FABP4 as a key determinant of metastatic potential of ovarian cancer

1Primary cytoreductive surgery followed by adjuvant chemotherapy is the standard treatment for ovarian cancer. Several reports have established a link between residual disease after surgery and shorter overall and progression-free survival, as well as poor response to adjuvant chemotherapy–. Incomplete resection can result from the presence of numerous, dense nodules that simply cannot be removed, distant tumor metastasis, location of tumor near critical organs (e.g., porta hepatis), and extensive mesenteric involvement. Thus, while a surgeon’s skill is important, residual disease could occur because the intrinsic biology of the tumor creates a highly metastatic and infiltrative disease pattern, making complete resection infeasible,. A large retrospective study showed that initial disease distribution could also affect patient survival despite aggressive cytoreductive efforts. Considering the impact of residual disease on patient survival, many studies are now focusing on the development of predictive models for residual disease. However, currently there is little understanding of the biological mechanisms that could lead to residual disease in ovarian cancer. To address this knowledge gap, in the current study, we explored the underlying tumor biology responsible for the aggressive pattern of tumor metastasis.

2We have previously shown that high expression of fatty acid binding protein 4 (FABP4) can be a reliable molecular predictor of residual disease in high-grade serous ovarian cancer. Hence, we decided to investigate the mechanisms that lead to upregulation of FABP4 in ovarian cancer, as well as its downstream effects that lead to metastasis. FABP4 is implicated in atherosclerosis, diabetes, inflammatory response, and angiogenesis–. In prostate and ovarian cancers, FABP4 acts as a key mediator between adipocytes and cancer progression,. However, FABP4 regulation and the functional implications of FABP4 overexpression in increasing the metastatic potential of ovarian cancer cells remain to be investigated.

3Here, we examined the biological effects of increased FABP4 expression in ovarian cancer. Through extensive bioinformatics analyses, we identified hypoxia mediated downregulation of miR-409-3p as a key regulator of FABP4. FABP4 siRNA or a miR-409-3p mimic led to significant inhibition of tumor metastasis. Downstream of FABP4, we observed that FABP4 can affect several metabolites as well as metastasis-related pathways in ovarian cancer. Collectively, this study provides an understanding of the causal biology of metastatic potential in ovarian cancer.

4To investigate the functional role of FABP4 in ovarian cancer progression, we first knocked down FABP4 in HeyA8 MDR ovarian cancer cells (which have high FABP4 expression) using small interfering RNA (siRNA). FABP4 silencing led to a significant reduction in both migration and invasion of HeyA8 MDR cancer cells (Fig. 1a, b, Supplementary Fig. 1a and b) (p < 0.01). Conversely, ectopic expression of FABP4 in A2780-ip1 ovarian cancer cells led to increased migration (p < 0.05; Fig. 1c) and invasion of cancer cells (p < 0.01; Fig. 1d, Supplementary Fig. 1c and d). The effect was consistent when tested using a second siRNA sequence against FABP4 (Supplementary Fig. 1e and f), as well as another ovarian cancer cell line (Ovcar 5; Supplementary Fig. 1g and h).

5Next, we examined the effect of FABP4 on tumor progression in orthotopic mouse models of ovarian cancer. We first transfected A2780-ip1 cells that have low endogenous FABP4 levels with a lentivirus expressing either a control vector or FABP4-expressing vector. We then injected the cells into the ovaries of athymic nude mice. This orthotopic model mimics the ovarian cancer metastasis pattern seen in patients,. Compared with the control mice, mice injected with cells ectopically expressing FABP4 had a significantly higher tumor burden (~3-fold, p < 0.01; Fig. 1e), as well as an increased number of distant metastatic nodules (~3-fold, p < 0.01; Fig. 1f). We examined hematoxylin and eosin-stained sections of tumor nodules from both groups. In the control group, the cancer cells did not invade normal organs (muscle, pancreas), however, after ectopic expression of FABP4, the cancer cells were found to be deeply invading the normal tissues (Fig. 1g and Supplementary Fig 1i). The tumor metastasis rate with FABP4 overexpression was significantly higher at sites such as the diaphragm, liver, and pelvis, whereas tumors in the control mice were mainly observed at the site of injection (ovary) and in the mesentery (Fig. 1h). Expression of FABP4 was examined in the tumor sections using immunohistochemistry, and expression of FABP4 was higher in sections from the ectopic FABP4 group than in those from the control group (Supplementary Fig. 1j). We also noted a positive correlation between the expression of FABP4 and the invasive capacity of ovarian cancer cell lines in vitro (Supplementary Fig 1k). Cell lines with higher FABP4 expression (HeyA8 MDR and Ovcar 5) had higher invasive potential compared to A2780 and SKOV3 cells.

6We next investigated the effects of FABP4 knockdown in orthotopic mouse models of ovarian cancer. We first injected HeyA8 MDR-Luc cells into the ovaries of nude mice. Seven days after cell injection, DOPC nanoliposomes containing control siRNA or FABP4 siRNA were administered intraperitoneally twice per week. The mice that received FABP4 siRNA had a significant decrease in aggregate tumor weight (80%, p < 0.01) as well as number of metastatic nodules (75%, p < 0.01) compared with control mice (Fig. 1i, j). The effect was also observed in the pattern of metastasis. The treatment group had a significantly lower rate of metastasis, and the control group had widespread metastases to the diaphragm, liver, spleen, lymph node, and pelvic areas (Fig. 1k). Bioluminescence images also supported the finding that FABP4 siRNA led to decreased metastasis in treated mice (Fig. 1l). The treatment did not affect mouse body weight (Supplementary Fig. 2a). We then checked expression of FABP4 in the tumors collected from the control and treatment groups. FABP4 expression was significantly decreased after treatment with FABP4 siRNA, as observed in Western blot analysis (Supplementary Fig. 2b) and immunohistochemical staining (Supplementary Fig. 2c).

7We also examined an additional orthotopic mouse model in which Ovcar 5 cells were injected directly into the ovary. Similar to the previous experiment, FABP4 siRNA treatment resulted in significant reduction in tumor weight (p < 0.05; Supplementary Fig. 2d) and the number of nodules (p < 0.05; Supplementary Fig. 2e). Metastasis was also reduced at various sites, including the peritoneum, mesentery, and diaphragm (Supplementary Fig. 2f). The treatment had no effect on the body weight of the mice (Supplementary Fig. 2g).

8Next, we investigated the potential mechanisms by which FABP4 is regulated in tumors. We first sought to determine whether there is any association between copy number and mRNA expression of FABP4. We examined The Cancer Genome Atlas (TCGA) datasets across platforms (Affymetrix, Agilent, RNASeqv2) and found no correlation between copy number and mRNA expression (Supplementary Fig. 3a). We also examined whether microRNAs (miRNAs) could regulate FABP4 levels. Our miRNA selection process is presented in Fig. 2a. We probed the TCGA dataset and selected samples from chemotherapy-naive cases of high-grade serous ovarian cancer. In these samples, we looked for miRNAs that were associated with FABP4 levels. To narrow our search further, we used the maximum information coefficient score, where a score of 0 indicates no association and a score of 1 indicates perfect association between a gene and a miRNA. Thirty-two miRNAs with a score >0.2 were selected (Supplementary Table 1). We then selected miRNAs from this group that were predicted to bind to the 3′ untranslated region (UTR) of FABP4 by miRNA–gene interaction prediction software. Mir-409-3p was predicted to target FABP4 by nine different prediction software programs.

9Next, we transfected HeyA8 MDR cells with miR-409-3p mimic, which led to a significant decrease (85%, p < 0.01) in FABP4 expression (Fig. 2b). This effect was also observed at the protein level (Fig. 2c). Similar effect was observed in another cell line (Ovcar 5, p < 0.01, Supplementary Fig. 3b; Ovca 432, p < 0.05, Supplementary Fig. 3c). We also transfected ovarian cancer cells (Ovcar 3) with anti-miR-409-3p and then examined FABP4 expression. Inhibition of miR-409-3p led to a significant increase in FABP4 expression (p < 0.05; Supplementary Fig. 3d).

10To determine whether miR-409-3p binds to the 3′UTR of FABP4, we conducted a luciferase assay. The predicted binding site for miR-409-3p on FABP4 is presented in Supplementary Fig 3e. We also mutated the binding site; the original sequence and mutated sequences are shown in Fig. 2d. In cells treated with a miR-409-3p mimic, the relative luciferase activity was significantly lower for cells transfected with FABP4 3′UTR than in cells transfected with an empty control. When FABP4’s 3′UTR binding site was mutated, abrogation of binding activity was observed (Fig. 2e), suggesting that miR-409-3p binds to the 3′UTR of FABP4.

11We next compared the expression of miR-409-3p in various ovarian cancer cell lines and non-transformed ovarian epithelial cells (HIO180). Compared with the control cells, all ovarian cancer cell lines showed low expression of miR-409-3p (Fig. 2f). We also analyzed the expression of FABP4 in these cell lines to determine the correlation between FABP4 and miR-409-3p at the basal level (Supplementary Fig 3f) and observed an inverse association between miR-409-3p and FABP4 in the cell lines. We then probed a Gene Expression Omnibus dataset (GSE15190) and compared the expression of miR-409-3p in serous ovarian tumor tissues with that of normal ovarian tissue samples. The dataset revealed that miR-409-3p is expressed at significantly lower levels in ovarian tumors than in normal ovarian samples (Fig. 2g). We also examined 25 patient samples for correlations between miR-409-3p and FABP4 expression. In situ hybridization staining revealed a clear inverse association in the staining pattern of FABP4 and miR-409-3p (Fig. 2h).

12To investigate whether the functions of miR-409-3p are mediated through FABP4, we performed a rescue experiment where we stably expressed a FABP4 construct lacking the 3′UTR in SKOV3ip1 cells so that it would not be sensitive to a miR-409-3p mimic. Compared to controls, SKOV3ip1 cells with ectopically expressed FABP4 were significantly more invasive (p < 0.05). Treatment of the cell line ectopically expressing FABP4 with miR-409-3p mimic did not result in a significant decrease in invasion (Fig. 2i), suggesting that FABP4 is an important target of miR-409-3p.

13To determine whether miR-409-3p has tumor-suppressive properties, we transfected HeyA8 MDR cells with control miRNA or miR-409-3p mimic and conducted migration and invasion assays. Transfection of the miR-409-3p mimic led to a significant decrease in the migratory and invasive potential of cancer cells (p < 0.01; Fig. 3a, b). In contrast, reduction of miR-409-3p with anti-miR-409-3p resulted in increased migration (p < 0.05) and invasion of cancer cells (Supplementary Fig. 4a and b). The effect of miR-409-3p mimic transfection was also observed with the Ovcar 5 cells. Migration (Supplementary Fig 4c) and invasion (Supplementary Fig 4d) potential of Ovcar 5 cells were significantly reduced after miR-409-3p transfection (**p < 0.01).

14To examine the effects of miR-409-3p in vivo, we used the HeyA8 MDR orthotopic mouse model. Starting seven days after cell injection, DOPC nanoliposomes containing control miRNA or miR-409-3p were administered intraperitoneally twice weekly. Compared with controls, mice treated with miR-409-3p mimic showed significant reduction in tumor weight (~60%, p < 0.01; Fig. 3c) and number of tumor nodules (~60%, p < 0.01; Fig. 3d). The treated mice also showed a different pattern of metastasis. Control mice had widespread metastasis in lymph nodes, diaphragm, pelvis, perisplenic, and perihepatic areas, whereas the incidence of metastasis was significantly lower in treated mice (Fig. 3e). Importantly, treatment with miR-409-3p mimic led to significant reduction of FABP4 protein (Fig. 3f, g) and RNA levels (p < 0.05; Supplementary Fig. 4e). There was no deleterious effect of the treatment on the mice weight (Supplementary Fig. 4f).

15To identify the effects of the tumor microenvironment on FABP4 expression, we next examined various factors, including co-cultures with fibroblasts, macrophages, mesothelial cells and hypoxia treatment (Supplementary Fig. 5a). Of the factors examined, hypoxia (1% O2) consistently led to a significant increase in FABP4 expression in cancer cells at 24 and 48 h (Fig. 4a). We also examined the expression of miR-409-3p after hypoxia treatment and observed low expression in hypoxic compared with normoxic cells at 24 and 48 h (Fig. 4b). We observed the same effect in an additional cell line (Ovca 432; Supplementary Fig. 5b and c); upon exposure to hypoxia, the expression of miR-409-3p was significantly reduced, whereas FABP4 expression was significantly increased.

16We then assessed the effect of hypoxia on miR409-3p expression in tumors treated with bevacizumab. This therapy, which targets vascular endothelial growth factor, has been shown to lead to hypoxia in tumors. Hence, we compared the miR-409-3p expression between bevacizumab-treated and control tumors. Expression of miR-409-3p was lower in tumors from mice treated with bevacizumab compared to control tumors (Fig. 4c). We then stained the tumor sections taken from these mice with CA9 (a hypoxia marker) and FABP4; we observed co-localization of FABP4 with CA9 (Fig. 4d). We further analyzed the TCGA and Tothill ovarian cancer datasets and observed a significant positive correlation between FABP4 and PLAU (a gene from hypoxia metagene signature) across all platforms (Fig. 4e). The correlations of FABP4 with other genes from the metagene signature are presented in Supplementary Table 2.

17To identify signaling pathways downstream of FABP4, we conducted reverse phase protein arrays (RPPA) after silencing FABP4 in HeyA8 MDR cells. The heat map of the RPPA data (Fig. 5a) showed that several proteins (Supplementary Fig. 6a and Supplementary Data 1) were downregulated after FABP4 knockdown. We further used Netwalker software to analyze the protein data, and the results showed sub pathways pertaining to metastasis being significantly downregulated in the FABP4 siRNA group (Fig. 5b).

18To explore the effects of FABP4 on metabolic changes in human ovarian cancer tissues, we first analyzed previously published gene expression and metabolomics data of high grade serous ovarian carcinoma tissues–. From those results, several metabolites were specifically expressed in either low or high FABP4 expressing patient tumor samples (Supplementary Table 3). Next, we evaluated metabolic and lipid trends related to FABP4 gene expression by desorption electrospray ionization mass spectrometry imaging (DESI-MSI). In this study, a total of 31 high-grade ovarian cancer tissue samples were analyzed by DESI-MSI in both negative and positive ion modes to investigate a broad range of lipid and metabolites related to FABP4 expression. Using this approach, only the data specific to the tumor region were selected and evaluated, thus eliminating the inclusion of stromal lipids and metabolites.

19SAM analysis (Significance Analysis of Microarrays) was conducted to identify statistically significant m/z values (mass to charge ratios) detected by DESI-MSI characteristic of low- or high-FABP4 expression. Positive SAM scores represent higher relative abundance in samples with low-FABP4-expression, and negative SAM scores represent higher relative abundance in samples with high-FABP4-expression. Negative ion mode data revealed 361 and 289 monoisotopic m/z values associated with low or high FABP4 expression, respectively, with false discovery rate (FDR) of <3.8%. From these, 238 m/z values were tentatively identified by high mass accuracy measurements and tandem MS analysis (Supplementary Table 4) as small metabolites, fatty acids (FA), ceramides (Cer), glycerolipids (GL), monoacylglycerophosphates (PA), glycerophosphoethanolamines (PE), glycerophosphoglycerols (PG), glycerophosphoinositols (PI), glycerophosphoserines (PS), and cardiolipins (CL). Molecular formulas, detected m/z and mass errors in ppm are provided in Supplementary Table 5 for the species identified. Fatty acid and metabolite composition was observed to significantly differ between the two expression groups. For example, higher unsaturation and oxidation of fatty acid species was selected as characteristic of high FABP4 expression. A significant number of m/z values related to high FABP4 expression corresponded to GL, PE, PG, and lysophospholipid species, such as LysoPE, LysoPG, and LysoPI. On the other hand, CL species were almost exclusively selected with the low-FABP4-expression type. These variations in lipid composition suggest alterations in metabolism and biosynthesis due to the changes in the expression of the FABP4 gene.

20We then applied the Lasso (Least absolute shrinkage and selector operator) method to evaluate if predictive statistical models could be built to classify samples as having low- or high-FABP4 expression based on metabolite composition. The samples were randomly divided into a training and validation set of samples (50:50). Results based on negative ion mode data yielded an overall agreement of 81.6% and 0.79 area under the curve (AUC) value for the training set performance. The remaining independent set of samples were classified with an overall agreement of 56.1% and AUC value of 0.60 (Table 1). In positive ion mode data, an overall agreement of 74.2% for CV (AUC = 0.73) and 61.7% agreement for the validation set (AUC = 0.58) (Table 1) were achieved. To improve prediction of FABP4 expression, positive and negative ion mode data were combined to build a comprehensive model. Using this approach, 66.7% of high-FABP4 patients and 100% of low-FABP4 patients were successfully classified in patient-by-patient analysis (Fig. 5c and Table 1).

21To further investigate direct changes in lipid and metabolite levels due to the changes in FABP4 expression, we analyzed in vivo tissues from high-grade ovarian cancer mouse models presented in Fig. 1i–l. Using DESI-MSI, we compared tumor tissues of FABP4 siRNA group (low FABP4 expression, n = 3) to control siRNA samples (high FABP4 expression, n = 3). SAM analysis revealed 627 monoisotopic m/z values characteristic of FABP4 silencing, with FDR < 5.5%. From the selected m/z values, a total of 184 species (59 increased in low FABP4, 125 increased in high FABP4) were identified as metabolites, FA and complex lipids, with attributed SAM scores associated to FABP4 expression (Supplementary Table 6). Overall, the majority of the lipid classes received negative SAM scores, thus being associated to high FABP4 expression. For example, FA and monoacylglycerophosphates (PA) were selected exclusively for characterization of the control siRNA samples. On the other hand, Cer species were associated with low expression of the FABP4 gene. These trends were also reflected by the relative abundances observed in the ion images as shown in Fig. 5d, where all the lipid species shown except for the Cer d42:1 displayed higher relative abundances within the high FABP4 samples. Remarkably, comparison between the metabolic species selected by SAM analysis for both human and mouse samples resulted in a total of 76 common m/z values related to high or low FABP4 expression (Table 2, Supplementary Table 7). Among these, various FA species were associated with high expression of the FABP4 gene in both the mice and human samples. Interestingly, common GL and PI species for both animal samples were specific to the high FABP4 class, while Cer species were characteristic of tissues with low FABP4 expression. Differences in FA chain length and saturation level were also observed within the same GP class. For example, longer FA chain in PE and lysoPE species were associated with high FABP4 expression, while longer FA chain in PS lipids were characteristic to tissue samples with lower FABP4 expression. These similar metabolic trends observed in both human and mice samples strongly suggest and corroborate the effects of the FABP4 gene in metabolism and biosynthesis.

22High FABP4 expression was significantly correlated with poor overall (Fig. 6a) and progression-free (Fig. 6b) survival (p < 0.01) in the Tothill dataset for both FABP4 probe sets. The multivariate analysis for overall (Supplementary Fig 6b) and progression-free (Supplementary Fig 6c) survival as well as the univariate analysis for overall (Supplementary Fig 6d) and progression-free (Supplementary Fig 6e) survival are presented.

23Having established the importance of FABP4 as a candidate target for ovarian cancer treatment, we next wished to identify clinically approved drugs that could be used to inhibit FABP4. Tamoxifen, a selective estrogen receptor modulator, has been shown to inhibit FABP4 in macrophages. We checked whether it can be used to target FABP4 in ovarian cancer cells. We used physiologically relevant concentrations of tamoxifen to determine its effect on FABP4 expression and function in ovarian cancer cells. At a concentration of 3.5 µM, tamoxifen decreased the expression of FABP4 by approximately 50% (p < 0.05; Supplementary Fig. 7a). As a result of this downregulation, the ability of cancer cells to take up free fatty acids was also impaired. The cells treated with tamoxifen exhibited lower uptake compared to untreated control cells (p < 0.01; Supplementary Fig. 7b). The treatment also affected the migratory potential of cancer cells and resulted in a significant decrease in the number of migrated cells (Supplementary Fig. 7c). The effect of tamoxifen on FABP4 was also observed in A2780 cells where tamoxifen treatment significantly downregulated FABP4 expression (Supplementary Fig 7d) as well as the ability of cells to migrate (Supplementary Fig 7e) (**p < 0.01).

24Systematically coupling bioinformatics, experimental models and clinical data, we elucidated aspects of the underlying biology of ovarian cancer that could lead to aggressive metastasis of ovarian cancer. High expression of FABP4 is a known predictor of residual disease. Our preclinical studies show that ectopic expression of FABP4 increases tumor progression while in vivo experiments with DOPC nanoliposomes show that silencing FABP4 reduces the extent of metastasis. Our findings thus provide a promising therapeutic strategy that can potentially be useful for patients at high risk of being left with residual disease at initial surgery. Previous studies have shown that FABP4 present in the stromal cells can lead to cancer progression by providing a source of energy to cancer cells or increasing angiogenesis,. However, our study presents cancer cell specific FABP4 as a crucial player in promoting invasive and infiltrative metastasis in ovarian cancer. Metabolic changes in ovarian cancer have been studied before,. However, without spatial information, it is difficult to determine whether the metabolic alterations are occurring in the tumor or stromal compartment. To focus on the tumor compartment, we employed the DESI-MS imaging technique which allowed us to identify the chemical species present in the tissue as well as their spatial distribution. Analysis of the data, specific to tumor compartment of patient samples revealed that a unique metabolic profile is associated with the expression of FABP4. Higher unsaturation and oxidation of fatty acids and lysophospholipids were observed in higher relative abundance in samples with higher FABP4 expression. Several studies have suggested that unsaturated fatty acids play a key role in pathways leading to tumor progression by activating the beta catenin pathway, downregulating PTEN or increasing cancer cell adhesion–. Fatty acid oxidation has also been shown to increase metastasis in breast and ovarian cancer models,. Lysophospholipids have also been known to stimulate cancer cell migration and they are also considered as potential biomarkers for ovarian cancer,. High expression of FABP4 can thus regulate various metabolites and protein pathways that can lead to aggressive metastasis of ovarian cancer. This data elucidating the key metabolic determinants of ovarian cancer metastasis thus lay a strong foundation for future studies which can focus on the development of therapeutic strategies targeting these species.

25Previous studies have identified PPARγ and endothelial DLL4-NOTCH as regulatory factors for FABP4,,. Our findings provide a new understanding of FABP4 regulation by miR-409-3p. MiR-409-3p has been known to target several genes such as MGMT, HMGN5 (glioblastoma),, beclin, NLK and GAB1 (Colon cancer)–, AKT, ZEB1 (Breast cancer),, c-Met (Lung cancer), radixin, PHF10 (gastric cancer),. In these cancer types, miR-409-3p has been known to act as a tumor suppressor whereas in prostate cancer miR-409-3p has been shown to increase EMT and stemness markers in cancer cells. The role of miR-409-3p in ovarian cancer, however, was not known. Here, we demonstrate that miR-409-3p acts as an effective tumor suppressor in ovarian cancer through its effects on FABP4. The antitumor properties of miR-409-3p stem from its inhibitory effects on FABP4, as shown by the fact that restoration of FABP4 rescued the tumor suppressive effects of miR-409-3p on cancer cells. Importantly, we show that the DOPC nanoliposomes containing a miR-409-3p mimic effectively reduce tumor burden and metastasis in vivo. Hypoxia has been linked with ovarian cancer progression,. However, the exact role of hypoxia in tumor metastatic patterns was not well understood. We found that hypoxia can downregulate miR-409-3p, thus potentially removing its inhibitory effect on FABP4, which subsequently results in increased FABP4 levels. Recent findings have also suggested that HIF-1α can directly bind to the promoter region of FABP4. The cross-talk between these regulators of FABP4 and their specific place and importance in the FABP4 pathway presents an interesting avenue for future studies.

26Previous studies have shown that increased fatty acid uptake by the cells lead to increased motility and EMT,. Since FABP4 is a key molecule involved in free fatty acid uptake, inhibiting its ability to uptake fatty acid could be a potential therapeutic avenue to inhibit cancer metastasis. We thus identified tamoxifen as a potential drug of interest that could inhibit FABP4 and affect migration of ovarian cancer cells. Although, additional research is warranted, our data suggest tamoxifen as a promising drug for high grade serous ovarian cancer treatment, possibly in a maintenance setting.

27In summary, our study provides a new understanding of a previously unrecognized mechanism (Fig. 7) by which FABP4 leads to altered tumor metastasis patterns and a higher extent of disease in ovarian cancer. Collectively, our data offer candidate targets for therapeutic interventions for ovarian cancer patients.

28All cancer cell lines were maintained at 37 °C and 5% CO2 in culture with RPMI-1640 or Dulbecco’s modified Eagle medium supplemented with 15% fetal bovine serum and 0.1% gentamicin sulfate (Gemini Bio-Products, Calabasas, CA, USA). HeyA8MDR, A2780, SKOV3ip1, Ovcar 5, and Ovcar 3 were received from the MDA Characterized Cell Line Core (CCLC) or directly from ATCC. Ovca 432 was received from the originator, Dr. Ronny Drapkin.

29All siRNA transfections were conducted using Lipofectamine2000 as a transfecting agent. SiRNA concentration of 100 nM was used, and the ratio of Lipofectamine (Life Technologies, Carlsbad, CA, USA) to a specific siRNA was 3:1. The cells were treated with siRNAs for 4 h in serum-free media before incubation in complete media for the specified time frame. For miRNA transfections (mimic and anti-miRNA), RNAiMAX (Life Technologies, Carlsbad, CA, USA was used as a transfection agent, and the ratio of RNAiMAX to a specific miRNA was 2:1. The concentration of miRNA-mimic or anti-miR inhibitor was 40 nM concentration and the transfections were conducted in serum-free conditions. After 4 h, the cells were incubated in complete media for the specified time frame. The sequences for siRNAs and miRNAs are listed in Supplementary Table 8.

30Female athymic nude mice were purchased from Taconic Farms (Hudson, NY) and housed in pathogen-free conditions. The mice were cared for according to the guidelines of the American Association for Accreditation for Laboratory Animal Care International and the US Public Health Service Policy on Humane Care and Use of Laboratory Animals. All in vivo experiments and protocols were approved by MD Anderson’s Institutional Animal Care and Use Committee.

31To establish the tumors, 1 × 106 cells/mouse A2780 cells, 8 × 105 cells/mouse HeyA8 MDR-Luc cells, or 1 × 106 cells/mouse Ovcar 5-Luc cells were injected into the ovary. No therapeutic intervention was conducted in the experiment in which FABP4-overexpressing cells were injected. For all therapeutic experiments, a siRNA or miRNA dose of 200 μg kg−1 was used, and the treatments were started 1 week after cell injections. The mice were divided into two groups: control and treatment, 10 mice/group. Mice in the control group received control siRNA or control miRNA incorporated into neutral DOPC liposomes. Mice in the treatment group received FABP4 siRNA or a miR-409-3p mimic incorporated into DOPC liposomes. The doses were given twice weekly intraperitoneally. The mice were monitored daily for any toxic effects. Luciferase imaging was conducted to observe the effect of treatments on the metastasis as described previously. Briefly, imaging and data acquisition were performed using the IVIS Spectrum in vivo imaging system coupled to Living Image Software (Xenogen). The mice were first anesthetized in an acrylic chamber with a mixture of 1% isoflurane. They were then injected intraperitoneally with luciferin potassium salt (15 mg ml−1) in PBS at a dose of 150 mg kg−1 body weight. A digital grayscale image was initially acquired, which was then overlaid with a pseudocolor image representing the spatial distribution of detected photons emerging from active luciferase present within the animal. Signal intensity was expressed as a sum of all photons detected per second. Once a mouse in any group became moribund, all mice were euthanized. Mouse weight, tumor weight, number of nodules, and locations of metastasis were recorded. Tumor tissues were then frozen in optimal cutting temperature media, fixed in formalin for paraffin embedding, or snap-frozen.

32Incorporation of siRNA or miRNA into DOPC liposomes was achieved as previously described. Briefly, DOPC and siRNA or miRNA were mixed at a ratio of 1:10 (w/w) in the presence of tertiary butanol. Tween 20 was added to the mixture at a ratio of 1:19. The mixture was vortexed and frozen in an acetone/dry ice bath and lyophilized. The lyophilized preparations were hydrated with PBS at room temperature to a concentration of 200 µg of siRNA or miRNA kg−1 per injection per mouse.

33TCGA cases were included if the patients had high-grade disease and no history of pre-treatment. TCGA gene expression quantification were produced using justRMA applied to the CEL files from TCGA (Affymetrix HT HG-U133A arrays, n = 598). TCGA miRNA microarray level 3 data (Agilent 8 × 15 K Human miRNA-specific microarray) was obtained from the Data Access Matrix. Data was available for 799 miRNAs. There were 541 samples meeting the inclusion criteria which had both gene expression and miRNA data available. The association between each probe for the gene of interest and each available miRNA was assessed using the Maximal Information Coefficient (MIC) computed using the publicly available MINE software of Reshef et al.

34We retrieved miRNA–target interaction information from miRWalk2.0 (http://www.umm.uni-heidelberg.de/apps/zmf/mirwalk/) that hosts miRNA–target predictions from twelve programs. We selected the cases predicted by at least seven algorithms (half of the total number of programs checked + 1). The interaction between miR-409-3p and FABP4 (NM_001442) (3′UTR) was predicted by nine programs (miRWalk, Microt4, miRanda, miRMap, miRNAMap, PITA, RNA22, RNAhybrid, Targetscan). We used Perl to sort the information available and Latex to present the sites most probable to interact.

35Luciferase assays were conducted as described previously. Briefly, a GoClone pLightSwitch luciferase reporter for the 3′UTR of FABP4 was purchased from Switchgear Genomics (Menlo Park, CA, USA). HeyA8 MDR cells were transfected with control miRNA or miR-409-3p mimic (100 nM; Life Technologies, Carlsbad, CA, USA) with the help of FuGENE HD transfection agent. The cells were also transfected with 3′UTR reporter constructs and Cypridina TK controls (pTK-Cluc). After 24 h, the LightSwitch Dual Luciferase assay kit was used and the luciferase signal was measured using a microplate luminometer, as per the manufacturer’s guidelines (Biotek, Winooski, VT, USA). Luciferase activity was normalized using the Cypridina TK control, and an empty 3′UTR construct was used as a negative control. The ratios were then normalized to the scrambled control miRNA. Mutant FABP4 3′UTR was created for the predicted binding site mentioned in Fig. 2e using a QuickChange lightning multi-site-directed mutagenesis kit (Agilent Technologies, Santa Clara, CA, USA) using the primers mentioned in Supplementary Table 8 [AKS1] [PS2]. [SK3] The mutation was then confirmed using Sanger DNA sequencing before the mutant FABP4 3′UTR was used for the luciferase assay.

36Protein lysates from tumor tissues or cultured cells were prepared using modified RIPA buffer containing proteinase and phosphatase inhibitors. Protein concentrations were determined using the BCA protein assay reagent kit (Pierce Biotechnology, Rockford, IL, USA). Lysates were loaded and separated on SDS-polyacrylamide gels. The proteins were then transferred to a nitrocellulose membrane. The membrane was blocked at room temperature for 1 h in 5% milk powder in Tris-buffered saline with Tween 20 (TBST) and then incubated at 4 °C overnight with primary antibodies: FABP4 (catalog number HPA002188; SigmaAldrich) and vinculin (catalog number V9131; Sigma) or beta-actin (catalog number A5316, Sigma) as loading control. After washing with TBST, the membranes were incubated with horseradish peroxidase-conjugated donkey anti-rabbit IgG (for FABP4) or sheep anti-mouse IgG (for vinculin) (catalog numbers NA934 and NA931; 1:2000; GE Healthcare) at room temperature for 2 h. An enhanced chemiluminescence detection kit (catalog number NEL104001EA; Pierce Biotechnology) was used to visualize the horseradish peroxidase signal. Full blots are included in Supplementary Fig. 8.

37Corning Transwell inserts (Costar, MA, USA) were coated with 0.1% gelatin (migration) or defined basement membrane matrix (invasion). Defined basement membrane matrix (for invasion) was prepared in a 10 ml stock solution with laminin (50 µg ml−1), 1 ml of type IV collagen (50 µg ml−1), 0.2 ml of gelatin (2 mg ml−1), and 4 ml and 4.8 ml of PBS. In the upper chamber, HeyA8 MDR and Ovcar 5 cells (0.7 × 105) suspended in 200 μl of serum-free media were added 48 h after siRNA/miRNA transfections or tamoxifen treatments. In the lower chamber, complete media for cells containing 10% fetal bovine serum (500 μl) was added as a chemo-attractant. The chambers were incubated at 37 °C in 5% CO2 for 6 h (migration assay) or 24 h (invasion assay). After incubation, the cells in the upper chamber were removed with cotton swabs. Cells were fixed, stained, and counted using light microscopy. Cells from five random fields were counted. Experiments were done in triplicate.

38High-grade ovarian tumor samples were obtained from the Cooperative Human Tissue Network (CHTN) and MD Anderson Tissue Bank under an approved IRB protocol and written consent was obtained for the use of patient samples for research. The details regarding the quality control for the samples obtained from CHTN can be found at https://www.chtn.org/quality.html

39The detailed method can be found in Wyaman et al, 2009. Briefly, 4 samples of HOSE (normal primary ovarian surface epithelial cells taken from postmenopausal women) and 19 serous epithelial ovarian cancer samples were analyzed. The tumor samples were high-grade and contained more than 70% malignant epithelial cellularity. The cloning frequency of a specific miR expressed as a fraction of total reads from a given sample was used to compare the relative expression of miRNAs between samples.

40In situ hybridization was done as previously described. High grade serous ovarian cancer tissues which were chemonaïve were selected. Xylene and an ethanol dilution series were used for deparaffinization and rehydration of the formalin-fixed, paraffin-embedded tissue sections. Tissue sections were digested with 15 µg ml−1 proteinase K for 20 min at room temperature and then loaded onto Ventana Discovery Ultra (Tucson, AZ) for in situ hybridization analysis. The tissue slides were then incubated with a double-digoxigenin-labeled miRCURY LNA miRNA probe (Exiqon, Woburn, MA, USA) for 2 h at 55 °C. Three percent H2O2 was used to inactivate endogenous peroxidases. After incubation with polyclonal anti-digoxigenin antibody and horseradish peroxidase-conjugated secondary antibody (Ventana), a tyramine-conjugated fluorochrome (TSA) reaction was performed for 12 min. Sequential TSA rounds were performed for the detection of proteins using the same protocol. Slides were mounted with antifading ProLong Gold Solution (Life Technologies).

41Paraffin-embedded sections (5 µm) of tumor tissues were cut and used for detection of FABP4. Formalin-fixed sections were deparaffinized by sequential washings with xylene, 100% ethanol, 95% ethanol, 80% ethanol, and PBS. After antigen retrieval, endogenous peroxidase was used to block the slides. Nonspecific binding was prevented by incubating the slides with 4% fish gelatin. This was followed by incubation with primary antibody overnight at 4 °C. The next day, the slides were washed with PBS and incubated with suitable secondary antibody (goat anti-rabbit horseradish peroxidase antibody diluted in protein block solution [JB4]) for 1 h at room temperature. The slides were then washed with PBS followed by development with 3,3′-diaminobenzidine. The nuclei were stained with Gill’s hematoxylin solution and the slides were then mounted.

42For florescence staining, optimal cutting temperature-embedded frozen tissue samples were used. Nonspecific protein blocking was achieved using 4% fish-gelatin solution in tris-buffered saline with tween-20. Slides were incubated with primary antibody (CA9 or FABP4) overnight at 4 °C. For immunofluorescence, Alexa 594 or Alexa 499 were used for secondary antibody staining. Nuclear staining was performed using Hoechst 33342 (Molecular probes).

43Total RNA was extracted from the tumor tissues using the Direct-Zol RNA extraction kit (Zymo Research, Irvine, CA, USA). RNA was then quantified using a NanoDrop spectrophotometer method and the 260 nm/280 nm ratios were checked to determine quality. RNA (1 µg per sample) was reverse-transcribed into cDNA using the Verso cDNA kit (Thermo Scientific, West Palm Beach, FL, USA) according to the manufacturer’s protocol.

44Quantitative real-time PCR was performed on an ABI7500 PCR system (Applied Biosystems, Warrington, UK) using 1 µl of cDNA for each sample. SYBR green (Applied Biosystems) was used to detect the products, and 20 pmol of primer was used for the reaction. All reactions were carried out with 20 µl of reaction mix and performed in triplicate. We used the following primers: for FABP4, 5′-TGATGATCATGTTAGGTTTGGC-3′ (forward) and 5′-TGGAAACTTGTCTCCAGTGAA-3′ (reverse); for 18 S, 5′CGCCGCTAGAGGTGAAATTC3′ (forward) and 5′TTGGCAAATGCTTTCGCTC3′ (reverse). The following conditions were used for PCR: 50 °C for 2 min, then 95 °C for 15 min, followed by 40 cycles at 95 °C for 1 min each. All reactions were analyzed using the ABI7500 Applied Biosystems PCR software (v.2.0.5). The cycle threshold values of the target genes were initially normalized to the cycle threshold values of 18 S rRNA, and melt curves were checked to determine the specificity of the reactions.

45For miRNA quantifications, Taqman miRNA assays (Life Technologies) were used and reverse-transcription real-time PCR was performed, according to the manufacturer’s instructions. RNU6B was used as a housekeeping gene.

46TCGA mRNA microarray (Agilent 244 K Custom Gene Expression G4502A-07, Affymetrix Human Genome U133A 2.0 Array, Affymetrix Human Exon 1.0 ST Array) and RNASeqv2 level 3 and clinical data were retrieved from Broad GDAC Firehose http://gdac.broadinstitute.org/. Putative copy-numbers from GISTIC were retrieved from cbio portal (http://www.cbioportal.org/). To find the relationship between FABP4 expression and copy number, we first employed a Shapiro–Wilk test and verified that the data do not follow a normal distribution. The nonparametric Kruskal–Wallis test was applied and no relationship between FABP4 and copy-number could be established. A box-and-whisker plot (Box plot represents first (lower bound) and third (upper bound) quartiles, whiskers represent 1.5 times the interquartile range) was used to visualize the data (log2(x)).

47The Spearman’s rank-order correlation test was applied to measure the strength of the association between FABP4 and the genes from the Winter et al. hypoxia metagene. We imposed a cut-off of functional relevance on the Spearman correlation coefficient in absolute value of 0.2 based on the method published previously.

48Protein lysates were isolated from HeyA8 MDR cells, 72 h after transfecting them with control siRNA or FABP4 siRNA. Modified RIPA buffer containing proteinase and phosphatase inhibitor was used for the lysis. Samples were then denatured by 1% SDS with β-mercaptoethanol and were analyzed by MD Anderson Core facility. The heat map was created using normalized log2 median centered values and the data was normalized using the average of the control values. The heat map was generated using http://bioinformatics.mdanderson.org/testchm/.

49The metabolomics data and the gene expression data (normalized log values) were taken from Zand et al., 2016–. The gene expression was analyzed using qRT-PCR and the metabolites data was analyzed using ultrahigh performance liquid chromatography/tandem mass spectrometry (UHLC/MS/MS2) in those studies. This information was obtained from database. High grade serous ovarian cancer tumor samples (n = 61) were analyzed. The samples were divided into high or low FABP4 groups based on the median expression of FABP4 in all samples. Similarly, a median value for each metabolite was calculated for high or low FABP4 samples. The comparison is presented as antilog values.

50DESI-MS imaging was conducted as described previously. A 2D Omni Spray (Prosolia Inc., Indianapolis, IN) coupled to an LTQ-Orbitrap Elite mass spectrometer (Thermo Scientific, San Jose, CA) was used for tissue imaging. Recently, DESI-MSI has been used to characterize the metabolic profiles of ovarian high-grade serous cancer, borderline serous tumors and normal tissues, and allowed identification of potential predictive markers of ovarian cancer aggressiveness. DESI-MS imaging was performed in the negative and positive ion mode from m/z 100–1500, using a hybrid LTQ-Orbitrap mass spectrometer which allows for tandem MS experiments, high mass accuracy (<5 ppm mass error), and high mass resolution (240,000 resolving power) measurements. The spatial resolution of the imaging experiments was 200 µm. Spatially accurate ion images were assembled using BioMap and MSiReader software. The histologically compatible solvent system dimethylformamide:acetonitrile (DMF:ACN) 1:1 (v/v) was used for negative ion mode analysis, at a flow rate of 1.2 µl min−1. For positive ion mode analysis, pure ACN was used, at a flow rate of 3 µl min−1. The N2 pressure was set to 185 psi. For ion identification, high mass resolution/accuracy measurements using the same tissue sections analyzed were conducted. Tandem MS analyses were performed using both the Orbitrap and the linear ion trap for mass analysis.

51The same tissue sections analyzed by DESI-MS imaging were subjected afterwards to standard H&E staining protocol. Light microscopy images of the H&E stained slides were taken using the EVOS FL Auto Cell Imaging System (Invitrogen, Thermo Fisher Scientific, Waltham, MA, USA).

52Survival analysis was performed in R (version 3.2.5) using Tothill data set. In order to visualize survival (overall or progression free) differences between groups of patients based on FABP4 levels, we used the log-rank test to find the cut-off with the most significant (lowest p-value) split into high and low mRNA level groups. Kaplan–Meier plots were generated based on these cutoffs. The median survival time of each group is presented in brackets. The numbers of patients at risk in low and high mRNA groups at different time points are presented at the bottom of the graph. For overall survival, the cut-off to optimally separate the patients in the high versus low (min p-value) group was 0.54 for both probes associated to FABP4. For progression free survival, the cut-off to optimally separate the patients in the high versus low (min p-value) group was 0.55 for 203980_at and 0.61 for 235978_at.

53The relationship between overall survival, respectively progression free survival and covariates (mRNA expression levels and clinical parameters (age and stage)) was examined using a Cox proportional hazard model. A multivariate Cox proportional hazard model was fitted, including the clinical parameters and mRNA expression significant in the univariate analysis.

54The drug was dissolved in DMSO and physiologically achievable concentrations were used for the experiments. HeyA8 MDR cells were plated in 24 well plates and serum starved before tamoxifen treatment. The cells were collected, and RNA was extracted for qRT-PCR.

55The experiment was conducted as per the manufacturer’s instructions (Free Fatty Acid Uptake Assay Kit (Fluorometric, ab176768) Cambridge, MA, USA). The cells were serum starved and treated with tamoxifen at a desired concentration. Then the fatty acid dye loading solution was added (100 µl per well of a 96 well plate). The florescence signal was measured at excitation per emission 485 per 515 nm using a bottom read mode after 30 min incubation at room temperature.

56MS data corresponding to the areas of interest were extracted using MSiReader software. The m per z range was discretized by performing hierarchical clustering and cutting the resulting dendrogram at distance 0.05. Peaks appearing in more than 10% of the pixels were kept for analysis. Logistic regression was performed with Lasso regularization using the “glmnet” package (26) in the R language. Regularization parameters were determined by 3-fold cross-validation analysis. The data were randomly equally divided into training and validation sets of samples, 50-50 per patient basis. The training set was used to build a model by cross-validation (CV) based on assigned statistical weights to specific lipid and metabolites important for classification of samples as having either high- or low-FABP4 expression. For other assays, Student’s t-test was performed to examine the difference between the control and treatment groups. A P value less than 0.05 was deemed statistically significant. All the statistical tests were two-sided. Error bars represent standard error from triplicates.

57RPPA data is provided in supplementary data. All other remaining data are available within the article and supplementary files.