Polyol Pathway Links Glucose Metabolism to the Aggressiveness of Cancer Cells

Differentiation and phenotypical heterogeneity have a fundamental impact on the aggressiveness of tumors (1, 2). Epithelial-to-mesenchymal transition (EMT) is a dedifferentiation process that cancer cells use to acquire an invasive, chemoresistant phenotype as well as the properties of cancer stem cells (CSC; refs. 3, 4). EMT/CSCs are therefore therapeutically very attractive, but a further understanding of the underlying biological processes is required to develop novel effective therapeutics (5). Some cancer types, and the corresponding CSC population in particular (6), are highly dependent on glucose metabolism and aerobic glycolysis, which they use as a major pathway for biosynthesis (7), and the molecular connection between glucose metabolism, EMT and CSCs has recently started to emerge (6, 8, 9). Importantly, epidemiologic and experimental studies have connected sucrose/fructose consumption with increased cancer risk (10–12), but the molecular determinants are still undefined and the impact of other glucose-related pathways has been poorly investigated. By means of an unbiased transcriptomic analysis aimed at identifying genes highly correlated with aggressive EMT-driven tumors, we found the aldo-keto reductase family 1, member B1 (AKR1B1). This gene encodes for a specific member of the aldo-keto reductase superfamily catalyzing the reduction of glucose to sorbitol (13), the first step of the polyol pathway (PP). Using different models, we provide evidence that the PP can functionally regulate EMT, representing a novel link between glucose metabolism and tumor aggressiveness.

All cell lines were cultured in media supplemented with 5% FBS, 1% penicillin/streptomycin and 1% l-glutamine (Sigma) at 37°C and 5% CO₂ in a humidified incubator. Cells were from the National Cancer Institute (those part of the NCI-60 panel, except MDA-MB-231 and HCT116) or from ATCC, were STR-profiling authenticated and used between passages 3 and 15, were examined for the presence of mycoplasma and maintained in Plasmocin to prevent contamination (detection kit and reagents from Invivogen). Breast (T47D, MCF-7, MDA-MB-231, BT549, Hs578T), ovarian (Ovcar -4,-5,-8, HeyA8), lung (A549, Calu-1, H1299), colorectal (KM12 and Colo205), as well as renal CAKI-1 and glioblastoma U251 cancer cell lines were cultured in RPMI (Sigma). Other colorectal cancer cell lines (HCT116, HT29) were cultured in McCoy 5A (Gibco). 293T and KPC(Z) cell lines were cultured in DMEM (Sigma).

For EMT induction, cells were treated with 10 ng/mL TGFβ (R&D Systems) for 2–3 days or with 50 or 100 mmol/L glucose (Sigma) for 3–6 days. Inhibition of TGFβ signaling was performed with a neutralizing TGFβ 1, 2, 3 antibody (R&D Systems, 1D11, 1 μg/mL) in the presence of either 11 mmol/L (concentration of standard media formulation) or 100 mmol/L glucose for 6 days. The corresponding IgG1 isotype antibody (R&D Systems, 11711, 1 μg/mL) served as a control. To test the sensitivity toward cisplatin, cells were treated with 350 or 500 μmol/L cisplatin (Sigma) the day after plating, using DMSO as control.

Overexpression experiments were performed as described previously (14) using the EX-C0237-Lv105 construct (Genecopoeia) for AKR1B1. Stable knockdown of human AKR1B1 and SORD expression was achieved using Sigma MISSION lentiviral transduction particles (TRCN0000288741 and TRCN0000288812 for AKR1B1-; TRCN0000028052, TRCN0000028069 and TRCN0000028106 for SORD- knockdown) and nontargeting controls. Cells were selected in media containing 3 μg/mL puromycin (Sigma).

Transient transfections were performed using siPORT NeoFX transfection agent (Applied Biosystems) according to the manufacturer's instruction. A total of 8 × 10⁴ cells were transfected in 12-well plates with 50 nmol/L siRNA (Ambion).

Cell proliferation and migration were monitored using live time-lapse images recorded with the IncuCyte ZOOM live cell analysis system (Essen BioScience). As readout for proliferation, the occupied area (% of confluence) is plotted over time. For migration, homogeneous scratch wounds were created using the 96-pin WoundMaker (Essen BioScience) and relative wound density is plotted over time.

For protein isolation, cells were lysed in RIPA buffer with Halt Protease & Phosphatase Inhibitor Cocktail (both Thermo Fisher Scientific), followed by centrifugation at 12,000 rpm. Proteins were quantified with the Pierce BCA Protein Assay Kit (Thermo Fisher Scientific), separated by SDS-PAGE (10%) and transferred onto a PVDF membrane (Thermo Fisher Scientific). Proteins were detected using AKR1B1 (Thermo Fisher Scientific, polyclonal), SORD (Thermo Fisher Scientific, polyclonal), E-cadherin (Cell Signaling Technology, 4A2), vimentin (Cell Signaling Technology, D21H3), ZEB1 (Sigma, polyclonal), Fra-1 (Cell Signaling Technology, D80B4), Smad3 (Cell Signaling Technology, C67H9), pSmad3 (Cell Signaling Technology, C25A9), β-actin HRP-conjugated (Cell Signaling Technology, 8H10D10), and species-specific HRP-conjugated secondary antibodies (Southern Biotech). Bands were detected using the Pierce ECL Western Blotting Substrate (Thermo Fisher Scientific), X-ray CL-XPosure films (Thermo Fisher Scientific) and the automatic film processor CP1000 (AGFA).

Immunofluorescence staining of cells grown on cover slips was performed as described previously (14). Signals were visualized using Leica DM5500B fluorescence microscope and Leica Application Suite-X software.

To assess clonogenic ability, a colony formation assay was performed using cells plated at a low density (1,000 cells/well), as described previously (14).

Bioenergetics of control and knockdown cell lines were determined using the XFe96 Extracellular Flux Analyzer (Seahorse Bioscience/Agilent Technologies). Cells were seeded in specialized cell culture microplates at a density of 2,000 cells/well and cultured for 72 hours. One hour before the measurement, cells were incubated at 37°C in a CO₂-free atmosphere. For the determination of glycolytic parameters, basal extracellular acidification rate (ECAR; indicative of glycolysis) was first determined under glucose-free conditions. Second, the rate of glycolysis was calculated using the ECAR after glucose supplementation (10 mmol/L). For the determination of respiratory parameters, basal oxygen consumption rate (OCR, indicator for mitochondrial respiration) was measured. All experiments were performed in pentaplicates. Raw values were normalized to the total protein content for each well.

The proportion of cancer stem cells was determined using the Aldefluor Kit (StemCell Technologies) according to the manufacturer's instruction. Propidium iodide (PI; Sigma) staining was performed as described previously (14). Cells were analyzed with FACSCalibur (BD Biosciences) and FlowJo (Treestar Inc, version 8.8.6) or CytoFLEX and CytExpert 2.0 (Beckman Coulter).

Tumorsphere assays were performed by seeding 20,000 cells/well in ultra-low attachment plates (Sigma) in complete MammoCult Medium (StemCell Technologies) supplemented according to the manufacturer's instructions for 8 days. Analysis was performed by counting the number of spheres (1/20 dilution in quadruplicates) or lysing proteins for Western blot analysis.

RNA was isolated using miRNeasy Mini Kit (Qiagen) according to the manufacturer's instruction. Five-hundred nanograms of RNA were used to perform reverse transcription with the Tetro cDNA synthesis Kit (Bioline) according to the manufacturer's instructions. DNA was amplified with gene-specific primers using TaqMan Universal Master Mix II (Thermo Fisher Scientific). Expression values were measured using ABI Prism 7300. Data were analyzed using the ΔΔC_t method and normalized to GAPDH. Results are shown as relative fold expression compared with control cells.

Before and during the library preparation, the quality of RNA was analyzed using a 2100 Bioanalyzer system (Agilent Technologies). Barcoded RNA sequencing libraries were prepared from 100 ng total RNA using Illumina's TruSeq stranded mRNA kit according to the manufacturer's instructions. Libraries were subjected to single-end sequencing (101 bp) on a HighSeq-2500 platform (Illumina). Quality filtering was performed using cutadapt v. 1.9.1; then reads were mapped against the human reference genome using the STAR aligner v. 2.5.2b, and a STAR genome directory created by supplying the Ensembl gtf annotation file (release 87) for GRCh37. Read counts per gene were obtained using featureCounts program v. 1.5.1 and the Ensembl gtf annotation file. Following analyses were performed using R version 3.3.1. In particular, differential expression analysis was performed with the DESeq2 package v.1.12.3. The total number of expressed genes was 19.165. Data are deposited in the GEO database with the accession number GSE106169. Genes were considered differentially expressed when they were at least two fold up-/downregulated and if their Benjamini–Hochberg adjusted P value was less than 0.01. The statistical significance of the overlap between differentially expressed genes was assessed by means of the hypergeometric distribution test. Gene-set enrichment analysis was performed with the online application available at http://software.broadinstitute.org/gsea/index.jsp).

Lists of differentially expressed genes between shAKR1B1 and shSORD versus control cells were compared to calculate the degree of overlap. A previously generated signature comprising 494 genes differentially expressed during EMT in different cancers (15) was compared with protein-coding genes differentially expressed upon AKR1B1 and SORD knockdown considering the directionality, that is, genes downregulated in shAKR1B1 and shSORD were compared with genes upregulated during EMT and vice versa. Overlaps are shown as Venn diagrams. The significance of the overlap between the gene lists was tested by a hypergeometric distribution test as implemented in http://nemates.org/MA/progs/overlap_stats.html. The knockdown (KD) score in patients was defined as the sum of z scores of the commonly upregulated genes minus the sum of z scores of the commonly downregulated genes upon AKR1B1 and SORD knockdown for each patient. z scores were calculated in SPSS software. The separation of patients based on their KD score was done from lower and upper quartiles for GSE14333 and GSE7390, and from median for GSE17710. Kaplan–Meier plots were generated in GraphPad software. Statistical significance of the separation between the two curves was assessed by log-rank test. The lists of genes used in all the analyses are provided in the Supplementary Table.

Chromatin immunoprecipitation was performed as described previously (16), except additional crosslinking with 1.5 mmol/L ethylene glycol bis(succinimidyl succinate) (EGS) for 30 minutes before fixing with 1% formaldehyde. One-hundred micrograms of chromatin per experimental condition was incubated with 5 μg anti-ZEB1 (Santa Cruz Biotechnology, H102) or 5 μg normal rabbit IgG control antibody (Santa Cruz Biotechnology). After precipitation and decrosslinking, DNA was purified with a QIAquick PCR Purification Kit (Qiagen). AKR1B1 promoter regions in ZEB1 bound chromatin were quantified using real-time PCR. Primers used: EPCAM, forward (5′-GCC AGG TAA AAG CTC AAA GG-3′) and reverse (5′-GCG GGA ACT GGA TAG AGG A-3′); AKR1B1-1 forward (5′-AGC CGT CTC CTG CTC AAC-3′) and reverse (5′-CGG AGA GTG TGA GGC GAG-3); AKR1B1-2 forward (5′-CGC TTT CCC ACC AGA TAC AG-3′) and reverse (5′-TAG TGG CAG CGGATT CTT TC-3′); AKR1B1-3 forward (5′-CCT AGA GTG GGG TGC AAA GA-3′) and reverse (5′-TCC TTT TGC AAA GCA CCT TC-3′); HPRT forward (5′- TGA GAG TTC AAG TTG AGT TTG GA-3′) and reverse (5′-TGA TAA TTT TAC TGG CGA TGT CA-3′).

For measurement of extracellular TGFβ, culture supernatants were collected 48 hours after plating in serum-free media and centrifuged at 1,000 rpm and 4°C for 10 minutes. ELISA was performed using the human TGFβ 1 ELISA kit (Sigma) according to the manufacturer's instruction. Absorption rates were measured with the SpectraMax190 (Molecular Devices) and normalized to the cell number at the time of collection.

Formalin-fixed paraffin embedded (FFPE) surgical specimens of tumor tissues of 59 NSCLC patients completely resected between 2005 and 2006 at the San Luigi Hospital (Turin, Italy) were consecutively collected. The main patients' characteristics are shown in Supplementary Fig. S3A. None of the patients received presurgical chemo/radiation therapy. All cases were reviewed and classified (WHO classification) using anonymized samples. The studies were conducted in accordance with the Declaration of Helsinki. The Research Ethics Committee of the San Luigi Hospital, University of Turin (Turin, Italy), had approved the retrospective use of solid tumor tissues for IHC staining (approvals no.167/2015 and 204/2016).

Mouse specimens were obtained from a murine model of AOM-induced colon tumorigenesis with an intestinal epithelial cell–specific p53 deletion (p53ΔIEC) and corresponding wild-type animals (17). Two investigators (I. Rapa and M. Volante) independently reviewed the hematoxylin and eosin–stained FFPE tissue sections to identify tumors. Of the n = 30 p53ΔIEC tumors, n = 24 were noninvasive and n = 6 were locally invasive. Samples were stained and scored as described below. All mouse experiments were reviewed and approved by the Regierungspräsidium Darmstadt (Darmstadt, Germany).

Tissue sections were deparaffinized and rehydrated with a series of decreasing ethanol concentrations. Antigen retrieval was performed using a pressure cooker in citrate buffer (pH 6) for 30 minutes. For AKR1B1 detection, endogenous peroxidase activity was blocked using 3% H₂O₂ for 15 minutes, followed by blocking of nonspecific binding sites with 1% BSA in TBS-T for 30 minutes. The primary antibody incubation (AKR1B1 from Thermo Fisher Scientific at 1:25, or the E-cadherin from Neomarkers at 1:30) was performed for 1 hour. After washing, all sections were incubated with rabbit Dako REAL Envision-HRP for 40 minutes. DAB substrate (Dako) was used for visualization of antibody binding. Sections were counterstained with hematoxylin, rehydrated, and mounted. Immunoreactivity scorings were evaluated independently by two pathologists with an H-score, which was generated from the following equation: H-score = ΣPi (i + 1), where “I” represents the intensity of staining (0–3+), and “Pi” stands for the percentage of stained tumor cells (0%–100%).

Cells were harvested for metabolite analysis by direct scraping in 80% methanol as described in ref. 18. For sample preparation, the cell suspension was thawed, vortexed, and spiked with 10 μL internal standard containing ¹³C₆-glucose at a concentration of 1mmol/L. The sample was then centrifuged (9,560 × g, 6 minutes, 4°C) and the supernatant was removed. The cell pellet was washed twice with 200 μL 80% methanol, and the supernatants were combined and dried in a vacuum evaporator (CombiDancer). Sugars were analyzed by GC-MS after methoximation and silylation using the derivatization protocol and instrumental setup recently described (18). An injection volume of 1 μL and splitless injection were employed. Quantification was performed using calibration curves with ¹³C₆-glucose as internal standard. Intracellular metabolite levels were normalized to protein amount. The protein pellet obtained after extraction was resuspended in a NaH₂PO₄ buffer (20 mmol/L with 1.2 % SDS) and the protein amount was determined using the FluoroProfile Protein Quantification Kit (Sigma).

In the transcriptomic analysis (performed on Novartis and U133 gene array datasets), correlations were assessed by Pearson's statistics, statistical difference between groups was assessed with Mann–Whitney U tests, and post hoc analysis was performed with the Bonferroni method for multiple comparisons. Significant associations with clinical–pathologic variables were tested with the Mann–Whitney U and the Kruskal–Wallis tests. In vitro experiments were performed in replicates, independently repeated and representative results are shown. t tests in the in vitro experiments were performed comparing groups of different conditions with replicates. In Western blot experiments, integrated densitometry data were obtained using the ImageJ software. Statistical significance in proliferation and migration assay was analyzed with two-way ANOVA by GraphPad Prism 7.0. Results were considered statistically significant if P < 0.05 (*, P < 0.05; **, P < 0.01; ***, P < 0.001).

A mesenchymal-like phenotype is a common feature of dedifferentiated aggressive cancers (5, 19). To identify genes associated with EMT in an unbiased fashion, we performed a transcriptomic analysis using two gene-array datasets from the NCI-60 panel of cancer cell lines (Fig. 1A). For each dataset, we generated a vimentin/E-cadherin gene ratio (VIM/CDH1) and the two ratios were found to be strongly correlated (Supplementary Fig. S1A). The ability of this gene ratio to differentiate between epithelial-like and mesenchymal-like cancer cells was validated dividing cells by EMT status according to a previously reported Western blot quantification (Fig. 1B; ref. 20). Pearson statistics was then used to identify genes most significantly positively correlating with VIM/CDH1 in both datasets (mesenchymal genes, Fig. 1A). The genes present in both lists were reranked for the fold level of increase (mesenchymal- versus epithelial-like) and the final list included a total of 35 genes (Fig. 1C; Supplementary Fig. S1B). Several known mesenchymal markers were present, like fibronectin (FN1), TGFβ1 (TGFB1), and N-cadherin (CDH2), as well as recently identified EMT determinants, like LOXL2 (21) and VCAN (Supplementary Fig. S1B; ref. 22). Among the top-ranked genes, we focused our attention on the aldo-keto reductase family 1, member B1 (AKR1B1), which exhibited in average 7-fold higher mRNA expression in mesenchymal-like cells (Fig. 1D). This was also further confirmed analyzing the dataset from the Cancer Cell Line Encyclopedia (CCLE) on more than 900 cancer cells, dividing the cells in epithelial- and mesenchymal-like based on the VIM/CDH1 gene ratio (Fig. 1E). This gene encodes an enzyme catalyzing the reduction for numerous aldehydes, most importantly glucose as part of the polyol pathway. Other members of the same superfamily without any connection to the PP, such as AKR1B10 (71% sequence identity; ref. 23) did not show a correlation with EMT (Supplementary Fig. S1C), suggesting that the role in EMT could be a specific property of the B1 member.

To validate these findings at the protein level, we quantified AKR1B1 in lysates from a panel of cancer cells (mainly belonging to the NCI-60). The results of this analysis (Fig. 1F) clearly indicated a significant increase in the expression of AKR1B1 in mesenchymal-like cells (indicated by no E-cadherin, high vimentin, and ZEB1 expression), with the appearance of a specific band at 37–38 kDa. Few exceptions were observed, but overall the analysis of the protein level confirmed the data from mRNA levels and emphasizes the possible role of this enzyme in EMT.

The transcription factor ZEB1 is a master regulator of EMT (24). Gene array analysis from the NCI-60 panel indicated a significant positive correlation between AKR1B1 and ZEB1 expression (Fig. 1G). To functionally explore this association, we measured mRNA and protein levels in MDA-MB-231 cells stably knocking down ZEB1, a condition that suppresses EMT morphology and markers expression and found that AKR1B1 was significantly downregulated (Fig. 1H–J), strengthening its association with EMT. Conversely, analysis of GEO profiling data from breast cancer cells in which miR-200c (ZEB1 inhibitor and EMT-suppressing miRNA) was overexpressed (25) indicated a significant Akr1b3 (AKR1B1 mouse ortholog) reduction, along with EMT suppression (Supplementary Fig. S2A). A significant AKR1B1 downregulation was also obtained knocking down ZEB1 with siRNA and shRNA in lung cancer A549 cells (Supplementary Fig. S2B and S2C). Similarly, protein analysis of cell lines derived from pancreatic tumors of Kras^G12D;Trp53^R172H;Pdx1-Cre (KPC) mice and their Zeb1 knockout counterpart (KPCZ; ref. 24), indicated higher AKR1B1 levels in the mesenchymal-like KPC and a trend toward a lower expression in the KPCZ, see Supplementary Fig. S2D and S2E. To understand whether ZEB1 was directly involved in AKR1B1 regulation, we performed chromatin immunoprecipitation and transfection experiments, but the results failed to indicate a direct regulation, as no significant binding to AKR1B1 promoter was detected (Supplementary Fig. S2F), and no AKR1B1 upregulation followed ZEB1 overexpression (Supplementary Fig. S2G and S2H).

To test the association between AKR1B1 and EMT in patient tissues, samples from a cohort of consecutively resected non–small cell lung cancer (NSCLC) patients (n = 59, Supplementary Fig. S3A) were stained by IHC and a significant negative correlation between AKR1B1 and E-cadherin was found (Fig. 2A and B). Also, a trend for a higher AKR1B1 expression in tumors with a higher N status was observed (extent of lymph nodal involvement, Supplementary Fig. S3B). Interestingly, analysis of comparative profiling (GEO profiles) on airway epithelial cells isolated from nonsmoker versus smoker subjects, revealed a striking increase in AKR1B1 levels, suggesting a possible role on lung tumorigenesis (Fig. 2C). Survival analysis from large datasets of NSCLC, ovarian, and gastric cancer patients confirmed a strong negative prognostic role for AKR1B1 (Fig. 2D; Supplementary Fig. S3C).

To test the in vivo relevance of AKR1B1 in tumorigenesis and invasiveness, an IHC analysis was performed on samples from a mouse model of AOM-induced colon tumorigenesis with an intestinal epithelial cell–specific Trp53 deletion (p53^ΔIEC; Fig. 2E; ref. 17). This model has been previously characterized as shifting from a noninvasive to an invasive EMT-like phenotype upon p53 inactivation, (17), in line with p53′s EMT-suppressing properties (26). Samples were scored for AKR1B1 expression and the results indicated significantly higher expression in invasive tumors from p53^ΔIEC mice as compared with both p53-deficient noninvasive or wild-type counterparts (Fig. 2F–H). All together, these results obtained in tissues confirmed the association from cultured cells and strengthened the correlation of AKR1B1 with aggressive EMT-driven tumors.

To understand whether AKR1B1 plays a role in cancer EMT, we subjected A549 cells to shRNA-mediated knockdown using two independent sequences targeting AKR1B1, and Western blot analysis indicated that the EMT phenotype was suppressed (Fig. 3A). Knockdown cells acquired a cobblestone shape (Fig. 3B), and immunofluorescence staining confirmed increased E-cadherin and reduced vimentin expression (Fig. 3C). To check whether EMT reversal affected migratory ability, a hallmark of EMT, a wound-healing assay was performed and indicated a significant suppression of motility in shAKR1B1 cells (Fig. 3D). Migration experiments were also analyzed with a matrix (relative wound density) that normalizes for proliferation. This is because, along with a reduced motility, a significant reduction in growth rate (Fig. 3E) and colony formation (Fig. 3F) was observed, in line with previous reports (27). The reduction in proliferation was comparable with that measured in shZEB1 cells (Supplementary Fig. S3D), possibly linking the loss of growth potential to the EMT suppression. No significant difference was observed in the relative proportion of cells in each phase of the cell cycle (Fig. 3G; Supplementary Fig. S3E). Similar results in terms of E-cadherin expression and growth suppression upon AKR1B1 knockdown were obtained in the ovarian cancer cell line Ovcar-5 (Fig. 3H and I). Conversely, AKR1B1 overexpression in MCF-7 cells reduced E-cadherin and slightly increased ZEB1 (Fig. 3J), while stimulating proliferation (Supplementary Fig. S3F). Altogether, these results indicated that AKR1B1 has a direct impact on EMT-associated phenotypes and proliferation.

EMT and the CSC phenotype are strictly functionally associated (28). We analyzed GEO profiling data of ovarian cancer cells grown as monolayers compared with 3D spheroids, a condition used to enrich for CSCs (29), and AKR1B1 was found to be the top-ranked upregulated gene (Fig. 4A). We therefore compared the AKR1B1 protein levels in HeyA8 (ovarian) and A549 cells maintained in adherent or spheroid-forming conditions, and an increased AKR1B1 expression in the tumorspheres was detected (Fig. 4B). Consistently, shAKR1B1 cells displayed reduced ability to form spheres (Fig. 4C; Supplementary Fig. S3G). FACS analysis of activated aldehyde dehydrogenase H1 (ALDH1), an established CSC marker (30), indicated a significant reduction of ALDH1⁺ cells upon knockdown (Supplementary Fig. S3H). Comparable results were also obtained treating cells with the AKR1B1 inhibitor sorbinil (Fig. 4D–F), confirming the involvement of AKR1B1 in cancer stemness.

AKR1B1 belongs to the polyol pathway (PP), a two-step enzymatic reaction converting excess intracellular glucose into fructose, and sorbitol dehydrogenase (SORD) is the second-in-line enzyme (Fig. 5A; ref. 13). We therefore investigated the impact of SORD loss on EMT. As a result, shRNA-mediated knockdown of SORD was found to significantly block EMT and CSC marker expression (Fig. 5B and C), and to inhibit colony formation (Fig. 5D), migratory ability, and growth (Fig. 5E and F) to a similar extent as the AKR1B1 knockdown. Comparable results were also obtained from other cancer cells (Fig. 5G; Supplementary Fig. S4A–S4C) and SORD gene expression revealed a prognostic role in NSCLC (Fig. 5H). Importantly, both PP-deficient cells still displayed sensitivity towards cisplatin and the combination of knockdown and drug treatment induced an almost complete proliferation arrest (Supplementary Fig. S4D). Assessment of cells' metabolic phenotype with the Seahorse analyzer indicated a comparable significant loss of activity in both knockdown cells, affecting basal respiration (Supplementary Fig. S4E) as well as glycolysis, a condition that could not be rescued by the addition of external glucose (Fig. 5I). Interestingly, SORD knockdown was found to induce a significant reduction in AKR1B1 levels in some cells (e.g., Ovcar-5 in Fig. 5G), indicating the possible existence of a cell- or tumor-specific coregulation of these genes. However, a comprehensive analysis of AKR1B1/SORD coexpression levels in tumors from the TCGA dataset indicated tissue-specific patterns and a higher prevalence of a negative correlation (Supplementary Fig. S5).

To better characterize changes associated with the loss of PP genes at an unbiased and high-throughput level, we performed RNA sequencing profiling of shAKR1B1 and shSORD cells compared with scrambled controls (Fig. 6A). The results indicated a very significant overlap (P < 0.0000001) between the genes up- and downregulated in both knockdowns (Fig. 6B) and were successfully validated by qPCR (Fig. 6C). Gene-set enrichment analysis (GSEA) revealed that the pathway most significantly altered in the PP-deficient cells (overlapping genes) was EMT (Fig. 6D). The comparison with a previously reported EMT signature comprising 494 coding genes (15) indicated a very significant overlap (Fig. 6E; Supplementary Table), further confirming the functional association with EMT. Importantly, a signature generated with the genes most differentially downregulated in shAKR1B1/shSORD cells (see Supplementary Methods) proofed to be strongly prognostic in different types of cancer (Fig. 6F).

Among the list of EMT genes suppressed by the knockdown of PP enzymes, we noticed the presence of several TGFβ-related genes, including TGFB1 (Fig. 7A). This was verified by a further comparison with a previously reported TGFβ-specific gene signature (P < 0.0001; ref. 31). TGFβ is a cytokine with an established major role in promoting EMT (32). Reduced levels of the soluble active form of TGFβ were measured in the supernatant of knockdown cells (Fig. 7B), and exogenous TGFβ addition rescued EMT/CSC markers in PP-deficient cells (although with delayed kinetics, Fig. 7C–E). On the basis of this evidence and on the previously reported connection between glucose uptake (33), metabolism (8), and EMT, we decided to further explore the interplay between glucose, PP, and TGFβ-driven EMT. First, we tested whether excess glucose had polyol-generating and EMT-promoting effects. Cells grown in the presence of high glucose levels displayed a marked increase in PP metabolites, as evaluated by GC-MS analysis (Fig. 7F; Supplementary Fig. S6A), along with a significant increase in EMT markers (see the vimentin/E-cadherin ratios), and AKR1B1 expression levels (Fig. 7G; Supplementary Fig. S6B), and an augmented migratory ability (Fig. 7H). Then, we hypothesized that glucose could promote EMT via the PP by an autocrine TGFβ stimulation. High glucose treatments were therefore repeated in the presence of a TGFβ-neutralizing antibody, and this was sufficient to abrogate the glucose-induced increase of EMT markers (Fig. 7I). Augmented TGFβ signaling following stimulation with excess glucose was further confirmed by an upregulation of phosphorylated Smad3 (Supplementary Fig. S6C) and RUNX2 transcript levels (Supplementary Fig. S6D) as well as a significant increase in the Fra-1 protein (FOSL1 gene; Supplementary Fig. S6E), previously shown to control TGFB1 gene and EMT (34). Conversely, exogenous overexpression of AKR1B1 in epithelial-like cells was capable of increasing TGBF1 and RUNX2 (Supplementary Fig. S6F). Finally, PP-deficient cells exposed to high glucose were found more refractory to EMT induction compared with control cells (Fig. 7J). The knockdown of PP enzymes also consistently reduced cells' ability to increase the proportion of ALDH1⁺ cells upon glucose treatment (Supplementary Fig. S6G). We could therefore conclude a model in which excess glucose is partly metabolized in the PP supporting EMT via TGFβ autocrine stimulation (see Fig. 7K).

The study of metabolic processes represents an emerging field in cancer biology (35) with a great translational potential (14, 36–38). Here, we describe by several approaches that a glucose-metabolizing pathway is strongly correlated and functionally associated with cancer aggressiveness and the cancer differentiation program. Our data suggest that interfering with targetable cancer-relevant metabolic pathways may represent a valuable strategy to block or attenuate the EMT/CSC process and cancer malignancy.

PP has been implicated in the development of diabetic complications via the generation of reactive oxygen species (13). The activity of AKR1B1 has been previously linked to EMT in mouse normal lens epithelium (39, 40) or in other normal cells (41). In cancer, AKR1B1 has been found to have a role in the growth and the aggressiveness of cancer cells (27), but was never associated with cancer cells' EMT, CSCs, or related pathways, until recently. Wu and colleagues, in fact, reported that AKR1B1 can induce EMT via PGF2α synthesis and NFκB activation in triple-negative breast cancer cells (42). On the basis of the data here presented, which point at the PP as strongly connected to EMT as well as the plasticity of other types of cancer cells, it is possible to speculate that the same enzyme has multiple ways to regulate EMT (in a tissue-specific manner). The expression of SORD has been previously detected in cancer cells (43, 44), but was never functionally investigated, and the current data indicate an important role in cancer growth and differentiation worth of further investigations.

AKR1B1 deficiency in mice has been found to protect from AOM-induced preneoplastic aberrant crypt foci formation (45). The presence of an intestinal epithelial cell–specific p53 deletion was previously demonstrated to initiate EMT and promote invasive disease in this model (17), and we could here show a significant increase in AKR1B1 protein levels in samples from p53-deleted mice, particularly evident in invasive tumors. The in vitro and in vivo data (including from patients samples) of the current work strongly encourage further in vivo experimentations on AKR1B1-deficient mice to conclusively determine the role of AKR1B1 and associated pathways in tumorigenesis, in the EMT phenotype, and in cancer stemness. Importantly, an unbiased analysis of global RNA profiles in PP-deficient cells revealed a dramatic loss of EMT and TGFβ signature genes, most of which were previously found altered in the p53-deleted invasive mouse model (17), and this observation prompted us to investigate the interaction between glucose, PP, and TGFβ in more detail. Several epidemiologic and experimental studies have connected sucrose/fructose consumption with increased cancer risk (10–12). A few previous reports suggested that glucose uptake (33) and metabolism (8), as well as fructose (46) could be connected to the occurrence of EMT, metastasis formation, and with a more advanced stage in tumor progression (47), but the functional details were not defined. The current data highlight the role of alternative glucose metabolism pathways in cancer dedifferentiation and aggressiveness clearly pointing at a direct role of glucose as an EMT promoter, as previously reported in some noncancerous physiologic contexts (48–50). A number of open questions remain to be answered. For instance, the detailed molecular mechanisms downstream of PP need to be identified for a deeper comprehension of the link between glucose/fructose and tumorigenesis. In addition, the exact causes for the observed loss of proliferation and metabolic activity in PP-deficient cells remain to be determined. Furthermore, the observed tissue-specific changes in the AKR1B1/SORD coexpression pattern (Supplementary Fig. S5) suggest the importance of investigating the transcriptional regulation of these EMT-promoting genes, as previously done for breast cancer (42), and the existence of a possible balance between them that controls cancer differentiation.

Noteworthy, PP genes and PP-related signatures were here found to be highly prognostic in various human cancer types, further confirming their important role in cancer progression. Targeting the expression or the activity of PP enzymes (alone or in combination with chemotherapy) may therefore represent a novel effective therapeutic strategy for the treatment of aggressive tumors.

M. Volante reports receiving a commercial research grant from Novartis Pharma and Ipsen and has received speakers bureau honoraria from Novartis Pharma and MesoScale Discovery. No potential conflicts of interest were disclosed by the other authors.

Conception and design: A. Schwab, T. Brabletz, P. Ceppi

Development of methodology: A. Schwab, A. Siddiqui

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): A. Schwab, A. Siddiqui, M.E. Vazakidou, F. Napoli, M. Böttcher, B. Menchicchi, U. Raza, A.M. Krebs, I. Rapa, K. Dettmer-Wilde, A.B. Ekici, S.A.K. Rasheed, D. Mougiakakos, P.J. Oefner, O. Sahin, M. Volante, F.R. Greten, P. Ceppi

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): A. Schwab, M. Böttcher, U. Raza, Ö. Saatci, F. Ferrazzi, K. Dettmer-Wilde, A.B. Ekici, D. Mougiakakos, O. Sahin, T. Brabletz, P. Ceppi

Writing, review, and/or revision of the manuscript: A. Schwab, B. Menchicchi, M.J. Waldner, A.B. Ekici, S.A.K. Rasheed, M. Volante, P. Ceppi

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): M.J. Waldner, P. Ceppi

Study supervision: P. Ceppi

Other (critically discussed the work): I. Rapa

This work was supported by the Interdisciplinary Center for Clinical Research (IZKF) of the University of Erlangen-Nuremberg, the Deutsche Krebshilfe grant number 70112536, the IALSC Lung Cancer Young Investigator Award (to P. Ceppi), and by the Clinical Research GroupKFO262 funded by the German Research Foundation. This work was partially presented at the American Association for Cancer Research 2017 annual meeting. Special thanks to Dr. H. Wurdak (University of Leeds) for critical discussion.

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.