A Multi-Step Immune-Competent Genetically Engineered Mouse Model Reveals Phenotypic Plasticity in Uveal Melanoma

Abstract

Uveal melanoma (UM) is a highly aggressive intraocular malignancy with limited therapeutic options for metastatic disease. Existing transgenic UM mouse models inadequately recapitulate human disease progression, while transplant models lack immune competence for studying the tumor immune microenvironment and therapeutic interventions. To address these limitations, we developed a genetically engineered mouse model incorporating stepwise genetic alterations implicated in human UM progression. Spatiotemporally controlled expression of mutant GNAQ^Q209L from the endogenous locus induced choroidal nevi with limited penetrance. Concomitant BAP1 deletion enhanced nevus formation, while further MYC activation led to fully penetrant intraocular tumors with the potential to disseminate. Single-cell RNA sequencing revealed malignant cells segregated into melanocytic and neural crest-like subpopulations characterized by distinct transcriptional and biosynthetic programs. Trajectory analyses inferred dedifferentiation from the melanocytic toward the neural crest-like state during tumor progression. Comparison to human UM revealed commonalities with highly aggressive class 2 UM, including gene expression signatures and copy number gains affecting genes that map to human chromosome 8q beyond the activated MYC allele, suggesting cooperative effects of multiple drivers in this chromosomal region. The tumor microenvironment featured immunosuppressive macrophage populations and exhausted T cells, closely resembling human UM. This physiologically relevant, immune-competent model provides a platform for investigating UM biology, functionally characterizing candidate driver genes, and developing immune-based therapeutic strategies.

INTRODUCTION

Uveal melanoma (UM) is a rare subtype of melanoma but the most common ocular malignancy (1). Approximately half of UM patients develop metastases, primarily in the liver, resulting in a median overall survival of less than one year (2,3). Current therapeutic options for metastatic UM patients remain highly limited (4), emphasizing the urgent need for a deeper understanding of the genetic and molecular mechanisms driving this disease to develop more effective treatments.

UM has a distinct mutational landscape compared to other melanoma subtypes, characterized by mutually exclusive activating hotspot mutations in GNAQ or GNA11, which encode Gα subunits of heterotrimeric G-proteins (5). Notably, GNAQ/11 mutations are also found in choroidal nevi, yet only about 1 in 8,800 of these nevi progress to UM (6,7), indicating that additional genetic alterations are necessary for malignant transformation. Most UM undergo subsequent nearly mutually exclusive prognostic mutations in BAP1, SF3B1 or EIF1AX, associated with high, intermediate and low metastatic risk, respectively. Additionally, canonical chromosome copy number alterations (CNA) are common, including monosomy 3, gains of 6p and 8q, and losses of 1p, 6q and 8p (8). While very little is understood about the role of these chromosomal aberrations in UM, it has been speculated that the oncogene MYC, located at chromosome 8q24, may play a role in UM progression (9).

UM can be classified into distinct prognostic subtypes based on a 15-gene expression profile (15-GEP) with Class 1 tumors being associated with EIF1AX and SF3B1 mutations and low to intermediate metastatic rate, while Class 2 tumors are associated with BAP1 loss-of-function mutations and a high propensity for metastasis (10). Chromosome 8q gains occur in both Class 1 and Class 2 UM, but the number of extra copies of 8q tends to be greater in Class 2 tumors, correlating with increased metastatic potential (11).

This landscape of genetic alterations provides a foundation for developing genetically engineered mouse models (GEMMs) that can facilitate preclinical UM research and serve as platforms for testing therapeutic strategies. Cre-inducible conditional knock-in alleles of mutant GNAQ (12) and GNA11 (13) have been developed, and when crossed to melanocyte-specific Cre mice, they induce rapid ocular melanoma formation (12,13). These findings establish GNAQ/11 as initiating oncogenes in UM. However, single mutant GNAQ/11 mice develop UM within weeks, in contrast to the slower progression seen in human patients, where cooperating alterations are required. This discrepancy reflects that these models do not fully recapitulate the genetics and biology of human UM, limiting their utility for studying UM progression. Possible reasons for this discrepancy include inappropriate timing of mutant GNAQ/11 activation during embryonic development and non-physiological expression levels from ectopic promoters. Moreover, these models develop undesired cutaneous melanomas, precluding long-term studies aimed at modeling metastatic progression following enucleation of the primary tumor. These limitations underscore the need for more refined GEMMs that better reflect human UM development.

In this study, we report the development of a conditional Gnaq-mutant allele expressed at physiological levels from its endogenous promoter. Restricting mutant Gnaq activation to the eye resulted in choroidal nevi, and concomitant BAP1 deletion enhanced the penetrance of nevus formation. The addition of an activated MYC allele, a candidate oncogene located on chromosome 8q24, led to fully penetrant intraocular tumor formation with features of human UM. Cell lines derived from this triple-mutant model formed tumors in orthotopic transplants and exhibited metastasis. Furthermore, scRNAseq of murine tumors revealed intra-tumoral phenotypic heterogeneity, as well as inferred copy number alterations that correspond to many of the genes lost on chromosome 3 and gained on chromosome 8q in human UM. Cross-species comparison confirmed similarities to human Class 2 UM, reinforcing the relevance of this model for studying disease progression. This model will facilitate preclinical research into the pathogenesis of UM and enable the identification of potential drivers and vulnerabilities.

MATERIALS AND METHODS

Generation of a GNAQ^Q209L (Gnaq^CA) allele

A 9,217bp genomic fragment from the murine Gnaq locus, including 4,491bp upstream and 4,596bp downstream of exon 5, were inserted into pKO2.2 (Addgene plasmid #22676, RRID:Addgene_22676) by InFusion cloning. Exon 5 was then mutated at nucleotide 626 (A→T) of the Gnaq coding sequence by site-directed mutagenesis to create the Q209L hotspot mutation. A minigene cassette was generated by GeneArt (ThermoFisher) and inserted into an NdeI site 649bp upstream of exon 5. This minigene cassette contained the following elements: (i) a Splice Acceptor followed by the partial Gnaq cDNA from exon 5 to exon 7 and two SV40 polyA signals; (ii) a PGK promoter followed by the Blasticidin resistance marker and a BGH polyA signal in the reverse orientation; (iii) FRT sites flanking the PGK promoter and the BGH polyA to enable Flp recombinase-mediated removal of the Blasticidin selection cassette; and (iv) loxP sites flanking the upstream Splice acceptor and the downstream FRT site to enable Cre recombinase-mediated removal of the entire minigene cassette. The final targeting vector was sequence verified by Plasmidsaurus whole plasmid sequencing. V6.5-C10 embryonic stem cells (14) on a mixed C57BL/6 × 129S4/SvJae (RRID:MGI:2159769, RRID:MGI:2164439) background were targeted by homologous recombination in Moffitt’s Gene Targeting Core and selected in 10 μg/ml Blasticidin. Clones were screened by Southern blot using a 5’ external probe (NheI digest), a 3’ external probe (AflII digest), and a Blasticidin internal probe (HindIII digest). The presence of the Q209L mutation was confirmed by Sanger sequencing in Southern blot-positive clones. Clones were injected into Balb/c (RRID:MGI:2161072) blastocysts and transferred into pseudopregnant CD1 females. High contribution chimeras were then bred to C57BL/6 (RRID:MGI:2159769) females to achieve germline transmission.

Animals

To generate a genetically engineered mouse strain harboring conditional alleles for Gnaq, Cas9, and Myc, we performed a multi-generational breeding scheme using the following strains: GNAQ^Q209L (Gnaq^CA) mice were generated in this study as described above, Rosa26-LSL-Cas9-EGFP (Cas9^CKI) mice were purchased from The Jackson Laboratory (#028551, RRID:IMSR_JAX:028551), H11-LSL-Myc^T58A-Luciferase (Myc^CKI) mice were obtained from Elsa Flores. Gnaq^CA were maintained on a mixed C57BL/6 × 129S4/SvJae (RRID:MGI:2159769, RRID:MGI:2164439) background. Cas9^CKI and Myc^CKI strains were maintained on a C57BL/6J (RRID:MGI:2159769) background. Initial crosses were performed to generate double homozygous Gnaq^CA; Cas9^CKI breeders, which were then intercrossed to homozygous Myc^CKI mice to obtain the desired triple mutant animals. Genotyping was performed on genomic DNA extracted from tail biopsies using PCR with allele-specific primers (listed in Supplementary Table S1). Only mice carrying the correct allele combinations were used for experiments, and equal numbers of male and female mice were enrolled in the various experimental cohorts and no sex-specific phenotypic differences were observed. All procedures involving animals were approved by the Institutional Animal Care and Use Committee (IACUC) and conducted in accordance with institutional and national guidelines.

Mouse embryonic fibroblasts

Mouse embryonic fibroblasts (MEFs) were isolated from Gnaq^CA embryos at E13.5 and infected with lentiviral Cre. RNA was extracted from MEFs using TRIzol (Invitrogen) following protocols supplied by the manufacturer. cDNA was generated with PrimeScript RT Master Mix (Takara). Gnaq exon 5 was PCR amplified with allele-specific primers (listed in Supplementary Table S1) using GoTaq Green Master Mix (Promega) and subjected to Sanger sequencing. MEFs isolated from Gnaq^CA; Cas9^CKI mice were infected with pL-CLB. Genomic DNA (gDNA) was isolated from MEFs using a standard Proteinase K digestion and ethanol precipitation method. Bap1 exons 1–3 were amplified by PCR using primers listed in Supplementary Table S1 from gDNA with GoTaq Green Master Mix (Promega) and subjected to Sanger sequencing.

Lentivirus production, concentration, and suprachoroidal administration

pl-CMV-Cre-EF1a-luciferase-U6-sgBap1 (pL-CLB), which expresses Cre recombinase from the CMV promoter, firefly luciferase from the EF1α promoter, and a U6-driven sgRNA targeting Bap1, was generated based on a pLenti construct. The sequence of the sgRNA is listed in Supplementary Table S1. To produce lentivirus supernatants, HEK293T cells were transfected with lentiviral vector and delta8.2 and VSVG helper plasmids at a 9:8:1 ratio using JetPRIME transfection reagent. Virus-containing supernatant was collected at 48 and 72 hours post-transfection, filtered through a 0.45 μm PVDF filter, and stored at 4°C. Supernatants were concentrated using Lenti-X Concentrator (Takara Bio): viral supernatant was mixed with 1/3 volume of Lenti-X reagent, incubated at 4°C overnight, and centrifuged at 1,500 × g for 45 minutes at 4°C. Pellets were resuspended in sterile PBS (50x concentration) and aliquots stored at −80°C. Viral titer was quantified using the Lenti-X qRT-PCR Titration Kit.

Suprachoroidal administration of concentrated lentivirus was performed in adult mice (6–8 weeks old) under isoflurane anesthesia. A 33-gauge beveled Hamilton syringe was inserted tangentially through the sclera to access the suprachoroidal space. 2 μL of concentrated lentivirus (titer ~10⁷-10⁸ TU/mL) was slowly injected over 30–60 seconds. To prevent reflux, the needle was held in place for 30 seconds post-injection before removal. The eye was treated with topical erythromycin ophthalmic ointment post-procedure. Efficient gene delivery was verified by bioluminescence imaging (IVIS) at 3–5 days post-injection.

Orthotopic and metastatic transplant of murine uveal melanoma

Uveal melanomas were harvested from Gnaq^CA; Bap1^CKO; Myc^CKI mice when ocular enlargement was visibly apparent. Mice were euthanized, and eyes were enucleated using fine scissors and forceps under a dissection microscope. After removal, each eye was placed in ice-cold sterile PBS. Under a stereomicroscope, extraocular tissues (conjunctiva, optic nerve), lens, and vitreous were removed. Tumor tissue arising from the choroid and adjacent structures was dissected out with fine-tipped forceps and immediately transferred into the enzyme solution. Enzymatic dissociation was performed using the Tumor Dissociation Kit (Miltenyi Biotec, RRID:SCR_020285) according to the manufacturer’s instructions. Tissues were incubated at 37°C for 30–45 minutes with intermittent gentle agitation. The resulting cell suspension was washed with PBS containing 2% FBS and centrifuged at 300 x g for 5 minutes. Cells were resuspended in sterile PBS, counted using a hemocytometer, and assessed for viability by trypan blue exclusion. Cell viability of >80% was observed after dissociation. All downstream injections were performed immediately following dissociation to minimize ex vivo artifacts.

Orthotopic tumor engraftments were performed in 6–8-week-old NSG mice or mixed C57BL/6 × 129S4/SvJae background mice under isoflurane anesthesia. A 33-gauge beveled Hamilton syringe was inserted tangentially through the sclera to access the suprachoroidal space. 20,000 freshly dissociated uveal melanoma cells in 2 μL of sterile PBS were slowly injected over 30–60 seconds. To prevent reflux, the needle was held in place for 30 seconds post-injection before removal. The eye was treated with topical erythromycin ophthalmic ointment post-procedure. Tumor development was tracked by in vivo imaging, and mouse eyes bearing uveal melanoma tumors were collected at defined time points for histologic assessment.

To assess metastatic potential, 1 million freshly isolated uveal melanoma cells in 100 μL sterile PBS were injected into the lateral tail vein of anesthetized NSG mice using a 30-gauge insulin syringe. Metastasis development was tracked by in vivo imaging and histologic assessment at defined time points. Liver, lung, and other organs were harvested at endpoint for histologic evaluation of metastases.

Experimental liver metastases were generated by intrasplenic injection of uveal melanoma cells as previously described (15). Mice were anesthetized with 3% isoflurane for induction and maintained at 2% via nosecone on a warming pad. Sustained-release buprenorphine (0.1 mg/kg, s.c.) was administered 30 min before surgery. Following shaving and disinfection of the left flank, a ~1.5 cm subcostal incision was made to expose the spleen. A single-cell suspension of uveal melanoma cells (5×10⁵ in 100 μL PBS) was injected slowly (~30–40 s) into the splenic parenchyma using a 30 G needle. Injection was confirmed by transient splenic paling, and pressure was applied for 1 min to prevent bleeding. Immediately after injection, the spleen was removed by cauterizing the splenic ligaments (splenectomy). The peritoneum was closed with absorbable sutures and the skin sealed with tissue glue or wound clips. Topical bupivacaine (0.12%, 50 μL) was applied along the incision, and 150–300 μL sterile saline was administered intraperitoneally for hydration. Mice were maintained on warming pads until recovery and monitored daily for 3 days. Transient weight loss (<10%) within 24 h typically resolved within 48 h. All procedures followed institutional animal care and ethical guidelines.

Immunohistochemistry and Periodic Acid–Schiff (PAS) Staining

Mouse eyes bearing UMs and organs bearing metastasis were harvested and fixed in 10% neutral-buffered formalin for 48–96 hours at room temperature. Embedding, sectioning, and H&E staining were performed by IDEXX BioAnalytics.

For standard Immunohistochemistry (IHC), slides were deparaffinized in xylene and rehydrated through graded ethanol. Antigen retrieval was performed using citrate buffer (pH 6.0) in a microwave oven for 12 minutes. Endogenous peroxidase activity was quenched using 3% hydrogen peroxide for 10 minutes. Sections were blocked in 2.5% normal horse serum for 1 hour at room temperature and incubated overnight at 4°C with primary antibodies: Tyrosinase (1:100, Thermo Fisher, PA5–86066, RRID:AB_2802867), Melan-A (1:100, Thermo Fisher, MA5–13232, RRID:AB_10984369), MYC (1:100, ABclonal A19032, RRID:AB_2862524). Detection was performed using HRP-conjugated secondary antibodies (Vector Labs, RRID:AB_2631198) and DAB substrate (Vector DAB Peroxidase Substrate Kit, RRID:AB_2336520), followed by hematoxylin counterstaining. Slides were dehydrated and mounted with permanent mounting media.

Fluorescent Immunohistochemistry (F-IHC) was performed using antibodies against RPE65 (1:150, GeneTex, GTX13826, RRID:AB_372073) and SOX10 (1:50, Cell Signaling Technology, 78330, RRID:AB_3698115). After antigen retrieval and blocking as above, sections were incubated overnight at 4°C with primary antibodies diluted in 1% BSA in PBS. The next day, slides were washed and incubated with species-appropriate Alexa Fluor-conjugated secondary antibodies (1:1,000, Thermo Fisher, RRID:AB_2535813, RRID:AB_2534073) for 1 hour at room temperature. Slides were coverslipped using ProLong Gold Antifade Mountant with DAPI (Thermo Fisher, P36931) and imaged using a fluorescence microscope (EVOS-Auto, RRID:SCR_026039).

PAS staining was performed using the Periodic Acid–Schiff Stain Kit (Sigma-Aldrich) according to the manufacturer’s instructions. Briefly, paraffin sections were deparaffinized and rehydrated, then oxidized in 0.5% periodic acid for 5 minutes, rinsed, and incubated in Schiff’s reagent for 15 minutes. Sections were counterstained with hematoxylin, dehydrated, and mounted. PAS-positive structures were visualized as magenta.

Multiplex Immunohistochemistry (mIHC)

Mouse primary uveal melanoma and liver metastasis tissues were stained for GLI2 (1:100, Invitrogen, PA1–28838, RRID:AB_2111904), Melan-A (1:100, Abcam, ab210546, RRID:AB_2889292), TRP1 (1:100, Abcam, ab83774, RRID:AB_2211142) , MYC (1:100, ABclonal, A19032, RRID:AB_2862524), and DAPI with an Automated Opal 9-Color IHC Kit and quantified in Moffitt’s Advanced Analytical and Digital Laboratory:

Multiplex Immune Panel Procedure:

Formalin-fixed and paraffin-embedded (FFPE) tissue samples were immunostained using the AKOAYA Biosciences OPAL 9-Color Automation IHC kit (Waltham, MA) on the BOND RX autostainer (Leica Biosystems, Vista, CA, RRID:SCR_025548). The OPAL 9-color kit uses tyramide signal amplification (TSA)-conjugated to individual fluorophores to detect various targets within the multiplex assay. Sections were baked at 65°C for one hour and then transferred to the BOND RX (Leica Biosystems). All subsequent steps (ex., deparaffinization, antigen retrieval) were performed using an automated OPAL IHC procedure (AKOYA). OPAL staining of each antigen occurred as follows: heat induced epitope retrieval was achieved with EDTA pH 9.0 buffer for 20 min at 95°C before the slides were blocked with AKOYA blocking buffer for 10 min. Then slides were incubated with primary antibody at RT for 30 min followed by OPAL HRP polymer and one of the OPAL fluorophores during the final TSA step. Individual antibody complexes are stripped after each round of antigen detection. This was repeated five more times using the remaining antibodies of the panel. After the final stripping step, DAPI counterstain was applied to the multiplexed slide and was removed from BOND RX for coverslipping with ProLong Diamond Antifade Mountant (ThermoFisher Scientific). Autofluorescence slides (negative control) were included, which used primary and secondary antibodies omitting the OPAL fluorophores and DAPI. All slides were imaged with the Phenolmager HT (Akoya Biosciences, RRID:SCR_023772).

Quantitative Image Analysis:

Multi-layer TIFF images were exported from InForm (AKOYA) and loaded into HALO Image Analysis Platform version 4.0 (Indica Labs, New Mexico, RRID:SCR_018350) for quantitative image analysis. For the quantitative fluorescent phenotype analysis, tissues were segmented into individual cells using the DAPI marker. For each marker, a positivity threshold within the nucleus or cytoplasm is determined based on visual intensity and expected staining localization. After setting a positive fluorescent threshold for each staining marker, the entire image set is analyzed with the created algorithm. The generated data includes positive cell counts for each fluorescent marker in cytoplasm or nucleus, and percent of cells positive for the marker.

Western blot analysis

Fifteen micrograms of protein were separated on NuPAGE 4% to 12% precast gels (Thermo Fisher Scientific) and transferred to nitrocellulose membranes. Membranes were blocked in 5% non-fat dry milk in TBST and incubated with one of the following primary antibodies overnight at 4°C: Gαq (1:1000, Cell Signaling Technology, 14373, RRID:AB_2665457), ERK1/2 (1:2,000, Cell Signaling Technology, 4695, RRID:AB_390779), p-ERK (Thr202/Tyr204) (1:1,000, Cell Signaling Technology, 4370, RRID:AB_2315112), YAP (1:1,000, Cell Signaling Technology, 4912, RRID:AB_2218911), p-YAP (Ser127) (1:1,000, Cell Signaling Technology, 4911, RRID:AB_2218913), BAP1 (1:1,000, Abcam, ab255611). Anti-beta-actin (1:3,000, Invitrogen, AM4302, RRID:AB_2536382) was blotted as a loading control. Membranes were washed 3 times with TBST for 10 min, followed by incubation with HRP-conjugated secondary antibodies (1:3,000) for 1 hour at room temperature. After three washes in TBST, chemiluminescence substrate (1:1) was applied to the blot for 4 min and chemiluminescence signal was captured using a LI-COR imaging system (RRID:SCR_023227).

Single-cell sequencing

Single-cell RNA-sequencing was performed using the 10X Genomics Chromium System (10X Genomics, RRID:SCR_025146) by Moffitt’s Molecular Genomics Core. Cell viability and counts were obtained by AO/PI dual fluorescent staining and visualization on the Nexcelom Cellometer K2 (Nexcelom Bioscience LLC, RRID:SCR_018656). Cell viability >80% was used for inclusion in sequencing. Cells were then loaded onto the 10X Genomics Chromium Single Cell Controller to encapsulate approximately 10,000 cells per sample. Single cells, reagents, and 10X Genomics gel beads were encapsulated into individual nanoliter-sized Gelbeads in Emulsion (GEMs), and reverse transcription of polyadenylated mRNA was performed in each droplet at 53°C. The cDNA libraries were then completed in a single bulk reaction by following the 10X Genomics Chromium GEM-X Single Cell 3’ Reagent Kit v4 user guide, and 50,000 sequencing reads per cell were generated on the Illumina NovaSeq6000 instrument (RRID:SCR_016387). Demultiplexing, barcode processing, alignment, and gene counting were performed using the 10X Genomics CellRanger v8 software (RRID:SCR_023221).

Single-cell RNA-seq processing, integration, clustering, and annotation

Single-cell RNA-seq data were processed following a standard 10x Genomics–Seurat workflow, with study-specific parameters as described below and adapted from previously described (16). Illumina BCL files were demultiplexed to FASTQ and processed with Cell Ranger v7.1.0 (10x Genomics; RRID:SCR_023221), aligning reads to the mouse GRCm39 transcriptome using STAR (RRID:SCR_004463) (17). Gene-barcode matrices were imported into Seurat v4.0 (RRID:SCR_016341) (18). Quality control filtering was applied to remove low-quality cells and genes. Cells with fewer than 200 detected genes or with more than10% UMIs derived from mitochondrial were removed. Genes detected in fewer than 3 cells were also removed. After filtering, a total of 24,455 cells from three mice were retained for analysis. For each sample, raw UMI counts were log-normalized (NormalizeData) and the 5,000 most variable genes were identified (vst, FindVariableFeatures). To correct for batch effects across samples, data were integrated using Seurat anchor-based integration (18,19) (FindIntegrationAnchors/IntegrateData) with 8,000 anchors and 40 PCs (CCA-based integration). Integrated data were scaled (ScaleData) regressing out percentage of mitochondrial UMIs, S and G2/M phase scores, and total UMI counts. An SNN graph was built using the first 40 PCs, and Louvain clustering (FindClusters) was performed at resolution of 1.2 (major cell-type analysis; 34 clusters). UMAP embeddings were generated with RunUMAP (default settings). Differential expression was computed with FindAllMarkers (test.use=“wilcox”, logfc.threshold=0.25, min.pct=0.2); genes with Bonferroni-adjusted P<0.05 and avg logFC>0.25 were considered significant. Clusters were annotated using canonical markers: melanoma cells (neural crest-like: Mitf, Mlana, Tyr, Gli2, Meis1, Sox5; melanocytic: Mitf, Mlana, Tyr, Dct, Tyrp1), B cells (Cd79a, Cd79b, Cd22), T/NK cells (Cd3d, Cd3e, Cd3g), monocytes (Ly6c2, Vcan, Plac8), neutrophils (S100a8, S100a9), mregDC (Ccr7, Fscn1, Ccl22, Cacnb3), multiciliated epithelial cells (Crocc2, Foxj1, Deup1, Mcidas), endothelial cells (Pecam1, Cdh5, Tie1), inflammatory CAF (Col1a1, Col1a2, Col5a1, Cxcl5, Il11, Saa3), myofibroblast-like CAF (Col1a1, Col1a2, Col5a1, Actg2, Cd248, Myh11), activated CAF (Col1a1, Col1a2, Col5a1, Adamts2, Col6a3, Tnc), Arg1+ macrophages (Cd68, Mrc1, Arg1, Ccl9, Frt3), C1qa+ macrophages (Cd68, Mrc1, C1qa, C1qb, C1qc, Fcrls), RPE (Rbp1, Rlbp1, Ttr), retinal bipolar cells (Isl1, Neurod4, Pcp2), photoreceptors (Gnat1, Gngt1, Rho), and Müller glia (Aqp4, Gfap, Sox2).

Annotation of melanoma cells

Melanoma cells were annotated as Neural Crest-like (Gli2, Glis3 Meis1, Sox5) and Melanocytic (Mlana, Dct, Tyrp1, Slc24a5) and these two major populations were further annotated into nine subtypes. Neural Crest-like melanoma cells: TRIO-high (Trio, Ngf, Camta1, Arhgap10) and MITF-high (Tcf4, Nfia, Mift, Gsk3b); Melanocytic melanoma cells: Epithelial-like (Prdx1, Krt18, Gpnmb, Atp1b1), Mesenchymal-like (Vim, Pcolce, Mfap4, Acta2), Proliferative (Top2a, Stmn1, Pcna, Mki67), Hypoxia (Ndufa4l2, Higd1a, Egln3, Bnip3), Stem-like (Pdgfc, Igfbp2, Gdf15, Gas6), Vascular-mimicry (Scin, Prtg, Pecam1, Itgb6), and RPE-like (Ttr, Rpe65, Rlbp1, Rdh5, Rbp1). Differential expression analysis was performed to compare two major populations of melanoma cells, as well as the 9 subtypes, using FindMarkers function in Seurat with logfc.threshold=0.25, min.pct=0.2, and test.use=”wilcox”. Genes were ranked based on -log10(p-value)*(sign of log2(fold-change)). Pre-ranked GSEA (RRID:SCR_003199) was performed on gene rankings using R package fgsea (bioRxiv 2021:060012) with 10,000 permutations, against Hallmark, REACTOME (RRID:SCR_003485), and GO databases from MsigDB (RRID:SCR_016863) (20–22). The normalized enrichment scores (NES) of gene sets were visualized using heatmaps. To further characterize the malignant cells, we compared gene expression of 9 subtypes to cell identity signatures previously reported in mouse cutaneous melanoma studies. Enrichment scores of these signatures were calculated for each malignant cell using the AUCell algorithm implemented in SCENIC (RRID:SCR_017247) (23), and the scores were visualized in heatmaps.

Copy number variation patterns in malignant cells were extracted using InferCNV (24) R package v1.20.0 (RRID:SCR_021140). Normal immune cells identified above (B, T/NK, Neutrophils, mregDC and Monocytes) were selected as “reference” cells for de-noise control. InferCNV analysis was performed using “denoise” mode to correct for batch effects from different mice, with tumor_subcluster_partition_method = ‘qnorm’, HMM=TRUE, and analysis_mode = ‘subclusters’. The “cluster by group” parameter was turned off to allow the observation cells to cluster unbiasedly based on CNA patterns. Two major CNA branches were identified in melanoma cells, corresponding to the two main melanoma cell types: neural crest-like melanoma cells and melanocytic melanoma cells.

To compare CNV patterns between mouse and human UM, we mapped CNV results obtained from mouse UM cells as described above to the TCGA uveal melanoma dataset (TCGA-UVM). CNV data for TCGA-UVM were downloaded from cBioPortal (RRID:SCR_014555). HOMDEL and HETLOSS events were defined as loss, and GAIN and AMP events were defined as gain. Mouse genes were mapped to their human orthologs using the Mouse Genome Informatics database (www.informatics.jax.org, RRID:SCR_006460). Genomic locations of human orthologs were retrieved using the biomaRt package in R (RRID:SCR_019214). The CNV status of these human orthologs was then summarized in the TCGA-UVM dataset to evaluate conservation of CNV patterns between species.

RNA velocity analysis was performed to infer the dynamic states of malignant cells. Initially, spliced and unspliced transcript abundances were quantified from BAM files generated from CellRanger count using Velocyto (RRID:SCR_018167) (25). The resulting loom files were merged across samples. Seurat meta data was integrated with the loom file and malignant cells were extracted for downstream analysis. The dynamical model of RNA velocity in Python module scVelo (RRID:SCR_018168) (26) was applied to estimate transcriptional rates and compute velocity vectors for individual cells in “dynamical” mode. Briefly, we used scv.tl.recover_dynamics() to infer gene-specific transcriptional kinetics from spliced and unspliced mRNA counts. Velocity vectors were computed using scv.tl.velocity() in dynamical mode, followed by scv.tl.velocity_graph() to estimate the transition probabilities between cells. Latent time—a continuous pseudotime derived from the velocity field—was computed with scv.tl.latent_time(), representing the internal clock of cells along their differentiation trajectories. The velocity fields were visualized on the batch-corrected UMAP projection, with arrows overlaid to indicate the direction and magnitude of transcriptional changes. To identify genes associated with the latent time progression, malignant cells were grouped into 100 bins based on their estimated latent time. Marker genes specific to each bin were identified using the FindAllMarker function in Seurat (RRID:SCR_016341) and visualized in a heatmap.

SCENIC analysis

Transcription factor motif analysis was performed using pySCENIC (RRID:SCR_025802) (23) on a normalized expression matrix of the 18,901 melanoma cells with default setting. In addition, option --mask_dropouts was used. Using the mouse v10 database (https://resources.aertslab.org/cistarget/databases/mus_musculus/mm10/; https://resources.aertslab.org/cistarget/motif2tf/), 323 transcription factors were identified in 18,901 melanoma cells. The resulting matrix was z-scored using the scale function from base R. Transcription factors with the top 15 highest AUC scores were visualized in a heatmap with hierarchical clustering.

Myc-luciferase transgene sequence search

We performed a BLAT search between the Firefly Luciferase sequence and 90 bp 10X raw reads from all three samples. 10X reads that matched the Firefly Luciferase sequence were defined as those with 100% identity and an alignment length of at least 89 bp. Cells containing “FireflyLuc reads” were extracted and counted how many such reads were associated with each unique cell ID. The presence of any FireflyLuc reads in a given cell ID was recorded as “FireflyLuc Yes”; absence was marked as “FireflyLuc No”. The number of FireflyLuc reads per unique cell ID was recorded as LucReadNum, which was used to represent the expression level of the Myc^T58A-IRES-Luciferase transgene in each cell. We compared gene expression of the 18,901 melanoma cells to the mouse MYC targets signatures in Hallmarks. Enrichment scores of these signatures were calculated for each malignant cell using the AUCell algorithm (RRID:SCR_021327) implemented in SCENIC (RRID:SCR_017247) (23), and the scores were visualized in heatmaps. All this information was visualized by Seurat v4.0 using FeaturePlot (RRID:SCR_016341). Spearman correlation was performed between Myc and other genes to identify genes that have the highest correlation with Myc expression.

Data availability

Sequencing data supporting the findings of this study have been deposited in the Gene Expression Omnibus (GEO) at GSE316353. All other data supporting the findings of this study are available from the corresponding author upon request. Custom scripts used for scRNA-seq, TCGA dataset, and GSEA analysis were written in R (v4.2.0) and are available in GitHub at https://github.com/liuxiaoxian/Uveal_Melamona, or upon request from the corresponding author. No proprietary code was used. The public data analyzed in this study were obtained from Gene Expression Omnibus (GEO) at GSE139829 and TCGA Uveal Melanoma dataset at https://portal.gdc.cancer.gov/projects/TCGA-UVM.

RESULTS

Generation of a conditional, physiologically expressed Gnaq^Q209L allele

To model the genetics and biology of human UM, we generated a conditional Gnaq^Q209L allele using a “mini-gene” strategy. Exon 5 of Gnaq was replaced with a loxP-flanked cDNA cassette containing wildtype exons 5–7. Before Cre recombination, Gnaq transcripts therefore endogenous exons 1–4 followed by cDNA-encoded exons 5–7. Immediately downstream, we inserted a mutant exon harboring the Q209L hotspot mutation. Cre-mediated excision of the mini-gene juxtaposes endogenous exon 4 to mutant exon 5, generating the Gnaq^Q209L transcript. This allele, termed Gnaq^CA, thus expresses wildtype GNAQ prior to Cre activity and mutant GNAQ^Q209L from the endogenous locus after recombination (Fig. 1A).

Figure 1. Generation of a genetically engineered mouse model of choroidal nevi driven by Gnaq mutation and Bap1 loss.

A, Schematic diagram of the Gnaq conditional activation (Gnaq^CA) allele (Created in BioRender. Xu, X. (2026) https://BioRender.com/ccwj1ps). B, Mouse embryonic fibroblasts (MEFs) isolated from Gnaq^CA mice were infected by lentiviral Cre. PCR amplification and Sanger sequencing of Gnaq exon 5 from cDNA confirmed expression of the Q209L mutation. C, Western blot of Cre-infected Gnaq^CA MEFs revealed changes in Gαq downstream signaling but physiological Gnaq expression. D, Schematic diagram of Gnaq conditional activation (Gnaq^CA) in the uvea by in situ lentiviral delivery of Cre and Luciferase (pL-CL) (Created in BioRender. Xu, X. (2026) https://BioRender.com/ccwj1ps). E, Representative in vivo bioluminescence imaging of a Gnaq^CA mouse infected with pL-CL. F, Quantification of frequency of choroidal hyperplasia in Gnaq^CA mice at the indicated timepoints. G, Representative image of H&E staining showing localized choroidal hyperplasia in a Gnaq^CA mouse. H, Representative fluorescent IHC showing Sox10 and Rpe65 expression in uveal tracts of wildtype and Gnaq^CA mice. I, Schematic diagram of Gnaq conditional activation and Bap1 conditional CRISPR knockout (Gnaq^CA; Bap1^CKO) in the uvea by in situ lentiviral delivery of Cre, Luciferase, and sgBap1 (pL-CLB) (Created in BioRender. Xu, X. (2026) https://BioRender.com/ccwj1ps). J, Quantification of choroidal hyperplasia and tumors in Gnaq^CA; Bap1^CKO mice at 7 months. K and L, Representative images of H&E staining of choroidal hyperplasia (K) and micro-tumor (L) in Gnaq^CA; Bap1^CKO mice. M, Representative fluorescent IHC showing Sox10 and Rpe65 expression in the uveal tract of a Gnaq^CA; Bap1^CKO mouse.

To validate physiological expression and Cre responsiveness, we derived mouse embryonic fibroblasts (MEFs) from Gnaq^CA/+ mice and infected them with adenoviral Cre. Sanger sequencing confirmed heterozygous expression of the mutant exon (Fig. 1B). Cre administration induced hyperactivation of Gq-dependent pathways, as shown by increased ERK phosphorylation and reduced YAP phosphorylation, without altering total GNAQ levels (Fig. 1C). Thus, the Gnaq^CA allele enables Cre-inducible, physiological expression of GNAQ^Q209L to model oncogenic GNAQ activation in UM.

Melanocyte-specific activation of GNAQ^Q209L

To investigate the effects of GNAQ^Q209L, we generated Gnaq^CA; Tyr-CreERt2 compound mutant mice for tamoxifen-inducible GNAQ^Q209L induction in melanocytes (27). All Gnaq^CA; Tyr-CreERt2 mice developed extensive skin hyperpigmentation, especially on non-hair bearing skin, and multiple nevi in the absence of Tamoxifen administration by 2–3 months of age (Supplementary Fig. S1A). Several nevi per mouse progressed to rapidly growing pigmented or amelanotic tumors by 4–6 months (Supplementary Fig. S1B,C). Although leakiness of Tyr-CreERt2 has been observed in the Braf^V600E; Pten^FL/FL background, the extent and kinetics of tumorigenesis in Gnaq^CA; Tyr-CreERt2 mice in the absence of cooperating mutations was unexpected. The skin phenotype in Gnaq^CA; Tyr-CreERt2 mice emerged before overt ocular abnormalities were observed. When the skin disease required euthanasia, ocular histology revealed extensive pigmented hyperplasia of the uveal tract (Supplementary Fig. S1D). SOX10 and RPE65 staining confirmed hyperplastic cells were melanocytes rather than retinal pigmented epithelium (Supplementary Fig. S1E). Thus, Gnaq^CA activation robustly induces uveal melanocytic hyperplasia. However, as in prior mutant Gnaq or Gna11 models (12,13), the extensive skin disease limits the utility of melanocyte-specific Cre drivers for studying GNAQ^Q209L-driven uveal melanoma.

Development of a uveal melanoma mouse model

To restrict GNAQ^Q209L expression to the eye and avoid recombination in cutaneous or organ-resident melanocytes, we generated a lentivirus expressing Cre and Luciferase (pL-CL) (Fig. 1D). Suprachoroidal delivery (Fig. 1D) resulted in efficient infection of the uveal tract, confirmed by in vivo bioluminescence (Fig. 1E). Histologic evaluation at 3, 6, and 9 months revealed localized choroidal thickening with increasing penetrance, reaching 33% at 9 months (Fig. 1F,G). Choroidal melanocytic hyperplasia was evident without violation of Bruch’s membrane or effect on the retinal pigment epithelium, and no cutaneous lesions occurred. SOX10 and RPE65 staining confirmed increased melanocyte numbers in choroidal hyperplasia (Fig. 1H). Thus, lentiviral Cre delivery enables spatiotemporal control of GNAQ^Q209L activation and produces a phenotype resembling human choroidal nevi.

To enable rapid introduction of secondary mutations, we crossed Gnaq^CA mice with R26-LSL-Cas9 (28) mice and suprachoroidally delivered the pL-CL virus containing a Bap1-targeting sgRNA (pL-CLB) to induce CRISPR-mediated in situ knockout of Bap1 (Gnaq^CA; Bap1^CKO; Fig. 1I, Supplementary Fig. S1F,G). All Gnaq^CA; Bap1^CKO mice exhibited localized choroidal hyperplasia by 7 months (Fig. 1J,K). While Bap1 deletion increased the penetrance of hyperplasia, it rarely progressed to melanoma, with only one small SOX⁺/RPE65^- choroidal tumor observed (Fig. 1L,M). Thus, concurrent GNAQ^Q209L activation and Bap1 deletion efficiently induces choroidal nevi but drives UM only infrequently, suggesting that additional genetic events are required for robust tumorigenesis in this model.

Progressive chromosome 8q copy number gains occur throughout UM evolution, starting in low-grade Class 1 tumors and increasing in number in Class 2 UM and metastatic tumors (29). Given the location of MYC on chromosome 8q and its overexpression upon 8q gain (30), we tested MYC as a cooperating oncogene in UM formation. Gnaq^CA; R26-LSL-Cas9 mice were crossed to a Cre-dependent stabilized MYC allele (Myc^T58A) (31) and infected suprachoroidally with pL-CLB to generate Gnaq^CA; Bap1^CKO; Myc^CKI mice (Fig. 2A). These triple-mutant mice developed aggressive, fully penetrant UM with a median survival of ~3 months (Fig. 2B). Robust in vivo bioluminescence signal from primary tumors (Fig. 2C) enables longitudinal tumor monitoring and, potentially, tracking of metastasis. At endpoint, tumors filled the vitreous cavity (Fig. 2D), caused significant proptosis and often extended extraocularly (Supplementary Fig. S2A), consistent with locally invasive disease. Tumors grew into the sub-retinal space, displacing the retina and filling the posterior chamber. Tumors grew as largely unpigmented epithelioid sheets with focal pigmentation, occasional nested architecture and areas of necrosis with corresponding variable inflammatory responses. Cells exhibited round-to-oval nuclei with prominent nucleoli and granular cytoplasm, with some tumors displaying areas of partial cytoplasmic clearing (Fig. 2E). IHC confirmed clonal SOX10 positivity, RPE65 negativity, and a Periodic Acid–Schiff-negative pattern (Fig. 2F,G), consistent with UM.

Figure 2. MYC promotes UM formation.

A, Schematic diagram of Gnaq conditional activation, Bap1 conditional CRISPR knockout, and Myc conditional knock-in (Gnaq^CA; Bap1^CKO; Myc^CKI) in the uvea by in situ delivery of pL-CLB (Created in BioRender. Xu, X. (2026) https://BioRender.com/ccwj1ps). B, Kaplan-Meier overall survival curve of Gnaq^CA; Bap1^CKO; Myc^CKI mice after pL-CLB administration. C, Representative in vivo bioluminescence imaging showing strong signal from a primary intraocular tumor. D and E, Representative images of H&E staining of intraocular tumors from Gnaq^CA; Bap1^CKO; Myc^CKI mice. F, Representative fluorescent IHC showing Sox10 and Rpe65 expression in a tumor from a Gnaq^CA; Bap1^CKO; Myc^CKI mouse. Tumors are clonally positive for Sox10 and negative for RPE65. G, Representative image of Periodic acid–Schiff (PAS) staining of a tumor from a Gnaq^CA; Bap1^CKO; Myc^CKI mouse. Tumors displayed minimal PAS positivity. H, Representative image of multiplex IHC (mIHC) showing Gli2, Melan-A, and Tyrp1 expression in an unpigmented tumor from Gnaq^CA; Bap1^CKO; Myc^CKI mice. I and J, Representative images of H&E staining of a pigmented tumor from Gnaq^CA; Bap1^CKO; Myc^CKI mice. K, Representative image of multiplex IHC (mIHC) showing Gli2, Melan-A, and Tyrp1 expression in a pigmented tumor from Gnaq^CA; Bap1^CKO; Myc^CKI mice. L, Representative images of IHC using a melanoma cocktail (Tyrosinase + Melan-A) to detect disseminated UM cells in the livers of Gnaq^CA; Bap1^CKO; Myc^CKI mice. M, Representative images of IHC using GFP and Myc to detect disseminated UM cells in the livers of Gnaq^CA; Bap1^CKO; Myc^CKI mice.

Multiplex IHC revealed intermixed yet spatially distinct GLI2⁺ (neural crest-like (32–34)) and TRYP1⁺/Melano-A⁺ (melanocytic) populations occupying discrete niches within the tumor microenvironment (Fig. 2H), revealing lineage heterogeneity and prompting single-cell transcriptomic analysis (below). A subset of tumors exhibited a strongly pigmented phenotype resembling lesions in Bap1-deficient mice, with expansion of pigmented choroid, Bruch’s membrane breach, and posterior chamber masses containing both pigmented and occasional non-pigmented cells (Fig. 2I,J). The nuclei of pigmented and non-pigmented cells appeared similar, with open chromatin and prominent nucleoli, and mitoses were readily identified (Fig. 2J). These tumors were enriched for TYRP1⁺/Melan-A⁺ differentiated cells, whereas GLI2⁺ cells were rare (Fig. 2K).

Tumor DNA sequencing confirmed CRISPR-mediated editing of Bap1 (Supplementary Fig. S2B), and Western blotting showed marked Bap1 reduction compared with normal retina/RPE (Supplementary Fig. S2C), with residual signal likely reflecting stromal or partially edited cells. MYC expression was weak to moderate (Supplementary Fig. S2D), within a physiological range and lower than in canonical MYC-driven cancers (35–37). Collectively, histologic and molecular analyses demonstrate that Gnaq^CA; Bap1^CKO; Myc^CKI mice develop intraocular tumors that recapitulate key features of human uveal melanoma.

Metastatic uveal melanoma models

Given the association of BAP1 loss-of-function and chromosome 8q gain with high metastatic risk in human UM, we examined the livers of Gnaq^CA; Bap1^CKO; Myc^CKI mice for evidence of dissemination. Although no gross or histologic metastases were detected, IHC for TYR+MLANA revealed rare positive cells in the livers in two of seven mice (Fig. 2L; Supplementary Fig. S2E). These cells also expressed GFP (from the Cas9 allele (28)) and MYC (Fig. 2M), consistent with rare UM cell dissemination to the liver. Because rapid intraocular tumor growth limits long-term metastasis analysis, we evaluated metastatic potential of Gnaq^CA; Bap1^CKO; Myc^CKI tumor cells in transplant models.

Orthotopic tumors were generated by suprachoroidal injection of primary UM cells (QBM12) into immunocompromised NSG and immunocompetent C57BL/6 mice (Fig. 3A). Transplant tumors in both backgrounds reproduced the histologic features of the primary lesions and exhibited mixed GLI2⁺ and TYRP1⁺/Melan-A⁺ populations, clonal SOX10 positivity, and RPE65 negativity (Fig. 3B,C and Supplementary Fig. S2F-H). Tail vein injection of tumor cells produced metastases in the lungs (5/7), livers (2/7), and kidneys (1/7) of NSG mice (Fig. 3D,E), all recapitulating parental tumor morphology (Fig. 3F-H; Supplementary Fig. S2I-K).

Figure 3. Transplant models for UM liver metastasis.

A, Schematic diagram of orthotopic transplant of Gnaq^CA; Bap1^CKO; Myc^CKI primary tumor cells in immunocompetent and immunocompromised mice (Created in BioRender. Xu, X. (2026) https://BioRender.com/g7sncqs). B, Representative images of H&E staining of orthotopic transplant of Gnaq^CA; Bap1^CKO; Myc^CKI primary tumor cells (QBM12) in immunocompetent mice. C, Representative image of multiplex IHC (mIHC) showing Gli2, Melan-A, and Tyrp1 expression in a QBM12 orthotopic transplant tumor. D, Schematic diagram of metastatic transplant of Gnaq^CA; Bap1^CKO; Myc^CKI primary tumor cells in NSG mice via tail vein injection (Created in BioRender. Xu, X. (2026) https://BioRender.com/g7sncqs). E, Representative in vivo bioluminescence imaging showing signal from QBM12 cells 75 days after tail vein inoculation. F-H, Representative images of H&E staining of metastasis in lung (F), kidney (G), and liver (H). I, Schematic diagram of liver metastatic transplant of Gnaq^CA; Bap1^CKO; Myc^CKI primary tumor cells in NSG mice via intrasplenic injection followed by splenectomy (Created in BioRender. Xu, X. (2026) https://BioRender.com/g7sncqs). J, Representative in vivo bioluminescence imaging showing signal from QBM40 cells 30 days after tail vein inoculation. K, Representative images of H&E staining of metastasis in liver. L and M, Representative image of multiplex IHC (mIHC) showing Gli2, Melan-A, and Tyrp1 expression in liver metastasis formed by QBM12 (L) and QBM40 (M).

To explore organ tropism, we derived metastatic cell lines through serial in vivo passage, generating liver-tropic and lung-tropic variants (Supplementary Fig. S2L,M). In vitro, liver-tropic cells exhibited an amoeboid morphology, whereas lung-tropic cells were more elongated and mesenchymal (Supplementary Fig. S2L,M). To model hepatic colonization directly, we performed intrasplenic injection followed by splenectomy (Fig. 3I). Six of eight NSG mice developed multiple hepatic nodules (Fig. 3J). These metastases comprised primarily non-pigmented malignant cells forming expansile nodules with occasional pigmented clusters (Fig. 3K). Multiplex IHC revealed either intermingled GLI2⁺ and TYRP1⁺/Melan-A⁺ cells (Fig. 3L) or spatial segregation with TYRP1⁺/Melan-A⁺ cells predominating (Fig. 3M). Collectively, these findings demonstrate that GNAQ^Q209L expression, BAP1 inactivation, and MYC activation drive intraocular tumors that phenocopy human UM and possess intrinsic potential to disseminate.

Phenotypic characterization of murine uveal melanoma by single cell RNA sequencing

We performed scRNA-seq on three Gnaq^CA; Bap1^CKO; Myc^CKI tumors to characterize the cellular composition and phenotypic landscape of murine uveal melanoma (Supplementary Fig. S3A). After quality control and doublet removal (Supplementary Fig. S3B), 24,455 cells were retained for downstream analysis. Approximately 75% of cells in each tumor were classified as malignant (Fig. 4A, Supplementary Fig. S3A-D). All tumors exhibited immune cell infiltration dominated by macrophages/monocytes and neutrophils, with macrophages partitioning into Arg1+ (Arg1, Ccl9, Flrt3) and C1qa+ (C1qa, C1qb, C1qc, Fcrls) subsets (Fig. 4A, Supplementary Fig. S3C,D). T/NK cells and B cells were less abundant (Fig. 4A, Supplementary Fig. S3C). Several immune exhaustion markers were detected, including Lag3 in T cells and Havcr2 in mature regulatory dendritic cells (mregDC), monocytes, and macrophages (Supplementary Fig. S3E), consistent with previously reported T cell states in Class 2 UM (38). Cancer-associated fibroblasts (CAF) were present in all tumors and further classified into inflammatory (Cxcl5, Il11, Saa3), myofibroblast-like (Actg2, Cd248, Myh11), and activated CAF subsets (Fig. 4A, Supplementary Fig. S3C-D). Additional stromal populations included retinal cells (photoreceptor cells, bipolar cells, Müller glial cells), retinal pigmented epithelium (RPE) cells, multiciliated epithelial cells, and endothelial cells (Fig. 4A, Supplementary Fig. S3C-D).

Figure 4. Characterization of phenotypic diversity in murine UM.

A, UMAP plot showing malignant and microenvironmental cell types in UMs from Gnaq^CA; Bap1^CKO; Myc^CKI mice analyzed by scRNA-seq. Cell types are color-coded. B, UMAP plot showing the two major subtypes in UMs from Gnaq^CA; Bap1^CKO; Myc^CKI mice, Melanocytic and Neural Crest-like (color-coded). C, Violin plot showing the expression of Melanocytic and Neural Crest-like subtype markers. D, UMAP plot showing phenotypic clusters among the malignant UM cells from Gq^CA; Myc^CKI; Bap1^CKO tumors. Clusters are color-coded. E, Dot plot showing the expression of cluster markers, with dot color indicating the average expression levels and dot size indicating the percentage of cells in each cluster expressing the respective marker. F, Heatmap showing the relationship of murine UM cell states to murine cutaneous melanoma phenotypic states based on the relative expression of the indicated gene signatures.

Consistent with the multiplex IHC evidence of lineage heterogeneity (Fig. 2H,K), scRNA-seq revealed two major malignant cell states, Melanocytic and Neural Crest-like, occupying distinct regions of UMAP space (Fig. 4B). The Melanocytic state was defined by high expression of mature melanocytes markers (Tyrp1, Dct, Mlana, and Slc24a5), whereas the Neural Crest-like state expressed markers of neural crest progenitors, premelanoblasts, and melanoblasts (Gli2, Glis3, Meis1, and Sox5) (Fig. 4C, Supplementary Fig. 3D). Marker-based subclustering further resolved the major Melanocytic and Neural Crest-like states into multiple minor phenotypes (Fig. 4D,E). Within the Neural Crest-like compartment, we identified TRIO-high (Trio, Ngf, Arhgap10, Camta1) and MITF-high (Mitf, Gsk3b, Nfia, Tcf4) clusters corresponding to neural crest-like cells and early melanoblasts, respectively. Within the Melanocytic compartment, we detected Epithelial-like (Krt18, Atp1b1, Gpnmb, Prdx1), Mesenchymal-like (Vim, Acta2, Mfap4, Pcolce), Proliferative (Mik67, Pcna, Stmn1, Top2a), Hypoxia (Bnip3, Egln3, Higd1a, Ndufa4l2), Stem-like (Gas6, Gdf15, Igfbp2, Pdgfc), and RPE-like (Rbp1, Rdh5, Rlbp1, Ttr) clusters. A vascular-mimicry population (Itgb6, Pecam1, Prtg, Scin), a state associated with Class 2 GEP (39) and chromosome 3 loss and 8q gain (40), contained cells from both Neural Crest-like and Melanocytic lineages (Fig. 4D,E). All major and minor malignant subtypes were comparably represented across the three tumors and showed similar quality metrics (Supplementary Fig. S4A-C). Notably, these phenotypic states partially overlapped with those observed in Nras^Q61K; p16Ink4a^−/− murine cutaneous melanoma (41), suggesting shared transcriptional and phenotypic programs across melanoma types despite being induced by distinct oncogenic drivers (Fig. 4F, Supplementary Fig. S4D). These analyses reveal a complex tumor microenvironment with malignant cells exhibiting various phenotypic states and an immune microenvironment reminiscent of that seen in human UM.

Melanocytic and Neural Crest-like UM cells exhibit distinct gene expression profiles

Given that MYC promotes intraocular tumor development in our UM model, we examined the expression of Myc and its target genes across malignant cell states. Endogenous Myc was moderately higher in Melanocytic cells than in Neural Crest-like cells (Fig. 5A,B), and expression of the ectopic Myc transgene (inferred from linked Luciferase expression) was lost in a subset of Neural Crest-like cells (Supplementary Fig. S5A). Endogenous Myc expression was not significantly changed among the two Neural Crest-like and seven Melanocytic minor subtypes (Supplementary Fig. S5B). These patterns suggest that the Neural Crest-like and Melanocytic states differentially rely on MYC activity. Accordingly, MYC target gene expression (defined by the GSEA_Hallmark geneset) was significantly higher in Melanocytic cells (Fig. 5C). Gene set enrichment analysis of Hallmark signatures identified enrichment of MYC targets, oxidative phosphorylation, reactive oxygen species (ROS) pathway, mTORC1 signaling, UV response, fatty acid metabolism, and apoptosis in Melanocytic cells (Fig. 5D). Gene ontology (GO) analysis further revealed enrichment of ribosome and mitochondrial signatures in the Melanocytic state (Fig. 5E). Indeed, genes most strongly correlated with Myc expression included numerous ribosomal protein genes, and these expression correlations were conserved in human UM (Fig. 5F). Consistently, 84 of 90 ribosomal genes and 74 of 78 mitochondrial ribosomal genes were upregulated in the Melanocytic compared to the Neural Crest-like state (Fig. 5G). These findings indicate that Melanocytic and Neural Crest-like UM subtypes exhibit distinct transcriptional programs that correlate with Myc expression.

Figure 5. MYC activity across the UM progression continuum.

A, UMAP plot showing Myc expression in malignant UM cells from Gnaq^CA; Bap1^CKO; Myc^CKI tumors. B, Violin plots comparing endogenous Myc expression between the Melanocytic and Neural Crest-like subtypes. C, UMAP plot showing expression of Myc targets in murine UM cells, color coded by average expression of canonical Myc targets from “GSEA_Hallmark_MYC_targets_v1/v2”. D and E, Gene set enrichment analysis (GSEA) of Hallmark signature (D) and gene ontology (GO) (E) in Melanocytic and Neural Crest-like UM cells. F, Spearman correlation between Myc expression and the indicated genes in murine UM from the scRNA-seq analysis and human UM from the UM dataset form TCGA (TCGA-UVM). G, Scatter plot showing the expression changes of 90 ribosome genes and 78 mitochondrial ribosome genes between Melanocytic and Neural Crest-like cells. H, UMAP of human UM scRNA-seq dataset, color coded by Class 1 and Class 2 primary tumors. I and J, Violin plots showing expression of MYC (I) and MYC targets (J) in Class 1 and Class 2 human UM. K, Percentage of Class 1 and Class 2 human UM cells positive for the Melanocytic or Neural Crest-like signatures. L, Violin plot showing MYC expression in Class 2 human UM cells positive for the Melanocytic or Neural Crest-like signatures.

To determine whether these murine states align with human UM phenotypes, we analyzed a scRNA-seq dataset of Class 1 and Class 2 primary human UM (38) (Fig. 5H). MYC expression was significantly higher in Class 2 than Class 1 tumors (Fig. 5I, Supplementary Fig. S5C), accompanied by increased MYC target gene expression (Fig. 5J, Supplementary Fig. S5D). Both murine-derived Melanocytic and Neural Crest-like signatures were present in human UM, but distributed distinctly across classes. Class 1 tumors predominantly expressed the Melanocytic signature but lacked the Neural Crest-like signature, whereas Class 2 tumors contained fewer Melanocytic cells and a distinct population of Neural Crest-like cells (Fig. 5K, Supplementary Fig. S5E). Both Class 1 and Class 2 tumors contained cells positive for both signatures (Fig. 5K, Supplementary Fig. S5E), though whether this represents transitional states remains to be determined. As in the mouse model, Neural Crest-like Class 2 cells expressed significantly lower MYC levels than Melanocytic cells (Fig. 5L). Analysis of 80 TCGA UM tumors grouped by established Class 1 versus Class 2 criteria (11) similarly showed elevated Neural Crest-like markers (GLI2, NGF) and reduced melanocytic differentiation markers (DCT, PMEL) in Class 2 tumors (Supplementary Fig. S5F-I). Additionally, Class 2 tumors exhibited upregulation of Stem-like markers (GAS6, GDF15, IGFBP2) (Supplementary Fig. S5J–L), supporting a shift toward a dedifferentiated, plastic cell state in aggressive UM.

Four genes in the 15-GEP prognostic test (CDH1, ECM1, HTR2B, RAB31) are overexpressed in Class 2 UM (10,42), three of which (Htr2b, Ecm1, Rab31) are significantly upregulated in murine Neural Crest-like cells (Supplementary Fig. S5M-O), consistent with the expression profile of BAP1-deficient Class 2 UM. Furthermore, Pros1 and Mertk, an immunosuppressive ligand-receptor pair upregulated by BAP1 loss in human UM (43,44), were highly expressed in murine Neural Crest-like cells and tumor-associated macrophages (Supplementary Fig. S5P, Q). These findings indicate that the Gnaq^CA; Bap1^CKO; Myc^CKI model recapitulates transcriptional features of human Class 2 UM, with the Neural Crest-like state particularly resembling the Class 2 subtype.

Transcription factor programs underlying UM cell state diversity

To delineate transcriptional programs shaping UM cell states, we determined transcription factor (TF) motif enrichment among genes expressed across malignant subpopulations. The top 15 enriched motifs formed distinct activity modules across cells that correspond to diverse transcriptional states (Fig. 6A). Motifs for KLF4, JUNB, EGR1, FOSB, YBX1, JUND, ATF4, ATF3, JUN, and MYC were enriched in Melanocytic cells (Fig. 6A). This was consistent with higher Myc expression (Fig. 5A,B) and predominant expression of the AP-1 family members FOSB, JUN, JUNB, and JUND in the Melanocytic subtype (Fig. 6B-E). In contrast, RBPJ, KLF6, RUNX, TEAD1, and FOXP2 motifs were enriched in Neural Crest-like cells, corresponding to elevated Yap1 and Tead1 (Fig. 6F, G). Interestingly, Fosl1 expression correlated strongly with Tead1 and Runx1 rather than with other AP-1 family members (Fig. 6G-I), consistent with reports that the gene product of Fosl1, FRA1, may form atypical complexes with TEAD1 to promote therapy resistance (45,46). These TF programs indicate that divergent AP-1- and YAP-dependent regulatory circuits underlie Melanocytic and Neural Crest-like states and reflect signaling heterogeneity downstream of oncogenic Gαq.

Figure 6. The transcription factor landscape across murine UM subtypes.

A, Heatmap showing Z-score normalized transcription factor (TF) motif activity across mouse UM single cells from Gnaq^CA; Bap1^CKO; Myc^CKI tumors, as inferred by SCENIC analysis of the scRNA-seq data. Hierarchical clustering of both cells and motifs reveals distinct regulatory programs associated with Melanocytic and Neural Crest-like UM states. B-I, UMAP showing the expression of Jun (B), Jund (C), Fosb (D), Fos (E), Yap1 (F), Tead1 (G), Runx1 (H), and Fosl1 (I) in mouse UM cells from Gnaq^CA; Bap1^CKO; Myc^CKI tumors.

Trajectory analysis reveals UM progression to a dedifferentiated state

To infer the dynamic transitions between transcriptional states within the malignant cell populations, we performed RNA velocity analysis using spliced and unspliced transcript abundances (Fig. 7A). This analysis revealed a strong directional flow from Melanocytic to Neural Crest-like cells (Fig. 7A). Two major routes were identified: the first path originated from the Mesenchymal-like cluster and arrived at the TRIO-high cluster via the MITF-high cluster while the second path originated from the Hypoxia cluster, passed through the Vascular-mimicry, Epithelial-like, and Stem-like clusters, and also ended at the TRIO-high cluster (Fig. 7A).

Figure 7. Progression trajectory analysis of murine UM.

A, RNA velocity analysis based on spliced and unspliced transcript abundances revealed directional transitions across mouse UM cell states from Gnaq^CA; Bap1^CKO; Myc^CKI tumors. Velocity vector fields projected onto the UMAP embedding indicate major evolutionary paths. B, Latent time analysis shows the latency of single UM cells, color coded by pseudotime. C, Expression dynamics of representative genes plotted along latent pseudotime. D-F, UMAP plots showing the expression of Foxo3 (D), Cd44 (E), and Bach2 (F) in murine UM cells.

To further reconstruct the continuous trajectory of transcriptional changes, we computed latent time, providing an unsupervised ordering of cells along inferred differentiation paths (Fig. 7B). Cells with the earliest latent time values predominantly resided within the Hypoxia, Vascular-mimicry, Stem-like, Epithelial-like, and parts of the Mesenchymal-like clusters. Intermediate latent times were associated with the Proliferative cluster and parts of the Mesenchymal-like cluster from the Melanocytic subtype, and the MITF-high cluster from Neural Crest-like subtype. The latest latent times were enriched in cells from the TRIO-high cluster. This latent time progression aligns with the RNA velocity analysis, suggesting a potential dedifferentiation trajectory from the Melanocytic to the Neural Crest-like subtype.

Gene expression dynamics aligned with these temporal patterns. Early latent time exhibited high expression of proliferation and cell cycle genes (Cenpe, Cenpf, Mki67, Ccnb1, Top2a), followed by induction of immediate-early genes (Fos, Egr1, Jun) (Fig. 7C), indicating activation of stress-response pathways and transcriptional remodeling. Late latent times showed a shift toward a more progenitor-like and migratory state, characterized by the induction of Ngf, Cd44, Foxo3, Prune2 (Fig. 7C-E). These genes are associated with cell survival, motility, and stemness, hallmarks of neural crest-like properties. Additional late-expressed genes such as Nedd4l, Pparg, and Bach2 further supported the emergence of a plastic, less differentiated state (Fig. 7C, F). Collectively, these gene expression dynamics support a model in which proliferative melanocytic cells progressively dedifferentiate into Neural Crest-like states through intermediate stress-adaptive and progenitor-like stages.

Genomic instability in murine UM correlates with transcriptional states and recapitulates copy number alterations in human UM

To detect copy number alterations (CNA) in murine UM, we applied inferCNV analysis to the scRNA-seq data using non-malignant immune cells as diploid reference. This analysis revealed two CNA-defined tumor subpopulations, CNA Branch 1 and CNA Branch 2, that differed in the extent of aneuploidy (Fig. 8A,B). CNA Branch 1, enriched for Neural Crest-like cells, exhibited broad chromosomal amplifications and deletions, while Branch 2, enriched for Melanocytic cells, was comparatively diploid (Fig. 8B,C). Functional state annotation further revealed that CNA Branch 1 consisted predominantly of TRIO-high and half of MITF-high cells, with minor contributions from Stem-like and Vascular-mimicry states, whereas CNA Branch 2 encompassed nearly all Melanocytic subtypes (Vascular mimicry, Stem-like, Proliferative, Hypoxia, Epithelial-like, Mesenchymal-like, and RPE-like) (Fig. 8D), suggesting a link between dedifferentiation and genomic instability.

Figure 8. Copy number alterations correlate with UM subtypes.

A, Inferred copy number alteration (CNA) profiles of individual malignant cells from murine UM tumors using inferCNV, using immune cells as diploid references. Heatmap shows relative genomic alteration levels across chromosomes, and hierarchical clustering of UM cells revealed two distinct CNA-defined UM branches. B, UMAP plot of mouse UM cells colored by CNA branch assignment. C and D, Proportional composition of CNA branch within each major subtypes (C) and phenotypic clusters (D). E and F, Heatmap of genes frequently amplified (E) or deleted (F) in both murine and human UM, color coded by frequency of chromosomal gain (E) or loss (F) in mouse UM cells or TCGA-UVM samples.

We observed a coordinated increase in gene expression across regions of mouse chromosomes 15, 1, and 5 and decreased expression of mouse chromosomes 3, 14, and 16, syntenic to parts of human chromosomes 8q and 3, respectively (Fig. 8A). These CNAs potentially recapitulate chromosome 8q gain and loss of chromosome 3 in human Class 2 UM. Cross-species gene orthology mapping to the TCGA uveal melanoma dataset identified 1,832 genes with CNAs in >50% of Branch 1 or Branch 2 cells that also exhibit alterations in human UM, 189 of which are altered in >40% of human UM cases. Of these, 105 amplified genes mapped mostly to human chromosome 8q (103 on 8q, 2 on 6p) (Fig. 8E) and 84 deleted genes localized exclusively to human chromosome 3 (Fig. 8F). While all 105 gained genes were shared by both CNA branches, only 12 of the 84 lost genes overlapped across branches, suggesting distinct trajectories of genomic alterations. These findings indicate that our UM model displays CNA patterns that recapitulate hallmark genomic features of human Class 2 UM.

Melanocytic state predominance and T-cell exhaustion in early UM tumors

Gnaq^CA; Bap1^CKO; Myc^CKI mice that developed intraocular tumors rapidly within 1–2 months after induction, and we performed scRNA-seq on four such tumors. After quality control and doublet removal, 30,768 cells were retained for downstream analysis. Malignant cells comprised 80–90% of each tumor (25,976 cells total) (Supplementary Fig. S6A-C). In addition to the malignant compartment, all four tumors showed prominent immune infiltration dominated by T cells (2,364 cells) and NK cells (648 cells) (Supplementary Fig. S6A,C). T cells expressed high levels of exhaustion markers (Lag3, Pdcd1, Ctla4), while conventional dendritic cells, monocytes, and macrophages expressed Havcr2 (TIM-3), consistent with exhausted immune states previously reported in the Class 2 UM microenvironment (38)(Supplementary Fig. S6D). CAFs and other stromal populations, such as retinal cells (photoreceptor, bipolar, and Müller glia), RPE cells, multiciliated epithelial cells, and endothelial cells, were present but less abundant than in slower growing tumors (Supplementary Fig. S6A). Distinct from the prior scRNA-seq analysis of tumors harvested at 3–4 months, where malignant cells segregated into the Melanocytic and Neural Crest-like subtypes (Fig. 3B), these early and fast-developing tumors consisted almost exclusively of Melanocytic cells (Supplementary Fig. S6A,E). Within this dominant lineage, we identified transcriptionally distinct clusters representing Progenitor-like (Eya1, Lgr4, Notch2, Tead1), MITF-high hybrid (Dct, Mitf, Oca2, Tyr), RPE-like (Rbp1, Rlbp1, Ttr), Hypoxia-associated (Bnip3, Egln3, Higd1a, Ndufa4l2), Vascular-mimicry (Col4a6, Pecam1, Egr1, Mdk), Stress-responsive (Atf3, Cdkn1a, Junb, Nr4a1), Mesenchymal-like (Cd44, Col3a1, Mmp2, S100a4), Antigen-presenting (Cd74, H2-Ab1, Psmb8, Stat1), Cycling Antigen-presenting (Cd74, H2-Ab1, Psmb8, Stat1, Ccnb2, Ki67, Stmn1, Top2a), and Proliferative (Ki67, Ccnb2, Stmn1, Top2a) states (Supplementary Fig. S6E,F). Despite the dominance of a Melanocytic identity, Progenitor-like and MITF-high hybrid clusters retained expression of neural crest-associated genes (Supplementary Fig. S6G), resembling patterns observed in human UM (Fig. 5K). Together, these analyses revealed that early and rapidly developing tumors are more homogeneous and melanocytic yet already display T-cell exhaustion and subtle lineage plasticity, reflecting the immunosuppressive microenvironment of human UM.

DISCUSSION

Current UM mouse models, while establishing GNAQ/11 as initiating oncogenes, develop aggressive tumors within weeks and require euthanasia due to concurrent cutaneous disease (12,13). These features limit their utility for studying immune interactions and metastatic progression. Our model overcomes these limitations by (i) providing spatiotemporal control over oncogene activation confined to the eye, (ii) modeling stepwise genetic progression that mirrors human disease evolution (47,48), (iii) preserving immune competence to enable analysis of tumor-immune interactions, and (iv) yielding transplantable cell lines with distinct organ tropism.

Our model follows a genetic progression model for UM development, where GNAQ^Q209L activation induces choroidal nevi with limited penetrance, BAP1 loss enhances nevus formation, and MYC activation promotes the formation of fully penetrant intraocular tumors resembling human UM. This stepwise progression models aspects of the multistage evolution observed in human UM, where GNAQ/11 mutations are found in benign nevi, while BAP1 loss and chromosome 8q gains are associated with malignant progression and metastatic risk. While it remains unclear whether MYC is the key gene on 8q that drives the evolutionary pressure to gain more copies of 8q during human UM progression, our model provides experimental evidence that MYC expression, which may occur through chromosome 8q gains in human UM (11,30,48,49), can contribute to the formation of intraocular tumors resembling UM in the mouse. Moreover, triple mutant Gnaq^CA; Bap1^CKO; Myc^CKI mice exhibited hepatic dissemination of tumor cells, indicating that this model recapitulates not only primary tumor formation but also early steps in metastatic progression. Intravenous transplantation of primary UM cells led to metastasis in multiple distant organs, demonstrating the intrinsic metastatic potential of UM cells in this model. Moreover, intrasplenic inoculation of primary UM cells reliably produced hepatic metastases, resembling those observed in patients, offering a practical and reproducible approach for studying the mechanisms of liver colonization and metastatic progression in uveal melanoma.

Single-cell transcriptomic profiling revealed a complex and physiologically relevant tumor microenvironment composed of both malignant and non-malignant populations. Critically, this immune-competent models preserves the native tumor-immune ecosystem, with approximately 12% of tumor cells being immune infiltrates. The immune landscape of our model prominently featured C1q+ and Arg1+ macrophage subtypes as well as Lag3+ exhausted T cells and Havcr2+ dendritic cells (38,50–54), closely resembling the immunosuppressive microenvironment characteristic of human Class 2 UM. The presence of cancer-associated fibroblasts (including inflammatory, myofibroblast-like, and activated subtypes), combined with the diverse immune infiltrate, creates a tumor microenvironment that accurately reflects the complexity of human UM. This enables investigation of tumor-immune interactions and immunotherapies that are currently being explored clinically but lack adequate preclinical models for testing.

The malignant cells segregated into two major states: Melanocytic and Neural Crest-like. This phenotypic bifurcation recapitulates cellular hierarchies previously reported in cutaneous melanoma models (41), suggesting that despite differing anatomical locations and driver mutations, melanocytic malignancies may converge on common transcriptional programs during tumorigenesis. Sub-clustering further revealed phenotypic diversity, including proliferative, mesenchymal, epithelial-like, hypoxic, stem-like, and RPE-like states, reflecting the morphological heterogeneity observed in murine UMs and consistent with the known plasticity of melanoma cell states (55). The Melanocytic and Neural Crest-like cell populations exhibited distinct features, indicating different functional roles in tumor progression. Melanocytic cells showed high expression of MYC target genes and downstream metabolic, ribosomal, and mitochondrial programs, consistent with MYC coordinating cellular growth with anabolic metabolism and protein synthesis (56). The robust correlation between MYC and ribosomal gene expression observed in both our model and the TCGA-UVM dataset underscores a conserved regulatory axis in UM pathobiology. In contrast, Neural Crest-like cells exhibited lower MYC target expression and upregulation of genes associated with immune evasion (e.g., Pros1, Mertk) (43,44) and key clinical biomarkers of aggressive Class 2 UM, including Htr2b, Ecm1, and Rab31 (10,42). Intriguingly, while MYC is upregulated in human Class 2 compared to Class 1 UMs, the MYC expression pattern in Melanocytic and Neural Crest-like subpopulations mirrors the findings in our mouse model, potentially indicating that MYC downregulation is associated with UM progression. This is supported by the clinical observation that high MYC expression levels correlate with better survival outcomes in UM patients (57), suggesting a nuanced role of MYC in UM progression. Moreover, MYC governs cell state plasticity in other contexts, including small cell lung cancer (SCLC) and olfactory neuroblastoma (ONB) (58–60), and whether MYC plays a similar role in driving cell states in UM will be determined in future studies.

Transcriptional network analysis revealed further divergence between these states. Melanocytic cells showed enrichment of MYC and AP-1 binding motifs, aligning with elevated expression of proliferation-related and metabolic genes. The AP-1 complex, which has been implicated in melanocyte lineage maintenance and melanoma progression (47,61), is a canonical effector of MAPK signaling and activated downstream of mutant GNAQ, suggesting these cells are transcriptionally poised for proliferation and biosynthesis. Neural Crest-like cells exhibited enrichment of motifs associated with YAP-TEAD signaling, RUNX1, FOXP2, RBPJ, and KLF6, corresponding with elevated expression of Yap1 and Tead1. This is notable as YAP is a key downstream target of GNAQ/GNA11 signaling (62,63), has been shown to drive dedifferentiation and stem-like properties (64,65), and may thus play a potential role in promoting cellular plasticity and metastatic competence in our model. The co-expression and motif co-enrichment of FRA1 (FOSL1) with TEAD1 is particularly intriguing, as their non-canonical interaction has been implicated in drug resistance (45,46).

The phenotypic segregation paralleled distinct patterns of genomic instability. Neural Crest-like cells showed broad chromosomal amplifications and deletions, suggesting dedifferentiation is associated with increased genomic instability. In contrast, Melanocytic cells displayed a more diploid-like profile, reflecting a genomically more stable phenotype. Cross-species mapping identified a substantial overlap between copy number alterations in murine and human UM, particularly gains of chromosome 8q and losses of chromosome 3 (66,67). Among the ~600 protein-coding genes on 8q, 105 orthologous genes underwent copy number gains in our mouse model—an unexpectedly high concordance that underscores the fidelity of this model in recapitulating human UM genomics. Importantly, this broad amplification pattern suggests that multiple drivers beyond MYC are critical for UM progression. While MYC is a well-established oncogene on 8q, the coordinated amplification of numerous genes in the 8q11–13 and 8q21–24 regions indicates that UM progression likely depends on the cooperative effects of multiple 8q-resident genes rather than MYC alone. The fact that our model recapitulates this broad 8q amplification pattern suggests that the selective pressure for 8q gain in UM reflects the need for multiple cooperating oncogenes in this chromosomal region. Deletions corresponding to human chromosome 3 varied between Melanocytic and Neural Crest-like cells, suggesting distinct evolutionary trajectories or selective pressures in these subtypes. The biological significance of this divergence remains unclear but may reflect differential lineage constraints within each subtype. The distinction in CNA patterns further suggests that genetic instability may drive cell state transitions in UM, providing a mechanistic link between dedifferentiation, chromosomal aberrations, and tumor progression.

Our trajectory analysis suggested a directional transition from the Melanocytic to the Neural Crest-like state, accompanied by dynamic regulation of key transcriptional programs. Dedifferentiation toward a Neural Crest-like state has also been observed in cutaneous melanoma and is associated with immune evasion, metastasis, and resistance to therapy (41,68–70). The induction of immediate early genes (Fos, Egr1, Jun) at intermediate stages suggests transcriptional adaptation to cellular stress (71), while the late expression of stem cell-associated genes (Foxo3, Cd44, Ngf, Prune2) points to acquisition of stem-like and migratory traits—features commonly attributed to invasive or therapy-resistant cells (72,73). However, without definitive lineage tracing, we cannot exclude the alternative possibilities that these populations diverge from a common ancestor early during tumor evolution or arise independently from distinct cells of origin.

Analysis of early, fast-growing tumors revealed a predominantly Melanocytic transcriptional program with minimal representation of Neural Crest-like cells, indicating less intratumoral heterogeneity at the initial stages of tumor evolution. This predominant Melanocytic state likely reflects rapid tumor outgrowth before extensive dedifferentiation or lineage diversification occurs. This lends further support to the notion that UM cells in the Melanocytic state are biosynthetically active and fast-cycling while Neural Crest-like cells adopt a less proliferative and possibly more invasive state. Additionally, despite higher immune cell infiltration compared with later-stage tumors, the T-cell compartment displayed pronounced exhaustion, mirroring the immunosuppressive phenotype characteristic of advanced human UM (38,74). These findings suggest that immune dysfunction emerges early during tumor development, even in histologically and transcriptionally homogeneous lesions, underscoring the potential of this model to study the temporal evolution of tumor-immune interactions in uveal melanoma.

An elegant recent study by Kenny and colleagues established a zebrafish uveal melanoma model through choroid-targeted electroporation of oncogenic GNAQ^Q209L and demonstrated that Mitfa-independent, tfec⁺/pax3a⁺ progenitor cells are highly susceptible to transformation (bioRxiv 2025.05.05.652300). This zebrafish model provides a strong complement to our immune-competent mouse model by independently recapitulating GNAQ-driven tumorigenesis within the native ocular microenvironment. Despite species differences, both models converge on similar malignant cell states, including Melanocytic, Progenitor-like, and Neural Crest-like populations, highlighting conserved transcriptional hierarchies in GNAQ-mutant UM. Together, these findings suggest that intratumoral heterogeneity in uveal melanoma may arise through at least two complimentary mechanisms: (i) progressive dedifferentiation along a trajectory from Melanocytic to Neural Crest-like, as inferred from our trajectory analysis, and (ii) malignant transformation of developmentally different precursors such as melanocytes, melanoblasts, or neural crest-derived progenitor cells, as suggested by the findings in the zebrafish model. The concordance between these models strengthens a unified, cross-species framework for understanding lineage plasticity and tumor evolution in UM.

In summary, our study introduces a biologically and preclinically relevant mouse model of UM that recapitulates many key features of the human disease, including cell state heterogeneity, genomic alterations, and metastatic potential. Our work demonstrates that overexpression of MYC can substantially increase the malignant potential of these mouse tumors, although this remains to be verified in human UM. Interestingly MYC can cooperate with GNAQ mutation and BAP1 loss to enhance malignant transformation and its spontaneous downregulation by tumor cells is associated with a transition to the Neural Crest-like state. The identification of distinct Neural Crest-like and Melanocytic subtypes, each with unique transcriptional signatures and copy number landscapes, adds to a growing appreciation of UM as a heterogeneous and dynamic disease, as also highlighted by scRNA-seq analyses of human (38) and zebrafish UM (bioRxiv 2025.05.05.652300). This immune-competent model provides a powerful platform for advancing UM research by enabling studies of drivers of metastatic progression such as PRAME (75) and GDF15 (bioRxiv 2025:2025.05.07.652654), tumor-immune interactions, and preclinical testing of immunotherapies – critical needs given the limited treatment options for metastatic disease.

Limitations of this study

While our mouse model represents a significant advance in UM research, we acknowledge certain limitations. The lentiviral Cre delivery approach may potentially result in tumors with atypical cellular heterogeneity, as suggested by the RPE-like and epithelial-like malignant populations identified in our scRNA-seq analysis. This heterogeneity could reflect the biology of human UM, which also displays phenotypic plasticity, but might also result from non-specific recombination events in different ocular cell types such as RPE cells. Despite this limitation, our model successfully recapitulates molecular signatures and phenotypic states of human UM, particularly the aggressive Class 2 subtype. Future iterations of the model will incorporate lineage-specific Cre drivers to improve resolution in defining the initiating cell population.

SIGNIFICANCE STATEMENT.

Development of a mouse model resembling the genetic progression and phenotypic plasticity of human uveal melanoma offers an immune-competent preclinical platform for understanding disease immunogenomics and testing immune and targeted treatments.