Glucose Deprivation Promotes Pseudohypoxia and Dedifferentiation in Lung Adenocarcinoma

Epigenetic adaptation allows cancer cells to overcome the tumor-suppressive effects of glucose restriction by inducing de-differentiation and an aggressive phenotype, which could help design better metabolic treatments.

Abstract

Increased utilization of glucose is a hallmark of cancer. Sodium-glucose transporter 2 (SGLT2) is a critical player in glucose uptake in early-stage and well-differentiated lung adenocarcinoma (LUAD). SGLT2 inhibitors, which are FDA approved for diabetes, heart failure, and kidney disease, have been shown to significantly delay LUAD development and prolong survival in murine models and in retrospective studies in diabetic patients, suggesting that they may be repurposed for lung cancer. Despite the antitumor effects of SGLT2 inhibition, tumors eventually escape treatment. Here, we studied the mechanisms of resistance to glucose metabolism-targeting treatments. Glucose restriction in LUAD and other tumors induced cancer cell dedifferentiation, leading to a more aggressive phenotype. Glucose deprivation caused a reduction in alpha-ketoglutarate (αKG), leading to attenuated activity of αKG-dependent histone demethylases and histone hypermethylation. The dedifferentiated phenotype depended on unbalanced EZH2 activity that suppressed prolyl-hydroxylase PHD3 and increased expression of hypoxia-inducible factor 1α (HIF1α), triggering epithelial-to-mesenchymal transition. Finally, a HIF1α-dependent transcriptional signature of genes upregulated by low glucose correlated with prognosis in human LUAD. Overall, this study furthers current knowledge of the relationship between glucose metabolism and cell differentiation in cancer, characterizing the epigenetic adaptation of cancer cells to glucose deprivation and identifying targets to prevent the development of resistance to therapies targeting glucose metabolism.

Significance:

Epigenetic adaptation allows cancer cells to overcome the tumor-suppressive effects of glucose restriction by inducing dedifferentiation and an aggressive phenotype, which could help design better metabolic treatments.

Graphical Abstract

Introduction

Rewiring of energy metabolism is a hallmark of cancer, and the increased dependency on glucose is a common characteristic of cancer cells (1, 2). Malignancies require glucose for quick production of energy and for macromolecule biosynthesis (3). Oncogenic signaling induces metabolic rewiring, creating tumor-specific metabolic vulnerabilities. Kras-induced metabolic rewiring is specifically geared toward increased glycolysis (4), and glucose uptake via Glut1 and Glut3 is required for Kras-driven oncogenesis (5). Much interest and research effort has focused on targeting glucose uptake and glycolysis for therapeutic purposes (6–9). However, the clinical results with metabolic therapies targeting glucose have been disappointing, likely due to cancer cell plasticity and adaptation mechanisms (10). The complex adaptations induced in cancer cells by glucose restriction are not known. Here, we report the discovery that glucose restriction induces dedifferentiation and promotes aggression of lung adenocarcinoma (LUAD) cells.

Recent studies have highlighted how metabolic requirements evolve during cancer progression (11). We previously identified a metabolic vulnerability specific to early-stage LUAD, which relies on sodium-glucose transporter 2 (SGLT2) for growth and progression (12). Because early lesions of the LUAD spectrum rely on SGLT2 for glucose uptake, inhibition of this transporter with an antidiabetes drug, canagliflozin, significantly delays the development of LUAD, prolongs survival, and reduces tumor growth in genetically engineered murine models and patient-derived xenografts (12). Over the past decade, evidence has begun to emerge from epidemiologic studies showing that diabetic patients treated with SGLT2i have improved cancer outcomes (13, 14), confirming the favorable effects of these drugs against cancer demonstrated by our preclinical studies. However, the tumors eventually escape the treatment and undergo a dedifferentiation process, suggesting that glucose restriction induces a phenotypic switch in LUADs.

Glutamine restriction in the tumor core is known to cause cancer dedifferentiation, due to reduced alpha-ketoglutarate (αKG; ref. 15). αKG is a key modulator of cell differentiation both in normal development and in cancer (16). αKG availability has a direct impact on gene expression, because it is required for the activity of the Jumanji C domain (JMJD)-containing histone demethylases and the ten-eleven-translocation enzymes, involved in DNA demethylation (17). Therefore, αKG depletion leads to histone and DNA hypermethylation, with repression of differentiation-related genes.

Multiple histone demethylases depend on αKG for their enzymatic activity (18, 19). However, H3K27 trimethylation plays a major role in regulating glutamine restriction-dependent cell dedifferentiation (20). H3K27 is methylated by the polycomb repressor complex 2 through the activity of histone methyltransferase enhancer of zeste homolog 2 (EZH2), and it is demethylated by JMJD3 and UTX. Knockout of UTX promotes Kras-driven LUAD in an EZH2-dependent way (21), whereas EZH2 overexpression drives LUAD development independently of Kras (22), suggesting that this pathway is relevant in LUAD progression.

Here, we show that glucose deprivation in LUAD causes cancer dedifferentiation similar to that caused by glutamine restriction. Glucose deprivation reduces αKG availability, limiting αKG-dependent histone demethylase activity, and induces histone hypermethylation, driving LUAD to a poorly differentiated state and a highly aggressive phenotype. Surprisingly, we found that glucose restriction-induced dedifferentiation and increased aggressiveness require activation of the hypoxia-inducible factor 1α (HIF1α) signaling. EZH2 is involved in upregulation of HIF1α by direct repression of the proline hydroxylase PHD3, which initiates HIF1α degradation in normoxia. HIF1α induces Slug activity and epithelial-to-mesenchymal transition (EMT), leading to a highly aggressive and metastatic phenotype. Targeting the EZH2/HIF1α/Slug axis pharmacologically potentiates the effect of SGLT2 inhibitors in murine LUAD. Finally, we describe a transcriptional signature regulated by the hypoxia pathway whose expression in human LUAD confers a significantly worse prognosis.

Materials and Methods

Cell lines

Human A549 (RRID: CVCL_0023), NCI-H358, hereby called H358 (RRID: CVCL_1559), MCF7 (RRID: CVCL_0031), and PANC1 (RRID: CVCL_0480) cells were purchased from ATCC and maintained in Roswell Park Memorial Institute (RPMI) 1640 medium (Corning, #10-040CV) supplemented with 10% FBS and 5% penicillin–streptomycin. Cells were cultured in a humidified incubator at 37°C and 5% CO₂. Murine 2953A cells were established in our lab from a lung KP tumor by enzymatic tissue dissociation and maintained in DMEM (Corning, #10-017CV) supplemented with 10% FBS and 5% penicillin–streptomycin. Cells were cultured in a humidified incubator at 37°C and 5% CO₂. All cells were tested for Mycoplasma after being received from ATCC and used within 10 passages.

Patient-derived organoids culture

Patient-derived organoid (PDO) lines were established in our lab from fresh human lung tumor procured from surgical pathology after obtaining written informed consent. The collection of samples was conducted in accordance with recognized ethical guidelines and approved by the University of California Los Angeles (UCLA) Institutional Review Board (IRB #10-001096, PI: Steven Dubinett). The tissues were minced with surgical scissors into small (<0.5 mm) fragments and digested in lung dissociation mix (1,500 μg/mL Collagenase A, 100 μg/mL Collagenase Type IV, 100 μg/mL DNase I, 100 μg/mL Dispase II, 9.2 μg/mL Elastase, 1,250 μg/mL Pronase in Hank's Balanced Salt Solution) for 2 cycles of 10 minute incubation on orbital shaker followed by mechanic dissociation by pipetting. When the tissue was completely dissolved, single cells were filtered through a 70-μm cell strainer (Falcon). The enzymes were inactivated with FBS (final 3%). Pellets were resuspended first in ACK buffer (Gibco) for erythrocyte lysis and next in Advanced DMEM/F12 (1× Glutamax, 10 mmol/L HEPES, 1× Anti-Anti) to stop the reaction. Cells were counted and seeded at 75,000 cells per 50 μL of Cultrex Basement Membrane Extract (RGF BME; R&D Systems) and cultured in DMEM/F12 (Thermo Fisher 11320033) supplemented with R-Spondin 1 (PeproTech #120-38, 500 ng/mL), FGF7 (PeproTech #100-19, 25 ng/mL), FGF10 (PeproTech 100-26, 100 ng/mL), noggin (PeproTech 120-10C, 100 ng/mL), A83-01 (Tocris #2939, 500 mmol/L), Y-27632 (Abole #Y-27632, 5 μmol/L), SB202190 (Sigma #S7067, 500 nmol/L), B27 supplement (gibco #17504-44, 1×), N-acetyl-cysteine (Sigma #A9165-5g, 1.25 mmol/L), nicotinamide (Sigma #N0636, 5 mmol/L), glutamax (Invitrogen #12634-034, 1×), Hepes (Invitrogen #15630-056, 10 mmol/L), penicillin/streptomycin (Invitrogen #15140-122, 100 U/mL and 100 μg/mL, respectively), and primocin (Invivogen #Ant-pm-1, 50 μg/mL).

Mouse models

All experiments performed in mice were approved by the UCLA Institutional Animal Care and Use Committee and were carried out according to the guidelines of the Department of Laboratory Animal Medicine at UCLA. For our imaging and therapeutic trials in genetically engineered mouse models we used Kras^LSL/G12D (Kras^tm4Tyi), p53^fl/fl (Trp53^tm1Brn), and Rosa26^{LSL/luciferase} (ROSA26Sor^tm1(Luc)Kael) mice (KP_luc mice, back-crossed into FVB background) bred in our laboratory, as previously reported (12). The breeders were kindly provided by Dr. David Shackelford (UCLA).

For tail-vein injection assays we used syngeneic wild-type mice (FVB background). KPluc mice were bred in our colony at UCLA. Syngeneic wild-type mice were purchased from Jackson lab.

In vitro studies

All experiments in cell lines were performed in biological triplicate. Cells were seeded at the same confluence and grown for 5 days (and for some experiments overnight or up to 9 days or 30 days, as indicated in the figures and text) in different concentrations of glucose: high (20 mmol/L), low (1 mmol/L) and for some experiments physiologic (5 mmol/L) and murine physiologic (10 mmol/L), as indicated. We used mediums with glutamine and without glucose: RPMI (Corning, #10-043CV) for A549 and H358 cells, DMEM (Thermo F. Scientific, #11566025) for 2953A cells, supplemented with 10% FBS, 1× penicillin–streptomycin and complemented with D-glucose (Gibco, #A2494001) to the desired final concentration, as indicated in the figures. Breast cancer (MCF7) and pancreatic cancer (PANC1) cell lines were seeded and grown up to 5 days in RPMI-1640 complemented with D-glucose to a final concentration of either 20 mmol/L or 1 mmol/L. All αKG rescue experiments were conducted using 10 mmol/L dimethyl α-ketoglutarate (Sigma, #13192-04-6) and α-Mannitol (Sigma, #M9647) as osmotic control.

For the glucose measurement in the growth medium, A549 and H358 cells were incubated in culture media supplemented with either 20 mmol/L (high) or 1 mmol/L (low) D-glucose for different time points, up to 24 hours. Glucose concentration was measured in culture media using a Glucose Colorimetric Kit (Thermo F. Scientific, #EIAGLUC).

For the glutamine deprivation experiment, A549 and H358 cell lines were seeded and grown up to 5 days in DMEM without glutamine and glucose (Corning, #17-207CV) and complemented with L-glutamine (Gibco, #35050061) to a final concentration of either 6 mmol/L (high) or 2 mmol/L (low) and D-glucose to a final concentration of either 10 mmol/L (high) or 1 mmol/L (low).

PDO lines were cultured for 3 weeks in Advanced DMEM/F12 (Gibco, #12634010) without glucose (Thermo F. Scientific, #A2494301) complemented as described above for PDO culture, and with D-glucose to a final concentration of 20 mmol/L (high), 5 mmol/L (physiologic), or 1 mmol/L (low) glucose for RNA extraction and RT-PCR. α-Mannitol (Sigma, #M9647) was used as osmotic control in PDOs incubated in low glucose. High- and low-glucose culture media were replaced every two days.

In vivo studies

Lung tumors were induced by intranasal administration of AdenoCre (purchased from University of Iowa Viral Vector Core) as previously described (12), or by transthoracic injection of AdenoCre in KP_luc mice. All treatment trials were started 2 weeks after tumor induction, and the results of biological duplicates were pooled for statistical analysis. In all trials, mice were assigned to therapeutic groups so that there were no significant differences in baseline tumor burden, age, sex, or weight among the different groups.

In the trial for assessment of tumor differentiation by Western blot, lung tumors were induced by intranasal administration of AdenoCre. After induction, mice (n = 3 per group) were treated with empagliflozin (10 mg/kg/d by oral gavage) in 0.5% hypromellose (Sigma, #H7509) or vehicle for 6 weeks, followed by tumor collection and total protein extraction.

In the trial for assessment of tumor differentiation by RT-PCR, mice (n = 3) received intranasal AdenoCre and treatment with empagliflozin (10 mg/kg/d by oral gavage) in 0.5% hypromellose (Sigma, #H7509) or vehicle for 4 weeks, followed by tumor collection and RNA extraction. RT-PCR was performed in technical triplicates, yielding a total of 9 experimental points per group.

For the measurement of in vivo H3K27 trimethylation, mice (n = 4 per group) received transthoracic AdenoCre and were treated for 3 weeks with either placebo (0.5% hypromellose) or empagliflozin (10 mg/kg/d by oral gavage) followed by lung collection and histone extraction for ELISA assay. H3K27 trimethylation was also assessed by Western blot. After tumor induction, mice (n = 2 per group) were treated with empagliflozin (10 mg/kg/d by oral gavage) in 0.5% hypromellose (Sigma, #H7509) or vehicle for 6 weeks, followed by tumor collection. Tumor cells were sorted using CD326 MicroBeads (Miltenyi Biotec, #130-105-958). Histone extraction was performed on EpCam-positive tumor cells.

In the combination treatment trial, mice received the first bioluminescence imaging (BLI) measurement two weeks after AdenoCre administration for measurement of the baseline tumor burden using PerkinElmer IVIS Spectrum In Vivo Imaging System (RRID: SCR_020397). Four therapeutic groups were established as follows: (i) placebo (n = 11), receiving oral gavage of vehicle (DMSO 5%, PEG 30%, Tween80 10%); (ii) SGLT2 inhibitor empagliflozin (10 mg/kg/d, n = 13); (iii) EZH2 inhibitor tazemetostat (125 mg/kg/b.i.d., n = 10; MedChem Express, #HY-13803); and (iv) empagliflozin and tazemetosat (n = 19). Each group received bi-daily administration by oral gavage.

Tail-vein injection assays

For our studies on the investigation of cell aggressiveness due to glucose deprivation, we used syngeneic FVB mice. Two experiments were performed, in biological duplicate.

In the first experiment, we used the murine cell line, 2953A, previously established in our lab from a KP_luc tumor, which has a transgenic luciferase expression. 2953A cells were cultured in either high (20 mmol/L) or low (1 mmol/L) glucose for at least 1 month. In cells incubated in low glucose, α-mannitol was used as osmotic control. For low-glucose cells, to exclude clonal effects, we expanded three different clones of low glucose–resistant cells. The day before the inoculation, cells were transfected with either pooled siRNAs targeting HIF1α or control siRNA. Cells were trypsinized and resuspended in cold PBS, followed by tail-vein inoculation (250,000 cells/mouse) in syngeneic mice (n = 8). After 1 week, tumor burden was measured by bioluminescence performed on an IVIS Spectrum In Vivo Imaging System (PerkinElmer) 10 minutes after intraperitoneal injection of luciferin (150 mg/kg). Bioluminescence acquisitions were analyzed by Living Image software (RRID: SCR_014247).

In the second experiment, KP_luc mice injected transthoracically with AdenoCre were treated for 3 weeks with either placebo (n = 3) or empagliflozin (n = 3) and sacrificed for lung collection. Tumors were identified in the murine lungs by ex vivo bioluminescence. Cell suspension was obtained by enzymatic dissociation and homogenized using a 40-μm sterile cell strainer (Biologix, #15-1040). Tumor single cells were resuspended in cold PBS followed by sorting using CD326 MicroBeads (Miltenyi Biotec, #130-105-958) to isolate the EpCam-positive epithelial cells. Finally, sorted tumor cells were injected by tail-vein inoculation (1,000,000 cells/mouse) into syngeneic mice (n = 6 per group). Tumor burden and number of tumors were measured by bioluminescence at 13 weeks after inoculation. For each mouse, one whole-lung section containing all 5 lung lobes was used for tumor quantification. Tumor number and area were quantified with ImageJ.

IHC staining

Mouse lungs were collected and inflated with 10% formalin in phosphate-buffered saline. After 24 hours, formalin was replaced with 70% ethanol. All tissues were paraffin-embedded and sliced into 4-μm sections in the Translational Pathology Core Laboratory at UCLA.

For IHC staining, the slides were deparaffinized by overnight incubation at 65°C, followed by rehydration by serial passages in xylenes (three washes of 5 minutes in 100% xylenes) and decreasing concentrations of ethanol (two washes in 100% ethanol, two washes in 95%, one wash in 80%, one wash in 70%, and one wash in water). Antigen retrieval was performed for 23 minutes in either AR6 (pH 6.0) buffer (Akoya, #AR600250ML) or AR9 (pH 9.0) buffer (Akoya, #AR900250ML). Blocking was performed with 2.5% normal horse serum for 30 minutes at room temperature, followed by incubation with primary antibodies for 45 minutes at room temperature for mouse antibodies and overnight at 4°C for rabbit antibodies. Incubation with secondary antibody (ImmPRESS Peroxidase Polymer Anti-Rabbit IgG Reagent, #MP-7401; ImmPRESS Peroxidase Polymer Anti-Mouse IgG, #MP-7422) was performed at room temperature for 30 minutes, followed by incubation with ImmPACT 3,3′-diaminobenzidine (DAB; Vector Lab, #SK-4103-100) under the previously established conditions. Counterstain was performed with Harris’ hematoxylin (New Commer Supply, #1201A) diluted 1:5 in water, followed by a rehydration step and mounting slide. After the staining, digital images of the slides were obtained with an Aperio ScanScope (RRID: SCR_006355) slide scanner (Leica Biosystems; RRID: SCR_018041).

The quantification was performed using the QuPath software (RRID: SCR_018257). Briefly, regions of interest were drawn blindly in each slide to include each tumor present in a whole-lung section, followed by cell detection and quantification of the DAB signal using a constant threshold for all samples in the same experiment. Staining intensity in each cell was classified as negative (0), weak (1), median (2), or strong (3). The results were expressed as H score. Antibodies were purchased from Cell Signaling Technology (CST; see Supplementary Table S1).

Metabolomics assay

A549 and H358 cells were cultured overnight in a medium complemented with either high (20 mmol/L) or low glucose (1 mmol/L). For each condition, we prepared 3 plates for LC-MS analysis. Then, cells were rinsed with ice-cold 150 mmol/L NH4AcO at pH 7 and incubated with precooled 80% methanol at −80°C for 60 minutes. Next, cells were scraped, transferred into a new tube on ice, and spun down at 16,000 × g for 15 minutes at 4°C. Afterward, the supernatants, containing the metabolites extracted, were transferred in a glass vial and then they were dried down at 30°C in an evaporator (Genevac EZ-2 Elite, RRID: SCR_017367). The pellets, which are still in the tubes on ice, were resuspended with 3 volumes of RIPA buffer, and protein concentration was determined by BCA Protein Assay (Thermo F. Scientific, #23225).

LC-MS was performed by UCLA Metabolomic Center by using the Q Exactive HF Hybrid Quadrupole-Orbitrap Mass Spectrometer (RRID: SCR_020425) coupled to a Thermo Scientific Dionex Ultimate 3000 system (RRID: SCR_020563) chromatography systems. Samples were normalized by protein content. Data collected was processed with Thermo Scientific TraceFinder 4.1 (RRID: SCR_023045). Histograms were generated based on normalized relative amounts of each replicate of both groups. Statistical analysis was performed by analysis of variance (ANOVA).

Small interfering RNA transfection

For small interfering RNA (siRNA) knockdown of human EZH2, HIF1α, HIF2α, PHD3, and HIFAL cells were cultured in either high (20 mmol/L) or low glucose (1 mmol/L) and transfected with 25 pmol/L of siRNAs for each target and control siRNA using Lipofectamine RNAi Max reagent (Thermo F. Scientific, #13778075).

Because the total incubation time in low glucose was 5 days, two separate transfections were performed on day 1 before starting the high vs. low-glucose incubation, and at day 3. On day 5, cells were collected, followed by total protein extraction or RNA extraction. Each siRNA was used individually or pooled. Each pooled siRNA was transfected either as a single pool or in combination with another pooled siRNA.

For the long-term (30 days) exposure to low glucose, cells were incubated long-term without transfection, and siRNAs were transfected 5 days and retransfected 3 days before harvesting.

For tail-vein injection, cell clones isolated from long-term low-glucose culture were transfected with a pool of two siRNAs targeting HIF1α or control the day before tail-vein injection. We confirmed the knockdown efficiency of mouse HIF1α knockout 3 and 7 days after RNA transfection.

All siRNAs were purchased from QIAGEN and Dharmacon (see Supplementary Table S2).

Plasmid transfection

The A549 cell line was transfected with 10 μg of EZH2 (pCMVHA hEZH2, RRID: Addgene_24230) and PHD3 (plp6.3-Egln3, RRID: Addgene_79115) plasmid using Lipofectamine 3000 (Thermo F. Scientific, #L3000015). Each plasmid was transfected either individually or in combination. After three days, cells were harvested, lysed with RIPA buffer, and analyzed by SDS-PAGE and Western blotting.

ELISA

The levels of trimethylation on H3K27me3 in tumors treated with empagliflozin were detected by ELISA kits (Active Motif, #53106) according to the manufacturer's instructions. The experiment was carried out with 4 replicates for each group. After tumor collection, histone extraction was performed.

We used H3K27me3 and H3 total (Active Motif, #53110) antibodies for plate coating. The absorbance was detected by Varioskan Lux (Thermo F. Scientific, #VL0000D0). The H3K27me3 levels were normalized to H3 total and presented as optical densitometry (OD) 450 nm of a ratio of H3K27me3/H3.

Total protein extraction

For total protein extraction, cells were harvested, washed twice with ice-cold PBS-EDTA (0.5 mm EDTA), lysed using RIPA buffer (50 mmol/L Tris-HCl pH 7.6, 150 mmol/L NaCl, 0.1% SDS, 0.5% C24H39NaO4, 1% NP-40, 2 mm EDTA, 50 mm NaF) for 15 minutes on ice and centrifuged at 13,000 rpm for 30 minutes at +4°C.

The resulting protein extracts were quantified using BCA Protein Assay (Thermo F. Scientific, #23225) followed by analysis by SDS-PAGE and Western blotting.

Histone extraction

Histones were extracted from A549, H358, and 2953A cell lines using the Histone extraction kit (Abcam, #ab113476) according to the manufacturer's instructions.

Western blotting

SDS-PAGE and Western blotting analyses were performed using standard protocols. For antibody information, see Supplementary Table S1. Western blot images were detected by iBright Imaging Systems (RRID: SCR_017632). The protein detected was normalized to Actin or H3 total antibodies, the latter in the case of histone extract. Each immunodetection derived from the same membrane was performed with the same exposure times according to the manufacturer's antibody guidelines and was cropped only for presentation purposes; this is indicated by a dotted line. Antibodies were purchased from CST, Novus Biologicals, and GeneTex.

RNA extraction

Total RNA was extracted from A549, H358, and 2953A by using TRI Reagent Solution (Applied Biosystem, #AM9738), according to the manufacturer's instruction. RNA concentration was assayed by NanoDrop 3300 Fluorospectrometer (RRID: SCR_015804). Then, 1 μg of RNA was treated with DNase I (Thermo F. Scientific, #MAN0012000) and used for cDNA preparation.

RT-qPCR

cDNA was prepared using 1 μg of RNA with SensiFast RT Kit (Meridian Biosciences, #BIO-65053). The SYBR green-based RT-PCR kit (Applied Biosystem, #A25742) was performed using human and mouse primers (see Supplementary Table S3). mRNA levels were normalized to GAPDH (ΔCt = Ctgene of interest−Ct GAPDH) and presented as relative mRNA expression (2^ΔC_t). All primers were designed by NCBI Primer-BLAST (RRID: SCR_003095) and purchased from Integrated DNA Technologies.

RNA-seq and data analysis

A549 and H358 cell lines were incubated in triplicate with high glucose (20 mmol/L), low glucose (1 mmol/L), and low glucose + dm-αKG (10 mmol/L). Total RNA was extracted as previously described. RNA-seq was performed by Med Genome. Sample quality control was performed using Thermo Fisher Qubit 2.0 Fluorometer (RRID:SCR_020553) and Agilent 4150 TapeStation System (RRID:SCR_019394). For library preparation, TruSeq Stranded Total RNA kit (Illumina, #20020597) was used, and libraries were sequenced on Illumina NovaSeq 6000 Sequencing System (RRID:SCR_016387). The FASTQ data generated were used for gene-expression analysis, performed as described by Nassa and colleagues (23). Briefly, the raw sequence files generated (.fastq files) underwent quality control analysis using FastQC (RRID: SCR_014583) and adapter sequences were removed using Trimmomatic v0.38 (RRID: SCR_011848; ref. 24). Filtered reads were aligned on the human genome (assembly hg38) considering genes present in GenCode Release 35 (GRCh38.p12) using STAR v2.7.6a (RRID: SCR_004463; ref. 25) with standard parameters. Quantification of expressed genes was performed using featureCounts (RRID: SCR_012919; ref. 26) and differentially expressed genes were identified using DESeq2 (RRID: SCR_015687). A given RNA was considered expressed when detected by at least ≥10 raw reads.

Differential expression was reported as |fold change| (FC) ≥ 1.5 along with associated adjusted P ≤ 0.05 computed according to Benjamini–Hochberg. Gene set enrichment analysis (GSEA; RRID:SCR_003199) was performed to examine pathway enrichment for the differential expressed genes with the Molecular Signature Database “Hallmarks” gene set collection (27). Only those with an FDR ≤0.25 have been selected. Functional analysis was also performed with Ingenuity Pathway Analysis (IPA; QIAGEN, RRID: SCR_008653). Only pathways with a P value ≤ 0.05 were considered for further analysis.

Chromatin immunoprecipitation sequencing and data analysis

A549 cell lines were incubated in triplicate with either high glucose (20 mmol/L) or low glucose (1 mmol/L) for 5 days. A total of 15 ×10⁶ cells were fixed, lysed to isolate nuclei, sonicated, and diluted as described by Nassa and colleagues (23). An aliquot of nuclear extract was taken as input to be used as a control for sequencing and data analysis. For H3K27me3 and H3K4me3 pulldown, 50 μL of equilibrated Dynabeads M-280 Sheep Anti-Rabbit IgG (Thermo F. Scientific, #11203D) were incubated overnight at 4°C with 10 μg of the antibodies chosen for immunoprecipitations (see Supplementary Table S1). Bead washing, elution, reverse cross-linking, and DNA extraction were then performed as described by Tarallo and colleagues (28). Size distribution of each chromatin immunoprecipitation (ChIP) DNA sample was assessed by running a 1 μL aliquot on an Agilent High Sensitivity DNA chip using an Agilent 4150 TapeStation System (RRID:SCR_019394). The concentration of each DNA sample was determined by using Quant-IT DNA Assay Kit-High Sensitivity (Thermo F. Scientific, #Q32851) and Thermo Fisher Qubit 2.0 Fluorometer (RRID:SCR_020553). Purified ChIP and input DNAs (6 ng each) were used as the starting material for sequencing library preparation by using the TruSeq ChIP Sample Prep Kit (Illumina, #IP-202-1012) and were sequenced (single read, 1 × 75 cycles) on an Illumina NextSeq 550 System (RRID: SCR_016381).

Quality control of the sequenced reads was performed using FastQC (RRID: SCR_014583). Reads were aligned to the reference genome assembly (hg38) using Bowtie (RRID:SCR_005476; ref. 29) allowing up to one mismatch and considering uniquely mapped reads. Signal artifact blacklist regions were filtered out using BEDtools (RRID: SCR_006646).

For each biological replicate, peak calling was performed using HOMER (RRID: SCR_010881; ref. 30) setting the following parameter: -P 0.01 –F 2.5 –style histone. Only peaks common to at least two replicates were considered for further analysis. Annotation of peaks to the nearest gene was performed using the annotatePeaks.pl function of HOMER, whereas annotation of peaks to cis-Regulatory Elements was performed with BEDtools, using the track available in SCREEN (31).

To find differentially regulated features between low glucose versus high glucose the getDifferentialExpression.pl function of HOMER was used, setting the parameter “–edgeR.” Only features showing an adjusted P ≤ 0.05 and |FC| ≥ 1.3.

Prediction of potential transcription factor binding sites on selected genes associated with histone peak was performed using CiiiDER (32).

ChIP-qPCR

The A549 cell line was incubated in either high (20 mmol/L) or low (1 mmol/L) glucose for 5 days. Chromatin was isolated as described previously starting from 15 × 10⁶ cells. Before immunoprecipitation, an aliquot of chromatin extract was taken as input to be used as control of qPCR. ChIP was carried out by overnight incubation of chromatin at 4°C with 50 μL of Dynabeads Protein G (Thermo F. Scientific, #10003D), precoated with 5 μg of anti-EZH2 (CST, #5246). Beads washing steps, DNA elution, and extraction were performed as previously described. Before DNA elution, an aliquot of beads for each condition was conserved for Western blot assay, resuspended in sample buffer, and boiled at 90°C for 5 minutes.

0.2 ng of DNA was used to amplify PHD3, MYT-1, and GAPDH promoter regions (see Supplementary Table S3).

Data analyses were presented as percentages of input. The experiment was repeated a second time, and the data were pooled as a biological replicate.

Quantification and statistical analysis

GraphPad Prism 8.0 software (RRID: SCR_002798) for statistical analysis. Analysis for significance was performed by parametric or nonparametric Student t test when only two groups were compared and by one-way ANOVA when three or more groups were compared.

For the treatment trials, in order to compare tumor volume (log scale) between groups (high glucose vs. low glucose, control siRNA vs. HIF1α siRNA), we ran a general linear model with terms for group, experiment, and group × experiment interaction. We then extracted the pairwise contrast from the model to compare the groups with 95% confidence intervals. P values < 0.05 were considered statistically significant, and all analyses were run using IBM SPSS Statistics (RRID: SCR_019096).

QuPath (RRID: SCR_018257) was used for signal quantification in IHC staining. It was expressed as H score. Kaplan–Meier curves were performed with Kaplan–Meier Plotter (RRID: SCR_018753), using the lung cancer section. For the selected genes only, the best probe was used. All data set available were used as cohorts. For Kaplan–Meier plot performed using multiple genes, the option “use mean expression of selected genes” was set.

Data, code, and materials availability

The RNA-seq raw data are publicly available in ArrayExpress (RRID: SCR_002964) repository under accession number: E-MTAB-11253.The ChIP sequencing (ChIP-seq) raw data are publicly available in ArrayExpress (RRID: SCR_002964) repository under accession number: E-MTAB-11678. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

All animal models and cell lines developed in our lab (2953A and glucose-restricted clones) will be made available upon request to the corresponding author.

Results

Glucose restriction induces dedifferentiation in LUAD

We previously showed that pharmacologic inhibition of SGLT2 in murine LUAD results in the reduction of the tumor burden and improvement of survival (12). However, mice eventually develop aggressive tumors and die even during the treatment. To find out the mechanisms of resistance, we used our previously developed conditional genetically engineered model driven by Kras^G12D mutation and p53 deletion (KP_luc mice; ref. 12). In this model, lung tumors are induced by intranasal administration of adenovirus encoding the Cre recombinase (AdenoCre). Two weeks after tumor induction, we treated the KP_luc mice with empagliflozin, a selective SGLT2 inhibitor, and we collected the lungs at week 8 after AdenoCre inhalation for histology. Visual inspection of the hematoxylin and eosin staining of lung sections showed that the lesions in the SGLT2-treated mice were more poorly differentiated than those in the placebo group. The KP_luc tumors recapitulated a heterogeneous morphology typical of human LUAD, with coexistence of different components: papillary, acinar, solid, and micropapillary (Fig. 1A). Pathologic quantification performed by a board-certified pathologist (Supplementary Table S4) showed that empagliflozin changed the differentiation state of the tumors, significantly reducing the solid component and increasing the micropapillary component of murine tumors (Fig. 1B). The micropapillary component is an independent predictor of poor prognosis in LUAD (33). According to the most recent grading system of the International Association for the Study of Lung Cancer, invasive pulmonary adenocarcinoma are graded based on the histologic pattern: grade 1 (well-differentiated) tumors are lepidic predominant with ≤ 20% high-grade patterns; grade 2 (moderately differentiated) tumors are acinar or papillary predominant with ≤20% high-grade patterns; grade 3 (poorly differentiated) tumors have ≥ 20% of high-grade pattern. High-grade patterns are solid, micropapillary, and complex glandular (34). Based on this classification, the treatment with empagliflozin caused an increase in high-grade tumors from 39.6% to 50.9% (Supplementary Table S4). In addition, some of the tumors in the empagliflozin group had a clear cell component, which was completely absent from placebo tumors (Fig. 1A; Supplementary Table S4). Clear cell carcinoma is a very rare variety of LUAD with uncertain prognostic significance (35). The pathologic assessment suggested that empagliflozin treatment changed the differentiation state of the tumors.

Figure 1.

SGLT2 inhibition causes tumor dedifferentiation in LUAD. A–F, KP_luc mice carrying LUADs were treated with either placebo or empagliflozin (10 mg/kg/d) from week 2 to week 8 after tumor induction. A, Representative images of the histologic patterns identified. Scale bar, 100 μm. B, Quantification of the percentage of different histologic types in hematoxylin and eosin slides from the two treatment groups, performed by a board-certified pathologist. C, Western blot on whole tumor extracts from three tumors for each group. D, KP_luc mice were treated for 4 weeks with either placebo or empagliflozin, followed by tumor collection and RNA extraction. RT-PCR was performed to measure the changes in type II alveolar cell marker SpC and tumor differentiation markers FoxA2, Ttf1, and HmgA2. Significance was evaluated by Student t test. E, PDOs were established from fresh surgical specimens of lung adenocarcinoma. All organoids were incubated in high (20 mmol/L), physiologic (5 mmol/L), or low (1 mmol/L) glucose for three weeks, followed by RT-PCR. Significance was evaluated with Student t test. F–H, A549 and H358 cells were incubated in different concentrations of glucose, as indicated, followed by Western blot (F and G) or RT-PCR (H). G, Quantification of the Western blot signal was performed with ImageJ. H, The correlation between RT-PCR signal and glucose concentration was evaluated by Pearson method for FOXA2 and TTF1, and by Spearman method for HMGA2 and GLUT1. *, P < 0.05; **, P < 0.01; ***, P < 0.001; ****, P < 0.0001; n.s., not significant.

To quantify more accurately the effect of glucose restriction on tumor differentiation, we performed a Western blot on the KP_luc murine tumors with antibodies against FoxA2 and Ttf-1, two major markers of differentiation and regulators of cell state transitions in LUAD (36, 37). The tumors treated with empagliflozin showed reduced expression of Ttf-1 and FoxA2 (Fig. 1C). To confirm this phenotypic switch, we repeated empagliflozin treatment, collecting the tumors at week 6 after tumor induction, the first time point in which mice develop macroscopically dissectible tumors, for RT-qPCR analysis. We observed a significant reduction in surfactant protein C (SpC), a marker of type II alveolar cells (ATII), which are the cell of origin of LUAD (Fig. 1D). We also observed reduced expression of markers of club cells (Scgb3A2), type I alveolar cells (Pdpn), and ciliated cells (FoxJ1), although the basal levels of these markers were two orders of magnitude lower than that of SpC (Supplementary Fig. S1A). We also observed a significant reduction of differentiation markers FoxA2 and Ttf1 and, conversely, an increase of dedifferentiation marker HmgA2 in tumors treated with empagliflozin compared with placebo (Fig. 1D). As the tumors have a conditional luciferase transgene, we validated luciferase expression (Supplementary Fig. S1B), confirming the proper activation of the conditional model in the collected tumor tissue and showing no difference between the two groups.

To confirm the effect of glucose deprivation on cell differentiation in a more clinically relevant model, we established PDOs from 3 surgical samples of early-stage lung cancer. The pathologic features of the tumors are summarized in Supplementary Table S4. To explore the differences in differentiation within a wide range of glucose exposure, we cultured the newly established PDOs in high (20 mmol/L), physiologic (5 mmol/L), or low (1 mmol/L) glucose for three weeks, with the addition of mannitol as osmolarity control, and extracted RNA for RT-PCR. This experiment showed that markers of well-differentiated LUAD, FOXA2 and TTF1, were progressively and significantly downregulated in all three PDOs as the glucose concentration in the medium was lowered from 20 to 5 and to 1 mmol/L, whereas poor differentiation markers HMGA2 and GLUT1 were upregulated in all three PDOs as the glucose concentration was lowered (Fig. 1E). We have previously shown that GLUT1 is expressed in poorly differentiated LUADs (12). GLUT1 can be upregulated by glucose restriction in an αKG- and NF-kB-dependent way (38). GLUT1 upregulation can provide an alternative mechanism of glucose supply in cancer cells, conferring resistance to SGLT2 inhibitors.

To investigate more in detail the effects of glucose deprivation on cell differentiation, we performed in vitro experiments in two human cell lines, A549 and H358, which have been described as moderately differentiated (39–41) and express high levels of FoxA2, as well as a murine cell line, 2953A, which was established in our lab from a KP_luc lung tumor. To find out the optimal glucose concentration to use in vitro, we measured glucose concentration in the medium in A549 and H358 cells at different time points. We observed that if we incubated the cells in 20 mmol/L glucose, which is a commonly used concentration for cell culture, sugar consumption by cancer cells rapidly reduced the concentration of glucose in the medium to physiologic concentrations of 6 to 8 mmol/L within 4 hours, whereas cells incubated in 5 mmol/L glucose maintained the concentration constant for 12 hours, followed by a sharp reduction of glucose in the medium (Supplementary Fig. S1C). Therefore, we decided to compare 20 mmol/L (high glucose), 5 mmol/L (physiologic glucose), and 1 mmol/L (low glucose), changing medium every 12 hours to maintain the physiologic concentration of 5 mmol/L constant. Measurement of glucose in the blood and tumor interstitial fluid of mice showed that blood glucose levels averaged 13 mmol/L, whereas glucose concentration in the tumor interstitial fluid was around 1 mmol/L (Supplementary Fig. S1D), confirming that the low-glucose concentration of 1 mmol/L mimics realistically the concentration of glucose in an in vivo setting. Western blot analysis showed a progressive and significant reduction of differentiation markers FOXA2 in both cell lines and of TTF1 in H358 cells (this marker is not expressed in A549 cells) as the glucose concentration was lowered from 20 to 5 to 1 mmol/L (Fig. 1F–G).

We also performed RT-PCR for markers of differentiation and hypoxia with 4 different glucose concentrations: 20, 10, 5, and 1 mmol/L (Fig. 1H). The RT-PCR analysis showed a significant correlation between differentiation markers and glucose concentration. Differentiation markers FOXA2 and TTF1 exhibited a linear positive correlation with glucose concentration, whereas dedifferentiation markers HMGA2 and GLUT1 were exponentially negatively correlated with glucose concentration in the medium (Fig. 1H). Taken together, these results suggest that glucose restriction, induced by SGLT2 inhibition or by incubation in low glucose culture, leads the tumors toward a poorly differentiated phenotype.

Glucose restriction leads dedifferentiation in LUAD cell lines and drives a more aggressive phenotype

To investigate the relationship between glucose metabolism and cell differentiation, we performed liquid chromatography–mass spectrometry (LC-MS) analysis to detect changes in metabolites caused by incubation of A549 and H358 cells overnight in low glucose. We focused specifically on metabolites involved in glycolysis and the TCA cycle. As expected, low glucose significantly reduced the levels of glycolytic and TCA cycle metabolites in both A549 and H358 cells, except for Acetyl-CoA, which showed no significant changes. Representative metabolites of the two pathways are presented in Fig. 2A (A549 cells) and Supplementary Fig. S1E (H358 cells). A list of all the metabolites is reported in Supplementary Table S5.

Figure 2.

Glucose restriction causes LUAD dedifferentiation, due to low αKG and histone hypermethylation, and increases cell aggressiveness. A, A549 cells were incubated in high (20 mmol/L) or low (1 mmol/L) glucose for 24 hours, and mass spectrometry analysis was performed. The results are expressed as metabolite amounts (AUC, area under the curve) normalized by total protein amounts. The statistical significance of the observed changes was evaluated by one-way ANOVA. B and C, A549, H358, and 2953A cells were cultured for 5 days in a medium containing high or low glucose with or without αKG, as indicated. B, Western blot analysis of FoxA2 and TTF1 (only for H358, because the other cell lines did not express this marker). β-Actin was used as a loading control. C, RT-PCR analysis of expression of FOXA2 and GLUT1 in H358 cells. GAPDH was used as a housekeeping gene. D–I, Murine LUAD cells 2953A were incubated in either high (20 mmol/L) or low (1 mmol/L) glucose for at least one month (D). Three clones from cells incubated in low glucose were picked and cultured separately. Both high-glucose and low-glucose cells were inoculated in syngeneic mice by tail-vein injection to measure the development of lung metastases. The tumor burden was measured by BLI. Both representative pictures of single mice (E) and quantification of the signal (F) are reported. Tumor areas (G) and number of tumors (H) were measured in histologic sections of murine lungs stained with hematoxylin and eosin. One whole slide with 5 lung lobes was analyzed per mouse. I, Representative pictures of the hematoxylin and eosin stains. J–P, KP_luc mice carrying LUADs were treated with either placebo or empagliflozin (10 mg/kg/d), starting 2 weeks after tumor induction by transthoracic injection of AdenoCre. At week 5 after tumor induction, mice (n = 3) were sacrificed for tumor isolation, single-cell dissociation, sorting of epithelial cells, and i.v. injection in syngeneic mice (J). Tumor burden was estimated by BLI for 13 weeks after reinjection. K, Quantification of ex vivo BLI at week 5 after the end of the treatment trial. L, Quantification of the in vivo BLI at week 13 after reinjection in syngeneic mice. M, Representative images of BLI at week 13. N–P, Tumor areas (N) and number of tumors (O) were measured in histologic sections of murine lungs stained with hematoxylin and eosin (P). Significance was evaluated by Student t test. *, P < 0.05; **, P < 0.01; ***, P < 0.001; ****, P < 0.0001; n.s., not significant.

To test the effect of glucose depletion on LUAD differentiation markers, we performed a time-course experiment by growing A549 and H358 cells in low glucose (1 mmol/L) for up to 9 days. We observed that glucose restriction downregulated the expression of differentiation marker FoxA2 (Supplementary Fig. S1F) and upregulated the expression of poor differentiation marker HmgA2 (Supplementary Fig. S1G) after 5 days of treatment, confirming that glucose restriction induces LUAD dedifferentiation.

αKG plays a fundamental role in cell differentiation because it is a cofactor of histone and DNA demethylases (17). Depletion of glutamine causes cancer dedifferentiation due to reduced αKG (20), inhibition of αKG-dependent demethylases and DNA/histone hypermethylation. To test if the reduction of FOXA2 and TTF1 in low glucose was due to αKG depletion, we incubated A549, H358, and 2953A cells in low glucose for 5 days, with or without addition of the cell membrane-permeable precursor dimethyl-αKG (dm-αKG, 10 mmol/L), which gets converted to αKG in the cytosol (42). We used mannitol as an osmolarity control in low-glucose culture conditions. As expected, incubation in low glucose caused a consistent reduction in FOXA2 and TTF1, rescued by dm-αKG, whereas addition of αKG in high glucose did not affect FoxA2 expression, excluding an off-target effect (Fig. 2B). Analysis of mRNA levels by qPCR showed significant inhibition of TTF1 in H358 cells (Supplementary Fig. S1H) and of FoxA2 in A549, H358 (Fig. 2C), and 2953A cells (Supplementary Fig. S1I), rescued by dm-αKG. In addition, low glucose increased GLUT1 expression, and dm-αKG reduced its levels (Fig 2C; Supplementary Fig. S1I).

Because poorly differentiated tumors are typically more aggressive than well-differentiated lesions, we investigated whether glucose deprivation caused a shift toward a more aggressive phenotype in LUAD. To this aim, we incubated the 2953A cell line, which has transgenic luciferase expression, for one month in either high (20 mmol/L) or low (1 mmol/L) glucose. We expanded three clones of low-glucose growing cells, followed by tail-vein inoculation in syngeneic mice (Fig. 2D). One week after injection, we detected that the high-glucose cells developed a very small tumor burden as estimated by BLI, whereas the low-glucose cells showed significantly higher BLI signal (Fig. 2E and F). Histologic analysis of lung sections from the mice sacrificed one week after i.v. injection showed that the tumors formed by glucose-deprived cells had significantly larger areas than those formed by cells grown in high glucose, whereas the number of tumors was not significantly different in the high vs. low glucose cells (Fig. 2G–I). To test if in vivo SGLT2 inhibition can also increase the aggressiveness of cancer cells, we treated KP_luc mice with empagliflozin from week 2 to week 5 after tumor induction by transthoracic injection of AdenoCre, followed by tumor collection, dissociation into single cells, sorting of epithelial cells, and reinoculation in syngeneic mice by tail-vein injection (Fig. 2J). Ex vivo BLI showed that treatment with empagliflozin significantly reduced the tumor burden at week 5 (Fig. 2K); however, when the same number of placebo and empagliflozin-treated cells were injected in syngeneic mice, the empagliflozin-treated cells formed larger tumors than the placebo group (Fig. 2L–M). Similarly to what was observed for the cells incubated in vitro in low vs. high glucose, cells that had been previously treated with empagliflozin formed significantly larger tumors than the placebo cells, whereas the number of tumors was not significantly different between the two groups (Fig. 2N–P). These data showed that glucose starvation, while slowing tumor growth for reduced proliferation (12), eventually induces a more aggressive phenotype in LUAD cells.

H3K27 trimethylation is required for cancer dedifferentiation in response to glucose deprivation

Previous studies have shown that glutamine deprivation causes cancer dedifferentiation due to unbalanced histone methylation (20, 43). To assess the effect of glucose deprivation on histone modification, we incubated A549, H358, and 2953A cells in high (20 mmol/L), physiologic (5 mmol/L), or low (1 mmol/L) glucose for 5 days, followed by histone extraction and Western blot for histone marks. We observed a progressive increase of trimethylation of H3K4, H3K9, and H3K27 as the glucose concentration was reduced in all three cell lines. Conversely, H3K27 acetylation was progressively reduced by low glucose (Supplementary Fig S2A). Interestingly, administration of dm-αKG was able to reverse the H3K4, H3K9, and H3K27 hypermethylation and the H3K27 hypoacetylation (Fig. 3A). Because H3K27me3 has a major role in regulating glutamine restriction-dependent cancer dedifferentiation (20), we focused on this histone mark. To confirm that glucose restriction caused H3K27 hypermethylation in LUAD cells in vivo, we induced tumors by transthoracic injection of AdenoCre and treated KP_luc mice with empagliflozin for 1 week, followed by collection of tumors and ELISA assay for H3K27me3. We observed that SGLT2 inhibition caused a significant increase of trimethylation on H3K27 compared with the placebo group (Fig. 3B). To measure this histone mark in cancer cells, we collected tumors from mice treated with either placebo or empagliflozin from week 2 to week 8 after tumor induction, followed by sorting for EpCam-positive epithelial cells and Western blot. This test showed increased H3K27me3 mark in empagliflozin-treated than placebo tumor cells (Fig. 3C). Because EZH2 is the only known methyltransferase targeting H3K27, these results suggest that glucose restriction causes αKG reduction and deficient activity of αKG-dependent histone demethylases, resulting in an unbalanced activity of methyltransferase EZH2 and hypermethylation on H3K27.

Figure 3.

EZH2 is required for cell dedifferentiation induced by glucose restriction. A, A549, H358, and 2953A cells were cultured for 5 days in a medium containing high or low glucose with or without dm-αKG, as indicated. Histone marks were analyzed on histone extracts by Western blot, as indicated. Total histone 3 (H3) was used as a loading control. B, KP_luc mice were treated with empagliflozin (10 mg/kg/d) for 4 weeks starting at week 2 after tumor induction. The tumors were collected and used for ELISA assay for quantification of H3K27me3. The results are expressed as a ratio of H3K27me3 over total histone 3. C, KP_luc mice were treated with empagliflozin from week 2 to week 8 after tumor induction, then tumors were collected and tumor cells were isolated by tissue dissociation and magnetic sorting for EpCam-positive cells. Western blot was performed on histone extracts from the sorted tumor cells. D–F, A549 and H358 cells were cultured for 5 days in RPMI containing either high (20 mmol/L) or low glucose (1 mmol/L), with or without transfection of siRNAs targeting EZH2 (D and E) or treatment with EZH2 inhibitor GSK126 (F). D, Western blot for FOXA2 and EZH2 on whole-cell extracts and H3K27me3 on histone extracts. Actin was used as the loading control for the whole-cell extracts and H3 for the histone extracts. E, RT-PCR for FOXA2 and GLUT1 in A549 and H358 cells. Significance was evaluated by Student t test. F, Western blot for FOXA2 and H3K27me3 in A549 cells treated with GSK126. G–J, KP_luc mice were enrolled in a treatment trial with four groups: (i) placebo; (ii) empagliflozin (10 mg/kg/d by oral gavage); (iii) tazemetostat (125 mg/kg/b.i.d. by oral gavage); (iv) empagliflozin + tazemetostat. Treatment was carried on from week 2 to week 8 after tumor induction. Tumor burden was estimated by bioluminescence imaging. Representative images (G) and quantification in all the groups with pooled two biological replicates (H) are reported. Significance was measured using generalized estimating equation models with terms for time, group, and time-by-group interaction. I, Representative pictures of IHC stains for FoxA2 and Ttf1 in the tumors in the four treatment groups. J, Quantification of the IHC stain by QPath analysis. Multiple tumors (all tumors found in a whole-lung slide) were measured per mouse. Significance was evaluated by Student t test. Error bars, mean ± SEM. *, P < 0.05; **, P < 0.01; ***, P < 0.001; ****, P < 0.0001; n.s., not significant.

To confirm the role of EZH2 in glucose restriction-induced dedifferentiation, we used siRNAs. We first tested the efficiency of two different siRNAs targeting EZH2 separately by RT-PCR (Supplementary Fig. S2B). We incubated A549 and H358 cell lines in either high or low glucose and transfected the cells with either siRNAs targeting EZH2 or control siRNA, followed by protein extraction and Western blot analysis. As expected, incubation in low glucose reduced the expression of FoxA2 and H3K27me3 mark in both cell lines, and EZH2 knockdown rescued FoxA2 expression and H3K27 methylation in low glucose, whereas EZH2 knockdown did not have any effect in the cells incubated in high glucose (A549: Fig. 3D; H358: Supplementary Fig. S2C). RT-PCR analysis confirmed that low glucose incubation reduced expression of FoxA2 and increased the expression of Glut1 in both cell lines, and both were rescued by EZH2 knockdown, whereas the siRNAs targeting EZH2 had no effect on the basal levels in high glucose, excluding an off-target effect (Fig. 3E). Finally, we incubated the two cell lines in high or low glucose, with additional treatment with GSK126 (AZZOTA Scientific LLC, #C-1158), a selective EZH2 inhibitor (44). We tested different concentrations (Supplementary Fig. S2D) and we observed that a concentration of 0.25 μmol/L inhibited EZH2 activity and rescued the expression of FoxA2 and trimethylation of H3K27 in both cell lines (A549: Fig. 3F; H358: Supplementary Fig. S2E). Overall, these data confirm that EZH2 plays an important role in glucose restriction-induced cancer cell dedifferentiation.

Our data suggest that glucose deprivation causes dedifferentiation due to insufficient activity of αKG-dependent histone demethylases. To confirm this hypothesis, we performed RNA interference experiments with siRNAs targeting the two histone demethylases active on H3K27: UTX and JMJD3. Western blot analysis showed that knockdown of each of these demethylases in high glucose culture can mimic the effect of low glucose, causing reduction of FoxA2 levels in both A549 and H358 cells (Supplementary Fig. S2F), confirming the importance of H3K27me3 on the maintenance of cell differentiation and suggesting that both demethylases are required for cell differentiation in LUAD.

Therefore, we sought to test the hypothesis that EZH2 inhibition, by preventing glucose restriction-induced dedifferentiation, would potentiate the effect of SGLT2 inhibition in LUAD. We treated KP_luc mice with SGLT2 inhibitor empagliflozin, EZH2 inhibitor tazemetostat, or both from week 2 to week 8 after tumor induction, following the tumor burden by bioluminescence imaging. This trial showed that both empagliflozin and tazemetostat significantly reduced the tumor burden compared with the placebo group, and the combination treatment significantly reduced the tumor burden compared with empagliflozin single agent (Fig. 3G and H). To measure the effect of the treatment on cell differentiation, we stained lung slides for FoxA2 and Ttf1. Empagliflozin caused a significant reduction in Ttf1 and FoxA2 expression, which was completely rescued by cotreatment with tazemetostat (Fig. 2I and J). These data confirm that EZH2 inhibition has an additive effect on SGLT2 inhibitor therapy and rescues glucose restriction-induced dedifferentiation.

Glucose restriction changes gene expression profiles in LUAD cells, inducing cell proliferation block and dysregulation of cellular differentiation

Because we observed an important role of glucose in maintaining cell differentiation in LUAD, to find the pathways regulated by glucose deprivation, we analyzed the effect of glucose deprivation on gene-expression profiles in A549 and H358 cells. We incubated both cell lines in high glucose, low glucose, or low glucose with supplementation of dm-αKG for five days, followed by RNA-seq analysis. Incubation in low glucose caused upregulation of 2,394 genes in A549 cells and 1,335 genes in H358 cells, and downregulation of 2,024 genes in A549 cells and 942 genes in H358 cells (Supplementary Table S6). Of these, 518 were consistently upregulated and 439 downregulated in both cell lines (Supplementary Fig. S3A). Addition of dm-αKG to the low-glucose culture rescued the expression of the majority of glucose-dependent genes in A549: 1,634 up and 1,431 downregulated genes were rescued by dm-αKG (Supplementary Fig. S3B). Fewer genes were rescued in H358 cells: 510 upregulated genes and 370 downregulated genes (Supplementary Fig. S3C). We sought to find genes that were commonly regulated by low glucose and rescued by dm-αKG in the two cell lines, without excluding from the analysis genes that were barely below the threshold of significance. Therefore, we decided to continue our analysis on genes that were significantly regulated in at least three of the four conditions (low glucose vs. high glucose, low glucose + dm-αKG vs. low glucose, in both cell lines; Supplementary Table S6, where the class number indicates the number of conditions in which the gene is significantly regulated). A minority of genes (75 in total) were regulated discordantly by low glucose in the two cell lines, and we excluded these from our further analyses. We identified 524 genes upregulated and 445 genes downregulated by low glucose consistently in both cell lines and rescued by dm-αKG. We focused our analysis on these glucose-regulated, αKG-dependent 969 genes.

Figure 4A shows volcano plots of the glucose-regulated, αKG-dependent genes in A549 and H358 cells. Visual inspection of the plots showed that the most downregulated genes in both cell lines were involved in cell cycle and mitosis (MCM genes, E2F8, ESCO2, SPC25, RRM2, KIF20A, CDC20, and several histone subunits). Conversely, the most upregulated genes included genes involved in neuronal differentiation (BEX2, PCDH1, AKNA, UNC5B, CUX1, FLRT1, PSAP, APP, SPX, and PHRP), hematopoietic differentiation (SLFN5, LCN2), and hypoxia-regulated genes (NUPR1, UNC5B, LAMP3, RRAGD, GDF15, DNAJC2, and DDIT3), suggesting that low glucose stopped cell proliferation and mitosis, along with dysregulation of cell differentiation and activation of hypoxia signaling. Consistently, IPA showed inhibition of functions related to cell proliferation (cell-cycle control of chromosomal replication, kinetochore metaphase signaling pathway, cyclins and cell-cycle regulation, cell-cycle regulation by BTG family proteins), DNA repair (nucleotide excision repair), as well as epithelial cell identity (epithelial adherens junction signaling). Conversely, IPA showed activation of stress pathways (senescence pathway, endoplasmic reticulum stress pathway, NRF2-mediated oxidative stress response, autophagy, unfolded protein response), pathways related to neuronal differentiation (neuregulin signaling, synaptic long-term depression) and hematopoietic differentiation (Fcγ receptor-mediated phagocytosis in macrophages and monocytes, PI3K signaling in B lymphocytes), and hypoxia (HIF1α signaling; Fig. 4B; Supplementary Table S7). Figure 4C shows representative downregulated genes associated with epithelial cell identity and cell cycle, and upregulated genes associated with stress and hypoxia signaling. Consistently with the IPA results, GSEA showed downregulation of cell proliferation pathways (G₂–M checkpoint and others), and upregulation of stress pathways (unfolded protein response and others), as well as the hypoxia pathway (Fig. 4D; Supplementary Table S8). We confirmed by RT-PCR the low glucose-dependent downregulation of some genes associated with epithelial cell identity (TUBB4A and TUBB4B) and upregulation of hypoxia-dependent genes (VEGFB and HK2), and all were rescued by αKG (Fig. 4E). To confirm the relevance of these gene-expression changes in vivo, we performed RT-PCR in tumors treated with empagliflozin from week 2 to week 8 after tumor induction in KP_luc mice. This analysis confirmed downregulation of Tubb4a and Tubb4b and upregulation of Vegfr and Hk2 in murine tumors (Fig. 4F). Overall, RNA-seq analysis suggested that glucose restriction caused a proliferation block and activation of stress signals in LUAD cell lines, associated with dysregulation of cell differentiation, with downregulation of epithelial markers and upregulation of genes associated with neuronal and hematopoietic lineages, as well as the hypoxia pathway. These observations were very interesting because quiescence and increased response to stress and hypoxia can be associated with resistance to therapy (45).

Figure 4.

Glucose restriction affects gene expression patterns in LUAD cells. A549 and H358 cells were incubated in high, low glucose, or low glucose plus αKG for 5 days, followed by RNA extraction of RNA-seq analysis. We focused our analysis on the genes that were commonly up- or downregulated in both cell lines and that were rescued by αKG. A, Volcano plots of genes up- and downregulated in both cell lines. B, Representative pathways identified by IPA in the genes up- or downregulated by glucose in both cell lines. C, Heat map of selected genes up- or downregulated by low glucose and rescued by αKG in both cell lines. D, Representative GSEA plots of pathways enriched in low glucose vs. high glucose. E, RT-PCR analysis was performed on selected genes identified from gene expression analysis in A549 and H358 cell lines incubated in high (20 mmol/L) and low (1 mmol/L) glucose, with or without the supplement of αKG. α-Mannitol was used as osmotic control. F, RT-PCR in KP_luc tumors from mice treated with either placebo or empagliflozin from week 2 to week 8 after tumor induction. *, P < 0.05; **, P < 0.01; ***, P < 0.001; ****, P < 0.0001; n.s., not significant.

Glucose restriction induces shifts in chromatin marks associated with EMT

Because we observed an important role of EZH2 in glucose restriction-induced dedifferentiation, we decided to investigate the global distribution of the H3K27me3 mark in A549 cells incubated for 5 days in low (1 mmol/L) vs. high (20 mmol/L) glucose by ChIP-seq. As expected, incubation in low glucose caused a massive repositioning of H3K27me3 in the genome: 12,662 sites present in high glucose were not present in low glucose; conversely, 9,773 sites were present in low glucose but not in high glucose (Fig. 5A). A minority of the sites (5,538) were present both in high and low glucose. Figure 5B shows a heat map with changes in H3K27me3 genomic locations in high versus low glucose. A panoramic analysis of the localization of these H3K27me3 sites in the genome showed that most of these sites were intronic and intergenic (Fig. 5C). Supplementary Table S9 reports a more detailed description of the intergenic sequences with enriched H3K27me3 in high and low glucose, including repetitive sequences (LINE and SINE), long terminal repeats, and other nonspecified intergenic regions. H3K27 methylation in large intergenic regions has been associated with X chromosome inactivation (46) and with silencer elements (47). We next wanted to analyze the changes in H3K27me3 on the genes that were regulated by low glucose in A549 cells in the RNA-seq experiment. As expected, the H3K27me3 mark was significantly increased in the genes downregulated by low glucose (Fig. 5D), whereas there was no change in the genes that were upregulated by low glucose (Fig. 5E). These results confirmed that changes in EZH2 activity are prevalent in the cellular response to glucose deprivation, and are involved in silencing of genes associated with cell proliferation and epithelial differentiation.

Figure 5.

Glucose restriction causes the repositioning of H3K27me3 on repressed genes and enrichment of H3K4me3 on Slug target genes. A549 cells were incubated in either high (HG) or low (LG) glucose for 5 days, followed by immunoprecipitation with specific antibodies targeting H3K27me3 and H3K4me3. A and B, Global distribution of the H3K27me3 mark in high vs. low glucose, expressed as Venn diagram (A) and heat map (B). C, Functional classification of the H3K27me3-enriched sites in high and low glucose. D and E, Changes in H3K27 trimethylation in genes that were downregulated (D) or upregulated (E) by low glucose in the RNA-seq experiment. F and G, Global distribution of the H3K4me3 mark in high vs. low glucose, expressed as Venn diagram (F) and heat map (G). H, Functional classification of the H3K4me3-enriched sites in high and low glucose. I, Comparison of ChIP-seq, RNA-seq, and RT-qPCR analysis at 5 days and at 30 days of incubation in low glucose on a subset of 11 genes that showed a significant increase in H3K4me3 mark in low vs. high glucose. J, Western blot analysis of EMT transcription factors SNAIL, SLUG, and ZEB1 in A549 and H358 cells incubated for 5 days in a medium containing different concentrations of glucose with or without dm-αKG, as indicated.

Because the RNA-seq analysis showed upregulation of high numbers of genes in the cells exposed to low glucose, we also sought to characterize the changes in an activation mark, H3K4me3, induced by incubation of A549 cells in low glucose. We observed that a minority of H3K4me3 sites were changed by incubation in low glucose (Fig. 5F–G). As expected, most enrichment sites were associated with promoters or introns (Fig. 5H). Alignment of the ChIP-seq data with the RNA-seq data showed no significant enrichment of H3K4me3 on the promoters of genes that were activated in the RNA-seq experiment, suggesting that this histone mark is not associated with gene activation after 5 days of glucose starvation. However, we reasoned that changes in this activation mark could cause changes in gene expression at longer time points. We decided to focus on the promoters of the genes on which the H3K4me3 mark is increased after incubation in low glucose. We identified 14 genes that displayed higher recruitment of H3K4me3 on their promoter in low versus high glucose (Supplementary Table S10). Transcription factor analysis of these promoters (Supplementary Table S11) showed significant enrichment of binding sites for Snail, Slug, and ZEB1 in the promoters of these genes, suggesting that glucose restriction induces activation of EMT. Interestingly, the promoters of 12 of the 14 genes also contained the binding site for HIF1α, confirming an important role of the hypoxia pathway in the response to glucose restriction. However, only 3 of the 14 genes were upregulated by low glucose in the RNA-seq experiment. We reasoned that these genes may be primed for activation after 5 days of glucose restriction, but require longer incubation times to be upregulated. Therefore, we performed RT-PCR analysis of the 14 genes in A549 cells incubated in low glucose for 5 days versus long-term incubation in low glucose for at least one month. As expected, 11 of the 14 genes were significantly (P < 0.01; t test) upregulated by longer incubation in low glucose (Fig. 5I), whereas 3 were not detectable in A549 cells by RT-PCR analysis. These observations suggested that glucose restriction primes certain genes associated with EMT for activation at later time points.

It has been previously reported that glutamine deprivation causes activation of Slug, but not Snail or ZEB1 (48). To test the activation of Snail, Slug, and ZEB1 in our system, we incubated A549 and H358 cells in high versus low glucose for 5 days, with or without supplementation of dm-αKG, followed by Western blot. Glucose deprivation caused an increased level of Slug in both cell lines, rescued by dm-αKG, whereas it surprisingly caused a reduction in Snail and ZEB1 levels in A549 (this factor was not expressed in H358 cells), and this inhibition was partially rescued by dm-αKG (Fig. 5J). This experiment suggests that Slug, but not ZEB or Snail, is upregulated by low glucose, and it may be responsible for driving the more aggressive phenotype observed in glucose-restricted cells.

EZH2 causes cancer cell dedifferentiation by regulating HIF1α signaling in LUAD cells

Because the hypoxia pathway was significantly upregulated in the RNA-seq analysis, and HIF1α activation has been associated with EMT induction, we decided to focus on this pathway. To examine the activation of hypoxia-related pathways by glucose deprivation, we performed Western blot analysis in A549 and H358 cells after incubation for 5 days in low glucose, with or without the addition of dm-αKG. Both the two major HIF isoforms, HIF1α and HIF2α, were upregulated by low glucose in A549 (Fig. 6A) and H358 (Supplementary Fig. S4A) cells, and this was reversed by dm-αKG, whereas dm-αKG did not have any effect on the cells growing in high glucose. Furthermore, RT-PCR showed upregulation of both HIF isoforms by low glucose, rescued by dm-αKG in A549 (Fig. 6B) and H358 cells (Supplementary Fig. S4B). These results showed that low glucose can upregulate the expression of HIF isoforms at both the protein and mRNA levels.

Figure 6.

EZH2 causes dedifferentiation by regulating Hif1α signaling in LUAD. A and B, A549 cells were incubated in high or low glucose, with or without dm-αKG, as indicated. HIF1α and HIF2α expression was evaluated by Western blotting (A) and RT-PCR (B). C, KP_luc mice were given AdenoCre to induce tumors, followed by treatment with either placebo or empagliflozin as in Fig. 1D, followed by collection of the lungs and RNA extraction for RT-PCR. Significance was measured by Student t test. D, PDOs from three patients were incubated in high vs. low glucose, as in Fig. 1E, followed by RT-PCR. Significance was measured by Student t test. E, A549 cells were incubated in low glucose and transfected with either control or siRNAs targeting HIF1α and HIF2α in either high or low glucose, followed by Western blotting for HIF1α, HIF2α, and FOXA2. F–H, A549 cells were incubated in different concentrations of glucose, as in Fig. 1F–H, followed by Western blot (F and G) or RT-PCR (H). G, Quantification of the Western blot signal was performed with ImageJ. H, The relation between RT-PCR signal and glucose concentration was evaluated by Pearson method for PHD3 and by Spearman method for HIF1α. I, KP_luc mice carrying LUADs were treated with either placebo or empagliflozin (10 mg/kg/d) from week 2 to week 8 after tumor induction (same experiment as in Fig. 1C; the actin control was the same as Fig. 1C). Western blot was performed on whole tumor extracts from three tumors for each group. J, HIF1A and PHD3 mRNA expression level was evaluated by RT-PCR in A549 cells transfected with or without siRNAs targeting EZH2. K, A549 cells were transfected with pooled siRNAs for EZH2 and/or PHD3. Knockdown efficiency and changes in HIF1α and FOXA2 expression were evaluated by Western blotting. L, The Western blot signal for HIF1α was measured by ImageJ. Significance was evaluated by Student t test. M, HIF1α expression was evaluated in cells transfected with EZH2 and PHD3 expression vectors. The overexpression of the two proteins was also confirmed by Western blotting. N, A549 cells were incubated in either high or low glucose for 5 days, followed by ChIP assay with either an antibody targeting EZH2 or normal IgG as negative control. qPCR was performed on the precipitated DNA with primers targeting the promoter of PHD3, MYT-1 (a canonical EZH2 target), or GAPDH (an unrelated promoter as negative control). The results are reported as a percentage of input. Significance was evaluated by Student t test comparing the values in high vs. low glucose for each target promoter. *, P < 0.05; **, P < 0.01; ***, P < 0.001; ****, P < 0.0001; n.s., not significant.

To confirm the activation of the hypoxia pathway in vivo and more clinically relevant models, we performed RT-PCR for HIF1α in the KP_luc murine tumors of mice treated with empagliflozin and in the three PDOs incubated in high (20 mmol/L), physiologic (5 mmol/L) and low (1 mmol/L) glucose (the same experiments as in Fig. 1). As expected, empagliflozin treatment of mice caused a significant upregulation of HIF1α RNA levels (Fig. 6C), and in PDOs the levels of HIF1α increased progressively as the glucose concentration in the medium was reduced from 20 to 5 and to 1 mmol/L (Fig. 6D).

To evaluate if HIFs were involved in glucose restriction-induced dedifferentiation, we incubated A549 cells in low glucose with or without the transfection of siRNAs targeting either HIF1α or HIF2α. The efficiency of siRNA knockdown was also confirmed by Western blot for both proteins. Interestingly, we observed that the HIF1α, but not the HIF2α knockdown, increased the expression of LUAD differentiation marker FOXA2 in low glucose, but did not have any effect on the basal levels of FOXA2 in high glucose in A549 (Fig. 6E) and H358 cells (Supplementary Fig. S4C), suggesting that activation of HIF1α repressed FOXA2 in low glucose.

We next aimed to investigate the mechanism of HIF1α upregulation by glucose restriction in LUAD. HIF1α is negatively regulated by prolyl hydroxylases (PHD), which induce proteasome-dependent degradation of HIF1α under normoxic conditions (49). To investigate which PHD isoform was affected by low glucose, we performed RT-PCR for the three known PHDs in A549 and H358 cells incubated in high and low glucose with and without dm-αKG for 5 days. Only the expression of PHD3 was repressed by low glucose and rescued after the addition of dm-αKG in both A549 and H358 cells (Supplementary Fig. S4D). Incubation in high (20 mmol/L), physiologic (5 mmol/L), and low (2 mmol/L) glucose for 5 days caused a progressive and significant reduction of PHD3 and increase in HIF1α protein as the concentration of glucose decreased in A549 (Fig. 6F and G) and H358 (Supplementary Fig. S4F and S4G). We also observed a progressive activation of the EMT marker N-cadherin in low glucose (Fig. 6E and F; Supplementary Fig. S4E and S4F). RT-PCR analysis also showed a direct linear correlation between PHD3 expression and glucose concentration, and a significant inverse exponential correlation between HIF1α and glucose concentration in A549 (Fig. 6H) and in H358 (Supplementary Fig. S4G) cells. Consistently, PHD3 protein was downregulated and HIF1α protein was upregulated in murine tumors treated in vivo with empagliflozin (Fig. 6I). It has been previously demonstrated that loss of PHD3 promotes dedifferentiation in breast cancer through a hydroxylase-dependent mechanism (50), and PHD3 is strongly upregulated in well-differentiated human pancreatic cancer cells compared with less-differentiated tumors (51), acting as a marker of differentiation. Consistently, PHD3 knockdown with two different siRNAs caused increased VEGF expression (a marker of HIF activation) and reduced FoxA2, suggesting dedifferentiation in high glucose (Supplementary Fig. S4H). Therefore, we hypothesized that EZH2 upregulates HIF1α and leads to dedifferentiation in low glucose via inhibition of PHD3.

To investigate the role of EZH2 in HIF1α regulation by glucose, we incubated cells in low glucose and performed transfection with either control or siRNAs targeting EZH2, followed by RT-PCR analysis. EZH2 knockdown reduced HIF1α and increased PHD3 expression in low glucose in A549 (Fig. 6J) and H358 cells (Supplementary Fig. S4I), whereas it did not have any effect in high glucose. This result was also confirmed by a Western blot. PHD3 was upregulated and HIF1α downregulated by EZH2 silencing in low glucose in A549 (Fig. 6K) and H358 cells (Supplementary Fig. S4J). As expected, EZH2 siRNA also increased the expression of differentiation marker FOXA2, suggesting that EZH2 inhibits cell differentiation through PHD3 inhibition and HIF1α stabilization in low glucose. To verify this mechanism, we also transfected cells with an siRNA pool targeting PHD3. PHD3 knockdown in low glucose did not cause significant changes in HIF1α and FOXA2 compared with control. This is not surprising because PHD3 is downregulated in low glucose (Fig. 6F and G; Supplementary Fig. S4E and S4F), and inactive for low αKG, which is an essential cofactor of PHD3 activity. However, double knockdown of both EZH2 and PHD3 protein resulted in significantly increased expression of HIF1α compared with single transfection of siRNA for EZH2 (Fig 6L; Supplementary Fig. S4K). PHD3 knockdown also reversed the upregulation of FOXA2 caused by siEZH2 in low glucose in A549 (Fig. 6K) and H358 (Supplementary Fig. S4L). This result suggested that HIF1α activation and cell dedifferentiation induced by EZH2 in low glucose are mediated by inhibition of PHD3.

To confirm that EZH2-dependent inhibition of PHD3 can cause HIF1α stabilization, we overexpressed EZH2 and/or PHD3 in A549 cells incubated in high glucose media. EZH2 overexpression caused reduced PHD3 and increased HIF1α expression. PHD3 overexpression did not change the basal level of HIF1α, which was already basally low. Interestingly, the double overexpression of EZH2 and PHD3 abrogated the HIF1α induction caused by EZH2 overexpression, bringing back HIF1α to control levels in A549 (Fig. 6M) and H358 cells (Supplementary Fig. S4M). These results suggest that the unbalanced activation of EZH2 in low glucose causes LUAD dedifferentiation by inhibition of PHD3 and consequent activation of HIF1α. Interrogation of the Gene Transcription Regulation Database (GTRD; ref. 52) showed recruitment of EZH2 on two sites in the promoter of the EGLN3 gene, which encodes for PHD3 (Supplementary Fig. S5A). To verify that EZH2 is directly responsible for the transcriptional repression of PHD3 in low glucose, we performed a ChIP-qPCR assay. We incubated A549 cells in either high or low glucose for 5 days, followed by ChIP with an antibody specific for EZH2. This experiment showed that EZH2 recruitment was increased in low glucose both on the PHD3 promoter and on the canonical EZH2 target MYT-1, whereas no recruitment was observed on an unrelated promoter (GAPDH), or in the sample precipitated with control IgG (Fig. 6N). Similar results were obtained in H358 cells (Supplementary Fig. S5B). This result confirmed that EZH2 directly repressed the expression of PHD3 in cells incubated in low glucose.

Because the HIF pathway is associated with EMT and we have observed Slug induction by incubation in low glucose, we tested the hypothesis that Slug is downstream of the EZH2–HIF1α axis here identified. We incubated A549 cells in either high or low glucose, and in low glucose with control siRNA or pooled siRNAs targeting either EZH2 or HIF1α. As expected, incubation in low glucose caused increased levels of HIF1α and Slug and reduction of Snail (Fig. 7A). HIF1α knockdown completely rescued, and EZH2 knockdown partially rescued, the changes induced by low glucose in both Slug and Snail levels in A549 (Fig. 7A) and H358 cells (Supplementary Fig. S5C). This experiment confirmed that Slug activation in low glucose is dependent on HIF1α. Our above-shown data showed that incubation for 5 days in low glucose primes lung cancer cells for EMT, but EMT-associated genes are upregulated at later time points. Therefore, we aimed at investigating the upregulation of Slug by long-term incubation in low glucose. We performed RT-PCR for Slug canonical inhibition target E-cadherin and the EMT marker vimentin after 30 days of low glucose incubation. As expected, long incubation in low glucose repressed E-cadherin and increased vimentin expression in A549 (Fig. 7B) and H358 cells (Supplementary Fig. S5D). The results showed that incubation in low glucose caused activation of Slug and subsequent EMT, with inhibition of E-cadherin and upregulation of vimentin at 30 days. To confirm the role of HIF1α in low glucose-induced EMT, we incubated the cells in low glucose for at least 30 days, followed by transfection with siRNAs targeting HIF1α. As expected, E-cadherin expression was increased, and vimentin was reduced, by siHIF1α in A549 (Fig. 7C) and H358 cells (Supplementary Fig. S5E). To confirm the relevance of this pathway in vivo, we performed Western blot for E-cadherin in KP_luc tumors from mice treated with empagliflozin for 6 weeks and observed downregulation of E-cadherin in empagliflozin-treated tumors compared with placebo (Fig. 7D). IHC for E-cadherin in lungs of KP_luc mice treated with empagliflozin and/or EZH2 inhibitor tazemetostat from week 2 to week 8 after tumor induction showed a significant downregulation of E-cadherin, which was significantly rescued by cotreatment with tazemetostat (Fig. 7E and F), confirming that this pathway is active in vivo. These data confirmed that pseudohypoxia in low glucose activates the EMT pathway in LUAD cells.

Figure 7.

EZH2 and Hif1α drive EMT in LUAD. A, A549 cells were incubated for 5 days in low glucose, with transfection of either control siRNA or siRNAs targeting HIF1α and EZH2, followed by Western blot analysis as indicated. B, RT-PCR for E-cadherin and vimentin was performed in A549 cells incubated in high glucose and low glucose for 30 days, and in cells incubated in low glucose for 30 days with transfection of siRNAs for HIF1α. Significance was measured by Student t test. C, After a 30-day incubation in low glucose, cells were transfected with siRNAs targeting HIF1α and RT-PCR. D, KP_luc mice carrying LUADs were treated with either placebo or empagliflozin (10 mg/kg/d), from week 2 to week 8 after tumor induction (same experiment as in Fig. 1C; the actin control was the same as Fig. 1C). Western blot was performed on whole tumor extracts from three tumors for each group. E and F, KP_luc mice carrying LUADs were treated with placebo, empagliflozin, and/or tazemetostat as in Fig. 3G–J. E, Quantification of the E-cadherin signal, expressed as H score, measured by QuPath analysis. Error bars, mean ± SEM. F, Representative images. Significance was measured by Student t test. **, P < 0.01; ****, P < 0.0001.

Our data suggested that the regulation of HIF1α by the EZH2–PHD3 pathway was transcriptional, because both RNA and protein levels of HIF1α were stimulated in low glucose. However, HIF1α is known to be regulated posttranscriptionally by PHD3, through prolyl-hydroxylation and proteasome degradation. Importantly, the activity of PHD3 also depends upon αKG availability. However, the direct hypoxic stabilization of HIF1α is an early event (within hours of hypoxia exposure), whereas the time point we examined here was after 5 days. After so long, more indirect mechanisms are likely to occur. In particular, it has been recently shown that increased HIF1α activity triggers a positive feedback loop with transcription of the long noncoding RNA HIFAL, which ultimately activates HIF1α transcriptionally (53). We therefore measured the expression of HIFAL and detected a significant activation of HIFAL in low glucose, completely rescued by αKG (A549: Fig. 8A; H358: Supplementary Fig. S5F). To verify that HIFAL is responsible for HIF1α induction after 5 days of low glucose exposure, we transfected A549 (Fig. 8B) and H358 cells (Supplementary Fig. S5G) with siRNAs targeting HIFAL and observed that HIFAL knockdown in low glucose caused reduction of HIF1α and GLUT1 expression and induction of FoxA2 at the protein level (A549: Fig. 8C; H538: Supplementary Fig. S5H) and the RNA level (A549: Fig. 8D; H358: Supplementary Fig. S5I). These results suggest that sustained HIF1α activation in low glucose exposure requires an HIFAL-dependent positive feedback with transcriptional upregulation of HIF1α.

Figure 8.

HIF1α signaling is associated with a more aggressive phenotype. A, A549 cells were incubated in high or low glucose with or without supplementation of dm-αKG, followed by RT-PCR for HIFAL. B–D, A549 cells were incubated in low glucose for 30 days, with transfection of two siRNAs targeting HIFAL or control siRNAs, five and two days before RNA extraction. RT-PCR (B and D) and Western blot (C) were performed with the indicated makers. E–G, Model of HIF1α regulation by low glucose. In normal conditions of oxygen and glucose, HIF1α is targeted for degradation by αKG-dependent hydroxylation by PHD3 (E). When cells are incubated in low glucose for a short time (2 hours), the lack of αKG causes HIF1α stabilization, which can be rescued by αKG supplementation (F). Longer exposure to low glucose (5 days) causes a more complex mechanism of HIF1α activation, with transcriptional upregulation by the long noncoding RNA HIFAL and transcriptional repression of PHD3 by EZH2 recruitment on the PHD3 gene promoter (G). H–M, Three different low glucose-induced clones of murine LUAD cells 2953A were expanded from a long-term incubation in low (1 mmol/L) glucose for at least one month, as in Fig. 2D–I. One day before the injection in syngeneic mice, cells were transfected with two pooled siRNAs targeting HIF1α and siRNA control. The day after, the cells were injected in syngeneic mice by tail-vein injection to measure the development of lung metastases (H). The tumor burden was measured by bioluminescence. We report representative bioluminescence pictures of single mice (I) and quantification of the bioluminescence (J). Lung hematoxylin and eosin sections (K) were used to measure tumor areas (L) and number of tumors (M) as in Fig. 2G–I. Significance was evaluated by Student t test. This experiment was performed together with the experiment presented in Fig. 2D–I. The bioluminescence and tumor measurement data for the low-glucose clones is the same as that presented in Fig. 2F and 2G–H, respectively. N, The expression of the hypoxia gene set identified in our RNA-seq in A549 and H358 cells was measured in human LUAD by TCGA analysis. O, The correlation of hypoxia markers (HIF1α and GLUT1) with markers of differentiation (TTF1 and FOXA2), as expressed in transcripts per million (TPM), was measured in TCGA samples. P, Kaplan–Meier overall survival curve of TCGA data showing a statistically significant difference in survival in patients expressing high vs. low levels of genes involved concomitantly in HIF1A signaling. *, P < 0.05; **, P < 0.01; ***, P < 0.001; ****, P < 0.0001.

To confirm the hypothesis that short-term and long-term glucose deprivation cause activation of HIF1α through different mechanisms (protein stabilization and transcriptional regulation, respectively), we performed a time-course experiment with incubation of A549 cells in low glucose for different time points, from 2 to 48 hours, showing that the increase in HIF1α protein levels is an early event, starting at 2 hours after incubation in low glucose, and continuing for up to 48 hours (Supplementary Fig. S6A). We also explored earlier time points (30′, 1 hour and 2 hours) in both A549 and H358 cells, showing that 2 hours is the earliest time point when HIF1α is activated (Supplementary Fig. S6B). To verify the mechanism of early activation of the HIF1α pathway, we treated A549 and H358 cells with both αKG and EZH2 knockdown and performed Western blot after 2 hours of incubation in low glucose. This experiment showed that after short-time glucose deprivation, the activation of HIF1α is rescued by αKG, but not by EZH2 knockdown (Supplementary Fig. S6C). Taken together, these data suggest that glucose deprivation activates the hypoxia pathway through different mechanisms: at early time points (2 hours), HIF1α is regulated by αKG but not by EZH2, whereas at later time points (5d) HIF1α is regulated both by αKG and by EZH2. We assume that at 2 hours, HIF1α stabilization depends on the direct inactivation of PHDs by low αKG availability (Fig. 8E and F), whereas at later time points, the activation of positive feedback loops including HIFAL upregulation causes transcriptional upregulation of HIF1α and EZH2 recruitment on the PHD3 promoter causes transcriptional repression of PHD3, with a combination of transcriptional and posttranscriptional activation of HIF1α (Fig. 8G).

We next investigated the role of HIF1α in the observed phenotypic shift toward a more aggressive phenotype. To this aim, we used murine 2953A cells incubated for at least 1 month in a culture medium containing low glucose (1 mmol/L). We transfected three clones of low-glucose 2953A cells with either control or pooled siRNAs against HIF1α, followed by tail-vein inoculation in syngeneic mice (Fig. 8H). A control experiment with two different concentrations of the siRNAs targeting HIF1α showed that the RNAs were still silenced 7 days after transfection (Supplementary Fig. S6D). One week after injection, we detected that the cells transfected with HIF1α siRNAs had significantly lower BLI signal than the control cells (Fig. 8I and J). Histology on the lungs collected one week after the tumor cell injection showed a significantly reduced number of tumors and tumor areas in lungs injected with the HIF1α-knockdown cells compared with the parental clones (Fig. 8K–M). These results suggest that HIF1α activation by glucose restriction is responsible for a transition of LUAD cells toward a more aggressive phenotype.

To test if these epigenetically driven changes in cell differentiation are reversible, we rechallenged A549 and H358 cells that had been cultured long-term (>30 days) in low glucose with higher glucose concentrations: physiologic (5 mmol/L) and high (30 mmol/L) glucose for either 5 or 30 days. We observed a reduction of the H3K27me3 mark and HIF1α expression and increased FoxA2 expression in both physiologic and high glucose at 30 days after rechallenge compared with the low-glucose cells, whereas at 5 days the rescue was partial (Supplementary Fig. S6E). These data suggest that the observed starvation-induced dedifferentiation is reversible in the long term.

We finally sought to investigate the clinical relevance of the HIF pathway in human LUAD by interrogating The Cancer Genome Atlas (TCGA) database. First, we wanted to find out which percentage of LUADs have the hypoxia pathway activated. We looked at the expression of the genes identified in our RNA-seq analysis as upregulated by low glucose in the hypoxia GSEA (Fig. 4D). This analysis showed that in total, 8% of tumors have the hypoxia pathway activated, as defined by an ssGSEA score > 1.5. This percentage is significantly higher (15%) in stage IV patients (Fig. 8N), suggesting that activation of the hypoxia pathway is associated with more advanced disease. To confirm that the GSEA hypoxia pathway is correlated with HIF1α expression in patient tumors, we measured HIF1α mRNA expression in hypoxia-activated (ssGSEA score > 1.5) and hypoxia-repressed (ssGSEA score < −1.5) tumors, and found a significantly higher expression of HIF1α in hypoxia-activated tumors (Supplementary Fig. S6F). We also compared HIF1α expression in tumors with the most frequent oncogenic mutations in LUAD (KRAS, EGFR, and ALK), and did not find any difference in HIF1α expression across different oncogenotypes (Supplementary Fig. S6G). To confirm that the activation of the hypoxia pathway is associated with dedifferentiation, we performed Pearson analysis to compare the cumulative expression of hypoxia markers (HIF1α and GLUT1) with markers of differentiation (FOXA2 and TTF1). As expected, we observed a significant negative correlation between markers of differentiation and hypoxia (Fig. 8O).

We performed Kaplan–Meier analysis to measure the effect of hypoxia-regulated genes identified in our RNA-seq experiment on LUAD disease-free (DFS) and overall (OS) survival. Supplementary Table S12 shows the hazard ratio (HR) and P value of hypoxia-regulated genes whose overexpression was associated with a significant HR (> 1.5 or < 0.5). These included, as expected, glycolytic genes (SLC2A1, HK2, ENO1, ENO2, and ALDOC) and canonical HIF1α targets (VEGFA and VEGFB). We identified a group of 16 genes associated with HR >1.5 (ALDOC, ANGPTL4, CAVIN1, DDIT3, ENO1, ENO2, FAM162A, HK2, LDHA, PPP1R15A, PPP1R3C, SLC2A1, TPI1, VEGFA, VEGFB, and VHL), and measured the cumulative effect of overexpression of these genes on LUAD patient survival. Cumulative expression of these genes was associated with an HR of 1.86 (P < 0.0001) for DFS (Supplementary Fig. S6H) and of 2.55 (P < 0.0001) for OS (Fig. 8P), confirming an important role of the activation of the hypoxia pathway on the clinical behavior of LUAD.

Starvation-induced dedifferentiation may be a general mechanism shared by different cancers

We finally aimed to check if glucose restriction-induced dedifferentiation is restricted to lung cancer or generalized to other cancer types. Western blot analysis showed that incubation in low glucose for 5 days induced FoxA2 downregulation, rescued by αKG, in a breast cancer (MCF7) and pancreas cancer (PANC1) cell line (Supplementary Fig. S6I). This result confirmed that the described pathway is not restricted to LUAD, but is present in other tumors as well.

Glutamine depletion has previously been shown to cause H3K27 hypermethylation and dedifferentiation in the core regions of melanoma (20). We wanted to test if glutamine starvation also caused HIF1α activation in our system. We incubated A549 and H358 cells in either low glucose (1 mmol/L) or low glutamine (2 mmol/L), or both low glucose and glutamine, for 5 days, followed by Western blot analysis of HIF1α expression. As expected, glutamine deprivation caused an increase in HIF1α level, which was even more pronounced than that caused by glucose deprivation (Supplementary Fig. S6J). This is expected, because αKG derives directly from glutamine. Interestingly, restriction of both glutamine and glucose did not cause a further increase in HIF1α activation. Taken together, these data confirmed that pseudohypoxia is a common response mechanism to glutamine and glucose starvation.

Discussion

Here, we report a novel mechanism by which glucose restriction, while blocking cell proliferation and delaying the growth of lung tumors, induces dedifferentiation and increases the aggressiveness of LUAD cells. Glucose deprivation causes a reduced abundance of TCA cycle metabolites, with insufficient activity of αKG-dependent histone demethylases and unbalanced activity of methyltransferase EZH2, leading to repression of PHD3 and hyperactivation of HIF1α. HIF1α activation leads to Slug activation and EMT, eventually causing an aggressive/metastatic phenotype. We envision at least two different scenarios in which this mechanism can be relevant. In the first scenario, glucose deprivation can occur in cancers even in the absence of treatments, as a consequence of insufficient vascularization. In this case, glucose deprivation is accompanied by glutamine deprivation and hypoxia, reinforcing the activation of the HIF pathway and accelerating the progression of LUAD toward a more aggressive and dedifferentiated phenotype. Consistently, our results showed that overexpression of certain genes of the hypoxia pathway in human LUAD is associated with significantly reduced survival.

In the second scenario, glucose deprivation is caused by treatment with glucose transport inhibitors or glycolytic inhibitors, which are novel experimental strategies against lung cancer (8, 12). Our work provides evidence that these treatments are effective in reducing the tumor burden, but are also likely to cause an epigenetic adaptation of cancer cells to glucose restriction, leading to an unexpected and unintended dedifferentiation and increased aggressiveness of treated tumors driven by pseudohypoxia. SGLT2 inhibitors are FDA approved for diabetes, heart failure, and chronic kidney disease (54). Since the first SGLT2 inhibitors were introduced a decade ago, epidemiologic evidence is starting to emerge showing improved cancer outcomes in diabetic patients treated with SGLT2 inhibitors (13, 14), confirming the beneficial effects of these drugs against cancer. Recently, the first prospective phase I study of the SGLT2 inhibitor dapagliflozin for pancreatic cancer has been published (55), increasing the enthusiasm for SGLT2 inhibitors as a new anticancer strategy. However, our preliminary data showed that chronic glucose restriction causes starvation-induced dedifferentiation, eventually promoting a more aggressive phenotype. This could be a cancer adaptation and a resistance mechanism that may emerge in some patients after chronic administration of SGLT2 inhibitors and may limit the clinical utility of these drugs against cancer. Because millions of patients are prescribed SGLT2 inhibitors on a chronic basis, some of which are bound to develop lung or other premalignant lesions, it is of paramount importance to characterize in detail the action of SGLT2 inhibitors against cancer and to elucidate the epigenetic adaptations of LUAD to chronic SGLT2 inhibition. Our in vivo reinjection experiments showed that glucose-starved cells, once exposed to normal blood glucose concentrations, are able to grow faster than non-starved cells. This suggests that there may be a rebound effect if the treatment is stopped, as suggested by Park and colleagues in their clinical trial with dapagliflozin (55). However, our rechallenge experiment showed that starvation-induced dedifferentiation is a reversible phenomenon, and cells that had been grown in low glucose long-term and had gone through dedifferentiation can reactivate the differentiation markers when reexposed to long-term physiologic or high glucose. Based on these results, we envision starvation-induced dedifferentiation as a dynamic process, for which cells go through phases of dedifferentiation when deprived of glucose or glutamine, and then stabilize and redifferentiate when rechallenged with higher glucose concentrations (for instance, for GLUT1 upregulation). This plasticity is typical of an epigenetically regulated process. Indeed, we showed that combination treatment with an EZH2 inhibitor, tazemetostat, significantly improves the response of LUAD to SGLT2 inhibition and prevents starvation-induced dedifferentiation. Combination treatments with epigenetic modulators or HIF inhibitors are therefore important strategies as metabolic therapies are moved to the clinic. The challenge will be to find the right therapeutic window to prevent starvation-induced dedifferentiation without causing significant side effects due to whole-body inhibition of H3K27 methylation. In the future, alternative strategies to increase the intracellular αKG levels (inhibition of oxoglutarate dehydrogenase, direct supplementation of αKG) may also be implemented to prevent starvation-induced dedifferentiation.

Because our murine models and the cell lines used for this study are Kras-dependent, we currently do not know if the observed phenotypes are applicable to other lung cancer oncogenotypes. One of the PDOs used for the study has EGFR exon 19 deletion (Supplementary Table S4), but we do not feel this evidence is enough to generalize our observations to other oncogenotypes of lung cancer. Some observations suggest that glucose starvation-induced dedifferentiation may not be limited to Kras-dependent cancers. Our previous observation showed that SGLT2 expression is associated with tumor differentiation, regardless of the mutational background (12). The TCGA analysis presented here showed that the activation of the hypoxia pathway is not different in tumors carrying KRAS, EGFR, or ALK mutations. Finally, we also observed downregulation of differentiation marker FOXA2 in breast cancer cells, which do not depend on KRAS mutation. However, further studies are warranted to verify that starvation-induced dedifferentiation is a general phenomenon not restricted to KRAS-mutant cancers.

Cancer dedifferentiation is known to occur because of glutamine deprivation in the tumor microenvironment (20). Our data suggested that glucose deprivation causes a similar phenotypic switch from an early to poorly differentiated state in LUAD. We showed that αKG depletion causes an unbalanced activity between histone demethylase and methyltransferase, resulting in hypermethylation on histone marks. In particular, increased trimethylation on H3K27, carried out by EZH2, can affect the expression of a wide variety of signaling pathways involved in the cellular transition to a more aggressive phenotype. Our work led to the discovery that a key mediator of the aggressive phenotype caused by glucose (and glutamine) deprivation is pseudohypoxia due to PHD3 inhibition by EZH2, and Slug activation. Even if we have substantial evidence that αKG deficit is a major driver of starvation-induced dedifferentiation, we cannot exclude that other mechanisms could be involved in this phenomenon.

Our ChIP-seq analysis showed that glucose deprivation causes a massive repositioning of EZH2-dependent modulation of histone of trimethylation of H3 lysine-27 in the genome, with predominance of effect in intergenic regions. The role of this histone mark on intergenic regions is only starting to be elucidated. Some of these intergenic loci are associated with silencer elements; the role of the majority of these intergenic sites is still to be discovered (47). In addition, we observed that glucose depletion affects the trimethylation in different histone marks, such as H3K4 and H3K9, as well as H3K27 acetylation. This suggests that several histone alterations can be potentially involved in the observed phenotypic transition. Enrichment of H3K4me3 on a subset of Slug target genes associated with EMT poises these genes for activation, promoting a more aggressive and metastatic phenotype. Further studies to dissect the role of different histone modifications and DNA methylation in glucose-induced cell dedifferentiation are warranted.

We found that the HIF signaling pathway played a relevant role in glucose restriction-induced dedifferentiation. This is very interesting because HIF1α can drive dedifferentiation and cancer stem cell phenotype (56). Our data show that EZH2 activation represses PHD3 expression due to direct repression with recruitment of EZH2 on the PHD3 promoter. In addition, PHD3 regulates the stability of HIF1α protein by inducing its ubiquitination and degradation. However, our data show induction of HIF1α at both the RNA and protein levels, suggesting a transcriptional regulation of this gene by glucose restriction. HIF stabilization by hypoxia and by glucose deprivation occurs early within hours, without changes in HIF1α mRNA levels (57, 58). A short-term exposure to low glucose stabilized the HIF1α protein in an EZH2-independent manner, likely because the PHD proteins responsible for targeting HIFs for degradation are αKG-dependent enzymes (59, 60). Long-term regulation of HIF1α involves more complex combinations of transcriptional and posttranscriptional mechanisms, with the involvement of feed-forward loops and long noncoding RNA-mediated regulation (53). EZH2-mediated regulation of HIF1α observed in our studies is likely involved in long-term regulation of HIF1α expression, linking glucose starvation with changes in cell differentiation state.

In conclusion, we discovered a novel mechanism of glucose restriction–induced cancer dedifferentiation, mediated by unbalanced activity of histone methyltransferase EZH2 and activation of the hypoxia pathway and EMT, leading to an aggressive phenotype. These results should be considered in designing new anticancer metabolic therapies targeting glucose uptake and glycolysis.

Supplementary Material

Supplementary Data

List of Supplementary Data

Figure S1

Figure S1

Figure S2

Figure S2

Figure S3

Figure S3

Figure S4

Figure S4

Figure S5

Figure S5

Figure S6

Figure S6

Table S1

Table S1

Table S2

Table S2

Table S3

Table S3

Table S4

Table S4

Table S5

Table S5

Table S6

Table S6

Table S7

Table S7

Table S8

Table S8

Table S9

Table S9

Table S10

Table S10

Table S11

Table S11

Table S12

Table S12

Authors' Disclosures

J. Yanagawa reports grants from LUNGevity, personal fees from AstraZeneca, OncLive, and IDEOlogy outside the submitted work. S.M. Dubinett reports other support from EarlyDiagnostics and LungLife AI, Inc., outside the submitted work. C. Scafoglio reports a patent for combination metabolic-epigenetic treatment for early lung cancer pending. No disclosures were reported by the other authors.

Authors' Contributions

P. Saggese: Conceptualization, data curation, software, formal analysis, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. A. Pandey: Resources, data curation, validation, investigation, visualization, methodology. M. Alcaraz Jr: Investigation. E. Fung: Data curation, validation, investigation, visualization, methodology. A. Hall: Investigation. J. Yanagawa: Resources. E.F. Rodriguez: Formal analysis. T.R. Grogan: Formal analysis. G. Giurato: Data curation, software, formal analysis. G. Nassa: Data curation, formal analysis, investigation, methodology. A. Salvati: Data curation, formal analysis, investigation, methodology. O.S. Shirihai: Conceptualization. A. Weisz: Conceptualization, supervision, funding acquisition. S.M. Dubinett: Conceptualization, supervision. C. Scafoglio: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing.