Retrodifferentiation of Human Tumor Hepatocytes to Stem Cells Leads to Metabolic Reprogramming and Chemoresistance

Hepatocellular carcinoma (HCC) is the third most common cause of cancer-related deaths worldwide. Despite recent progress in treatment, tumor recurrence remains high. HCCs show heterogeneous cell populations, suggesting that liver tumors may originate from different cell types (1). Although mature hepatocytes have long been considered as the target of oncogenic transformation, it is now proposed that HCCs may arise from different cells of the hepatic lineage, that is, hepatocytes, hepatoblasts, and hepatic progenitors (2, 3). Furthermore, the presence of cancer stem cells (CSC) in HCCs derived from these cell types suggest that any cell of the hepatic lineage could acquire stem cell properties. In vitro experiments have demonstrated that the inflammatory microenvironment promotes the retrodifferentiation of tumor-derived hepatocytes toward bipotent progenitor cells (4) and human and mouse hepatocytes can undergo reversible ductal metaplasia in response to injury in vivo (5). These new insights on the plasticity of differentiated tumor cells add to the complexity of the origin of CSCs and reinforce the hypothesis of their contribution to the cellular heterogeneity and chemoresistance in HCC (6). In addition, CSCs are spared by cancer chemotherapy that usually affects differentiated and dividing cells to reduce the tumor mass.

CSCs share properties with normal adult stem cells. They are able to self-renew by symmetric division but can also undergo asymmetric division, which preserves the stem cell pool, and give rise to a progeny of transit-amplifying cells that commit to a lineage differentiation program. CSCs are thought to give rise to all cell types within the tumor and can promote malignant progression. They exhibit drug resistance by having a low proliferation rate and high expression of xenobiotic transporters (7). Similar dysregulations of signaling pathways for stemness, proliferation, and differentiation in normal stem cells have been identified in CSCs and can contribute to tumor formation (8). Recent data unveiled unexpected metabolic heterogeneity within the tumor mass (9) and suggested that CSCs could harbor distinct metabolic phenotypes from differentiated tumor cells in various tissues. However, the metabolic profile of CSCs is controversial. Metabolic plasticity could allow CSCs to alternatively rely on either glycolysis or oxidative phosphorylation (OXPHOS) rather than acquiring a stereotypical energy metabolism profile (10). Therefore, a better understanding of the functional properties of retrodifferentiated cells is crucial for developing new therapeutic approaches that aim to eradicate both CSCs and differentiated cancer cells.

Among HCC cell lines, some retain a poorly differentiated phenotype such as HuH7 and HepG2 cells, whereas others display reversible differentiation program, for example, HBG-BC2 (BC2) and HepaRG cell lines (11, 12). HepaRG cells differ from BC2 cells in that they display a higher potential for plasticity, which ranges from liver progenitors to highly differentiated hepatocyte- and biliary-like cells. Furthermore, differentiated HepaRG cells are able to retrodifferentiate into progenitor cells when seeded at low density (12). Recently, we have demonstrated that inflammatory cytokines and TGFβ promote the retrodifferentiation of mature HepaRG hepatocyte-like cells and that retrodifferentiated progenitors express a genetic program typical of HCCs with aggressive behavior and unfavorable outcome (4). Here, we show that differentiated HepaRG hepatocyte-like cells not only retrodifferentiate to HCC progenitor cells, but also can acquire stem cell features under particular conditions. The HepaRG-derived stem cells expressed a transcriptomic program that was similar to proliferative, bad outcome HCCs in two separate cohorts totaling 457 patients. Emergence of the CSC phenotype is associated with a modulation of invasion, inflammation, redox homeostasis, and energy metabolism, initiated, in part, by angiopoietin like-4 (ANGPTL4) and mitochondrial pyruvate dehydrogenase kinase (PDK) 4. We demonstrate that the upregulation of PDK4 was associated to chemoresistance that could be successfully reversed by the PDK4 inhibitor dichloroacetate (DCA), in four HCC cell lines. Therefore, targeting stem cells by restoring mitochondrial pyruvate oxidation in combination with chemotherapy appears as a promising approach to treat HCC.

HepaRG and BC2 cells were originally isolated in the laboratory (11, 12) and grown in William E medium containing 0.025 g/L pyruvic acid and supplemented with 10% FCS, 5 μg/mL insulin, and hydrocortisone hemisuccinate (5 × 10⁻⁵ and 7 × 10⁻⁷ mol/L, respectively). HepG2 and HuH7 cells, obtained from ATCC and European Collection of Authenticated Cell Cultures, were grown in Eagle minimum essential medium (MEM) and DMEM, respectively, both supplemented with 10% FCS. For the sphere culture, cells were seeded in the same media with 2% FCS and 50-fold reduced concentration of insulin and hydrocortisone hemisuccinate, for 5 days. Then, cells were detached and seeded in ultra-low attachment plates in DMEM/F12 medium supplemented with 20% knockout serum replacement, 1 mmol/L l-glutamine, 1% nonessential amino acids, 0.1 mmol/L β-mercaptoethanol, and 4 ng/mL fibroblast growth factor 2 (FGF2, Miltenyi Biotec) for 5 more days. Medium was renewed at day 1 and 3. The use of human embryonic stem (hES) cells H9 (WA09 obtained from WiCell) was approved by the French Biomedicine Agency (no. RE10-013 I/C/R). Cell line authentication was performed with AuthentiFiler PCR Amplification Kit, Genotype Comparison Tool v1.0 (Life Technologies) and Mycoplasma testing was performed twice a year with MycoAlert Mycoplasma Detection Kit (Lonza).

At day 4, HepaRG, HuH7, and HepG2 cells were suspended at 1 × 10⁶ cells/mL in Hank balanced salt solution supplemented with 3% FCS and 10 mmol/L HEPES, and incubated for 90 minutes with 20 μg/mL Hoechst 33342 (Sigma-Aldrich), with or without the P-glycoprotein inhibitor verapamil (50 μmol/L, Sigma-Aldrich). Then, 1 μg/mL propidium iodide (Pharmingen) was added and cells were filtered through a 40-μm strainer to obtain a single-cell suspension. Cell analysis and sorting were performed using Becton Dickinson FACSAria III. Fluorescence emission of Hoechst 33342 was measured through 450DF20 BP (Hoechst blue) and 675DF20 BP (Hoechst red) optical filters. Dead cells were discriminated using propidium iodide labeling. Cells able to efflux Hoescht 33342 are referred to as side population (SP) cells.

mRNAs were obtained from biological replicates of H9 hES cells (n = 3), HepaRG-SP (n = 4), HepaRG-Spheres (n = 4), and HepaRG cells recovered 4 or 10 days after seeding in standard medium [HepaRG-D4 (n = 4) or HepaRG-D10 (n = 4), respectively]. RNA purity and integrity were evaluated with an Agilent Bioanalyser (Agilent Technologies). Genome-wide expression profiling was performed using the low-input QuickAmp Labeling Kit and human SurePrint G3 8 × 60K pangenomic microarrays (Agilent Technologies). Gene expression data were processed using Feature Extraction and GeneSpring software (Agilent Technologies).

The data discussed in this publication have been deposited in NCBI's Gene Expression Omnibus (GEO) and are accessible through Series accession number GSE75752 (http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE75752).

To determine genes significantly deregulated between all different conditions, a one-way ANOVA was performed with a P < 0.001 and a fold change (FC) > 4. To determine genes significantly deregulated between Spheres and HepaRG-D10 cells or SP and HepaRG-D10 cells, a t test (Bonferroni FWER correction) with a P < 0.03, FC > 2 and P < 0.05, FC > 3 was used, respectively. Different fold-changes and P values were used to obtain a comparable numbers of genes in both conditions. Gene Cluster 3.0 software (Pearson metrics with centroid distances) was used to cluster the selected genes and conditions and TreeView 1.6 to visualize the clustering.

Gene set enrichment analysis (GSEA) was used to check whether an a priori defined set of genes showed statistically significant, concordant differences between two biological states. GSEA was performed by using the Java-tool developed at the Broad Institute (Cambridge, MA). Unsupervised GSEA was done with the whole C2 collection of curated gene sets from the molecular signatures database (MSigDB). Enrichment score (ES) was determined after 1,000 permutations, as described previously (13).

The study protocol complied with French laws and regulations and was approved by INSERM's Institutional Review Board (number 01-036) in the context of the National Network of Liver Biological Resource Centers. Sample collection was reported to the Ministry of Education and Research (no. DC-2008-338). All the patients provided written informed consent before study entry. Tissue microarray (TMA) comprises 25 patients with HCC and 5 histologically normal livers, all in triplicate as described previously (14). Briefly, TMA construction was done with a MiniCore3 tissue arrayer (Alphelys). Five-micron microtome sections were processed for IHC with a Discovery XT from Ventana Medical Systems (Roche) slide staining system. Antibodies used were purchased from Santa Cruz Biotechnology: hepatocyte nuclear factor 1β (HNF1B, sc-8986), Prolab: CD44 (20282), Sigma Aldrich: leucine-rich repeat-containing G-protein coupled receptor 5 (LGR5, HPA012530), lunatic fringe homolog (LFNG, AV44923), hyaluronan and proteoglycan link protein 1 (HAPLN1, HPA019105), insulin-like growth factor binding protein 5 (IGFBP5, I7660), Indian hedgehog (IHH, AV45230), and ProteinTech: PDK4 (12949-1-AP) and ANGPTL4 (18374-1-AP). Stained slides were converted into high-resolution digital data with a NanoZoomer digital slide scanner (Hamamatsu). Digital slides were viewed using NDP.view software (Hamamatsu). The signal was independently read by four observers (K. Fekir, O. Musso, F. Cabillic, and A. Corlu) who were blinded to any information. Discrepancies were resolved by consensus reading. A 4-point scale (1-2-3-4) reflecting increasing signal intensity was used. Focal and general expressions of the proteins were considered.

Adherent cells or spheres were cultured for 3 days in medium deprived of serum and supplemented with cisplatin (Accord Healthcare France SAS) or sorafenib (Interchim), alone or in combination with DCA.

Two independent HCC cohorts referred to as “the Roessler's cohort” (n = 247 HCCs, GEO Series accession number GSE14520) and The Cancer Genome Atlas liver hepatocellular carcinoma (TCGA-LIHC, n = 210 HCCs) cohort were used in this study. For the Roessler cohort, raw feature data from GEO were normalized and log₂ intensity expression summary values for each probe set were calculated using Robust Multiarray Average (RMA, packages Affy). Each dataset was assessed for the detection of outliers and inter-array, experiment-related batch effects by principal component analysis and density plots. Because a clear-cut inter-array experiment-related batch effect was observed within GSE14520, this effect was corrected by the COMBAT method (package sva). Then, probes being detected over the background noise in at least one HCC were retained and quantile normalized (package preprocessCore). For the TCGA-LIHC cohort, normalized RNAseq data were downloaded from the Genomic Data Commons website and the patients were selected according to the method detailed in Desert and colleagues 2017 (15). Then, hierarchical cluster analysis were performed with the Ward's method using 1-Pearson correlation as a distance metric, based on the expression of the genes up- or downregulated in HepaRG-SP or HepaRG-Spheres.

Results are expressed as means ± SD. Data were analyzed using GraphPad Prism software (Version 5.0; GraphPad). The significance was evaluated by Student t test for comparison of two conditions or by one-way ANOVA for comparison between groups.

See Supplementary Materials and Methods for further information.

Cell phenotypic and functional plasticity is known to promote intratumor heterogeneity, which in turn contributes to chemotherapy resistances (16). To gain insight into the role of cell plasticity in chemoresistance, we first performed an unsupervised search for gene expression signatures matching the transcriptomic program of 48-hour–retrodifferentiated HepaRG hepatocyte-like cells to HCC progenitor cells (GSE52989; Supplementary Fig. S1A). GSEA revealed that among the top 50 signatures, 38% were related to solid and hematopoietic tumors, and 17% corresponded to stem cells, including the hES cell signature (Fig. 1A). Thus, CSCs can emerge from highly differentiated hepatocyte-like HCC cells.

To further characterize stemness features of HCC cells, we studied anchorage-independent growth and xenobiotic efflux of HepaRG, BC2, HepG2, and HuH7 HCC cell lines. Culture of the four cell lines under nonadherent serum- and EGF-free conditions resulted in the formation of 3D spheres (Fig. 1B). Spheres upregulated stem cell markers (OCT4, NANOG, CD90) when compared with cells cultured during the same period in adherent conditions (D10). Interestingly, HepaRG and BC2 cells, which are both able to reversibly differentiate, shared specific features: they formed smaller spheres, upregulated stem cell markers, and downregulated levels of hepatocyte makers (albumin, aldolase B) in comparison with HepG2 and HuH7 cells (Fig. 1B). Flow cytometry analysis demonstrated that the HepaRG-progenitor cells included 0.6% of SP cells (Fig. 1C), whereas SP cells accounted only for 0.04% in HuH7 and 0.01% in HepG2 cells. Interestingly, SP cells from the HepaRG cell line were able to differentiate to mature hepatocyte-like and biliary-like cells (Fig. 1D), as attested by cytochrome P450 3A4 and cytokeratin 19 staining, respectively (Supplementary Fig. S1B). Finally, BrdU pulse-chase experiments with two rounds of cell division without BrdU revealed late mitotic figures wherein only one nucleus was BrdU-positive, showing that asymmetric division occurred within the HepaRG-progenitor population (Fig. 1E).

Among the four HCC cell lines studied, HepaRG cells harbored the highest plasticity potential and stemness features. Thus, these cells were selected for transcriptomic analyses. Gene profiles of HepaRG-Spheres (Spheres) and HepaRG-SP cells (SP) were compared with those of hES cells (H9), HepaRG-progenitors (D4), and differentiating HepaRG cells (D10; Supplementary Fig. S1A). We identified 5,866 significantly deregulated genes by one-way ANOVA (P < 0.001, FC > 4). Hierarchical clustering of the differentially expressed genes showed two main branches classifying the samples according to the degree of stemness (Fig. 2A). One branch included D10, whereas the second one grouped hES cells, SP, Spheres, and D4. Interestingly, in this stem/progenitor cluster, the gene profiles of SP and Spheres were closer to the gene profiles of hES cells than that of D4. Accordingly, unsupervised GSEA showed significant enrichment of signatures related to stem cells, hepatoblastoma, cancer recurrence, and metastasis in the gene profiles obtained from Spheres and SP compared with D10 (Fig. 2B). Notably, both SP and Sphere signatures discriminated two subpopulations of patients with HCC within two public transcriptomic datasets, consisting of 210 and 247 patients with HCC, respectively (Fig. 2C). Patients harboring SP or Sphere signatures had a decreased overall survival (Fig. 2C). They also displayed high serum alpha-fetoprotein (AFP) level, tumor grade, Tumor Node Metastases (TNM), Cancer of the Liver Italian Program (CLIP), and Barcelona Clinic Liver Cancer (BCLC) stages, as well as metastasis signatures (Supplementary Table S1). Conversely, these patients showed low rates of periportal-type HCC signature, which highlights well-differentiated, favorable outcome tumors (15). Supervised GSEA confirmed that the SP and Sphere signatures were negatively and positively correlated with the gene expression signatures defining the periportal-type and STEM-type HCC subclasses, respectively (Fig. 2D; Supplementary Table S2) described by Desert and colleagues (15).

To gain more insight into the signaling pathways activated in these HCC stem cells, we compared the gene profiles of Spheres and SP versus D10 (Fig. 3A; Supplementary Table S3). Ingenuity Pathway Analysis (IPA) revealed that top upregulated molecules in Spheres were the proinflammatory cytokines IL11, IL6, IL1β, which trigger the retrodifferentiation process, and ANGPTL4 known to activate ERK and PI3K/AKT signaling pathways (17). Noteworthy, the main activated networks in both Spheres and SP were ERK and PI3K/AKT pathways while TGFβ and TNFα were the upstream regulators (Fig. 3A; Supplementary Fig. S1C and S1D). Consistent with IPA results, a connectivity map approach identified the inhibitor of PI3K LY-294002 among the top 7 ranked molecules that are likely to reverse the gene expression profile induced during spherogenesis (Supplementary Table S4).

Analysis of 788 common genes differentially expressed in Spheres and SP versus D10 (P < 0.05, FC > 2) performed by GSEA confirmed significant enrichment of signatures related to stem cells, inflammatory response, angiogenesis and tumorogenesis, in both Spheres and SP (Fig. 3B; Supplementary Tables S5 and S6). Besides activation of ERK and PI3K, IPA identified PPARG and PPARG coactivator 1α (PPARGC1A) associated to downregulation of numerous genes involved in the tricarboxylic acid (TCA) and OXPHOS. Decreased expression of PPARGC1A, known as a nodal regulator of mitochondria biogenesis and function, could be related to changes in mitochondrial pathways in both Spheres and SP cells (Fig. 3B). A downregulation of genes involved in OXPHOS [cytochrome c oxidase (COX), NADH:Ubiquinone Oxidoreductase (NDUF), succinate dehydrogenase (SDH), and ubiquinol cytochrome c oxidoreductase chain (UQRC) families] was also observed with concurrent upregulation of glucose transporters (GLUT3, 5, and 6) and glycolytic enzymes [Phosphofructokinase, Platelet (PFKP), pyruvate kinase M (PKM)] (Fig. 3C; Supplementary Fig. S1E). Moreover, poor labeling of mitochondria by Mitotracker red revealed a reduced mitochondrial membrane potential in SP compared with non-SP cells, and suggested a reduced mitochondrial activity although the mitochondrial mass, assessed with COX subunit II (COX2) staining, was higher (Fig. 3D). Lower mitochondrial membrane potential was associated with lipid accumulation in Spheres and SP (Fig. 3E). In parallel to PPARGC1A decrease, upregulation of PPARG was associated with increased transcription of its target gene PDK4, an inhibitor the pyruvate dehydrogenase (PDH) complex (Fig. 3C). In keeping with the role of PDK4, increased phosphorylation of the E1-α subunit of PDH complex, high lactate production, low pyruvate consumption, and low ATP content were observed in HepaRG-Spheres (Fig. 3F). Of note, due to the limited number of available SP cells, the later experiments were only performed on Spheres.

Unlike most other genes related to cell metabolism that were repressed, ANGPTL4 and PDK4 were found highly upregulated in HepaRG-Spheres and -SP (Fig. 3C). Interestingly, both were also significantly upregulated in BC2-Spheres along with stem cell markers (Fig. 1C; Supplementary Fig. S1F). This prompted us to evaluate the expression of ANGPTL4 and PDK4 by TMA-based immunochemistry in human HCCs (n = 25) and normal livers (n = 5). In normal liver, ANGPTL4 and PDK4 were predominantly detected in the cytoplasm of hepatocytes (Supplementary Fig. S1G). In HCCs, ANGPTL4 and PDK4 showed similar staining patterns: both proteins were diffusely detected in the tumor (Fig. 4A; Supplementary Fig. S1H). However, scattered cells showing upregulated expression of ANGPTL4 and PDK4 could be observed within the tumor parenchyma and intratumor fibrous tracts. Immunostaining of serial sections showed a similar staining pattern with respect to the progenitor/stem cell markers CD44, LGR5, HNF1B, IHH, LFNG, IGFBP5, and HAPLN1 (Fig. 4B; Supplementary Fig. S1H). Of note, ANGPTL4/PDK4-positive cells located within intratumor fibrous tracts were positive for anti-hepatocyte antibody (Supplementary Fig. S1H). In situ protein expression was semiquantitatively rated on a 4-point scale (1–4) by 4 observers in 25 HCCs. ANGPTL4 and PDK4 scores were highly correlated (Fig. 4C; Supplementary Fig. S1I) and associated with CSC markers, as shown by network correlation analysis (Fig. 4D). In addition, HCCs without tumor necrosis showed high protein expression scores of ANGPTL4 and PDK4 (Fig. 4E).

ANGPTL4 is a multifaceted secreted protein that could play a role in anoikis resistance by stimulating NADPH oxidase, therefore leading to activation of the ERK and PI3K prosurvival pathways. In agreement, we found in Spheres and SP cells increased expression of the integrin β1 subunit ITGB1, which is known to bind ANGPTL4 and transduce its signal (Supplementary Fig. S1J; ref. 17). Hence, to evaluate the role of ANGPTL4 on stemness acquisition and metabolism adaptation, we performed loss-of-function experiments in HepaRG cells only using Spheres, because the limited number of available SP cells did not allow us to carry out concomitant analyses. Transduction with lentivirus encoding shRNAs against ANGPTL4 (ShANGPTL4) greatly diminished synthesis and secretion of ANGPTL4, whereas it neither affected spherogenesis nor viability assessed by lactate dehydrogenase (LDH) release (Fig. 5A and B). ANGPTL4 knockdown reduced ERK phosphorylation, slightly decreased PI3K expression, whereas phosphorylated AKT was increased (Fig. 5A). Concomitantly, we detected a downregulated expression of metalloproteinases (MMP) MMP2, MMP13, anti-inflammatory cytokine IL11 and proangiogenic factor VEGFA (Fig. 5C). These observations were consistent with reduced invasion capacity of the ShANGPTL4-Spheres (Fig. 5D). In metabolic terms, pyruvate and lactate extracellular concentrations as well as ATP and lipids contents were unchanged (Fig. 5E; Supplementary Fig. S1K). Interestingly, downregulation of PDK4 by ShANGPTL4 (Fig. 5C) may favor mitochondrial activity as suggested by higher the MitoTracker staining (Fig. 5F). In contrast, COX2 staining was decreased by ShANGPTL4, thus suggesting that the compensatory increase in mitochondrial mass was no longer required for the restoration of mitochondrial activity.

To specify the role of ERK and PI3K, downstream effectors of ANGPTL4/integrin signaling, we sought to target both pathways, which were found strongly activated in Spheres, using specific inhibitors. Although the sphere number was reduced and mild cell toxicity was detected by higher LDH release in the culture medium, HepaRG-progenitors treated with the MEK1/2 inhibitor U0126 (10 μmol/L) preserved their spherogenesis ability (Fig. 6A). After treatment, ERK phosphorylation was markedly reduced, whereas the total amount of ERK was unchanged (Fig. 6A). Unlike OCT4 and NANOG, mRNA levels of ANGPTL4, PDK4, VEGF, IL6, IL11, MMP1, 2, 9, and 13 were decreased by U0126 (Fig. 6B). Consistent with the sharp downregulation of MMPs, a strong inhibition of the invasive potential of the spheres was observed (Fig. 6C). The ATP content was decreased, concomitantly to cell viability, while the lactate and pyruvate extracellular concentrations were not modified (Fig. 6D).

HepaRG-progenitors treated with the PI3K inhibitor LY294002 (10 μmol/L) showed an inhibition of AKT phosphorylation that resulted in alteration of sphere formation. Moreover, the release of LDH during the culture revealed reduced cell viability (Fig. 6E). In remaining cells, RNA levels of ANGPTL4, PDK4, VEGF, IL6, IL11, PDK4, MMP1, 2, 9, and 13 were downregulated compared with those of untreated spheres, whereas OCT4 and NANOG expression was unchanged (Fig. 6F). Of note, 10 μmol/L LY294002 did not affect HepaRG-D10 cell viability. The low number of viable cells after LY294002 treatment did not allow us to perform metabolism and invasion assays.

Because ANGPTL4 and PDK4 were found upregulated in stem/progenitor cells, we investigated their possible involvement in the resistance to standard chemotherapies. When exposed to cisplatin or sorafenib for 3 days, the viability of HepaRG, BC2, HuH7, and HepG2 cells was reduced by 50% to 80% compared with untreated cells. In contrast, HepaRG- and BC2-Spheres were totally chemoresistant. However, a moderate efficacy of cisplatin could be noted against HepG2- and HuH7-Spheres (Fig. 7A). Interestingly, in HepaRG-D4 cells that were resistant to cisplatin or sorafenib, we observed an increased expression of OCT4, NANOG, ANGPTL4, and PDK4 (Fig. 7B). Moreover, among these markers, PDK4 expression was found significantly increased in all HCC Spheres (Supplementary Fig. S1F). Hence, we hypothesized that inhibition of PDK4 could reduce the chemoresistance by reactivating the mitochondrial PDH complex. Accordingly, in ShANGPTL4-HepaRG-Spheres, reduced expression of PDK4 sensitized cells to sorafenib (Fig. 7C). Therefore, we evaluated the effects of the chemical PDK inhibitor DCA against HCC cell lines. DCA (50 mmol/L) induced drastic decrease in cell viability in HepaRG-Spheres, which expressed the highest levels of PDK4 (Fig. 7B and C; Supplementary Fig. S1F) and the co-treatment DCA/chemotherapy did not provide significant additional effect. Low expression levels of antioxidant enzymes such as hemoxygenase2 (HMOX2), superoxide dismutase 1 and 3 (SOD1 and 3), and glutathione peroxidase 1 and 4 (GPX1 and 4) in HepaRG-Spheres (Supplementary Fig. S1L) may enhance reactive oxygen species (ROS) toxicity and contribute to efficacy of DCA alone (Fig. 7C). On the contrary, DCA had no effect on cell viability of HepaRG-D10 and showed mild efficiency on HepaRG-D4, which expressed low and intermediate PDK4 levels, respectively. Knocking-down PDK4 expression with siRNA in HepaRG-D4 also reduced cell viability (Fig. 7C; Supplementary Fig. S1M). Cotreatment DCA/chemotherapy was more effective than chemotherapy alone on both HepaRG-D4 and HepaRG-D10 (Fig. 7C). In addition, efficacy of DCA and cotreatment DCA/chemotherapy was confirmed in BC2, HepG2, and HuH7 HCC cell lines (Fig. 7C) that all expressed PDK4 at various levels (Supplementary Fig. S1F).

Cell heterogeneity is one of the main features of HCC and likely contributes to drug resistance. The tumor microenvironment can significantly influence cell fate. Therefore, by understanding the mechanisms underlying how tumor cells adapt to their environment, the molecular basis of chemoresistance may be understood. We previously demonstrated that the crosstalk between the proinflammatory cytokines TNFα, IL6, and TGFβ triggers retrodifferentiation of hepatocyte-like HCC cells into proliferating liver progenitors whose gene signatures match proliferative HCC subclasses that have a poor prognosis (4). In this work, we show that liver progenitors retrodifferentiate toward even more immature cells endowed with stem properties and increased chemoresistance. Both Spheres and SP expressed gene signatures related to stemness, inflammation, HCC metastasis, and recurrence. Notably, HCCs expressing the Sphere or SP signatures were associated with unfavorable outcome.

Gene expression profiling revealed that cytokines (TNFα, IL6, TGFβ, IL11), whose expression in HCCs are associated with poor prognosis (18–21) and involved in retrodifferentiation, were expressed in Spheres and SP at higher levels than in progenitor cells. This emphasizes the role of an autocrine regulation by cytokines in cell fate conversion within the tumor and the maintenance of stem cell properties. It also highlights that the acquisition of stem properties was associated with an important modulation in cell metabolism. To support these findings, among the upregulated genes in both Spheres and SP, we identified ANGPTL4, a molecule described as a potential modulator of the cross-talk between metabolism and cancer (22), and PDK4, an enzyme responsible of cell metabolic plasticity (23).

ANGPTL4 is a circulating protein predominantly produced in the liver and adipose tissue, which regulates lipid and carbohydrate metabolism, redox homeostasis, and attenuates the inflammatory response (24). The biological significance of ANGPTL4 in cancer remains controversial. In the liver, ANGPTL4 expression has been found to be associated with cancer cell extravasation, enhanced tumorigenesis, and poor prognosis (25). In patients with HCC, serum ANGPTL4 levels correlate with tumor progression and metastasis (25, 26). In contrast, ANGPTL4 expression has been shown to be significantly lower in HCC tissue than in nontumor tissue. The downregulation of ANGPTL4 RNA levels has been associated with advanced tumor stage, high alpha-fetoprotein levels, poor differentiation, and tumor recurrence (27, 28). Furthermore, ANGPTL4 may suppress HCC progression and metastasis in mice through disruption of a tumor-favorable microenvironment (27). This discrepancy may result from the dual expression pattern of ANGPTL4 in human HCCs. Using IHC, we observed a diffuse staining pattern of variable intensity in the tumor hepatocyte cytoplasm but also a very strong signal in scattered cells expressing CSC markers within the tumor parenchyma and intratumor fibrous tracts. The hypothesis that ANGPTL4 is involved in malignant progression was supported by in vitro knockdown of ANGPTL4 in HCC cells expressing CSC features.

ANGPTL4 expression was able to regulate IL11 but not IL6 production. Initially considered as an anti-inflammatory cytokine promoting cell adaptation to inflammatory environment, IL11 has been recently reported to present protumorigenic properties by protecting epithelial cells from death (29) and activating oncogenic signaling pathways (30). By interacting with integrins (predominantly ITGB1 and ITGB5), tumor-derived ANGPTL4 also sustains anchorage-related signals in the absence of extracellular matrix and cell–cell contacts. It stimulates NADPH oxidase–dependent production of ROS to maintain a high anion superoxide (O₂.⁻)/hydrogen peroxide (H₂O₂) ratio that activates the ERK and PI3K prosurvival pathways (17). We demonstrated that ANGPTL4 contributes to ERK activation and PI3K signaling, leading to upregulation of MMP-2 and 13 allowing tumor invasion and VEGF-A allowing angiogenesis. We also confirmed that PI3K/AKT and ERK signaling pathways are crucial for cell survival, inflammatory response, while stem cell marker expression was unchanged.

The PPARγ, TGFβ, and HIF1α pathways are able to induce ANGPTL4 expression (22, 31). In the small HepaRG-Spheres (less than 200 μm diameter), no significant accumulation of HIF1α was detected (Supplementary Fig. S1N). In addition, high protein expression scores of ANGPTL4 were found in human HCCs without tumor necrosis, suggesting that tumors enriched in stem/progenitor cells are metabolically fit to withstand the cancer microenvironment. In terms of PPARγ and TGFβ signaling, they are both activated in Spheres and SP, and also described to regulate PDK4 expression (32, 33). In addition, we observed that inhibition of ERK or PI3K, known to affect PPARγ transcriptional activity (34), decreased both ANGPTL4 and PDK4 expression.

Silencing ANGPTL4 in Spheres also decreased PDK4 expression and increased mitochondrial membrane potential, emphasizing its potential role in mediating an interplay between inflammation, metastasis, and metabolism (22). The PDH complex manages the competition between glucose and fatty acid oxidation by controlling the conversion of pyruvate into acetyl-coenzyme A. Therefore, by inactivating the PDH complex, PDK4 plays a pivotal role in metabolic plasticity because it restricts the supply of acetyl-coenzyme A from glucose oxidation to feed the TCA cycle. We noted low pyruvate consumption and high lactate production in the HepaRG-Spheres. However, reduced OXPHOS gene expression, mitochondrial ATP synthesis, and membrane potential, associated with accumulation of lipids suggested a lack of metabolic switch toward long-chain fatty acid oxidation to feed the TCA cycle. ANGPTL4 has been shown to inhibit lipoprotein lipase activity and contributes to intracellular lipase degradation (35, 36). The increased expression of phosphofructokinase and glucose transporters GLUT3, 5, and 6 combined with the decreased expression of PPARGC1A, a nodal regulator of mitochondrial biogenesis and function, suggest that both SP and Spheres use preferably the glycolytic pathway.

Mitochondria are the primary cellular source of ROS, in particular O₂.⁻, which is readily dismutated to the freely diffusible and cytotoxic H₂O₂ (37). Therefore, a reduction in OXPHOS could favor the survival of stem cells as a result of decreased mitochondrial H₂O₂ release (38). In addition, decreased mitochondrial activity is expected to impair the intrinsic apoptotic pathway (39). Consistent with these observations, metabolism reprogramming associated with a glycolytic profile is observed in hES and induced pluripotent stem cells, which leads to lower levels of ROS than in their progeny (40). During somatic reprogramming to pluripotency, it was demonstrated that glycolytic shift and inhibition of pyruvate transport into mitochondria occurs as a consequence of the induction of the expression of mitochondrial enzyme PDK1, which inhibits PDH (41). In the same way, recent reports suggest that a decrease in gene expression and protein amounts of the PDH complex in breast tumor–initiating cells is linked to reduced mitochondrial activity and higher glycolysis (42). A more active glycolysis over OXPHOS was detected in theCD133⁺ PLC/PRF/5 human HCC cell line compared with CD133⁻ cells (43) and in liver tumor-initiating stem–like cells in mice (44).

Recently, reports have suggested a causal link between metabolic change or mitochondrial dysfunction and chemoresistance. Consistent with this hypothesis, HCC-Spheres were resistant to cisplatin and sorafenib treatments compared with adherent cultures. The greatest cisplatin resistance of HepaRG and BC2 compared with HuH7 and HepG2 Spheres may result from the presence of proliferative cells in high-growing HuH7 and HepG2 Spheres. Moreover, HepaRG-D4–resistant cells strongly expressed OCT4, NANOG, ANGPTL4, and PDK4. These data suggest that these drugs spare cells with stem cell characteristics. Importantly, DCA, which is known to redirect glucose metabolism from glycolysis to TCA and OXPHOS (38), succeeded in increasing the effect of chemotherapies against all HCC cell lines assessed. The chemotherapeutic effect of DCA may originate from the inability of cells to counteract the increased ROS production following mitochondria reactivation. This effect was especially notable in Spheres, probably because of their low levels of antioxidant enzymes.

Despite initial favorable responses with sorafenib and improved survival with regorafenib (45), tyrosine kinase inhibitors do not completely block HCC progression (46–48). Therefore, identifying new therapeutic targets remains a challenge. We have demonstrated that cisplatin- and sorafenib-resistant HCC cells exhibit stem cells properties and have revealed that ANGPTL4/PDK4 contribute to the maintenance of CSC features. Importantly, we provide the proof of concept that targeting the enzymes involved in metabolic reprogramming could be useful in association with chemotherapy to better control HCC recurrence.

No potential conflicts of interest were disclosed.

Conception and design: K. Fekir, F. Morel, F. Cabillic, A. Corlu

Development of methodology: K. Fekir, R. Désert, D. Glaise, F. Morel, O. Musso, F. Cabillic, A. Corlu

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): K. Fekir, R. Désert, Y. Daniel, D. Glaise, C. Rauch, F. Morel, O. Musso, F. Cabillic, A. Corlu

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): K. Fekir, H. Dubois-Pot-Schneider, R. Désert, Y. Daniel, D. Glaise, C. Rauch, F. Morel, O. Musso, F. Cabillic, A. Corlu

Writing, review, and/or revision of the manuscript: K. Fekir, H. Dubois-Pot-Schneider, R. Désert, B. Fromenty, O. Musso, F. Cabillic, A. Corlu

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): K. Fekir, Y. Daniel, D. Glaise, C. Rauch, F. Morel, O. Musso, A. Corlu

Study supervision: F. Cabillic, A. Corlu

The authors thank Drs. P. Loyer and C. Aninat for thoughtful discussions and critical reading of the manuscript; Drs. L. Sulpice and D. Bergeat, Rennes University Hospital, for patient data; and Drs. Mireille Desille and Bruno Turlin, CRB-Santé, Biosit, Biogenouest, Rennes, France. The authors also thank the “Genomique Santé,” “ImPACcell,” “flow cytometry,” “H2P2” Core facilities from Biogenouest, SFR Biosit, University of Rennes1 and the “flow cytometry” Core facility from Paul Brousse Hospital, Villejuif, France. This work was supported by the Institut National de la Santé et de la Recherche Médicale (to B. Fromenty, O. Musso, D. Glaise, C. Rauch), the Centre National de la Recherche Scientifique (to A. Corlu), the University of Rennes 1 (to F. Cabillic, Y. Daniel), the « Association pour la Recherche sur le Cancer », the «Ligue contre le cancer - Comités d'Ille-et-Vilaine, des Côtes d’Armor, de Loire Atlantique et de Vendée», Cancéropôle Grand Ouest (CONCERTO), and Institut National du Cancer (grant no. 12688). K. Fekir was funded by the «Conseil Régional de Bretagne», INSERM (CESOCA) and the Ligue contre le cancer-Comité d’Ille et Vilaine. H. Dubois-Pot-Schneider was funded by the Seventh Framework Programme (FP7) LIV-ES (grant no. 223317) and the «Contrat Plan Etat-Région». R. Désert was funded by the «Conseil Régional de Bretagne» and INSERM (VALCIPRE).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.