Machine learning identifies glycosphingolipid signature linking immune dysregulation and clinical prognosis in uveal melanoma

Abstract

Purpose

To investigate glycosphingolipid biosynthesis (GSB) dysregulation in uveal melanoma (UVM) and develop a machine learning-driven prognostic signature bridging GSB activity, tumor microenvironment, and clinical outcomes.

Methods

Using TCGA and GEO cohorts, GSB activity was quantified via Gene Set Variation Analysis (GSVA). Differential expression analysis, least absolute shrinkage and selection operator regression, and Cox regression were used to identify prognostic GSB-related genes and establish the glycosphingolipid biosynthetic risk score (GBRS). Kaplan-Meier survival analysis, receiver operating characteristic (ROC) curve analysis, and C-index were used to evaluate GBRS. Immune infiltration was assessed using ssGSEA and ESTIMATE algorithms. Drug sensitivity (IC50) correlations were evaluated via oncoPredict.

Results

We uncovered pervasive GSB dysregulation across cancer types, with GSB score showing superior prognostic accuracy to traditional clinical variables in UVM. An eight-gene prognostic signature was identified and refined to a three-gene GBRS comprising CA12, SLC44A3, and BHLHA15. GBRS independently predicted overall survival and demonstrated predictive value for disease-specific survival, progression-free interval, and metastasis-free survival. GBRS correlated positively with CD8⁺ T cell and M2-like macrophage infiltration. Meta-analysis confirmed CA12 as a significant prognostic risk factor, with sphingolipid/glycosphingolipid synthesis pathways enriched in its co-expressed genes.

Conclusions

GBRS is a robust prognostic biomarker for UVM, reflecting immune dysregulation and therapeutic vulnerabilities.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-026-04494-3.

Introduction

Uveal melanoma (UVM), the most common malignant intraocular tumor in adults, is known for its high metastatic potential, aggressive nature, and associated high mortality [1]. Over the course of disease progression, up to 50% of UVM patients will develop distant metastases [2]. Once metastasis occurs, the median survival time for these patients plummets to a mere 18 to 20 months [3, 4]. Despite recent advancements in therapeutic strategies, there has been no substantial improvement in UVM patient survival rates [2], highlighting the critical need for early diagnosis and the development of novel therapeutic targets to enhance patient outcomes.

Glycosphingolipids (GSLs) are membrane-associated glycolipids involved in cell signaling, intercellular communication, and cell recognition [5]. Dysregulated GSL metabolism has been implicated in tumor progression and anti-tumor immunity across several malignancies [6, 7]. However, despite such findings in cutaneous melanoma and other cancer types [8, 9], the biological and prognostic relevance of glycosphingolipid biosynthesis in UVM has not been systematically investigated.

In this study, we assessed the dysregulation of glycosphingolipid biosynthetic activity and its prognostic impact in pan-cancer, discovering that it was a significant and independent risk factor in UVM. Therefore, we endeavored to develop a new indicator, glycosphingolipid biosynthetic risk score (GBRS), to predict the efficacy of therapeutic interventions and prognosis in UVM, and sought to explore the correlation between GBRS and both the tumor microenvironment and drug sensitivity. This research may help optimise precision treatment and further improve the clinical outcomes of UVM patients.

Methods

Data acquisition and processing

RNA-sequencing expression data and corresponding clinical metadata for the Cancer Genome Atlas (TCGA) cohort were downloaded utilizing the “TCGAbiolinks” package. Formalin-fixed paraffin-embedded samples within the TCGA dataset were excluded due to their demonstrated inferior performance in sequencing analyses [10, 11]. Disease-free interval (DFI), disease-specific survival (DSS), and progression-free interval (PFI) information of TCGA cohorts were acquired from the UCSC Xena database. Microarray gene-expression profiles were retrieved from the Gene Expression Omnibus (GEO) repository using the “GEOquery” package.

Gene set variation analysis and functional enrichment analysis

Gene sets of the Kyoto Encyclopedia of Genes and Genomes (KEGG), HALLMARK, Gene ontology (GO) and the glycosphingolipid biosynthesis pathway (version 2023.2) were retrieved from the Molecular Signatures Database (MSigDB) [12]. Gene set variation analysis (GSVA) was performed using the “gsva” function of the GSVA package [13]. The “limma” package was employed to identify differentially enriched pathways across high- and low-risk groups, with significant criteria of |Log2 Fold Change| > 0.25 and adjusted p-value < 0.001. Gene set enrichment analysis (GSEA) and overrepresentation analysis was conducted using the “enrichKEGG” function from the “clusterProfiler” package [14].

Differentially expressed analysis

The “DESeq2” package was used to discern differentially expressed genes (DEGs) according to the criteria of |log2 (fold change) |> 1 and adjusted p-value < 0.05. Genes were filtered if the proportion of samples with non-zero gene expression was greater than 80% or if the average raw count was less than 1, to improve DEGs detection sensitivity [15].

Risky model construction and validation

Univariate Cox regression analysis was conducted to evaluate the prognostic significance of the glycosphingolipid biosynthesis (GSB) signatures. Subsequently, Least Absolute Shrinkage and Selection Operator (LASSO) regression, combined with multivariate Cox regression analyses, was applied to further shrink the candidate genes to construct the most optimal signature. The “glmnet” package was utilized to determine the optimal penalty parameter value (lambda.min) through 1000 iterations of cross-validation [16]. Patient-specific risk scores were calculated by integrating the coefficients and expressions of the prognostic genes, as follows: ∑(Coef_i * Expr_i), where Coef_i denotes the coefficient of gene i, and Expr_i signifies the expression level (log2[TPM]) of gene i in patient i.

Immune cell infiltration analysis

Single-sample gene set enrichment analysis (ssGSEA) method, implemented in R package GSVA, was leveraged to quantify the relative abundance of 28 immune cells in the TCGA-UVM cohort [17]. The ESTIMATE algorithm was harnessed to derive Immune Scores, Stromal Scores, and ESTIMATE Scores [18]. Additionally, immune cell infiltration data from multiple algorithms, including TIMER, quanTIseq, MCP-counter, xCell, CIBERSORT, CIBERSORT-ABS, and EPIC, were retrieved from the Tumor IMmune Estimation Resource (TIMER) 2.0 database for further correlation validation [19].

Prediction of drug sensitivity and immunotherapy analyses

The prediction of the half maximal inhibitory concentration (IC50) was conducted using “oncoPredict” package, with GDSC v2 data serving as the training set [20]. A lower IC50 value indicates greater sensitivity to the drug in cancer treatment.

Statistical analysis

All data processing, statistical analysis, and visualizations were conducted in R software version 4.2.0. Correlations between continuous variables were assessed using Spearman’s rank correlation coefficients. The Wilcoxon rank-sum test was employed to ascertain the statistical divergence between variables across distinct groups. Cox regression and Kaplan-Meier survival analyses were conducted using the “survival” package. Meta-analysis was executed with the “meta” package. Time-dependent area under the receiver operating characteristic curve (ROC) analyses were performed using the “timeROC” package to assess the diagnostic specificity and sensitivity of the variables. All statistical tests were two-tailed, with a significance threshold set at p < 0.05.

Results

Dysregulation and prognostic impact of glycosphingolipid biosynthetic activity

In an effort to understand the alterations in GSB associated with cancer, we quantified GSB activity scores across all TCGA samples using GSVA method. Our findings revealed a significant decrease in GSB scores in tumor tissues compared to their corresponding normal tissues in 15 different cancer types (Fig. 1A). Further analysis of the prognostic implications of GSB scores through univariate Cox regression indicated that GSB scores could serve as a prognostic biomarker in five specific cancer types, including uveal melanoma (UVM), stomach adenocarcinoma (STAD), lung adenocarcinoma (LUAD), low-grade glioma (LGG), and sarcoma (SARC) (Fig. 1B-C). Notably, in UVM, the GSB score was identified as a significant prognostic risk factor for all three survival indices. Multivariate Cox regression analysis confirmed that the GSB score is an independent prognostic risk factor for overall survival (OS) in UVM patients (Fig. 1D). Compared to tumor staging, GSB score demonstrated superior predictive performance, with a 3-year survival area under the curve (AUC) of 0.796 and the C-index of 0.710, while tumor staging yielded values of 0.622 and 0.626, respectively. Additionally, the combination of GSB scores with stage status further enhanced prognostic predictive accuracy, as indicated by increased AUC values and a higher C-index (Fig. 1E). These results suggest that the GSB score emerges as an effective predictor of prognosis in UVM, with dysregulations in GSB activity potentially playing a critical role in the initiation and progression of UVM.

Fig. 1.

Glycosphingolipid Biosynthetic Activity Dysregulation. (A) Relative glycosphingolipid biosynthesis (GSB) activity scores in the TCGA pan-cancer cohort. (*p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001) (B) Prognostic impact of GSB scores on OS, DSS, DFI and PFI as determined by univariate Cox regression. Red denotes a risk factor for cancer prognosis, while blue indicates a protective factor. Only p-values < 0.05 are displayed. (C) Forest plot of the OS prognostic effect of GSB scores using univariate Cox regression. (D) Multivariate analysis of OS with respect to clinicopathological characteristics and GSB scores in the TCGA-UVM cohort. (**p < 0.01) (E) AUC of time-dependent ROC analysis and C-index for GSB scores in the TCGA-UVM cohort

Construction of glycosphingolipid biosynthetic prognostic gene signature

To address the computational complexity of the GSB score in clinical applications, we aimed to develop a simplified and tractable linear risk model to enhance the efficiency of prognostic prediction. Initially, we conducted differential expression analysis twice to filter genes involved in GSB. Firstly, patients were stratified into high- and low-GSB groups based on the median GSB score (Supplementary Fig. 1A). Secondly, we calculated the prognostic score integrating both GSB score and stage status through multivariate Cox regression, followed by differential expression analysis between high- and low-risk groups (Supplementary Fig. 1B). Applying a cutoff criterion of absolute log2 Fold Change ≥ 1.0 and adjusted p-value < 0.05, we identified 792 DEGs by intersecting the results (Fig. 2A). Additionally, we performed correlation analysis between gene expression and GSB score, identifying 1506 genes significantly negatively correlated and 3223 genes significantly positively correlated with GSB score. The intersection of differential and correlation analyses yielded 713 potential GSB-related genes (Fig. 2B), predominantly enriched in immune-related pathways (Supplementary Fig. 1C).

Fig. 2.

Prognostic Model Construction. (A) Venn diagram of DEGs from two rounds of analysis. (B) Venn diagram of DEGs and genes with significant expression correlation to GSB scores. (C) Cross-validation of the prognostic impact of GSB-related genes in LASSO. (D) Forest plot of the OS prognostic effect of the GBRS signature genes by multivariate Cox regression. (E) Forest plot of the OS prognostic effect of GBRS and other clinicopathological characteristics by multivariate Cox regression. (F) Time-dependent ROC analysis for predicting OS at 1, 3, and 5 years using GBRS in the TCGA-UVM cohort. (G) Kaplan-Meier survival curves for OS in low- and high-GBRS groups of the TCGA-UVM cohort

Subsequently, to construct a risk model for UVM patients, univariate analysis was conducted on all these GSB-related genes in the discovery cohort (TCGA-UVM), leading to the identification of 470 genes significantly associated with prognosis (p-value < 0.05). Then, these genes underwent LASSO-Cox proportional hazards regression and 10-fold cross-validation. The LASSO coefficient profile plot against the log(k) sequence revealed 8 optimal coefficients using the minimize k method (Fig. 2C). Finally, we incorporated these 8 core genes into a multivariate stepwise Cox regression analysis, resulting in a gene model comprising CA12, SLC44A3 and BHLHA15, which exhibited the optimal regression efficiency (Fig. 2D).

External validation of glycosphingolipid biosynthetic risk score

Based on the three-gene prognostic model, we established the GBRS using the following formula:

GBRS = (1.48 × CA12 expression level) + (3.07 × BHLHA15 expression level) + (0.37 × SLC44A3 expression level).

As depicted in Fig. 2E, the GBRS is an independent prognostic factor irrespective of gender, age, and stage. ROC curve analysis demonstrated the discriminatory capacity of the GBRS, with 1-, 3-, and 5-year AUCs of 0.88, 0.96, and 0.94 in TCGA-UVM cohort (Fig. 2F). Stratifying patients in the TCGA-UVM cohort into high-GBRS and low-GBRS groups revealed significantly poorer prognosis for high-GBRS patients (Fig. 2G). To evaluate the potential impact of clinical characteristics on the predictive efficacy of the GBRS, we performed survival analyses across patients with diverse clinical and pathological profiles. Consistently, in all subgroups of age (≥ 60 or < 60 years), sex (male or female), or pathological stage (stage II or stage III/IV), patients in the high-GBRS group showed significantly shorter median OS and higher mortality rates than those in the low-GBRS group (Supplementary Fig. 2).

To validate the robustness and efficacy of the GBRS across diverse populations, we conducted ROC curve analysis and survival analysis in external cohorts GSE84976 and GSE44295. The ROC curves consistently showed high prognostic predictive accuracy, with AUC values of 0.85, 0.92, 0.87, 0.89 for 1/3/5/10-year predictions in GSE84976, and 0.92, 0.78, 0.75 for 1/3/5-year predictions in GSE44295 (Fig. 3A, C). High-GBRS patients demonstrated significantly poorer metastasis-free survival (MFS) compared to low-GBRS patients (Fig. 3B and D). Moreover, the GBRS also performed well in assessing DSS and PFI in the TCGA-UVM cohort (Fig. 3E-H), with AUC values for predicting 1-year, 3-year, and 5-year survival rates all exceeding 0.75. These findings indicate that the GBRS possesses robust prognostic evaluation capabilities.

Fig. 3.

Validation of the GBRS. Time-independent ROC curves of GBRS for MFS in GSE84976 (A), GSE44295 (C), and DSS, PFI in TCGA-UVM (E, G) cohorts. Kaplan-Meier analyses for MFS in GSE84976 (B), GSE44295 (D), and DSS, PFI in TCGA-UVM (F, H) cohorts

Considering the relatively low expression of BHLHA15 (not detected in one-third of samples from the GSE44295 cohort, recorded as missing (NA)), we developed a dual-gene prognostic prediction model using only CA12 and SLC44A3. The dual-gene prognostic model exhibited robust performance across various endpoints not only in the TCGA cohort, encompassing OS, DSS, and PFI (Fig. 4 A-F), but also in external validation cohorts such as GSE84976, GSE22138, and GSE44295 in MFS (Fig. 4G-L). The AUC values of the ROC curve exceeded 0.6 in all cohorts. Patients classified in the high-GBRS group experienced significantly poorer survival outcomes compared to those in the low-GBRS group. While the ROC and C-index values of the dual-gene model slightly decreased in predicting OS in the TCGA-UVM and MFS in GSE44295 cohorts, there was no such decrease in predicting MFS in GSE84976 cohorts (Supplementary Fig. 3 A-F). Decision curve analysis (DCA) and integrated discrimination improvement (IDI) analysis indicated no statistically significant difference between the dual-gene prognostic model and GBRS (Supplementary Fig. 3G-I). Since BHLHA15 was not included in the platforms of GSE22138 microarray datasets, a direct comparison of the dual-gene model with the GBRS was not feasible.

Fig. 4.

Dual-Gene Model Validation. Time-independent ROC curves of the dual-gene model for OS, DSS, PFI in TCGA-UVM (A, C, E), and MFS in GSE84976 (G), GSE22138 (I), GSE44295 (K) cohorts. Kaplan-Meier analyses for OS, DSS, PFI in TCGA-UVM (B, D, F) cohort, and MFS in GSE84976 (H), GSE22138 (J), GSE44295 (L) cohorts

In conclusion, based on these comprehensive results, we assert that the GBRS exhibits substantial superiority in prognostic accuracy and performance, providing a valuable tool for risk stratification and clinical decision-making in UVM patients.

Dissection of tumor microenvironment based on GBRS

To elucidate the biological functions and signaling pathways associated with the GBRS, we conducted a comprehensive gene set variation analysis using KEGG and Hallmark gene sets in the TCGA-UVM and GSE84976 cohorts. The results revealed significant upregulation of multiple inflammatory and immune-related pathways in patients with higher GBRS, including the IL6-JAK-STAT3 pathways, IL2-STAT5 pathways, TNFα-NF-κB pathways, IFN-α response pathways, IFN-γ response pathways, cytokine-cytokine receptor interaction and NOD-like receptor signaling pathway. Additionally, pathways implicated in tumor growth and metastasis, such as angiogenesis, cell adhesion molecules, and Notch signaling, were markedly enriched in high-GBRS patients (Supplementary Figs. 4, 5). Gene Set Enrichment Analysis (GSEA) confirmed these observations, highlighting significant enrichment of immune response pathways in high-GBRS patients (Fig. 5A, B). Furthermore, ESTIMATE analysis revealed a robust positive correlation between GBRS and both immune scores and ESTIMATE scores (Fig. 5C-D). Collectively, these results suggest that GBRS may serve as a potential biomarker for immune infiltration and response.

Fig. 5.

Relationship between tumor microenvironment and GBRS. (A-B) GSEA indicating activation of multiple immune response pathways in high-GBRS samples of TCGA-UVM and GSE84976 cohorts. (C-D) Correlation analysis between GBRS, immune score, and ESTIMATE score in TCGA-UVM and GSE84976 cohorts. (E) Distribution of 28 immune cell subset infiltration abundances between high and low-GBRS groups in TCGA-UVM and GSE84976 cohorts (top 25% vs. bottom 25%, *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001)

Building on these insights, we employed the ssGSEA method to identify immune cell infiltrates associated with GBRS. Our analysis disclosed a significant enrichment of diverse immune cell populations in the high-GBRS group, including CD4 T cells, CD8 T cells, type 1 T helper cells, NK cells, immature dendritic cells, mast cells, and myeloid-derived suppressor cells (MDSCs) (Fig. 5E). Additionally, we conducted correlation analyses between GBRS and the infiltration abundance of various immune cells, utilizing methods including CIBERSORT, CIBERSORT-ABS, EPIC, MCP-COUNTER, QUANTISEQ, TIMER, and XCELL. Notably, GBRS showed a robust positive correlation with the levels of infiltrating CD8 T cells and M2-like macrophages (Supplementary Fig. 6A). In line with this, a significant positive correlation was observed between markers of CD8 T cells and M2-like macrophages and GBRS (Supplementary Fig. 6B, C). These findings collectively imply a potential enhancement of immunotherapy efficacy in the high-GBRS group, underscoring the need for further investigation into this area.

Efficacy of GBRS in predicting drug sensitivity

To explore the relationship between the GBRS and drug sensitivity, we calculated the IC50 values for various drugs and assessed their correlation with GBRS in both the TCGA-UVM and GSE84976 cohorts. Drugs with consistent correlation patterns are illustrated in Fig. 6. Our analysis revealed a positive correlation between GBRS and the IC50 values of ERK_2440, ulixertinib, VX-11e, BI-2536, SB505124, daporinad, sinularin, sepantronium bromide, and UMI-77. Conversely, the IC50 values of ribociclib, KU-55,933, PF-4,708,671, palbociclib, BMS-754,807, dactolisib, ABT737, and AZ960 were negatively correlated with GBRS. Furthermore, we found that the drug sensitivities of olaparib and LY-294,002 were positively associated with GBRS, while trametinib and cobimetinib demonstrated a negative correlation with GBRS in the CellMiner database (Supplementary Fig. 7). These findings imply that high-GBRS UVM patients may exhibit resistance to standard chemotherapy regimens but could potentially show sensitivity to other FDA-approved therapeutic agents.

Fig. 6.

Correlation between drug sensitivity and GBRS. (A) Dot plot illustrating the relationship between IC50 values and GBRS, with circle size representing the p-value of the correlation. Golden yellow indicates positive correlation with GBRS, and dark green indicates negative correlation. (B) Bubble plot of the relationship between IC50 values and GBRS signature genes, with circle size representing the p-value of the correlation and color indicating the correlation coefficient

Prognostic and functional analysis of CA12

Among the three GBRS signature genes, CA12 demonstrated a significant correlation with tumor stage and was identified as an independent prognostic risk factor, separate from factors such as gender, age, BMI, tumor stage, and tumor size, for OS, DSS, and PFI (Supplementary Fig. 8). The sensitivity of ERK_2440, ulixertinib, and VX-11e, which are significantly positively correlated with GBRS, also showed a notably positive correlation with CA12 expression levels (Fig. 6B). In the GSE44295 cohort, excluding samples with missing values, CA12 expression levels were lower in 90.9% (50/55) of uveal melanoma samples compared to normal melanocytes (Fig. 7A).

Fig. 7.

Expression Distribution, Prognostic Effect and Functional Analysis of CA12. (A) Expression distribution of CA12 across different cell types. (B) Forest plot of meta-analysis showing the prognostic effect of CA12. (*p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001) (C-G) Functional enrichment results of CA12 co-expressed genes in TCGA-UVM (C), GSE22138 (D), GSE39717 (E), GSE84976 (F), and GSE44295 (G) cohorts

Consequently, we conducted an extensive prognostic and functional analysis of CA12 across multiple UVM cohorts from GEO and TCGA databases. The meta-analysis confirmed CA12 as a prognostic risk factor for OS, DSS, PFI, and MFS (Fig. 7B). Furthermore, by screening CA12 co-expressed genes in these cohorts, we found that they were significantly enriched in glycan, sphingolipid, and glycosphingolipid metabolism pathways (Fig. 7C-G).

Discussion

Uveal melanoma (UVM) remains a major clinical challenge due to its high metastatic propensity and poor response to existing systemic therapies [21, 22]. Given the roles of glycosphingolipids (GSLs) in oncogenic signaling, membrane dynamics, and immune regulation, their biosynthetic pathways represent biologically meaningful candidates for prognostic modeling and therapeutic exploration [23–25].

In this study, we demonstrate that glycosphingolipid biosynthesis (GSB) activity is a strong prognostic determinant in UVM, with a GSB-derived score outperforming AJCC staging. Although GSB levels are generally reduced in tumors compared with normal tissues across cancers, higher GSB activity within UVM tumors predicts worse outcomes. This is consistent with UVM’s unique biology: elevated GSB likely reflects tumors with enhanced membrane remodeling, hypoxia-related signaling, and chronic inflammation, aligning with the well-recognized “immune paradox” in which CD8⁺ T-cell and macrophage infiltration correlates with adverse prognosis rather than immune control [26, 27].

To further quantify this biology, we identified 713 GSB-related genes through Cox and LASSO analyses, and refined them through Cox and LASSO analyses, arriving at a compact three-gene signature—CA12, SLC44A3, and BHLHA15—that consistently provided optimal prognostic performance in both TCGA and multiple GEO cohorts. Each component of the signature has biologically plausible relevance. BHLHA15, a transcription factor belonging to the basic helix-loop-helix (bHLH) family, hijacks the PTEN/AKT signaling pathway by directly regulating SNAI1, affecting anoikis resistance in cutaneous malignant melanoma cells [28]. SLC44A3, encoding a phosphatidylcholine transporter, has been implicated in drug sensitivity and resistance in hepatocellular carcinoma [29] and UVM [30]. CA12 belongs to the carbonic anhydrase family, which can catalyze the reversible hydration of carbon dioxide [31]. In UVM, CA12 has emerged as a prognostic biomarker associated with hypoxia [32] and metabolism [33]. Yin et al. found that hypoxia induces the upregulation of CA12, and knocking down CA12 significantly inhibits cell migration, colony formation, and proliferation [32], yet these findings, which were observed in a single UVM cell line (MUM2B), necessitate validation in other UVM cell lines. Notably, CA12-centered co-expression meta-analysis consistently highlighted enrichment of sphingolipid and glycosphingolipid pathways across five cohorts, suggesting that CA12 may be a key modulator linking hypoxia, GSL metabolism, and UVM progression. These findings warrant further mechanistic exploration in diverse UVM models.

The tumor microenvironment (TME) evolves dynamically during tumor progression, enabling cancer cells to evade immune surveillance and resist therapy [34]. In several cancer models, inhibiting glycosphingolipid (GSL) synthesis enhances antitumor immunity by strengthening NK- and CD8⁺ T-cell activity through IFNγ-dependent mechanisms [35], and eliglustat has been shown to increase tumor antigen visibility and improve CD8⁺ T-cell access [36]. However, UVM presents a distinct immune landscape. Histopathological and single-cell studies indicate that inflammatory infiltration—particularly CD8⁺ T cells, exhausted/regulatory T-cell subsets, and M2-polarized macrophages—is frequently associated with adverse clinical outcomes rather than effective immune control [26, 27]. Within this disease-specific context, the higher levels of CD8⁺ T-cell and macrophage infiltration observed in GBRS-high tumors are more likely to reflect an immunosuppressed, chronically inflamed microenvironment characterized by T-cell dysfunction and macrophage-driven immunosuppression. This UVM-specific “immune paradox” provides a plausible explanation for the coexistence of elevated GSB activity, increased immune infiltration, and poor prognosis. Future studies using single-cell transcriptomics and spatial immunophenotyping will be essential to clarify the functional states of these immune populations in relation to GSB pathway activity.

The associations between GBRS and predicted drug sensitivity should be interpreted with caution, as oncoPredict-derived IC50 estimates are based on cell line–trained transcriptomic models and are therefore exploratory rather than clinically validated. GBRS showed positive correlations with the IC50 values of several ERK-pathway inhibitors (e.g., ERK_2440, ulixertinib, VX-11e), consistent with the limited clinical benefit of monotherapy with MEK/ERK inhibitors such as selumetinib and ulixertinib in UVM [37, 38]. In contrast, negative correlations with compounds including ribociclib, KU-55,933, and PF-4,708,671 suggest that alternative, potentially actionable agents may warrant further investigation in GBRS-high tumors, although none have been systematically evaluated in UVM.

Mechanistically, several factors may link GSB activity to differential therapeutic response. GSLs regulate membrane lipid-raft organization, receptor clustering, and intracellular signaling, thereby modulating drug uptake, efflux, and downstream pathway activation [25, 39]. Because glycosphingolipids can mask tumor antigens and impair T-cell recognition [35, 36], GBRS-associated candidate agents may perform better when combined with tebentafusp, a gp100×CD3 TCR-bispecific fusion protein that remains the only systemic therapy with a Phase III overall-survival benefit in metastatic UVM [40]. Additionally, the upregulation of CA12 in GBRS-high tumors suggests an acidic, hypoxia-driven microenvironment, which may further influence drug distribution, target conformation, and immune effector function [41]. Together, these mechanisms provide a biologically coherent rationale for the observed drug–GSB associations, although all findings remain exploratory. Rigorous validation through gene-modulation experiments, pharmacologic assays in UVM models, and in vivo functional testing will be required to establish causality.

In conclusion, this study provides the first comprehensive evaluation of glycosphingolipid biosynthesis in UVM and establishes a robust prognostic framework for assessing tumor biology and the microenvironment. Despite strong performance in both the training and external validation cohorts, several limitations should be acknowledged. Our analyses rely on retrospective TCGA and GEO datasets, which are subject to sample-selection bias, technical variability, and incomplete clinical annotation, particularly with respect to treatment history and long-term outcomes. Furthermore, the modest size of available UVM cohorts limits subgroup analysis. Because GBRS is derived from bulk RNA-sequencing data, differences in sequencing platforms and preprocessing protocols may affect reproducibility and hinder immediate clinical implementation. Unmeasured confounders, including prior treatments and comorbidities, may also influence survival associations. Thus, before GBRS can be applied in routine practice, analytical optimization using clinically feasible platforms (e.g., qPCR-based assays) and prospective multi-center validation will be essential. Finally, the mechanistic roles of the GBRS genes—particularly the potential involvement of CA12 in glycosphingolipid metabolism—require further investigation through dedicated in vitro and in vivo studies.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Junjie Chen and Haoran Guo designed the research. Haoran Guo performed the statistical analysis and created the figures. Bai Li and Haoran Guo wrote the main manuscript text.Zhi Zhang contributed to the interpretation of the results and provided critical feedback on the manuscript. Zhi Zhang and Junjie Chen supervised the project, secured funding, and reviewed the manuscript. All authors discussed the results, contributed to the writing of the manuscript, and approved the final version.

Funding

The authors declare that no funds, grants, or other support were received during the preparation of this manuscript.

Data availability

The datasets analyzed during the current study are available in public repositories. The TCGA-UVM dataset can be accessed via the TCGA data portal (https://portal.gdc.cancer.gov/). The GEO datasets (GSE84976, GSE44295, GSE22138, and GSE39717) are available from the Gene Expression Omnibus (GEO) repository (https://www.ncbi.nlm.nih.gov/geo/).

Declarations

Ethics approval and consent to participate

The authors declare that the study did not involve human participants, human material, or human data.The authors declare that the study did not involve animals. This study utilized publicly available de-identified data from databases such as TCGA and GEO; therefore, no separate ethical approval was required.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.