KDM4 Inhibition Targets Breast Cancer Stem–like Cells

Breast cancer is the leading cause of cancer-related death among women worldwide (1). Among the subtypes of breast cancer, triple-negative disease is associated with a particularly poor prognosis and limited therapeutic options (2). During breast cancer treatment, therapy resistance and metastatic dissemination are the main problems that have to be faced (1). Of note, breast cancer stem cells (BCSC) have been suggested to be responsible for both therapy resistance and metastatic dissemination (3, 4). Until now, these resistant cancer stem cell populations have only been poorly characterized and targeted therapeutics have yet to be identified.

It has been shown that alterations of epigenetic regulators such as the KDM4 family members control tumor cell proliferation particularly in aggressive breast cancers (5) and dysregulation of KDM4 demethylases has been documented in a variety of cancers including breast cancer (6). The KDM4 subfamily is comprised of KDM4A, B, C, and D and belongs to the Jumonji C (JmjC) domain-containing family of histone demethylases (7). KDM4 demethylases catalyze removal of the repressive H3K9me3 mark and that of H3K36me3, a mark linked to transcriptional elongation (8), thereby regulating a range of crucial pathways. These findings highlight KDM4 demethylases as potential therapeutic targets for breast cancer treatment. Consequently, we set out to test potent and selective drug-like KDM4 inhibitors. In order to validate inhibitors on cancer stem cells (CSC) from triple-negative breast cancer (TNBC), we established an efficient three-dimensional (3D) cultivation method allowing for growth of CSCs from patient tumor tissue without prior fluorescence-activated cell sorting (FACS) or murine xenografts to enrich for CSCs. We used defined conditions including a serum-free culture medium, a rho kinase inhibitor, Matrigel, and a low oxygen environment to isolate and enrich for BCSCs from individual patient tumors after neoadjuvant chemotherapy.

All patients were operated at the Department of Obstetrics and Gynecology at the University Medical Centre Freiburg. Tumor tissue specimens for BCSC isolation and paraffin embedding were obtained from pathologists of the tumor bank of the Comprehensive Cancer Centre Freiburg. All experiments were performed in accordance with the Declaration of Helsinki. We confirm that all experimental protocols were approved by the Institutional Review Board in the Ethics vote 307/13 (independent Ethics Committee University of Freiburg). Written informed consent was obtained from each patient.

All primary breast cancer tumors were collected from individuals who had received chemotherapy and were classified as triple-negative. Primary BCSC lines were isolated by mechanical dissociation of the tumor material followed by enzymatic digestion in 5 mL DPBS (Gibco) supplemented with 6 units DNAse I (Machery-Nagel) and 1 mg liberase (Roche) for 1 h at 37°C. Afterward, the digestion medium was diluted with 10 mL DPBS and filtered through a cell strainer (40 μm, BD). Following centrifugation at 200 ×g for 5 minutes, the supernatant was discarded and the cell pellet was washed with MEBM (Gibco). Subsequently, if red blood cells were visible in the pellet, 2 mL ACK Lysis-buffer (Gibco) was added to the cell pellet. After 1 minute of incubation at room temperature, the suspension was filled up to 6 mL with MEBM and centrifuged at 200 ×g for 5 minutes. The supernatant was discarded and the pellet was resuspended in 1 mL MEBM and filtered through a 40-μm strainer. Following centrifugation at 200 ×g for 5 minutes, the supernatant was discarded and the remaining cell pellet suspended in MSC medium. Isolated cells were counted and 2 × 10⁴ cells in 200 μL of a 1:1 mixture of MSC medium and Matrigel (ice cold, Corning, 354230) were plated per well in a 24-well low attachment plate (Corning). After solidification of the Matrigel at 37°C for 30 minutes, each well was topped up with 500 μL of MSC medium. The cells were cultured at 37°C under low oxygen conditions (3% O₂, 5% CO₂, 92% N₂). 3D cells stably proliferating cells were cultured and expanded in 2D. All primary BCSC lines were isolated in 2014 and authenticated by the high-throughput Multiplex human cell authentification test (MCA) developed at the DKFZ in 2016 (9). Mycoplasma tests were conducted every 3 months via PCR detection utilizing the positive control provided free of charge by the Leibniz-Institute DSMZ. Experiments with cells were conducted in a passaging window of 15 passages.

The mammary stem cell (MSC) medium is composed of mammary epithelial basal medium (Gibco, 31331-028), supplemented with 1 × B27 (Gibco, 17504-044), 1 × amphotericin B (Sigma-Aldrich, A2942), and 1 × penicillin–streptomycin (Gibco, 15140122). Furthermore, epidermal growth factor (20 ng/mL, PeproTech, AF-100-15), heparin (4 μg/mL, Sigma-Aldrich, H3149), fibroblast growth factor (20 ng/mL, PeproTech, AF-100-18B), gentamicin (35 μg/mL, Gibco, 15750-045), and rho kinase inhibitor (500 nmol/L, Calbiochem, 555552) were added.

BCSCs were cultured as spheres in a 3D environment as described above. One milliliter medium was added after 2 days. Cells were split once a week using Dispase (Corning) to solve residual Matrigel and Accutase for sphere dissociation. To expand BCSCs in a 2D environment, 4 × 10⁵ cells were seeded in 2 mL MSC medium containing 2% Matrigel (ice-cold) in a 10 cm culture dish. After solidification of the Matrigel at 37°C for 30 minutes, the dish was topped up with 8 mL of MSC medium. Cells were grown under low oxygen conditions as described above. Medium was changed after 3 days. Cells were split once a week.

Cells were detached by Accutase and counted. A total of 3 × 10³ single BCSC1 and 1 × 10³ single BCSC2 cells were seeded into individual wells of 96-well ultra-low attachment plates (Corning, 3474) in serum-free MSC medium containing 1% methylcellulose (Sigma, M0512). After 7 days, all spheres were counted to evaluate the sphere-forming capacity except for experiments including QC6352 treatment, where only spheres over 50-μm diameter were counted.

Cells were detached by Accutase and counted. Single BCSC1 and BCSC2 cells (1 × 10³) were seeded as triplicates in 50% Matrigel into individual wells of 24-well ultra-low attachment plates (Corning) in serum-free MSC medium. QC6352 and QC6688 were dissolved in DMSO and paclitaxel in 0.9% saline solution. After 7 days, spheres over 50 μm diameter were counted for QC6352- and QC6688-treated and control cells and spheres over 20 μm diameter were counted for paclitaxel-treated and control cells. For the assessment of secondary sphere formation, wells with 4 × 10⁴ single BCSC1 and BCSC2 spheres were seeded and treated as described above. After 7 days, BCSC1 and BCSC2 cells were split and counted as described above. From these, 1 × 10³ single BCSC1 and BCSC2 cells were seeded in triplicates as described above to assess secondary sphere formation in the absence of QC6352, QC6688 or paclitaxel. After 7 days, spheres over 50 μm diameter were counted for QC6352- and QC6688-treated and control cells, and spheres over 20 μm diameter were counted for paclitaxel-treated and control cells.

High titer lentiviral stock (CMV-NLS-mCherry) was obtained from the Sanford Burnham Prebys Medical Discovery Institute. Lentiviral particles were added at a multiplicity of infection (MOI) of 5 to BCSC1 and BCSC2 in MSC medium. Cells were cultured as described. For cell proliferation assay, BCSC1mCherry and BCSC2mCherry were detached by Accutase and counted. 384-well plates (Greiner, 781091) were coated with 10 μL of MSC medium containing 2% of Matrigel (Corning, 354230). After incubation at 37°C for 30 minutes to solidify the Matrigel, 1 × 10³ single cells were seeded per well in 384-well. Inhibitor was added 24 hours later to the indicated final concentrations. Subsequently, daily cell number assessment for the indicated number of days was performed utilizing the ScanR microscope-based imaging platform (Olympus) and ScanR software 6. At each time point, mCherry-positive cell nuclei in 9 sectors of each well were assessed with a 10× lens._.

High titer adenoviral stocks (shRNA Ctrl (Ad-GFP-U6-scrambled-shRNA; #1122N), shRNA KDM4A (Ad-GFP-U6-h-KDM4A-shRNA; #shADV-212841), shRNA KDM4B (Ad-GFP-U6-h-KDM4B-shRNA; #shADV-212842), shRNA KDM4C (Ad-GFP-U6-h-KDM4C-shRNA; #shADV-212844), and shRNA KDM4D [Ad-GFP-U6-h-KDM4D-shRNA(#70); #shADV-212848)] were obtained from Vector BioLabs. Adenoviral particles were added at an MOI of 300 to BCSC1 cells and MOI of 150 to BCSC2 cells in MSC medium. Cells were harvested 3 days after infection for Western blot analyses and after 5 days for Chromatin immunoprecipitation (ChIP) and ChIP-seq analyses. For cell proliferation assays, BCSC1 were infected with an MOI of 300. After 5 days, BCSC1 were seeded into 384-well plates and reinfected with an MOI of 300. To generate xenografts, BCSC1 were infected with an MOI of 300. After 24 hours, BCSC1 were detached, reinfected with an MOI of 300 and 1 × 10⁵ cells transplanted into immunocompromised mice as described below.

ChIP experiments were performed essentially as previously described (10). BCSC1 cells were either cultured for 18 hours in the absence or presence of 5 × 10⁻⁸ mol/L QC6352 or cells were infected 5 days before harvesting with adenovirus expressing either shRNA against KDM4A or scrambled control shRNA (Ad-GFP-U6-hKDM4A-shRNA and Ad-U6-RNAi-GFP, respectively, Vector Biolabs) according to the manufacturer's instructions. Immunoprecipitation was performed with specific antibodies (anti-KDM4A (#5766, lot 021110, Schuele Laboratory), anti-H3K9me3 (Diagenode, #C15410056, lot A1675-001P), anti-H3 (#ab1791, lot GR300976-1, Abcam), spike-in antibody (#61686, Active Motif), rIgG (#C15410206, lot RIG001L, Abcam) in the presence of spike-in chromatin (#53083, Active Motif) on GammaBind G-Sepharose (GE-Healthcare). For PCR, 2 μL out of 70 μL isolated DNA were used. Primer sequences were as follows: EGFR 5′-cagagctcatcctggccaac-3′ and 5′-ttctgtctgcacacttggca-3′. Libraries were prepared from immunoprecipitated DNA according to standard methods. ChIP-seq libraries were sequenced using a HiSeq 2000 (Illumina) and mapped to the hg19 reference genome using bowtie 2 (11). Data were further analyzed with the peak finding algorithm MACS 1.42 (12) using input as control. Normalization to spike-in chromatin was performed according to Ref (13). All peaks with FDR greater than 0.5% were excluded from further analysis. The uniquely mapped reads were used to generate the genome-wide intensity profiles, which were visualized using the IGV genome browser (14). HOMER (15) was used to annotate peaks, to calculate overlaps between different peak files, and for motif searches. The genomic features (promoter, exon, intron, 3′UTR, and intergenic regions) were defined and calculated using Refseq and HOMER. Data are deposited under GSE95294.

All animal studies and experiments were performed in accordance with German Animal Welfare regulations and in accordance with an Institutional Animal Care and Use Committee as described in the animal protocol G13/114. NOD/SCID females (4–5 weeks old) were anesthetized using an isoflurane inhalator. A small sagittal incision (no longer than 1.0 cm) on the shaved and sterilized abdomen allowed access to the mammary gland #4 on both sides. Indicated numbers of BCSCs were mixed with 1 × 10⁶ irradiated fibroblasts [newborn human foreskin fibroblasts (NuFF), p11, GlobalStem, GSC-3002] and suspended in a 1:1 mixture of Matrigel (Corning, 354230) and MSC medium in a total volume of 40 μL per gland. The mixture was injected into the mammary fat pad of the #4 gland on both sides of the animal. Each transplant was localized distal to the lymph node in the gland. Surgical incisions were sealed by suturing with a 5/0 thread (Ethicon, Z995). Animals were monitored twice weekly for weight and tumor growth, which was determined by caliper measurement. Tumor volumes were calculated using the formula 4/3 × π × r³.

Ultrasound measurements of xenograft tumors in NOD/SCID mice were performed using a small animal high-resolution ultrasound system (Vevo3100) and transducer (MX550D) with 40 MHz (VisualSonics). For 3D tumor modeling, the transducer was moved along the tumor automatically with a step size of 0.076 mm. Tumors were visualized with Vevo LAB (Version 1.7.1) at start and end of treatment.

Immediately before treatment, QC6352 was dissolved in 50% polyethylene glycol (Sigma-Aldrich)/50% DPBS (pH 9, Gibco) with sonication (Diagenode bioruptor) until a clear solution was formed. When tumors reached a palpable size of 3 mm³, mice were treated with vehicle (control) or QC6352. The inhibitor was administered daily to mice via oral gavage at 10 mg/kg. Control animals received vehicle only. Animals were monitored twice weekly for weight and tumor growth.

Cells were detached by Accutase, dissociated and counted. The wells of a black 384-well plate (Greiner) were coated with 10 μL of MSC medium containing 2% of Matrigel (354230, Corning). After incubation at 37°C for 30 minutes, to solidify the Matrigel, 1 × 10³ single cells were seeded as described above. The inhibitor was added 24 hours later. Following 96 hours of incubation, cells were washed with DPBS and fixed with ice-cold methanol for 15 minutes at −20°C. Cells were washed with DPBS, stained with DAPI, and counted using the ScanR microscope-based imaging platform (Olympus).

Total RNA was isolated from patient tumor material, xenografts, and cells using the Universal RNA Purification Kit (Roboklon) according to the manufacturer's instructions. Isolated RNAs were processed with the Ambion WT Expression Kit (Ambion) as described by the manufacturer and hybridized to Illumina HT-12 v4 Expression Bead Chips following the standard Illumina protocols. Expression data were processed and quantile normalized using the R Bioconductor Beadarray package (16) in version 2.22. Only probesets mapping to an Entrez ID applying the Bioconductor package illuminaHumanv4.db (Version 1.26) were considered for further downstream analysis. In case of multiple probesets matching the same Entrez ID, we selected the probeset having the highest interquartile range across all samples. The dendrogram depicts a complete-linkage hierarchical clustering based on the Euclidean distance between the samples. Data are deposited under GSE95042.

To analyze the expression of established CSC markers, cells were detached and counted as described above. Cells (1 × 10⁵) were washed with staining buffer (DPBS + 1% BSA) and stained for 20 minutes at room temperature in the dark with the following antibodies diluted in staining buffer: anti-CD24 (eBioscience, 46-0247; 1:100), anti-CD44 (eBioscience, 12-0441-81; 1:1000), anti-EpCAM (eBioscience, 660 50-9326; 1:100), and anti-CD49f (eBioscience, 46-0495; 1:200). Cells were analyzed using BD LSR Fortessa and FlowJo software (Version 6).

Patient tumor tissue specimens were fixed in 10% formalin and embedded in paraffin. Two-μm-thick paraffin-embedded tissue sections were mounted onto glass slides. All slides were stored for two days at 58°C in a drying chamber, subsequently deparaffinized using xylene and hydrated with ethanol. Human and corresponding xenograft tumor tissue were stained using the following antibodies: anti-ER (clone EP1, code IR084, Dako); anti-PR (clone PgR 636, code IR068, Dako); anti-HER2 (code A0485, Dako); anti-Ki67 (clone MIB-1, code IR626, Dako); anti-vimentin (clone V9, code IR630, Dako); anti–E-cadherin (clone NCH-38, code IR059, Dako), and anti-cytokeratin 8/18 (clone EP17/EP30, code IR094, Dako). For the horseradish-based peroxidase detection EnVision Flex Peroxidase-Blocking Reagent (Dako, SM801), EnVision Flex+ Rabbit (LINKER; Dako, K8019) or EnVision Flex+ Mouse (LINKER; Dako, K8021) and EnVision Flex/HRP (Dako, SM802) were used. Counterstaining was performed with hemalum before adding a coverslip. As internal positive control, patient-derived physiological mammary gland was used for ER, PR, Ki67 (nuclear staining), cytokeratin 8/18 and E-cadherin (membranous, cytoplasmic staining). The mammary gland-surrounding myoepithelial layer was used as internal control for vimentin. For HER2, tissue specimens from HER2-positive breast cancer patients (score 3 according to ref. 17) were included for every HER2 staining session as external positive control. TNBC was defined as ER-, PR-, and HER2-negative (score < 2; ref. 18).

Experiments were performed as previously described (10). The following antibodies were used: anti-KDM4A (#5766, lot 021110, Schuele Laboratory), anti-KDM4B (#4662, lot 962009, Schuele Laboratory), anti-KDM4C (#23855, lot 23062015, Schuele Laboratory), anti-KDM4D (#ARP35946, lot 001, Aviva Systems Biology), anti-EGFR (#2232S, lot 16, Cell Signaling Technology).

Cells were cultured in the presence of vehicle or 5 × 10⁻⁸ mol/L QC6352. RNA was isolated as previously described (19). Quantitative RT-PCR was performed using the Abgene SYBR Green PCR Kit (Invitrogen) according to the supplier's protocol. HPRT was used for normalization. Primer sequences for HPRT were described previously (20). Other primers were as follows: VCAN: 5′-ACTGTGGATGGGGTTGTGTT-3′, 5′-CTGCGTCACACTGC TCAAAT-3′; PRR5: 5′-CGGGACAAGATTCGCTTCTA-3′, 5′-AGCGCATCC TCTAGCTTCAC-3′; ATF4: 5′-CCAACAACAGCAAGGAGGAT-3′, 5′-GTGTC ATCCAACGTGGTCAG-3′; EGR1: 5′-TGACCGCAGAGTCTTTTCCT-3′, 5′-CACAAGGTGTTGCCACTGTT-3′; FST: 5′-GGAAAACCTACCGCAATGAA-3′, 5′-GAGCTGCCTGGACAGAAAAC-3′; EGFR: 5′-CCAACCAAGCTCTCTTGA GG-3′, 5′-GCTTTCGGAGATGTTGCTTC-3′.

BCSC1 cells were cultured for 18 hours in the absence or presence of 5 × 10⁻⁸ mol/L QC6352. RNA was isolated as described above. RNA samples were sequenced by the standard Illumina protocol to create raw sequence files (.fastq files) at the sequencing core facility of the DKFZ. Reads were aligned to the hg19 build of the human genome using TopHat version 2 (21). The aligned reads were counted with the homer software (analyzeRNA) and DEG's were identified using EdgeR (22). Data are deposited under GSE95294.

An intravenous dosing solution was made by dissolving QC6352 into phosphate buffered saline. The pH was adjusted to 9 by dropwise addition of 1 N NaOH. An oral dosing suspension was made by dissolving QC6352 into 0.5% methylcellulose. The intravenous and oral dosing solutions were administered at 5 and 10 mg/kg, respectively, to female CD-1 mice. Pharmacokinetic parameters were calculated as the average of groups consisting of three animal.

Data are represented as mean and SD or SEM as indicated. Significance was calculated by a two-tailed Student t test or one-way ANOVA as indicated with GraphPad Prism Version 6. P values below 0.05, 0.01, and 0.001 are indicated in figures as *, **, and ***, respectively.

The ability of QC6352 to inhibit the activity of different KDM family members was determined in 384-well plate format under the following reaction conditions: 2 nmol/L enzyme, 300 nmol/L H3K9me3, H3K4me3, or H3K36me2 biotin-labeled peptides (Anaspec), 100 μmol/L alpha-ketoglutaric acid in assay buffer containing 50 mmol/L HEPES, pH 7.3, 0.005% Brij35, 0.5 mmol/L TCEP, 0.2 mg/mL BSA, 50 μmol/L sodium L-ascorbate, and 2 μmol/L ammonium iron(II) sulfate. Reaction product was determined quantitatively by TR-FRET after the addition of detection reagent phycolink streptavidin-allophycocyanin (Prozyme) and Europium-anti-H3K9me2, H4K4me2, or H3K36me1 antibody (PerkinElmer) in the presence of 5 mmol/L EDTA in LANCE detection buffer (PerkinElmer) at a final concentration of 50 nmol/L and 1 nmol/L, respectively. To initiate the assay reaction, 2 μL of 6 nmol/L enzyme and 2 μL of 11-point serially diluted QC6352 in 3% DMSO were added to each well for 60 minutes, followed by the addition of 2 μL of a mixture of 900 nmol/L biotin labeled peptide and 300 μmol/L α-ketoglutaric acid. After incubation at room temperature for 20 minutes, the reaction was terminated by addition of 6 μL of 5 mmol/L EDTA in LANCE detection buffer containing 100 nmol/L Phycolink streptavidin-allophycocyanin and 2 nmol/L europium-labeled antibody. After 60 minutes of incubation at room temperature, plates were read by EnVisionMultilabel Reader in TR-FRET mode (excitation at 320 nm, emission at 615 and 665 nm). A ratio was calculated (665/615) for each well and fitted to determine inhibition constant (IC₅₀).

As shown in Table 1, we established four BCSC lines (BCSC1-4), which originate from four independent breast tumor samples lacking estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2) proteins. BCSCs could be cultivated in a 2D and 3D environment and grow as epithelial clusters and spheres, respectively (Fig. 1A and Supplementary Fig. S1A). In vitro clonogenic assays such as sphere formation assays have been developed to study proliferation, self-renewal, and differentiation of cell populations at the single-cell level (23). Thus, to verify their stem cell potential BCSC1 and BCSC2 were challenged in an anchorage-independent growth assay. BCSC1 and BCSC2 demonstrated a sphere-forming capacity of 10.8% and 16.2%, respectively, providing evidence for the potential of BCSC1 and BCSC2 to self-renew (Fig. 1B and Supplementary Fig. S1B). CSCs from breast cancer have been described to express specific surface markers such as CD49f, EpCAM, and CD44 while lacking expression of CD24 (24, 25). We analyzed BCSC1 and BCSC2 for these stem cell markers and found 58% CD24⁻/CD44⁺ and 96% EpCAM⁺/CD49f⁺ cells in BCSC1 and 5% CD24⁻/CD44⁺ and 97% EpCAM⁺/CD49f⁺ cells in BCSC2 (Fig. 1C and D; Supplementary Fig. S1C and S1D). The flow-cytometric analysis clearly showed that BCSC1 and BCSC2 harbor a stem cell population. To further corroborate the stem cell potential of the BCSCs we performed limiting dilution orthotopic xenografts in immunocompromised NOD/SCID mice, the current gold standard assay for self-renewing CSCs. As few as 1 × 10³ BCSC1 and BCSC2 were able to form tumors demonstrating that both cell lines contain tumorigenic CSCs (Fig. 1E and F; Supplementary Fig. S1E). In summary, we used three independent methods namely sphere formation, flow cytometry analysis, and xenografts to demonstrate stemness, self-renewal capacity, and tumorigenicity of BCSC1 and BCSC2.

To further address the ability of these cells to regenerate the original patient tumor in a xenograft tumor model we analyzed tumors derived from BCSC1 and BCSC2 transplants histologically and genetically. Immunohistochemical analyses of BCSC1 and BCSC2 xenograft tumors using the mammary epithelial marker cytokeratin 8/18, E-cadherin, and vimentin as well as the proliferation marker Ki67 indicated that the xenograft tumors share a similar pattern with their parental patient tumors (Fig. 1G and Supplementary Fig. S1F). Matching the parental tumor, the BCSC xenografts were devoid of ER, PR, and HER2 protein expression (Table 1, Fig. 1H, and Supplementary Fig. S1G). Unsupervised hierarchical clustering analysis of RNA microarray data showed that the BCSC tumor xenografts share a close expression profile with the parental patient tumors, indicating preservation of the respective molecular tumor subtype (Fig. 1I). BCSC lines clustered within the corresponding host tumor and xenograft subtype depicting a close correlation between the three entities (Fig. 1I). Taken together, our data demonstrate that BCSCs can be isolated from TNBC patient tissue using optimized culture conditions and faithfully recapitulate the patient tumor in xenografts.

To evaluate the expression levels of the KDM4 family members in BCSC1 and BCSC2 we performed Western blot analysis. As shown in Fig. 2A, we detected robust expression of KDM4A in both BCSC1 and BCSC2. In contrast, expression levels of KDM4B, C, and D were more heterogeneous (Fig. 2A). To see whether KDM4s control proliferation of BCSC1 we performed adenoviral shRNA-mediated knockdown of KDM4A, B, C, and D in BCSC1 and monitored proliferation in real time. As shown in Fig. 2B, depletion of KDM4A impaired proliferation of BCSC1 cells. In contrast, knockdown of KDM4B, C, or D did not influence proliferation of BCSC1 (Supplemental Fig. S2A–S2E).

Next, we wondered whether knockdown of KDM4A affects growth of BCSC1 xenografts. Therefore, BCSC1 were infected with adenovirus encoding either shRNA control (shRNA Ctrl) or shRNA against KDM4A (shRNA KDM4A) and implanted into the fat pads of immunocompromised NOD/SCID mice. Importantly, upon knockdown of KDM4A, tumor growth, and final tumor weight of the BCSC1-derived xenografts was strongly reduced (Fig. 2C–E). In contrast, the control knockdown of KDM4D (shRNA KDM4D) in BCSC1 did not significantly affect tumor growth and final tumor weight of the xenografts (Supplementary Fig. S2F–S2H). Together, our data demonstrate that the histone demethylase KDM4A controls proliferation and xenograft tumor growth of BCSC1. Furthermore, our bioinformatic analyses indicate that relapse-free survival over time decreases for patients with TNBC that express high levels of KDM4A in comparison to patients with TNBC that express low levels of KDM4A (Supplementary Fig. S2I). These findings indicate that targeting KDM4 might be a therapeutic option to limit expansion of BCSC populations.

Based on our observations that KDM4A controls proliferation and xenograft tumor growth of BCSC1 we decided to test whether the drug-like KDM4 inhibitor QC6352 (also called compound 6; ref. 26) might qualify for the treatment of BCSC-originating tumors (Fig. 3A). As a first step, prior to any investigation of either biological function or underpinning mechanism of action, we assessed the selectivity profile of the inhibitor. QC6352 was evaluated against the KDM4 subfamily and other JmjC domain-containing histone demethylases such as KDM2A, 2B, 5B, 6A, and 6B. QC6352 inhibited the demethylase activity of KDM4A, B, C, and D at nanomolar concentrations with IC₅₀ values between 35 and 104 nmol/L (Fig. 3B). QC6352 showed 100-to 300-fold selectivity over KDM2 and KDM6 demethylases and only weak inhibition of KDM5B (Fig. 3B). Finally, QC6352's favorable pharmacokinetic properties warranted further in vivo investigation (Fig. 3C). In summary, QC6352 presented as an orally available, potent, and selective KDM4 inhibitor.

When investigating the in vitro efficacy of QC6352 in BCSCs we observed that concentrations as low as 10 nmol/L inhibited BCSC1 and BCSC2 cell proliferation (Fig. 3D and E; Supplementary Fig. S3A and S3B). To characterize and define the effect of QC6352 on stem cell potential (self-renewal and differentiation) we tested whether QC6352 treatment might interfere with the sphere-forming capacity of both BCSCs. Consequently, BCSCs were plated as single cells to ensure clonality and inspected for one week. In a concentration-dependent manner, QC6352 dramatically reduced the anchorage-independent sphere-forming capacity of BCSC1 and BCSC2 (Fig. 3F and Supplementary Fig. S3C). We then isolated single cells from 1 week QC6352-treated spheres and evaluated for anchorage-independent sphere formation in a secondary assay in the absence of QC6352. Of note, the secondary sphere-forming capacity was blocked even in the absence of inhibitor (Fig. 3G and Supplementary Fig. S3D). In contrast, the differentiation-inducing LSD1 inhibitor QC6688 (27) neither affected primary nor secondary sphere formation (Supplementary Fig. S3E and S3F). In addition, chemotherapeutics such as the taxane paclitaxel impaired proliferation of BCSC1 in primary sphere formation but did not interfere with stem cell potential and thus allowed secondary sphere formation (Supplementary Fig. S3G). Taken together, our data demonstrate that QC6352 blocks proliferation and self-renewal of BCSCs.

To unravel the molecular mechanism underlying the action of the inhibitor, we performed transcriptome analyses to identify genes that were differentially regulated upon QC6352 treatment. BCSC1 were cultivated in the presence or absence of QC6352 and subjected to RNA-seq. Our analysis identified a total of 580 differentially regulated genes (Fig. 4A). Among them, 254 were upregulated and 326 were downregulated (Fig. 4A). To see whether these genes are direct KDM4A targets, we performed ChIP-seq in BCSC1 cells with anti-KDM4A antibody. The analysis shown in Fig. 4B identified 46,822 high confidence KDM4A peaks. Motif analysis using HOMER indicates enrichment of FOXA3, FOXL2, FOXA1, FOXP1, or FOXF1 motifs at KDM4A locations (Supplementary Fig. S4A). Only 3,221 (6.9%) KDM4A locations were observed in BCSC1 treated with shRNA against KDM4A, thus confirming specificity of the KDM4A antibody (Supplementary Fig. S4B). This finding prompted us to intersect the KDM4A cistrome with the QC6352 transcriptome. Among the 580 differentially regulated genes, KDM4A was present at the promoter of 258 genes (44%; Fig. 4C). Pathway analysis for these 258 genes revealed that 30 of these genes, including EGFR, belong to the “EGF receptor signaling pathway” (Fig. 4D). We verified by qRT-PCR analysis that treatment with QC6352 reduced the expression levels of genes such as versican (VCAN), proline rich 5 (PRR5), activating transcription factor 4 (ATF4), early growth response 1 (EGR1), follistatin (FST), and importantly EGFR (Fig. 4E). EGFR is an emerging therapeutic target that is associated with poor clinical outcome of TNBC (28). To unravel the importance of EGFR signaling in growth of BCSCs we treated BCSC1 and BCSC2 with erlotinib, a specific EGFR inhibitor. Of note, treatment with erlotinib blocked proliferation of both BCSC1 and BCSC2 (Supplementary Fig. S4C–S4F). Furthermore, stem cell potential measured by anchorage-independent sphere-forming capacity of both BCSC1 and BCSC2 was dramatically reduced upon treatment with erlotinib (Supplementary Fig. S4G and S4H). Together, these data demonstrate that EGFR, at least in part, controls growth and stem cell potential of BCSCs. As shown by Western blot analysis, the protein levels of EGFR were reduced in both BCSC1 and BCSC2 upon treatment with QC6352 (Fig. 4F and Supplementary Fig. S4I). Because EGFR is a direct KDM4A target, we tested whether knockdown of KDM4A affects EGFR protein levels. As depicted in Fig. 4G and Supplementary Fig. S4J, shRNA-mediated knockdown of KDM4A led to reduced levels of EGFR in both BCSC1 and BCSC2. Of note, knockdown of KDM4B, C, or D in BCSC1 did not affect the levels of EGFR (Supplementary Fig. S4K–S4M). Taken together, the data show that EGFR expression in BCSCs is blocked by QC6352 via inhibition of KDM4A.

Because KDM4A is a demethylase that erases the repressive H3K9me3 mark, we hypothesized that an increase in H3K9me3 should be observed upon inactivation of KDM4A by QC6352. To this end, we performed genome-wide ChIP-seq with H3K9me3 antibody in BCSC1 cells that were treated with either vehicle or QC6352. To allow for normalization of H3K9me3 ChIP-seq tags, we added spike-in chromatin during the ChIP procedure (13). The analysis shown in Supplementary Fig. S5A identified 141,722 high confidence H3K9me3 peaks in vehicle-treated cells and 144,266 peaks in cells treated with QC6352. Importantly, we observed a global increase of the H3K9me3 reads over the KDM4A peaks (Fig. 5A). Similarly, on the EGFR promoter we also observed increased levels of the repressive H3K9me3 mark over the KDM4A peak subsequent to inactivation by QC6352 (Fig. 5B). To demonstrate that the increase of H3K9me3 levels at the EGFR promoter upon treatment with QC6352 is due to inhibition of KDM4A, we infected BCSC1 cells with adenovirus encoding either shRNA control (shRNA Ctrl) or shRNA against KDM4A (shRNA KDM4A) and performed ChIP assays. ChIP-qPCR analyses indicate that in BCSC1 infected with shRNA KDM4A the levels of H3K9me3 at the EGFR promoter increase concomitantly to a decrease in KDM4A when compared to control cells (Fig. 5C). In summary, these results demonstrate that treatment of BCSCs with QC6352 targets EGFR via inhibition of the KDM4A demethylase activity.

Next, we wondered whether QC6352 might affect growth of BCSC1 and BCSC2 xenografts. Therefore, 1 × 10⁵ cells were implanted into the fat pads of immunocompromised NOD/SCID mice. To mimic the clinical treatment situation of a preexisting tumor we allowed growth of the xenografts to a size of 3 mm³ before starting treatment. Mice carrying tumors were then treated with QC6352 for 21 days at 10 mg/kg per os. Importantly, QC6352 strongly affected tumor growth and final tumor weight of both BCSC1 and BCSC2 xenografts (Fig. 6A–E and Supplementary Fig. S6A–S6E). Treatment with QC6352 was well tolerated and did not affect body weight of the mice (Supplementary Fig. S6F and S6G). Furthermore, we analyzed whether treatment with QC6352 affected expression of KDM4A target genes in the BCSC1 xenograft tumors in a similar manner as observed in cell culture (Fig. 4E). As displayed in Fig. 6F, expression of VCAN, PRR5, ATF4, EGR1, FST, and EGFR was similarly affected in BCSC1 xenograft tumors of QC6352-treated mice as in vitro. Taken together, our data demonstrate that treatment with the KDM4 inhibitor QC6352 blocks BCSC xenograft tumor growth.

In recent years, CSCs from solid tumors were identified using surface marker profiles and subsequent limiting dilution orthotopic xenografts. The first description of BCSCs by Al-Hajj and colleagues identified that the BCSC population is marked by CD24^low/CD44^high cells (24). In addition, several other surface molecules such as CD61, CD49f, and EpCAM identifying BCSCs were proposed (29, 30). Importantly, due to a lack of appropriate culture conditions BCSCs cannot be cultivated continuously in vitro in a stem cell state (31, 32). To address this issue and to stabilize the CSC phenotype, we used a 3D Matrix and applied a Rho Kinase inhibitor in a low oxygen environment. Thus, our culture conditions allowed cultivation of BCSCs directly from the patient tumor without the need for initial separation of tumor tissue and CSCs by FACS or the expansion of the patient's tumor tissue in murine xenografts. Limiting dilution transplants of our cultivated BCSCs regenerated the original patient's tumor. The xenograft tumors had an almost identical immunohistochemical pattern compared with the tumor of origin, which is comparable to results obtained from direct patient xenografts (24, 33). In addition, transcriptome analyses uncovered a strong clustering of primary tumors, xenografts, and cell lines. Thus, BCSCs faithfully reproduce the original patient's tumor and are therefore an ideal cellular platform to test novel therapeutics.

Because breast cancer progression was shown to be associated with alterations of KDM4 demethylase family members (8, 34–36), we set out to characterize the expression profiles of the four KDM4 family members, namely KDM4A, B, C, and D in BCSC1 and BCSC2. BCSC1 and BCSC2 are characterized by a robust expression of KDM4A. In contrast, expression levels of KDM4B, C, and D were more heterogeneous. While depletion of KDM4A impaired proliferation of BCSC1 and blocked growth of BCSC1 tumor xenografts, knockdown of the other KDM4 family member neither affected cell proliferation nor xenograft tumor growth. These findings are in accordance with previous studies suggesting that KDM4A controls proliferation of breast cancer cells (37, 38). Furthermore, by showing for the first time that KDM4A controls proliferation and self-renewal of BCSCs isolated from chemotherapy-resistant triple-negative breast tumors we validate KDM4A as a therapeutic target.

In recent years diverse KDM4 inhibitors have been identified, which act either as α-ketoglutarate mimics, a cofactor essential for the enzymatic function of KDM4s, or as inhibitors of the catalytic site (8). These molecules showed inhibitory effects on KDM4s in vitro and in cell culture models (8). However, these KDM4 inhibitors did not qualify as therapeutic agents. In contrast, QC6352 is a drug-like KDM4 inhibitor that is potent, selective, orally available and presents favorable pharmacokinetic properties.

In summary, we established an advanced culture method that allows isolation and growth of BCSC lines isolated from individual patient tumors after neoadjuvant chemotherapy. We demonstrate that BCSC xenografts faithfully recapitulate parental patient tumors and that BCSCs, BCSC xenografts, and the parental tumors share a highly similar transcriptome and phenotype profile. Therefore, our models are ideal tools for the identification and validation of novel therapeutics. In line with this idea, we identified the histone demethylase KDM4A as a therapeutic target for BCSC-originating tumors. Consequently, we showed that the orally available, potent, and selective KDM4 inhibitor QC6352 abrogates expression of target genes via inhibition of the KDM4A demethylase activity, thereby blocking proliferation, sphere-forming capacity in vitro, and xenograft tumor growth of BCSCs in vivo. Thus, modulation of KDM4 activity is a promising therapeutic strategy for the treatment of TNBC.

R. Schüle is a consultant/advisory board member for Celgene. No potential conflicts of interest were disclosed by the other authors.

Conception and design: E. Metzger, J. Xu, M.B. Wallace, J. Maurer, R. Schüle

Development of methodology: E. Metzger, S.S. Stepputtis, B.-T. Preca, M. Follo, T. Kanouni, J. Maurer, R. Schüle

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): E. Metzger, S.S. Stepputtis, J. Strietz, D. Willmann, A. Allen, P. Bronsert, E. Stickeler, J. Maurer, R. Schüle

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): E. Metzger, S.S. Stepputtis, J. Strietz, B.-T. Preca, D. Willmann, N. Iovino, M. Follo, M. Boerries, J. Xu, J. Maurer, R. Schüle

Writing, review, and/or revision of the manuscript: E. Metzger, S.S. Stepputtis, J. Xu, M.B. Wallace, J.A. Stafford, T. Kanouni, J. Maurer, R. Schüle

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): E. Metzger, S. Urban, D. Willmann, F. Zenk, A. Proske, E. Stickeler, J. Maurer, R. Schüle

Other (design and synthesis of QC6352): M.B. Wallace

We thank Hauke Busch for expert advice and help with bioinformatics analyses. Furthermore, we thank Robert Oshima and Pedro Aza-Blanc for providing reagents and the ZTZ Core Facility for expert advice and technical support. We are obliged to Astrid Rieder, Jessica Pfannstiel, Lisa Füner, and Bettina Herde for providing excellent technical assistance.

This work was supported by grants of the European Research Council (ERC AdGrant 322844) and the Deutsche Forschungsgemeinschaft SFB 992, 850, 746 to R. Schuele and by Deutsche Forschungsgemeinschaft SFB 850 and funding by DKTK to J. Maurer and M. Boerries. M. Boerries is funded by the German Federal Ministry of Education and Research within the framework of the e:Med research and funding concept, DeCaRe (FKZ 01ZX1409B).

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