Abstract

   The regulation of complex cellular activities in palmitate treated
   HepG2 cells, and the ensuing cytotoxic phenotype, involves cooperative
   interactions between genes. While previous approaches have largely
   focused on identifying individual target genes, elucidating interacting
   genes has thus far remained elusive. We applied the concept of
   information synergy to reconstruct a “gene-cooperativity” network for
   palmititate-induced cytotoxicity in liver cells. Our approach
   integrated gene expression data with metabolic profiles to select a
   subset of genes for network reconstruction. Subsequent analysis of the
   network revealed insulin signaling as the most significantly enriched
   pathway, and desmoplakin (DSP) as its top neighbor. We determined that
   palmitate significantly reduces DSP expression, and treatment with
   insulin restores the lost expression of DSP. Insulin resistance is a
   common pathological feature of fatty liver and related ailments,
   whereas loss of DSP has been noted in liver carcinoma. Reduced DSP
   expression can lead to loss of cell-cell adhesion via desmosomes, and
   disrupt the keratin intermediate filament network. Our findings suggest
   that DSP expression may be perturbed by palmitate and, along with
   insulin resistance, may play a role in palmitate induced cytotoxicity,
   and serve as potential targets for further studies on non-alcoholic
   fatty liver disease (NAFLD).

Introduction

   Accumulation of lipids (primarily triglycerides) in hepatocytes is
   considered a hallmark and pre-requisite for development of
   non-alcoholic fatty liver disease (NAFLD) [33][1]. With increasing
   prevalence of obesity among adults and children, NAFLD has become the
   most common form of chronic liver disease in developed countries
   [34][2]. Studies have shown that diets rich in saturated fats
   contribute significantly towards increasing the risk of NAFLD [35][3],
   [36][4]. Consequently, a large number of studies have focused on the
   effects of free fatty acids (FFA) on hepatocytes to understand the
   pathogenesis of NAFLD and related liver diseases (reviewed in [37][5]).
   In recent years, systems biology approaches utilizing high-throughput
   (microarray) measurements have been applied to gain significant
   insights into the cytotoxic effects of FFAs on hepatocytes using either
   static [38][6], [39][7], [40][8], [41][9] or dynamic gene expression
   profiles [42][10]. While these efforts have focused predominantly on
   identifying individual target genes, some researchers have suggested a
   more complex scenario, whereby genes cooperate in regulating cellular
   events in response to FFA treatment [43][11].

   In this study, we analyzed the cytotoxic effects of palmitate
   treatment, the most common FFA in the diet, on HepG2 cells. First, we
   selected a subset of genes affected by FFA treatment by mapping their
   gene expression to metabolite profiles [44][8], [45][12]. This allowed
   the integration of multi-level data and further helped to alleviate the
   computational burden associated with the analysis of a large set of
   gene expression data. Next, we used an integrative methodology to
   reconstruct a gene cooperation network using the concept of
   “Information Synergy” [46][13]. The underlying principle of information
   synergy states that if two genes cooperate to affect a phenotype, then
   the joint expression of these two genes should provide more information
   on the phenotype than the sum of the information contributed
   independently by each of the genes. Thus, the gain in information or
   information synergy could be used to quantitatively asses the
   cooperative effect of any two genes on a phenotype. To help elucidate
   the processes that may be altered we analyzed the pathways that were
   enriched in the network.

   Our search for over-represented pathways in the synergy network
   recovered insulin signaling pathway as the most significantly enriched
   pathway. This is notable given the fact that almost all patients
   diagnosed with NAFLD have concomitant insulin resistance [47][14].
   Ruddock et al. confirmed that palmitate treatment induces resistance to
   insulin signaling in hepatocytes [48][15]. We further expanded our
   search to the neighbor genes of the insulin signaling pathway in the
   synergy network and recovered desmoplakin (DSP), a junction protein, as
   the top neighbor. The DSP protein is an obligate component of
   functional desmosomes at intercellular junctions. DSP together with
   plakoglobin and plakophilin forms the intracellular desmosomal plaque
   [49][16], which serves as an anchor for keratin intermediate filament
   attachment [50][17]. In general, junction proteins are known to be
   inhibited during chronic liver diseases, including cirrhosis and
   hepatitis [51][18]. It has also been shown that the loss of expression
   and abnormal localization of junction proteins, including DSP, is
   correlated with progression of HCC [52][19]. However, the effects of
   palmitate on DSP protein expression have not been previously reported.

   Here, we investigated the effects of palmitate treatment on the
   expression of DSP in HepG2 cells. Furthermore, since DSP was recovered
   as a neighbor of insulin signaling pathway, we analyzed the effect of
   insulin treatment on DSP expression. We found that palmitate treatment
   leads to the loss of DSP expression and treatment with insulin enhanced
   the recovery of DSP expression following palmitate treatment. Thus, our
   study indicates that the junction protein DSP, in synergy with insulin
   signaling, may be a novel target in the pathology of fatty liver
   disease.

Materials and Methods

Datasets

   HepG2 cells purchased from American Type Culture Collection (ATCC) were
   exposed to different FFA treatments for 48 hours. Gene expressions of
   these treated cells were then profiled with cDNA Microarrays at the Van
   Andel Institute (Grand Rapids, MI) and pre-processed by GenePix Pro 5.0
   (Microarray Acquisition and Analysis Software from Molecular Devices,
   LLC). The fluxes of 28 metabolites were measured using either HPLC or
   kits, and cytotoxicity was evaluated based upon LDH levels in the
   supernatants and in the cell lysates. The procedures of fatty acid
   treatments, microarray analysis, cytotoxicity and metabolic fluxes
   measurement were described in our previous works [53][8], [54][12]. The
   microarray dataset has been deposited to GEO website [55][20], with the
   query number of [56]GSE26885.

Gene selection based on the trend of the metabolites profiles

   FFAs modulate multiple intracellular metabolic pathways, many of which
   are involved in the pathophysiology of fatty liver disease. The fluxes
   of various metabolites were used to identify four representative trends
   across the treatment conditions: bovine serum albumin (BSA, control),
   Palmitate and Oleate, as discussed in [57][21] (see [58]Figure S1). The
   gene expressions were first logrithmized with base 2, and the
   expressions for the probes designed for same genes were averaged out.
   The genes annotated with unknown functions (i.e. “EST/hypothetical
   proteins” or “ORF of unknown functions”) were removed from the dataset.
   Then, the expression patterns of the remaining genes were matched to
   the four representative metabolite trends across BSA, Palmitate and
   Oleate. We further removed the genes not differentially expressed
   between any two of the three conditions based on two-sample t-test
   (p-value cutoff set at 0.05). Finally, 610 genes remained (see
   [59]Table S1) for further analysis.

Calculation of information synergy

   An information-theoretic measure was used to quantify the synergy
   between the genes. Given two genes, G1 and G2, and a phenotype P (toxic
   or non-toxic in our case), the information synergy between G1 and G2
   with respect to the phenotype P is defined as:
   graphic file with name pone.0028138.e001.jpg

   where I(G1;P) is the mutual information between G1 and P, I(G2;P) is
   the mutual information between G2 and P, and I(G1,G2;P) is the mutual
   information between (G1,G2) and P. This equation reflects the
   definition of synergy, the additional contribution provided by the
   “whole” as compared to the sum of the contributions of the individual
   “parts”. Mutual information (I) was calculated based on a
   clustering-based method for continuous expression data [60][22], on an
   information synergy scale of [−1 1]. A positive synergy score indicated
   that two genes jointly provided additional information on the
   phenotype, a negative synergy score indicated that the two genes
   provided redundant information about the phenotype, and a zero score
   indicated that the two genes provided no additional information on the
   phenotype.

   A permutation test was performed to assess the statistical significance
   of the synergy score of the gene pairs. The phenotypes were randomly
   permutated to un-correlate the phenotype with the gene expression
   profiles, and then the synergy scores were re-calculated based on the
   permutated phenotype for all gene pairs. This process was repeated 1000
   times to calculate the p-values of the synergy score for each gene
   pair. A Benjamin-Hochberg false discovery rate procedure [61][23] was
   applied to adjust the p-values for all the gene pairs and thereby
   control the expected false discoveries. Finally, 4376 out of the 185745
   pairs from the 610 genes were identified as significant based on a
   cutoff of 0.05 for adjusted p-values.

Statistical test for over-represented pathways and pathway-gene associations

   A web platform GENECODIS [62][24] was used to identify the KEGG
   pathways overrepresented in the synergy network. The hyper-geometric
   test was used to calculate the p-values of the KEGG pathways in the
   network, followed by Benjiamin-Hochberg FDR control procedure to obtain
   adjusted p-values, and finally the adjusted p-value cutoff was set at
   0.05 to determine the significantly over-represented pathways.

   Likewise, hyper-geometric test was also used to determine the
   significant neighboring genes for each over-represented pathway in the
   network. Given one pathway and one of its neighboring genes in synergy
   network, the p-value of this pathway-gene association was calculated as
   below:
   graphic file with name pone.0028138.e002.jpg

   Where x is the number of pathway genes connected to the given gene in
   the network; N is the number of genes in the network; n is the number
   of pathway genes in the network; k is the number of genes connected to
   the given genes in the network. For each pathway, the p-values of its
   neighbor were adjusted by Benjiamin-Hochberg procedure, and the
   adjusted p-value cutoff was set at 0.05 to determine the significantly
   associated neighbors.

Differential Correlation Network

   A differential correlation network was constructed by obtaining
   gene-pairs which show significant difference in the correlation
   coefficients between phenotypes [63][25]. Pathway enrichment analysis
   and pathway-gene association analysis were performed as described above
   for synergy network.

Cell culture and treatment

   HepG2 cells were cultured in Dulbecco's Modified Eagle's Medium (DMEM)
   (Invitrogen) containing 10% fetal bovine serum (Invitrogen) and 2%
   Penicillin-streptomycin (Invitrogen). Cells were incubated at 37°C and
   in 10% CO[2] atmosphere. Confluent cells were treated with 0.7 mM
   palmitate (Sigma) dissolved in DMEM culture medium for 48 hours. Cells
   were treated with 1 nM insulin in culture medium for 24, 48 and 72
   hours. DMEM culture medium was used as control for all treatments.

Immunofluorescence

   Cells were fixed in 3.7% formaldehyde for 20 min. Cells were then
   washed twice with PBS and treated with 5% Triton-X for 10 min. at room
   temperature to permeabilize cell membrane. This was followed by washing
   twice with PBS and incubation with 1% BSA to block non-specific protein
   interactions. The cells were then incubated with 5 µg/ml anti-rabbit
   DSP primary antibody (Abcam ab71690) overnight at 4°C. After washing
   twice with PBS, cells were incubated in AlexaFluor-488 (Molecular
   Probes A11001) conjugated goat anti-rabbit secondary antibody (green)
   for 1 hour at room temperature. Cells were washed with PBS and cell
   nuclei were stained (blue) using DAPI (Molecular Probes D1306) at a
   concentration of 300 nM for 5 min. Coverslips were mounted in ProLong
   Gold (Molecular Probes [64]P36934) and incubated in the dark for 24
   hours.

   Confocal imaging was performed on an Olympus FluoView 1000 Inverted
   IX81 microscope, using a 40X oil objective. Blue and green images were
   taken sequentially, using a Kalman average of 2X. The intensity graph
   was constructed by normalizing the total green fluorescence intensity
   to total blue fluorescence intensity (to account for the number of
   cells) in 3 separate fields of view for each sample. Multiple
   comparisons between fluorescence levels across different treatment
   conditions were performed using one-way ANOVA followed by Tukey's HSD
   post-hoc analysis with p-value cut-off set at 0.05.

Results

1 Information synergy network

   Based upon the concept of information synergy, we reconstructed a
   gene-cooperation network composed of 570 genes with 4376 connection
   edges. Topological analysis shows that this network follows a power-law
   distribution (see [65]Figure S2) that is different from the bell-like
   distribution of a random network. The reconstructed network suggested
   potential gene targets and pathways that may play important roles in
   the induction of the cytotoxic phenotype.

Positive information synergy indicates joint association with phenotype

   We interrogated gene pairs in terms of their ability to jointly
   distinguish phenotypes. As described above, positive, zero and negative
   information synergies represent gene pairs that provide synergistic, no
   and redundant information on the phenotype, respectively. Examples of
   gene pairs for each level of information synergy, demonstrating how
   they jointly provide information on the phenotypes, are illustrated in
   [66]Figure 1.

Figure 1. Joint expression of gene pairs with different information synergy.

   [67]Figure 1
   [68]Open in a new tab

   Blue and red points represent nontoxic and toxic samples, respectively.
   From top to bottom: Gene pairs with positive synergy jointly
   discriminate the phenotypes (indicated by the black dashed lines –
   joint information), whereas either of the genes alone provide little
   information in discriminating the phenotypes (indicated by the data
   points projected to the axes – marginal information). In contrast,
   genes with zero information synergy appear to be independent of the
   phenotypes, both jointly and marginally. Gene pairs with negative
   information synergy are marginally informative on the phenotype, such
   that either of the genes can provide information on the phenotype, but
   the additional information provided by the other gene does not enhance
   the discrimination of the phenotype, i.e. information provided by both
   genes about the phenotype is redundant.

   For the gene pairs with significant positive synergy in [69]Figure 1
   (top row), neither of the two individual genes was strongly correlated
   with the phenotype in the univariate marginal distribution. On the
   other hand, the gene pairs were significantly correlated to the
   phenotype, i.e. their expressions jointly help to distinguish the
   phenotypes. The two phenotypes are clearly discernible with the
   existence of a gap, previously denoted as a “gap” pattern [70][26].
   Additional correlation patterns, including “substitute” and “on/off”
   [71][26], were also observed in a small number of positive gene pairs
   (see [72]Figure S3 for explanations of these correlation patterns).
   Thus, both genes in the pair with positive synergy would have been
   ignored by a correlation based one-gene-at-a-time approach which
   considers only those genes that are highly correlated to the
   phenotypes.

2 Enriched pathways related to palmitate induced cytotoxicity

   The pathways enriched in the synergy network were explored through
   enrichment analysis. 42 pathways were significantly enriched in the
   synergy network (see [73]Table S2), and the top 10 pathways are listed
   in [74]Table 1.

Table 1. Top ten KEGG pathways ranked by their p-values enriched in the
synergy network.

   Pathway names #genes in network p-value Affected by Palmitate
   04910: Insulin signaling pathway 15 6.68E-07 Known [75][15]
   04520: Adherens junction 10 2.10E-05 Known [76][29], [77][30]
   00310: Lysine degradation 7 0.00014 Known [78][12]
   00330: Arginine and proline metabolism 6 0.00053 Known [79][12]
   04916: Melanogenesis 8 0.00556
   04120: Ubiquitin mediated proteolysis 9 0.00644 Known [80][28],
   [81][65]
   04010: MAPK signaling pathway 13 0.00658 Known [82][66]
   05210: Colorectal cancer 7 0.00716
   04310: Wnt signaling pathway 9 0.00942 Known [83][67]
   04810: Regulation of actin cytoskeleton 11 0.00981 Known [84][30],
   [85][68]
   [86]Open in a new tab

   As seen in [87]Table 1, cellular activities such as insulin signaling,
   adherence junction/cytoskeleton regulation, amino acid metabolism, and
   ubiquitin-mediated proteolysis are highly enriched in the synergy
   network. Indeed, it has been suggested in the literature that palmitate
   treatment affects several of these enriched pathways, including insulin
   signaling inhibited by palmitate in hepatomal cells [88][15], ER stress
   and the ubiquitin mediated protease pathways [89][27], [90][28], as
   well as adherence junction and cytoskeleton structure [91][29],
   [92][30]. Palmitate has been also shown to modulate the metabolism of
   various amino acids [93][12] which may be important players in
   palmitate-related cytotoxicity. For example, arginine metabolism
   provides substrate for nitric oxide synthetase (NOS) and palmitate
   treatment was found to enhance cell death by increasing NOS activity
   and NO production [94][31].

3 Pathway-gene association analysis

   In this section, we investigated the associations between the enriched
   pathways and their neighbors in the synergy network. Neighbors were
   defined as genes in the network connected to at least one gene of the
   enriched pathway. The connections in the network represent cooperative
   relationships between the genes. Therefore, we assumed that the
   neighbor genes, especially those connected to multiple members of an
   enriched pathway may function cooperatively with the enriched pathway
   in the associated phenotype.

Neighbor genes relevant to insulin signaling pathway

   We evaluated the biological relevance of insulin signaling, the most
   significantly enriched pathway, and its neighbor genes. In the synergy
   network, there are a total of 262 neighbor genes (see [95]Table S3)
   connected to the insulin signaling pathway ([96]Figure 2 and [97]Figure
   S4), and most of these neighbor genes (more than 90%) are connected to
   one or two insulin signaling genes. We focused on only those neighbor
   genes that are connected to 3 or more insulin signaling genes in the
   synergy network. For each of these neighbor genes, the statistical
   significance of its association with the insulin signaling pathway was
   evaluated by the hyper-geometric test. The neighbor genes significantly
   associated with insulin signaling are listed and ranked according to
   their p-values in [98]Table 2.

Figure 2. Sub-network for insulin signaling genes and their neighbor genes in
the synergy network.

   [99]Figure 2
   [100]Open in a new tab

   Yellow nodes represent insulin signaling genes, and pink, green, blue
   and black nodes represent the neighbor genes connecting to one, two,
   three and more insulin signaling genes, respectively. The fifteen
   insulin genes, the top three neighbor genes (in [101]Table 2) and the
   four genes connecting to more than three genes were labeled with gene
   names for reference.

Table 2. Neighbors significantly associated with insulin signaling pathway.

   Gene Connections to insulin signaling Degree (overall network) p-value
   Association with insulin signaling
   DSP 3 12 4.18E-03 Unknown
   ACADSB 3 12 4.18E-03 Unknown[102]^1
   PANK3 3 15 7.50E-03 Known, [103][69]
   SLC39A3 3 18 9.60E-03 Known, [104][70], [105][71], [106][72]
   ICA1 3 20 1.24E-02 Known, [107][73]
   TTYH1 3 30 2.94E-02 Known, [108][74]
   EP400NL 3 31 2.94E-02 Unknown
   EGFR 3 36 3.59E-02 Known, [109][75], [110][76]
   C6orf150 3 37 3.59E-02 Unknown
   FAM69B 3 37 3.59E-02 Unknown
   HOMER2 8 180 4.84E-02 Known, [111][77], [112][78]
   [113]Open in a new tab
   ^1

   ACADSB is involved in the metabolism of fatty acids and branch chained
   amino acids [114][79]. Prolonged treatment with long chain FFAs,
   including palmitate, increases FA oxidation [115][12], [116][80], which
   has been proposed to serve as a protective mechanism against the
   potential toxic effects of long chain fatty acids [117][80], [118][81].
   Insulin, on the other hand, down-regulates FA oxidation in various cell
   types [119][82], [120][83]. Since palmitate can impair the insulin
   signaling pathway [121][84], it is reasonable that ACADSB emerges as a
   top neighbor of the insulin signaling pathway.

4 DSP: the top neighbor

   The junction protein, DSP, was recovered as the top-neighbor of insulin
   signaling pathway from the synergy network. Loss of DSP expression has
   been reported in more severe forms of liver disease, i.e. cirrhosis and
   HCC, but this has not been implicated in the pathology of fatty liver
   disease. Hence, we investigated the effects of palmitate treatment on
   the expression of DSP in HepG2 cells. Since, DSP was recovered as a
   neighbor of the insulin signaling pathway, we also studied the effect
   of insulin treatment on DSP expression. HepG2 cells were grown to
   confluency and treated with palmitate containing media. We observed
   cellular expression of DSP by measuring the fluorescence levels of DSP
   with immunostaining and confocal microscopy ([122]Figure 3). Relative
   intensity of DSP against nuclear staining was used to quantify and
   normalize the amount of DSP under each treatment condition ([123]Figure
   4). Finally, we compared changes in relative DSP expression between
   treatment conditions and determined statistically significant pairs
   using Tukey's HSD post-hoc analysis ([124]Figure 4).

Figure 3. Immuno-fluorescence images of HepG2 cells stained for DSP (green)
and cell nuclei (blue) obtained using confocal microscopy (see [125]methods).

   [126]Figure 3
   [127]Open in a new tab

   Scale bars represent 50 µm. A) Untreated, control cells grown in
   regular growth media. B) Cells treated with palmitate for 48 hours show
   decrease in DSP levels C) Cells treated with palmitate for 48 hours and
   recovered in normal growth media for 72 hours show partial recovery of
   DSP expression. Cells treated with palmitate for 48 hours and recovered
   in growth media with insulin for D) 24 hours and E) 48 hours show
   partial recovery, whereas for F) 72 hours show complete recovery of DSP
   expression.

Figure 4. Quantitative effects of palmitate and insulin treatment on DSP
expression (see [128]methods).

   [129]Figure 4
   [130]Open in a new tab

   Bars indicate relative expression of DSP under various treatment
   conditions. Palmitate treatment significantly decreases DSP expression
   where as subsequent treatment with insulin restores DSP expression.
   Lines represent pairs of condition where changes in DSP level are
   significant. * indicates significant difference (p<0.05) from control
   cells. ** indicates significant difference (p<0.05) from palmitate (48
   hours) treated cells.

Palmitate reduces DSP expression

   First, we evaluated the effect of palmitate treatment on DSP
   expression. We treated HepG2 cells with 0.7 mM palmitate for 48 hours
   and observed the fluorescence levels of DSP ([131]Figure 3B). When
   compared to control ([132]Figure 3A), we found that palmitate treated
   cells showed a significant decrease in DSP expression levels (p<0.05).

Insulin enhances recovery of DSP expression

   Next, we examined the recovery of DSP expression in palmitate treated
   cells by removing palmitate after 48 hours of treatment and adding
   regular growth medium. The cells were allowed to grow for another 72
   hours ([133]Figure 3C) and showed some recovery of DSP expression with
   time, albeit statistically different from control (p<0.05). To assess
   the effect of insulin on the recovery of DSP expression, we treated
   HepG2 cells with physiological concentration of insulin (1 nM)
   [134][32] for up to 72 hours, following the removal of palmitate after
   48 hours of treatment. Fluorescence levels of DSP in the cells were
   analyzed every 24 hours. As seen in [135]Figure 3D–E, cells treated
   with insulin for 24 and 48 hours show partial recovery of DSP levels
   compared to palmitate treated cells (p<0.05), but cells treated with
   insulin for 72 hours show a prominent increase in DSP levels (p<0.05)
   ([136]Figure 3F). In fact, the levels of DSP after 72 hours of insulin
   treatment are statistically similar to the levels in the control cells
   (p>0.05). The results suggest the expression of DSP that was lost upon
   palmitate treatment recovered to basal levels by treating the cells
   with insulin. Although, cells did recover some expression of DSP after
   removal of palmitate from the medium, the recovery was not restored to
   basal levels.

   Thus, our experiment supports the association of the palmitate-induced
   cytotoxic phenotype with the DSP-insulin signaling pathway in our
   network. Namely, that DSP and insulin signaling act synergistically to
   ensure the proper function of adherence junctions and the lost or
   reduction in DSP expression and insulin signaling by palmitate
   treatment contributes to the cytotoxic effect of palmitate. Thus
   suggesting that in liver cells DSP may be a novel target of
   palmitate-induced cytotoxicity.

5 Differential correlation analysis fails to identify DSP

   Extant methods based on differential correlation have been used to
   uncover phenotype-specific gene interactions [137][33]. However, the
   quality of the data (i.e. sample size or noise level) would affect the
   results obtained, i.e. in the correlation coefficients calculated,
   making it difficult to determine whether the difference in the
   correlation across the phenotypes are real changes or simply an
   artifact due to the size or noise level. In contrast, the information
   synergy approach takes the phenotype data directly into account, thus
   the influence of the difference in the data quality across phenotypes
   is less of a concern than in the correlation-based methods. We compared
   the performance of information synergy approach to a standard
   differential correlation approach [138][25] for reconstructing
   phenotype-specific networks. The statistics for this comparison are
   shown in [139]Table 3. As seen in [140]Table 3, the insulin signaling
   pathway was recovered as the top pathway in the differential
   correlation network. But, DSP was not found to be significantly
   associated with the insulin signaling pathway.

Table 3. The statistics of the two networks and association b/w DSP and
insulin signaling.

   Synergy network Differential correlation network
   Statistics for whole network # genes in the network 570 610
   # connections in the network 4376 87054
   # insulin signaling genes in the network 15 15
   Rank of insulin signaling in term of statistical significance in the
   network 1st 1st
   Statistics for DSP Degree of DSP in the network 12 210
   # insulin signaling genes connecting to DSP 3 3
   FDR adjusted p-value of the association b/w DSP and insulin signaling
   4.18E-3 0.882
   Rank of DSP among the genes connected to at least three insulin
   signaling genes based on statistical significance 1st 549th
   [141]Open in a new tab

   We further investigated the ability of genes paired with DSP to
   distinguish the two phenotypes. The joint expressions of these gene
   pairs are shown in [142]Figure 5. One gene (ACACA) was identified by
   both synergy analysis and differential correlation. From [143]Figure 5,
   it is evident that the two phenotypes are distinctly separated in the
   synergy analysis plots whereas the phenotypes in the correlation
   analysis plots do not separate distinctly. Thus, the synergy analysis
   was more effective in identifying gene pairs that can distinguish
   between phenotypes.

Figure 5. The expressions of gene pairs, namely of insulin signaling genes
and DSP.

   [144]Figure 5
   [145]Open in a new tab

   Shown are three pairs of insulin signaling genes connected to DSP in
   the synergy analysis (top) and three pairs obtained from the
   differential correlation analysis (bottom). Red and blue points
   represent toxic (treated with palmitate) and nontoxic samples,
   respectively.

Discussion

   We presented an integrative methodology to reconstruct a gene network
   that demonstrates cooperativity plays an important role in complex
   diseases. There are three major contributions from this analysis.
   First, we presented the idea of information synergy and applied this
   method to microarray data from palmitate treated HepG2 cells: a
   condition relevant to fatty liver disease. Second, we showed that DSP
   is a novel target of palmitate treatment. Third, we showed that insulin
   impacts DSP expression, and palmitate, known to antagonize insulin
   signaling, also negatively impacts DSP, contributing to the cytotoxic
   effect of palmitate. Overall, we suggest that a synergy network-based
   gene and pathway association analysis has the potential to reveal novel
   targets or mechanisms underlying biological processes, exemplified by
   the uncovering of DSP as a novel player in palmitate-induced
   cytoxicity.

   The concept of synergy is intuitive and relevant to biological systems.
   For example, two transcription factors may play a cooperative role in
   determining the expression of a common gene. Similarly, interactions
   may occur between signaling proteins, i.e. cross-talk, and such
   associations may also be seen in receptor mediated transduction.
   Information synergy aims to capture cooperative effects, and in this
   study, on the impact of these effects on the induction a particular
   phenotype. There are several examples of synergistic interactions
   implicated in the pathology of fatty liver disease. For example, Smith
   J.J et al. showed that two groups of transcription factors worked in a
   combinatorial manner to control the transcriptional responses to fatty
   acids [146][34], thus supporting that cooperativity is important in
   regulating cellular processes. Similarly, co-expression of HNF4 and
   GATA transcription factor could synergistically activate the
   expressions of two genes involved in cholesterol homeostasis, and their
   binding sites are essential for maximal synergism [147][35]. Peroxisome
   proliferator activated receptors (PPARs), one of the best characterized
   nuclear receptors that mediate transcriptional activities of long-chain
   unsaturated fatty acids, was found to selectively cooperate with fatty
   acid binding proteins (FABPs) to regulate transcription [148][36].
   Sterol receptor element-binding protein (SREBP-1c), a gatekeeper of
   lipotoxicity [149][37], functions together with BETA2, LXRs and SP1, to
   regulate various cellular processes, such as insulin signaling
   [150][38], lipid synthesis [151][39], and cholesterol biosynthesis
   [152][40]. It is evident from these examples that identifying genes
   with cooperative effects [153][13], instead of limiting efforts to
   individual genes, has the potential to reveal novel genes or
   interactions, and ultimately help elucidate the mechanism of
   progression of complex diseases such as NAFLD. Along this line,
   differential correlation based approaches have been applied to identify
   linear relationships between genes and phenotype. However, information
   synergy has the distinct advantage in their ability to capture gene
   pairs with different types of dependency (not limited to linear
   correlation), as long as the gene pairs provide additional information
   on the phenotype. This makes it particularly appealing for complex
   scenarios such as uncovering gene cooperation in the induction of a
   phenotype or disease, where the interactions may not be linear. Thus,
   information synergy is a promising approach for constructing
   phenotype-specific networks, and provides a complementary approach to
   correlation-based methods.

   Although synergistic gene pairs have been shown valuable in
   discriminating phenotypes, interpreting the roles of the individual
   gene pairs remains a challenge due to the lack of sufficient functional
   or structural annotations for many of the genes, thereby formulating
   plausible hypotheses of the gene pairs more difficult. In contrast,
   organizing the individual gene pairs into a network (synergy network)
   and then performing gene module level analysis [154][41], i.e.
   identifying the pathways over-represented in network or pathway-gene
   associations, could help to hint at potential mechanisms as shown in
   our study. Nevertheless, investigating individual gene pairs would
   still be valuable in the future. Since the gene pairs identified from
   the microarray data do not necessarily interact directly with each
   other, future investigation of individual gene pairs coupled with
   physical interaction data (i.e. protein-protein and protein-DNA
   interaction) should further improve the analysis.

   The application of information synergy methodology to microarray data
   obtained from FFA-treated HepG2 cells revealed insulin signaling as the
   most significantly enriched biological pathway. Thus, our analysis
   successfully recovered insulin signaling as a hallmark characteristic
   pathway involved in fatty liver disease [155][42]. We further examined
   and confirmed the effects of palmitate treatment on DSP, the top
   synergistic neighbor of insulin signaling pathway.

   Through synergy analysis we found that expression of DSP is reduced by
   palmitate treatment in HepG2 cells. DSP is a desmosomal protein
   essential for intercellular attachment by desmosomes. In a study of
   adhesion molecules in HCC, Cao et al. found that moderately
   differentiated hepatocytes tend to show a reduction in DSP protein
   expression, whereas poorly differentiated hepatocytes show complete
   loss of DSP protein expression [156][43]. It has been suggested that
   DSP expression levels are inversely correlated with cell growth and
   division. As shown in a study with squamous cell carcinoma, inhibition
   of EGF (epidermal growth factor) using antagonists led to an increase
   in DSP protein levels [157][44]. Another study has suggested that DSP
   may play a role in tumor-matrix interaction as it was found to be a
   major component of exosomes secreted by mesothelioma cells [158][45].
   Other studies have also suggested that loss of desmosomes may be
   associated with epithelial to mesenchymal transition (EMT) by means
   which hepatocytes may acquire malignant characteristics [159][46].
   Although, such associations were reported in HCC, our results suggest
   that the impact on DSP may be an early event in the pathogenesis of
   fatty liver disease. In other words, reduced expression of DSP may
   contribute to the progression to more severe forms of liver diseases.

   From [160]Table 1 adherens junction was identified as a highly enriched
   category in our synergy network. Treatment with palmitate has been
   shown to affect cellular adherence and cytoskeletal structure in
   hepatocytes [161][29], [162][30], although the exact mechanism and
   whether DSP is involved in mediating these processes is unknown. Since,
   keratins require DSP at the desmosomal plaque for anchoring to the cell
   membrane [163][17], our results open up the possibility that loss of
   cytoskeletal structure may be associated with the loss of DSP.

   Previous studies have shown that inherited mutations in DSP are a
   characteristic feature of a number diseases of the skin and heart
   including, skin fragility or wooly hair syndrome [164][47],
   arrhythmogenic right ventricular dysplasia [165][17], [166][48], lethal
   acantholytic epidermolysis bullosa [167][49], and dilated
   cardiomyopathy with wooly hair and keratoderma [168][50]. It is notable
   that in recent years some research groups have suggested the risk of
   cardiovascular disease (CVD) may be elevated in NAFLD patients
   [169][51], [170][52], [171][53]. Given that alterations in DSP
   expression are associated with heart disease and since DSP levels are
   reduced by palmitate treatment, it raises the possibility that DSP may
   serve as a potential link between fatty liver disease and heart
   disease.

   We also explored the relationship between DSP and insulin signaling and
   found that insulin treatment enhances the recovery of DSP expression
   lost due to palmitate treatment. It is known that palmitate can induce
   resistance to insulin signaling in hepatocytes [172][15] and several
   mechanisms have been proposed to explain this observation. Studies have
   suggested that PKCA (Protein Kinase C, alpha isoform), known to be
   activated by lipids (diacylglycerols) [173][54], shares an antagonistic
   relationship with insulin signaling, and that knock-out of PKCA
   enhances insulin signaling [174][55]. Interestingly, the activation of
   PKCA also affects the dynamics of desmosomal complex at plasma membrane
   [175][56], [176][57]. Junction plakoglobin co-localizes with DSP at the
   desmosomal plaque. It has been shown in a study with mice
   cardiomyocytes that inhibition of DSP expression leads to change in
   localization of plakoglobins from cytoplasm to nucleus. This was shown
   to antagonize Wnt/β-catenin signaling pathway, and trigger the
   accumulation of fat droplets [177][58]. It is noteworthy that Wnt
   signaling is a significantly enriched pathway in the synergy network
   (See [178]Table 1). Wnt signaling is frequently attenuated in HCC
   patients and is associated with poor prognosis [179][59]. Furthermore,
   recent studies have confirmed that activation of Wnt signaling enhances
   insulin sensitivity [180][60], [181][61]. This suggests that palmitate
   induced loss of DSP protein can further enhance insulin resistance
   through the Wnt/β-catenin signaling pathway. Thus, it may not be a
   coincidence that DSP was recovered as a synergistic pair to the insulin
   signaling pathway in the network.

   According to one estimate, about one-third of the general population in
   the United States suffers from NAFLD [182][62]. Follow up studies of
   patients suffering from this burgeoning disease have shown that their
   survival is significantly lower than that of the general population
   [183][63]. Furthermore, NAFLD can progress to the more severe
   steatohepatitis (NASH) in a number of patients, which may further
   develop into cirrhosis and eventually into liver carcinoma (HCC)
   [184][64]. Unfortunately, the factors responsible for the pathology and
   progression of liver disease in these patients remain poorly
   understood. Our studies confirmed a potential link between palmitate,
   insulin signaling and DSP, and are currently further investigating the
   mechanisms contributing towards progression to more severe forms of
   liver disease.

Supporting Information

   Figure S1

   Four Representative Trends of the Metabolites. Eleven metabolites
   differed significantly across the three conditions (treated by BSA,
   Palmitate and Oleate), and four representative trends were extracted
   from these metabolites. Trend I: BSA < Palm and Palm > Ole; Trend II:
   BSA > Palm and Palm < Ole; Trend III: BSA < Palm < Ole; Trend IV: BSA >
   Palm > Ole.

   (TIFF)
   [185]Click here for additional data file.^ (694.1KB, tif)
   Figure S2

   Degree distribution of synergy network and Random network. The random
   network was generated based on Erdös-Rényi model, with same number of
   nodes and edges as synergy network. The degree distribution in synergy
   network is clearly different with that in random network.

   (TIF)
   [186]Click here for additional data file.^ (1.3MB, tif)
   Figure S3

   Different types of correlation patterns captured by synergy analysis.
   Blue and red points represent nontoxic and toxic samples, respectively.
   These patterns were specified in [187][30], and their definitions were
   given below: (1) Gap: gene positively correlated; phenotype associated
   with the difference of gene expression; Substitute: gene negatively
   correlated; phenotype associated with the sum of gene expression;
   (3)On/off: turning on or off both genes lead to same phenotype.

   (TIF)
   [188]Click here for additional data file.^ (534KB, tif)
   Figure S4

   Subnetwork for insulin signaling pathway (all nodes labeled with gene
   symbols).

   (TIF)
   [189]Click here for additional data file.^ (1.6MB, tif)
   Table S1

   The list of 610 genes selected for synergy analysis.

   (DOC)
   [190]Click here for additional data file.^ (77.7KB, doc)
   Table S2

   All KEGG pathways significantly enriched in the synergy network.

   (DOC)
   [191]Click here for additional data file.^ (77.6KB, doc)
   Table S3

   All neighbor genes significantly associated with insulin signaling.

   (DOC)
   [192]Click here for additional data file.^ (105.1KB, doc)

Footnotes

   Competing Interests: The authors have declared that no competing
   interests exist.

   Funding: This research was supported in part by the National Institutes
   of Health (NIH) (R01GM079688-01 and R21RR024439), the National Science
   Foundation (NSF) (CBET 0941055 and CBET 1049127), and the MSU
   Foundation. The funders had no role in study design, data collection
   and analysis, decision to publish, or preparation of the manuscript.

References