Fatty acid oxidation is a druggable gateway regulating cellular plasticity for driving metastasis in breast cancer.

Cell state transitions control the functional behavior of cancer cells. Epithelial-to-mesenchymal transition (EMT) confers cancer stem cell-like properties, enhanced tumorigenicity and drug resistance to tumor cells, while mesenchymal-epithelial transition (MET) reverses these phenotypes. Using high-throughput chemical library screens, retinoids are found to be potent promoters of MET that inhibit tumorigenicity in basal-like breast cancer. Cell state transitions are defined by reprogramming of lipid metabolism. Retinoids bind cognate nuclear receptors, which target lipid metabolism genes, thereby redirecting fatty acids for β-oxidation in the mesenchymal cell state towards lipid storage in the epithelial cell state. Disruptions of key metabolic enzymes mediating this flux inhibit MET. Conversely, perturbations to fatty acid oxidation (FAO) rechannel fatty acid flux and promote a more epithelial cell phenotype, blocking EMT-driven breast cancer metastasis in animal models. FAO impinges on the epigenetic control of EMT through acetyl-CoA-dependent regulation of histone acetylation on EMT genes, thus determining cell states.

INTRODUCTION

Breast cancer is a heterogeneous disease that is classically stratified by receptor expression status; this necessitates different treatment strategies to achieve optimal disease control (, ). While estrogen receptor–positive (ER+) and ERRB2/HER2-positive tumors typically respond well to targeted therapies, the triple-negative breast cancer (TNBC) subtype, as defined by the absence of ER, progesterone receptor, and HER2 expressions, lacks targeted approaches and remains refractory to standard-of-care chemotherapy (, ). Limited progress has been made in carving new adjuvant regimens in the clinical management of TNBCs, and phase 2 and 3 trials for therapeutic agents against TNBCs have been short of promising (, ). This underscores the need to explore more effective therapeutic strategies for TNBCs. The activation of epithelial-mesenchymal transition (EMT) program in cancer cells has often been attributed to the gain-of-resistance phenotypes (), in part through the acquisition of cancer stem cell–like properties (–). Several studies have conceptually suggested that the reversal of the EMT program, i.e., a mesenchymal-epithelial transition (MET), may cause cancer cells to regain sensitivity to chemotherapy (, ).

While receptor status has been one of the mainstays of clinical classification, prognosis, and management decisions, gene expression profiling has been applied to categorize breast cancer into several “intrinsic” molecular subtypes: luminal A, luminal B, HER2 enriched, basal-like, and claudin low (). Although there remains some debate, it has been observed that ER+ tumors tend to be luminal-like and express “epithelial” marker genes such as E-cadherin, whereas TNBCs bear a basal-like gene signature and express “mesenchymal” markers that include claudins and vimentin (, ). During cell state transitions, whether the gain of epithelial or mesenchymal features is connected to luminal- or basal-like cancer, respectively, remains unclear. It may be possible that the MET, beyond changes in phenotypic cell states, may be capable of converting basal-like breast cancer into one that is more luminal-like, characterized by the expression of ER or luminal cytokeratins.

The EMT program is well established to be regulated by alterations in extracellular and intracellular signaling pathways. For instance, extrinsic signals such as transforming growth factor–β (TGFβ) and platelet-derived growth factor lead to the activation of intracellular signal transduction cascades that ultimately result in the activation of EMT-associated transcription factors such as TWIST, Snail, Slug, and ZEB1 (). In contrast, pathways that promote the MET appear more diverse and poorly defined. For example, forskolin-induced MET acts through the activation of cyclic adenosine 3′,5′-monophosphate (cAMP) pathway that results in epigenetic reprogramming (, ), while TGFβ inhibitors have been reported to mediate MET in germline stem cells (). In both instances, the precise mechanisms remain elusive. Furthermore, although it is, in principle, possible that the biochemical or genetic modulation of potential pathways such as cAMP activation or TGFβ receptor inhibition may help promote a MET, their potential for clinically relevant applications may not be forthcoming.

While the signaling pathways and transcription regulators that control cell states are well understood, other emerging aspects, particularly metabolism that contributes toward cancer cell function, remain far less understood. Altered metabolism has reemerged as a hallmark of tumors, underscoring the role of metabolic reprogramming in regulating tumor cell behavior, as well as their therapeutic implications (, ). Numerous studies have focused on evaluating the metabolic differences between cancer and normal cells, noting the heightened demand of glucose channeled to glycolysis that drives cancer cell proliferation (–). Recent advances have highlighted the distinct cancer metabolic dependencies on diverse metabolites such as glutamine, lactate, and methionine for growth and tumor initiation (–). Concurrently, lipids represent yet another class of metabolites that have gained attention for their pivotal influence on tumor progression. Fatty acid transporter, CD36, has been reported to contribute to tumor initiation (), and a myriad of studies have expounded on the role of lipogenesis via fatty acid synthase to be vital to tumorigenesis (, ). Palmitate channeled to an alternative fatty acid desaturation pathway involving fatty acid desaturase 2 promotes tumor plasticity (). Nonetheless, the mechanistic manner by which fatty acid oxidation (FAO) contributes toward cell state plasticity to drive metastatic progression remains unclarified. More recently, a high-resolution metastasis map further uncovered lipid metabolism to be a feature of metastases, although the mechanistic underpinnings and the connection to cell states are not clear (). Given the phenotypic diversity and plasticity of cancer cells, there remains much to be explored regarding the association between rewiring fatty acid metabolism and the precise control of cancer cell plasticity (, ). The shifts of cancer cells between cell states ostensibly necessitate different energetic requirements to support these functions.

RESULTS

High-throughput small-molecule library screens identify retinoids as promoters of the MET

The acquisition of a mesenchymal cell state in breast cancer is often associated with more aggressive and invasive phenotypes that are also chemoresistant and cancer stem cell–like (, , ). We reasoned that the conversion of a mesenchymal to a more epithelial cell state through the process of a MET may reduce the tumorigenicity and chemoresistance of cancer cells and improve therapy outcomes. To identify small-molecule compounds that have the ability to promote MET, we used a mesenchymal breast cell line, NAMEC8, which is responsive to perturbation in cell states, coupled to promoter firefly luciferase reporters as readouts for cell state changes (Fig. 1A). We tested several gene promoters that include CD24, CD44, CD166, CDH1, ZEB1, and SERPINE1, which have been previously associated with either an epithelial or mesenchymal cell phenotype. From other studies, TGFβ was shown to promote EMT, whereas the inhibition of the TGFβ pathway could either block EMT or moderately induce a MET-like phenotype. Thus, we used the TGFβR inhibitor, A83-01, as a positive control to evaluate the responsiveness of these gene promoters to changes in cell states (fig. S1A). A z score was calculated on the basis of the responses to A83-01; SERPINE1prom-luc had the lowest z score, giving a clear distinction of cell state changes (fig. S1B).

Fig. 1.

High-throughput chemical library screens identify retinoids as promoters of the MET.

(A) Schematic showing generation of reporter cell lines for MET screens using small-molecule compound libraries. (B) Dual-Glo Luciferase assay of 41 compounds performed at five-point dose response. Compounds are classified according to physiological function into nine classes. Data are represented as mean FF/Ren (n = 3). Gray boxes represent doses that resulted in <50% cell viability from DMSO. (C) Protein expression of EMT and lineage markers in NAMEC8 cells treated with increasing concentrations of bexarotene (left) or TTNPB (right). (D) Immunofluorescence staining of EMT and lineage markers in DMSO- versus bexarotene-treated NAMEC8 cells. Images are taken at ×100 magnification. Scale bars, 50 μm. (E) Mammosphere-forming assay of DMSO- versus bexarotene-, TTNPB-, or A83-01–treated NAMEC8 cells. Images were taken at ×50 magnification. Scale bars, 100 μm (left). Bar graph (right) shows number of spheres >100 μm. (F) Protein expression of EMT and lineage markers in BC2.2, BC22.1, BC29.1, MetBr-007 DR, BC24.1, and BC33.1 cells: DMSO versus bexarotene treatment. (G) Protein expression of EMT and lineage markers in NAMEC8 Ctrl shRNA and shRXRα cells treated with DMSO or bexarotene. (H) Protein expression of EMT and lineage markers in NAMEC8 cells treated with DMSO or bexarotene, in the presence or absence of 0.2 μM HX531. (I) Protein expression of EMT and lineage markers in NAMEC8 cells treated with DMSO or bexarotene, in the presence or absence of 0.1 μM AGN193109. (J) Dose-response graphs (top) and IC50 values (bottom) of DMSO- versus bexarotene-treated NAMEC8 cells to varying concentrations of 5-FU. (K) Dose-response graphs (top) and IC50 values (bottom) of DMSO- versus bexarotene-treated SUM159 PACR cells to varying concentrations of 5-FU. Graphs represent data as means ± SEM (n = 3). **P < 0.01,***P < 0.001.

We screened four chemical compound libraries [LOPAC (Library of Pharmacologically Active Compounds), anticancer, kinase inhibitor, and bioactive lipids] containing >2700 compounds to identify classes of molecules with MET-promoting properties (fig. S1C). Notably, in the mesenchymal NAMEC8 SERPINE1prom-Luc reporter cells, we further incorporated a Renilla luciferase reporter that was under the control of a constitutive cytomegalovirus (CMV) promoter as a readout for cell numbers. The purpose was to identify compounds that alter cell states without adversely affecting cell viability (Fig. 1A). From this primary screen, the “hit rate” was 2.19% (61 compounds). Z score was also calculated and plotted (fig. S1D). These were validated in a secondary screen and dose-response measurements, demonstrating a concordance rate of 83.6% (51 compounds), thus underscoring the robustness of the reporter-based screen (fig. S1E and table S1). At least eight classes of compounds were represented, including TGFβ inhibitors, which we earlier used as a positive control. Compounds such as adenosine/adrenergic, γ-aminobutyric acid (GABA) and N-methyl-d-aspartate (NMDA) receptor agonists and antagonists, and some neurotransmitters, which have not been previously implicated in cell state transitions, were identified as potential regulators. Among the classes, retinoids appeared to induce the most marked reductions in the SERPINE1prom-Luc activity, doing so in a dose-dependent manner (Fig. 1B).

We chose to focus on retinoids as MET-inducing agents for several reasons. First, they represent the most potent class of MET-inducing compounds in our screens. Second, they are clinically viable agents that are being used in acute promyelocytic leukemia (APL) as differentiating agents, although their utility and efficacy toward other cancers remain unclear (–). While bexarotene has been shown to inhibit cancer metastasis and angiogenesis, the mechanistic underpinnings by which retinoids induce changes in cell states or promote differentiation beyond direct transcriptional regulation have not been well elucidated (). For validation studies, as well as biochemical and pathway characterization, we chose to use a RAR (retinoid acid receptor) agonist (TTNPB) and a retinoic X receptor (RXR) agonist (bexarotene) as these induced the greatest down-regulations of the SERPINE1prom-Luc activity in a dose-dependent manner (Fig. 1B and table S1). TTNPB is an analog of retinoic acid that potently and selectively activates RARs [median effective concentration (EC50) = 5.1, 4.5, and 9.3 nM for RARα, RARβ, and RARγ, respectively], while bexarotene is a selective RXR agonist (EC50 values are 33, 24, and 25 and 33 nM for RXRα, RXRβ, and RXRγ, respectively) (, ). TTNPB and bexarotene have been used in mouse studies of mammary tumorigenesis, in which the agonists demonstrated retinoid receptor selectivity (). Consistent with the down-regulation of the SERPINE1prom-Luc reporter activity, which marked a loss of a mesenchymal cell state, treatment of NAMEC8 and another mesenchymal TNBC cell line, SUM159, with bexarotene or TTNPB caused a marked change in morphology from scattered, spindle, and fibroblastic-like cells to an epithelial-like morphology, characterized by tight clusters of epithelial islands (fig. S1F). This morphological change was associated with the dose-dependent down-regulation of mesenchymal signature markers (Zeb1, Fibronectin, SERPINE1, Slug, and Snail) and the concomitant gain of epithelial marker E-cadherin (Fig. 1, C and D, and fig. S1, F to H). We noted a switch in cytokeratin expression from KRT5, typically associated with basal-like TNBC, to KRT8/18, which is expressed in the luminal ER+ subtype of breast cancer. Mucin1 (MUC1), a characteristic marker of luminal breast cancer not commonly found in TNBC, was up-regulated as well. These data provided the first indications that beyond a MET, these basal-like TNBC cells might have gained luminal-like features. Because the mesenchymal phenotype has been shown to be associated with a cancer stem cell-like state, we examined the expression of CD44 and mammosphere-forming ability of the retinoid-treated cells. Following bexarotene or TTNPB treatment, both NAMEC8 and SUM159 cells showed significant reductions in CD44 expression and decreased mammosphere-forming capabilities without affecting in vitro proliferation (Fig. 1E and fig. S1, I to K). TGFβ inhibition by A83-01 was observed to induce a weaker MET, as seen by lesser reductions in mesenchymal EMT markers, CD44 expression, and mammosphere formation (Fig. 1E and fig. S1, I to L).

To further assess the effects of retinoids beyond classic cancer cell lines, which tend to be homogeneous, we generated a panel of patient-derived tumor cell lines from biopsied or resected triple-negative and ER+ breast tumors (table S2). Consistent with our previous findings, bexarotene treatment of the patient-derived TNBC cell lines (BC2.2, BC22.1, BC29.1, and MetBR 007-DR) increased the expression of E-cadherin, while reducing Slug expression, to varying extents. We continued to observe the switch in cytokeratin expression from KRT5 to KRT8/18, along with the gain in MUC1 expression (Fig. 1F and fig. S1, M and N). Mammosphere-forming abilities of patient-derived cancer cells were also decreased upon retinoid treatment (fig. S1O). Unexpectedly, in three of these tumor cell lines, we noted the up-regulation of ERα expression, a hallmark of ER+ luminal breast cancer. Estrogen response element (ERE)–luciferase reporter was also increased upon treatment with bexarotene or TTNPB in NAMEC8 cells (fig. S1P). In support of this, ER target genes such as ABCA3, NRIP1, TFF1, and CREB1 were strongly elevated by bexarotene treatment in both NAMEC8 and BC2.2 cells (fig. S1Q). Since the otherwise basal-like cancer cells had now gained ERα expression, we wondered whether they also acquired sensitivity to anti-ER therapy that is administered for patients with breast cancer harboring ER+ tumors. Bexarotene-treated cancer cells appeared to have gained sensitivity to low doses of the selective ER modulator, raloxifene hydrochloride (fig. S1R). Nonetheless, we noted that the difference in drug sensitivity was a moderate twofold, thus suggesting that the conversion from a basal to a truly luminal phenotype may be incomplete. When bexarotene was applied to the patient-derived tumor cell lines (BC24.1 and BC33.1) generated from ER+ luminal-like breast tumors, there was no marked change in the expression of the EMT/lineage markers (Fig. 1F and fig. S1M). Immunofluorescence staining of E-cadherin in an ER+ patient-derived cell line (BC24.1) treated with bexarotene was carried out, which showed that expression levels and membrane localization of E-cadherin were not altered significantly (fig. S1S). This suggested that retinoids exert their effects in mesenchymal-like but not epithelial-like breast cancer cells.

Next, we wanted to establish that bexarotene-induced MET is an on-target effect of RXRα signaling. A short hairpin RNA (shRNA) targeting RXRα was used to achieve a stable knockdown in NAMEC8. The gain in epithelial-luminal markers and loss of mesenchymal-basal markers induced by bexarotene were now decreased in the RXRα-knockdown cells (Fig. 1G and fig. S1T). We also used a RXRα-specific antagonist (HX531) and observed a similar abrogation in retinoid-induced MET, complemented by a restoration of mammosphere-forming ability and a mesenchymal-like cell morphology (Fig. 1H and fig. S1, U and V). Taking into account the dimerization of RXR and RAR, as well as having observed a similar MET induction with RAR agonists, we used a RAR receptor antagonist (AGN193109) and observed a similar blockade of retinoid-induced MET, demonstrating that retinoid-induced MET is mediated by cooperativity of both RXR and RAR receptors (Fig. 1I and fig. S1, U and V). Proliferation rates of the drug-treated cells were similar with control-treated cells (fig. S1V). Together, these confirmed that retinoid-induced cell state changes were driven by on-target retinoid receptor signaling.

The gain of a mesenchymal cell phenotype through EMT confers resistance to broad chemotherapy. We reasoned that the reverse phenomenon, a MET, may sensitize otherwise mesenchymal cancer cells to chemotherapeutic drugs, thereby overcoming resistance. Not unexpectedly, in both mesenchymal-like cancer cell lines (NAMEC8 and SUM159) and patient-derived line BC2.2, exposure to bexarotene was able to confer susceptibility to 5-flurouracil (5-FU) or paclitaxel, both standard of care for TNBC, by at least four- to sixfolds (Fig. 1J and fig. S1, W and X).

To evaluate this in the preclinical setting, we generated paclitaxel-resistant (PACR) SUM159 cells, which had a >20-fold increase in paclitaxel-resistance compared to the parental cells (fig. S1Y). Upon treatment with bexarotene, SUM159 PACR cells switched from a mesenchymal-like to an epithelial-like phenotype, which was accompanied by a reduction in mammosphere-forming ability, suggesting a decrease in cancer stem cell–like property (fig. S1Z). They also gained sensitivity to an alternative chemotherapeutic, 5-FU, thus indicating that drug-resistant cancer cells may respond to such a therapeutic strategy (Fig. 1K). In the context of breast cancer patient management, the use of retinoids as maintenance therapy in combination with chemotherapy as a first line of treatment in treatment-naïve patients may thus be a potential treatment strategy. This is currently being tested in our ongoing clinical trial (ClinicalTrials.gov identifier: NCT04664829).

Retinoid administration promotes MET in vivo and reduces the tumorigenicity of TNBC cells in animals

We reasoned that the conversion from a mesenchymal to an epithelial cell state may cause basal, mesenchymal-like breast cancer cells to lose tumorigenicity. Ex vivo treatment of SUM159 cells with bexarotene or TTNPB severely crippled their tumorigenic potential following transplantation into immunocompromised NSG [NOD-scid IL2Rgamma(null)] mice (Fig. 2A and fig. S2A). When a luminal, epithelial-like breast cancer cell line, MCF7-Ras, was similarly exposed to bexarotene or all-trans retinoic acid (ATRA), no impact on tumor growth was observed (Fig. 2B and fig. S2B). Notably, the reductions in tumorigenicity were not attributed to the effect of retinoids on cell proliferation as ex vivo treated cells maintained similar growth capacity with control cells (fig. S2C). This clearly underscores the utility of retinoid-induced cell state changes in basal but not luminal breast cancers.

Fig. 2.

Retinoid therapy promotes MET and reduces the tumorigenicity of TNBC cells in animals.

(A) Tumor sizes of SUM159 pretreated with DMSO, bexarotene, or TTNPB before subcutaneous injection into NSG mice (top). Data are represented as mean diameter ± SEM, n = 6 for DMSO, n = 8 tumors for bexarotene and TTNPB. Representative images of tumors from mice sacrificed at week 7 after implantation (bottom). (B) Tumor sizes of MCF7-Ras pretreated with DMSO, bexarotene, or ATRA before subcutaneous injection into NSG mice (top). Data are represented as mean diameter ± SEM (n = 4 for DMSO and n = 6 tumors for bexarotene and ATRA). Representative images of tumors from mice sacrificed at week 9 after implantation (bottom). (C) Top: Schematic of workflow for mouse xenograft experiment. Middle: Tumor sizes in control mice versus bexarotene-fed mice from NAMEC8-Ras (left) and BC2.2 (right). Data are represented as mean diameter ± SEM [n = 6 for control, n = 4 for bexarotene (NAMEC8-Ras), n = 7 for control and bexarotene (BC2.2)]. *P < 0.05 and **P < 0.01. Bottom: Representative images of NAMEC8-Ras and BC2.2 tumors formed in control mice versus bexarotene-fed mice. (D) Hematoxylin and eosin (H&E) staining of NAMEC8-Ras tumors formed in control mice versus bexarotene-fed mice. Images were taken at ×100 magnification. Scale bars, 100 μm. (E) Immunofluorescence staining of EMT and lineage markers in NAMEC8-Ras tumor sections from control mice versus bexarotene-fed mice. Images are taken at ×200 magnification. Scale bars, 50 μm. (F) Immunofluorescence staining of EMT and lineage markers in BC2.2 tumor sections from control mice versus bexarotene-fed mice. Images are taken at ×200 magnification. Scale bars, 50 μm. (G) Gene expression of retinoic acid response genes in NAMEC8-Ras (top) and BC2.2 (bottom) tumors formed in bexarotene-fed mice relative to tumors in control mice. Graphs represent data as means ± SEM (n = 3). *P < 0.05, **P < 0.01, ***P < 0.001, and ****P < 0.0001. Photo credit: Ser Yue Loo, Genome Institute of Singapore.

To evaluate the efficacy of retinoids in tumor-bearing animals, we first transplanted NAMEC8-Ras cells and two other patient-derived TNBC cell lines (BC2.2 and BC22.1) into NSG mice and allowed the tumors to reach ~3 mm in diameter by week 4. One group was dosed with bexarotene and the other group with vehicle control via oral gavage daily for the duration of the study. Bexarotene-fed tumor-bearing mice showed significant reductions in tumor growth rate and burden compared to vehicle-treated mice (Fig. 2C and fig. S2, D and E). To understand the cause of this growth reduction, we harvested residual tumors for analyses. Pathological analysis of hematoxylin and eosin (H&E) staining revealed control tumors to be grade 3 carcinoma with ductal features, whereas tumors from bexarotene-treated mice were discohesive and splayed apart. The latter also showed the presence of spindled cells with focal necrosis and fibrovascular cores, suggesting possible papillary or pseudopapillary components (Fig. 2D and fig. S2F), thus indicating a possible shift in cell states. To confirm this, we performed immunofluorescence staining of EMT markers. Tumors from bexarotene-treated mice now expressed E-cadherin and lost Fibronectin, indicating that MET had occurred in vivo as a result of retinoid treatment. Furthermore, there was a switch in cytokeratin expression from the basal-associated KRT5 to the luminal-associated KRT8/18 in these tumors (Fig. 2, E and F). Since retinoids activate target genes such as STRA6, CRAPBP2, DHRS3, ANGPTL4, IP6K3, and ACSL5 through RXR and RAR receptors, which bind DNA, we analyzed their expressions in tumors of treated animals. All target gene expressions were up-regulated by at least twofolds (Fig. 2G), confirming that retinoid treatment was associated with on-target RXR- and RAR-mediated gene expression changes in tumors.

Lipid metabolism genes are regulated by retinoid receptors

RAR and RXR are DNA binding transcription factors and exert their influence through modulating target gene expression in the presence or absence of retinoids. Reported pathways associated with retinoid signaling include TGFβ, NFκB (nuclear factor κB), and Nrf2 (–). To understand the manner by which RAR and RXR reprogram cell state transitions, we performed chromatin immunoprecipitation sequencing (ChIP-seq) to identify their compendium of target genes in an unbiased manner. We exposed the mesenchymal NAMEC8 and SUM159 cancer cells to short-term (7 days) and long-term (14 days) retinoid treatment, followed by the genome-wide analyses of target gene binding sites for RXR, RAR, and p300. Exposure to bexarotene, an RXR ligand, led to a subset of genes that became bound by RXR, as well as RAR, since both could form heterodimers. These genes furthermore showed increases in H3K27ace (histone 3, lysine-27 acetylation) and p300 binding (Fig. 3, A and B, and fig. S3, A and B). For instance, well-known retinoid target genes such as STRA6 and DHRS3 among others, which were originally not bound by RXR in the mesenchymal cell state, were bound upon bexarotene-induced MET (Fig. 3C and fig. S3C).

Fig. 3.

Lipid metabolism genes are regulated by retinoid receptors.

(A) Global heatmap of ChIP-seq peaks in NAMEC8 treated with DMSO, ST (short term; 7 days) bexarotene, or LT (long term; 14 days) bexarotene with RXRα, RARα, H3K27ace, and p300 antibodies. (B) Plots showing the occupancy of RXRα, RARα, H3K27ace, and p300 at gene bodies upon ST and LT bexarotene treatment in NAMEC8. (C) Gene tracks depicting increases in RXRα and H3K27ace binding at STRA6, DHRS3, PLIN1, PLIN2, and ACSL5 loci and decreases in RXRα and H3K27ace binding at MGLL locus upon ST and LT bexarotene treatment in NAMEC8. The x axis shows chromosome position with the gene structure drawn below, and each scale unit represents 5 kb. The y axis shows genomic occupancy in units of revolutions per minute per base pair. (D) Gene Ontology (GO) pathway enrichment analyses of genes that have increased binding to RXRα with bexarotene treatment in NAMEC8. (E) Plot of super-enhancers ranked by increasing H3K27ace signals in NAMEC8 treated with bexarotene. (F) Heatmap of lipid-associated genes altered upon bexarotene treatment in NAMEC8, SUM159, MetBr-007DR, BC29.1, and BC22.1 cells. Data represent log2 fold change. (G) Gene expression of lipid-associated genes involved in TAG synthesis and degradation upon bexarotene treatment relative to DMSO in NAMEC8 (top) and BC2.2 cells (bottom). Graphs represent data as means ± SEM (n = 3). *P < 0.05, **P < 0.01, ***P < 0.001, and ****P < 0.0001.

We next sought to identify the functional pathways regulated by RXR by mapping the top RXR-bound genes using Gene Ontology analyses. This revealed lipid metabolism to be a major pathway that was perturbed by retinoid treatment—a phenomenon that was previously reported in mice but for which the mechanistic underpinnings remained unexplained (Fig. 3D) (). To further test whether the RXR-bound genes were associated with transcriptional changes caused by RXR activation, we performed RNA sequencing (RNA-seq) of bexarotene-treated cells to detect gene expression changes. Consistent with the above analyses, genes that were up-regulated by at least more than twofold showed clear enrichment in pathways relating to lipid metabolism (fig. S3, D and E). Super-enhancers are features of keystone genes that help define cell states and cell lineages. Analyses of H3K27ace marks upon retinoid-induced MET revealed key genes that contained super-enhancers, among which were retinoid target genes such as DHRS3, RXR, and GDF15, as well as lipid metabolism genes such as B4GALT5 and ACSL1 (Fig. 3E and fig. S3F). For instance, lipid metabolism genes such as PLIN1, PLIN2, ACSL5, and MGLL demonstrated clear increases in RXR recruitment and concomitant H3K27ace modifications upon bexarotene treatment (Fig. 3C and fig. S3C); this further underscored their putative roles in MET.

To determine which specific aspects of lipid metabolism might be altered during cell state changes, we examined the list of retinoid target genes across multiple cancer lines treated with bexarotene to identify consistent up- or down-regulation patterns (Fig. 3F). Many of the target genes were involved in the synthetic pathway of triacylglycerides (TAGs) from fatty acids. Genes involved in TAG synthesis (ACSLs, AGPATs, GPAT, DGATs, and PLINs) were up-regulated, whereas MGLL, a lipase involved in TAG degradation, was strongly down-regulated (Fig. 3G). Similar changes in TAG synthesis and degradation were also observed in TTNPB-treated NAMEC8 cells (fig. S3G). Thus, these results strongly pointed to the reprogramming of fatty acid metabolic processes during retinoid-induced MET.

MET restricts FAO by channeling fatty acids toward lipid storage

A major fate of fatty acids is their conversion and storage as TAGs. Key lipid metabolism enzymes, such as long-chain fatty acid-CoA (coenzyme A) ligase (ACSL), glycerol-3-phosphate acyltransferase (GPAT), 1-acylglycerol-3-phosphate (AGPAT), prostatic acid phosphatase (PPAP), and diglyceride acyltransferase (DGAT), which were all up-regulated upon retinoid treatment, might help drive the conversion of fatty acids to intermediates such as diacylglycerides (DAGs) and TAG (Fig. 4, A and B). To demonstrate this, we determined the fate of fatty acids and their downstream intermediates using lipid-based mass spectrometry (MS)–based lipidomics upon changes in cell states. Following bexarotene treatment, lipid species such as phosphatidylinositols, phosphatidylglicerols (PGs), phosphatidylcholines, phosphatidylethanolamines, and phosphatidylserines, which are the downstream metabolites of fatty acids, were markedly increased. Accumulation of DAG and TAG species, which may be stored in lipid droplets, was also observed (Fig. 4A). A clear abundance of lipid droplets was observed in mesenchymal cancer cells that had undergone bexarotene-induced MET (Fig. 4C and fig. S4, A and B), suggesting that the fate of fatty acids might differ in the mesenchymal and epithelial cell states. Studies of retinoid therapy in patients with APL have reported hypertriglyceridemia, supporting our observations of increased lipid storage upon retinoid treatment (). Together, this supports the observation of rewired lipid metabolism toward TAG synthesis during the process of retinoid-induced MET.

Fig. 4.

MET restricts FAO by channeling fatty acids toward lipid droplet storage.

(A) Relative lipid content, measured by LC-MS/MS, in DMSO- versus bexarotene-treated NAMEC8. Genes up-regulated upon bexarotene treatment as validated by RT-PCR (Fig. 3G) are colored in red. (B) Protein expression of lipid-associated genes involved in TAG synthesis and degradation in DMSO- versus bexarotene-treated NAMEC8. (C) Lipid droplet staining in NAMEC8, SUM159, and BC2.2 cells: DMSO- versus bexarotene-treated. Images were taken at ×200 magnification. Scale bars, 50 μm. (D) Oxygen consumption rate (OCR) of DMSO- versus bexarotene-treated NAMEC8, in the presence or absence of palmitate. Arrows depict rate of FAO: DMSO (black) versus bexarotene (red). (E) Relative levels of acylcarnitines, measured by LC-MS, in DMSO- versus bexarotene-treated SUM159. (F) OCR of DMSO- versus A83-01–treated NAMEC8, in the presence or absence of palmitate. Arrows depict rate of FAO: DMSO (black) versus A83-01 (red). (G) Protein expression of EMT and lineage markers in NAMEC8 under the following conditions. Treatment with DMSO, bexarotene, FSG67, or FSG67 + bexarotene (left). Treatment of Ctrl sh or shAGPAT cells with DMSO or bexarotene (middle). Treatment with DMSO, bexarotene, amidepsine A, or amidepsine A + bexarotene (right). (H) Mammosphere-forming assay of NAMEC8 cells under the following treatment conditions as in (G). Images were taken at ×50 magnification. Scale bars, 100 μm. (I) Relative lipid content, measured by LC-MS/MS, in HMLE versus NAMEC8. Genes down-regulated in NAMEC8 relative to HMLE as validated by RT-PCR (fig. S4G) are colored in blue. (J) Graph showing FAO rates in a panel of breast cancer cell lines. (K) Cell viability assay of BT549 (left) and MCF7 (right) grown for 6 days in basal medium (DMEM with 1% FBS and 2.5 mM glucose) or basal medium supplemented with 25 mM glucose, 20 μM palmitate, or 20 μM lipoic acid. Graphs represent data as means ± SEM (n = 3). *P < 0.05, **P < 0.01, ***P < 0.001, and ****P < 0.0001.

Since a more epithelial cell state appeared to be associated with greater lipid storage, we reasoned that in the mesenchymal cell state, fatty acids might be used as an energy source through alternative biochemical pathways such as FAO. FAO is the catabolic process by which fatty acids are broken down in the mitochondria to produce acetyl-CoA for fueling the tricarboxylic acid (TCA) cycle. Here, we used the difference in maximal respiration as a readout for FAO activity. Using palmitate as a source of fatty acid, nutrient-starved control-treated cells residing in the mesenchymal state clearly metabolized palmitate for energy production. By contrast, bexarotene-treated cells, which had transited to the epithelial state, showed a lower FAO even when supplemented with palmitate (Fig. 4D and fig. S4, C and D). To further assess FAO activity, we measured acylcarnitines, which are intermediates of FAO. Using liquid chromatography–MS (LC-MS), lower levels of acylcarnitines were detected in bexarotene-treated cells (Fig. 4E). Reduced FAO activity was also observed in another cellular model of TGFβ inhibitor–induced MET (Fig. 4F and fig. S4, E and F). These affirmed the observation that mesenchymal cancer cells metabolically favored FAO, whereas the transition to the epithelial state shunted fatty acids toward lipid storage.

We next sought to functionally test whether intermediate lipid enzymes involved in the TAG synthesis pathway were necessary for the regulation of MET. Pharmacological inhibition of GPAT (using FSG67) and DGAT (using amidepsine A), as well as shRNA knockdown of AGPAT9, all resulted in clear reductions in the accumulation of lipid droplets during retinoid treatment when compared to vehicle treatment or control shRNA knockdown (fig. S4G). This was accompanied by the retention of FAO activity upon inhibition of TAG synthesis and retinoid treatment (fig. S4H). We observed a partial blockade in retinoid-induced MET when GPAT, DGAT, or AGPAT9 was functionally disrupted, as the gain in epithelial/luminal markers such as E-cadherin, KRT18, and MUC1 was suppressed, and there was a smaller reduction in mesenchymal/basal markers Zeb1 and KRT5 (Fig. 4G). Associated with the block in MET were the corresponding rescues in mammosphere-forming ability of mesenchymal NAMEC8 cells, clearly underscoring the importance of the TAG synthesis pathway in the epithelial cell state (Fig. 4H and fig. S4I). Changes to mammosphere formation were not the result of effects on proliferation rates of the drug-treated cells as they were similar to control-treated cells (fig. S4J).

Having observed the switch in the fate of fatty acid utilization in retinoid-induced MET, we wondered whether cancer cells in their native cell states also manifested such metabolic preferences. Comparing two isogenic cell lines stably residing in the epithelial [human mammary epithelial cell line (HMLE)] or the mesenchymal (NAMEC8) cell state, the former showed greater accumulation of lipid species to favor TAG production and, ultimately, lipid droplet formation (Fig. 4I and fig. S4K). This could be due to lower expression of TAG synthesis enzymes, coupled to an increased expression of the TAG degradation enzyme, MGLL, in mesenchymal NAMEC8 compared to HMLE (fig. S4, L and M). NAMEC8 also exhibited a higher dependency on FAO as reflected by higher FAO rates compared to epithelial HMLE (fig. S4N). Likewise, classic mesenchymal cancer cell lines (BT549, MDA-MB-468, and SUM159) had higher FAO activity than epithelial cancer cell lines (MCF7 and T47D) (Fig. 4J). Higher levels of acylcarnitines were also detected in mesenchymal cancer cells (NAMEC8 and SUM159) compared to epithelial HMLE (fig. S4O). A growth advantage was observed in the mesenchymal cancer cell line BT549 when supplemented with fatty acids palmitate or lipoic acid under nutrient-deprived conditions, while epithelial cancer cell line MCF7 did not exhibit such metabolic preferences (Fig. 4K), thus indicating that the ability to harness fatty acids for FAO could be a unique adaption of mesenchymal cells.

FAO results in the epigenetic regulation of EMT genes

If cancer cells display metabolic dependencies on FAO for maintenance of a mesenchymal phenotype, we then hypothesized that inhibiting β-oxidation might be able to promote or stabilize an epithelial cell phenotype through restricting access to this metabolic pathway. Carnitine palmitoyltransferase I (CPT1) is a key mitochondria enzyme responsible for catalyzing the transfer of an acyl group from long-chain fatty acyl-CoA to form acyl carnitines. In essence, this allows for the transport of fatty acids into the mitochondria for fueling the TCA cycle. We first showed that exposing mesenchymal cancer cells to a CPT1 inhibitor, etomoxir, alone or together with bexarotene completely inhibited FAO activity (fig. S5A). Although bexarotene alone could cause a MET, etomoxir on its own did not. An increase in TAG synthesis enzymes and a decrease in MGLL were observed with etomoxir treatment alone, priming cells for acquisition of an epithelial cell state (fig. S5B). Combination treatment of bexarotene and etomoxir exerted a more pronounced MET phenotype that was more durable and sustained even when the drugs were withdrawn for 7 days (Fig. 5A and fig. S5, C and D). Cells that had been treated with the combination therapy also sustained the inhibitory effect on mammosphere formation after drug withdrawal (Fig. 5B and fig. S5E). To confirm that the effects of etomoxir are on-target, we knocked down CPT1A, the major isoform of CPT1, in NAMEC8. Loss of CPT1A phenocopied the effects of etomoxir, showing similarly increased expression of TAG synthesis enzymes and reduction of MGLL (fig. S5B); this could contribute to increased lipid droplet abundance upon bexarotene treatment (Fig. 5C and fig. S5F). Reinforcement of bexarotene-induced MET was observed in CPT1A-knockdown cells as shown by amplified decreases in mesenchymal/basal proteins and increases in epithelial/luminal proteins (Fig. 5D), along with greater inhibition of mammosphere formation (Fig. 5E). Thus, FAO inhibition was sufficient to potentiate the induction of MET.

Fig. 5.

Induction of MET is epigenetically regulated by FAO.

(A) Protein expression of EMT and lineage markers in NAMEC8 treated with DMSO, bexarotene, etomoxir, or bexarotene + etomoxir. (B) Mammosphere-forming assay of NAMEC8 treated with bexarotene, etomoxir, bexarotene + etomoxir, −bexarotene, and −(bexarotene + etomoxir). Graph showing number of spheres >100 μm (right). (C) Lipid droplet staining in Ctrl sh, CPT1A-sh1, and CPT1A-sh2 NAMEC8 treated with DMSO or bexarotene. Images were taken at ×100 magnification. Scale bars, 50 μm. (D) Protein expression of EMT and lineage markers in Ctrl sh and shCPT1A NAMEC8 treated with DMSO or bexarotene. (E) Mammosphere-forming assay of Ctrl sh, CPT1A-sh1, and CPT1A-sh2 NAMEC8 treated with DMSO or bexarotene (left). Protein expression showing CPT1A knockdown and graph showing number of spheres >100 μm (right). (F) Lipid droplet staining in vector control and CPT1A-1– and CPT1A-2–overexpressing NAMEC8 treated with DMSO or bexarotene. Images taken at ×200 magnification. Scale bars, 50 μm. (G) Protein expression of EMT and lineage markers in vector control and CPT1A-overexpressing NAMEC8 treated with DMSO or bexarotene. (H) Mammosphere-forming assay of vector control and CPT1A-1– and CPT1A-2–overexpressing NAMEC8 treated with DMSO and bexarotene (left). Protein expression showing CPT1A overexpression and graph showing number of spheres >100 μm (right). (I) Protein expression of H3 acetylation in NAMEC8 treated with increasing doses of bexarotene. (J) Protein expression of acetyl-lysine and H3K27 acetylation in NAMEC8 treated with DMSO, bexarotene, bexarotene + etomoxir, or bexarotene + etomoxir + LAA. (K) qPCR of EMT gene promoters in ChIP of H3K27 acetylation in NAMEC8 treated with bexarotene + etomoxir or bexarotene + etomoxir + LAA. (L) Protein expression of EMT-associated markers in NAMEC8 treated with DMSO, bexarotene, bexarotene + etomoxir, or bexarotene + etomoxir + LAA. (M) Mammosphere-forming assay of NAMEC8 treated with DMSO, bexarotene, bexarotene + etomoxir, LAA, or bexarotene + etomoxir + LAA (left). Graph showing number of spheres >100 μm (right). Images of mammosphere-forming assays were taken at ×50 magnification. Scale bars, 100 μm. Graphs represent data as means ± SEM (n = 3). *P < 0.05, **P < 0.01, ***P < 0.001, and ****P < 0.0001.

If β-oxidation was critical for the maintenance of the mesenchymal cell state, we then reasoned that the increased activity of CPT1A through its overexpression would overcome the effects of bexarotene and the induction of an epithelial cell phenotype. We overexpressed CPT1A in SUM159 cells, and notably, this was sufficient to overcome bexarotene-dependent inhibition of FAO activity (fig. S5G). In two cell lines overexpressing CPT1A, cells clearly showed the absence of lipid droplets even upon bexarotene treatment (Fig. 5F and fig. S5, H and I), suggesting that retention of β-oxidation in CPT1A-overexpressing cells could have prevented a reversion to an epithelial cell state. This was demonstrated by the maintenance of a mesenchymal-like phenotype in CPT1A-overexpressing cells upon bexarotene treatment, as evidenced by the retention of a mesenchymal-like cell morphology (fig. S5J), partial maintenance of mesenchymal/basal markers, and suppressed induction of epithelial/luminal markers upon bexarotene treatment (Fig. 5G). The decrease in mammosphere formation by bexarotene treatment was also rescued by CPT1A overexpression, suggesting that the enhanced import of fatty acids into the mitochondria for FAO was sufficient to partially block bexarotene-induced MET (Fig. 5H and fig. S5K). The partial block in MET by CPT1A overexpression revealed that apart from FAO inhibition, other additional cellular processes identified in the ChIP-seq pathway analyses could participate in retinoid-induced MET (Fig. 3D). Nonetheless, while an upstream trigger of cell state transition by retinoids was necessary, the reprogramming of lipid metabolism was an essential gateway for determining cell states.

To connect fatty acid metabolism to regulation of EMT gene transcription, we reasoned that this might happen through epigenetic control of histone acetylation events. FAO generates acetyl-CoA, a substrate for histone acetylation. We first analyzed the perturbation of acetylated histone marks during retinoid-induced MET.Acetylation of H3K14, H3K18, and H3K27, which control active gene transcription, was decreased upon bexarotene treatment in a dose-dependent manner (Fig. 5I). With combination treatment of bexarotene and etomoxir, pan-histone acetylation was decreased; H3K27ace was further reduced compared to bexarotene alone. To demonstrate that the decrease was the result of FAO inhibition by etomoxir, which led to reduction in acetyl-CoA, supplementation of acetyl-CoA was carried out by addition of lithium acetoacetate (LAA). This partially rescued the suppression of acetylation levels upon combination treatment (Fig. 5J). To determine whether decreases in H3K27ace levels led to chromatin modifications of EMT gene promoters, ChIP of H3K27ace was carried out during bexarotene + etomoxir treatment, with or without LAA. A reduction in the enrichment of H3K27ace was observed on the promoters of mesenchymal genes upon bexarotene + etomoxir treatment, but this was rescued by the addition of LAA (Fig. 5K). The rescue of H3K27ace on the promoters of EMT genes likely accounted for their sustained expression, which were decreased during bexarotene + etomoxir treatment (Fig. 5L). The restoration of acetyl-CoA levels was also sufficient to rescue the inhibition of mammosphere formation (Fig. 5M). Treatment of mesenchymal cells with LAA alone did not have significant effects on cell morphology, EMT status, or fatty acid metabolism (fig. S5, L to O). Similar proliferation rates were also observed across drug treatments (fig. S5P). Together, heightened FAO activity in the mesenchymal state contributed toward the maintenance of H3K27ace on promoters of EMT genes, thereby directly connecting fatty acid metabolism to the epigenetic regulation of EMT.

Impairment of FAO rechannels lipids toward the support of an epithelial cell state and blocks EMT

Since MET programs the fate of fatty acids toward lipid storage, we reasoned that the reverse process, an EMT, would lead to opposite outcomes. Using three different cell lines (HMLE-Twist ER, HMLE-Zeb1, and MCF7-Slug+Sox9) containing the inducible expression of EMT transcription factors TWIST, ZEB1, and SLUG, respectively, we observed substantial abundance of lipid droplets when cells resided in the epithelial state, which were lost as cells transited to the mesenchymal state (Fig. 6A and fig. S6A). This was corroborated by a decreased expression of TAG synthesis enzymes and an increased expression of MGLL in induced HMLE-Twist ER and HMLE-Zeb1 cells (Fig. 6B). Using HMLE-Twist ER cells, the activation of an EMT program increased FAO rates in epithelial cells, which otherwise exhibited low FAO (Fig. 6C and fig. S6B). Thus, activation of EMT reprogrammed the metabolic features of epithelial cancer cells from lipid storage to FAO. In TGFβ-induced EMT in a panel of epithelial cell lines including lung cancer A549 and liver cancer Hep3B cells, we also observed a decrease in lipid droplet accumulation as epithelial cells underwent EMT (fig. S6C). They now also showed an increased rate of FAO, demonstrating that reprogramming of lipid metabolism is a characteristic of EMT across different cell types (Fig. 6D and fig. S6, D and E).

Fig. 6.

Impairment of FAO rechannels lipids toward the support of an epithelial cell state and blocks EMT.

(A) Lipid droplet staining in HMLE-Twist ER, HMLE-Zeb1, and MCF7-Slug+Sox9 cells: Control versus 4OHT-induced or DOX-induced. Images were taken at ×100 magnification. Scale bars, 50 μm. (B) Protein expression of lipid-associated genes involved in TAG synthesis and degradation in HMLE-Twist ER and HMLE-Zeb1: Control versus 4OHT-induced or DOX-induced. (C) OCR in HMLE-Twist ER control versus 4OHT-induced cells in the presence or absence of palmitate. Arrows depict rate of FAO: Control (black) versus 4OHT-induced (red). (D) OCR in DMSO- versus TGFβ-treated A549 in the presence or absence of palmitate. Arrows depict rate of FAO: DMSO (black) versus TGFβ (red). (E) Protein expression of EMT markers in HMLE-Twist ER, ± 4OHT induction and etomoxir treatment. (F) Protein expression of EMT markers in MCF7-Slug+Sox9, ±DOX induction and etomoxir treatment. (G) Lipid droplet staining in MCF7-Slug+Sox9 Ctrl sh and shCPT1A cells, ± DOX induction. Images were taken at ×200 magnification. Scale bars, 20 μm. (H) Protein expression of EMT markers in MCF7-Slug+Sox9 Ctrl sh and shCPT1A cells ± DOX induction. (I) Protein expression of acetyl-lysine and H3K27 acetylation in HMLE-Zeb1 cells treated with DMSO, DOX, DOX + etomoxir, or DOX + etomoxir + LAA. (J) Protein expression of mesenchymal markers in HMLE-Zeb1 cells treated as in (I). (K) Top left: Workflow schematic for mouse xenograft experiment. Right: Immunofluorescence staining of E-cadherin and vimentin in primary tumors formed in MCF7-Slug+Sox9 control versus DOX versus etomoxir + DOX groups (×200 magnification; scale bars, 50 μm), whole-mount fluorescence lung images (×100 magnification; scale bars, 100 μm), and histology of the lung sections (×10 magnification; scale bars, 50 μm) from lung metastases. Bottom left: Graph shows number of metastases per lung. Data are represented as means ± SEM (n = 5 technical replicates). **P < 0.01. (L) Protein expression of EMT and lineage markers in primary tumors from the experiment in (K).

Since the shifts in metabolic reprogramming during EMT translated to changes in FAO, we hypothesized that the inhibition of CPT1 might have consequences on EMT progression. Treatment with etomoxir blocked HMLE-Twist ER and HMLE-Zeb1 cells from undergoing a complete EMT, as observed by the retention of an epithelial-like cell morphology and weaker up-regulation of mesenchymal markers upon EMT induction (Fig. 6E and fig. S6, F and G). This was replicated in another inducible EMT cell line model, MCF7-Slug+Sox9, whereby etomoxir partially blocked Slug+Sox9-induced EMT in breast cancer cells, as well as in models of TGFβ-induced EMT using epithelial A549 lung and Hep3B liver cancer cells (Fig. 6F and fig. S6, H to J). Consistent with this, knockdown of CPT1A in MCF7-Slug+Sox9 partially blocked the loss in lipid droplets upon EMT induction, indicating maintenance of an epithelial cell state (Fig. 6G and fig. S6K). This is supported by weaker inductions of mesenchymal markers upon EMT activation in shCPT1A cells, demonstrating a partial blockade of EMT (Fig. 6H and fig. S6L). The contribution of acetyl-CoA from FAO toward histone acetylation was further analyzed by supplementation with LAA. There were negligible changes by LAA treatment on A549 and Hep3B cells, which reside in the epithelial cell state, as observed by morphology, protein expressions of EMT markers, and lipid droplet formation (fig. S6, M to O), indicative of lower demands for acetyl-CoA in this cell state. Upon doxycycline (DOX)-induced EMT in HMLE-Zeb1 cells, there were increased H3K27ace levels, likely contributed by enhanced FAO, which could promote the transcriptional activation of EMT genes. Inhibition of FAO by etomoxir decreased H3K27ace levels, blocking the up-regulation of mesenchymal markers under DOX-induced EMT. H3K27ace levels could, however, be rescued with LAA supplementation (Fig. 6I); this helped maintain the expression of mesenchymal markers (Fig. 6J). This underscores acetyl-CoA production by FAO as critical to epigenetically support the acquiring of a mesenchymal phenotype during EMT.

Last, we sought to functionally test whether the direct perturbation of metabolic pathways, in this case, FAO, could alter the course of EMT-induced metastasis in animal models in vivo. We used the cell line model MCF7-Slug+Sox9, whereby DOX-inducible concomitant expression of the EMT transcription factors, Slug and Sox9, in preestablished MCF7 tumor xenografts could strongly induce EMT in vivo and promote the metastasis of otherwise poorly metastatic tumor cells from the fat pads to the lungs of NSG mice (Fig. 6K) (). Here, MCF7-Slug+Sox9 cells were transplanted into the fat pads of NSG mice, and drug treatment was started 1 week after implantation. One group of mice was administered with etomoxir (20 mg/kg per day) before DOX administration. Another group of mice were sham-treated with vehicle control, followed by DOX. After 10 weeks, immunofluorescence and Western blot analyses showed that the primary tumors of control-treated mice were epithelial in nature, exemplified by high expressions of E-cadherin and KRT18 as well as low expressions of Fibronectin, Zeb1, Slug, and KRT5. In contrast, tumors of DOX-treated mice showed hallmarks of an EMT that included the loss of E-cadherin and simultaneous up-regulation of mesenchymal markers. However, tumors of DOX-treated mice that also received etomoxir showed retention of epithelial features, thus demonstrating a blockade of EMT in vivo (Fig. 6, K and L). In addition, Western blot analysis of lipid metabolism enzymes in the primary tumors showed decreases in the expression of TAG synthesis enzymes and an increase in MGLL in the tumors of DOX-treated mice, which were abrogated in etomoxir-treated mice (fig. S6P). These alterations in the expressions of lipid metabolism enzymes were consistent with the changes in cell state within the primary tumors (Fig. 6L). Together, these results demonstrated the effect of FAO inhibition via etomoxir in blocking the EMT program in vivo.

Extensive metastatic nodules were detected in the lungs of DOX-treated mice, whereas the counterparts that received etomoxir had as many metastatic lesions as control-treated mice (Fig. 6K). This provided the first proof of concept that FAO is necessary for the activation of an EMT program to spawn metastases, and the pharmacological inhibition of this crucial metabolic pathway may be useful for controlling metastasis in breast cancer.

DISCUSSION

The ability to precisely manipulate cell states represent one emerging approach to control the response of cancer cells to therapy. In APL and cutaneous T cell lymphoma, retinoids have been applied as a first-line adjuvant with curative intent (). In solid tumors, however, we demonstrated that retinoids act as an inducer of the shift in cell states, which primes otherwise resistant breast cancer cells to respond better to standard-of-care treatments, and may thus be clinically useful in combination therapies including immunotherapy. In addition, retinoid-induced MET also shifts the cells from the more aggressive mesenchymal state to the less metastatic epithelial state. In small cohort studies of patients with metastatic breast cancer, retinoid therapy did not result in response, possibly due to the lack of tumor molecular subtype stratification and patient selection (, ). With new information arising from this study, basal-like TNBCs, which tend to be more mesenchymal in nature, are more likely to respond. This highlights the need for biomarkers that could stratify potential responders from nonresponders.

Mechanistically, there may be several routes to achieve the conversion of a MET phenotype. In our studies, retinoids are the most potent class of such inducers identified in an unbiased manner. They exert their effects through rewiring lipid metabolism during cell state conversion (Fig. 7A). We show this to be mediated through RXR and RAR signaling. From other studies, RXR has been shown to heterodimerize with peroxisome proliferator–activated receptor γ (PPARγ) to mediate changes in lipid metabolism (, ). Thus, while we demonstrated that retinoid-induced lipid metabolic changes are predominantly mediated by RXR/RAR signaling, it remains plausible that RXR/PPAR signaling may be involved in this process. A few recent studies have delved into how lipid metabolism can affect tumor plasticity and metastasis (). For instance, LACTB impinges upon mitochondrial lipid metabolism to mediate tumor cell differentiation (), while inhibition of fatty acid synthase sensitizes APL cells to ATRA-mediated differentiation (). In our study, the mesenchymal cell state is associated with increased FAO, while rechanneling lipid flux toward TAG synthesis and lipid storage is a crucial aspect of retinoid-induced MET. This interplay between FAO and lipid storage may be a manifestation of metabolic plasticity during cell state changes. The link between FAO and tumor progression is supported by studies where Myc-overexpressing breast cancer displayed an increased bioenergetic reliance on FAO for growth (), and the tumorigenic potential of TNBC cells was dependent on FAO-driven Src activation (). Furthermore, FAO was also a critical metabolic adaptation for cancer metastasis to lymph nodes (), as well as being induced in response to acidosis of the tumor environment to ultimately support anoikis resistance and metastasis (, ). Inhibiting FAO alone was not sufficient to induce a MET, suggesting that potent “drivers” such as retinoids are required to activate cell state changes. Nevertheless, overexpression of CPT1A alone was able to partially block bexarotene-induced MET, and the effects of bexarotene could be prolonged by etomoxir treatment, thus demonstrating that lipid metabolic reprograming is a feature that helps control cell state transitions.

Fig. 7.

Rewiring of fatty acid metabolism supports cell state transitions.

(A) Model illustrating differing fatty acid metabolism driving epithelial and mesenchymal cell states. In the mesenchymal cell state, FAO is high, contributing acetyl-CoA for acetylation of histones on EMT genes. Upon induction of MET, the epithelial cell state is characterized by inhibited FAO and accumulation of lipid stores.

Why do the differences in cancer cell states impose unique metabolic requirements? In cancer, cell states control proliferation, invasiveness, and stress tolerance, which necessitate strict requirements for specific metabolites. For instance, asparagine availability promotes EMT-driven metastasis or tumor growth (, ), whereas cisplatin-resistant cancer cells appear to depend on glutamine for nucleotide biosynthesis (). Pertaining to lipids, dietary lipids boost metastasis of human oral carcinomas in a CD36-dependent manner (), while biosynthesis of monounsaturated fatty acid sapienate promotes tumor cell plasticity (). However, the specific roles of such lipid-associated metabolites in mediating cell differentiation remain unclarified. In embryonic stem cells, glutamine and α-ketoglutarate (αKG) support pluripotency through elevated αKG-to-succinate ratios that promote demethylation of DNA and histones (). TGFβ-induced inhibition of fatty acid synthesis enzyme (ACC1) contributed to acetyl-CoA–driven activation of SMAD2 and EMT activation (). In acidosis-adapted tumors, cancer cells were shown to rely on FAO for acetyl-CoA, which hyperacetylates mitochondrial proteins essential for tumor growth ().

In our study, channeling fatty acids toward FAO for energy production is a metabolic manifestation of mesenchymal cells, and the generation of acetyl-CoA results in the epigenetic regulation of EMT target genes through histone acetylation. This forms the mechanistic link between FAO and metabolic regulation of gene expression programs in distinct cell states. The enhancement of FAO may allow cancer cells in the mesenchymal state to tolerate fluctuations in nutrient levels during invasion and metastasis, whereas the more proliferative epithelial state may preferentially adopt aerobic glycolysis even under glucose-rich conditions, which is typical for bulk tumor cells. Pharmacological inhibition of FAO in the context of transcription factor–driven EMT could block the acquisition of a mesenchymal cell state and metastatic seeding of breast cancer cells in vivo. This highlights FAO as a metabolic “gateway” that needs to be activated for cell state transitions to occur. The addiction to specific metabolic pathways during distinct steps of cancer progression may be exploited to target specific cell states and opens up new avenues for antimetabolite therapeutic strategies.

MATERIALS AND METHODS

Cell lines and cell culture

HMLEs were generated as described and immortalized using retroviral vectors to express the catalytic subunit of the human telomerase enzyme, hTERT, and the SV40 large T antigen (). NAMEC8 was generated as described (). Briefly, HMLE cells were grown to 50% confluency, followed by differential trypsinization for 1 min with 0.05% trypsin. Detached cells were carefully collected and replated at approximately 200 cells per well of a 24-well plate. Upon expansion, wells were screened for populations that were morphologically mesenchymal and can be stably propagated. Clonal and nonclonal NAMEC cell lines were generated. NAMEC8 is a nonclonal line derived from at least 200 mesenchymal cells. HMLE cells and NAMECs were maintained in Mammary Epithelial Cell Growth Medium (MEGM) (Lonza).

HMLE-Twist ER cells were generated as described (). For EMT induction, 100 nM 4-hydroxytamoxifen in ethanol was added to the culture medium every 48 hours to induce Twist expression. Control cells were treated with ethanol. MCF7-Slug+Sox9 cells were generated as previously described (). HMLE-Zeb1 cells were generated by infection with FUW-Zeb1 and rTTA. DOX was added to the culture medium every 48 hours to induce Zeb1 expression. To generate tumorigenic cells from NAMEC8, cells were infected with pWZL-Blast-Ras (). MCF7, MCF7-Ras, SUM159, BT549, MCF10A, and T47D were maintained as previously described (). A549 and Hep3B were cultured in Dulbecco’s modified Eagle’s medium (DMEM; Gibco) with 10% fetal bovine serum (FBS), 2 mM l-glutamine, and 1% penicillin-streptomycin (Gibco). All cell cultures were maintained at 37°C with 5% CO2.

Primary human cells

Breast tissue biopsies from patients with cancer were obtained from the National Cancer Centre Singapore (NCCS) with patients’ consent and under an Institutional Review Board (IRB)–approved protocol (2020-149). Tissue was digested with collagenase IV (1 mg/ml; Gibco) and Dispase II (1 mg/ml; Thermo Fisher Scientific) in F12 media for 2 hours while shaking in a 37°C incubator. Cell suspension was then put through a cell strainer and centrifuged for 10 min at 1500 rpm. Cells were then washed twice and resuspended in media before seeding on irradiated 3T3 feeders as previously described (). When confluent, subculturing was carried out by first lifting the feeders using 0.15% trypsin for 2 min. After removing feeder cells, 0.25% trypsin was added to detach tumor progenitor cells.

Patient-derived breast cancer xenograft establishment and therapy

Primary human breast cancer samples were obtained from the NCCS with patients’ consent and under an IRB-approved protocol (2020-149). All animal experiments were approved by the Agency for Science, Technology and Research Singapore, Biological Resource Centre IACUC (Institutional Animal Care and Use Committee) (protocol number 171286). Patient-derived breast tumor fragments (approximately 3 mm by 1 mm by 1 mm) were inserted bilaterally into the inguinal mammary fat pads of 6- to 8-week-old NSG female mice for initial establishment of tumors and subsequently expanded in NSG mice once established. In this study, established tumor fragments of BC2.2 were implanted into cohorts of 6- to 8-week-old female NSG mice. For BC22.1, 1 × 106 cells were injected subcutaneously into both flanks of NSG mice.

Generation of reporter cell line for screening of compounds

pGreenFire1-mCMV was cut using Cla I and Spe I and gel-purified. Promoter sequences of genes Zeb1, SERPINE1, CDH1, CD24, CD44, and CD166 were amplified from human genomic DNA with predesigned primers containing Cla I and Spe I restriction cut sites using AccuPrime Taq DNA Polymerase (Thermo Fisher Scientific). Polymerase chain reaction (PCR) products were then digested with restriction enzymes and gel-purified. Vector was ligated with PCR insert using Quick Ligase (Fermentas) and transformed into Stbl2 competent cells. Colonies were picked and inoculated into TB media with ampicillin. Plasmids were then isolated using a plasmid miniprep kit (QIAGEN) and sent for sequencing (1stBASE Laboratories). After verifying the plasmid sequence, transfection into human embryonic kidney–293 (HEK293) cells was carried out to generate lentivirus. Transfection was carried out using FuGENE 6 (Promega). Viral supernatant was harvested at 24 and 48 hours. NAMEC8 cells were then infected twice with virus using polybrene (8 μg/μl) and checked for good transduction efficiency using the Dual-Glo Luciferase System (Promega).

Screening of compound libraries

Stock plates of compound libraries were prepared by GIS CHiP facility at 1 mM concentration in dimethyl sulfoxide (DMSO). NAMEC8-SERPINE1prom-Luc cells were seeded into 384-well plates at a density of 700 cells per well in a total volume of 50 μl, using an automated dispensing system (BioTek). Compounds (0.5 μl) were then added to the wells the following day using a robotic pin transfer with a Bravo workstation (Agilent Technologies). Cells were then grown for 3 days, and promoter luciferase activity was measured using the Dual-Glo Luciferase System (Promega).

Dual-Glo Luciferase assay

Assay was carried out according to the manufacturer’s instructions. To measure luciferase activity in 384-well plates, media was first removed from the plate. Firefly luciferase activity was determined by adding 30 μl of Dual-Glo Luciferase substrate (Promega) to the wells and incubated for 10 min before measuring luminescence using a TECAN M1000 Pro plate reader. Dual-Glo Stop & Glo substrate (15 μl) was then added to the wells, incubated for 10 min, and measured luminescence on the plate reader to determine Renilla luciferase activity. Promoter luciferase activity is then calculated by taking firefly luciferase over Renilla luciferase.

ERE-luciferase activity assay

ERE-luciferase construct was provided by A. P. Kumar, CSI, Singapore. To measure the ERE-luciferase activity, cells were seeded 24 hours before transfection. ERE-luciferase (1 μg) and NSV-Renilla (0.1 μg) were packaged using Lipofectamine 2000 and transfected into NAMEC8. Twenty-four hours after transfection, cells were reseeded into 96-well plates for drug treatment. Firefly and Renilla luciferase activities were measured with the Dual-Glo Luciferase System (Promega) 24 hours after treatment.

Lentiviral shRNA and overexpression constructs

Lentiviral shRNAs were obtained from The RNAi Consortium (TRC) collection from Broad Institute. The clone IDs for the shRNAs used are TRCN0000072181 [green fluorescent protein (GFP) control], TRCN0000330783 (RXRα shRNA1), TRCN0000036280 (CPT1A shRNA1), TRCN0000036283 (CPT1A shRNA2), and TRCN0000203516 (AGPAT9 shRNA). Each shRNA was cloned into lentiviral plasmid pLKO.1 (Addgene).

For overexpression of CPT1A, the open reading frame (ORF) of CPT1A was cut from pCMV6-CPT1A (OriGene, catalog no. RC200485) and cloned into pMN-GFP vector (gift from Y. Qiang laboratory, GIS). Positive clone was confirmed by sequencing and was used for retroviral production. For overexpression of Zeb1, the ORF of Zeb1 was cloned into FUW-LPT2 vector. Positive clone was confirmed by sequencing and was used for lentiviral production.

For production of lentivirus, plasmids were packaged into HEK293T cells with packaging plasmids using FuGENE 6 (Promega). For production of retrovirus, plasmids were packaged into PlatA cells with packaging plasmids using FuGENE 6 (Promega). Virus was harvested at 48 hours and transduced into target cells. Infected cells were selected with puromycin (2 μg/ml) for 7 days.

RNA isolation, reverse transcription, and real-time PCR analysis

Cells were rinsed twice in ice-cold phosphate-buffered saline (PBS). Total RNA was extracted using TRIzol (Invitrogen) and column-purified with RNeasy kits (QIAGEN). Complementary DNA synthesis was performed with 500 ng of total RNA at 37°C for 2 hours using the SuperScript III First-Strand Synthesis System (Thermo Fisher Scientific) and subsequently diluted fivefold. mRNA levels were measured with gene-specific primers listed below using the ABI Prism 7900HT Sequence Detection System 2.2 (Applied Biosystems). Results were normalized to glyceraldehyde-3-phosphate dehydrogenase (GAPDH) and analyzed using SDS 2.2 software.

Proliferation assays

To measure cell growth rates, 1000 cells were seeded onto 96-well plates in triplicate. Cell viability was measured using a CellTiter-Glo reagent (Promega) or CellTiter 96 Aqueous One Solution Cell Proliferation Assay (MTS) (Promega) according to the manufacturer’s instructions.

Protein extraction and immunoblotting

Cells were harvested and scraped from culture dishes in ice-cold PBS. They were then centrifuged at 12,000 rpm for 5 min at 4°C. The pellets were then snap-frozen in liquid nitrogen or lysed in ice-cold radioimmunoprecipitation assay lysis buffer (Thermo Fisher Scientific) and left on ice for 30 min. Subsequently, the lysates were centrifuged at 12,000 rpm for 10 min at 4°C. The supernatants were then aspirated and placed in fresh tubes. Protein concentrations were determined using Coomassie Blue. Total protein (50 to 80 μg) was mixed with NuPAGE LDS sample buffer and heated at 70°C for 10 min. Samples were loaded into NuPAGE 4 to 12% Bis-Tris gels (precast polyacrylamide gel; Invitrogen). Gel electrophoresis was performed at 200 V for 1.5 hours and then blotted onto nitrocellulose membranes (Trans-Blot Turbo transfer pack; Bio-Rad). The membranes were blocked with 5% skimmed milk in tris-buffered saline (1stBASE) with 0.1% Tween 20 (Santa Cruz Biotechnology) and then incubated with primary antibodies overnight at 4°C. Thereafter, the membranes were incubated with a horseradish peroxidase–conjugated secondary antibody at room temperature for 1 hour. Last, the immunoreactive proteins were visualized by using an enhanced chemiluminescent reagent (Thermo Fisher Scientific, 34075) with the ChemiDoc Imager (Bio-Rad, 17001401). For detection of secreted SERPINE1 protein, cells were seeded in a six-well plate, and media was collected after 24 hours. Twenty microliters of media was used for immunoblotting.

Fluorescence-activated cell sorting

Cells were trypsinized until they detached from the cell culture plates. Neutralization media (DMEM containing 20% FBS) was then added to inactivate the trypsin. The cells were then centrifuged at 1200 rpm for 3 min, and the cell pellet was resuspended in PBS/2% FBS. Cells were stained with DAPI (4′,6-diamidino-2-phenylindole; 1:2000) and filtered through a 40-μm strainer before sorting. Fluorescence-activated cell sorting (FACS) was performed on BD Aria Fusion cell analyzer.

Staining of CD44

Cells were harvested as per FACS and resuspended in 100 μl of solution containing CD44-APC (BD Biosciences) antibodies diluted in 2% FBS in PBS (1:40 dilution). Cells were stained for 45 min with vortexing every 10 min to ensure even staining. After antibody incubation, cells were washed once and resuspended in 300 μl of 2% FBS in PBS. Cells were then filtered through a 40-μm strainer and analyzed on BD LSR Fortessa X-20 cell analyzer.

Mammosphere assay

For mammosphere culture, cells were seeded onto 96-well ultralow-adherence plates at 700 cells per well in MammoCult medium (STEMCELL Technologies) supplemented with heparin (4 μg/ml), hydrocortisone (0.5 μg/ml), and 1% methylcellulose. Spheres (>100 μm) were counted at 7 to 10 days later.

Immunofluorescence staining of cells

Cell cultures were fixed in 4% paraformaldehyde and permeabilized with 0.25% Triton X-100, followed by blocking with 1% bovine serum albumin (BSA) in PBST (Phosphate-buffered saline with Tween 20). Cells were incubated in specific primary antibodies, followed by the appropriate secondary antibodies conjugated with Alexa Fluor 488 or 594 (Molecular Probes). Cell nuclei were then stained with DAPI and then viewed and imaged with an Axio Observer D1 epifluorescence microscope with a built-in AxioCam MR3 camera (Zeiss) using the optimal filters and ×100 to ×200 magnifications.

Generation of SUM159 PACR cells

SUM159 cells were exposed to stepwise increases to PAC starting from 10 nM. Media with fresh drug was changed every 2 days, and cells were split whenever they reached 80 to 90% confluency. Concentration of PAC was increased by 10 nM at every split until cells were able to survive at 100 nM. Cells were then subject to a cell viability assay to measure median inhibitory concentration (IC50). IC50 values were determined using GraphPad Prism software. Cells were considered resistant to PAC when there was at least a 10-fold increase in IC50 value compared to the parental cell line.

In vitro measurement of IC50

Cells were seeded in 96-well plates at a density of 2000 cells per well. Drugs (5-FU/PAC) were added in a 10-point dose-response manner the following day. Cells were allowed to grow for 3 days before cell viability was measured using the CellTiter-Glo reagent (Promega). Data were normalized to vehicle controls. Isobologram plots were generated, and IC50 values were determined using GraphPad Prism software.

In vivo experiments using NSG mice

All animal experiments were approved by the Agency for Science, Technology and Research Singapore, Biological Resource Centre IACUC (protocol number 171286). For pretreated SUM159 and MCF7-Ras cells, 1 × 106 cells were suspended in 100 μl of media containing 50% Matrigel and injected subcutaneously into both flanks of female NSG mice. The NSG mice were checked for tumor by palpation once a week, and tumors were harvested when they reached ~15 mm in diameter. For treatment in NSG mice, 1 × 105 NAMEC8-Ras cells were suspended in 100 μl of media containing 50% Matrigel and injected subcutaneously into both flanks of NSG mice. The NSG mice were checked for tumor formation by palpation, and when tumors were about 3 mm, mice were randomized into two groups: vehicle (5% DMSO in corn oil) and treatment (bexarotene, 100 mg/kg per day, oral dosing). Tumor size was measured every week using Vernier calipers, and mice were sacrificed when tumors reached ~15 mm or after ~4 weeks of treatment. For patient-derived cells, 2 × 106 cells were injected subcutaneously per site in NSG mice.

Orthotopic tumor transplantations of MCF7-Slug+Sox9 cells labeled with the tdTomato fluorescent protein were carried out as described previously (). Briefly, 1 × 106 cells were resuspended in 30 μl of media containing 50% Matrigel and injected into mammary fat pads of NSG mice. Mice were randomly divided into three groups (control, DOX, and etomoxir + DOX). Mice were administered vehicle control [0.9% (w/v) NaCl] or etomoxir (20 mg/kg) via intraperitoneal injections daily for 3 weeks. DOX was then administered through drinking water for 2 weeks. Mice were sacrificed 10 weeks after DOX treatment was stopped and were analyzed for lung metastases under a fluorescence microscope.

H&E staining of tumor sections

The tumor sections were fixed overnight in 4% paraformaldehyde, immersed in 70% ethanol, and sent to the Advanced Molecular Pathology Laboratory, IMCB for paraffinization and H&E staining.

Immunofluorescence staining of tumor sections

The tumor sections were fixed overnight in 4% paraformaldehyde, immersed in 70% ethanol, and sent to the Advanced Molecular Pathology Laboratory, IMCB for paraffinization. For immunofluorescence staining, slides were deparaffinized and rehydrated in PBS. The tissues were surrounded with a hydrophobic barrier using a barrier pen and blocked with the blocking buffer (1% BSA, 1% horse serum, and 0.2% Triton X-100 in PBS) for 1 hour. They were then washed with PBS and incubated in specific primary antibodies at 4°C overnight. The slides were washed with PBS and incubated with the appropriate secondary antibodies conjugated with Alexa Fluor 488 or 594 (Molecular Probes) at room temperature for 1 hour. The tissues were then mounted on the slides using VECTASHIELD Mounting Medium with DAPI (Vector Laboratories) and coverslip. The coverslips were permanently sealed around the perimeter with nail polish for prolonged storage. Slides were viewed and imaged with an Axio Observer D1 epifluorescence microscope with a built-in AxioCam MR3 camera (Zeiss) using the optimal filters and ×100 to ×200 magnifications. The fluorescence image analysis and the fluorescence overlay image were obtained with the AxioVision Rel. 4.8 image software.

Chromatin immunoprecipitation sequencing

ChIP assay was carried out as described previously (). Briefly, the cells were cross-linked with 1% formaldehyde for 10 min at room temperature, and formaldehyde was then inactivated by the addition of 125 mM glycine. Chromatin extracts containing DNA fragments with an average size of 200 to 500 base pairs were immunoprecipitated using anti-RARα, anti-RXRα, anti-H3K27ace, or anti-p300 antibody. The ChIP-enriched DNA was then decross-linked and used for library preparation using an Illumina TruSeq ChIP kit followed by analysis using Illumina sequencing or used for real-time PCR.

RNA sequencing

Cells were treated with bexarotene or DMSO and lysed with TRIzol. Total RNA was extracted using the Qiagen RNeasy kit (#74106). RNA integrity and concentrations were measured on a bioanalyzer, and libraries were prepared using the Illumina TruSeq RNA kit followed by analysis using Illumina sequencing.

Genomic data analyses

For ChIP-seq data processing, reads were aligned to human genome (hg19) using Burrows-Wheeler Aligner. The aligned reads were deduplicated using Picard Markduplicates.jar. Peak calling was performed using MACS2 peak-caller with default parameters. Differential binding was performed using DiffBind. H3K27ac super-enhancers were identified using the ROSE algorithm as described (, ).

For RNA-seq data processing, reads were aligned to the human reference genome GRCh37.p13 using TopHat2 (), and read counting was performed using featureCounts (). Lowly expressed genes (i.e., those with less than 10 counts in less than 60% of the samples) were excluded from further analysis. The library sizes were then scaled by trimmed mean of M values method (). Differentially expressed genes were identified using R package limma () and put through the GREAT software for Gene Ontology pathway analysis.

Lipid extraction

Lipids from cell pellets were extracted using butanol:methanol (1:1) as extraction solvent. The extraction solvent included an internal standard mixture consisting of 17:0 LPC (lysophosphocholine), C14 LPE (lysophosphoethanolamine), C17 ceramide, C17 phosphocholine, C17 phosphoserine, C17:0 PG, C8-ceramide, DMPE (1,2-dimyristoyl-glycero-3-phosphoethanolamine), DMPG [1,2-dimyristoyl-glycero-3-phospho-(1′rac-glycerol)], DMPS (dimyristoyl-glycero-phosphoserine), 12:0 DG (1,2-dilauroyl-sn-glycerol), C8 GluCer (d-glucosyl-β1-1′-N-octanoyl-d-erythro-sphingosine), 24:0 SM (N-lignoceroyl-d-erythro-sphingosylphosphorylcholine), 24:1 SM (N-nervonoyl-d-erythro-sphingosylphosphorylcholine), and d5-TAG 48:0. All the standards were obtained from Avanti Polar Lipids (Alabaster, AL, USA) except for d5-TAG 48:0 that was obtained from CDN Isotopes. All standards were added to a final concentration of 100 ng/ml. One hundred microliters of the extraction solvent including the standards was added to 1 × 107 cells in a vial containing the cell pellet. The mixture was vortexed for 15 s. The samples were then sonicated in a sonicator bath for 60 min at room temperature. The mixture was then centrifuged at 14,000g for 5 min. The supernatant was transferred to new tubes and kept at −80°C until the analysis.

Lipidomic analysis

Samples were analyzed with an LC–tandem MS (LC-MS/MS) system (Agilent 1290 UHPLC connected to Agilent 6490 QQQ). A reversed-phase column (Agilent ZORBAX RRHD Eclipse Plus C18, 2.1 mm by 50 mm, 1.8 μm) was used for separation. Gradient elutions were performed with mobile phase A (40% acetonitrile/60% water with 10 mM ammonium formate) and mobile phase B (90% isopropanol/10% acetonitrile with 10 mM ammonium formate). Quality control (QC) samples were generated by pooling an aliquot of the lipid extracts from each sample. A QC sample was injected every 10 samples during the analysis to monitor the performances of the experimental procedures. Lipids were quantified in positive multiple reaction monitoring mode. MS parameters are as follows: gas temperature, 300°C; gas flow, 5 liters/min; sheath gas flow, 11 liters/min; and sheath gas temperature, 250°C. Data were extracted using MassHunter Quantitative Analysis software (Agilent). Lipids were normalized using both internal standards and cell number. Graphs and statistical analysis were performed using GraphPad Prism.

Extraction of metabolites for acylcarnitine detection

Samples for acylcarnitine quantitation were prepared by first quenching 107 cells with ice-cold 150 mM NaCl (Merck, ACS Reagent Grade) solution. Cells were pelleted by centrifugation, and the supernatant was removed by aspiration. A single-phase liquid extraction protocol was performed to extract metabolites. Briefly, 10 μl of internal spike and 400 μl of cold methanol [100% (v/v); Thermo Fisher Scientific, Optima grade] were added to the tube containing cells. Cells were lysed by scraping the tube on a tube rack and further mixed by vortexing for 30 s. Cell debris were spun to the bottom of the tube by centrifugation at 14,000g for 20 min at 4°C. Supernatant containing metabolites was transferred to a new tube. Further extraction of metabolite was performed by adding 100 μl of cold methanol to the cell pellet and mixed by vortexing for 30 s. Supernatant was collected after tube was centrifuged at 14,000g for 10 min at 4°C. The combined supernatant was dried down completely using a vacuum concentrator (CentriVap, Labconco). Metabolites were reconstituted in 50% (v/v) methanol before LC-MS analysis.

LC-MS for acylcarnitine detection

A series of acylcarnitine standards were performed on the same LC-MS setup to obtain retention time for metabolite matching in the sample analysis. For identification of acylcarnitine species, 4 μl of extracts (in 50% methanol) was drawn into the sample loop at 0.4 ml/min flow rate using 99% of buffer A [1 mM ammonium formate, Sigma-Aldrich, LiChropur grade (pH 3)], and sample directly flowed from the sample loop onto a mixed mode column (ACQUITY Atlantis PREMIER BEH C18 AX column, 1.7 μm, 2.1 mm by 100 mm, Waters, Milford, MA), column thermostated at 45°C. Metabolite separation started after 0.5 min using a linear gradient of 1 to 80% (v/v) buffer B [50% (v/v) isopropanol (Thermo Fisher Scientific, Optima grade)/50% (v/v) acetonitrile (Supelco, LiChrosolv grade)/1 mM ammonium formate (pH 3)] over 8 min, with a wash step of 98% buffer B for 3 min before column equilibration for 2 min in 1% buffer B. LC (ACQUITY UPLC System, Waters) and MS (Q Exactive Hybrid Quadrupole-Orbitrap, Thermo Fisher Scientific) acquisition were controlled by Xcalibur version 2.2 (Thermo Fisher Scientific). MS scans were acquired in the mass range of 70 to 1050 mass/charge ratio at a resolution of 70,000 using HESI (Heated Electrospray Ionization) ion source heated to 400°C and spray voltage of 1500 V (positive mode). AGC (Automated Gain Control) target was set to 1 × 106.

Raw files of all LC-MS runs were converted to mzXML file format using an open source software, MSConvert (http://proteowizard.sourceforge.net/tools.shtml) where MS level was set to 1. Converted files (.mzXML) were then processed for retention time adjustment (aqueous spike) and metabolite matching (in-house libraries) using in-house–built R scripts. A pooled QC mixture was used, and only metabolites with a coefficient of variation of less than 30% were selected for further analyses. Relative quantitation was based on peak intensities that were normalized against total ion count within each sample given the assumption that most of the metabolite in a sample remained the same. Graphs and statistical analysis were performed using GraphPad Prism.

Lipid droplet staining

Intracellular lipid droplets were visualized with fluorescent Bodipy dye. Cells in culture were fixed in 4% paraformaldehyde for 15 min at room temperature, rinsed twice with PBS, permeabilized with 0.2% Triton X-100 in PBS for 10 min, and incubated with Bodipy diluted in PBS (final concentration of 1 μg/ml) for overnight at 4°C. Cells were washed with PBS and stained with DAPI. Fluorescence images were obtained with an Axio Observer D1 epifluorescence microscope with a built-in AxioCam MR3 camera (Zeiss) using the optimal filters and ×100 to ×200 magnifications. Fluorescence intensity of representative images was quantified using ImageJ.

FAO assay

FAO rate was measured using the mitochondrial stress test on the XF96 analyzer (Seahorse Bioscience) according to the manufacturer’s instructions. Cells were seeded in 96-well XF96 cell culture plates at a density of 9000 cells per well in complete media. Cells were then incubated with XF modified DMEM assay medium (Seahorse Bioscience) supplemented with 1 mM glutamine, 0.5 mM glucose, 1% FBS, and 0.5 mM carnitine the next day for 18 hours. On the day of the assay, cells were incubated with FAO assay medium [KHB: 111 mM NaCl, 4.7 mM KCl, 1.25 mM CaCl2, 2 mM MgSO4, and 1.2 mM sodium phosphate buffer supplemented with 2.5 mM glucose, 0.5 mM carnitine, and 5 mM Hepes (adjusted to pH 7.4)] 45 min before the assay in a non-CO2 incubator. Seahorse XF Cell Mito Stress Test compounds (2 μM oligomycin, 2 μM carbonyl cyanide p-trifluoromethoxyphenylhydrazone (FCCP), and 2 μM rotenone). For inhibition of FAO, etomoxir (40 μM) was added 15 min before the assay. Just before starting the assay, XF Palmitate-BSA FAO substrate or BSA was added to the cells. The plate was then put into the XF96 analyzer, and the XF Cell Mito Stress Test was run using the recommended protocol. The rate of FAO was calculated by the difference in maximal respiration when Palmitate-BSA FAO substrate was supplemented relative to BSA control.

Statistical analyses

GraphPad Prism (GraphPad software) v7.0 was used for statistical analyses. All data are presented as means ± SEM except when specified otherwise. Student’s t test (two-tailed) was used to compare two groups and calculate P values. P < 0.05 was considered statistically significant. Significance levels are indicated in relevant figure legends.