The single-cell data for UM utilized in this study were sourced from the NCBI Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/), specifically under the registration number GSE139829 (21), which comprises 11 UM samples (GSM4147091-GSM4147101). Additionally, relevant clinical data for UM were obtained from The Cancer Genome Atlas (TCGA) (https://portal.gdc.cancer.gov/) for further analysis.

UM scRNA-seq data were processed and downscaled for clustering using the “Seurat” package (version 4.3.0) (22). To eliminate potential doublet cells, we employed the “DoubletFinder” R package (version 2.0.3) (23–26). Cells were filtered according to specific criteria to exclude low-quality data: (1) total gene transcript count (nCount) per cell ranging from 500 to 100,000; (2) number of transcribed genes per cell (nFeature) between 300 and 6,000; (3) mitochondrial gene ratio less than 25%; and (4) erythrocyte gene count ratio under 5%.

Using the “NormalizeData” function, we normalized the high-quality UM single-cell data to identify the top 2,000 highly variable genes, subsequently applying the “FindVariableFeatures” function (27–29). We then employed the “ScaleData” function (30, 31). Additionally, the “CellCycleScoring” function was utilized to evaluate cell cycle effects. Dimensionality reduction was conducted using the “RunPCA” function, with the first 30 valid principal components selected (32, 33). The “Harmony” R package (version 0.1.0) was employed to mitigate batch effects, followed by clustering analysis using the “FindNeighbors” and “FindClusters” functions (34, 35). Clusters were labeled based on existing literature, and a Uniform Manifold Approximation and Projection (UMAP) plot was generated for visualization (36, 37). The Ro/e algorithm was also implemented to assess tissue preferences among different cell clusters.

To effectively characterize the differentially expressed genes (DEGs) within each UM tumor subcluster, we employed the “FindAllMarkers” function in Seurat, coupled with the ‘ClusterProfiler’ R package for Gene Ontology Biological Process exploration (38). Additionally, we conducted Gene Set Enrichment Analysis (GSEA) for each UM tumor subcluster, focusing on Gene Ontology Biological Process (GOBP) enrichment entries and highlighting the pathways with the highest Normalized Enrichment Scores (NES).

We utilized CytoTRACE analysis (39) to evaluate and predict the relative differentiation status of UM tumor subclusters, with scores ranging from 0 to 1, reflecting a positive correlation with stem cell properties. Furthermore, we applied the “Slingshot” R package (version 2.6.0) (40) to infer the cellular lineage of the UM tumor subclusters. By employing the “getLineage” and “getCurves” functions, we mapped the differentiation trajectories of each subcluster, assessing the dynamic changes in gene expression over time. Based on the results of the pseudotemporal analysis, alongside previous findings, we identified key UM subclusters integral to tumor development.

We conducted a comprehensive analysis of the cellular metabolic pathway profiles of UM tumor subclusters using the R package scMetabolism. The top five metabolism-related pathways were visualized, allowing us to examine their distribution on the UMAP map as well as their expression across various tissue types.

The pySCENIC package (version 0.10.0) in Python (version 3.7) was employed to investigate the enrichment of transcription factors (TFs) and their regulatory activity within each UM subcluster. Initially, we applied the GRNBoost method to compute linkage weights between each TF and its potential target genes. This was followed by DNA sequencing analysis to identify possible direct binding targets. Subsequently, the regulatory activity in each cell was assessed using AUcell, allowing us to screen for the top five transcription factors with the highest activity, for which specificity scores were calculated and visualized.

To evaluate cellular communication among different UM subclusters, we utilized the “CellChat” R package (version 1.6.1) (41). This analysis inferred cellular interactions at the level of signaling pathways and receptor-ligand dynamics, with outgoing and incoming signaling patterns visualized to depict varying signal intensities across cell types.

To enhance clinical relevance, we constructed a novel prognostic model for UM using bulk RNA-seq data from UM patients. We conducted univariate Cox analysis on the top 100 candidate genes from key subgroups to identify those significantly associated with UM prognosis (42–44). To mitigate the risk of overfitting, LASSO regression analysis (45–48) was employed to determine the optimal λ value. And multivariate Cox analysis was performed to determine the coefficients for each gene according to the calculated formula: RISK SCORE = ∑ i n Xi × Yi (X: coefficient, Y: gene expression level). Based on the optimal cut-off value of the PCOLCE TCs Risk Score, the UM cohort was categorized into high and low scoring groups, and Kaplan-Meier survival curves were generated to predict outcomes (49–51). Furthermore, we incorporated clinical factors such as age, race, tumor stage, and survival time to construct a nomogram for predicting overall survival (OS) using the “rms” R package, which was validated through ROC curve (52–54) and decision curve analyses (DCA).

We assessed immune infiltration in tumor tissues utilizing various algorithms, including ESTIMATE, CIBERSORT, EPIC, and Xcell. To investigate the correlation between immune cell populations and prognosis-related genes, we employed the “corrplot” R package (version 0.92) for correlation analysis. Additionally, immune cell abundance was quantified using the “CIBERSORT” R package (version 0.1.0) (55). To further elucidate the mechanisms of immune escape, we calculated Tumor Immune Dysfunction and Exclusion (TIDE) values across different groups.

After identifying differentially expressed genes (DEGs) across scoring groups, we performed Gene Ontology (GO) enrichment analysis (56, 57), encompassing Gene Ontology Cellular Component, Biological Process, and Molecular Function, to uncover potential differences in biological functions among the groups. We also utilized the “GSVA” R package to evaluate the biological properties and molecular distinctions of the groups, providing insights into their alterations in cellular behavior and function. Furthermore, we employed the “maftools” R package (58) to analyze the somatic mutation profiles of UM patients from the TCGA database, assessing the correlation between risk scores and tumor mutation burden (TMB) using the Spearman correlation test.

We obtained the relevant drug sensitivity dataset from The Genomics of Drug Sensitivity in Cancer (GDSC) (https://www.cancerrxgene.org/). Utilizing the “pRRophetic” package (version 0.5) (59, 60), we calculated IC50 values to evaluate the drug sensitivity across different groups.

MP65 and 92-1 cell lines were procured from the American Type Culture Collection (ATCC). Cells were maintained in RPMI-1640 medium supplemented with 10% fetal bovine serum and 1% penicillin/streptomycin. Cultures were incubated under standard conditions of 37°C, 5% CO2, and 95% humidity.

CITED1 knockdown was achieved using small interfering RNA (siRNA) constructs sourced from GenePharma (Suzhou, China). Transfection was conducted following the manufacturer’s instructions for Lipofectamine 3000 RNAi Max (Invitrogen, USA). Cells were seeded in 6-well plates at 50% confluence and transfected with either a negative control (si-NC) or CITED1 knockdown constructs (Si-CITED1-1 and Si-CITED1-2), employing Lipofectamine 3000 RNAiMAX for each transfection.

Cell viability of the transfected MP65 and 92-1 cells was assessed using the CCK-8 assay. Cells were inoculated at a density of 5 × 10³ per well in a 96-well plate. Following incubation, 10 μL of CCK-8 solution (A311-01, Vazyme) was added to each well, and the plate was incubated in the dark at 37°C for 2 hours. Absorbance was measured at 450 nm using an enzyme-labeled instrument (A33978, Thermo) to evaluate cell viability on days 1, 2, 3, and 4. The average optical density (OD) values were calculated and represented in a line graph.

Transfected cells were seeded at a density of 1 × 10³ per well in a 6-well plate. Following a wash with PBS, the cells were fixed with 4% paraformaldehyde and stained with 0.1% crystal violet (Solarbio, China). Colony formation was then quantitatively analyzed.

Cells were inoculated into a 24-well plate chamber (Corning, USA), with or without matrix gel (BD Biosciences, USA). The cell suspension was added to the upper chamber of a Costar Transwell plate, while medium supplemented with serum was added to the lower chamber. After a 48-hour incubation in a cell incubator, the cells were fixed with 4% paraformaldehyde and stained with crystal violet. Migration and invasion abilities were subsequently assessed under a microscope.

Transfected cells were cultured in 6-well plates until they reached 95% confluence. A sterile 200 μL pipette tip was used to create linear scratches in the cell monolayer, followed by gentle washing with PBS. The culture medium was replaced, and images of the scratches were captured at 0 and 48 hours to measure the wound width.

Transfected MP65 and 92-1 cells were seeded at a density of 3 × 10³ per well in a 6-well plate and incubated overnight. A 2× EdU working solution was prepared by adding a 10 mM EdU solution in serum-free medium, followed by a 2-hour incubation at 37°C. Cells were fixed with 4% paraformaldehyde for 30 minutes and treated with glycine (2 mg/mL) and 0.5% Triton X-100 for 15 minutes. Following fixation, the cells were incubated with 1 mL of 1× Apollo and 1 mL of 1× Hoechst 33342 for 30 minutes at room temperature. Cell proliferation was quantified using fluorescence microscopy.

All analyses were conducted using R software (version 4.3.0). The Wilcoxon test and Pearson’s correlation coefficient were employed to assess significance between groups (*P< 0.05, **P< 0.01, ***P< 0.001).

We obtained single-cell sequencing data for UM from the GEO database (GSE139829), comprising a total of 11 samples (GSM4147091-GSM4147101). After implementing quality control measures, we identified 87,529 high-quality cells derived from two tissue types: eight primary UM patients (T) and three metastatic UM patients (M) ( Supplementary Figure 1A ). These high-quality cells were classified using the Seurat framework, resulting in the identification of 39 Seurat clusters ( Supplementary Figure 1B ). Based on marker gene expression, these clusters were annotated into eight cell types: T_NK, endothelial cells (ECs), tumor cells, fibroblasts, photoreceptor cells, retinal pigment epithelial (RPE) cells, plasma cells, and myeloid cells ( Supplementary Figure 1C ). Notably, we focused on the tumor cell cluster, which predominantly originated from primary UM patients, although a subset was also derived from metastatic UM patients ( Supplementary Figure 1D ). The distribution of cell types across the two tissue sources was illustrated in Supplementary Figure 1E , revealing that 83.2% of tumor cells came from tumor tissues, while 67.1% were from metastatic tissues. Additionally, T_NK cells comprised 9.5% from tumor tissue and 18.0% from metastatic tissue. The top five genes expressed in each cell type are displayed in Supplementary Figure 1F , with CD74, C1QA, C1QC, C1QB, and HLA-DRA identified as the most prominent genes in the tumor cell cluster.

To investigate the heterogeneity of tumor cells in UM, we conducted an in-depth analysis of the tumor cell clusters. We classified UM tumor cells into six distinct subclusters based on marker gene expression: C0 WSB1+ TCs, C1 EEF1A1+ TCs, C2 HSPB7+ TCs, C3 XIST+ TCs, C4 GNG11+ TCs, and C5 PCOLCE+ TCs. The distribution of these subclusters was illustrated in the central UMAP plot of Figure 1A . Our analysis of the tumor cell origins (Groups M and T) and cell cycle phases (G1, G2M, and S) revealed that the C3 XIST+ TCs and C1 EEF1A1+ TCs subclusters were predominantly derived from Group T, with very few originating from Group M. Additionally, we assessed the copy number variation (CNV) core and Cell Stemness AUC for each UM subcluster ( Figures 1A, E ). The results indicated that C3 XIST+ TCs exhibited a higher CNV core, whereas C0 WSB1+ TCs displayed a greater Cell Stemness AUC. The InferCNV analysis results for the UM tumor cell subclusters were presented in Supplementary Figure 2 .

The top five genes for each UM subcluster were identified and are shown in Figure 1B : C0 WSB1+ TCs (FOSB, JUN, FOS, ZFP36, HLA-B), C1 EEF1A1+ TCs (RPS26, RPS12, RPSA, RPL32, CHCHD10), C2 HSPB7+ TCs (CHCHD10, NDUA4L2, GAPDH, RPS4Y1, NDRG1, A2M), C3 XIST+ TCs (MALAT1, NEAT1, MTRNR2L12, MT-ND6, XIST), C4 GNG11+ TCs (GNG11, HHATL, RPS2, RPSA.1, SNHG7), and C5 PCOLCE+ TCs (GDF15, S100A4, CDKN1A, GEM, IGFBP7).

Analysis of tissue type preferences for each UM subcluster revealed that the C5 subcluster had the highest affinity for tumor tissue ( Figure 1C ). The specific proportions of sample sources for each subcluster were depicted in the scale diagram in Figure 1D , indicating that the C5 subcluster was primarily derived from T5, while the C0 subcluster originated from multiple patient samples. Furthermore, the analysis of the identified genes (WSB1, EEF1A1, HSPB7, XIST, GNG11, PCOLCE) highlighted that EEF1A1, associated with the C1 subcluster, was expressed across nearly all UM subclusters ( Figure 1F ).

The top five DEGs for each UM subcluster were illustrated using volcano plots ( Figure 1G ), highlighting the most significantly up- and down-regulated genes in each subcluster. Subsequent analysis of these DEGs for GOBP enrichment aimed to identify biological processes that were overrepresented in each subcluster ( Figure 2A ). This analysis enhanced our understanding of the functional roles of DEGs within the various UM subclusters.

Building on the GOBP-enriched entries, we conducted GSEA for each UM subcluster to evaluate the significance of specific biological pathways and identify the entries with the highest NES in each subcluster ( Figure 2B ). The primary enriched biological processes were as follows: C0: antigen processing and presentation of peptide antigen; C1: cytoplasmic translation; C2: response to hypoxia; C3: developmental growth involved in morphogenesis; C4: cytoplasmic translation; and C5: positive chemotaxis. These findings indicate distinct functional specializations among the UM subclusters, reflecting their unique roles within the tumor microenvironment.

For instance, the enrichment observed in the C0 subcluster suggested a potential association with immune surveillance or immune evasion mechanisms, while the prominence of “cytoplasmic translation” in the C1 and C4 subclusters indicated heightened protein synthesis, likely linked to rapid cell proliferation or specific metabolic requirements. Overall, these results underscored the functional heterogeneity of UM subclusters, each possessing unique biological roles that might have influenced tumor behavior and therapeutic response.

To further investigate the heterogeneity within each tumor subcluster of UM, we employed various methods to analyze their developmental trajectories and differentiation potential. CytoTRACE analysis revealed that the C5 subcluster exhibited the highest CytoTRACE score, indicating greater developmental potential, while the C1 subcluster displayed the lowest score, reflecting a higher degree of differentiation and reduced developmental capacity ( Figure 2C, D ). These findings provided critical insights into the differentiation status and functional heterogeneity of UM cells.

We also conducted Slingshot analysis to elucidate the developmental trajectories of the UM subclusters. Lineage1 originated from C3 XIST+ TCs, progressing sequentially through C1 EEF1A1+ TCs, C4 GNG11+ TCs, C2 HSPB7+ TCs, and C0 WSB1+ TCs, ultimately culminating in C5 PCOLCE+ TCs ( Figure 2E ). Integrating these results with the CytoTRACE analysis, we identified the C5 PCOLCE+ TCs subcluster as a key player among UM malignant cells, positioned at the terminal stage of the developmental trajectory and possessing high differentiation potential. Therefore, we showed great research interest in the C5 PCOLCE+ TCs subcluster.

The naming genes for each UM subcluster within Lineage1 were displayed in Figure 2F . Notably, EEF1A1, the naming gene for the C1 subcluster, maintained high expression throughout the proposed temporal trajectory, whereas PCOLCE, associated with the C5 subcluster, consistently exhibited low expression levels, with a notable increase at the terminal stage. This suggested that the C5 PCOLCE+ TCs subcluster may play a critical role in the late stages of UM tumor progression, with changes in gene expression indicating its unique functions in tumor differentiation and potential metastasis. Given the upregulation of PCOLCE at the terminal stage, we hypothesized that it may contribute to regulating the tumor microenvironment, facilitating extracellular matrix remodeling, or interacting with other tumor-promoting factors to drive the progression of UM malignant cells.

The top five metabolism-related pathways of the UM subclusters were illustrated in Figure 3A , including oxidative phosphorylation, glycolysis/gluconeogenesis, and glutathione metabolism. The distribution and expression levels of these pathways across each UM subpopulation were presented in Figures 3B and 3C . Notably, C4 GNG11+ TCs and C5 PCOLCE+ TCs exhibited elevated expression across all five metabolic pathways ( Figure 3B ). Among these, the C5 subcluster displayed particularly high expression levels in oxidative phosphorylation, glutathione metabolism, and cysteine and methionine metabolism when compared to other UM subclusters ( Figure 3C ). Additionally, the expression of these metabolic pathways across different tissue sources (T and M) was depicted in the violin plot of Figure 3D , revealing significantly higher expression levels in T tissues, with all results demonstrating statistical significance.

To identify the core TFs within the UM subclusters, we conducted a SCENIC analysis. Utilizing pySCENIC, we inferred the gene regulatory networks associated with these subclusters. The top five TFs and their relative expression levels across each UM subcluster were presented in Figure 4A . The most active TFs identified were GTF2B in C0 WSB1+ TCs, OTX2 in C1 EEF1A1+ TCs, RELA in C2 HSPB7+ TCs, ELF2 in C3 XIST+ TCs, PPARG in C4 GNG11+ TCs, and HES6 in C5 PCOLCE+ TCs. Figure 4B ranks these transcription factors based on their Regulon specificity scores (RSS). In the C5 PCOLCE+ TCs subcluster, the top five TFs identified were HES6, E2F3, HLTF, NFYC, and GTF2B, with their specific expression levels detailed in Figure 4C . Notably, HES6 exhibited high expression in the C3 XIST+ TCs, C4 GNG11+ TCs, and C5 PCOLCE+ TCs subclusters.

Cell communication encompassed the capacity of cells to receive, process, and transmit signals, playing a crucial role in coordinating various biological activities such as development, differentiation, and inflammation. This interaction was primarily facilitated by ligand-receptor complexes, enhancing cell-to-cell communication. In our study, we employed the “CellChat” R package to quantitatively infer and analyze the communication networks among UM cells. The results revealed extensive communication links across different cell types, particularly focusing on the interactions between the C5 PCOLCE+ TCs subcluster and other clusters. Notably, the number and intensity of interactions between C5 PCOLCE+ TCs and fibroblasts were elevated, suggesting a significant signaling exchange between these cell types ( Figure 4D ). This robust interaction implies that C5 PCOLCE+ TCs may play a pivotal role in tumor progression and related biological processes, such as tissue remodeling and extracellular matrix regulation, through their communication with fibroblasts in the tumor microenvironment. Figure 4E illustrated the afferent and efferent signal strengths of all UM cell interactions, highlighting that nearly all UM subclusters are associated with CD99. Furthermore, within the incoming signaling patterns, the C5 PCOLCE+ TCs subgroup exhibited strong connections to CCL, SEMA6, LAMININ, and MK, underscoring the importance of these pathways in cell signal transduction.

To enhance clinical decision-making for UM and to gain deeper insights into patient prognosis, we developed a novel prognostic model centered on the key C5 PCOLCE+ TCs subgroup. We began by conducting univariate Cox analysis on the top 100 candidate genes from this subgroup to identify those with significant prognostic relevance ( Supplementary Figure 3 ). Subsequently, we performed LASSO regression analysis on the selected prognostic genes, yielding a lambda.min of 0.064 and identifying eight crucial genes ( Figure 5A ): APOE, ARC, CITED1, COX6C, S100A4, ATP5I, APOA1BP, and C4orf48. Multivariate Cox regression analysis revealed that APOE acted as a protective gene, while the majority were associated with increased risk ( Figure 5B ).

Using the expression levels of these genes and their respective coefficients, we constructed a new UM prognostic model, termed the PCOLCE TCs Risk Score (PTRS). The calculation formula was defined as: PCOLCE TCs Risk Score (PTRS)= ∑ i n Xi × Yi (X: coefficient, Y: gene expression level). Based on the optimal cutoff value of the PTRS, we categorized UM patients into Low PTRS and High PTRS groups to analyze survival outcomes. Consistent with our expectations, results indicated that the High PTRS group had a poorer prognosis ( Figure 5C ). Additionally, to evaluate the predictive capability of the PTRS model, we generated ROC curves, yielding AUC values of 0.803, 0.930, and 0.923 at 1, 3, and 5 years, respectively ( Figure 5D ), thereby demonstrating the model’s high sensitivity and specificity.

We conducted multivariate Cox regression analysis incorporating clinical factors alongside the PTRS to evaluate its potential as an independent risk factor. The analysis revealed a hazard ratio (HR) for PTRS of 30.42, with a 95% confidence interval (CI) of 6.48–142.80 and a P-value< 0.001 ( Figure 5E ). This finding underscored the significant association of PTRS with prognosis, suggesting its role as an independent prognostic indicator for UM patients and providing robust support for clinical decision-making.

To enhance prognostic accuracy, we developed a nomogram integrating clinical factors—such as gender, age, and tumor clinical stage (T, N, M)—alongside PTRS to predict overall survival (OS) for UM patients at 1, 3, and 5 years ( Figure 5F ). Both the PTRS subgroups and tumor clinical stage significantly influenced OS outcomes. Decision Curve Analysis (DCA) further evaluated the reliability of this UM prognostic model ( Figure 5G ), confirming its utility as a clinical decision support tool. The C-index for the nomogram model, displayed in Figure 5H , indicated predictive accuracy ranging from 0.5 to 1 for OS at 1, 3, and 5 years, demonstrating the model’s robust predictive ability across various time points. Notably, the C-index approached 1 over time, reinforcing its reliability for long-term prognostic assessments.

Additionally, Figure 5I illustrated the correlation between the eight prognostic genes within the PTRS model and both OS and PTRS, revealing that OS was negatively correlated with most of these genes. This comprehensive analysis highlighted the potential of the nomogram in providing valuable prognostic information for clinical practice.

We investigated the survival and mortality status, as well as the levels of the PTRS, across the Low and High PTRS Groups over time, alongside the overall distribution of the eight prognostic genes within the PTRS model ( Figure 6A ). The specific distribution of these genes in different PTRS cohorts was illustrated using ridge plots combined with boxplots ( Figure 6B ). Notably, APOE exhibited higher expression in the Low PTRS Group, whereas the other seven genes (ARC, CITED1, COX6C, S100A4, ATP5I, APOA1BP, and C4orf48) were more highly expressed in the High PTRS Group, with all differences being statistically significant.

Based on the expression levels of the eight PTRS model genes, the UM patient cohort was stratified into high and low expression groups for each gene ( Figure 6B ). The analysis revealed that the High APOE Group experienced better survival outcomes compared to the Low APOE Group, with this difference also being statistically significant. In contrast, the remaining seven PTRS model prognostic genes were associated with poorer prognosis in the High expression group, further establishing APOE as a protective factor while indicating that the other genes serve as risk factors.

The tumor immune microenvironment significantly influenced tumor progression and affected the response to therapy. To further investigate the heterogeneity among different PTRS cohorts, we analyzed immune infiltration in both Low and High PTRS Groups. Figures 7A, B illustrated the distribution of 22 types of immune-infiltrating cells across these cohorts, with macrophages representing a substantial proportion relative to other immune cells. Figure 7C detailed the specific distribution levels of various immune cells within the Low and High PTRS Groups.

Moreover, we examined the correlation between immune cell types, prognostic genes, OS, and PTRS within the PTRS model, as shown in Figure 7D . M1 macrophages exhibited a positive correlation with most prognostic genes in the PTRS model while being negatively associated with OS. In contrast, resting CD4 memory T cells were negatively correlated with most prognostic genes but positively correlated with OS. Figure 7E displayed the correlations between immune cells and PTRS, highlighting that resting CD4 memory T cells were significantly negatively correlated with PTRS, whereas follicular helper T cells showed a significant positive correlation. This suggested their involvement in pro-tumorigenic immune regulatory processes that may facilitate tumor growth or immune evasion. These findings underscored the notion that various immune cell types not only differ in abundance but also maintain a complex dynamic balance in their functional roles within the tumor microenvironment.

Additionally, we calculated the immune stromal scores for the different PTRS groups, revealing that the Immune Score, ESTIMATE Score, and Stromal Score were significantly higher in the High PTRS Group compared to the Low PTRS Group ( Figures 7F–H ). These results indicated a greater degree of immune infiltration and stromal component activity within the High PTRS tumor microenvironment, suggesting a more intricate structural composition associated with tumor progression and treatment resistance. Conversely, the Low PTRS Group exhibited higher Tumor Purity values ( Figure 7I ), implying a simpler microenvironment with less immune and stromal disruption, correlating with improved treatment responsiveness and better prognosis. Notably, TIDE scores did not significantly differ between the two groups ( Figure 7J ).

Finally, we assessed the correlations between immune checkpoint genes and the eight PTRS-related genes, OS, and PTRS, with results depicted in bubble plots ( Figure 7K ). The analysis revealed significant positive correlations between the constructed PTRS and several genes (COX6C, CITED1, ARC, S100A4) with most immune checkpoint genes. Figure 7L illustrated the expression differences of immune checkpoint genes across PTRS cohorts, showing that most were more highly expressed in the High PTRS Group. We also integrated various algorithms, including ESTIMATE, CIBERSORT, EPIC, and Xcell, to quantify immune infiltration across the PTRS groups, with the results presented in Figure 7M .

We identified DEGs across various PTRS cohorts, presenting the results using volcano plots ( Figure 8A ). These plots effectively illustrated the distribution of significantly up- and down-regulated genes, facilitating the visualization of notable expression differences among the PTRS groups. To further investigate the global variability in gene expression profiles, we conducted principal component analysis (PCA), which revealed that Dim1 and Dim2 accounted for 25.2% and 10.3% of the variance, respectively. This indicated that these principal components capture a substantial portion of the intergroup gene expression differences ( Figure 8B ), reaffirming the significant heterogeneity among the PTRS cohorts.

We subsequently performed Gene Ontology (GO) enrichment analysis on the DEGs, categorizing them into three main classifications: Cellular Component, Biological Process, and Molecular Function ( Figure 8C ). The GO enrichment results highlighted significant associations with several immune-related biological processes, particularly in pathways such as immune response regulation, cell adhesion, and cell signaling at the plasma membrane. These processes were crucial for modulating the tumor microenvironment and immune responses, suggesting that immune regulation may play a vital role in the varying PTRS groups.

The mutation waterfall plot revealed that GNAQ exhibited the highest mutation frequency at 49%, followed closely by GNA11 at 44% ( Figure 8D ). An analysis of CNV events among the PTRS model prognostic genes showed that most genes lacked CNV alterations, although some gains were noted in ARC and COX6C ( Figure 8E ). However, no significant differences in tumor mutation burden (TMB) values were observed between the different PTRS groups ( Figure 8F ). Based on TMB and PTRS values, the UM cohort was classified into four groups: High PTRS-High TMB, High PTRS-Low TMB, Low PTRS-High TMB, and Low PTRS-Low TMB. Notably, the High PTRS-High TMB group exhibited the worst survival prognosis (P< 0.0001) ( Figure 8G ). The correlation between TMB and PTRS was weak, with R = -0.042 and P = 0.71 ( Figure 8H ).

We further conducted gene set variation analysis (GSVA) utilizing gene sets from the Molecular Marker Database (MsigDB) to elucidate the biological characteristics of the different PTRS groups. The GSVA results, depicted in Figure 8I , demonstrated significant differences in multiple biological functions and pathways among the groups. Additionally, an in-depth analysis of GOBP revealed variations in specific biological processes linked to different PTRS groupings ( Supplementary Figure 4A ). Supplementary Figure 4B illustrated the Spearman correlation analysis between PTRS and the HALLMARK gene set, further elucidating molecular distinctions among the groups. These findings may provide new insights into the biology of UM and its therapeutic strategies.

To enhance clinical applicability, we systematically evaluated the sensitivity of various drugs in Low and High PTRS groups. We analyzed the differences in semi-inhibitory concentrations (IC50) for these drugs to inform individualized therapeutic approaches. Our assessment indicated that Paclitaxel and Bortezomib displayed higher IC50 values in the Low PTRS Group, suggesting reduced sensitivity to these agents. Conversely, Docetaxel exhibited higher IC50 values in the High PTRS Group, indicating a comparatively weaker effect in this cohort ( Figure 8J ).

In summary, we selected CITED1, the risk gene with the smallest absolute value in the PTRS prognostic model, for in vitro experiments to validate the applicability of the PTRS model in UM. We employed two UM cell lines, MP65 and 92-1, establishing both a negative control group and a CITED1 knockdown group for comparative analysis.

Cell viability assays ( Figure 9A ) demonstrated that knockdown of CITED1 significantly reduced cell viability, as indicated by CCK-8 results. Colony formation assays ( Figure 9B ) revealed a marked decrease in the number of colonies formed by both MP65 and 92-1 cell lines following CITED1 knockdown compared to the negative control group. In Transwell assays ( Figure 9C, D ), the migration and invasion capabilities of the CITED1-knockdown cell lines were significantly impaired.

Wound healing assays ( Figures 10A, B ) showed that the scratches in the CITED1 knockdown groups were considerably wider after 48 hours compared to the negative control, indicating a notable reduction in migratory capacity. EdU staining ( Figures 10C, D ) further corroborated these findings, revealing a decrease in cell proliferation in the CITED1 knockdown UM cell lines, emphasizing the critical role of CITED1 in cell growth.

Collectively, our experimental results indicated that knockdown of CITED1 significantly impaired cell viability, proliferation, migration, and invasion in both UM cell lines. These findings underscored the potential of CITED1 as a therapeutic target, highlighting its significance in the pathogenesis of UM and supporting the clinical feasibility of the PTRS prognostic model. Thus, targeting CITED1 and implementing the PTRS model may offer novel strategies and insights for the treatment of UM.