Novel prognostication biomarker adipophilin reveals a metabolic shift in uveal melanoma and new therapeutic opportunities

Metastatic uveal melanoma remains incurable at present. We previously demonstrated that loss of BAP1 gene expression in tumour cells triggers molecular mechanisms of immunosuppression in the tumour microenvironment (TME) of metastatic uveal melanoma. Adipophilin is a structural protein of lipid droplets involved in fat storage within mammalian cells, and its expression has been identified in uveal melanoma. We comprehensively evaluated adipophilin expression at the RNA (PLIN2) and protein levels of 80 patients of the GDC‐TCGA‐UM study and in a local cohort of 43 primary uveal melanoma samples respectively. PLIN2 expression is a survival prognosticator biomarker in uveal melanoma. Loss of adipophilin expression is significantly associated with monosomy 3 status and nuclear BAP1 losses in uveal melanoma tumours. Integrative transcriptomic and secretome studies show a relationship between transient loss of adipophilin expression and increased levels of tumour‐associated macrophages and hypoxia genes, suggesting PLIN2‐dependent changes in oxygen and lipid metabolism in the TME of low and high‐metastatic risk uveal melanoma. We designed four adipophilin‐based multigene signatures for uveal melanoma prognostication using a transcriptomic and secretome survival‐functional network approach. Adipophilin‐based multigene signatures were validated in BAP1‐positive and BAP1‐negative uveal melanoma cell lines using a next‐generation RNA sequencing approach. We identified existing small molecules, mostly adrenergic, retinoid, and glucocorticoid receptor agonists, MEK, and RAF inhibitors, with the potential to reverse this multigene signature expression in uveal melanoma. Some of these molecules were able to impact tumour cell viability, and carvedilol, an adrenergic receptor antagonist, restored PLIN2 levels, mimicking the expression of normoxia/lipid storage signatures and reversing the expression of hypoxia/lipolysis signatures in co‐cultures of uveal melanoma cells with human macrophages. These findings open up a new research line for understanding the lipid metabolic regulation of immune responses, with implications for therapeutic innovation in uveal melanoma. © 2023 The Authors. The Journal of Pathology published by John Wiley & Sons Ltd on behalf of The Pathological Society of Great Britain and Ireland.


INTRODUCTION
Uveal melanoma (UM) is the most aggressive intraocular tumour in adults. About 50% of UM patients will develop metastatic disease, typically in the liver, for which no treatments are available [1]. Immune checkpoint therapies achieved successful survival rates for metastatic cutaneous melanoma, although metastatic UM patients are unfortunately universally refractory to these treatments [2]. Thus, a better understanding of UM biology associated with existing risk factors can lead to the development of more effective treatments.
We have recently demonstrated that loss of the BRCA1-associated protein 1 gene (BAP1) in UM, regardless of the mutational load of this gene, is a core genetic transformation leading to immunosuppression of UM tumour microenvironment (TME), mainly coordinated by tumour-associated macrophages (TAMs) [3,4]. Among the various TAM subsets [5], lipidassociated TAMs induce immunosuppression by canonical M2-like pathways using fatty acid (FA) metabolism [5] and are correlated with a worse survival outcome [6]. The impact of BAP1 loss on cellular metabolism and immune functionshas been described [7,8], suggesting that BAP1 loss may modulate the immune TME of UM by governing metabolic routes that impact TAM biology. Indeed, reprogramming metabolism has been recognised as one of the hallmarks of the cancer [9], encompassing several mechanisms, including upregulated lipid metabolism and FA oxidation in the TME [10].
It has been suggested that BAP1-dependent control of lipid metabolism is associated with the upregulation of lipogenic pathways [11][12][13]. In primary UM (pUM), modified lipid metabolism was initially suggested after observing differences in the lipid droplet (LD) content in their TME using a LD marker called adipophilin (adipose differentiation-related protein, encoded by the gene PLIN2) [13]. Adipophilin is a critical structural protein responsible for developing, stabilising, and modifying LD. Adipophilin regulates cellular lipid metabolism by inhibiting LD dissociation and lipolysis [14][15][16][17], acquiring an important role in several metabolic diseases, including cancer [18].
Developing tailored metabolism-targeted approaches for BAP1-deficient UM is a current clinical need [19]. In this study, digital analysis of whole-slide images of pUM combined with their respective clinical data provided an expanded understanding of adipophilin-dependent lipid metabolism and its impact on patient survival. The study extends to a multi-omic analysis of UM metabolism using publicly available data from The Cancer Genome Atlas (TCGA)-UM study [20] and our local Liverpool Ocular Oncology Research Group (LOORG, Liverpool, UK) cohorts [3,21,22] to dissect the implications of adipophilin/PLIN2 biology in UM metabolism. We identified new prognostic metabolic signatures for UM and specific small molecules currently in preclinical and clinical development that have the potential to reverse the expression profile of these signatures using an in silico approach for future functional evaluation. Understanding the metabolic heterogeneity of UM will help the optimal development of novel metabolism-targeting approaches for BAP1-negative secondary UM aiming to impact BAP1-induced changes in TME and patient survival.

Ex vivo co-culture assay
Mel285 or MP46 cells were seeded in direct contact with human primary macrophages at a 1:4 ratio (8 Â 10 4 cancer cells and 3.2 Â 10 5 macrophages) in RPMI medium supplemented with 10% NBCS in a six-well plate, as described elsewhere [24]. Co-cultures were treated with half of the maximal inhibitory concentration (IC50) of carvedilol obtained from the viability assay. After a co-culture period of 24 h, all cells were harvested, and the RNA was isolated using the RNeasy Mini Kit (Qiagen, Germantown, MD, USA).

Reverse transcription-quantitative PCR (RT-qPCR)
Total RNA from co-cultures using human UM cells and macrophages was used for reverse transcription. cDNA was prepared from 1 μg RNA per sample, and qPCR was performed using iTaq Universal SYBR Green Supermix (Bio-Rad, Hercules, CA, USA) and the CFX96 Touch Real-Time PCR Detection system (Bio-Rad). The primer sequences used to amplify the adipophilin-related transcriptomic signatures are listed in supplementary material, Table S4. Relative expression levels were normalised to GAPDH expression as previously described [25].

Next-generation RNA-sequencing
Total RNA from Mel285 and MP46 UM cells lines was isolated as mentioned earlier. The RNA quality and integrity were assessed, and the RNA integrity number (RIN) of the samples used in this study was ≥9. Total RNA (100 ng) was used for the Illumina Stranded mRNA Library preparation (Illumina, San Diego, CA, USA). RNA-seq was performed by the Finnish Functional Genomics Centre (FFGC, Turku Bioscience, Finland) using an Illumina NovaSeq platform. A paired-end 2 Â 100 bp strategy was followed for each sample. The sequencing depth was 41 and 47 M reads per Mel285 and MP46 sample respectively. The quality of the sequenced reads was examined using FastQC (version 0.11.9) tool. Kallisto (version 0.48.0) was used to quantify the transcript per million (TPM) values using the human reference genome (GRCh38) [26]. The foldchange between the TPM values of the two cell lines was calculated using the following formula, and the heatmap was generated using the R pheatmap package: GeneTPM Mel285 -GeneTPM MP46 /GeneTPM MP46 .

Sample selection and preparation
pUM (n = 43) enucleation specimens were examined. The cases were selected according to tissue availability within the Liverpool Ocular Oncology Biobank with informed consent from the patients, who underwent the surgical procedure between 2016 and 2018 at the Liverpool Ocular Oncology Centre (LOOC) and Ethics Committee approval for the study (REC Ref 15/SC/0611). All cases were studied using conventional and immunohistochemical methods, as described previously [27], from the formalin-fixed paraffinembedded (FFPE) blocks. Pseudo-anonymised clinical and genetic information was available (Table 1). Adipophilin protein levels were quantified by immunohistochemistry using an automated machine learning strategy, as previously described [25]. All samples were assessed for nuclear BAP1 expression (nBAP1) using the mouse anti-human BAP1 antibody (0.5 μg/ml; Catalogue No. sc-28383, clone: C-4, Santa Cruz, Insight Biotechnology Ltd, Middlesex, UK) and for adipophilin using rabbit anti-human adipophilin polyclonal antibody (Catalogue No. LS-B4850; 1:500; Lifespan Biosciences, Seattle, WA, USA) and further detected using the 3-Amino-9-ethyl carbazole peroxidase substrate (AEC) or 3,3'-diaminobenzidine (DAB) chromogens, as described previously [27]. Digital slide scanning was performed using an Aperio slide scanner (Aperio CS2) at Â20 magnification. Image files were extracted using Aperio Imagescope software.

Genetics and clinical information
The 43 pUM enucleations were from 24 male and 19 female patients. The median age of patients at diagnosis was 71 years (range . The tumours had a mean largest basal diameter (LBD) of 15.0 mm and a mean ultrasound height of 8.3 mm. Nineteen cases (38%) had ciliary body involvement, and 13 (26%) had extraocular spread. Epithelioid cells were present in 42% UM (Table 1). Multiplex ligation-dependent probe amplification (MLPA) and microsatellite analysis (MSA) provided the information on chromosomal alterations for all tumours in this study [28]. Thirty two pUM were classified as M3 (64%), 11 were classified as disomy 3 (D3) (22%), and 6 had partial loss of chromosome 3 (12%) ( Table 1).

Machine learning digital analysis of full scanned images using NIS-Elements
The NIS-Elements Advanced Software (V4.20.23, Nikon, Melville, NY, USA) was used to customise the quantification of LD objects within digitally scanned slides as previously described [25,[29][30][31]. In brief, regions of interest (ROI) were applied manually to select the tumour areas. ROIs with tissue deformation were excluded based on size, circularity, and signal intensity (chromogen interference, empty spaces). The automated measurement tool generated binary regions based on colour Hue, Intensity, and Saturation (HIS), or the combination of red, green, and blue (RGB) colours. Varying levels of endogenous brown melanin were present in some UM samples and were subtracted from unstained serial section control images to those cases stained with DAB chromogen. The areas of the total tumour ROI and the binary adipophilin stained zones were measured in pixels squared (px 2 ) and further converted to frequency, assuming ROI (whole tumour) to be 100% of the total studied area.

GDC-TCGA data analysis
Publicly available 80 pUM RNA-seq and clinical data [20] were extracted from the TCGA-UM study [20]

UM-Secretome analysis and GDC-TCGA-UM studies
UM-secretome data from UM samples were reviewed from a previous LOORG cohort [21]. The matrix of proteins with significant fold-change expression scores identified in the UM secretome is associated with the patient's risk of developing metastasis: High Risk (3), Low Risk (2) and no risk (normal, 1). The risk associated with each protein (protein Clinical Score) is calculated by the difference between the Highest and Lowest mean condition of the protein expressed in the cohort, resulting in four possible scores: no risk (À2), low risk (À1), mid risk (+1) and high risk (+2) (supplementary material, Table S1). Proteins were sorted accordingly to their fold-change expression levels in the secreted media. Further, the name of the proteins was converted into gene codes and uploaded in the GDC-TCGA-UM cohort-linked with XenaBrowser web-based analysis tool [3] to generate the cluster expression profile of these genes, supervised by PLIN2. The Kaplan Meier survival analysis of each gene was recovered from the GDC-TCGA-UM data bank. The log-rank, p values, and survival profile (good and poor survival prognosis) were included in a matrix (supplementary material, Table S3). Genes matching clinical scores from LOORG secretome data and TCGA-UM data were used for network plasticity analysis.

Tissue immunofluorescence
Antigen retrieval was performed on sections of FFPE tissues (n = 12 pUM) using the Dako PT Link system (Dako, Agilent Technologies LDA UK Limited,

Network and transcriptomic metabolic plasticity analysis
The genes related to normoxia, hypoxia, lipid metabolism, and lipolysis supervised by BAP1 expression levels were individually analysed as a prognosticator marker in Kaplan-Meier curves at the Xena browser. The genes were marked as having good and poor prognoses based on their differential expression. The network plots of good and poor prognosis genes were generated using the circlize package in R [32]. The genes were associated with the hypoxia and lipid categories were collected from the literature and NanoString PanCan Immune Oncology multigene panel, as described previously [3]. The width of the links towards the hypoxia and lipid categories represents the variance of the gene in all the samples, which is calculated from the gene expression score obtained from the GDC-TCGA-UM cohort as previously described [32]. The links are coloured based on the log-rank p-value of the Kaplan-Meier curve.

L1000FWD drug prediction studies
The L1000FWD web-based application allows the query of gene expression signatures against signatures created from human cell lines treated with over 20,000 small molecules and drugs for the LINCS project [33]. The normalised RNA-seq data of identified multi-gene signatures supervised by PLIN2 ( Figure 4C and supplementary material, Table S3) was obtained from TCGA using the UCSC Xena Browser as previously described [3]. PLIN2-supervised multi-gene signatures were uploaded into the L1000FWD platform [33]. The drug signatures profile is shown in a scaffold network distribution where red clusters represent groups of drugs tested in different human cells with the potential to mimic the multi-gene signature that drives the metabolic shift in UM. Blue clusters represent the group of drugs with the highest potential to reverse the gene expression in the query.

Statistical analysis
Graphpad prism 5.0 software was used for statistical analysis. The Shapiro-Wilk normality test was performed for all group samples. Unpaired two-tailed Mann-Whitney U-test (non-parametric) was used to evaluate a significant difference between CD163 + adipophilin + cells and nBAP1 stained groups. Non-parametric two-tailed Spearman's correlation studies between two variables were used. Xena-browser was used to generate all Kaplan-Meier survival plots. The log-rank test was used in UCSC Xena to evaluate the significance of survival curves providing χ 2 and p values. Differences were considered significant when *p < 0.05, **p < 0.01, ***p < 0.001 and ****p < 0.0001.

Adipophilin predicts survival and clinical outcome in pUM
Previous findings by our group reported the expression of adipophilin in a pUM cohort (Fiorentzis cohort, n = 28). We grouped the pathological estimation of adipophilin expression (intensity Â proportion) of the Fiorentzis cohort into disomy 3 and monosomy 3 group status and observed that there is a decrease, but not significant, in adipophilin expression in monosomy 3 pUM compared with disomy 3 in this group cohort (supplementary material, Figure S1A). To evaluate if adipophilin expression is significantly reduced in monosomy 3 phenotype in a larger cohort, we initially revisited the GDC-TCGA-UM transcriptomic data for adipophilin gene expression (PLIN2). We examined the impact of differential expression of PLIN2 on UM patient survival and PLIN2 expression chromosome 3 copy number, BAP1 levels, and age distribution. We observed that lower expression of PLIN2 is significantly associated with poor UM patient survival ( Figure 1A), monosomy 3 status and BAP1 loss (patients with BAP1 levels lower than the median of the gene expression) ( Figure 1B). Interestingly, we observed that the older the patient, the lower the expression levels of PLIN2 ( Figure 1C). These clinical and genetic associations to adipophilin in the transcriptomic level (PLIN2) are in accordance with the ones observed in Fiorentzis cohort [13], showing for the first time a significant correlation between PLIN2 downregulation with reduced survival, as well as its association with BAP1 loss and monosomy 3 status and age. Then, we evaluated adipophilin expression at the protein level in a second larger local UM cohort (LOORG, n = 43), including clinical data (Table 1). This second cohort was previously well-phenotyped for nBAP1 expression using immunohistochemistry and chromosome 3 copy number variation. Primary FFPE sections stained for adipophilin by immunohistochemistry were fully scanned, and adipophilin levels were measured by applying a high-precision machine learning method based on the circularity and size of particles of interest, as previously described [23,25]. Representative images of adipophilin developed with AEC red chromogen are shown, including NIS-Elements particle identification threshold from the entire tumour area are shown (blue for nBAP1 negative and red for nBAP1 positive tumours) ( Figure 1D). Similar to the Fiorentzis and the GDC-TCGA-UM cohorts, adipophilin levels were Adipophilin as a metabolic biomarker for uveal melanoma 207 reduced in the group of patients with monosomy 3 and nBAP1 negative status ( Figure 1E). Interestingly, we also found a negative correlation between adipophilin expression and patient age (supplementary material, Figure S1B). Whether this correlation is due to the tumour metabolic evolution or a normal ageing phenotype needs further evaluation in a larger RNA-seq cohort of normal eye tissues with similar age ranges, something that is currently not available. However, a preliminary analysis using the RNA-seq of the NIH Eye Integration cohort (n = 8) [34] shows that PLIN2 levels are significantly higher in the retina of fetal eyes compared with the retina of adult eyes (supplementary material, Figure S1C), suggesting that the decrease of PLIN2 expression might be affected by ageing, and may be a contributing factor to an increased cancer risk with advancing age [35].
The fact that this clinical phenotype is associated with adipophilin in both RNA and protein levels from two independent cohorts (LOORG and GDC-TCGA-UM) suggests that adipophilin levels may be directly controlled by PLIN2 differential expression in UM. We did not observe a correlation between adipophilin expression and copy number variations in chromosomes 6p, 6q, 8p, and 8q (supplementary material, Figure S2A-D), nor with patients' tumour ciliary body extension (CBI), cell type, the occurrence of extraocular extension (EOE), gender, necrosis profile, tumour height, tumour LBD, and mitotic count (supplementary material, Figure S2E-L).

208
M Matareed, E Maranou et al contributing factor driving tumour immunosuppression in a TAM-dependent manner [3] suggest that BAP1 loss, adipophilin (and, thus, lipid metabolism), and TAM levels are linked variables. First, we evaluated whether CD163 as a macrophage gene marker could be associated with poor survival outcomes in the GDC-TCGA-UM cohort. We observed that the higher the expression of CD163, the lower the survival of UM patients (Figure 2A). Lipid consumption is one terminal metabolic feature of CD163 + M2-like TAMs [5,36]. It is an alternative metabolic pathway during oxygen deprivation in cancer, mainly regulated by the expression of hypoxia-inducible factors (HIFs) such as HIF-1α (gene HIF1A) [37]. We observed that the lower the levels of BAP1 in the GDC-TCGA-UM cohort, the higher the levels of CD163 and HIF1A and the lower the expression of PLIN2 lipid storage marker (red) ( Figure 2B). Indeed, LDs containing adipophilin appear to be Adipophilin as a metabolic biomarker for uveal melanoma 209 diffusely spread within the cytoplasm of nuclear BAP1-negative UM samples, suggesting potential lysis of the LD following increased levels of CD163 + macrophages ( Figure 2C,D). These findings suggest a relationship between loss of BAP1 expression and metabolic changes from normoxia toward lipid consumption during hypoxic conditions in high-risk UM, mainly featured by the decrease in available LDs containing adipophilin ( Figure 2E).

Increased protein secretion reveals new metabolic prognostication signatures in pUM
To evaluate whether the metastatic risk of UM was associated with a decrease in lipid storage markers and an increase in lipid consumption (lipolysis) pathways in high-risk UM, we revisited the secretome profile of a local LOORG cohort of pUM samples (n = 14) previously cultured ex vivo [21]. The secretome included secreted proteins with a clinical score associated with hypoxia and lipid metabolism regulation (supplementary material, Tables S1 and S2). We observed that the higher the secretory activity (measured by the average of protein fold-change expression), the higher the metastatic risk score of these proteins, suggesting an increase in metabolic activity in high-risk UM ( Figure 3A). Each secretome protein's risk distribution profile is associated with a clinical score, which ranges from À2 to 2, where À2 means no risk (blue) and +2 means high risk (red), shown in a heatmap sorted by the total expression level of individual proteins detected in the secretome. When proteins are sorted according to their expression levels, there is an evident shift from low-to high-risk clinical scores ( Figure 3B). The corresponding gene symbols of these proteins were further uploaded into Xena browser to interrogate the transcript levels for these proteins in the GDC-TCGA-UM cohort, which was further sorted according to BAP1 transcript levels. Surprisingly, the transcript profile for these proteins clustered into two main groups: upregulated genes following BAP1 expression (predictors of good outcome) and upregulated genes following BAP1 loss (predictors of poor outcome). Predictors of good outcomes are associated with survival (white), and predictors of poor outcomes are associated with death (black) ( Figure 3C). To validate the risk associated with the proteins identified in the secretome and identify those that could be part of a metabolic regulatory process in UM, we interrogated the impact on survival outcomes of these proteins in the GDC-TCGA-UM cohort (n = 80). All proteins and genes associated with poor outcomes in both cohorts were filtered for downstream analysis (supplementary material, Table S3). We performed a meta-analysis study on the functionality of these genes for their regulatory roles in the following metabolism subcategories: normoxia, hypoxia, lipid storage, and lipolysis. The category ranks were linked with the variance σ 2 of each gene expression within the GDC-TCGA-UM cohort to measure the plasticity of these genes (supplementary material, Table S2) and with the p value of Kaplan-Meier overall survival curves of each gene, to measure its impact on patient survival (supplementary material, Table S3). These variables were then uploaded in a network-category distribution cord plot, sorted primarily by survival impact ( p value) and variance scores (σ 2 , cord thickness) within each immune category. Genes that significantly impact the survival of UM patients with different statistical levels have cords coloured sky blue, orange, and red.
Interestingly, we observed that multigene predictors of good survival outcome in the cord plots clustered to categories of normoxia and lipid storage as one signature ( Figure 3D, left), as opposed to multigene predictors of poor survival that clustered to categories of hypoxia and lipolysis, as a second signature ( Figure 3D, right). The normoxia/lipid storage and hypoxia/lipolysis transcriptomic signatures were further evaluated between two UM cell lines with nBAP1 positive (Mel285) and negative (MP46) protein expression. We observed that some of the genes from both signatures produced the same differential expression profile of BAP1 positive and BAP1 negative UM samples from bulk RNA-seq data of the GDC-TCGA-UM cohort. For example, SNTB2 appears upregulated in pUM tumours with BAP1 loss as one of the poor outcome predictor signatures ( Figure 3C,D). Still, SNTB2 expression appears downregulated in the MP46 (nBAP1-negative) cell line compared with the Mel285 (nBAP1-positive) cell line ( Figure 3E, right panel). Such differences might be explained by the fact that some of the identified signatures are differentially expressed by other cellular components of the TME rather than the tumour cells. Indeed, we observed that SNTB2 and FN1 genes were upregulated in co-cultures of the MP46 cell line with macrophages differentiated from monocytes of human peripheral blood (supplementary material, Figure S3A), suggesting that BAP1 loss by tumour cells may trigger molecular processes that induce the expression of these signatures in stromal cells, such as TMAs. Therefore, 65.5% of upregulated normoxia/lipid storage signature can be represented by tumour cells, 34.4% can be represented by the TME in BAP1-positive tumours ( Figure 3E, left doughnut plot), 33.3% of upregulated hypoxia/lipolysis signature can be represented by tumour cells, and the TME can represent 66.7% of the signature in BAP1negative tumours ( Figure 3E, right doughnut plot). Representative Kaplan-Meier survival curves for identified genes from both metabolic signatures are SPP1 and LAMA4 (good survival predictors) and HTRA1 and SNTB2 (poor survival predictors) ( Figure 3F). These findings suggest that genes negatively correlated with BAP1 and PLIN2 levels in UM simultaneously associate with high protein metabolic index shifting from normoxia/lipid storage to hypoxia/lipolysis conditions.

Adipophilin-associated metabolic changes reveal novel therapeutic opportunities for UM
Discovering new adipophilin-based multigene metabolic signatures can establish the basis for developing  Adipophilin as a metabolic biomarker for uveal melanoma 211  novel tailored therapeutic approaches for targeting metabolism in BAP1-deficient UM [19]. Identifying repurposed drugs with the potential to reverse UM metabolic evolution could be of great value for UM therapy. The differential expression metabolic signatures that significantly impact UM survival and have high expression variance in the GDC-TCGA-UM cohort are represented in the heatmap of Figure 4A. These signatures were uploaded into the L1000FWD web-based platform to predict small molecules with the ability to reverse hypoxia/lipolysis expression profile toward normoxia/lipid storage. The drug signature profile is shown in a scaffold distribution plot, where red clusters represent groups of drugs tested in different human cells with the potential to mimic the multigene signature that drives the metabolic shift in UM. The blue clusters (c1, c2, c3, and c4) represent the group of drugs with the highest potential to mimic the metabolic transcriptome of BAP1-positive tumours or reverse the metabolic shift in BAP1-negative UM ( Figure 4B). Interestingly, drugs observed in these clusters tend to have a similar mechanism of action, which is mainly represented by adrenergic receptor agonism (c1), Pi3K, CDK and tubulin inhibitors (c2), MEK and RAF inhibitors (c3), and retinoid receptor agonism (c4) and all identified drugs that are mainly launched or in Phase 2 or 3 of clinical development ( Figure 4C). The top five drugs from each cluster, their mechanism of action, and their current clinical stage are presented in Table 2.
Therefore, we evaluated whether subdoses of carvedilol could reverse the expression of poor-outcome-predicting signatures or mimic/enhance the expression of good-outcome-predicting signatures in co-cultures of the BAP1negative MP46 cell line with human macrophages obtained from peripheral blood. A representative picture of UM cells co-cultured with human macrophages is shown in Figure 4E. As predicted, carvedilol restored PLIN2 levels, upregulated LAMA4 (normoxia/lipid storage signatures), and downregulated FN1 and SNBT2 (hypoxia/lipolysis signatures) ( Figure 4F and supplementary material, Figure S3C). These findings suggest that the predicted drugs have the potential to impact UM cell viability, restore the expression of PLIN2 in BAP1-negative tumours, and reverse the expression of metabolic signatures associated with poor survival outcomes.

DISCUSSION
Adipophilin has emerged as a prognostic biomarker in different cancer types, including colorectal, lung, breast, ovarian, clear cell renal cell carcinoma, and cutaneous malignant melanoma [40][41][42][43][44][45]. UM prognostication for predicting metastatic risk and patient management, involves the combination of clinical, morphological, Adipophilin as a metabolic biomarker for uveal melanoma 213 and genetic analyses [46,47]. Current genetic prognostic markers and biochemical pathways correlated to UM metastasis were recently reviewed [48]. In this study, we underscored the prognostic potential of adipophilin in UM for the first time, uncovering a network of adipophilin-dependent metabolic changes associated with BAP1 loss and clinical outcomes. Understanding UM-specific immunosuppression driven by BAP1 loss has been advanced by nCounter transcriptome sequencing of primary and metastatic UM samples [3]. Changes in BAP1 expression have recently been linked to significant effects on energy metabolism in UM [11][12][13]. Our findings provide insights into the role of adipophilin in lipid storage and lipid metabolism, where adipophilin levels are lost in the context of oxygen deprivation during hypoxia, promoting a metabolic immunosuppressive TME that favours UM progression and increased metastatic risk. When binding to LDs, adipophilin is stabilised and functions as a gatekeeper of LD, attenuating lipolytic metabolism [16,[49][50][51]. However, competition between adipophilin and other proteins of the PAT (perilipin, ADRP, TIP47) family leads to rapid degradation of adipophilin in LDs. Indeed, M2-like macrophages are associated with immunosuppression and tumour progression in UM with BAP1 loss [3,22]. Our co-culture approach using UM cell lines with human macrophages supports this relationship, where the expression of poor-outcome predictor signature SNTB2 increases in the context of BAP1 loss. Increased levels of CD163 + TAMs in the context of reduced adipophilin-positive LD may be associated with lipid catabolism favouring M2 polarisation and proliferation. As specialised phagocytic cells, M2 macrophages uptake lipids from both engulfed dying cells or LD in the TME, which are then processed by acid lipases within the lysosomes, leading to the generation of free FA that is subsequently transported into mitochondria for ATP generation under hypoxia conditions, reviewed in [52]. Also, insufficient oxygenation in the 'inflamed' TME induces the upregulation of HIF-1 [53], reprogramming TME metabolism into oxygenindependent mechanisms for energy generation such as de novo lipogenesis [54][55][56]. High levels of FA during hypoxia, in combination with PLIN2 downregulation, lead to liver lipotoxicity [57,58]. Indeed, the inability to store FA in LDs also leads to CD8+ T-cell exhaustion [59], while Tregs enhance their FA uptake and become resistant to lipotoxicity [60,61]. We propose that low levels of PLIN2 under hypoxia conditions are involved in impaired lipolysis control. Consequently, TAMs within the UM TME acquire a pro-tumorigenic phenotype facilitating immune suppression and, eventually, immunotherapy resistance. The regulatory role of BAP1 in lipid metabolism has also been reported in preclinical models [12]. Liver-specific deletion of BAP1 led to reduced levels of specific genes that included PLIN2 and resulted in lipid mismanagement, liver inflammation, and fibrosis [12,62]. Accordingly, our data suggest that adipophilin loss in BAP1-negative UM, which is also associated with poor outcomes, may be one of the contributing factors in making the liver a preferable metastatic site of UM cells.
Metabolic alterations in high-risk UM may be a suitable therapeutic opportunity for this devastating disease, which is universally refractory to immunotherapies [63]. Dissecting the molecular process and network links in metabolic heterogenous UM may provide tailored therapeutic approaches for BAP1-deficient UM [19]. In our drug-discovery screening approach using current identified metabolic UM signatures to query the L1000FWD application [33], we identified specific clusters of drugs predominantly homogenous in terms of their potential to reverse the hypoxia/lipolysis signature profile to normoxia/lipid storage profile in BAP1 negative UM. The main category of drugs includes adrenergic, retinoid, and glucocorticoid receptor agonists, HDAC, Pi3K, MEK, and RAF inhibitors.
Surprisingly, some of these drugs have already been tested in UM melanoma clinical trials, including HDAC inhibitors (NCT02068586, NCT03022565) [64,65], selumetinib, which impacted progression-free survival and response rate but not overall survival in UM [66], and in UM preclinical research, including carvedilol, which was recently described as eliciting anti-tumour responses in UM 3D spheroids [67], and vemurafenib, which induces autophagic cell death in UM cells with V600E in vitro [68].
The cytotoxic potential of five identified drugs was validated in UM cell lines. Bosutinib, a BCR-ABL small molecule inhibitor used for treating chronic myelogenous leukaemia, showed the strongest cytotoxic effect, and carvedilol, an adrenergic receptor antagonist, effectively modulated identified prognostication signatures. Importantly, BCR-ABL has been described as reducing the stability of BAP1 in chronic myeloid leukaemia [39], and the relationship of BCR-ABL with BAP1 in UM remains to be elucidated.
In conclusion, a decrease in adipophilin levels in highrisk UM is associated with a transitional state of normoxia and lipid storage to hypoxia and lipolysis in the TME of pUM. Such features may play a key role in BAP1-induced immunosuppression and metastasis development. Our findings uncover new grounds for developing therapeutic strategies for UM in view of its metabolic evolution associated with BAP1 loss and Author contributions statement MM, EM, SEC and CRF contributed to the conceptualisation of the study. MM, EM, SAK, AM, HK, SEC and CRF contributed to data curation, formal study analysis, study investigation, data visualisation, methodology validation, and wrote the manuscript. EM performed the research, and performed meta-analysis of literature metabolomics. SAK performed in vitro cytotoxicity studies. AM performed computational analysis of transcriptomic datasets, and developed computing methodologies. HK: enabled access to biobank resources and supported with funding acquisition. SEC undertook student supervision, funding acquisition, administration of local project and funding, and contributed to manuscript revision and editing. CRF conducted formal analysis of the entire study, supervised the students and research, contributed to funding acquisition, data validation, revised data visualisation, methodology development, wrote the original draft, performed the revision and editing and administred the project.   Table S1. LOORG pUM (n = 14) secretome protein list, including max fold-change and clinical scores Table S2. Immune metabolism categories with respective reference numbers. GDC-TCGA-UM gene variance levels (in decimal units) Table S3. Kaplan-Meier p value scores of good-and poor-survival predicting genes from GDC-TCGA-UM cohort Table S4. Primer sequences used for qPCR Adipophilin as a metabolic biomarker for uveal melanoma 221