Kinome‐wide transcriptional profiling of uveal melanoma reveals new vulnerabilities to targeted therapeutics

Metastatic uveal melanoma (UM) is invariably fatal, usually within a year of diagnosis. There are currently no effective therapies, and clinical studies employing kinase inhibitors have so far demonstrated limited success. This is despite common activating mutations in GNAQ/11 genes, which trigger signalling pathways that might predispose tumours to a variety of targeted drugs. In this study, we have profiled kinome expression network dynamics in various human ocular melanomas. We uncovered a shared transcriptional profile in human primary UM samples and across a variety of experimental cell‐based models. The poor overall response of UM cells to FDA‐approved kinase inhibitors contrasted with much higher sensitivity to the bromodomain inhibitor JQ1, a broad transcriptional repressor. Mechanistically, we identified a repressed FOXM1‐dependent kinase subnetwork in JQ1‐exposed cells that contained multiple cell cycle‐regulated protein kinases. Consistently, we demonstrated vulnerability of UM cells to inhibitors of mitotic protein kinases within this network, including the investigational PLK1 inhibitor BI6727. We conclude that analysis of kinome‐wide signalling network dynamics has the potential to reveal actionable drug targets and inhibitors of potential therapeutic benefit for UM patients.

2015). It is also markedly different to UM genetically, being characterized by activating BRAF mutations in ~35% of cases (Larsen et al., 2016), and an absence of GNAQ or GNA11 mutations.
Metastatic CoM containing BRAF mutations has been treated with some success with BRAF kinase inhibitors (Kao, Afshar, Bloomer, & Damato, 2016), although primary or secondary resistance is inevitable, and very few other treatment options are available. In metastatic UM, progress has been slower still (Carvajal et al., 2017), with kinase inhibitors showing very limited clinical efficacy, and no clear improvement in overall survival in phase II trials of the multikinase inhibitor sunitinib or the MEK inhibitor selumetinib (Blum, Yang, Komatsubara, & Carvajal, 2016;Carvajal et al., 2014). The therapeutic potential of modulating signalling pathways that lie downstream of constitutively active GNAQ/11, such as AKT, PKC, mTOR and the Hippo-YAP pathway is currently being evaluated (Carita et al., 2016;Feng et al., 2014;Shoushtari & Carvajal, 2016;Yu et al., 2014). In addition, targeting cell cycle-regulated protein kinases remains an attractive approach for therapy in a wide range of solid tumours (Otto & Sicinski, 2017), although, to our knowledge, this avenue has not been explored experimentally for metastatic UM.
A complementary (or alternative) therapeutic strategy in UM involves the repression of pro-proliferative transcriptional networks driving growth and survival in tumour cells. Mechanistically, this involves drug-mediated disruption of the interaction between bromodomaincontaining proteins, such as BRD4, and acetylated lysine residues on chromatin-bound histones. Bromodomain extra-terminal inhibitors (BETi), such as the MYC-gene targeting compound JQ1, compete for the acetyl-binding pocket of transcription factors such as BRD4, and block: (i) transcription of genes associated with G1/M progression (Dey, Nishiyama, Karpova, McNally, & Ozato, 2009); (ii) the recruitment of P-TEFb (Bisgrove, Mahmoudi, Henklein, & Verdin, 2007); (iii) the expression of the master oncogene MYC (Delmore et al., 2011); and d) activity of the BRD family themselves (Filippakopoulos et al., 2010). JQ1 is a potent cellular BETi with high selectivity for BET domains of BRD-related transcription factors, particularly bromodomains 1 and 2 of BRD4 (K d values ~50 nM and ~90 nM, respectively). Subsequently, the cytotoxic effects of JQ1 have been reported in a wide number of solid (Bandopadhayay et al., 2014;Cheng et al., 2013;Delmore et al., 2011) and haematological malignancies (Mertz et al., 2011;Ott et al., 2012). JQ1 exhibits extensive repressive effects on proliferative gene expression, and anticancer effects are thought to be achieved, in large part, by silencing MYC-dependent genes, such as the AP-1 transcription factor FOSL1 (Baker et al., 2015), and the pro-proliferative Forkhead family member FOXM1 (Zhang et al., 2016). In both GNAQ/11-mutant UM cell lines and a xenografted murine model, JQ1 arrests cell cycle progression in G1 phase in vitro, whilst apoptosis in UM is associated with the repression of Rad51 and Bcl-xL-dependent signalling pathways (Ambrosini, Sawle, Musi, & Schwartz, 2015). More recently, dual kinase-bromodomain inhibitors exemplified by the Polo-like kinase (PLK) inhibitors BI-2536 (Scutt et al., 2009;Steegmaier et al., 2007) and BI6727 (Rudolph et al., 2009) have been reported, raising the possibility that these rationally designed dual inhibitors might have widespread therapeutic potential (Ciceri et al., 2014).
In this study, we employ kinome-wide transcriptomics to discover a new mRNA expression signature that is shared between human pigmented uveal and conjunctival melanomas, irrespective of GNAQ/11 or BRAF mutational status and chromosome 3 copy number. This kinome "fingerprint," comprising very low levels of EGFR and high levels of CDK2 mRNAs, is also present in archived patient UM samples and in public data available from The Cancer Genome Atlas (TCGA) database, where four molecular and clinical UM subclasses have recently been recognized (Robertson et al., 2017). We also demonstrate that the cytotoxic BETi JQ1 strongly represses a FOXM1-dependent kinase transcriptional network containing cell cycle-related protein kinases in UM cells. Consistently, UM cell lines exposed to investigational antimitotic kinase inhibitors were cytotoxic, and selective UM susceptibility to the phase III PLK1 inhibitor BI6727/volasertib was observed. We conclude that kinome-wide analysis of transcription has the potential to expose vulnerabilities in signalling networks for evaluation as novel therapeutic targets in UM patients.

| Discovery of a kinome signature in UM and CoM cells
Given the overall low mutational burden in UM, epigenetic mechanisms have recently been proposed to shape the gene transcription networks that regulate UM metastasis (Decatur et al., 2016;Royer-Bertrand et al., 2016). To help understand steady-state and dynamic programmes of transcription in human UM kinomes, we explored mRNA levels in patient samples deposited in TCGA. Principal component analysis (PCA) (Ringner, 2008)  kinome expression in all UM samples, which was distinct from other types of tumour sample examined. Interestingly, although UM were most similar to cutaneous melanomas, a clear distinction in kinome profile was still evident between these two tumour types.
Human UM cell lines have been widely employed to compare transcription , signalling and drug responses in vitro and ex vivo. Sequencing of UM cell lines demonstrate common patterns of distinct genomic mutations in GNAQ/11 and BAP-1 genes, similar to tumour samples from patients (Table S1). These profiles are proving useful as both diagnostic and prognostic markers (Staby et al., 2017). We profiled four well-studied UM cell lines, including 92.1 (derived from a primary UM), which responds to the BETi JQ1 in vivo when implanted in a rodent tumour model (Ambrosini et al., 2015), Mel270 (derived from a primary UM), OMM2.5 (a liver metastasis from the same patient as Mel270) and OMM1, a subcutaneous metastasis. In addition, the CoM lines CRMM1  and CRMM2  were compared with model (non-ocular) cancer cell lines, including A375, derived from a cutaneous melanoma.
To examine human kinome mRNA expression profiles from experimental UM and CoM cell models, and to identify new potential networks for clinical intervention, we employed NanoString, an unbiased method of mRNA quantification (Kulkarni, 2011). This allowed for the simultaneous measurement of expression dynamics across >500 mRNAs, which together constitute the human kinome, a rich source of drug targets (Bago et al., 2016;Manning, Whyte, Martinez, Hunter, & Sudarsanam, 2002). were derived from the same patient. A second cluster in cell lines 92.1 and OMM1 contained Q209L mutations in the related GNAQ/11 homologues, and a third comprised the related CoM lines CRMM1 and 2, which differ in their BRAF mutational status (Figure 1b).
To deconvolute our whole-kinome data, we focused on kinase mRNAs that exhibited differential high or low expression levels in melanoma. We found extremely low levels of EGFR mRNA in all UM and CoM cell lines ( Figure 1c): it was minimally ~50-fold lower in ocular melanoma cells when compared with carcinoma cell lines and the skin melanoma cell line A375. In marked contrast, CDK2 transcript levels were always in the highest 1-2% of kinase mRNAs in all UM (Fig. S1) and always overexpressed at two-to fourfold higher levels in UM and CoM cell lines (Figure 1c). The mRNA encoding a regulator of MAPK pathways (TAOK3) was also highly expressed in UM and CRMM1 cells when compared to control cells (Fig. S1). As internal controls, we confirmed that BET domain containing BRD2 and BRD4 mRNAs were highly expressed in all cells lines examined, with no positive selection in melanoma cells (Fig. S1). As shown in Figure 1d, and in full agreement with our expression analysis in cells, EGFR mRNA was extremely low, and CDK2 very high, in all 80 human UM examined in the TCGA database. Quantitatively across all the samples, mean expression levels of CDK2 transcripts were ~250-fold higher than those of EGFR ( Figure 1d), revealing the enormous range in transcript levels detectable by this approach.
We next established that EGFR protein levels were essentially undetectable in both UM and CoM cells, in contrast to all other cancer cells tested, including a cutaneous melanoma. Conversely, CDK2 protein levels were markedly higher in the same UM and CoM lysates in all the cases examined (Figure 1e), consistent with our transcriptional data.
Next, we evaluated kinome expression profiles in nine formalinfixed and paraffin-embedded primary UM. Consistently, we observed exceptionally high levels of CDK2 and TAOK3 mRNAs, in contrast to very low expression of EGFR mRNA, and BRD2 and BRD4 mRNA levels were consistently amongst the highest 10% of transcripts examined ( Fig. S2a and b). Interestingly, there was no correlation between CDK2 (high) or EGFR (low) signatures and chromosome 3 status, a known strong prognostic marker in UM (van Beek et al., 2014;Damato et al., 2011), in the samples examined.
Our analysis suggests that kinome-wide mRNA data, and in particular, CDK2 and EGFR status, can be used to classify ocular melanoma cell types in vitro and in vivo, and might form the basis of a predictive network relevant to understanding differential kinase signalling pathways present in UM patients. To examine whether high expression of CDK2 in UM and CoM was predictive for sensitivity to chemical CDK2 inhibition, we exposed UM, CoM and control cancer cell lines derived from non-ocular sites to the preclinical CDK2 inhibitor SNS-032, the subject of a Phase 1 clinical trial (Tong et al., 2010). By comparing cell viability, we observed a broad, non-specific, inhibitory effect (SNS-032 EC 50 60-400 nM) in all cells tested irrespective of CDK2 mRNA or protein levels (Fig. S3a).
Moreover, the clinically approved selective CDK4/6 inhibitor palbociclib (Whittaker, Mallinger, Workman, & Clarke, 2017) exhibited a wide range of toxicity across all UM and CoM cell lines tested, indicating that its efficacy did not correlate with GNAQ/11 mutational status, or uveal compared to conjunctival origin (Fig. S3b).
F I G U R E 2 Assessing the sensitivities of UM to a panel of experimental and FDA-approved kinase and BET domain inhibitors. (a) Mean EC 50 values were obtained using MTT cell viability assays (mean ± SD from triplicate assays, N = 2). The indicated cell lines were incubated with increasing concentrations of the appropriate compound in a range between 10 nM and 10 μM, and compared to 0.1% (v/v) DMSO controls. For ease of analysis, green boxes highlight EC 50 values <1 μM, grey boxes EC 50 values between 1 μM and 10 μM and white boxes EC 50 values >10 μM. ND = not determined (b) 4,500 cells were exposed to increasing concentrations of JQ1, and cell viability was determined using MTT and SRB assays [Colour figure can be viewed at wileyonlinelibrary.com]

| Comparing responses of ocular melanoma cells to kinase and BET domain inhibitors
We next assembled a panel of FDA-approved inhibitors, including specific and pan-tyrosine kinase inhibitors (TKIs) and the UM candidate selumetinib (Komatsubara, Manson, & Carvajal, 2016) and calculated EC 50 values associated with cell killing. As expected, UM and CoM demonstrated limited sensitivity (EC 50 values >1 μM) to a variety of TKIs, including those targeting EGFR, whose expression is extremely low in these cells ( Figure 1c). Interestingly, the MEKi selumetinib exhibited higher cytotoxicity towards UM cells, particularly those derived from primary tumours (92.1 and Mel270), with mean EC 50 values lying in the sub-micromolar range. Of note, CoM cell lines were also relatively sensitive to the non-specific TKI ponatinib, in contrast to all UM cell lines examined. As a positive control, BCR-ABL-positive K562 CML cells exhibited low nanomolar sensitivity to approved clinical TKIs (Figure 2a), whose marked lack of effect in any UM cells was prominent, consistent with experimental clinical findings detailing a lack of imatinib response in KIT-positive UM (Calipel et al., 2014).
The resistance of tumour cells to kinase inhibitors can sometimes be overcome by exposing them to cellular BET domain inhibitors such as JQ1 (Kurimchak et al., 2016;Stuhlmiller et al., 2015). We therefore turned our attention to chemical targeting of bromodomain (BRD) proteins, which are known to regulate expression of oncogenic MYC (Delmore et al., 2011) that is frequently upregulated in UM. Chromosome 8q amplification, which encompasses the oncogenic MYC locus, is consistently observed in ~30% of UM patients (Caines et al., 2015;Parrella, Caballero, Sidransky, & Merbs, 2001). Patients with UM characterized by monosomy 3 and amplified 8q have a statistically significantly higher risk of developing metastatic disease (van Beek et al., 2014;Damato et al., 2011).
This risk is enhanced further should additional poor prognostic factors be present, such as ciliary body location, epithelioid cell type, high mitotic count and extraocular extension (Damato et al., 2011).
A recent study concluded that GNAQ/11 mutations conferred single-agent sensitivity to JQ1, which potently represses MYC transcription in GNAQ/11-mutated UM (Ambrosini et al., 2015). These observations suggested that MYC and/or GNAQ/11 mutational status may be important for the cellular response to broad transcriptional repressors such as JQ1. The very high levels of BRD2 and BRD4 mRNA observed in both UM and CoM cell lines and patient samples (Figs S1 and S2) prompted us to test the effect of the BET domain inhibitor JQ1 and JQ1(−), an inactive enantiomer, across UM, CoM and non-ocular cell types. As shown in Figure 2a and b, all UM-derived cells tested were sensitive to JQ1, but not JQ1 (-), exhibiting EC 50 values <500 nM. In addition, we found that the CoM cell lines CRMM1 and CRMM2 (WT at GNAQ/11 loci, Table   S1) were also sensitive to JQ1-induced cell death (Figure 2a and b). Interestingly, A375 and colorectal HCT116 cells were relatively resistant to JQ1, with calculated EC 50 values of >10 μM and 2.4 μM, respectively, suggesting a potential therapeutic window in UM and CoM for small molecules that interfere with signalling pathways controlled by BRD4-dependent transcription.

| Bioinformatic and pathway analysis of kinase signalling pathways in UM
Very little is known about dynamic intracellular signalling networks in UM that might contribute to the resistance or sensitization mechanisms to drugs established in other cancer cell models (Kurimchak et al., 2016;Stuhlmiller et al., 2015). For example, JQ1 has been demonstrated to modulate transcription in proliferative signalling pathways and decrease tumour volume in a mouse xenograft GNAQ mutant model, which coincides with downregulation of the anti-apoptotic F I G U R E 3 Quantification of JQ1-induced changes in human kinome mRNA levels in UM and CoM cell lines. (a) Kinome mRNAs were quantified using NanoString technology, following JQ1 exposure and compared to DMSO-treated controls. The number of kinases whose expression changed by more than twofold was counted and converted to a percentage of the 522 kinases analysed. Kinases were discarded from the analysis if JQ1(−) caused a similar response (although this was only documented in two cases). Kinases with increased or decreased expression relative to untreated controls are shown on the positive or negative axis respectively. (b) Ingenuity pathway analysis of the UM and CoM cell lines exposed to JQ1 for 24 hr. Red-or blue-shaded cells indicate pathways or functional terms that were predicted to be activated or repressed respectively, according to the expression changes observed following JQ1 exposure. Ingenuity pathway analysis provides both a p-value for enrichment of a category as well as a measure of how consistent the fold changes are with increased/decreased activity of that category (z-score). The least significant enrichment amongst the categories shown had a p-value of <.0001 [Colour figure can be viewed at wileyonlinelibrary.com] protein Bcl-XL (Ambrosini et al., 2015). As shown in Figure 3a, JQ1 exposure caused a highly significant (>2 fold) downregulation of ~ 100 kinase mRNAs (representing between 17 and 24% of the total kinome) in all four UM cell lines examined. In contrast, CRMM1/2 cell responses were characterized by up-and downregulation of relatively few kinase mRNAs. Interestingly, the whole-kinome response of A375 cells resembled the UM kinome response following exposure to JQ1, despite the limited cytotoxicity demonstrated by JQ1 in this cell line.
We next evaluated expression dynamics in the context of cancer signalling pathways, which has the potential to reveal novel singleagent or combinatorial opportunities for therapeutic intervention, which might avoid unwanted "off-target" effects of very broad transcriptional repressors such as JQ1. Ingenuity pathway analysis (IPA) confirmed that JQ1 effects were predicted to reduce signalling activity within a variety of proliferative/pro-survival pathways in UM, but not CoM, cells (Figure 3b). These include protein kinases that are specifically associated with metastatic HGF/c-Met, GαQ, NF-κB and melanocyte-specific signalling pathways. In addition, and consistent with the ability of JQ1 to reduce cell viability ( Figure 2) and promote apoptosis (Ambrosini et al., 2015), we noted a prominent number of kinase mRNAs associated with cell death signalling in each of the UM cell lines. In contrast, pathway responses were reversed in JQ1exposed CoM cells CRMM1 and CRMM2 (Figure 3b), despite an equal cellular sensitivity to JQ1 in vitro ( Figure 2a). As noted above, IPA analysis of A375 cell response following JQ1 exposure revealed close pathway similarities to those induced in UM cells, despite a lack of cellular toxicity to this compound.

| JQ1 reduces the expression of BRD4 and C-MYC in UM cells
We  (Hermeking et al., 2000), which is involved in G1 progression, were also reduced following JQ1 exposure (Fig. S4a). Despite BRD4 engagement in both ocular and cutaneous melanoma cell types, no correlation between target engagement and subsequent cell death could be established for JQ1. This finding is in line with a recent report (Ambrosini et al., 2015), where F I G U R E 4 Confirmation of cellular JQ1 target engagement by transcriptomics. (a) Quantified mRNA counts (mean ± SD) of BRD family members after incubation with DMSO (blue bars) or 500 nM JQ1 (red bars) or 2 μM JQ1 (green bars). Statistical significance between JQ1-treated cells and the controls was assessed using paired t tests. *p < .05, **p < .01, ***p < .001. (b) Endogenous MYC protein was detected by Western blot of lysates exposed to DMSO (control)

| JQ1 downregulates a FOXM1-dependent network of protein kinase mRNAs
To evaluate new potentially druggable responsive signalling pathways involved in JQ1-induced cell death that might be of relevance to inform UM patient therapeutics, we used the reconstruction of transcriptional networks (RTN) pipeline to evaluate signalling modules (Fletcher et al., 2013). Kinome-wide mRNA expression data were as- promoter is a known transcriptional target for the cancer-associated FOXM1 pathway (Chen et al., 2013;Raychaudhuri & Park, 2011).

| Targeting the FOXM1 transcription network with kinase inhibitors
Our finding that JQ1 reduced UM cell viability and downregulated FOXM1-regulated kinases suggested a new approach to target proliferative signalling in UM. We hypothesized that chemical repression of the activity of protein kinases in this network might be employed to mimic the widespread and potentially non-specific effects observed with JQ1. We therefore exposed UM, CoM, A375 and HCT116 cells to chemically distinct inhibitors of PLK1, including GSK461364 (Ferrarotto et al., 2016), BI6727/volasertib (Van den Bossche et al., 2016), the dual PLK1-BRD4 inhibitor BI2536 (Mross et al., 2012) and the investigational Aurora B inhibitor (AZD1152, barasertib (Wilkinson et al., 2007)). Cytotoxicity and cell viability were determined by combined SRB and MTT assay (Figure 6a and both exhibited hypersensitive to vemurafenib. These findings were confirmed by clonogenic assays, which showed specific effects of F I G U R E 6 UM cell lines are highly sensitive to the PLK1 inhibitor BI6727. (a) Cells were exposed to increasing concentrations of the stated compounds. Mean EC 50 values (nM) were obtained from two independent cell viability assays, SRB and MTT, each performed in triplicate. (b) Colonies formed from single cells were stained with methylene blue following 7 days exposure to increasing concentrations of BI6727 ranging between 1 nM and 30 nM. DMSO was employed as a mock compound control [Colour figure can be viewed at wileyonlinelibrary.com] BI6727 on UM cells (Figure 6b). Finally, when kinome-wide transcriptional responses to BI6727 in the UM cell line 92.1 were compared with the BETi JQ1 (Fig. S6b), clear mechanistic differences between the two compounds became apparent. For example, BI6727 had no effect on BRD4 mRNA levels or members of the FOXM1-regulated mRNA network that are downregulated by JQ1. This finding is consistent with recent data suggesting that BI6727 only binds to the BET domain of BRD4 in cells at concentrations two to three orders higher than the EC 50 value established here (Gjertsen & Schoffski, 2015).
The mechanism of BI6727 cytotoxicity in UM cell lines, therefore, appears to be distinct from that of JQ1 and the dual-PLK1/BETi BI2536, which we propose might underlie the window of selectivity identified in UM cells.

| DISCUSSION
In this study, we undertook a transcriptomic and chemical biology approach to evaluate the human kinome in two rare melanomas that arise either in the uveal tract (UM) or conjunctival membranes (CoM) of the eye. These melanomas receive less attention than metastatic cutaneous melanomas, whose mutational status has been exploited clinically with targeted BRAF kinase inhibitors. We focused on increasing our understanding of cellular protein kinase networks, as they represent important (but currently ineffective) druggable modules in metastatic UM downstream of mutated Gα-subunits of GNAQ/ GNA11. The importance of an unbiased understanding of kinome dynamics to guide therapeutic options is made clear by a recent dissection of responses to targeted kinase therapies in breast cancer models (Bago et al., 2016) and made relevant to UM by the recent discovery of pathway-specific kinase inhibitors in human preclinical models (Carita et al., 2016).
A multiplexed approach to simultaneously quantify mRNA expression of the entire human kinome in samples from four UM and two CoM cell lines facilitated the identification of a kinome mRNA signature that was specific to UM and CoM. Several kinase mRNAs are highlighted by our analysis: EGFR (low expression) and CDK2 and TAOK3 (high expression). In agreement with these findings, EGFR levels have previously been reported to be low or undetectable in immortalized UM cell lines (Amaro et al., 2013). In contrast, Amaro and colleagues reported high levels of EGFR in the cell lines "Mel285" and "Mel290". However, these cells are highly unusual, and do not display the typical antigens associated with melanoma (e.g., MelanA) (Amaro et al., 2013;van Dinten et al., 2005;Griewank et al., 2012). The clinical or signalling relevance of high basal mRNA levels of CDK2 and TAOK3, the latter potentially lying downstream of GNAQ/11 as a transducer of MAPK pathway signalling in UM, remains unknown, although clear co-segregation of these mRNAs as part of a unique UM signature ( Figure 1) might be useful for future diagnostic applications.
CDK2 protein kinase activity is modulated by cancer-associated Cyclin E and p27Kip regulatory partners and is classically associated with G1/S phase progression. In melanocytes, CDK2 levels are under the control of melanogenesis-controlled transcription factor (MITF) (Du et al., 2004), and we speculate that MITF is also responsible for the high levels of CDK2 observed in ocular melanomas (Figure 1) as previously established in distinct UM models (Yan et al., 2012). Consistent with a general proliferative role for CDK2, all UM and CoM cells studied exhibited sensitivity to the CDK2 inhibitor SNS032, although effects were not nearly as marked as those measured with Aurora B or PLK1 inhibitors. During the course of our studies, we confirmed that selumetinib (a MEKi currently undergoing phase III clinical trials in patients with metastatic UM) was modestly potent at killing cultured UM cells, including two cell lines derived from metastatic sites (Figure 2a).
However, no clear beneficial patient outcome has been observed following a phase II clinical trial of the TKi sunitinib (SUAVE) (Blum et al., 2016;Sacco et al., 2013) or evaluation of selumetinib alongside the alkylating agent temozolomide, where 97% of patients treated with selumetinib suffered adverse effects, with 37% requiring reductions in dosage, with little improvement in progression-free survival and none in overall survival reported (Carvajal et al., 2014).
Although JQ1 has not yet been tested as a targeted clinical agent in UM, this BETi exhibits very broad anticancer effects and has previously demonstrated inhibition of tumour growth in an animal model (Ambrosini et al., 2015), consistent with high levels of transcriptional repression revealed in our UM study (Figures 3 and 5). We extend these published findings by demonstrating that CoM cell models also exhibit sub-micromolar sensitivity to JQ1, despite GNAQ/11-wt status and the presence of a BRAF mutation in CRMM1, which sensitizes these cells to the V600E-selective BRAF inhibitors PLX4720 (Riechardt et al., 2015) and vemurafenib (PLX4032) (Figure 6a). When enrichment for mRNAs in specific canonical pathways and their associated cellular functions were analysed, the likely functional consequences on cell survival and proliferative mechanisms through kinome reprogramming by JQ1 in UM and CoM cells became apparent ( Figure 3b). Interestingly, although each of these ocular melanoma models were sensitive to similar concentrations of JQ1 in vitro (Figure 2), the "downstream" transcriptional effects on cell cycle, proliferation and apoptotic kinase networks were essentially reversed in CoM cells. These findings suggest that the specific mechanism of cell killing by JQ1 may vary between melanomas of different cellular origin, rather than simply as a function of GNAQ/11 status (Ambrosini et al., 2015).

| A JQ1-responsive transcriptional network that influences UM cell survival
The presence of a JQ1-regulated FOXM1 transcriptional network in model UM cells was further evaluated to reveal a similar strong correlation in UM patient samples between absolute levels of FOXM1 mRNA and that of the same family of proliferative protein kinases. CDK1, PLK1, Aurora A and Aurora B are individually ratelimiting for mitotic entry, satisfaction of components of the spindle assembly checkpoint and mitotic exit (Nigg, 2001;Otto & Sicinski, 2017). In agreement with our analysis (Figure 5), BRD4 siRNA phenocopies JQ1 exposure in UM (Ambrosini et al., 2015) and BRD4 siRNA potently suppresses Aurora B mRNA and protein (You et al., 2009). JQ1 has recently been shown to control a transcriptional FOXM1 network in ovarian cancer cells (Z. Zhang et al., 2016) and also controls expression of cell cycle genes such as murine PLK1 (Matzuk et al., 2012)

| Relevance of kinome profiling to future clinical strategies in UM
Our analysis raises the possibility that comparative kinome profiling of patient-derived samples could also be employed as a prognostic or predictive tool to help patient profiling in pigmented tumour types, including many UM and CoM. In order to help progress the treatment of metastatic UM, the potential exists for transcriptional profiling to be rapidly taken up as part of experimental and clinical protocols (Barlesi et al., 2016;Mashima et al., 2015;Park et al., 2016).

| Cell proliferation, clonogenic assay and cell viability assays
Cells were seeded in a 96-well plate at a density of 4,500 cells/well and left to adhere for 24 hr. Cells were exposed in triplicate to the indicated compound concentration, dissolved in a final concentration of 0.1% (v/v) DMSO. Cell viability and cytotoxicity were determined after 5 days by MTT assay or SRB assay, or after 7 days after fixation and staining in a 1% (w/v) methylene blue, 80% (v/v) methanol solution. Thiazolyl blue tetrazolium bromide (Sigma-Aldrich, St-Louis, Mo, USA) was dissolved in PBS at a concentration of 5 mg/ml and 5 μl in 100 μl media added to the cells for 3 hr. Reactions were stopped by addition of 50 μl acidified 10% SDS. Absorbance was measured at 570 nm. Cell viability was determined as the percentage of growth compared to control cells (incubated in 0.1% (v/v) DMSO alone), and EC 50 values were calculated using GraphPad Prism 6 software using data from three, or occasionally two, independent experiments, each performed in triplicate. Values are mean ± SD.

| NanoString nCounter analysis, siRNA and quantitative RT-PCR
Samples were prepared essentially as described previously (Bago et al., 2016). Briefly, the appropriate cells were collected, counted, washed in PBS and snap-frozen until required. Pellets were thawed and resuspended at a concentration of ~6,500 cells/μl in RLT buffer (Qiagen). The equivalent of ~10,000 cells (~1.5 μl) was employed for direct kinase mRNA quantification, without the need for further mRNA purification or amplification. mRNAs were hybridized to NanoString human kinome barcode probes and control code sets, and kinase mRNA levels quantified using the nCounter colour barcoding system after count normalization with internal housekeeping genes and eight negative controls. For total copy numbers, replicate normalized count values were averaged. p-values were calculated using a paired t test (GraphPad Prism software) by comparing treated and control count data for each mRNA. The complete UM kinome data sets are available as Supporting information data sets online. siRNA was performed using RNAiMAX reagents (Life Technologies) and Dharmacon ON-TARGETplus siRNA pools corresponding to human FOXM1 (ID:2395) or human BRD4 (ID:23476) in Mel270 and A375 cells. Complete cDNA was generated from total RNA using GoScript Reverse Transcription system (Promega), using 1 μg RNA per reaction and 0.5 μg of Random primer. qPCR was performed in triplicate using the Comparative C t (ΔΔC t ) method on an Applied Biosystems (AB) StepOnePlus machine, a Power SYBR Green PCR Master Mix and the following primer pairs. PLK1: 5′-AAGAGGAGGAAAGCCCTGAC-3′ and 5′-TTCTTCCTCTCCCCGTCATA-3′, FOXM1: 5′-TGCAGCTAG GGATGTGAATCTTC-3′ and 5′-GGAGCCCAGTCCATCAGAACT-3′, BRD4: 5′-CCATGGACATGAGCACAATC-3′ and 5′TGGAGAACA TCAATCGGACA-3′, GAPDH: 5′-TTCACCACCATGGAGAAGGC-3′ and 5′-CCCTTTTGGCTCCACCCT-3′. Expression levels were normalized to GAPDH mRNA using AB software.
Protein concentration was determined using a Pierce BCA Protein Assay Kit, and 30 μg of each sample was denatured in 1× Laemmli sample buffer, heated at 95°C for 5 min and then analysed by SDS-PAGE using 10% (v/v) polyacrylamide gels. Standard Western blotting procedures were followed. Membranes were incubated with the following primary antibodies MYC (D84C12), cyclin-D1 (#2922), FOXM1 (D12D5), PLK1 (Cell Signaling Technologies) and alpha tubulin (TAT-1, CRUK) overnight. Antibodies were detected with appropriate HRPconjugated secondary antibodies using ECL.

| Computational and pathway analyses
Computational analysis including PCA and data visualization was performed using the R software package for statistical computing (https://www.r-project.org/) and commercial Ingenuity Pathway Analysis software (Qiagen Biosciences).
The RTN pipeline was carried out in R using the RTN package with default parameters. This pipeline is able to reconstruct genomewide transcriptional networks from microarray or RNA-seq data using an information-based theoretical approach, which it uses to search for network hubs whose first neighbours are enriched with genes provided  Zhang et al., 2015). The network shown in Figure 5a represents a consensus of the analyses using gene signatures from the four UM cell lines.