Abstract

   Cancer stem cells (CSCs) drive tumor growth, metastasis, and
   chemoresistance. While emerging evidence suggests that CSCs have a
   unique dependency on lipid metabolism, the functions and regulation of
   distinct lipid species in CSCs remain poorly understood. Here, we
   developed a stem cell factor SOX9-based reporter for isolating CSCs in
   primary tumors and metastases of spontaneous mammary tumor models.
   Transcriptomic analyses uncover that SOX9^high CSCs up-regulate the
   ABCA12 lipid transporter. ABCA12 down-regulation impairs cancer
   stemness and chemoresistance. Lipidomic analyses reveal that ABCA12
   maintains cancer stemness and chemoresistance by reducing intracellular
   ceramide abundance, identifying a CSC-associated function of ABCA
   subfamily transporter. Ceramide suppresses cancer stemness by
   inhibiting the YAP-SOX9 signaling pathway in CSCs. Increasing ceramide
   levels in tumors enhances their sensitivity to chemotherapy and
   prevents the enrichment of SOX9^high CSCs. In addition, SOX9^high and
   ABCA12^high cancer cells contribute to chemoresistance in human
   patient-derived xenografts. These findings identify a CSC-suppressing
   lipid metabolism pathway that can be exploited to inhibit CSCs and
   overcome chemoresistance.
     __________________________________________________________________

   CSCs maintain stemness and enhance chemoresistance via the
   ABCA12-mediated expulsion of ceramide lipids that inhibit YAP-SOX9.

INTRODUCTION

   Breast cancer displays high degrees of intratumor heterogeneity (ITH),
   containing distinct subpopulations of cancer cells with diverse
   genetic, epigenetic, phenotypic, or functional characteristics in each
   tumor ([38]1, [39]2). ITH is one of the leading causes of therapeutic
   resistance to various treatments ([40]3). One major source of ITH is
   the ability of cancer cells to acquire functionally diverse cell states
   through differentiation and dedifferentiation processes ([41]4). In
   breast cancer, often only a small population of cancer cells are highly
   tumorigenic and show stem-like properties capable of both self-renewing
   and differentiating into less tumorigenic cells. These cancer stem
   cells (CSCs), defined by their tumor-initiating capacity, drive tumor
   growth, metastasis, and relapse ([42]5–[43]8). However, current
   therapeutic strategies do not effectively target CSCs. Therefore, a
   better understanding of the mechanisms controlling CSC induction and
   maintenance is needed, particularly in spontaneously developing primary
   tumors and metastases that recapitulate ITH. However, these studies
   have been hampered by a lack of specific CSC markers for spontaneous
   tumor models. Although combinations of cell-surface markers have been
   used to enrich CSCs, they are not unique to CSCs and identify rather
   heterogeneous cell populations ([44]9, [45]10).

   Metabolic reprogramming is one of the hallmarks of cancer
   ([46]11–[47]13). Emerging studies demonstrated that CSCs have unique
   metabolic phenotypes and exhibit metabolic plasticity between high
   levels of glycolysis and oxidative phosphorylation, depending on the
   tumor microenvironment ([48]14–[49]16). Furthermore, certain CSCs rely
   on high levels of fatty acid oxidation and cholesterol synthesis
   ([50]17–[51]19). However, whether any lipid pathways can inhibit cancer
   stemness remains poorly understood. As a prominent component of
   membrane lipids and a main class of bioactive lipid molecules, various
   sphingolipids play complex and often opposing functions in cancer. For
   example, ceramide, the central molecule of sphingolipid metabolism, can
   cause cell death and growth arrest. In contrast, its metabolic
   derivative, sphingosine1phosphate, has anti-apoptotic and
   tumor-promoting functions ([52]20). Thus far, it is unclear whether
   sphingolipids play a role in regulating cancer stemness.

   Master regulators of normal stem cell programs are often coopted by
   CSCs to promote cancer stemness. We and others have demonstrated that
   the transcription factor SOX9 acts as a key regulator of mammary stem
   and progenitor cells ([53]21–[54]26). In breast cancer cells, SOX9
   down-regulation inhibits tumor-initiating activity and metastatic
   ability ([55]21, [56]23, [57]26). Furthermore, SOX9 up-regulation
   promotes lineage plasticity and drives the dedifferentiation of luminal
   progenitor cells to generate basal-like breast cancer (BLBC) ([58]26).
   These findings suggest that SOX9 is a key regulator of breast cancer
   stemness and may serve as a specific marker for breast CSCs. In this
   study, we developed spontaneous mouse tumor models expressing a
   Sox9–green fluorescent protein (GFP) transgenic reporter and
   demonstrated that cancer cells expressing high levels of SOX9
   functioned as CSCs in both primary tumors and metastases in distinct
   mammary tumor models. Furthermore, through transcriptomic and lipidomic
   analyses of SOX9^high CSCs, we uncovered a critical dependency of CSCs
   on ceramide homeostasis, providing a strategy for targeting CSCs.

RESULTS

Sox9-GFP reporter recapitulates heterogeneous SOX9 expression in spontaneous
primary tumors and metastases

   Evaluation of SOX9 expression using a single-cell RNA sequencing
   (RNA-seq) dataset ([59]27) revealed substantial ITH of SOX9 levels in
   human breast tumors of various subtypes (fig. S1A). To investigate the
   function of cancer cells with different levels of SOX9, we used the
   C3/TAg BLBC mouse model carrying a Sox9-GFP reporter that expresses GFP
   under the control of the Sox9 enhancer and promoter region ([60]26).
   C3/TAg mice develop a small population (~2%) of Sox9-GFP^high cells in
   the premalignant mammary epithelium, and these cells serve as the
   precursors for tumor development ([61]26). However, SOX9 expression
   patterns in invasive C3/TAg tumors and distant metastases have not been
   determined. We found that C3/TAg tumors contained heterogeneous
   Sox9-GFP–negative and Sox9-GFP–positive cancer cells, and the
   GFP-positive cells can be further stratified into GFP-low and GFP-high
   subpopulations ([62]Fig. 1A). The Sox9-GFP expression faithfully
   reflected the endogenous SOX9 protein levels ([63]Fig. 1B). The
   spatially heterogeneous SOX9 expression in cancer cells was further
   confirmed by immunofluorescence on primary tumor sections
   ([64]Fig. 1C). Analysis of lung metastases in C3/TAg; Sox9-GFP mice
   showed that metastatic tumor cells also expressed variable levels of
   Sox9-GFP ([65]Fig. 1, D and E).

Fig. 1. Sox9-GFP reporter identifies cancer cells expressing distinct levels
of SOX9 in spontaneous primary tumors and metastases.

   [66]Fig. 1.
   [67]Open in a new tab

   (A) Flow cytometry showing the expression of Sox9-GFP in
   C3/TAg;Sox9-GFP tumors (n = 22). CD31, CD45, and Ter-119 were used as
   lineage markers. (B) SOX9 Western blot of cells sorted from
   C3/TAg;Sox9-GFP tumors as in (A). One representative experiment of
   three repeats was shown (n = 3). (C) Representative immunofluorescence
   of SOX9 in C3/TAg;Sox9-GFP primary tumors (n = 3). (D) Representative
   images of lung metastases in C3/TAg;Sox9-GFP mice (n = 4). (E) Flow
   cytometry showing the expression of Sox9-GFP in lung metastases in
   C3/TAg;Sox9-GFP mice (n = 4). (F) Representative immunofluorescence of
   SOX9 in PyMT;Sox9-GFP tumors (type 1: n = 6; type 2: n = 2). (G) Flow
   cytometry measuring Sox9-GFP in PyMT;Sox9-GFP tumors (type 1: n = 6;
   type 2: n = 2). (H) Representative images of lung metastases in
   PyMT;Sox9-GFP mice (type 1: n = 4; type 2: n = 2). (I) Flow cytometry
   measuring Sox9-GFP in lung metastases from PyMT;Sox9-GFP mice (type 1:
   n = 4; type 2: n = 2). DAPI, 4′,6-diamidino-2-phenylindole; PerCP,
   peridinin-chlorophyll-protein; APC, allophycocyanin; SSC, side scatter;
   EpCAM, epithelial cell adhesion molecule.

   We further evaluated the SOX9 expression in MMTV-PyMT (PyMT) tumors
   that model luminal B breast cancer. We observed two types of PyMT
   primary tumors and metastases based on SOX9 expression. Type 1 tumors
   were composed of a majority of SOX9^low tumor cells, while type 2
   contained a majority of SOX9^neg tumor cells ([68]Fig. 1, F to I). In
   both cases, SOX9 protein and Sox9-GFP were heterogeneously expressed in
   cancer cells of PyMT;Sox9-GFP primary tumors and lung metastases
   ([69]Fig. 1, F to I).

SOX9^high cancer cells were enriched in tumor-initiating activity

   To determine the functional difference of SOX9^high cells from other
   cancer cells, we sorted the top 2% Sox9-GFP^high cells from C3/TAg
   tumors by fluorescence-activated cell sorting (FACS) using the Sox9-GFP
   reporter based on the similar frequency of Sox9-GFP^high cells in
   hyperplastic glands ([70]26) and compared them with other cells by
   tumor organoid culture and in vivo transplantation. We observed that
   SOX9^high cells have a markedly higher ability to form tumor organoids
   than other cells ([71]Fig. 2A). In organoid culture, SOX9^high cells
   generated cells with heterogeneous levels of SOX9 same as in the
   original tumor (fig. S2A). Furthermore, SOX9^high cells in organoids
   were enriched in secondary organoid-forming ability, indicating that
   SOX9^high cells have long-term clonogenic potential and can
   differentiate to SOX9^neg/low cells in culture (fig. S2B). Using in
   vivo limiting dilution transplantation assay, we showed that only
   SOX9-expressing C3/TAg tumor cells could initiate new tumors, and the
   SOX9^high cells had a further ~4-fold increase in tumor-initiating
   capacity compared to the SOX9^low cells ([72]Fig. 2, B and C).

Fig. 2. SOX9^high cancer cells are enriched in tumor-initiating activity.

   [73]Fig. 2.
   [74]Open in a new tab

   (A) Representative images and quantification of organoid-forming
   efficiency of Sox9^neg, Sox9^low, and Sox9^high cells sorted from
   C3/TAg;Sox9-GFP primary tumors (n = 5). (B) Tumor-initiating cell (TIC)
   frequency tumor cells sorted from C3/TAg;Sox9-GFP primary tumors.
   Combination of two experiments using two different spontaneous tumors
   was shown. CI, confidence interval. (C) Tumor onset in NOD/SCID mice
   transplanted with 10,000 tumor cells sorted from C3/TAg;Sox9-GFP
   tumors. Combination of two experiments using two different spontaneous
   tumors was shown (n = 8 per group). (D) Organoid forming ability during
   serial replating of tumor cells sorted from PyMT;Sox9-GFP primary
   tumors (n = 4; circles indicate type-1 tumors, and triangles indicate
   type-2 tumors). (E) Representative images and quantification of
   organoid-forming efficiency of cells sorted from lung metastases in
   C3/TAg;Sox9-GFP mice (n = 4). (F) Representative images and
   quantification of organoid-forming efficiency during serial replating
   of cells sorted from lung metastases in PyMT;Sox9-GFP (n.d., not
   determined). Combination of two experiments using two different lung
   metastases was shown. (G) Correlation of stemness score and SOX9
   expression in breast tumors at the single-cell level as determined by
   CancerSEA [[75]GSE77308 ([76]29) and [77]GSE75688 ([78]27)]. All data
   are represented as means ± SEM. P values were determined by one-way
   ANOVA with Tukey’s test [(A), (E), and (F), first passage], Chi-square
   test in Extreme Limiting Dilution Analysis (ELDA) (B), log-rank test
   (C), one-way ANOVA with Tukey’s test (D), or unpaired two-tailed t test
   [(F), second and third passages]. ****P < 0.0001, ***P < 0.001, **P <
   0.01, and *P < 0.05; ns, not significant.

   We further characterized SOX9^high cells in PyMT tumors by organoid
   culture. By subjecting freshly sorted primary tumor cells to a serial
   organoid replating assay, we found that only SOX9^high cells maintained
   robust organoid-forming capacity during the serial replating, while
   SOX9^neg and SOX9^low cells gradually lost the ability to generate
   organoids after two to three passages ([79]Fig. 2D).

   We next compared SOX9^high and other cancer cells FACS sorted from lung
   metastases in C3/TAg and PyMT models as shown in [80]Fig. 1 (E and I).
   Such as in primary tumors, SOX9^high cells in C3/TAg metastases
   exhibited markedly higher organoid-forming capacity than other cell
   types ([81]Fig. 2E). Markedly increased organoid-forming ability was
   also found for SOX9^high cells in PyMT metastases ([82]Fig. 2F).
   Together, these data suggested that SOX9^high cancer cells in both
   primary tumors and metastases are enriched in CSCs, defined by their
   tumor-initiating ability. This notion is further supported by the
   correlation of high SOX9 expression with increased stemness scores at
   the single-cell level in multiple human cancer types, including breast,
   kidney, brain, and lung cancers, as revealed by CancerSEA analysis
   ([83]Fig. 2G and fig. S2C) ([84]27–[85]29).

SOX9^high tumor cells have increased metastasis-initiating capacity

   Analysis of breast patient-derived xenograft (PDX) single-cell RNA-seq
   data showed that SOX9 levels were positively correlated with
   metastasis, epithelial-mesenchymal transition, and invasion signatures,
   suggesting that SOX9^high cancer cells have higher metastatic potential
   ([86]Fig. 3A). Consistently, freshly sorted SOX9^high C3/TAg primary
   tumor cells exhibited increased transwell migration and invasion
   capacity compared to SOX9^low cells ([87]Fig. 3, B and C). Organoids
   formed by SOX9^high cells also showed more invasive morphology compared
   to that by SOX9^low cells when cultured in three-dimensional (3D)
   collagen gel ([88]Fig. 3D). Furthermore, in the experimental metastasis
   assay via tail vein injection, SOX9^high C3/TAg cells developed
   significantly more lung metastases than SOX9^low cells ([89]Fig. 3E).
   SOX9^high cells sorted from PyMT tumors similarly exhibited
   significantly higher metastatic ability compared to SOX9^neg/low cells
   ([90]Fig. 3, F and G). In these experiments, we isolated tumors from
   PyMT mice carrying a constitutively expressing CAG::mRFP1 (monomeric
   red fluorescent protein 1) transgene to facilitate the detection of
   metastases, although the mRFP expression was suppressed in some
   metastatic nodules. Together, these results suggest that SOX9^high CSCs
   contribute to metastasis in different breast tumor models.

Fig. 3. SOX9^high tumor cells have increased metastasis-initiating capacity.

   [91]Fig. 3.
   [92]Open in a new tab

   (A) Correlation between SOX9 levels and functional states in breast
   tumors at the single-cell level as determined by CancerSEA. EMT,
   Epithelial-Mesenchymal Transition. (B) Transwell migration assay of
   C3/TAg primary tumor cells (n = 5 tumors). (C) Transwell invasion assay
   of C3/TAg primary tumor cells (n = 5 tumors). (D) Representative images
   and frequencies of organoids with the invasive morphology in collagen
   gel (n = 3 tumors). (E) Lung metastasis-initiating ability of Sox9^low
   and Sox9^high C3/TAg primary tumor cells by tail vein injection. Each
   line indicates one experiment with a distinct tumor (n = 4). (F) Lung
   metastasis-initiating ability of Sox9^neg, Sox9^low, and Sox9^high PyMT
   tumor cells by tail vein injection. Metastatic nodules were determined
   after 14 weeks (triangles and circles representing two different
   donors). (G) Lung metastasis-initiating ability of Sox9^neg, Sox9^low,
   and Sox9^high PyMT tumor cells by tail vein injection. Metastatic
   nodules were determined 14 weeks after injection (n = 3). All data are
   represented as means ± SEM. P values were determined by two-tailed t
   test [(B), (C), (D), (E), and (G)] or one-way ANOVA with Tukey’s test
   (F). ***P < 0.001, **P < 0.01, and *P < 0.05.

SOX9^high tumor cells are enriched after chemotherapy

   Resistance to chemotherapy is a cardinal feature of CSCs ([93]6,
   [94]30–[95]32). To test whether SOX9^high tumor cells are
   chemoresistant, we treated tumor organoids derived from C3/TAg;Sox9-GFP
   tumors with vehicle or chemotherapeutic drugs, including doxorubicin,
   cisplatin, or paclitaxel. Each of the drugs induced substantial cell
   death after a 3-day treatment. However, the frequency of Sox9-GFP^high
   cells was significantly increased among tumor cells that survived the
   drug treatment (fig. S3, A and B). We then tested the response of
   SOX9^high cells in vivo by treating C3/TAg;Sox9-GFP allograft tumors
   with the chemotherapy drugs. While doxorubicin and cisplatin markedly
   inhibited tumor growth ([96]Fig. 4, A and E), the overall Sox9-GFP
   expression levels and the frequency of Sox9-GFP^high cells were
   significantly increased in doxorubicin or cisplatin-treated tumors
   ([97]Fig. 4, B to D and F to H). The frequency of SOX9^high cells was
   also increased significantly in paclitaxel-treated tumors, although it
   had a modest effect on tumor growth ([98]Fig. 4, I to L).

Fig. 4. Chemotherapy enriches SOX9^high tumor cells.

   [99]Fig. 4.
   [100]Open in a new tab

   (A) Growth of tumors treated with Vehicle (n = 4) or doxorubicin (n =
   2). (B to D) Flow cytometric profiles (B), Sox9-GFP mean fluorescence
   intensity (MFI) (C), and the percentage of SOX9^high cells (D) in
   tumors treated in (A). (E) Growth of tumors treated with vehicle (n =
   2) or cisplatin (n = 4). (F to H) Flow profiles (F), Sox9-GFP MFI (G),
   and the percentage of SOX9^high cells (H) in tumors treated in (E). (I)
   Growth of tumors treated with vehicle (n = 6) or paclitaxel (n = 4). (J
   to L) Flow profiles (J), Sox9-GFP MFI (K), and the percentage of
   SOX9^high cells (L) in tumors treated in (I). (M) Growth of tumors
   treated with vehicle (n = 8) or cisplatin (n = 10). (N to P) Flow
   cytometric profiles (N), Sox9-GFP MFI (O), and the percentage of
   SOX9^high cells (P) in tumors treated in (M). (Q) Growth of tumors
   treated with vehicle (n = 6), doxorubicin (n = 4), or paclitaxel (n =
   6). (R and S) Flow cytometric profiles (R) and the percentage of
   SOX9^high cells in tumors treated in (Q). (T) Growth of PDXs (H39:
   vehicle, n = 6; cisplatin, n = 8; H17: vehicle, n = 9; cisplatin, n =
   10). (U) IHC and quantification of SOX9 expression in PDXs treated in
   (T). (V) SOX9 expression in matched pre- and post-neoadjuvant tumors.
   All data are represented as means ± SEM. P values were determined by
   two-way ANOVA [(A), (E), (I), (M), (Q), and (T)], multiple unpaired t
   test (U), unpaired [(C), (D), (G), (H), (K), (L), (O), and (P)] or
   paired (V) t test, or one-way ANOVA (S). ****P < 0.0001, ***P < 0.001,
   **P < 0.01, and *P < 0.05.

   To assess the chemoresistant property of SOX9^high cells in additional
   tumor models, we treated PyMT;Sox9-GFP tumor organoids with
   doxorubicin, paclitaxel, cisplatin, or vehicle. Consistent with the
   observations in C3/TAg organoids, the drug treatment caused substantial
   cell death but greatly increased the frequency of Sox9-GFP^high cells
   in the surviving tumor cells (fig. S3, C and D). We then treated
   PyMT;Sox9-GFP tumors with cisplatin in vivo. In two cohorts of
   allografts generated from different PyMT spontaneous tumors, tumor
   growth was markedly inhibited by cisplatin, whereas Sox9-GFP expression
   and Sox9-GFP^high cell frequency were increased significantly
   ([101]Fig. 4, M to P, and fig. S3, E to G). Likewise, doxorubicin or
   paclitaxel treatment resulted in a significant enrichment of
   Sox9-GFP^high cells in PyMT tumors ([102]Fig. 4, Q to S).

   We further assessed the effect of chemotherapy on SOX9 expression in
   two human triple-negative breast cancer PDX models established
   previously ([103]33). Cisplatin inhibited the tumor growth of both PDX
   models ([104]Fig. 4T). However, SOX9 levels were significantly
   increased in both PDXs after cisplatin treatment, consistent with mouse
   tumor models ([105]Fig. 4U). Furthermore, by analyzing RNA-seq data of
   20 matching pre- and post-neoadjuvant chemotherapy breast cancer
   patient samples ([106]34), we found that SOX9 levels significantly
   increased in the post-treatment samples ([107]Fig. 4V). Together, these
   data showed that SOX9^high cancer cells are chemoresistant and enriched
   by chemotherapy in both mouse tumor models and human breast cancer,
   further supporting the notion that SOX9^high cells have properties of
   CSCs.

SOX9^high CSCs up-regulate the ABCA12 lipid transporter

   Specific identification of breast CSCs in spontaneous tumor models
   allowed us to investigate molecular determinants of cancer stemness in
   vivo. We performed RNA-seq comparing SOX9^high cells to other cancer
   cells freshly isolated from C3/TAg;Sox9-GFP tumors. While >2000 genes
   were significantly differentially expressed between SOX9^+ and SOX9^−
   cells, only 9 genes were found significantly different between
   SOX9^high and SOX9^low cells ([108]Fig. 5A and fig. S4A). Among them,
   only Abca12 and Mogat1 were consistently correlated with the Sox9
   levels, as confirmed by reverse transcription quantitative polymerase
   chain reaction (RT-qPCR) (fig. S4B). Abca12 was also the most
   up-regulated gene in SOX9^high luminal/basal hybrid cells compared to
   SOX9^low cells in C3/TAg premalignant lesions, whereas Mogat1 was not
   significantly changed (fig. S4C) ([109]26). In the normal mammary
   gland, Abca12 is exclusively expressed in fetal mammary stem cells, but
   not in adult lineage-committed epithelial cells, suggesting that
   increased Abca12 expression in C3/TAg tumors is associated with
   reactivation of embryonic multipotency (fig. S4D) ([110]35). FACS
   analysis confirmed that ABCA12 protein levels are up-regulated along
   with the increased SOX9 expression in C3/TAg tumor organoid cells (fig.
   S4, E and F). Furthermore, Abca12 expression was positively correlated
   with the Sox9 levels in PyMT tumors (fig. S4G). SOX9 overexpression in
   C3/TAg tumor cells also significantly elevated Abca12 levels
   ([111]Fig. 5B), although we cannot distinguish whether this is due to
   direct transcriptional regulation or an indirect effect of increasing
   CSC frequency by SOX9. Together, these data showed that ABCA12 is
   up-regulated in SOX9^high breast CSCs.

Fig. 5. ABCA12 lipid transporter is up-regulated in SOX9^high cells and
required for CSC activity.

   [112]Fig. 5.
   [113]Open in a new tab

   (A) Volcano plot of RNA-seq comparing SOX9^low and SOX9^high cells from
   C3/TAg;Sox9-GFP tumors (n = 4). (B) Sox9 and Abca12 mRNA levels in
   pLVX-ctrl and pLVX-SOX9 transduced C3/TAg organoids (n = 4). (C) UMAP
   showing SOX9 and ABCA12 levels in tumor cells in patients with breast
   cancer (BRCA) ([114]GSE176078). (D) ABCA12 expression in PDX tumors
   treated with vehicle or cisplatin as in [115]Fig. 4T (H39: n = 8 each
   group; H17: vehicle, n = 8; cisplatin, n = 9). (E) Organoid-forming
   efficiency of sgNT and sgAbca12 C3/TAg;Sox9-GFP tumor cells (n = 4).
   (F) Tumor-initiating cell frequency of sgNT and sgAbca12
   C3/TAg;Sox9-GFP organoid cells, as determined by ELDA (n = 2). (G)
   Growth of sgNT and sgAbca12 C3/TA;Sox9-GFP tumor (sgAbca12/cisplatin, n
   = 3; other groups, n = 4). (H and I) Sox9-GFP MFI (H) and percentage of
   Sox9-GFP^high cells (I) in sgNT and sgAbca12 C3/TAg;Sox9-GFP tumors
   treated in (G). (J) SOX9 Western blot of the indicated cells. (K)
   Organoid-forming efficiency of control (sgNT) or Abca12-overexpressing
   (CRISPRa-Abca12) C3/TAg;Sox9-GFP tumor cells. (L) Effect of cisplatin
   on organoid formation in control (sgNT) or Abca12-overexpressing
   (CRISPRa-Abca12 #b) C3/TAg;Sox9-GFP tumor cells. IC[50], median
   inhibitory concentration. (M) Organoid-forming efficiency of the
   indicated C3/TAg;Sox9-GFP tumor cells. (N) Organoid-forming efficiency
   of sgNT or sgABCA12-transduced human cell lines. (O) Doxorubicin dose
   response of sgNT and sgABCA12-transduced breast cancer cells in
   organoid culture. All data are represented as means ± SEM. P values
   were determined by unpaired two-tailed t test [(B), (D), (E), and (N)],
   one-way ANOVA with Dunnett’s (K) or Tukey’s test [(H) to (I) and (M)],
   Chi-square test in ELDA (F), ordinary two-way ANOVA [(L) and (O)], or
   two-way RM ANOVA (G). ****P < 0.0001, ***P < 0.001, **P < 0.01, and *P
   < 0.05.

   We investigated the expression of ABCA12 in human breast cancer.
   Single-cell RNA-seq analysis of 26 human breast tumors ([116]GSE176078)
   ([117]36) revealed a positive correlation between SOX9 and ABCA12
   expression at the single-cell level ([118]Fig. 5C). Analysis of The
   Cancer Genome Atlas data using GEPIA2 ([119]37) showed that ABCA12 is
   overexpressed in human breast cancer compared to normal breast tissue
   (fig. S4H). In addition, ABCA12 is one of the most highly up-regulated
   adenosine 5′-triphosphate (ATP)–binding cassette transporter genes in
   breast tumors compared to normal breast tissue (fig. S4I) ([120]38).
   ABCA12 level is markedly higher in breast tumors that retained residual
   disease after neoadjuvant chemotherapy than those that had
   pathologically complete responses (fig. S4J) ([121]39). Further
   supporting the role of ABCA12 in chemoresistance, the frequencies of
   ABCA12^+ cells were markedly increased in PDX tumors after cisplatin
   treatment ([122]Fig. 5D).

ABCA12 promotes cancer stemness and chemoresistance of SOX9^high cells

   To determine the function of ABCA12 in CSCs, we deleted Abca12 in
   C3/TAg;Sox9-GFP tumor organoids using the lentiCRISPRv2 vector, which
   led to a marked decrease in both ABCA12 RNA and protein levels (fig.
   S5, A and B). We found that, in three of the four tumor organoid lines,
   Abca12 deletion significantly inhibited organoid formation
   ([123]Fig. 5E). We further assessed the impact of Abca12 deletion on
   tumor-initiating ability by in vivo limiting dilution transplantation
   assay. In two independent tumor organoid lines, Abca12 deletion
   markedly reduced the tumor-initiating ability (8- to 30-fold)
   ([124]Fig. 5F).

   We next evaluated whether ABCA12 contributed to chemoresistance in
   SOX9^high cells. Cisplatin treatment of both wild-type (WT) and
   Abca12-deleted C3/TAg;Sox9-GFP tumor organoids led to pronounced cell
   death. However, SOX9^high cells were no longer enriched by cisplatin
   treatment in the Abca12-deleted organoids (fig. S5, C to E).
   Furthermore, the loss of Abca12 enhanced the inhibitory effects of
   cisplatin on tumor organoid formation (fig. S5F). To extend this
   finding in vivo, we transplanted a saturated number (3 × 10^5) of WT
   and Abca12-deleted C3/TAg tumor organoid cells to allow them to form
   tumors at similar take rates. Under this condition, WT and
   Abca12-deleted tumors grew at the same speed and responded to cisplatin
   similarly ([125]Fig. 5G). It is possible that CSCs can activate
   ABCA12-independent compensatory mechanisms to maintain stemness and
   chemoresistance when ABCA12 is chronically knocked out. Injecting large
   numbers of tumor cells may have facilitated the generation and/or
   selection of these ABCA12-independent CSCs. Nevertheless, in
   Abca12-deleted tumors, SOX9^high cells were no longer enriched by
   cisplatin treatment ([126]Fig. 5, H and I, and fig. S5G).

   We then tested the effect of ABCA12 overexpression using a CRISPR
   activation (CRISPRa) approach ([127]40). Coexpression of the
   dCas9-VP64-p65-Rta (VPR) transcriptional activator and single guide
   RNAs (sgRNAs) targeting the Abca12 promoter region markedly increased
   Abca12 expression in C3/TAg tumor cells (fig. S5H). Abca12
   overexpression led to a marked up-regulation of SOX9 and significantly
   increased organoid formation and resistance to cisplatin ([128]Fig. 5,
   J to L, and fig. S5, I and J). More remarkably, Abca12 overexpression
   was sufficient to induce organoid-forming ability in SOX9^neg non-CSCs
   ([129]Fig. 5M).

   To assess the ABCA12 function in human cells, we deleted ABCA12 in a
   panel of human breast cancer cell lines, including two luminal lines
   (T47D and MCF7) and three triple-negative lines (MDA-MB-231,
   MDA-MB-468, and HCC1937), with the lentiCRISPRv2 vector (fig. S5K).
   Notably, loss of ABCA12 significantly impaired the organoid formation
   ability in four of the five cell lines except for HCC1937 ([130]Fig. 5N
   and fig. S5L). Furthermore, ABCA12 knockout significantly increased the
   sensitivity to doxorubicin in most cell lines except MCF7
   ([131]Fig. 5O). These results reinforced the role of ABCA12 in
   promoting breast cancer stemness, although various cell lines have
   slightly different dependencies on ABCA12, likely due to heterogeneous
   genetic and epigenetic makeup of these cell lines.

Lipidomic analyses uncovered that ABCA12 regulates SOX9 expression and cancer
stemness by controlling ceramide abundance

   ABCA12 is a member of the ATP-binding cassette transporter family and
   functions as a lipid transporter ([132]41–[133]45), suggesting that
   SOX9^high CSCs may have unique lipid metabolism. Therefore, we
   performed untargeted liquid chromatography–mass spectrometry
   (LC-MS)–based lipidomics to compare SOX9^neg, SOX9^low, and SOX9^high
   cells sorted from C3/TAg;Sox9-GFP tumors. Overall, 1144 lipid species
   from 30 lipid classes were detected, indicating high sensitivity of
   lipidomics using small numbers of sorted cells (fig. S6A). The
   abundances of 230 lipid species from 19 lipid classes were
   significantly altered between SOX9^neg and SOX9^high cells (fig. S6B).
   Most of these lipid species were down-regulated in SOX9^high cells,
   except for 12 species that were up-regulated (fig. S6C). Among the 230
   down-regulated lipid species in SOX9^high cells, 18 of them also showed
   a consistent down-regulation comparing SOX9^high to SOX9^low and
   SOX9^low to SOX9^neg cells (fold change > 2, P < 0.05) ([134]Fig. 6A).
   Most of these lipid species (12 of 18) were ceramides and their
   derivatives, such as Hex1Cer, CerG3GNAc1, and sphingomyelin.
   Furthermore, all the significantly altered ceramide species between
   SOX9^neg and SOX9^high cells (P < 0.05) showed a gradually decreasing
   trend when SOX9 expression increased ([135]Fig. 6B). Lipid pathway
   enrichment analysis using Lipid Pathway Enrichment Analysis (LIPEA)
   showed significant down-regulation of sphingolipid signaling and
   metabolism pathways in SOX9^high CSCs, concordant with the role of
   ceramide as a central component of sphingolipid metabolism
   ([136]Fig. 6C).

Fig. 6. Lipidomic analyses reveal that ABCA12 reduces intracellular ceramide
abundance to maintain cancer stemness.

   [137]Fig. 6.
   [138]Open in a new tab

   (A) Lipid species significantly down-regulated in SOX9^high cells (fold
   change > 2) in C3/TAg;Sox9-GFP tumor (n = 5). (B) Level of ceramide
   lipids in Sox9^neg, Sox9^low, and Sox9^high C3/TAg tumor cells. (C)
   Lipid pathway enrichment analyses of the lipids identified in (A). AGE,
   advanced glycation end products; RAGE, receptor for AGE. (D) Flow
   cytometry measuring the abundance of ceramide in the SOX9^neg,
   Sox9^low, and Sox9^high C3/TAg tumor cells (n = 7). (E) Abundance of
   ceramide in the SOX9^neg, Sox9^low, and Sox9^high PyMT tumor cells (n =
   5). (F) Abundance of ceramide species in sgNT and sgAbca12 C3/TAg
   organoids (n = 3). (G) Ceramide levels in C3/TAg organoids treated with
   dimethyl sulfoxide (DMSO) or D-NMAPPD for 5 days as measured by flow
   cytometry (n = 5). (H) Effect of C6-ceramide or D-NMPAAD on C3/TAg
   organoid formation (n = 6). (I) Ceramide levels in PyMT organoids
   treated with DMSO or D-NMAPPD (10 μM) for 5 days (n = 5). (J) Effect of
   C6-ceramide or D-NMPAAD on PyMT organoid formation (n = 3). (K and L)
   Organoid-forming efficiency of C3/TAg [(K), n = 5] or PyMT [(L), n = 3]
   tumor cells pretreated with D-NMAPPD (20 μM) for 3 days. (M) Ceramide
   levels in C3/TAg organoids treated with DMSO or Fumonisin B1 (FB1; 10
   μM) for 5 days (n = 4). (N and O) Effect of FB1 on organoid formation
   by SOX9^neg/low C3/TAg [(N), n = 2] or SOX9^neg/low PyMT [(O), n = 3]
   tumor cells. All data are represented as means ± SEM. P values were
   determined by one-way ANOVA [(D) and (E)], or with Tukey’s test [(H),
   (J), (N), and (O)], or paired two-tailed t test [(G), (I), and (K) to
   (M)]. ***P < 0.001, **P < 0.01, and *P < 0.05.

   We then performed flow cytometry to measure ceramide levels in
   SOX9^neg, SOX9^low, and SOX9^high tumor cells from C3/TAg and PyMT
   tumors. Consistent with the lipidomics analysis, the ceramide level was
   markedly decreased in SOX9^high cells in both tumor models
   ([139]Fig. 6, D and E, and fig. S6, D and E). Similar differences were
   observed in C3/TAg tumor organoid cells (fig. S6F). Previous studies
   showed that ABCA12 is required for the secretion of ceramide by
   keratinocytes to form extracellular lipid layers that are essential for
   skin barrier function ([140]44). This suggests that high ABCA12
   expression in SOX9^high CSCs maintains low intracellular ceramide
   levels by efflux. To test this, we compared lipid profiles of WT and
   Abca12-deleted C3/TAg tumor cells by nontargeted LC-MS–based lipidomics
   and found that intracellular ceramides levels were indeed increased
   after Abca12 deletion ([141]Fig. 6F). These results suggest that
   SOX9^high CSCs maintain low ceramide levels via ABCA12-mediated efflux.

   We then explored whether ceramide levels affect CSC activity. We
   treated C3/TAg and PyMT tumor organoids with exogenous C6-ceramide or
   with the acid ceramidase inhibitor D-NMAPPD
   {N-[(1R,2R)-2-hydroxy-1-(hydroxymethyl)-2-(4-nitrophenyl)ethyl]-tetrade
   canamide; (1R,2R)-B13} to increase intracellular ceramide levels
   ([142]Fig. 6, G and I, and fig. S6, G and H) ([143]46). Both treatments
   markedly inhibited tumor organoid-forming ability ([144]Fig. 6, H and
   J). To further evaluate the long-term effect of ceramide on tumor
   stemness, we pretreated C3/TAg and PyMT tumor cells with D-NMAPPD and
   then measured their organoid-forming efficiency without further
   treatment. We observed that D-NMAPPD pretreatment inhibited organoid
   formation in both tumor models, suggesting that D-NMAPPD led to
   irreversible inhibition of cancer stemness ([145]Fig. 6, K and L).
   Conversely, we used fumonisin B1 (FB1), a ceramide synthase inhibitor,
   to reduce the intracellular accumulation of ceramides in C3/TAg cells
   ([146]Fig. 6M and fig. S6G). We found that FB1 treatment significantly
   increased the organoid-forming efficiency of SOX9^neg/low cells
   ([147]Fig. 6N). Similar effects of FB1 were confirmed in PyMT tumor
   organoids ([148]Fig. 6O).

Ceramide suppresses SOX9 expression in tumor cells by inhibiting the YAP/TAZ
pathway

   Increasing intracellular ceramide levels by D-NMAPPD greatly reduced
   the expression of Sox9-GFP reporter and endogenous SOX9 protein levels
   in C3/TAg and PyMT organoid cells, consistent with the decreased
   organoid-forming ability ([149]Fig. 7, A to C, and fig. S6, G to I).
   Ceramide has been shown as an inhibitor for the Yes-associated protein
   1 (YAP)/transcriptional coactivator with PDZ-binding motif (TAZ)
   signaling in hepatic stellate cells ([150]47). Therefore, we asked
   whether ceramide regulates the YAP/TAZ pathway to control breast cancer
   stemness. The elevation of ceramide abundance with C6-ceramide or
   D-NMAPPD treatment greatly inhibited YAP/TAZ signaling in C3/TAg and
   PyMT tumor organoids ([151]Fig. 7, D to G, and fig. S7, A and B).
   Furthermore, inhibiting YAP by verteporfin caused a marked reduction of
   the SOX9 protein level ([152]Fig. 7H) and organoid-forming ability of
   C3/TAg organoids ([153]Fig. 7I and fig. S7C).

Fig. 7. Ceramide down-regulates SOX9 expression by inhibiting the YAP/TAZ
pathway.

   [154]Fig. 7.
   [155]Open in a new tab

   (A and B) Sox9-GFP expression in C3/TAg (A) or PyMT (B) organoids
   treated for 5 days (n = 5). (C) SOX9 levels in D-NMAPPD–treated
   organoids (n = 3). (D) RT-qPCR measuring YAP target genes in C3/TAg
   organoids (n = 4) treated for 6 hours. (E) YAP and TAZ levels in
   D-NMAPPD–treated C3/TAg organoids (n = 3). (F and G) Immunofluorescence
   (F) and quantification of nuclear YAP (G) in C3/TAg cells treated for 2
   days (n = 100 cells per group). (H) SOX9 and YAP levels in C3/TAg
   organoids treated for 3 days. (I) Organoid-forming efficiency of C3/TAg
   cells (n = 3). (J) Sox9-GFP level in C3/TAg organoids treated for 5
   days. (K) SOX9 levels in organoids treated with myriocin (2 μM) or FB1
   (10 μM) for 5 days. (L and M) Immunofluorescence (L) and quantification
   (M) of nuclear YAP in C3/TAg cells as treated for 2 days (n = 100 cells
   per group). (G) and (M) share the same DMSO group. (N) YAP and TAZ
   levels in C3/TAg organoids as treated for 5 days. (O) Effect of YAP
   (S6A) on SOX9 levels in C3/TAg organoids. (P and Q) Effect of YAP (S6A)
   on SOX9 levels (P) and organoid-forming efficiency (Q) in sgAbca12
   C3/TAg organoids. (R) Cisplatin dose-response curve of the indicated
   cells. (S) Growth of C3/TAg;Sox9-GFP tumors treated as indicated. (T
   and U) Sox9-GFP MFI (T) and percentage of SOX9^high cells (U) in tumors
   as shown in (S). All data are represented as means ± SEM. P values were
   determined by paired t test [(A), (B), (D), and (J)], one-way ANOVA
   with Dunnett’s test [(G), (I), and (M)], Tukey’s test (Q), ordinary
   one-way ANOVA [(T) and (U)], or two-way ANOVA [(R) and (S)]. ****P <
   0.0001, ***P < 0.001, **P < 0.01, and *P < 0.05.

   In contrast, reducing ceramide levels by FB1 or myriocin (a serine
   palmitoyltransferase inhibitor) markedly increased SOX9 and YAP/TAZ
   levels ([156]Fig. 7, J to N, and figs. S6, G and J, and S7D).
   Expressing constitutively active YAP (S6A) ([157]48) led to the
   up-regulation of SOX9 protein levels ([158]Fig. 7O), supporting the
   role of YAP in SOX9 up-regulation. This is likely achieved by the
   binding of the YAP/TEA domain transcription factor (TEAD) complex to
   the SOX9 promoter, as previously reported in esophageal cancer
   ([159]49, [160]50). YAP activation also significantly increased
   organoid formation and chemoresistance (fig. S7, E and F). Furthermore,
   YAP activation restored SOX9 expression, organoid formation, and
   cisplatin resistance in ABCA12 knockdown cells ([161]Fig. 7, P to R,
   and fig. S7G), supporting the notion that ABCA12 enhances cancer
   stemness through activating the YAP pathway.

   We then investigated whether increasing ceramide levels by D-NMAPPD
   could suppress chemotherapy-induced enrichment of SOX9^high CSCs and
   inhibit tumor progression synergistically with cisplatin in C3/TAg
   allograft models. D-NMAPPD or cisplatin treatment alone led to a
   significant reduction in tumor growth, although neither could induce
   tumor regression ([162]Fig. 7S). The combination of two drugs led to
   marked tumor regression ([163]Fig. 7S). Furthermore, combinatory
   treatment completely blunted the increase of Sox9-GFP expression and
   Sox9-GFP^high cell frequency that would be induced by cisplatin
   treatment alone ([164]Fig. 7, T to U). We note that D-NMAPPD treatment
   showed a more pronounced effect than ABCA12 knockdown ([165]Fig. 5G) on
   tumor growth and response to cisplatin. It is possible that acute
   D-NMAPPD treatment directly increases ceramide levels that are less
   likely to be overcome by compensatory mechanisms that may occur in
   ABCA12 knockdown cells. Together, our results showed that high ceramide
   levels inhibit SOX9 expression by inactivation of the YAP/TAZ pathway
   and that increasing ceramide levels pharmacologically inhibits CSCs and
   overcomes chemoresistance.

DISCUSSION

   Understanding the determinants of CSCs in spontaneous tumors will help
   develop targeted strategies against these clinically important cell
   types. Using a CSC reporter based on the stemness regulator SOX9, we
   established an effective method for identifying CSCs in both primary
   tumors and metastases of different mammary tumor models. We showed that
   SOX9^high CSCs were metastasis initiating and chemoresistant,
   highlighting the need to target these cells. Specific isolation of CSCs
   from spontaneous tumors enabled molecular characterization of these
   cells by transcriptomic and lipidomic profiling. We found an important
   role of ceramide homeostasis in controlling breast cancer stemness and
   chemoresistance. SOX9^high CSCs expressed high levels of ABCA12, which
   promoted ceramide efflux to maintain low cellular ceramide. Elevating
   ceramide levels down-regulated SOX9 by suppressing YAP/TAZ signaling,
   inhibited cancer stemness, and overcame chemoresistance. These results
   uncovered a critical lipid metabolism mediated signaling circuitry for
   sustaining breast cancer stemness (fig. S7H).

   Reliable identification of the CSC state in spontaneously developing
   tumors has been challenging ([166]6, [167]51, [168]52). Several studies
   have used cell surface markers for enriching CSCs in transgenic tumor
   models, such as CD29^hiCD24^hi or CD29^hiCD24^med ([169]53, [170]54),
   Thy1^+CD24^+ ([171]55), and CD61^+ ([172]56). However, these markers
   are not specific or functionally linked to cancer stemness ([173]9,
   [174]10). Reporters controlled by core embryonic stem cell regulators,
   such as OCT4/SOX2 and NANOG, have been developed to identify breast
   CSCs in cancer cell lines ([175]57, [176]58). However, their utilities
   in spontaneous tumors remain unclear. Recently, the stem-SH2-containing
   5′-inositol phosphatase (s-SHIP) promoter activity has been used to
   enrich CSCs in C3/TAg tumors, but substantial tumor-initiating activity
   can still be found in s-SHIP promoter inactive cells, suggesting that
   it only marks a subset of CSCs ([177]59). Here, using the mammary stem
   cell regulator SOX9, we developed a specific CSC reporter for various
   spontaneous mammary tumor models. We found a strict association of
   cancer stemness with Sox9-GFP expression. Only Sox9-GFP^+ cancer cells
   could generate organoids in vitro and reinitiate tumors in vivo,
   whereas Sox9-GFP^− cells completely lacked organoid-forming and
   tumor-initiating activity. Among Sox9-GFP^+ cells, Sox9-GFP^high cells
   showed markedly higher CSC activity than Sox9-GFP^low cells. The SOX9
   reporter can identify CSCs from primary tumors and metastases of
   different mammary tumor models. SOX9^high CSCs contributed to
   metastasis and chemoresistance, highlighting the importance of these
   cells. Thus, our Sox9-GFP reporter provides a much-needed experimental
   tool for studying CSCs in spontaneous tumors.

   Accumulating evidence shows that CSCs rely on altered lipid metabolism.
   However, current understanding is limited to the CSC-promoting role of
   fatty acid oxidation and cholesterol synthesis ([178]17–[179]19,
   [180]60, [181]61). In this study, we found a critical need for CSCs to
   control their intracellular ceramide level. Ceramide serves as the hub
   for sphingolipid metabolism and has well-established roles in inducing
   apoptosis, autophagy, and cell cycle arrest ([182]62, [183]63). Here,
   we identified a function of ceramide in inhibiting stemness by
   suppressing stem cell regulator expression. In C3/TAg tumor cells,
   ceramide abundance was negatively correlated with SOX9 levels.
   Increasing ceramide levels markedly decreased SOX9 expression in tumor
   organoids, which, in turn, impaired tumor organoid formation.
   Conversely, lowering intracellular ceramide levels by blocking its
   synthesis up-regulated SOX9 expression and promoted organoid formation.
   We further demonstrated that ceramide inhibits the YAP/TAZ signaling
   that is required for SOX9 expression in breast CSCs.

   We uncovered a unique mechanism by which SOX9^high CSCs down-regulate
   intracellular ceramide levels via ABCA12 lipid transporter-mediated
   efflux. SOX9^high CSCs overexpressed ABCA12 and had low levels of
   ceramide. Abca12 knockout increased intracellular ceramide levels. This
   is consistent with previous studies showing that ABCA12 is required for
   extracellular ceramide secretion by keratinocytes to form a proper skin
   barrier ([184]43, [185]45). ABCA12 is one of the most significantly
   up-regulated ATP-binding cassette transporter genes in cancer compared
   to normal tissues ([186]38). High ABCA12 levels also correlate with
   reduced response to neoadjuvant chemotherapy in breast cancer patients
   ([187]39). Chemotherapy is known to induce ceramide generation, which
   sensitizes cells to cytotoxic agents ([188]20, [189]64, [190]65).
   Therefore, high ABCA12 expression could protect cells from chemotherapy
   by reducing the level of chemotherapy-induced ceramide. This is also
   likely to endow survival advantage of SOX9^high CSCs against
   chemotherapy. Supporting this notion, we showed that loss of ABCA12
   abolished the enrichment of SOX9^high CSCs by chemotherapy in vivo.
   Furthermore, increasing ceramide levels using acid ceramidase inhibitor
   D-NMAPPD sensitized tumors to chemotherapy and prevented enrichment of
   SOX9^high CSCs. These data suggest a potential strategy for targeting
   CSCs and overcoming chemoresistance.

MATERIALS AND METHODS

Mice

   NOD/SCID (#001303), NSG (#005557), and MMTV-PyMT (#002374) mice were
   purchased from the Jackson Laboratory (Bar Harbor, ME) and bred
   in-house. Sox9-GFP;C3/TAg mice were generated as described previously
   ([191]26). Sox9-GFP transgenic mice [Tg(Sox9-EGFP)EB209Gsat/Mmucd] were
   obtained from Mutant Mouse Resource & Research Centers and backcrossed
   to the FVB/n background. CAG::mRFP1 (#005884) mice were obtained from
   the Jackson Laboratory and backcrossed to the FVB/n background. Mice
   were housed in microisolator cages and a pathogen-free condition. All
   animal experiments were performed according to protocols (20170607 and
   00001266) approved by the Institutional Animal Care and Use Committee
   of Albert Einstein College of Medicine.

Tumor cell isolation

   Tumor cells were isolated and cultured as described previously
   ([192]66). Tumor samples were dissected, minced, and then digested with
   collagenase type 3 (300 U/ml; Worthington Biochemical, [193]LS004182),
   hyaluronidase (100 U/ml; Worthington Biochemical, [194]LS002592), 5 μM
   Y-27632 (Cayman Chemical Company, 10005583-10), and deoxyribonuclease I
   (DNase I) (10 μg/ml; Roche, 104159) in the Dulbecco’s modified Eagle’s
   medium (DMEM)/F-12 medium at 37°C for 2 hours on an orbital shaker.
   Afterward, the cell pellet was washed with 1× phosphate-buffered saline
   (PBS). Red blood cells (RBCs) were then lysed by RBC lysis buffer
   (eBioscience, 00-4300-54). Cells were further digested with 0.05%
   trypsin-EDTA for 5 min and neutral protease (1 U/ml; dispase)
   (Worthington Biochemical, LS02109) plus DNase I (100 μg/ml; Roche,
   104159) for 5 min. The dissociated cells were then filtered through
   40-μm cell strainers to obtain single cells.

Tumor organoid culture

   Tumor organoids were generated and cultured as described previously
   ([195]66). For measuring tumor organoid-forming efficiency, 200 to 500
   sorted primary tumor cells per well or 600 to 1000 sorted lung
   metastasis tumor cells per well were seeded in ultralow attachment
   96-well plates (Corning Life Sciences), with organoid culture medium
   containing Epicult (R)-B medium (STEMCELL Technologies) or Advanced
   DMEM/F-12 (Life Technology) supplemented with 1% l-glutamine, 5%
   heat-inactivated fetal bovine serum (FBS) (Sigma-Aldrich, F2442), 5%
   Matrigel (Corning Life Sciences, 354234), epidermal growth factor (10
   ng/ml; Sigma-Aldrich, E9644), basic fibroblast growth factor (20 ng/ml;
   EMD Millipore, GF003), heparin (4 μg/ml; Sigma-Aldrich, H4784), and 5
   μM Y-27632 (Cayman Chemical Company, 10005583-10). After 7-day culture,
   organoids greater than 100 μm in diameter were counted. For passaging
   organoids, 300,000 cells in 2 ml of organoid medium per well were
   seeded in six-well plates.

Cell lines and adherent cell culture

   MCF7, T47D, MDA-MB-231, and MDA-MB-468 cell lines were cultured in DMEM
   (Corning) with 10% FBS (VWR) and 1% (v/v) antibiotic-antimycotic
   solution (GE Healthcare). HCC1937 were maintained in RPMI 1640
   (Corning) with 10% FBS (VWR) and 1% (v/v) antibiotic-antimycotic
   solution (GE Healthcare). Cell lines were authenticated by short tandem
   repeat analysis, matching to the American Type Culture Collection
   database.

Orthotropic tumor cell implantation

   Female NOD/SCID (NOD.Cg-Prkdc^scid/J) mice at 6 to 8 weeks of age were
   injected in the inguinal mammary fat pad with 3 × 10^5 or 5 × 10^5
   Sox9-GFP;C3/TAg tumor cells or 1 × 10^6 PyMT;Sox9-GFP;RFP tumor cells
   suspended in 50 μl of 25% Matrigel diluted in PBS. Tumor onset was
   monitored by weekly palpation. Tumor volume was measured by caliper and
   calculated as V = (4/3) × π × (L/2) × (D/2) × (D/2). For in vivo
   limiting dilution assay of SOX9^neg, SOX9^low, or SOX9^high cells, 100,
   1000, and 10,000 sorted tumor cells were injected into the third and
   fourth mammary fat pad of NOD/SCID mice. Mice were monitored twice a
   week for a total of 4 months for tumor onset. For in vivo limiting
   dilution assay of nontargeting sgRNA (sgNT) and sgAbca12 cells, 500,
   5000, and 50,000 cells were injected into the third and fourth mammary
   fat pads of NOD/SCID mice. Mice were monitored twice a week for a total
   of 2 to 2.5 months for tumor onset.

PDX model

   Human PDX models were a gift from A. Patsialou and J. Condeelis
   ([196]33). Each tumor fragment (1 mm in diameter) was implanted in the
   fourth mammary pad of NSG recipient mice. The resulting tumors can be
   palpated 4 to 6 weeks following the transplantation. The PDX
   experiments were approved by the Institutional Review Board (IRB) of
   Albert Einstein College of Medicine (IRB# 2018-9307). Informed consent
   is not relevant because the PDX models were previously established, and
   we did not gather the tissue ourselves.

In vivo chemotherapy treatment

   When tumors reached 3 to 5 mm in diameter, mice received weekly
   intraperitoneal injections with cisplatin (6 mg/kg), paclitaxel (15
   mg/kg), doxorubicin (5 mg/kg), or vehicle [PBS or ethanol/cremophor
   EL/PBS (1:1:8)] for 3 weeks. Tumor volume was measured by caliper and
   calculated as V = (4/3) × π × (L/2) × (D/2) × (D/2) every week or twice
   a week.

In vitro drug treatment

   A total of 300,000 cells were seeded in poly-(2-hydroxyethyl
   methacrylate) (poly-HEMA)–coated ultralow attachment six-well plate, or
   500 cells were seeded in ultralow attachment 96-well plate. The test
   compounds—including doxorubicin (1 μM), cisplatin (0.5, 1, or 2 μM),
   paclitaxel (5 nM), D-NMAPPD (1, 10, or 20 μM), C6-ceramide (1 or 10
   μM), verteporfin (1, 3, or 10 μM), myriocin (2 μM), FB1 (10 μM), or
   vehicle—were applied.

Experimental metastasis assay

   A total of 80,000 to 200,000 sorted SOX9^low or SOX9^high cells from
   Sox9-GFP;C3/TAg tumors or 10,000 to 20,000 sorted SOX9^neg, SOX9^low,
   or SOX9^high cells from Sox9-GFP;PyMT;CAG::mRFP1 tumors were injected
   via tail vein into 4- to 6-week-old female NOD/SCID mice. Fourteen to
   16 weeks later, mice were perfused with 5 ml of PBS and lungs were
   harvested and imaged under the fluorescence microscope ex vivo.

Transwell cell migration and invasion assay

   For invasion assay, 30,000 cells were plated in Matrigel-coated (40 μl,
   2 mg/ml; Corning Life Sciences, 354234) upper chambers (pore diameter
   of 8 μm; Greiner Bio-One, 662638) with serum-free Epicult medium. For
   migration assay, 20,000 cells were plated in uncoated upper chambers
   with serum-free Epicult medium. In both assays, the lower chamber was
   added with Epicult medium plus 10% FBS. After incubation for 24 hours,
   cells on the upper side of the membrane were removed by wiping with a
   cotton swab. The membranes were stained with
   4′,6-diamidino-2-phenylindole (DAPI). Images were captured using a
   confocal microscope (×20 magnification).

3D collagen gel migration assay

   3D collagen gel migration assay was done on the basis of the protocol
   described previously ([197]67). Briefly, acid-solubilized rat-tail
   collagen I gels [collagen I (3 mg/ml), pH 7 to 7.5) were prepared.
   Eight-well chamber slides were coated with 200 μl of collagen I gel (3
   mg/ml) per well for 1 hour at 37°C. Sorted Sox9-GFP^low and GFP^high
   cells were cultured in the Epicult/Matrigel organoid culture for 1 week
   to form organoids and then transferred into collagen-coated chamber
   slides for 3 days. The images were taken, and the numbers of organoids
   with or without branching morphology were counted.

Western blot

   Cells were lysed with the radioimmunoprecipitation assay buffer
   containing 25 mM tris-HCL (pH 7.5), 150 mM NaCl, 1% sodium
   deoxycholate, 0.1% SDS, 1 mM EDTA, and 1.0% NP-40 on ice. Samples were
   resolved by SDS–polyacrylamide gel electrophoresis and transferred to
   polyvinylidene difluoride membranes. The Western blots were performed
   with antibodies against SOX9 (MilliporeSigma, AB5535), YAP (Cell
   Signaling Technology, 14074S), pYAP (S397) (Cell Signaling Technology,
   13619S), TAZ (Cell Signaling Technology, 72804S), β-actin (BD
   Biosciences, 612656), or histone H3 (Cell Signaling Technology,
   14269S). Original scans of the Western blot results are provided in the
   Supplementary Materials.

Flow cytometry and FACS sorting

   Multiparametric flow cytometric analyses were performed on an LSR II
   equipped with FACS Diva 6.1 software (BD Biosciences) and analyzed with
   FlowJo software v9.3 or 10.0.8 (FlowJo, LLC). Dead cells were excluded
   by forward scatter, side scatter, and DAPI (Sigma-Aldrich) staining.
   Lineage markers were used to exclude immune cells and endothelial cells
   in tumor cell analyses. Cell sorting was performed using MoFlow Astrios
   Cell Sorter (Beckman Coulter). The following antibodies were used in
   the current study: anti–epithelial cell adhesion molecule
   (EpCAM)–allophycocyanin (clone G8.8, BioLegend, 118214),
   anti–EpCAM–peridinin-chlorophyll-protein/Cy5.5 (clone G8.8, BioLegend,
   118220), biotinylated antibodies against lineage markers (CD45 (clone
   30-F11, BioLegend, 103104), CD31 (clone 390, BioLegend, 102404) and
   TER-119 (clone TER-119, BioLegend, 116204), anti-ABCA12 (Novus
   Biologicals, NB100-93466), and anti-ceramide (clone MID 15B4,
   Sigma-Aldrich, C8104) antibody. Biotinylated antibodies were further
   labeled with Strepavidin-V450 (BD Biosciences, 560797). Unless
   otherwise specified, all antibodies were used at a 1:100 dilution.
   Anti-goat or anti-mouse secondary antibodies conjugated with Alexa
   Fluor 633 or 549 and anti–immunoglobulin M (IgM) secondary antibody
   conjugated with Alexa Fluor 549 or DyLight 755 were used.

Immunofluorescence and immunohistochemistry

   Formalin-fixed and paraffin-embedded tissue sections or Optimal Cutting
   Temperature compound (OCT)–embedded frozen sections were stained with
   standard protocols. The following antibodies were used in this study:
   anti-YAP (1:200; Cell Signaling Technology, 14074S), anti-SOX9 (1:200;
   MilliporeSigma, AB5535), anti-Keratin 8 (1:60; Developmental Studies
   Hybridoma Bank, TROMA-I), anti–Keratin 14 (1:1000; BioLegend, 905304),
   anti–Keratin 14 (1:800; BioLegend, 906004), and anti-ABCA12 (1:100;
   Sigma-Aldrich, HPA043194). The ImmPRESS horseradish peroxidase (HRP)
   goat anti-rabbit IgG polymer reagent (Vector Laboratories, MP-7451) and
   the DAB/HRP Detection IHC Kit (Vector Laboratories, SK-4100) were used
   for immunohistochemistry. Fluorescent slides were imaged with an AXIO
   Examiner D1 microscope (Zeiss) with a confocal scanner unit (Yakagawa).
   SlideBook software 6.0 was used for the acquisition of images.
   Immunohistochemistry slides were scanned with a P250 High-Capacity
   Slide Scanner (3DHISTECH).

Quantitative RT-PCR

   Total RNA was isolated from cells using the Nucleo Spin RNA Kit
   (Macherey-Nagel AG), and cDNA synthesis was performed by the
   High-Capacity cDNA Reverse Transcription Kit (Applied Biosystems).
   Real-time PCR was performed with SYBR Green PCR Master Mix (Applied
   Biosystems) with Rpl13a as internal controls in an Applied Biosystems
   7900HT real-time PCR machine. Primer sequences are listed in table S1.

RNA-seq gene expression analyses

   RNA-seq libraries were prepared using the NEBNext Ultra II Directional
   Library Prep Kit (New England Biolabs, E7765) and sequenced by GENEWIZ.
   RNA-seq reads were aligned using Hisat2 ([198]68) to the reference
   mouse genome (GRCm38) using the M14 Gencode Basic CHR Annotations.
   Stringtie and the prepDE.py script were used to generate raw counts
   from aligned reads ([199]69). Differential expression was determined
   using DESeq2 ([200]70) (from Galaxy version 2.11.40.7+galaxy2). All
   data were accessible at Gene Expression Omnibus (GEO; [201]GSE220178).

Lentivirus production and lentiviral CRISPR-mediated gene knockout

   Lentivirus was prepared as described previously ([202]26). The mouse
   Sox9 cDNA was cloned in the pLVX-puro (Takara, 632164) lentiviral
   vector. The sgRNAs targeting ABCA12 were cloned into the lentiCRISPRv2
   vector (Addgene, #52961, a gift from F. Zhang). The sgRNA targeting
   mouse sequences are as follows: sgABCA12 #1-AAATGTAAAGACACGCCCTA,
   sgABCA12 #5-CAATTGCTTACACGCATCCT, and sgNT-GCGAGGTATTCGGCTCCGCG. The
   sgRNA targeting human ABCA12 is as follows:
   sgABCA12-ATAGAAGGCTGTCTGTTCGG. Cells were transduced with the
   lentiviral vectors and selected by puromycin.

CRISPR targeting efficiency measurement by TIDE assay

   A TIDE assay was used to assess editing efficiency by CRISPR-Cas9 as in
   the previous study ([203]71). Genomic DNA was extracted, and 100 ng of
   genomic DNA was used in a PCR amplification using SapphireAMP Fast PCR
   Master Mix (Clontech, catalog no. RR350B) and appropriate primers
   surrounding the sgRNA target region. The primer sequences for human
   ABCA12 are as follows: forward primer: -AAAGACACAGTCTCAGGAAAAGA,
   reverse primer-CTGTTGCTGAAGAAGATCATGGC.

Lentiviral CRISPR-mediated gene activation

   Tumor organoid cells were transduced with the lentiviral vector
   lenti-ER1a-dCas9-VPR-Puro (Addgene, #99373, a gift from K. Brennand),
   and puromycin was selected for 3 days. Cells were then transduced with
   sgNT or CRISPRa.Abca12 cloned into pLKO5.sgRNA.EFS.tRFP and sorted for
   RFP-positive cells. Guide RNAs were designed using the CRISPR-ERA
   ([204]http://crispr-era.stanford.edu) web tools. The sgRNA targeting
   sequences are as follows: sgAbca12 #a-GGGACCCTGGAAGGGCCAGT, sgAbca12
   #b-GTGGCAGGCGCTGGGTACTG, and sgNT-GCGAGGTATTCGGCTCCGCG.

YAP overexpression

   YAP1 (S6A)-V5 in pLX304 (Addgene, #42562, a gift from W. Hahn) and
   control pLX304 (Addgene, #25890, a gift from D. Root) plasmids were
   used. Lentivirus was prepared as described previously ([205]26). For
   the selection of virally infected cells, blasticidin (5 μg/ml) was used
   24 hours after infection.

Lipidomics analysis

   MS-based lipidomics was performed by Biological Mass Spectrometry Core
   Facility at University of Colorado School of Medicine as described
   previously ([206]72). Freshly sorted tumor cells (100,000) were
   centrifuged, and the pellet was flash-frozen in liquid nitrogen. Cells
   were extracted at 4°C in the presence of cold methanol at a
   concentration of 2 × 10^6 cells/ml. The samples were vortexed for 10
   min, placed at −20°C for 20 min, and then centrifuged at 18,213g for 10
   min at 4°C. The resulting supernatant was analyzed on a Thermo Vanquish
   Ultra High Performance Liquid Chromatography (UHPLC) coupled to a
   Thermo Q Exactive MS. The UHPLC was equipped with a Waters ACQUITY HSS
   T3 column and instrument methods were 15-min positive ion mode and
   17-min negative ion mode (both ESI) with top 15 ddMS^2 data
   acquisition. Features were annotated and integrated using LipidSearch
   (Thermo Fisher Scientific). Lipid metabolic pathway enrichment analysis
   was performed using the LIPEA software.

Quantification and statistical analysis

   Data are represented as means ± SEM. Statistical tests were performed
   with GraphPad Prism V8.0 (San Diego, CA). A two-tailed Student’s t test
   was used for comparisons of continuous variables between two groups.
   One-way analysis of variance (ANOVA) with Tukey’s or Dunnett’s test was
   used when three or more groups were compared. Two-way ANOVA was used to
   find the comparison between two independent variables. Kaplan-Meier
   survival curves were analyzed by using GraphPad Prism V8.0, V9.4.1, and
   V10.0.0 (San Diego, CA), and P values were calculated using the
   log-rank test.

Acknowledgments