Dysregulated paired related homeobox 1 impacts on hepatocellular carcinoma phenotypes

Hepatocellular carcinoma (HCC) is a major cause of cancer-related death. Paired related homeobox 1 (PRRX1) is a transcription factor that regulates cell growth and differentiation, but its importance in HCC is unclear. We examined the expression pattern of PRRX1 in nine microarray datasets of human HCC tumour samples (n > 1100) and analyzed its function in HCC cell lines. In addition, we performed gene set enrichment, Kaplan-Meier overall survival analysis, metabolomics and functional assays. PRRX1 is frequently upregulated in human HCC. Pathway enrichment analysis predicted a direct correlation between PRRX1 and focal adhesion and epithelial-mesenchymal transition. High expression of PRRX1 and low ZEB1 or high ZEB2 significantly predicted better overall survival in HCC patients. In contrast, metabolic processes correlated inversely and transcriptional analyses revealed that glycolysis, TCA cycle and amino acid metabolism were affected. These findings were confirmed by metabolomics analysis. At the phenotypic level, PRRX1 knockdown accelerated proliferation and clonogenicity in HCC cell lines. Our results suggest that PRRX1 controls metabolism, has a tumour suppressive role, and may function in cooperation with ZEB1/2. These findings have functional relevance in HCC, including in understanding transcriptional control of distinct cancer hallmarks.

overexpressed in breast, pancreas, head and neck squamous cell carcinoma, liver, and colon cancer [8][9][10][11][12][13][14]. In murine pancreatic cancer and human breast cancer cell lines, the knockdown of the isoforms PRRX1a or PRRX1b reduced migration and invasion, indicating functional similarities [9,14,15]. However, overexpression of PRRX1a led to the induction of genes involved in cell migration whereas PRRX1b was more involved in cell cycle regulation [9], supporting that these isoforms could have distinct molecular functions. The knockdown of PRRX1b inhibited proliferation as well as migratory and invasive capabilities of triple negative breast cancer cell lines [14]. In other studies, knockdown of PRRX1 reduced tumour volume of MDA-MB-231 mice xenografts [14], and its overexpression in colon cancer cell lines increased colony formation and anchorage independent growth [10]. In the HCC cell line HUH7, ectopic expression of PRRX1 induced resistance to 5-fluoruracil [12]. These data indicate a tumour-promoting role of PRRX1 in some cancer settings. On the contrary, PRRX1 has also been shown to exert tumour suppressor functions. For example, low expression of PRRX1 enabled metastatic colonization of lung by breast cancer cells [8]. HCC cells were also shown to migrate more upon PRRX1 knockdown [13]. In clinical contexts, PRRX1 has been associated to contradictory prognostic outcomes showing that high PRRX1 expression predicted improved overall survival in colorectal cancer and HCC [10,12,13], but was associated with metastasis and poor survival outcome in breast cancer [8].
Although the precise roles of PRRX1 and its isoforms, PRRX1a and PRRX1b, are unclear in human cancer, PRRX1 was often associated with cancer stemness and epithelial-mesenchymal transition (EMT) [8,9,15]. For example, the overexpression of PRRX1 in human colon cancer cell lines induced EPH receptor B2an intestinal stem cell marker [10], whereas the knockdown of PRRX1 in breast cancer increased stemness features [8]. Regarding EMT, overexpression of PRRX1 caused the upregulation of the EMT gene TWIST1 [8]. In addition, its knockdown reversed invasiveness [15]. Few studies have investigated PRRX1 expression in HCC [12,13,16], and its functions are largely unknown. Based on the contradictory reports on PRRX1, including on expression level in HCC, we set out to comprehensively analyze its expression and to predict its functions in human HCC. Making use of several HCC tissue gene expression datasets and applying functional assays in vitro, we find that PRRX1 is frequently upregulated in human HCC. We identified two EMT transcriptional factors, zinc finger Ebox-binding homeobox (ZEB1 and ZEB2) as novel PRRX1-related genes, and report that PRRX1 knockdown affects the phenotype of HCC cell lines, including metabolism.

Collection of liver cancer microarray datasets and analyses
Eight liver cancer microarray datasets were obtained from the National Center for Biotechnology Information Gene Expression Omnibus (NCBI GEO), whereas the TCGA liver cancer data was accessed via cBioPortal platform (http://www.cbioportal.org/) [17,18] and http://cancergenome.nih.gov. The expression values for PRRX1 probes were compiled for each dataset, and its differential expression in HCC compared to normal or adjacent non-tumour tissues was determined using Student T-test in GraphPad Prism version 6. Genes positively or negatively co-expressed with PRRX1 in liver cancer data were downloaded from cBioPortal platform and used for correlation analyses in TCGA and GSE14520 datasets. For PRRX1-high versus low expressing tumour comparison, TCGA and GSE14520 samples were ranked based on PRRX1 expression level and divided into two groups (based on median). The high (n = 186) vs low (n = 185) samples from TCGA were subsequently compared in R, using limma package. For Kaplan-Meier overall survival analysis of PRRX1 in combination of other genes, we first considered whether the genes have significant direct or inverse correlation with PRRX1 prior to subsetting samples for comparison. For example, for genes that met the criteria of direct correlation with PRRX1, the tumours with high expression of the gene and PRRX1 were compared with tumours with low expression of both. In contrast, for genes inversely correlated with PRRX1, tumours with high PRRX1 (n = 111) and low expression of the genes were selected and compared with tumours with low PRRX1 (n = 110) and high level of the gene. Analysis of cancer hallmark pathways was done with the 'GSEAPreranked' option in GSEA (http://software.broadinstitute.org/gsea/index.jsp).

Cell culture
The Hep3B, HLF and HUH7 cell lines were provided by Prof Kern (Pathology, Heidelberg). The SNU398 cells were obtained from Dr. Francois Helle (University of Picardie Jules Verne, France). The cell lines HLF, HUH7, Hep3B and SNU398 were cultured in Dulbecco Modified Eagle's medium (DMEM, high glucose, Lonza, BE12-709) supplemented with 2 mM glutamine, 10% fetal bovine serum (FBS), penicillin (100 U/ml), and streptomycin (100 μg/ml). The cells were cultured at 37°C in a humidified atmosphere containing 5% CO 2 . Hank's Balanced Salt Solution (HBSS) (Sigma-Aldrich, H6648) was used for cell washing steps. The cells were used for experiments between passage 2-10. All used cell lines were authenticated by short tandem repeat profiling (STR). PCR Mycoplasma Test (PromoCell, Huissen, Netherlands) was performed to confirm that cells are mycoplasma-free. siRNA transfection and RNA isolation were performed as described in the supplement.

MTT proliferation assay
Proliferation assays were performed with 3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide (MTT) assay. Briefly, 7.5-10 × 10 3 cells per well were seeded in quadruplicates in 48-well plates. The cells were incubated o/n and thereafter transfected with the respective siRNA oligos. After the indicated time, 25 μl of 5 mg/ml MTT reagent (Sigma Aldrich, USA) was added to the wells and incubated for 3-4 h at 37°C. The media was subsequently aspirated and 250 μl of MTT solubilizing reagent was added to dissolve the formed formazan crystals. The plate was then incubated o/n at 37°C. Absorbance was read at 560 nm with background correction at 670 nm using Infinite 200 Spectrophotometer (Tecan, Austria) and the data were normalized to the control wells.

Clonogenic assay
The HCC cells were initially seeded into 12 well plates and allowed to attach overnight. Culture medium was then replaced with medium containing siRNA/transfection reagent. After 24 h, the cells were trypsinized and seeded in triplicates in 6-well plates (2500 per well).
Medium was changed at day 4 and the experiment terminated at day 8. Cells were fixed for 5 min with methanol (Carl Roth, Karlsruhe, Germany) and stained for 15 min in 0.5% crystal violet (Alfa Aesar, Karlsruhe, Germany). The wells were then washed with running tap water, allowed to dry and photographed.
In vitro scratch assay 5 × 10 5 cells were seeded in triplicate into 12 well plates (Greiner Bio-One, 665,180), and incubated o/n, followed by siRNA transfection. Next day the wells were scratched with 200 μl pipette tips and reference points made using a needle. Images were taken from the wells at time points 0 and 24 h later using an inverted microscope (Leica, Wetzlar, Germany). ImageJ was used to measure the distance/gap between the two edges of the scratch. Migration was calculated as difference between gap distance at time 0 and 24 h divided by the distance at time 0.

Measurement of glucose consumption and lactate output
Cells (1.5 × 10 5 ) were seeded in triplicate into 12 well plates and allowed to attach o/n. Then, siRNA transfection was performed. 48 h after transfection, cell culture medium was collected and analyzed for glucose and lactate using the Roche Cobas C311 Chemistry Analyzer. Cells were lysed with RIPA buffer and data was normalized to protein concentration.

Metabolite analyses
One million of HUH7 and HLF cells were seeded in triplicate in 14.5 cm petri dishes and allowed to attach o/n. The next day, siRNA transfection was performed. After 24 h, medium was changed to growth medium for another 48 h. Thereafter, medium was removed, cells were washed with pre-warmed pure water and immediately snap frozen by adding liquid nitrogen to the petri dishes. Extraction of the samples was performed as follows: frozen HUH7 and HLF cells were extracted directly on petri dishes by adding pre-cooled 1 ml 50% methanol and 10 μl Ribitol (0.2 mg/ml; internal standard for the polar phase). All liquid containing cell debris was transferred to 2 ml reaction tubes on ice. To each tube, 0.5 ml 100% chloroform containing 0.1 mg/ml heptadecanoic acid (internal standard for organic phase) was added and samples were vortexed for 10 s. To separate polar and organic phases, as described previously [19,20], samples were centrifuged for 10 min at 11,000x g. For derivatization, 0.9 ml of the upper polar phase were transferred to a fresh tube and speed-vac dried without heating. Pellets of the aqueous phase after extraction were dissolved in 20 μl methoximation reagent containing 20 mg/ml methoxyamine hydrochloride in pyridine and incubated for 2 h at 37°C under vigorous shaking. For silylation, 35 μl N-methyl-N-(trimethylsilyl) trifluoroacetamide were added to each sample. After incubation for 45 min at 50°C, samples were analyzed by Gas Chromatography/Mass Spectrometry (GC/MS). A GC/MS-QP2010 Plus (Shimadzu, Germany) fitted with a Zebron ZB 5MS column (Phenomenex; 30 m × 0.25 mm × 0.25 μm) was used for GC/MS analysis. The GC was operating with an injection temperature of 250°C and 1 μl sample was injected with split mode (diluted 1:5). The GC temperature program for polar compounds started with 1 min hold at 40°C followed by a 6°C/min ramp to 210°C, a 20°C/min ramp to 330°C and a bake-out for 5 min at 330°C. Helium was used as carrier gas with consistent linear velocity. The MS was operated with ion source and interface temperatures at 250°C, a solvent cut time of 5 min and a scan range (m/z) of 40-700 with an event time of 0.1 s. Raw data were processed using the "GCMS solution software" (Shimadzu) and normalized to the internal standard Ribitol as well as to the cell number.

Statistics
Results are presented as mean ± SD unless indicated otherwise. Comparison of sample groups was performed using GraphPad Prism version 6 Software or R statistical software. Where applicable, t-test was applied for unpaired sample comparison, while one-way ANOVA was used for multiple comparisons. Statistical significance was defined with P < 0.05. Quantification of migration was done using ImageJ 1.5 (http:imagej.nih.gov/ij). Statistical significance is indicated as follows: * P < 0.05, ** P < 0.01, *** P < 0.001, and **** P < 0.0001.

PRRX1 is frequently upregulated in HCC
We analyzed PRRX1 expression in nine human HCC gene expression datasets (Table S1). Out of these datasets, PRRX1 level were upregulated in seven (P < 0.05), and not significantly changed in two datasets (Fig. 1a).
In addition, we compared the expression of PRRX1 across liver cancer cohorts in Oncominean online repository of manually curated cancer datasets (http:// oncomine.org). The Oncomine platform contains the cancer genome atlas (TCGA) liver cancer data and GSE14520 dataset used in our analysis, thus was also advantageous for cross-validating our observation. Indeed, in Oncomine platform PRRX1 was upregulated in three datasets, but not significantly changed in five other datasets (Fig. S1A). Hence, we concluded that when significantly altered, PRRX1 is upregulated in HCC.
We then compared the alteration frequency of PRRX1 in TCGA cancer datasets via cBioPortal (http:// cbioportal.org). PRRX1 alterations in liver cancer (i.e. HCC and cholangiocarcinoma) along with its alterations in lung, bladder and breast cancer cohorts, ranked very  Table S1. B PRRX1 expression in tumours showing its gene amplification. Data was analysed with Student's t-test. Amp: amplification C PRRX1 expression in tumours with TP53 and CTNNB1 mutation in TCGA liver cancer cohort. Data was analysed with Student's t-test. D Kaplan-Meier overall survival analysis (logrank test) based on PRRX1 expression in GSE14520 dataset. High PRRX1 n = 110; low PRRX1 n = 110. E PRRX1 expression in human HCC cell lines after 48 h culture. Normalized to the basal expression in HLF (lowest PRRX1 level). Bars indicate mean ± SD, n = 3 per group. F Ratio of PRRX1a to PRRX1b isoform expression in HCC cell lines after 48 h culture. Bars indicate mean ± SD, n = 3 per group high (within top 10% of all TCGA cohorts) (Fig. S1B). Of note, PRRX1 was amplified in~10% (n = 36) of TCGA HCC tumour samples, but this amplification did not strikingly overlap with known mutated genes in HCC, i.e. TP53 or CTNNB1 (Fig. S1C). As expected, HCC tumours with high amplification showed higher mRNA expression (P = 0.0007) compared to the tumours with no amplification (Fig. 1b). Further, we compared PRRX1 expression based on two frequently mutated genes in HCC (i.e. TP53 and CTNNB1). Accordingly, tumours with CTNNB1 mutation showed low PRRX1 level (P < 0.0001), whereas no expression change was observed with TP53 mutations (Fig. 1c), suggesting that the different oncogenic background influences PRRX1 level.
A prior study using the GSE14520 dataset showed an improved OS with high PRRX1 expression [12]. To gain insight on association of PRRX1 with clinicopathological variables, we analyzed clinical data from one of the two platforms used in GSE14520 (GPL3921, n = 225 HCC samples). Upon Kaplan-Meier overall survival (OS) analysis, we found a tendency towards improved outcome for patients with high PRRX1-expressing tumours, but this was not statistically significant (Fig. 1d). Also, PRRX1 was not associated with other clinical variables analyzed, e.g. tumour size, stage, alanine transaminase (ALT) and alpha-fetoprotein (AFP) level (Fig. S2A).
We further analyzed PRRX1 expression patterns in experimental HCC models. In multi-drug resistance (Mdr2) knockout micea liver fibrosis model that progresses to HCC in about 12 months [21,22] we observed higher Prrx1 expression in Mdr2 knockout mice than in age -matched controls and a lower expression of Prrx1 in older mice with HCC compared to the younger Mdr mice ( Figure S2B).
In HCC cell lines, PRRX1 expression was variable. Specifically, it was highest in SNU398, a cell line with epithelial-like appearance albeit reported to be poorly differentiated [23]. Other cell lines with epithelial features, HUH7 and Hep3B, showed higher PRRX1 level when compared to HLF cells that display more mesenchymal properties (Fig. 1e). In these four cell lines, we additionally analyzed expression of PRRX1 isoforms, a and b. With the exception of SNU398, the HCC cell lines (Fig. 1f) expressed less of the truncated isoform (i.e. PRRX1a) than the longer isoform PRRX1b. These analyses support that PRRX1 is frequently upregulated in liver tumours, while data from experimental models suggest that PRRX1 may be more expressed in well differentiated cells.
The top PRRX1-negatively correlated genes (n = 1022) were mostly involved in metabolism, a process that has not yet been associated with PRRX1 in HCC (Fig. 2b and  c). We found that prominent inversely correlated metabolic processes include glycine, serine and threonine metabolism (n = 16 genes, e.g., SHMT1, CHDH, GATM), fatty acid degradation (n = 16 genes, e.g., ECI1, ECI2, ACOX1), valine, leucine and isoleucine degradation (n = 16 genes, e.g. ACADSB, EHHADH, ECHS1), and peroxisome activity (n = 42 genes, e.g. ACOX1/2, HACL1, PEX1, 5, 6, 7, 16, and 19) (Fig. 2b). Metabolism also dominated the GO term 'biological processes' of the negatively correlated genes (Fig. 2c, Table S7). Further, the GO term 'cellular component' implicated peroxisomes and mitochondria (Fig. S3, Table S8)two organelles involved in antioxidant defenses and cellular respiration, respectively. Given the link to metabolism, we wondered if the PRRX1 inversely correlated genes included consistently altered metabolic genes in human HCC [24]. Indeed, overlap of all PRRX1-correlated genes with metabolic genes from human HCC revealed 135 common elements (Table S9, Fig. S5A). Of these, 124 metabolic genes were downregulated in HCC and several of them belonged to amino acid and fatty acid metabolism, and also small molecule transport (Fig. S5B). Thus, PRRX1 expression correlated with the downregulation of metabolic pathways in HCC. In cholangiocarcinoma and lung cancer, we also observed a negative correlation with metabolic pathways, suggesting that similar functions predicted for PRRX1 in HCC may apply to other cancer types (Fig. S4). Taken together, these data suggest that PRRX1 promotes cancer pathways and contributes to a suppressed metabolic gene program, as previously seen in HCC [4,24,25].
Integrative analyses identify EMT genes ZEB1 and ZEB2 as novel transcription factors related to PRRX1 Based on our in silico analyses of several HCC patient gene expression datasets, which linked PRRX1 to metabolism and cancer pathways, we first focused on its influence towards cellular plasticity and EMT. Thus, we stratified TCGA HCC dataset into high versus low PRRX1-expressing tumour samples, and analyzed the differential gene expression. We observed that the PRRX1-high tumours expressed several cancer-associated genes (e.g. MMP2/9, ZEB2, VIM) (Fig. 3a, Table S10). This expression pattern was consistent with that of the co-expressed gene list from cBioPortal (which generally reflected EMT, ECM, cancer and metabolic pathway alterations as earlier noted in Fig. 2). We compiled~150 genes involved in these processes (Table S11), including the topmost PRRX1 co-expressed genes (Fig. 2a), and sought to identify those correlated with PRRX1 in TCGA HCC data and another dataset (GSE14520). Of those genes, 19 significantly correlated with PRRX1 in both datasets ( Table 2). The EMT transcription factors ZEB1 and ZEB2 emerged alongside HNF1A, IL10, and 15 metabolic genes, e.g., gamma-glutamyltransferase 5 (GGT5), glucose transporter 2 (SLC2A2), NADPH oxidase 4 (NOX4), alcohol dehydrogenase 6 (ADH6), and hepatic lipase C (LIPC) (Fig. S6A). Kaplan-Meier OS analysis for each of the 19 genes in combination with PRRX1 showed that liver tumours with high PRRX1 and low ZEB1 (P = 0.0369), high ZEB2 (P = 0.0328), high GGT5 (P = 0.0015) or high SLC7A8 (P = 0.0016) showed improved OS outcome (Fig. 3b, Fig.  S6B). ZEB1 or ZEB2 did not predict OS when analyzed alone (Fig. S7A) indicating a synergistic impact with PRRX1. Since PRRX1 is associated with EMT [8,9,15], but no prior study had linked it with ZEBs in HCC, we performed further analysis on ZEBs.
Analyses of multiple HCC cohorts showed that both ZEB1 and ZEB2 expression positively correlated with PRRX1 in TCGA, GSE25097 and GSE55092 datasets. In addition, ZEB2 correlated positively with PRRX1 in the dataset GSE64041. On the other hand, we found one dataset each (GSE14520 and GSE36376) where ZEB1 and ZEB2, respectively, were negatively correlated with PRRX1 (Fig. 3c, Fig. S7B). Thus, PRRX1 is more often positively correlated with ZEB1 and ZEB2 in patients' HCC tumour samples. However, besides OS, PRRX1 combined with ZEB1 or ZEB2 expression was not significantly associated with several clinicopathological variables (Tables S12, S13). Next, we sought to determine whether ZEB1 and ZEB2 show a consistent expression pattern across the HCC cohorts. Out of nine datasets, we found that ZEB1 was high in six while ZEB2 was high in two datasets. ZEB2 was significantly low in five datasets (Fig. 3d, Fig.  S7C), thus indicating that ZEB1 is generally high in HCC while ZEB2 is often downregulated. We also analyzed ZEB1/2 expression in tumours with CTNNB1 and TP53 mutations, which are frequent in HCC. We found that ZEB1 is downregulated in TP53-mutated tumours and not changed in CTNNB1 mutations, whereas ZEB2 was upregulated in tumours with CTNNB1 mutation and not changed in TP53-mutated tumours (Fig. S7D), further supporting that the ZEBs often have a contrasting expression pattern.
In cell lines, we measured the basal mRNA level of ZEB1/2 relative to PRRX1 and observed a heterogeneous expression pattern. Specifically, poorly differentiated cell line HLF, which expresses PRRX1 at low level (Fig. 1e), showed comparatively higher levels of both ZEB1/2 (Fig. 3e). SNU398 cells, which express high PRRX1, showed low ZEB1 and high ZEB2, whereas HUH7 and Hep3B cells (well-differentiated cells) expressed higher ZEB1 and lower ZEB2 compared to PRRX1 expression (Fig. 3e). Taken together, we identify ZEB1 and ZEB2 as EMT genes correlated with PRRX1 and assumingly acting with it to influence tumour characteristics.

Modulation of PRRX1 alters HCC cell phenotype
We knocked down PRRX1 in HCC cell lines to experimentally validate its relationship with ZEB1/2 (Fig. 4a). Consistent with the positive correlation with ZEB1/2 seen in patient datasets, the knockdown of PRRX1 caused a significant downregulation of ZEB1 in all three cell lines tested. ZEB2 was less strongly regulated and changed significantly only in two cell lines (Fig. 4b). Furthermore, we tested the ability of the cells to migrate (as key EMT phenotype). Under basal conditions, the cell line HLF migrated more than HUH7, SNU398 and Hep3B cells. However, cell migratory response was variable upon PRRX1 knockdown, being increased in HUH7 and Hep3B cells, reduced in HLF and unchanged in SNU398 cells (Fig. 4c). This variability may be linked to the differentiation status of the cells. Further, HCC cell lines showed an increase in cell proliferation after 48 h (Fig. 4d) and clonogenicity after 8 days upon PRRX1 knockdown (Fig. 4e). Altogether, these findings suggest that modulating PRRX1 alters ZEB1/2 and is accompanied by various cellular phenotypes associated with cancer.

PRRX1 and HCC metabolism
Our previous work revealed consistent metabolic gene changes in HCC [24]. Thus, we further interrogated metabolic alterations, based on the negative correlation between PRRX1 expression and metabolic genes we observed in the array analyses. To first test a functional link, we measured glucose consumption and lactate output, which are indicative of the Warburg effect (aerobic glycolysis) and are well known to be elevated in cancer. Indeed, the knockdown of PRRX1 led to an increased glucose consumption in HUH7 and HLF, with a similar upward trend in SNU398 cells (Fig. 5a). In accordance, lactate secretion into the supernatant was increased in HUH7 and SNU398 significantly, and by tendency in HLF cells (Fig. 5a). Based on these findings, distinct gene expression patterns of key components of the glycolytic pathway were determined. A strong effect was monitored in HUH7, with hexokinases 1/2 being increased upon PRRX1 knockdown, as were PDHX, LDHA, and SLC1A5 (Fig. 5b). In contrast, lactate dehydrogenases B (LDHB) was reduced. Effects in HLF were less pronounced, but LDHA/B, which catalyze the conversion of pyruvate to lactate under anaerobic conditions, were strongly increased (Fig. 5b). Notable alterations of tricarboxylic acid cycle (TCA) genes were also detected in both cell lines. As shown in Fig. 5c, crucial enzymes of the TCA cycle were significantly increased in both cell lines upon PRRX1 knockdown, e.g. isocitrate dehydrogenases 3A/B (IDH3A/B), oxoglutarate dehydrogenase (OGDH), several succinate dehydrogenase complex flavoprotein subunits (SDH), fumarate hydratase (FH), and others (Fig. 5c, Fig. S8A). A few tested targets were reduced upon PRRX1 knockdown, e.g. phosphoenolpyruvate carboxykinase 1 (PCK1), which is responsible for the regulation of gluconeogenesis, and the succinate-CoA ligase GDP/ADP-forming subunit 1 (SUCLG1) in both cell lines (Fig. S8A). We also accessed amino acid metabolism genes, especially those involved in glutaminolysis and found them to be deregulated upon PRRX1 knockdown although for some the direction was cell type dependent (Fig. 5d, Fig. S8B). Considering glutaminase 1 (GLS1), which catalyzes the hydrolysis of glutamine to glutamate, the upregulation was consistent in both lines whereas glutamate-pyruvate transaminase 1 (GPT1) was strongly repressed (Fig. 5d). Based on the plethora of metabolic gene expression changes upon PRRX1 modulation, we performed mass spectrometrybased metabolomics to determine changes at the metabolite level. Gas chromatography-mass spectrometry revealed alterations in glycolysis, TCA cycle and amino acid metabolism, with the effect being strongest in HUH7 (the well differentiated cell) (Fig. 5e). While major changes in glycolysis were restricted to the increase of glucose-6-phosphate and fructose-6-phosphate (in HUH7), most analyzed metabolites in TCA cycle were increased. In contrast, HLF cells showed less significant metabolite changes and for all altered analytes, a reduction upon PRRX1 knockdown was observed. Consistent with these findings, an increase in several amino acids was found (except for serine that was significantly reduced) in HUH7 (Fig. 5e), whereas in HLF, the effects were minor. These complex changes on metabolite and gene expression levels are illustrated in Fig. 6a (HUH7) and 6B (HLF) in which affected components of glycolysis and TCA cycle are marked. Taken together, these data support that PRRX1 controls metabolism in HCC cells in a highly context dependent manner.

Discussion
Although previous studies investigated the abundance of the transcriptional co-activator PRRX1 in HCC [12,13], conflicting expression patterns were reported and PRRX1 co-regulated genes were not entirely known. In addition, the functions of PRRX1 in HCC remained largely unknown. In this study, we combined bioinformatics analysis of a large cohort averaging over 1000 HCC tumour samples to resolve the expression pattern of PRRX1 in HCC. The analysis was complemented with experiments to gain insight on the functional relevance of PRRX1 in this cancer type. We found that PRRX1 is frequently upregulated in human HCC cohorts, suggesting that its function is unilateral in HCC. But in our analysis, and consistent with the previous report by Hirata et al. [12], PRRX1 level did not correlate with several These genes significantly correlated with PRRX1 (P < 0.05) in two cohorts GSE14520 (n = 225 samples) and TCGA HCC (n = 371 samples). Pearson correlation analysis was used clinicopathological parameters of HCC patients. Although Hirata et al. found that high PRRX1 expression predicted longer OS, we observed a similar but marginal tendency that was not statistically significant. We attribute this discrepancy to differences in the number of samples used for the OS analysis. The study by Fan et al. [13] also pointed to a favorable outcome with high PRRX1 (determined by immunohistochemistry) as low expression correlated with vascular invasion, intrahepatic and distant metastasis as well as advanced tumour stage. Thus, based on the pattern of association with clinical parameters observed by others and us, PRRX1 can be considered as a tumour suppressor in HCC, at least in the advanced stages of the disease. PRRX1 is a typical example of an enigmatic gene that has context-specific functions in HCC. Consistent with this view, a tumour suppressor function seems incompatible with our observation of elevated PRRX1 in HCC cohorts. Of note, these cohorts also include TCGA data in which Hirata et al. [12] showed that high PRRX1 expression predicted better OS. Furthermore, the observation that PRRX1 expression is significantly altered in tumours with mutations in CTNNB1, but not TP53 implies that its expression patterns could substantially vary with additional molecular stratification.
HCC cell lines also provide evidence of variable PRRX1 level, with a notable low expression in HLF (a high migrating cell line), although not reflecting PRRX1 correlation to CTNNB1 (there is an extensive heterogeneity in the mutational landscape of the cell linesmaking it hard to attribute even their phenotypic differences to the various mutations [26]). Nevertheless, PRRX1 level did not clearly enable discrimination of well from poorly differentiated HCC cell lines unlike we and others have shown for other proteins, e.g. caveolin-1, albumin, alpha fetoprotein, SMADs, and WNT signaling targets [23,[27][28][29].
Evidence from prior studies in colon, breast, and pancreatic cancer support that PRRX1 has contextdependent functionsi.e. it can have a tumour promoter or suppressor function. Our experimental data with cell lines support a tumour suppressor function for PRRX1 in HCC. For example, through in silico analysis we predicted that PRRX1 likely acts by repressing metabolism in HCC. The knockdown of PRRX1 promoted an aerobic glycolytic phenotype (i.e. increased glucose consumption and lactate output). In a recent genome-wide association study linking type II diabetes sensitivity with gene expression patterns, the authors showed that PRRX1 is positively linked with insulin sensitivity, supporting the finding that PRRX1 is related to glucose metabolism [30]. In HUH7 cells, the TCA cycle was also strongly affected at gene and metabolite levels. These observations reflect a tumour suppressive function, even though more studies are required to validate the role of PRRX1 in cell metabolism and to define whether it is critical for HCC progression.
An increased proliferation rate upon PRRX1 knockdown also supports a suppressive function. The observation of these phenotypes, even in the cells with comparatively low basal PRRX1 (i.e. HUH7 and HLF), raises another possibility that certain functions of PRRX1 are conserved in the different HCC cell lines regardless of basal expression.
In agreement with other studies, we also observed that PRRX1 function is ambiguous, which likely depends on undetermined factors. For instance, Fan et al. showed an increased migration upon PRRX1 knockdown in HCC cell lines HEPG2 and SMMC7721 [13]. However, when tested across other cell lines in our work, migration was increased in Hep3B and HUH7 cells, unchanged in SNU398, and reduced in HLF cells. Similarly, with regards to the PRRX1 expression isoforms, most HCC cells expressed more PRRX1b isoform than PRRX1a. The exception was SNU398, which showed no change in proliferation and migration after PRRX1 knockdown. Thus, it could be speculated that PRRX1a, rather than PRRX1b is involved in its phenotypic roles in HCC. Further functional studies will be required to clearly resolve the conditions under which PRRX1 may exert specific functions in HCC and the exact contribution of the isoform ratio.
Our analysis of human HCC datasets indicated a direct correlation between PRRX1 and cancer-related processes, e.g. PI3K-Akt signaling pathway, focal adhesion, extracellular matrix, actin cytoskeleton activities and EMT. The exact role of PRRX1 in controlling the predicted cancer processes needs empirical validation, especially given that the knockdown of PRRX1 induced pro-cancer activities whereas its high expression in tumours had marginal clinicopathological correlation. PRRX1 has been described as EMT gene [8,9,15] and prior studies link ZEB1 with unfavorable clinical outcome in HCC [31][32][33]. Interestingly, we identified EMT genes ZEB1/2 as specific candidates that together with PRRX1 impact on clinical OS outcome. We found that PRRX1 correlated with ZEB1/2 in three HCC cohorts analyzed. Whereas none of the three genes independently predicted overall survival in the datasets we investigated, high Finally, an interesting observation is related to patients with CTNNB1 mutations. In these patients, PRRX1 expression was significantly lower whereas ZEB2 was increased. Also, tumours with TP53 mutations showed significantly reduced ZEB1 expression. It can be assumed that distinct cancer subtypes, defined by mutation status, will be associated with PRRX1/ZEB1/2 patterns and thus determine disease mechanisms and therewith outcome. Further work will have to shed light on this aspect.

Conclusion
We have provided evidence of consistent upregulation of PRRX1 in human HCC. We propose that this gene has anti-tumour functions, at least in advanced tumour stages based on patients' OS data and our experimental observations. We identify a novel function of PRRX1 in modulating glycolysis and TCA cycle, and that PRRX1 correlated with EMT transcription factors ZEB1/2 (especially with respect to patients' survival outcome). These findings will enable further in-depth mechanistic studies of the functional relevance of PRRX1 in human liver cancer, and its crosstalk with ZEB1/2.
Additional file 1: Figure S1. PRRX1 expression in HCC. Figure S2. Analysis of PRRX1 expression with respect to clinicopathological variables, and in experimental models. Figure S3. Cellular components in which are involved genes positively and negatively correlated with PRRX1 in TCGA liver cancer data. Figure S4. KEGG pathway annotation of PRRX1 correlated genes in various cancers where its alteration frequency is high. Figure S5. Annotation of PRRX1 inversely correlated genes. Figure S6. Correlation and survival analyses of PRRX1 and its co-expressed genes. Figure S7. Survival analysis and expression of ZEB1/2 also with respect to TP53 and CTNNB1 mutation. Figure 8S. PRRX1 and metabolic targets. Table S1. Human HCC microarrays. Table S3. KEGG pathway annotation of genes co-expressed (positively) with PRRX1. Table S4. KEGG pathway annotation of genes inversely correlated with PRRX1. Table S5. GO Biological processes for PRRX1 positively co-expressed genes. Table S6. GO Cellular components for PRRX1 positively co-expressed genes. Table S7. GO Biological processes for PRRX1 inversely correlated genes. Table S8. GO Cellular components for PRRX1 inversely correlated genes. Table S9. Metabolic targets in the PRRX1 co-expressed gene list. Table S11. 148 genes as potential candidates that likely cooperate or are coregulated with PRRX1. Table S12. ZEB1 expression in combination with PRRX1 with respect to the clinicopathological variables (n= number of patients). Table S13. ZEB2 expression in combination with PRRX1 with respect to the clinicopathological variables (n= number of patients).
Additional file 2: Table S2. PRRX1 positively and negatively correlated genes in TCGA liver cancer data.