Identifying the Transcriptional Drivers of Metastasis Embedded within Localized Melanoma

Malignant melanomas arise from pigment cells called melanocytes or their neural crest precursors (1–3). These cells can metastasize to distant organs and drive mortality in patients (4, 5). Although significant progress has been made in melanoma biology, effective therapies targeting metastasis remain suboptimal (5). Although driver mutations in genes such as BRAF and NRAS initiate tumors (6–8), the large number of passenger alterations due to ultraviolet radiation-induced mutagenesis confounds mutational profiles in metastasis (6, 9). The tumor microenvironment with cell types such as immune cells, keratinocytes, fibroblasts, and adipocytes adds further complexity to our understanding of metastasis (10). Recent work shows that metastasis is governed less by genetic mutations and more by epigenetic, transcriptional, and/or structural alterations (7, 8, 11, 12, 13).

Primary melanoma is a common disease, but the majority of these will not go on to metastasize. Thus, a major challenge in the field is predicting which tumors will eventually metastasize or not, as those higher-risk tumors need more careful follow-up and possible adjuvant therapy (14–17). A variety of clinical factors (i.e., Breslow thickness, ulceration, extent of tumor-involved lymph nodes) can provide such prognostic information (18, 19). Advances in imaging of these lesions, including machine learning/artificial intelligence–driven approaches, add further predictive power (20–22). However, as it is increasingly recognized that metastasis is at least partially governed by epigenetic and transcriptional changes (7, 8, 11, 12, 13), there have been large-scale efforts into developing gene signatures of localized tumors that predict which patients will metastasize (23–29).

Despite the potential predictive power of these signatures, little is known about the mechanism by which the genes within these signatures promote metastasis. In addition, it is often not clear which transcriptional changes are drivers versus passengers in the metastatic process. In part, this is because there are few genetically tractable models of spontaneous metastasis in vivo (30–32). Transplantation-based models such as mouse xenografts (including cell lines or patient-derived xenografts; ref. 33) are powerful, but these represent already transformed cells, rather than primary tumors. Genetically engineered mice such as the BRAF^V600E; Pten^−/− and the BRAF^V600E; myrAkt1; Cdkn2a^−/− model are very useful as they recapitulate all stages of tumor development (30–32, 34). The development of ES-cell–based chimeric mice (35) has further accelerated modeling of these alterations in vivo, but these can still be challenging to scale up to large numbers. For these reasons, we have developed a highly scalable zebrafish model that can be used to rapidly screen for metastasis modifiers and dissect the involved mechanisms. Building off previous work showing that transgenes can be directly electroporated into the skin of adult zebrafish (Transgene Electroporation in Adult Zebrafish, or TEAZ; ref. 36), we developed a new BRAF^V600E; tp53^−/−; pten^−/− melanoma model that has a 25% spontaneous metastasis rate, allowing us to then readily test novel genes that modify this rate. We analyzed transcriptomics data of human melanomas to identify candidate genes whose expression from the primary or locally relapsed tumor predicted distant metastasis. Using TEAZ (36), we performed an in vivo metastasis screen to model these alterations and identified GRAM domain containing 1B (GRAMD1B), a cholesterol transfer protein as a novel metastasis suppressor in human melanoma. Downregulation of GRAMD1B is a frequent event in the tumors of melanoma patients, is associated with higher metastatic incidence, and portends poor overall survival. Our data show that gramd1b loss accelerates tumor growth and metastasis in vivo. We show that GRAMD1B depletion results in dysregulated cholesterol transport in melanoma cells leading to intracellular sterol accumulation and suppression of de novo cholesterol synthesis. This accumulated cholesterol activates an invasive transcriptional program via AP-1 transcription factor to drive metastasis. Thus, our study demonstrates the utility of our metastasis platform as a scalable system to interrogate clinical alterations and implicates GRAMD1B as an important gene in melanoma metastasis.

To identify novel transcriptional changes that regulate metastatic dissemination, we obtained data from the AVAST-M clinical trial, the largest adjuvant study in high-risk primary melanoma (17, 23, 37). For this study, 337 patients with primary or locally advanced disease underwent bulk RNA sequencing (RNA-seq) of localized tumors along with extensive clinical follow-up to determine which patients ultimately developed distant metastasis. Differential expression (DE) analysis was performed on the samples from patients who did or did not develop metastasis to identify a transcriptional signature that predicts metastasis (Fig. 1A; Supplementary Table S1). This revealed a total of 61 transcripts that were predictive of distant metastatic relapse (FDR-corrected P < 0.1). This included 41 transcripts that predicted “pan-organ” metastasis and 20 transcripts that specifically predicted metastasis to the brain (Supplementary Fig. S1; Supplementary Table S1). PANTHER pathway classification revealed that genes in this signature were enriched in pathways related to apoptosis signaling, glutamate receptor pathway, and the p53 pathway, among others (Fig. 1B). Furthermore, DAVID Gene Ontology analysis (38, 39) indicated a suppression in immune cell networks (downregulation of genes such as PRF1, TRIM22, and KYNU) and apoptotic signaling (downregulation of genes such as FASLG and TNFRSF10A) in patients predestined for metastasis (Supplementary Table S1 and S2). To further stratify the role of the 61 predicted metastasis suppressor genes in melanoma, we also analyzed their DNA alterations and RNA expression profiles in an independent human melanoma dataset [ref. 40; The Cancer Genome Atlas (TCGA) Firehose Legacy cohort; Fig. 1C]. Based on this analysis along with a review of the literature in melanoma, we prioritized genes that were either downregulated, deleted, or had missense mutations in this independent cohort because these types of alterations have been linked to metastasis regulation (7, 8, 11–13). This led to a prioritized list of nine candidate metastasis suppressor genes (Fig. 1C) for screening in vivo.

Transcriptional changes associated with metastasis can occur either during primary tumor evolution or after the tumors have already disseminated. In predictive scoring systems (23–28), these changes occur within early-stage tumors, necessitating genetic model systems in which these genes can be tested during de novo tumor initiation and evolution. To achieve this, we utilized TEAZ (36) and CRISPR–Cas9 gene editing to build a baseline metastatic melanoma model driven by expression of oncogenic BRAF^V600E and loss of tumor suppressors tp53 and pten, common events occurring in human melanoma (refs. 31, 41; Fig. 2A). Previous work has shown that inactivation of TP53 is relevant in human melanoma because it often occurs via loss of CDKN2A, although wild-type p53 can also be required for the slow-cycling state, suggesting its role can be variable depending on context (42). In our model, germline mitfa: BRAF^V600Etp53^−/− casper animals (41, 43) were electroporated with a MAZERATI 2X-sgpten (MAS2X-sgpten) plasmid (ref. 44; Supplementary Fig. S2A) to knockout (KO) pten in a melanocyte-specific manner. Introduction of this plasmid results in mitfa minigene expression, melanocyte rescue, expression of melanocyte-specific Cas9 and 2 single-guide RNAs (sgRNA) against pten (44). Loss of pten resulted in aggressive melanomas (70% tumor incidence) in fully immunocompetent and optically transparent animals (Fig. 2B and C). Efficient depletion of Pten in tumors was confirmed by CRISPR sequencing (Supplementary Fig. S2B) and IHC [Fig. 2D; for phospho-AKT (Serine 473); Fig. 2E]. Tumors in this model histologically resembled human melanoma, invaded into the muscle, adipose tissue and spontaneously metastasized to distal sites in 25% of animals (Fig. 2E and F). Because we aimed to perform an in vivo loss-of-function screen for the AVAST-M genes, we first assessed if we could model additional CRISPR-induced loss-of-function events in this system. As a proof of principle, we knocked out slc45a2 (alb) gene in a melanocyte-specific manner (Supplementary Fig. S2C), which leads to impaired melanin synthesis in cells but should otherwise be neutral. Consistent with this, loss of alb resulted in tumors devoid of melanin (Fig. 2G and H). Moreover, CRISPR sequencing of the KO tumors revealed efficient genome editing at the alb locus (Supplementary Fig. S2D). Thus, this serves as proof of principle that our system can reliably model loss-of-function alterations in vivo. Finally, we further improved upon the efficiency of the system by electroporation of multiple smaller plasmids (Supplementary Fig. S2E) rather than the larger MAZERATI plasmid. This multiplasmid approach led to an ∼8-fold increase in melanocytes transformed (Supplementary Fig. S2F) and markedly accelerated tumor initiation (Supplementary Fig. S2G and S2H). In this model, animals developed melanomas within ∼6 weeks that metastasized within ∼10 weeks, making this platform ideal for screening.

Zebrafish orthologs of the AVAST-M candidate genes were retrieved from the Zebrafish Information Network (ZFIN; ref. 45), and 4 sgRNAs per gene were cloned into a MITF GFP 2X sgRNA plasmid. A TEAZ-based screen of these genes was performed using the above metastasis platform (Fig. 3A). Fluorescence imaging was used to quantify primary tumor area and metastatic burden over time (Fig. 3B and C). Among the 9 candidates, loss of gramd1b (GRAM Domain Containing 1B) led to significantly increased tdTomato⁺ primary tumor area at 4 and 8 weeks after TEAZ (Fig. 3B). Subsequent validation experiments using larger cohorts of animals confirmed that gramd1b functioned as a tumor suppressor in vivo (Fig. 3D). Loss of gramd1b was confirmed by surveyor assays (Supplementary Fig. S3A) and real-time PCR (Supplementary Fig. S3B). Because GRAMD1B was identified as a putative metastasis suppressor from human melanoma samples, we assessed if gramd1b KO tumors had altered metastatic potential. Fluorescent imaging of animals at 12 weeks after TEAZ (Fig. 3E) revealed that animals with gramd1b KO tumors had increased visceral metastasis (to the kidney marrow), compared with controls (Fig. 3F). Furthermore, IHC for tdTomato confirmed extensive infiltration of metastatic tumor cells into the kidney marrow of animals with KO tumors (Fig. 3G). Animals with KO tumors had significantly increased metastatic incidence as well as burden relative to control tumors (Fig. 3H and I). Interestingly, loss of two other genes, kif1a and gabrg2, significantly increased metastatic burden in animals without affecting primary tumor growth (Supplementary Fig S4A and S4B), which will be investigated in the future. Collectively, these data demonstrate that our metastasis platform can be used to screen clinical alterations and uncover novel metastatic modulators in human melanoma.

Because gramd1b loss in zebrafish had such a pronounced effect on metastasis of zebrafish melanoma cells, we sought to investigate its role in human melanoma. Analysis of the AVAST-M melanoma cohort samples showed that patients with low GRAMD1B expression had poor overall and progression-free survival in the AVAST-M melanoma cohort (Fig. 4A; Supplementary Fig. S5A). GRAMD1B expression was significantly associated with survival after correcting for clinical correlates such as sex, age, stage, and treatment in the AVAST-M melanoma cohort and an independent AVAST-M lymph node cohort (Fig. 4B; Supplementary Fig. S5B). To further validate its role in human melanoma, we analyzed three independent patient cohorts. In the skin cutaneous melanoma cohort (TCGA, n = 287) and the UCLA Metastatic Melanoma cohort (n = 38), deep deletions and missense mutations in GRAMD1B are frequent events (ref. 40; Supplementary Fig. S5C and S5D). In the Leeds Melanoma Cohort (n = 666; ref. 24), which includes stage I–III tumors, low GRAMD1B expression predicted poor survival in thicker tumors (Breslow > 3 mm; Fig. 4C) as well in ≧ stage IIB tumors regardless of Breslow thickness. These clinical data encouraged us to further focus on its functional role. GRAMD1B belongs to the GRAM domain 1 family of proteins (consisting of GRAMD1A, GRAMD1B, and GRAMD1C) and encodes a cholesterol transfer protein that mediates nonvesicular transport of cholesterol within cells (46, 47). Although GRAMD1B has been shown to transport cholesterol in mouse adrenal tissue and human embryonic kidney cells (46, 48), the biological significance of this protein in melanoma cells and cancer in general has not been well explored. To better understand its role in melanoma, we performed RNA-seq in human SKMEL2 cells to identify differentially expressed genes in GRAMD1B KO cells (Fig. 4D). Introduction of GRAMD1B sgRNA led to potent target depletion (Fig. 4D), whereas transcript levels of GRAMD1A and GRAMD1C remained unaltered (Supplementary Fig. S5E). Gene set enrichment analysis (GSEA) identified epithelial-to-mesenchymal transition (EMT) as one of the top enriched pathways in the KO cells (Fig. 4E). Specifically, there was a significant upregulation in EMT genes such as SERPINE2, CCN2, COL12A1, and MMP1 in the KO cells, previously reported to promote metastasis (refs. 49–54; Fig. 4F). This indicated that GRAMD1B loss may promote metastasis not only in zebrafish melanoma cells but also in human melanoma. To further functionally test this, we knocked out GRAMD1B in a panel of human melanoma cell lines (SKMEL2, A375, and A2058) using lentiviral expression of Cas9 and sgRNAs against GRAMD1B. Although we saw no defect in proliferation (Supplementary Fig. S6A–S6C), we found a significant increase in invasion in a Matrigel-coated transwell invasion assay in all 3 cell lines (Fig. 4G; Supplementary Fig. S6D and S6E). Similar results were seen upon shRNA-mediated knockdown of GRAMD1B in SKMEL2 cells (Supplementary Fig. S6F). Although GRAMD1B loss promoted invasiveness through Matrigel, the migratory ability of these cells in the absence of Matrigel was not altered (Supplementary Fig. S6G), suggesting the metastatic phenotypes may be specific to its function in ECM degradation and extravasation. Consistent with this, we performed qRT-PCR and observed an upregulation of multiple MMP genes, which are known to cleave ECM proteins, upon GRAMD1B KO in these 3 different human melanoma cell lines (Supplementary Fig. S6H–S6J). Finally, to test if GRAMD1B depletion alters the metastatic ability of human melanoma cells in vivo, we generated doxycycline (dox)-inducible NT or GRAMD1B shRNA SKMEL2 cells expressing TurboRFP and performed a tail-vein metastasis assay in NSG mice (Fig. 4H). We observed that GRAMD1B depletion potently increased metastasis to the liver (RFP⁺ tumor cells in the livers of 100% of animals) and lung (RFP⁺ cells in the livers of 90% of animals) compared with an NT control (no RFP⁺ tumor cells detected) within 3 weeks (Fig. 4I). Histologic examination via hematoxylin and eosin (H&E) staining confirmed the presence of metastatic nodules in the liver and lungs (Fig. 4J). Color deconvolution was performed to separate hematoxylin stains in mouse livers to quantify tumor-positive liver area. This analysis revealed that GRAMD1B depletion led to increased metastatic burden in the liver (Fig. 4K). Taken together, these data demonstrate that GRAMD1B suppresses the invasiveness of human melanoma cells in vitro and metastasis in vivo.

GRAMD1B represents a newly reported cholesterol transporter that mediates the nonvesicular transfer of accessible cholesterol from the plasma membrane (PM) to the endoplasmic reticulum (ER; refs. 46, 55, 56). GRAMD1B loss in mouse adrenal tissues did not affect total cholesterol levels, but led to decreased ER cholesterol, increased nuclear translocation of sterol regulatory element-binding transcription factor 2 (SREBP2), the master transcription factor regulating de novo cholesterol synthesis, and increased transcription of de novo synthesis genes (46). In contrast, another recent study reports that GRAMD1B is also required for lysosome to ER transfer of low density lipoprotein (LDL) cholesterol and that GRAMD1B knockdown leads to increased free cholesterol within cells (57). Thus, its precise role in mammalian cancer cells remains under investigation.

We analyzed our RNA-seq data and noted significant downregulation in cholesterol synthesis pathways in GRAMD1B KO cells (Fig. 5A). Mammalian cells maintain cholesterol homeostasis by tightly regulating its transport across organelles. Excess ER cholesterol binds SCAP and prevents the ER to Golgi transport of the SCAP–SREBP2 complex (58–60). This prevents proteolytic processing of active SREBP2 and results in suppression of transcription of de novo synthesis. Consistent with this, real-time PCR quantification confirmed the downregulation of SREBF2 as well as key de novo cholesterol genes such as HMGCR and HMGCS1 upon GRAMD1B loss in both SKMEL2 cells (Fig. 5B) and A375 cells (Supplementary Fig. S7A). Similar results were obtained using both sgRNA and shRNA approaches (Supplementary Fig. S7B). Furthermore, there was a decrease in nuclear SREBP2 protein levels upon GRAMD1B loss in these cells (Fig. 5C). Our data are, therefore, suggestive of an increase in cholesterol in the KO cells, similar to what is seen by directly supplementing excess cholesterol to these cells (Supplementary Fig. S7C and S7D).

These findings led us to hypothesize that GRAMD1B loss leads to free cholesterol accumulation in melanoma cells. To test this, we used dox-inducible NT or GRAMD1B shRNA in SKMEL2 cells (Supplementary Fig. S7E) and measured cholesterol levels in these cells using two independent assays. We stained NT and GRAMD1B shRNA cells with Filipin, an endogenous cholesterol probe, and observed a striking increase in Filipin staining in GRAMD1B knockdown cells (Fig. 5D and E). We also performed cholesterol extractions in three independent human melanoma cell lines and observed an increase in free cholesterol upon GRAMD1B depletion (Fig. 5F–H). Next, to test if gramd1b loss leads to free cholesterol accumulation in vivo, we performed cholesterol extractions from transgenic zebrafish melanomas as well as the blood of animals bearing NT or gramd1b KO tumors (Fig. 5I). We observed that gramd1b loss led to an increase in free cholesterol in tumor tissues (Fig. 5J), similar to what we saw in vitro. We did not see an effect on serum cholesterol levels (Fig. 5K), indicating that GRAMD1B acts at least in part in a cell-autonomous manner to regulate intracellular cholesterol levels. As most of the cholesterol in cells are located in the PM, we assessed whether this increase in free cholesterol affected PM composition and integrity. GRAMD1B knockdown did not significantly alter the abundance of choline-containing phospholipids in the cells (Supplementary Fig. S7F). To assess the consequences of GRAMD1B knockdown on membrane fluidity, we measured the lateral diffusion of lipophilic pyrene probes in NT and GRAMD1B knockdown cells. These probes undergo excimer formation upon spatial interaction at the cell membrane and exhibit change in their spectral properties compared with their monomer counterpart. We did not observe a significant change in membrane fluidity upon GRAMD1B knockdown (Supplementary Fig. S7G). Taken together, these findings suggest that GRAMD1B loss leads to free cholesterol accumulation in melanoma cells in vitro and in vivo.

As GRAMD1B loss led to an increase in free intracellular cholesterol, we hypothesized that the invasiveness of GRAMD1B KO cells is dependent on cholesterol accumulation. To test this, we depleted cholesterol in NT and GRAMD1B shRNA cells by culturing cells in lipoprotein-free FBS and in the presence of Simvastatin (10 μmol/L; Fig. 6A) and then assessed their invasiveness. GRAMD1B knockdown led to an increase in free cholesterol (Fig. 6A) and invasion (Fig. 6B), but cholesterol depletion abolished this increased invasiveness (Fig. 6B). Next, we sought out to determine how free cholesterol promotes melanoma cell invasion. Excess free cholesterol in cells can be metabolized into oxysterols to aid steroid and bile synthesis or esterified and stored in lipid droplets (61). Interestingly, the oxysterol 27-hydroxycholesterol (27-HC) has been implicated as a breast cancer metastasis driver (62–64). Consistent with this, we found that exogenous treatment of A375 and SKMEL2 cells with cholesterol and 27-HC was sufficient to stimulate invasiveness (Fig. 6C). To test if 27-HC could be a driver of the metastatic behavior in a GRAMD1B KO setting, we first assessed the expression of CYP27A1, the mitochondrial resident enzyme that synthesizes 27-HC from cholesterol in cells. Our RNA-seq results showed that CYP27A1 is the only oxysterol enzyme expressed in SKMEL2 cells and its expression was further increased upon GRAMD1B loss (Supplementary Fig. S8A). Real-time PCR and western blot analysis confirmed a robust increase in CYP27A1 mRNA (Supplementary Fig. S8B) and protein levels (Fig. 6D) upon GRAMD1B loss in SKMEL2 and A2058 cells. To directly test the role of CYP27A1 in melanoma cell invasion, we knocked down CYP27A1 in three human melanoma cell lines (Supplementary Fig. S8C; Fig. 6E) and assessed tumor cell invasion. We observed that CYP27A1 knockdown by itself potently suppressed invasion (Fig. 6F). Next, to definitively test the contribution of 27-HC to the invasiveness of GRAMD1B KO cells, we knocked down CYP27A1 in NT or GRAMD1B KO SKMEL2 cells (Fig. 6G and H). We observed that CYP27A1 knockdown diminished the invasiveness of GRAMD1B KO cells (Fig. 6I), suggesting that 27-HC is a key mediator of the metastatic phenotype upon GRAMD1B loss.

Sterols can drive invasion through changes in metabolism and cell signaling (65, 66), membrane synthesis (67, 68), or transcriptional programs (63, 64, 69). Previous studies have shown that activation of melanoma invasiveness and metastasis is largely governed by transcriptional cell state changes (70–75), including those associated with AP-1 as a key transcription factor linked to invasive/migratory cell states in melanoma (76, 77). To further investigate this, we performed Hypergeometric Optimization of Motif EnRichment (HOMER) analysis on GRAMD1B KO versus NT cells, in order to infer which transcription factors were driving changes in gene expression, as shown previously (70, 78, 79). This revealed that members of the AP-1 family—FOSL1, FOSL2, and JUNB—were among the topmost enriched motifs in the KO cells (Fig. 7A). FOSL1 belongs to the FOS family of proteins consisting of c-FOS, FOSL1, FOSL2, and FOSB (80). These encode leucine zipper proteins that dimerize with JUN proteins to form a stable AP-1 transcription factor complex (80). This led us to assess whether free sterol accumulation in KO cells activates AP-1. To test this, we first asked if cholesterol or 27-HC was sufficient to activate AP-1 using an AP-1 luciferase reporter assay. Treatment of human melanoma A375 and SKMEL2 cells with cholesterol or 27-HC potently activated the AP-1 reporter (Fig. 7B and C) but failed to activate a control luciferase reporter (Supplementary Fig. S8D and S8E). Importantly, GRAMD1B depletion also led to potent activation of AP-1 reporter (Fig. 7D) without affecting the control construct (Supplementary Fig. S8F).

The AP-1 complex and its subunits are regulated in response to growth factors, cytokines, and cell stress in mammalian cells (80). For example, cellular stress can activate the p38 or JNK pathway to regulate JUN transcription and AP-1 activation. Growth factor signaling, on the other hand, can activate ERK signaling to promote FOS transcription as well as phosphorylate the AP-1 complex to its active form (80). To better understand how AP-1 is activated under GRAMD1B depletion, we treated NT and GRAMD1B shRNA cells with chemical inhibitors of the JNK and ERK pathways (Supplementary Fig. S8G) and performed the AP-1 luciferase assay. GRAMD1B knockdown enhanced AP-1 activation as expected (Fig. 7E). Although the JNKi suppressed AP-1 activity in the NT cells, it failed to rescue AP-1 activation in GRAMD1B knockdown cells. In contrast, the ERKi led to a decrease in AP-1 activation in the knockdown cells (Fig. 7E). These data suggest that ERK signaling may be driving AP-1 activation in GRAMD1B KO cells. Indeed, loss of GRAMD1B led to potent phosphor­ylation of ERK1/2 in three independent melanoma cell lines (Fig. 7F). Additionally, real-time PCR analysis revealed that JUN, FOSL1, and FOSL2 transcripts were upregulated upon GRAMD1B loss in SKMEL2 cells (Supplementary Fig. S8H), further implicating the ERK pathway in AP-1 activation. Exogenous cholesterol and 27-HC also induced ERK activation (Fig. 7G; Supplementary Fig. S8I), but which of these was more potent differed across cell lines. In A375 and A2058, 27-HC was sufficient to increase pERK, whereas in SKMEL2, cholesterol itself was more effective in inducing pERK. We speculate this cell line specificity may be due to differences in uptake, as we found that the addition of cholesterol complexed with methyl cyclodextrin failed to induce pERK in A375 cells (Fig. 7G), but LDL cholesterol potently activated pERK in these cells (Supplementary Fig. S8I). Finally, to test the contribution of activated ERK signaling to the enhanced invasiveness in the KO cells, we treated NT and GRAMD1B knockdown SKMEL2 cells with a low dose (100 nmol/L) of ERKi for 24 hours and performed a transwell invasion assay. Treatment of GRAMD1B knockdown cells with the ERKi suppressed invasion without affecting cell proliferation (Fig. 7H; Supplementary Fig. S8J). Taken together, our data demonstrate that free sterols in GRAMD1B KO cells activate ERK signaling to promote an AP-1–mediated invasion program and metastasis in melanoma cells (Fig. 7I).

The incidence of localized melanoma far exceeds that of metastatic disease. Discerning which lesions will go on to metastasize is of urgent clinical need, both to avoid overtreatment of lesions with a benign course and to provide adjuvant therapy to those with a more aggressive course. Although clinical, histologic, imaging, and transcriptional signatures have all been used to make such predictions (18–28, 81–84), these do not provide mechanistic insight into how those signatures act to promote metastasis. Mouse models of melanoma to study these mechanisms are powerful but can be limited in their scalability for genetic perturbation (30, 32, 33). For this reason, we developed a zebrafish TEAZ-based metastasis platform as a rapid and scalable system for studying these human gene signatures. Through our screen, we identified GRAMD1B, a cholesterol transfer protein, as a novel metastasis suppressor in melanoma. Our work sheds light on an understudied protein in cancer and identifies a new role for interorganelle cholesterol trafficking in melanoma metastasis.

The cholesterol carrier APOE has been previously linked to metastasis in melanoma (85–88). Studies into the use of statins prior to melanoma diagnosis have shown it may reduce risk of melanoma recurrence (89), although the mechanism is not clear. Additionally, cholesterol homeostasis is a metabolic dependency in other tumor types (63, 67, 90, 91). Cholesterol in mammalian cells is obtained either from receptor-mediated uptake or from de novo cholesterol synthesis in the ER (61). Sixty percent to 80% of cellular cholesterol is transported and stored in the PM, where it is critical to maintain membrane structural integrity (61). It is also transported to the ER, where it is esterified and stored as lipid droplets and suppresses SREBP2-mediated de novo cholesterol synthesis, thus maintaining homeostasis (58, 61, 92). However, how organelles transfer cholesterol among each other and how this functions in cancer is not completely understood.

GRAMD1B represents a recently characterized cholesterol transfer protein, with a GRAM domain that recognizes unesterified accessible cholesterol and a StART-like domain that transfers cholesterol through membrane contact sites (MCS; ref. 48). A recent study identified that GRAMD1B may also transport carotenoids within retinal cells (93), necessitating a need to better understand its role in mammalian cells. GRAMD1B was shown to reside in the ER but upon cholesterol loading (complexed with methyl cyclodextrin or HDL cholesterol) forms MCS with the PM to transfer accessible cholesterol pools from the PM to the ER (46, 48, 56). There are conflicting data as to the consequences of GRAMD1B loss. Previous work showed that GRAMD1B loss leads to a block in PM to ER cholesterol transport, reduced cholesterol esterification in the ER, and activation of the SREBP2 pathway, which may be in part dependent upon phosphatidylserine in the inner leaflet of the PM (46, 48, 55). However, Höglinger and colleagues showed that GRAMD1B depletion in HeLa cells led to an increase in free cholesterol (57). Our data also report an increase in free cholesterol within cells upon GRAMD1B loss, without any significant alteration to levels of choline-containing phospholipids or membrane fluidity. The mechanisms of free cholesterol accumulation and transport upon GRAMD1B loss in melanoma require further exploration. Höglinger and colleagues also observed increased lysosome-mitochondria MCS in GRAMD1B knockdown cells, indicating that under GRAMD1B depletion, free cholesterol may move into the mitochondria instead of the ER (57). Another study observed an increase in mitochondrial cholesterol upon loss of GRAMD1C in renal cancer cells (94). These findings along with ours raise the intriguing possibility that GRAMD1B loss in melanoma leads to a relative increase in mitochondrial cholesterol. Once in the mitochondria, cholesterol can be converted into 27-HC through the action of CYP27A1 (sterol 27-hydroxylase enzyme; refs. 95, 96) and promote metastasis.

There is growing evidence that 27-HC, the stoichiometric downstream metabolite of cholesterol, can promote tumorigenesis of breast cancer stem cells (97) and metastasis of breast cancer cells through tumor-intrinsic mechanisms as well as recruitment of immune cells (62–64, 98, 99). Consistent with these findings, we observe that both cholesterol and 27-HC stimulate the invasiveness of melanoma cells. Importantly, we show that 27-HC is a key mediator of the effects of GRAMD1B loss on tumor cell invasion. The accumulated free sterols in turn activate an AP-1 tumor invasion program through potent ERK signaling to promote metastasis. This finding is of significance because AP-1 is linked to migratory and invasive states in melanoma (76, 77). Interestingly, 27-HC has also been reported to function as an endogenous selective estrogen receptor modulator (SERM; refs. 69, 100). Ligand-activated estrogen receptors can be recruited to AP-1 sites to promote gene transcription (101, 102). Whether 27-HC functions as a SERM in melanoma and plays a role in activating an ER/AP-1 invasion program needs to be explored. Potent ERK signaling in GRAMD1B KO cells potentiates AP-1 activity, at least partly by transcriptionally upregulating the AP-1 members JUN, FOSL1, and FOSL2. How free sterols activate ERK in melanoma cells requires future exploration. Our findings show that exogenous cholesterol and 27-HC can activate ERK signaling in A375 and A2058 melanoma cells, but only cholesterol itself activated pERK in SKMEL2 cells. Whether this is due to differences in uptake mechanisms versus intracellular processing to 27-HC is an important area for future exploration. It is also possible that free sterols promote tumor cell invasion through additional non-ERK mechanisms (Fig. 7I). These findings further highlight a need to study the consequences of cholesterol delivery and transport within cells on subsequent prometastatic cell signaling.

A recent study showed that GRAMD1B knockdown promotes breast cancer migration via activation of JAK–STAT signaling (103). Consistent with this, our transcriptomics data also revealed activation of the JAK–STAT pathway in GRAMD1B KO cells, suggesting additional mechanisms may be at play in the KO cells (Fig. 4D). Indeed, inflammatory signaling such as JAK–STAT and NF-KB pathways have been implicated in driving metastasis (104–107). The activation of inflammatory transcriptional signatures observed in GRAMD1B KO cells (Fig. 4D) highlights a need to additionally explore the contribution of the immune compartment to the metastasis observed in vivo. Our data indicate that the AP-1 tumor invasion transcriptional program represents at least one mechanism by which GRAMD1B loss promotes metastasis.

Given the large caseload of localized melanoma, our work highlights the importance of identifying which lesions will go on to metastasize and how those high-risk lesions can be therapeutically targeted. Recent work using adjuvant BRAF/MEKi or PD-1/PD-L1 therapies has shown substantial reductions in subsequent metastatic relapse and mortality (14–16, 108). Our data suggest that dysregulated cholesterol trafficking inside cells may be an additional important target and that mechanisms to modulate the synthesis of 27-HC may be one way of preventing relapse of high-risk disease.

An sgRNA plasmid (MITF GFP 2X-sgRNA), which harbors a mitfa:GFP insert as well as a 2X sgRNA cassette (Fig. 2A), was generated for the zebrafish screen. The 2X sgRNA cassette from MAZERATI 2X (MAS2X) sgRNA (RRID Addgene_118844), also called MiniCoopR 2X sgRNA (gift from Dr. Len Zon), was amplified and introduced into mitfa:GFP plasmid using In-fusion cloning. Site-directed mutagenesis was performed to introduce NheI sites spanning sgRNA2 of the plasmid. MAS2X sgpten was cloned from MAS2X sgRNA (Supplementary Fig. S2A; ref. 44, 109). The U6-sgRNA plasmid (RRID Addgene_64245) was used along with mitfa:Cas9 plasmid and MiniCoopR:tdTomato (previously developed in the lab). LentiCRISPRV2 Puro (RRID Addgene_52961), psPAX2 (RRID Addgene_12260), and MD2 (RRID Addgene_12259) plasmids were obtained from Addgene. TRIPZ shRNA plasmids were purchased from Dharmacon expressing a nontargeting (NT) shRNA or shRNAs against human GRAMD1B. Clone IDs and hairpin sequences are in Supplementary Table S3. Control luciferase plasmid (pGL3, RRID Addgene_48743) and Renilla luciferase plasmid (pRL-SV40, RRID Addgene_27163) were obtained from Promega (E1751 and E2231) and pGL3-AP-1 was from Addgene (RRID Addgene_40342). Details of the oligonucleotide sequences used for cloning and sequencing are in Supplementary Table S3.

A375 (RRID:CVCL_A2CR), A2058 (RRID:CVCL_1059), and SKMEL2 (RRID:CVCL_0069) cells were obtained in August 2013 from the ATCC, which performs cell line authentication testing using morphology, karyotyping, and PCR-based methods including short-tandem repeat profiling to confirm the identity of human cell lines. The cells were also tested for Mycoplasma using a luminescence-based assay (Mycoalert Mycoplasma Detection Kit, Lonza) in 2017. SKMEL5 (RRID:CVCL_0527) cells were obtained in 2021 from ATCC. Cells were cultured in DMEM High Glucose media (Invitrogen) supplemented with 5% FBS (Sigma-Aldrich) and 1% antibiotics, 1% glutamine (Invitrogen). ZMEL-LD cells were established previously in the lab (110) and a low passage culture was utilized for gene editing experiments. For exogenous sterol treatment assays, melanoma cells were treated with 25 μmol/L cholesterol (Sigma) or 10–25 μmol/L 27-HC (Sigma) for 48 hours. For chemical inhibitor experiments, melanoma cells were treated with 100 nmol/L-200 nmol/L of SCH772984 (ERKi) and JNK-IN-8 (JNKi) for 24 hours (invasion assays) or 48 hours (luciferase assays). A list of all chemicals, kits, and their source is available in Supplementary Table S4.

Zebrafish experiments were carried out in accordance with institutional animal protocols. Animals were housed in a temperature- (28.5°C) and light-controlled (14 hours on, 10 hours off) facility. Fish were fed thrice a day using brine shrimp and pelleted zebrafish food. All anesthesia was done using tricaine (Western Chemical Incorporated) with a stock of 4 g/L (protected for light) and diluted until the fish was immobilized. All procedures were approved by and adhered to Institutional Animal Care and Use Committee (IACUC) protocol #12-05-008 through Memorial Sloan Kettering Cancer Center. Fish strains used in this study are in Supplementary Table S4.

Data on differentially expressed human genes were extracted from the phase III adjuvant AVAST-M study (17, 37). The study (including the collection of DNA and RNA) was ethically approved in accordance with the Declaration of Helsinki (REC reference number 07/Q1606/15, March 16, 2007). Participants provided written informed consent to sampling of their tumor blocks during study recruitment. Extracted RNA was sequenced and aligned as previously described (23). RNA-seq data were available on 204 primary melanoma samples, of which 10 samples were removed from downstream analyses owing to a lack of data on all clinical covariates, resulting in a final n = 194 samples. Data were also available on 175 regional lymph node samples, of which 32 samples were removed from the downstream analyses owing to a lack of data on all clinical covariates, resulting in a final n = 143 samples.

DE analyses between early-stage melanoma tumors that became metastatic versus those that remained nonmetastatic over the 6-year study period were performed using the package DESeq2(111; v1.18; R v3.6.1). The negative binomial models we considered controlled for the clinical covariates stage, Breslow thickness, Eastern Cooperative Oncology Group Performance Status, and treatment, as well as for the library size (offset), full details are available in Garg and colleagues (23). In contrast to the DE analysis used to generate the signature described by Garg and colleagues, in this analysis, only the metastasis suppressor genes, whereby reduced expression within the primary tumor was associated with patients with metastases, were considered. Additionally, the covariate corrected DE was undertaken separately on primary melanomas (n = 194) and regional lymph node tumors (n = 143) and only the intersecting genes (downregulated in both datasets) were selected for further testing and are summarized in Supplementary Table S1. These (n = 41) genes include CCR5, PLA2G2D, TNFRSF10A, KIF19, GBP5, FASLG, SOCS1, TAS2R60, SLIT1, CLIC5, AIM2, PIK3R6, GRAMD1B, LAG3, UTS2, GLDC, GCH1, PRF1, ARRDC1, SMTNL1, TRABD2A, MFSD6L, RGS7, SFMBT2, KYNU, and TRIM22.

In contrast to the study by Garg and colleagues, an additional series of DE analyses were undertaken to generate a further list of genes associated with brain metastases, identified as an important area of unmet need. Only the metastasis suppressor genes, whereby reduced expression within the primary tumor associated with patients with brain metastases, were considered. Second, as opposed to the two phenotypes considered in the “pan-organ” metastases analysis, we considered 3 different brain metastasis phenotypes resulting in 3 DE analyses: (i) brain versus extracranial metastases, (ii) brain-only metastases (defined as metastases to the brain at the first presentation of metastatic disease in the absence of any extracranial disease sites) versus extracranial metastases, and (iii) brain-only metastases versus no distant metastases. Each of these analyses was undertaken separately on the primary melanoma and regional lymph node datasets. To be considered for functional testing, a gene needed to be significantly downregulated in only one of three brain metastasis DE analyses, but significantly downregulated in both the primary melanoma and lymph node datasets.

The TCGA Melanoma Firehose Legacy (n = 287 samples), Metastatic Melanoma UCLA 2016 dataset (n = 38 samples), and other available subcutaneous melanoma datasets were analyzed in cBioPortal (40) to generate an OncoPrint of AVAST-M gene alterations in these datasets.

The 61 genes from the AVAST-M signature were subject to PANTHER analysis using http://www.pantherdb.org/ (112, 113).

Zebrafish orthologs for genes from the metastasis signature were retrieved from ZFIN (45), University of Oregon, Eugene, OR 97403-5274; http://zfin.org/, and CRISPOR (114) tool was used to design 4 sgRNAs against each target gene. For the pilot screen, 9 of 61 genes were chosen. The sgRNAs (Supplementary Table S3) were consecutively cloned into MG2X using AarI sites and then the NheI site–based restriction digestion. Additionally, sgRNAs against ptena and ptenb were cloned into a U6-sgRNA plasmid (Supplementary Table S3).

TEAZ was utilized to generate melanomas as previously described (36). To develop a baseline model of metastasis, casper mitfa:BRAF^V600Etp53^−/− transgenic animals (4–5 months old) were electroporated with MAS2X sgpten (500 ng), mitfa:GFP (112 ng) and Tol2 (66 ng) plasmids to generate BRAF^V600E, tp53^−/−, pten^−/− tumors. For the metastasis screen, plasmids electroporated per fish included U6-ptena (40 ng), U6-ptenb (40 ng), and MiniCoopR-tdTomato (320 ng) to induce melanocyte rescue, MG2X-sgRNA_12, MG2X-sgRNA_34 (100 ng each), and Tol2 (88 ng) and mitfa:Cas9 (200 ng). Briefly, adult male fish were anesthetized with 0.2% tricaine and injected with 1 μL of the plasmid mix described above into the skin below the dorsal fin. Fish were electroporated and allowed to recover in freshwater. Electroporation was performed using CM830 Electro Square Porator from BTX Harvard Apparatus and Genepaddles 3 × 5 mm, with a voltage of 45 V, 5 pulses, 60 ms pulse length and 1 s pulse interval. Fish were imaged every week using brightfield and fluorescent imaging at 25× and 10× using a Zeiss Axios microscope. Primary tumor growth was measured at 4 and 8 weeks and quantified using ImageJ. Metastasis was monitored in the kidney marrow via fluorescent imaging at 10× at 12 weeks after TEAZ, in areas separate from the primary tumor as previously described (115). Manual scoring was performed based on fluorescent images to obtain metastatic incidence. Metastatic burden was calculated by quantifying the Corrected Total Cell Fluorescence in the kidney marrows of animals at 12 weeks after TEAZ after background correction.

Tumor-bearing zebrafish were euthanized in 0.2% tricaine and fixed in 4% PFA for 72 hours. Animals were transferred to 70% ethanol and subjected to sagittal sectioning to capture the head, kidney marrow, and primary tumor on a slide. Histology was performed by HistoWiz Inc. (histowiz.com) using a Standard Operating Procedure and fully automated workflow. Samples were processed, embedded in paraffin, and sectioned at 5 μm. IHC was performed on a Bond Rx autostainer (Leica Biosystems) with enzyme treatment (1:1,000) using standard protocols. Antibodies used were rat monoclonal F4/80 primary antibody (eBioscience, 14-4801, 1:200) and rabbit anti-rat secondary (Vector, 1:100). H&E staining was performed to histologically confirm melanomas. Sections were stained with a BRAFV600E antibody (ab228461, RRID AB_2833072), phospho-AKT antibody (CST4060, RRID AB_2315049), and antibodies against GFP (ab183734, RRID AB_2732027), and RFP (#600-401-379, RRID AB_2209751) to confirm the genotypes of melanomas and stain tumor cells. Animals with tumors expressing NT and gramd1b sgRNAs were sectioned and stained with RFP to visualize metastatic tumor cells in the kidney marrow. Histology of mouse tissues was similarly performed; H&E staining was used to histologically assess the presence of metastasis in liver and lung tissues. Bond Polymer Refine Detection (Leica Biosystems) was used according to the manufacturer's protocol. After staining, sections were dehydrated and film coverslipped using a TissueTek-Prisma and Coverslipper (Sakura). Whole slide scanning (40×) was performed on an Aperio AT2 (Leica Biosystems). Antibodies and concentrations used are listed in Supplementary Table. S4.

Primers were designed to amplify an amplicon of 200 to 280 bp in length where the mutation site targeted by sgRNA is within 100 bp of the beginning or end of the amplicon (Supplementary Table S3). Fish of interest were euthanized and dissected carefully under microscopy to isolate as much of the tumor as possible, 7 to 8 weeks after TEAZ. Tail tissue of fish was used as a negative control. Genomic DNA was isolated from tumor samples using the Qiagen DNeasy Blood and tissue kit. Promega GoTaq Green master mix was used to amplify the amplicon of interest. PCR products were visualized on a DNA gel and purified to submit samples for CRISPR sequencing. CRISPR sequencing results were analyzed using CRISPR Variants R package (116) to obtain gene editing frequency in tumor samples. For surveyor assays, genomic DNA was subject to the surveyor nuclease assay (IDT). Gene editing was confirmed on a 2% agarose DNA gel.

HEK 293T (1 × 10⁸) cells cultured in 10-cm dishes were cotransfected with lentiCRISPR V2 (1.5 μg), psPAX2 (0.6 μg), and pMD2.g (0.3 μg) helper plasmids using Qiagen Effectene reagent as per manual instructions. Lentiviral supernatant was collected 72 hours after transfection and filtered. Recipient cells were infected overnight with viral supernatant containing 10 μg/mL polybrene (Sigma-Aldrich) and replenished the next day with fresh media. After 48 hours, transduced cells were cultured in fresh media containing 1 μg/mL puromycin for 7 days and then utilized for experiments.

HEK 293T (1 × 10⁸) cells cultured in 10-cm dishes were cotransfected with psPAX2 (0.6 μg), pMD2.g (0.3 μg) helper plasmids along with Dharmacon's pTRIPZ dox-inducible shRNA plasmids or Sigma's pLKO shRNA plasmids (2 μg) using Qiagen Effectene reagent as per manual instructions. Lentiviral supernatant was collected 72 hours after transfection and filtered. Recipient cells were infected overnight with viral supernatant containing 10 μg/mL polybrene (Sigma-Aldrich) and replenished the next day with fresh media. After 48 hours, transduced cells were cultured in fresh media containing 1 μg/mL puromycin for 7 days and then utilized for experiments. The dox-inducible shRNA cells were pretreated with 4 μg/mL dox for 72 hours for most experiments.

Total RNA was extracted from cells, and 1 to 2 μg of RNA was subjected to cDNA synthesis using the Invitrogen SuperScript First-Strand Synthesis Kit. Real-time PCR analysis was performed on cDNA using a 2X PowerSybr master mix. mRNA levels were normalized to ACTIN, with gene expression levels measured using a standard curve for each set of primers crossing exon–exon junctions. Real-time PCR primers are listed in Supplementary Table S3.

SKMEL2 NT or GRAMD1B KO cells were cultured in 10-cm dishes and harvested for RNA isolation in triplicates (n = 3 per group) and used for cDNA synthesis and barcoding with Illumina adapters, followed by sequencing on Illumina NovaSeq 6000. RNA-seq reads in fastq format were quantified by Salmon (117) to a prebuilt index of the hg38 human transcriptome from Ensembl accounting for GC biases. Resulting count matrices were used as input for DESEQ2 (111) in R for normalization and DE analysis using the default parameters. Normalized counts were used as input for GSEA (118, 119). Heat maps were generated using the R package complexheatmap, and waterfall plot was generated using the R package mvhspatialplots (https://github.com/mvhunter1/mvhspatialplots). Homer analysis was performed based on a previously established protocol (78) on genes upregulated in GRAMD1B KO cells.

The bulk RNA-sequencing gene expression data from the AVAST-M cohort were normalized using variance stabilizing transformation. GRAMD1B gene expression signature was obtained by standardizing GRAMD1B gene expression in all the samples to have zero mean and unit variance. This standardized GRAMD1B expression signature was used as a continuous predictor in Cox regression models fitted by means of the coxph function of the survival package (v3.2-7) in R (v4.0.3). Samples were divided into groups based on median (standardized) GRAMD1B expression values for plotting Kaplan–Meier survival curves. Here, samples having higher than or equal to median expression values were grouped as “gigh,” otherwise grouped as “low.” The Kaplan–Meier plots were made using the survminer package (v0.4.8).

The independent replication of the survival was conducted using data from the Leeds Melanoma Cohort [European Genome-phenome Archive (EGA) under accession number EGAS00001002922; ref. 24]. This is a population-ascertained cohort study consisting exclusively of primary melanomas (stages I–III) coupled with long-term clinical follow-up. mRNA extracted formalin-fixed, paraffin-embedded tissue–derived primary tumors (n = 687 samples) were analyzed on the Illumina DASL-array platform. The survival time was calculated from the time of diagnosis to the time of last follow-up or time of death from melanoma, whichever occurred first and referred to as melanoma-specific survival. The final sample size was n = 666 after removing patients who died of other causes than melanoma. In order to provide optimal group balance for the visual representation of the survival curves, the analysis was stratified on Breslow thickness <3 mm or >3 mm. In order to better reflect samples from the AVAST-M study (late-stage tumor stage IIB–IIIC), the analysis was stratified on Breslow thickness <3 mm or >3 mm. The survival analysis was also considered for all ≥ stage IIB tumors, regardless of Breslow thickness. GRAMD1B expression was dichotomized using the median value. These analyses were conducted in STATA v16.

Cells and tissues were lysed in RIPA buffer (Invitrogen) containing Halt Protease Inhibitor Cocktail (Invitrogen) and homogenized via sonication. Proteins were quantified using the Bicinchoninic Acid Assay (Invitrogen), subject to separation using Biorad protein gradient gels, and transferred to a nitrocellulose membrane. The membranes were blocked for 1 hour at room temperature in 5% milk and probed with primary antibodies in 5% milk overnight at 4°C. After incubating the membrane with the appropriate secondary antibody conjugated to horseradish peroxidase for 1 hour at room temperature, protein levels were detected with SuperSignal Dura Extended Duration substrate (Thermo Fisher Scientific). Antibodies and concentrations used are listed in Supplementary Table S4.

SKMEL2 NT and GRAMD1B KO cells were seeded into a 0.5× BME-coated 96-well invasion chamber overnight as per (R&D Systems, 3455-096-K) manual instructions. The next day, invasion was measured, and data were normalized to invasion in the control group to obtain relative invasion. Because this kit was discontinued, invasion assays in Figs. 6 and 7 were performed using a matrigel assay system (Corning Biocoat) in a 6-well format. Briefly, 250,000 cells (A375) or 500,000 cells (SKMEL2) were seeded onto Matrigel-coated invasion chambers as per the manual. The next day, Hoechst dye (1:2000) was added to the bottom chamber to stain nuclei of invaded cells. Hoechst⁺ invaded cells were imaged using a Zeiss fluorescent microscope with at least 4 to 6 fields per well, and ImageJ was used to quantify the number of invaded cells. Relative invasion was calculated by normalizing all data to the number of invaded cells in the control group. For the ERKi invasion experiments, Hoechst⁺ cells in the upper chamber were also quantified. Relative proliferation was calculated by normalizing Hoechst+ cell numbers normalized to the NT group 24 hours after drug treatment. For migration assays, 80,000 cells were seeded per insert into Corning's transwell cell-culture inserts (CLS3422) in serum-free media. Cell migration to the bottom chamber containing complete media was assessed by quantifying Hoechst⁺ cells as described above.

Animals were cared for in accordance with guidelines approved by the Memorial Sloan Kettering Cancer Center IACUC and Research Animal Resource Center. 200,000 dox-inducible SKMEL2 NT and GRAMD1B shRNA cells were injected into the tail veins of 5- to 6-week-old NOD.Cg-Prkdcscid Il2rgtm1Wjl/SzJ female mice (10 mice per group). Mice were kept on dox food for the duration of the experiment and weights and clinical signs were assessed daily. Three weeks after cell injection mice were euthanized, and gross necropsies were performed to assess metastasis. Livers, lungs, kidneys, and ovaries were collected from all mice and fixed in 4% PFA. Representative images of the organs were taken via Zeiss Axios microscopy at 6× magnification. Tissues were then submitted to Histowiz for histologic characterization.

H&E sections from representative mice (n = 2 per experimental group) were subject to color deconvolution using the Qupath software H&E Brightfield setting. Sections were then thresholded to calculate tumor-positive area within 4 to 5 regions of interest of each liver lobe. Metastatic burden was calculated as % tumor-positive liver area.

One thousand SKMEL2 NT or GRAMD1B shRNA cells were pretreated with 4 μg/mL dox for 48 h and then seeded in 96-well plates in 4 μg/mL dox. Cell proliferation was measured over time using the Cyquant Direct cell proliferation kit as per manual instructions.

SKMEL2 NT and GRAMD1B shRNA cells were pretreated with 4 μg/mL dox for 3 days and seeded in 4-well chamber dishes (100,000 cells per well) in dox media overnight. The next day, cells were stained with Filipin according to Abcam's Cholesterol Cell-Based assay kit. Briefly, cells were fixed, and Filipin was added at 1:100 dilution in the staining buffer along with 1:100 Tubulin-Alexa488 (RRID AB_1502303) as a cytoplasmic marker, and chambers were incubated for 2h at room temperature. Chambers were washed thrice in PBS, mounted using Vectashield, and imaged on a confocal microscope (LSM880) using Airyscan, capturing at least 4 images at 60× for each group. Images were analyzed in ImageJ to quantify Filipin intensity within cells and the mean pixel value of Filipin signal in each group was calculated to depict intracellular cholesterol levels.

SKMEL2 NT or GRAMD1B shRNA cells (5 × 10⁶-10 × 10⁶) were trypsinized, counted, and centrifuged to obtain cell pellets for cholesterol extraction. Half of the cells were saved for protein extraction and quantification. The remaining half of the cells were washed 1× in PBS and then extracted with 200 μL of chloroform:isopropanol:NP-40 (7:11:0.1). Extracts were sonicated at an amplitude of 10%, with 10 total pulses of 1 second. Extracts were centrifuged for 10 minutes at 15,000 × g, and the organic phase was transferred to a new tube. Tubes were air dried at 55°C to remove chloroform. The samples were then dried in a speed-vac to remove the remaining organic solvent for 20 minutes. Dried lipids were resuspended with 200 μL of Cholesterol Assay Buffer by vortexing until homogeneous. A standard curve was generated using Cholesterol standards in the kit. Cholesterol levels were quantified using the fluorescent Abcam cholesterol/cholesteryl kit as per manual instructions. For experiments in Fig. 5H and I and Fig. 6A, we used Promega's cholesterol/cholesterol ester glo assay kit (J3190) as this was more convenient in a 96-well format and amenable to high-throughput analysis of intracellular cholesterol. For these experiments, 8,000 NT or KO cells were seeded in a 96-well plate with at least 6 wells per condition. The next day, 3 wells per condition were used to measure cholesterol as per the manual instructions. The remaining 3 wells were used to quantify the number of viable cells using Promega's CellTiter-Glo kit (G7570). Cholesterol luminescence values were normalized to cell viability and then a standard curve was used to obtain cholesterol concentration per sample (μmol/L).

2 × 10⁶ SKMEL2 NT or GRAMD1B shRNA cells were pretreated in 4 μg/mL dox-containing media for 4 days to induce GRAMD1B knockdown. On day 4, the cells were trypsinized, counted, and centrifuged to obtain cell pellets for phospholipid extractions. Extracts were suspended in 500 μL Abcam's phospholipid assay buffer and sonicated at an amplitude of 10%, with 10 total pulses of 1 second on ice. Samples were centrifuged for 10 minutes at 15,000 × g at 4°C and the supernatant was transferred to fresh tubes. Samples were diluted 1:5 in assay buffer and subject to Abcam's phospholipid assay kit as per manual instructions. A standard curve was used to calculate phospholipid levels in cells in pmol.

SKMEL2 NT or GRAMD1B knockdown cells were pretreated in 4 μg/mL dox-containing media for 4 days to induce GRAMD1B knockdown. On day 4, 10,000 cells were seeded in 96-well black plates. The next day, the assay was performed according to Abcam's membrane fluidity assay kit instructions. Excimer to monomer fluorescence was normalized to obtain relative membrane fluidity in these cells.

Adult male zebrafish ages 4- to 5-month-old were electroporated with plasmids as described before to generate NT and gramd1b KO melanomas. Animals were sacrificed 8 weeks after TEAZ, and blood was collected from 4 animals per condition using a previously described method (120) to assess blood cholesterol levels. Blood from animals was centrifuged at 21,000 × g for 1 minute, and the plasma supernatant was collected into fresh tubes for cholesterol analysis. Plasma was diluted 1:1,000 in Abcam's cholesterol assay buffer for the experiment. Simultaneously, melanomas were dissected from the same animals for tumor cholesterol experiments and extracted with 400 μL of chloroform:isopropanol:NP-40 (7:11:0.1). Extracts were sonicated at an amplitude of 10%, with 10 total pulses of 1 second. Extracts were centrifuged for 10 minutes at 15,000 × g, and the organic phase was transferred to a new tube. Tubes were air dried at 55°C to remove chloroform. The samples were then dried in a speed-vac to remove the remaining organic solvent for 20 minutes. Dried lipids were resuspended with 400 μL of Cholesterol Assay Buffer by vortexing until homogeneous. Samples were further diluted 1:50 in Cholesterol assay Buffer for the enzymatic assay. Abcam's cholesterol assay kit was then followed using manual instructions to measure total and free cholesterol in blood and tumor samples. A standard curve was used to calculate the concentration of cholesterol in samples.

On day 0, 100,000 NT or GRAMD1B shRNA SKMEL2 cells were pretreated in 4 μg/mL dox-containing media in 6-well dishes to induce GRAMD1B knockdown for 48 hours. On day 2, cells were switched to either dox-containing control media (DMEM, 10% FBS, 1% antibiotics, 4 μg/mL dox) or cholesterol depletion media (DMEM, 10% lipoprotein-free FBS, 1% antibiotics, 10 μmol/L simvastatin, and 200 μmol/L mevalonic acid, 4 μg/mL dox) for another 48 hours. Because the mevalonate pathway is critical for cell viability, mevalonate supplementation was included in the cholesterol depletion mixture as done in Trinh and colleagues (121). On day 4, cholesterol depletion was confirmed through Promega's cholesterol Glo assay, and cells were seeded for invasion assays.

On day 1, 3 × 10⁶ melanoma cells were seeded in 10-cm dishes for siRNA experiments in antibiotic-free media using the Dharmafect transfection reagent as per manual instructions. Briefly, cells were transfected with 40 μL siGENOME Control siRNA or CYP27A1 siRNA and 40 μL Dharmafect 1 per dish. The next day, the cells were replenished with fresh media containing antibiotics. Cells were harvested to seed for invasion assays and western blot/real-time analysis 72 hours after transfection.

A375 or SKMEL2 cells (40 × 10³) were seeded per well in 24-well plates in triplicate and transfected with a Renilla plasmid (40 ng), pGL3-AP-11 (400 ng), or pGL3-control plasmids (400 ng) per well using FuGENE HD (Promega) at a 3:1 FuGENE:DNA ratio. For sterol treatment assays, sterols were added at the time of transfection. For assays in shRNA cells, NT or GRAMD1B shRNA cells were pretreated in 4 μg/mL dox for 48 hours and then seeded in 24-well plates in dox media and subjected to transfection. Luciferase activity was measured 48 hours after transfection using Promega's Dual Glo assay kit and a luminescence plate reader (Synergy). Firefly luciferase activity was normalized to Renilla luciferase activity to obtain relative luciferase levels per sample.

Statistical comparisons were performed with the aid of GraphPad Prism 8.4.3, and the statistical details including sample size and type of statistical test performed can be found in the figure legends. Statistical analysis was performed on data pooled from at least n = 3 independent experiments, with associated technical replicates per experiment. P  > 0.05 is not considered statistically significant. *, P < 0.05; **, P < 0.01; ***, P < 0.001, and ****, P < 0.0001.

The code used to generate the plots in this paper is publicly available on GitHub (https://github.com/Manikgarg/GRAMD1B-analysis). The RNA-seq data generated in this study are uploaded to NCBI Gene Expression Omnibus and raw data as well as processed counts files are publicly available (accession number GSE202717). The Leeds Melanoma Cohort data (accession number EGAS00001002922) are available at the EGA. Plasmids will be available upon request. Normalized counts and results of DESEQ2 DE analysis and GSEA are listed in detail in Supplementary Tables S5–S7.

S. Suresh reports grants from the Melanoma Research Foundation, Memorial MSKCC TROT T32 training grant, Cycle for Survival, the Melanoma Research Alliance, The Debra and Leon Black Family Foundation, NIH Research Program Grants R01CA229215 and R01CA238317, NIH Director's New Innovator Award DP2CA186572, NCI Cancer Center Support Grant P30 CA08748, the Pershing Square Sohn Foundation, The Mark Foundation, the American Cancer Society, The Alan and Sandra Gerry Metastasis Research Initiative, the Harry J. Lloyd Foundation, Consano, the Starr Cancer Consortium, Wellcome Trust and CR-UK, the NIH Kirschstein-NRSA predoctoral fellowship F30CA254152, Medical Scientist Training Program grant from the NIH under award number T32GM007739, Predoctoral to Postdoctoral Fellow Transition Award (F99/K00) from the NIH under award number K00CA223016, CRUK (C588/A19167, C8216/A6129, C588/A10721), and the NIH (CA83115) during the conduct of the study. R. Rabbie reports other support from AstraZeneca outside the submitted work. R.M. White reports grants from the NIH, Cycle for Survival, the American Cancer Society, Pershing Square, The Mark Foundation, the Harry J. Lloyd Foundation, The Alan and Sandra Gerry Metastasis Research Initiative at the Memorial Sloan Kettering Cancer Center, and the Starr Cancer Consortium during the conduct of the study, as well as personal fees from Qiagen/N-of-One outside the submitted work. No disclosures were reported by the other authors.

S. Suresh: Conceptualization, resources, data curation, software, formal analysis, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. R. Rabbie: Conceptualization, resources, data curation, software, formal analysis, investigation, visualization, methodology, writing–original draft, writing–review and editing. M. Garg: Resources, data curation, software, formal analysis, validation, investigation, visualization, methodology, writing–review and editing. D. Lumaquin: Data curation, software, formal analysis, investigation, visualization, methodology, writing–review and editing. T.-S. Huang: Resources, methodology, writing–review and editing. E. Montal: Software, formal analysis, investigation, visualization, methodology. Y. Ma: Data curation, software, formal analysis, investigation, visualization, writing–review and editing. N.M. Cruz: Conceptualization, formal analysis, methodology, writing–review and editing. X. Tang: Data curation, formal analysis, investigation, writing–review and editing. J. Nsengimana: Data curation, formal analysis, writing–review and editing. J. Newton-Bishop: Data curation, formal analysis, methodology. M.V. Hunter: Software, visualization, methodology, writing–review and editing. Y. Zhu: Formal analysis, investigation. K. Chen: Formal analysis, investigation. E. de Stanchina: Conceptualization, data curation, formal analysis, investigation, visualization, methodology, writing–review and editing. D.J. Adams: Conceptualization, resources, supervision, funding acquisition, visualization, methodology, project administration, writing–review and editing. R.M. White: Conceptualization, resources, supervision, funding acquisition, writing–original draft, project administration, writing–review and editing.

S. Suresh was supported by a Melanoma Research Foundation Career Development Award grant number 719502 and an MSKCC TROT T32 training grant T32 CA16000. D.J. Adams and R. Rabbie were supported by the Wellcome Trust and CR-UK. D. Lumaquin was supported by NIH Kirschstein-NRSA predoctoral fellowship F30CA254152. Y. Ma was supported by a Medical Scientist Training Program grant from the NIH under award number T32GM007739. E. Montal was supported by the Predoctoral to Postdoctoral Fellow Transition Award (F99/K00) from the NIH under award number K00CA223016. M. Garg was supported by an EMBL doctoral fellowship. We thank the Aquatics Core facility, Memorial Sloan Kettering Cancer Center, for their support in this work. We also thank members of the White lab for their advice in this work. We acknowledge the use of the Anti-tumor Assessment Core, Integrated Genomics Operation Core, funded by an NCI Cancer Center Support Grant (P30 CA08748). The Leeds Melanoma Cohort was funded by CRUK grants C588/A19167, C8216/A6129, an C588/A10721 and NIH grant CA83115.

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.