February 2018

Publication date (Electronic): January 02 2018

Volume: 67

Issue: 2

Pages: 676-689

Affiliations

[1 ]The Liver Care Center and Divisions of Gastroenterology; Hepatology and Nutrition; Cincinnati OH

[2 ]Biomedical Informatics of Cincinnati Children's Hospital Medical Center and the Department of Pediatrics of the University Of Cincinnati College of Medicine; Cincinnati OH