Machine learning to construct sphingolipid metabolism genes signature to characterize the immune landscape and prognosis of patients with uveal melanoma

Background Uveal melanoma (UVM) is the most common primary intraocular malignancy in adults and is highly metastatic, resulting in a poor patient prognosis. Sphingolipid metabolism plays an important role in tumor development, diagnosis, and prognosis. This study aimed to establish a reliable signature based on sphingolipid metabolism genes (SMGs), thus providing a new perspective for assessing immunotherapy response and prognosis in patients with UVM. Methods In this study, SMGs were used to classify UVM from the TCGA-UVM and GEO cohorts. Genes significantly associated with prognosis in UVM patients were screened using univariate cox regression analysis. The most significantly characterized genes were obtained by machine learning, and 4-SMGs prognosis signature was constructed by stepwise multifactorial cox. External validation was performed in the GSE84976 cohort. The level of immune infiltration of 4-SMGs in high- and low-risk patients was analyzed by platforms such as CIBERSORT. The prediction of 4-SMGs on immunotherapy and immune checkpoint blockade (ICB) response in UVM patients was assessed by ImmuCellAI and TIP portals. Results 4-SMGs were considered to be strongly associated with the prognosis of UVM and were good predictors of UVM prognosis. Multivariate analysis found that the model was an independent predictor of UVM, with patients in the low-risk group having higher overall survival than those in the high-risk group. The nomogram constructed from clinical characteristics and risk scores had good prognostic power. The high-risk group showed better results when receiving immunotherapy. Conclusions 4-SMGs signature and nomogram showed excellent predictive performance and provided a new perspective for assessing pre-immune efficacy, which will facilitate future precision immuno-oncology studies.


Introduction
UVM accounts for 85% of all ocular melanomas and is the most common primary intraocular malignancy in adults (1).Approximately 85% of tumor cases arise from the choroid, with the remaining cases arising from the iris (3-5%) and ciliary body (5-8%) (2,3).More than 50% of patients with UVM develop systemic metastatic disease, with the liver being the most common site of metastasis in UVM (4).In addition, patients rarely achieve a good cure with surgery (5).As a result, patients with UVM have a very poor prognosis, with a 5-year mortality rate of 31% and a 15-year mortality rate of 45% (6).ICBhe diameter of the basal tumor, ciliary involvement and scleral expansion, non-random chromosomal aberrations and genetic mutations (e.g., BAP1 and SF3B1 mutations) are closely related to the prognosis of UVM (7)(8)(9) and are the main basis and foundation for prognostic grading, immunotherapy, radiotherapy and other treatment options.However, patients with the same clinical stage may have different clinicopathological features, suggesting that the prognosis of tumor patients based on traditional clinicopathological staging is not completely accurate (10,11).Therefore, to improve the quality of life of UVM patients, new prognostic biomarkers and molecular targets are needed to predict the prognosis of UVM patients and guide individualized treatment.
Sphingolipids are important components of biological membrane structure, maintaining the barrier function and fluidity of cell membranes (12).With the intensive study of sphingolipids in animals and yeast, sphingolipids and their metabolites have been found to be an important class of bioactive molecules, which are involved in regulating many important signaling processes such as cell growth, differentiation, senescence and programmed cell death (13).There is increasing evidence that sphingolipid metabolism is extensively involved in tumor proliferation, metastasis, angiogenesis and drug resistance, and plays a key role in the tumor immune microenvironment (14)(15)(16)(17)(18).In addition, it has been suggested that sphingolipids may be a potential tumorassociated antigen and are closely associated with tumor evolution and liver metastasis in UVM (19).Pelletier et al. found that UVM contains sphingolipids, which may be a target for monoclonal antibody therapy (20).With the continuous development of bioinformatics, biomarkers have been defined in various ways.Due to the unique role of sphingolipid metabolism in the tumor microenvironment, several studies have identified the potential of sphingolipidrelated genes or SMGs in the prognosis prediction of tumor tumors with a high degree of accuracy (21-24).Currently, the prognostic value of SMGs in UVM and the role of tumor immune microenvironment are unclear.Therefore, this study aimed to develop a novel SMGs-based approach to accurately predict prognosis and characterize the immune landscape of UVM patients.
In our study, we screened 4 reliable SMGs by machine learning, constructed a prognostic model based on the TCGA-UVM cohort, and went on to establish a risk score and comprehensively analyze the relationship between SMGs and immune microenvironment, immunotherapy, and chemotherapy sensitivity.We aimed to demonstrate the value of 4-SMGs for assessing the prognosis of UVM patients through a comprehensive analysis of genomic data, and to develop new tools to improve treatment options.

Method Patient data sources
We downloaded gene expression profiles and clinical data of TCGA-UVM cohort including 80 tumor patients from TCGA database (https://portal.gdc.cancer.gov/).The level 3 HTSeq-Fragments per kilobase million (FPKM) data of TCGA-UVM was converted to TPM (transcripts per million reads) according to the following formula: TPMn = FPKMn * 10 6 /(FPKM0 +… + FPKMm), where n represented gene n and m represented the total number of all genes, respectively.Then, we performed log 2based transformation of TPM.The sample size of UVM patients at the M stage and N stage varied greatly.These stages were consequently excluded from the analysis.The gene profiles and clinical data of 28 UVM patients in GSE84976 dataset were downloaded from the GEO database (https://www.ncbi.nlm.nih.gov/geo/).The GSE84976 was considered as an external validation dataset.

Consensus clustering analysis
To further elucidate the SMGs signature in UVM, all samples were divided into different clusters using "ConensusClusterPlus" R package (25).The "pheatmap" R package was used to show the differential expression and clinicopathological parameters of SMGs in different clusters.Gene set variation analysis (GSVA) analysis was performed using "c2.cp.kegg.v7.5.1.symbols.gmt"from the MSigDB database.Analysis of pathway differences across clusters using the "GSVA" R package (26).The single sample gene set enrichment analysis (ssGSEA) algorithm (27) was used to analyze the level of immune cell infiltration and the level of immune checkpoint expression between different clusters.

Model construction and validation
We obtained 97 SMGs (Supplementary Table 1) through the InnateDB portal (http://www.innatedb.com)(28).By performing univariate Cox regression analysis, we identified 27 genes associated with survival, followed by Least absolute shrinkage and selection operator (LASSO) regression analysis using 'glmnet' in R, with tenfold cross-validation to determine the optimal penalty parameter lambda.min=9.Nine genes were obtained.Support vector machine recursive feature elimination (SVM-RFE) is another machine learning method that uses the structural risk minimization principle while minimizing the empirical error as a way to improve learning performance (29).We used the SVM-RFE algorithm from the 'e1071' R package, with ten-fold cross-validation to obtain 13 valuable variables.Five public genes were extracted by Wayne diagram analysis, followed by using a stepwise multifactor COX regression model to identify and calculate the coefficients of the core genes.Finally, the risk signature of 4-SMGs was constructed.For each patient, the SMGs risk score was calculated as follows, risk score = Expression mRNA1 × Coef mRNA1 + Expression mRNA2 × Coef mRNA2 +… Expression mRNAn × Coef mRNAn .

Model formula
All UVM patients were given risk scores based on output model equations, and median value were calculated using the R package "survminer", classifying all UVM patients into low-risk and high-risk groups, and plotting survival curves for the two subgroups.The R package "pec" was adopted to calculate the Cindex.For assessing genetic traits' predictive power, receiver operating characteristic curve (ROC) curve analysis using the "time-ROC" R package was conducted.Decision curve analysis (DCA) of a multi-factor Cox regression model was plotted using the "ggDCA" R package.

Independent prognostic analysis and nomogram construction
We conducted univariate and multivariate Cox regression analyses to assess risk score as an independent prognostic factor.Using the "rms" R package, histograms were constructed using risk scores versus clinicopathologic characteristics to predict survival for patients in TCGA-UVM cohort.

Functional enrichment analysis
Through functional enrichment analysis of differentially expressed genes in UVM associated with SMGs, functional annotation and enrichment pathways have been explored.The analysis of Gene Ontology (GO) pathways was done using the "ClusterProfiler" R package, where P-value < 0.05 represents a statistically significant difference.GSVA was performed using "c2.cp.kegg.v7.5.1.symbols.gmt"from the MSigDB.Using "GSVA" R package to perform GSVA enrichment analysis.The "heatmap" R package was used to create heat maps.According to the "limma" R package, an adjusted P-value < 0.05 indicates statistical significance for subgroup differences.

Immunity analysis of the risk signature
Currently accepted methods, including XCELL (30,31), TIMER (32,33), QUANTISEQ (32,33), MCPCOUNT (34), EPIC (35), CIBERSORT (36) and CIBERSORT-ABS (37) were used to measure immune infiltration scores.Spearman correlation analysis was used to examine the correlation between immune cells and risk scores.Based on the immune cell characteristics of UVM patients, the ssGSEA method was adopted to differentiate patients at low-risk from those at highrisk.Using a list of 20 suppressive immune checkpoints derived from Auslander's study, we assessed the suppression of immune checkpoints between high-risk and low-risk groups (38).The "estimate" R package was used to calculate the immunological and mechanistic scores of the specimens from the RNA-seq data to assess the purity of the tumors.Evaluation and visualization of immunotherapy efficacy in UVM patients by "limma" and "ggpubr" R package.
Xu et al. developed a website that provided us with gene sets related to cancer and immunity (39) (http://biocc.hrbmu.edu.cn/TIP/) and a set of genes positively associated with anti-PD-L1 drug response was obtained from Mariathasan's study features (40).The R package "ggcor" for the analysis of correlations between risk scores and the two genetic traits mentioned above was used.ImmuCellAI (http://bioinfo.life.hust.edu.cn/ImmuCellAI) ( 41) is a portal that predicts tumor immune infiltration estimates and immunotherapy response, and we obtained corresponding data on immunotherapy in UVM patients.

Drug sensitivity
The "pRRophetic" R package was used to assess treatment response in high-risk and low-risk groups of patients, as determined by the half-maximal inhibitory concentration (IC50) of each UVM patient on the Genomics of Drug Sensitivity in Cancer (GDSC) (https://www.cancerrxgene.org/)(42).

TISCH analysis
A single-cell RNA sequencing database focused on TME is housed at the Tumor Immunization Single Cell Center (TISCH).Detailed cell type annotations are provided at the single cell level for further analysis of specific gene expression in different cell types.The specific gene expression in different cell types further reveals the variation of TME in patients with different UVM, thus explaining to some extent the heterogeneity of UVM.

Statistical analysis
Statistical analyses were performed using R software v4.1.3.Kaplan-Meier (KM) survival curves and log-rank test were used to compare Overall Survival (OS) between high-and low-risk groups.LASSO regression analysis and SVM-RFE for screening candidate SMGs.Stepwise multi-factor cox regression analysis was used to construct SMGs signature.Time-dependent ROC was used to evaluate the predictive performance of the model.Spearman correlation analysis was used to evaluate the correlation between risk score and immune cell infiltration.Wilcox test was used to compare the proportion of TIICs, immune checkpoints, and immune function between the two groups.P-values <0.05 were considered statistically significant and false discovery rate (FDR)<0.05 was considered statistically significant.

Result Consensus clustering identified the molecular subtypes of SMGs
The graphical flow chart outlines the main design of this study (Figure 1).We considered that the increasing trend of the cumulative distribution function (CDF) values relative to the consensus index indicated the presence of appropriate classification, and based on the CDF curve and the Delta area, k = 2 proved to be the best point to obtain the maximum difference between clusters when the clustering index "k" increased from 2 to 9, so we divided the UVM patients into two subgroups (Figures 2A, B).In addition, the consensus matrix is naturally a better visualization tool that can help to assess the composition and number of clusters.We plotted the color-coded heat map corresponding to the consensus matrix and found that it exhibited high intra-group correlation and low inter-group correlation when k=2, which strongly suggests that it is very appropriate to classify UVM patients into two subtypes (Cluster A and Cluster B) (Figure 2C).
To determine the difference in survival prognosis of patients in different clusters, the difference in OS between clusters was calculated according to the ClusterSurvival R package.We found an improved survival prognosis in cluster A compared to patients in cluster B (P<0.001) (Figure 2D).Principal component analysis (PCA) is often used to visualize the distribution of risk in different populations.Cluster A and Cluster B patients showed significant differences when based on different clusters (Figure 2E).In addition to this, we further explored the metabolic differences between Cluster A, B and SMGs, and the heat map showed that Cluster B had higher expression differences and clinical characteristics in SMGs (Figure 2F).To elucidate potential biological pathways, we performed enrichment analysis of different cluster samples using the Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway database and identified relationships with various cancer-related pathways, such as apoptosis, transporters, cell adhesion molecules and hematopoietic cell lineages (Figure 2G).Based on the fact that immunotherapy plays an important role in the treatment of tumors, to understand the distribution and correlation of the relative content of 23 TIICs (tumor-infiltrating immune cells) in this cohort, we calculated the level of immune cell infiltration in both clusters by the ssGSEA algorithm.It was found that there was a higher level of infiltration in most immune cells in Cluster B compared to Cluster A (Figure 2H).According to the results, the SMGs risk score model can classify different immune subtypes and thus influence the response to immunotherapy.Furthermore, due to the importance of immune checkpoints for the effectiveness of tumor immunotherapy and the fact that immune checkpoints are one of the important features of tumor microenvironment (TME).We explored the differences in immune checkpoint expression between the two groups and ultimately found that immune checkpoint gene expression was significantly upregulated in patients with Cluster B. Based on the above analysis, we concluded that Cluster B has higher effectiveness and sensitivity for immunotherapy (Figure 2I).

Construction and validation of 4-SMGs signature.
By univariate cox regression analysis, we obtained a total of 27 SMGs significantly associated with OS in UVM patients (Supplementary Table 2).We used two machine learning methods to identify candidate SMGs.For the LASSO regression analysis, 9 candidate genes were screened from the 27 SMGs with significant prognostic features (Figures 3A, B).For the SVM-RFE algorithm, the error was minimized when the number of features was 13 (Figures 3C, D).Five intersecting genes were obtained by Wayne diagrams for the above two methods (Figure 3E).Finally, 4 SMGs were identified as independent prognostic factors by the stepwise multivariate Cox analysis, including ARSH, GBA2, GLA and GLB1.Prognostic index (PI) = (-0.701*ARSHexp) + (-3.988*GBA2 exp) + (5.464*GLA exp) + (-2.985*GLB1 exp).We further explored the correlation between the expression of these 4-SMGs and risk scores and found that all 4-SMGs were closely associated with risk scores.Among them, ARSH, GBA2 and GLB1 had a significant negative correlation with the risk score, while GLA had a significant positive correlation with the risk score (Figure 3F).In addition, we calculated the prognostic risk score for each patient and divided the UVM patients into highrisk and low-risk groups based on the median score of score.(Figure 3G) demonstrates the distribution of 4-SMGs in the high-and low-risk groups, with ARSH, GBA2 and GLB1 being low expressed in patients in the high-risk group and GLA being highly expressed in high-risk patients.In the TGCA-UVM cohort, mortality in UVM patients increased with increasing risk (Figures 3H, I), with a better prognosis in the low-risk group (P<0.001)(Figure 3J).Time-dependent ROC curves were used to assess the accuracy of the model developed to predict OS in UVM patients.The time-dependent ROC curves showed 1-year The flowchart summarizes the main design of the present study.
AUC of 0.749, 2-year AUC of 0.875, 3-year AUC of 0.906, 4-year AUC of 0.927, and 5-year AUC of 0.925 (Figure 3K).PCA is often used to visualize the distribution of risk in different populations.When based on a risk model, high-and low-risk patients showed significant differences and showed a clear separation (Figure 3L).In the GSE84976 cohort, we demonstrated the same results as in the TCGA-UVM cohort.mortality in UVM patients increased with increasing risk (Figures 3M, N).KM survival analysis showed that low-risk patients had a better prognosis compared to high-risk patients (P<0.001)(Figure 3O).The time-dependent ROC curve showed a 2-year AUC of 0.783, a 3-year AUC of 0.789, a 5-year AUC of 0.893, an 8-year AUC of 0.929and a 10-year AUC of 0.927 (Figure 3P).PCA analysis showed that low-and high-risk patients exhibited significant differences, showing a clear separation (Figure 3Q).Based on these results, we can conclude that the construction of our prognostic model is quite superior.

Clinical correlation and survival analysis of SMGs in UVM patients
To analyze the correlation between high-and low-risk groups and clinical characteristics, heat maps were drawn based on the expression of clinical characteristics, risk scores, and 4-SMGs, the heat maps showed the association between the 4-SMGs identified in the prognostic risk model and the age, gender, clinical stage, T stage, and risk scores of all UVM patient samples in TCGA (Figure 4A).In addition, we further analyzed the difference in the proportion of patients with various clinicopathological characteristics between the high-risk and low-risk groups and found that SMGs had a significant impact on the proportion of patients with different clinicopathological characteristics (Figures 4B-E).To better understand whether the prognosis of patients in different clinical subgroups differed, a clinical analysis was performed on the entire sample subgroup, all samples were divided into different subgroups by age (>65 and ≤65 years), gender (male and female), clinical stage (II and III-IV) and T stage (T2 and T3-4) for further survival analysis.Survival times were significantly shorter in high-risk patients than in low-risk patients in all subgroups (Figures 4F-M).The currently identified risk model for SMGs also seems to be able to reliably predict the prognosis of certain subgroups of UVM based on their clinical characteristics.

Establishment of nomograms in combination with clinical characteristics
Considering the strong correlation between the constructed risk model and poor prognosis, we combined the OS of UVM patients and their clinical characteristics in univariate and multivariate Cox analyses to determine whether our prognostic characteristics constructed based on the 4-SMGs could be used as independent predictors of prognosis.Based on the results of univariate analysis of variables, age, stage, gender, T, and risk score, risk score was found to be significantly associated with prognosis in patients with UVM (P<0.001) (Figure 5A).Similarly, risk score remained the most reliable and independent predictor in the cohort after multifactorial analysis (P<0.001) (Figure 5B).To extend the clinical application and usability of the constructed risk model, we constructed Nomogram plots based on age, gender, clinical stage, T stage, and risk score as a predictor of 1-, 2-, and 3-year prognostic survival probabilities in patients with UVM.As a result, it was observed from the model results that the risk score had the greatest impact on predicting OS, indicating that the SMGs-based risk model could better predict the prognosis of UVM (Figure 5C).The calibration curves also showed a more satisfactory consensus between the predicted and observed values in terms of the probability of OS at 1-, 2-and 3-years (Figure 5D).Comparing Nomogram, risk and common clinicopathological features again, risk (AUC=0.917),as well as Nomogram (AUC=0.857),were better predictors of UVM prognosis than age, gender, clinical stage and T stage (Figure 5E).This implies that the predictive performance of our constructed SMGs signature was significantly better than other clinical features.

Functional enrichment analysis of SMGs
KEGG enrichment analysis and GO functional analysis were performed to assess differential genes in UVM to elucidate the relevance of bioactivity and signaling pathways to risk scores.The threshold FDR<0.05 and P<0.05 were used to select significantly enriched items (Figures 6A, B; Supplementary Table 3).Biological processes (BP) mainly included such as lymphocyte and T-cell differentiation, leukocyte adhesion, etc.The cellular component (CC) mainly included the plasma membrane signal transduction receptor complex, synaptic membrane intrinsic components and plasma membrane signal transduction receptor complex.6C; Supplementary Table 4), and among low-risk individuals, cystine and methionine metabolism, regulation of autophagy, and cancerrelated pathways were enriched, while in the high-risk group, pathway enrichment mostly involved immune function, including natural killer cell-mediated cytotoxicity, T-cytokine cytokine receptor interaction, sterol hormone biosynthesis and other related KEGG pathways.In summary, we were surprised to find a strong correlation between enrichment analysis results and immune response, and therefore we conducted a systematic analysis of the immune landscape in the UVM patients.Functional enrichment analysis of SMGs in TCGA-UVM.(A, B) Gene Ontology (GO) enrichment analysis was used to analyze the differential genes between UVM and normal samples.The x axis represents the ratio of genes associated with the term, and the y axis represents the pathway term.The p value of each item is colored according to the legend.(C) GSVA analysis between the high-risk cohort and the low-risk cohort with KEGG.(43).And TIICs are an important component of the TME, and their composition and distribution are closely related to tumorigenesis and development (44).First, we explored the correlation between risk score and infiltrating immune cell abundance according to XCELL, TIMER, QUANTISEQ, MCPCOUNTER, CIBERSORT, CIBERSORT-ABS and EPIC algorithms, where the CD8+ T cell, NK cell infiltration were all positively correlated with the risk score (Figure 7A).We then independently assessed immune cell infiltration between high and low-risk groups in UVM using the CIBERSORT algorithm and showed that the expression of plasma cells, B cells naïve, Monocytes, and Mast cells resting was significantly higher in the high-risk group (Figure 7B).Given the importance of checkpoint-based immunotherapy, we analyzed the expression of immune checkpoint genes in the high-and low-risk groups.Most of the immune checkpoint genes were found to be significantly upregulated in the high-risk group, including IDO1, CTLA-4, TIGIT, KIR3DL1, BTLA, CD28, etc. (Figure 7C), suggesting that patients in the high-risk group may have better efficacy with ICB therapy.Because immune cells with immune checkpoints can significantly affect immune function, we performed a comparison of ssGSEA scores for immune function, and multiple immune function scores were significantly greater in the high-risk group than in the low-risk group (Figure 7D).As infiltrating immune cells are an important component and one of the characteristics of the TME, alterations in the expression of immune cell types lead to changes in TME composition, so we analyzed the TME composition of UVM samples.The results showed that the immune score (P<0.001),ESTIMATE score (P<0.001) and Stromal Score (P<0.001) were lower in the low-risk group compared with the high-risk group, indicating that the overall immune level and immunogenicity of the TME were higher in the high-risk group (Figure 7E).More importantly, we obtained immunotherapy response outcomes for UVM patients through the immuneAI portal and found that patients with higher risk scores were more likely to benefit from immunotherapy (Figure 7F), and had a better predictive performance for 4-SMGs signature (Figure 7G).Since ICB response plays an important role in immune checkpoint therapy, we further analyzed the correlation between risk score and ICB response signature (Figures 7H, J) and found that among them, Systemic lupus erythematosus, Base excision repair, p53 signaling pathway, Proteasome, and Cytokinecytokine receptor interaction were significantly positively correlated, while significantly negatively correlated with Alcoholism and Spliceosome.The correlation analysis between risk score and tumor immune cycle steps was also performed, and it was found that only MDSC recruiting, Th2 cell recruiting, and Monocyte recruiting (step 4) were not significantly correlated with risk score, while other immune cycle steps were positively correlated with our risk score (Figures 7I, K).

4-SMGs signature predicts chemotherapy sensitivity
Based on the risk score, we further evaluated the potential value of 4-SMGs in predicting chemotherapy sensitivity through the GDSC database to enhance accurate drug therapy.A total of 51 chemotherapeutic or targeted agents with significant differences in chemosensitivity between high and low risk were analyzed by the "pRRophetic" R package (Supplementary Figure 1).Figure 8 showed nine common inhibitors or drugs.KRAS (G12C) Inhibitor-12, Daporinad, Selumetinib, Telomerase Inhibitor IX, Trametinib, Uprosertib and Vincristine had relatively high IC50 in the high-risk group.In contrast, Rapamycin and Rapamycin had lower IC50s in the high-risk group.

Correlation of SMGs with tumor microenvironment
We used the single-cell dataset UVM_GSE139829 from the TISCH database to analyze the expression of 4-SMGs in the immune microenvironment.There were 31 cell populations and 8 immune cell types in the UVM_GSE139829 dataset (Figures 9A, B), showing the distribution and number of various cell types (Figures 9C, D).Expression levels of each SMGs in immune cells were barely expressed in ARSH in the immune microenvironment (Figure 9E), GBA2, GLA, and GLB1 were expressed in various immune cells (Figures 9F-H).GLA was mainly expressed in Mono/Marco.

Discussion
Despite its low incidence, UVM is still the most common primary intraocular malignant tumor and is known for its high metastasis, high malignancy, and high mortality (45).Surgical removal is the first-line treatment of UVM.Although it has a great impact on the appearance and psychology of patients, the prognosis of UVM patients after surgery is also unsatisfactory, and the five-year survival rate is only 17%-53% (46).In addition, the 5-year and 10-year metastasis rates of UVM are about 25% and 34%, respectively.The mortality rate of UVM within 1 year after metastasis is 80%.Most patients with metastatic UVM have a survival time of 6-12 months after diagnosis, and there are almost no relevant management strategies and treatment measures for metastatic UVM (47)(48)(49).Therefore, the early diagnosis and treatment of UVM are particularly important to improve the prognosis of patients.The diagnosis and prognosis prediction of UVM is based on clinical manifestations and histopathological evaluation, but they are not enough to judge the heterogeneity and development trend of tumors.At the same time, they have limitations in predicting the prognosis and treatment response of UVM patients, while molecular prognosis prediction methods have better performance (50).Sphingolipid metabolism is a highly regulated intracellular process that controls the synthesis and degradation of bioactive lipids, including ceramide and sphingosine-1phosphate (51).A large number of studies have explored the relationship between sphingolipids and cancer.We noted that gangliosides in sphingolipids were closely associated with liver metastasis and host immune response in a human UVM nude mouse model and that anti-gangliosides antibodies could inhibit the spread of tumor cells in nude mice (52,53).In addition, sphingolipid metabolites promote tumor progression by promoting cell proliferation and stimulating chemotactic migration and invasion, such as SP1 in ovarian cancer (54)(55)(56)(57).This is exciting and may imply that sphingolipids play an important role in UVM proliferation, metastasis, and phenotype and can be used as a target for monoclonal antibodies and immunotherapy.Unfortunately, there are still gaps in this research.Given the great potential of sphingolipids for UVM, we used mRNA expression data from the TCGA-UVM dataset to identify important prognostic genes and constructed a multibiomarker prognostic model based on SMGs.Our results suggest that SMGs-based signatures can be used for risk stratification, prognosis prediction and immunotherapy efficacy assessment of UVM, thus providing a valuable reference for individualized treatment.
In this study, we integrated the expression profiles of 97 SMGs in the TCGA-UVM dataset and selected four genes (ARSH, GBA2, GLA and GLB1) to construct a new prognostic model for SMGs by using Lasso regression analysis, SVM-RFE, and stepwise multivariate COX regression analysis.The SMGs signature we constructed proved to be an independent prognostic factor for UVM.UVM patients were divided into high and low risk groups according to the median risk score, and significant prognostic differences were found between the two groups.In other reports, non-lysosomal b-glucosidase (GBA2) hydrolyzed GlcCer to release ceramide and glucose on the cytoplasmic side of ER and golgi membrane, potentially affecting tumor drug resistance (58, 59).Alpha-galactosidase (GLA) has long been considered a key link in Fabry's disease ( 60), but has rarely been studied in tumors, which seems to provide a way to link Fabry's disease and tumors.Lysosomal bgalactosidase (GLB1) hydrolyzed b-galactosidase from a sugar conjugate, representing the origin of the aging-associated bgalactosidase SA-SS-GAL, was reported to be a reliable biomarker of aging in prostate cancer (61).However, other studies have shown that the expression of lysosomal galactosidase b-1 (GLB1) is not associated with aging and is active in some types of cancer cells (62).The effect of GLB1 on various types of tumors is still inconclusive, and a large number of related mechanism studies are needed to elucidate.ROC curve and calibration curve analysis proved the outstanding predictive performance of SMGs signature.In order to expand the predictive ability of SMGs signature and prove its practical value in assessing the prognosis of UVM patients, we constructed a nomogram based on clinical factors and risk scores and found that SMGs signature has better predictive efficacy than clinicopathological features, which can provide a basis for clinicians to make decisions.
In order to have a comprehensive and clear demonstration of SMGs signature function, we carried out enrichment analysis of action pathways.The results showed that SMGs signature was associated with many immune or tumor related pathways, such as monocyte differentiation, lymphocyte differentiation, T cell differentiation, regulation of T cell activation, cell adhesion, cytokine and its receptor activity, etc.The cytotoxicity of lymphocytes to UVM cells with hematologic metastasis may imply that SMGs can promote lymphocyte differentiation to prevent UVM from migrating along the blood.Lentz et al. found that monocyte and melanocyte destruction invaded the uveal tract in UVM patients (63), suggesting that SMGs are associated with primary changes in UVM.UVM cells express PD-L1 to regulate T cell function by inhibiting IL-2 production, thus achieving immune escape (64).SMGs may activate T cell function, thereby achieving better efficacy of PD-1/PD-L1 targeted therapy.The enrichment of various adhesion pathways suggests that SMGs play an important role in UVM metastasis.We found many interesting results when we performed KEGG enrichment analysis.The low-risk group is enriched in the autophagy regulatory pathway, but the specific mechanism of this aspect is lacking.Given the importance of autophagy for tumor behavior and treatment, this represents a huge research prospect.UVM is protected from complement mediated cleavage in vitro and in vivo by expressing three types of CRP (CD46, CD55, and CD59) (65), while the high-risk group is enriched in complement related pathways.Therefore, it is hypothesized that patients in the high-risk group may prevent the immune escape of UVM by stimulating complement, but experimental verification is needed, which will be our next research direction.
The TME plays a key role in the occurrence, progression, metastasis and treatment resistance of tumors (66).Immune components in the TME play a key role in promoting immune escape and inflammation formation (67).A deeper understanding of immune infiltration in TME is essential to reveal the underlying molecular mechanisms and provide novel immunotherapy strategies to improve clinical outcomes (68)(69)(70)(71).Hence, We analyzed the immune cell infiltration in the high -  In addition, we also noted that mast cell infiltration was significantly higher in the low-risk group than in the high-risk group.Mast cells have many effects on tumors, including mast cell-mediated cytotoxicity, mast cell-directed angiogenesis, tissue remodeling in the adjacent environment, and immune cell recruitment (75).Some studies have suggested that mast cells have a great influence on tumor angiogenesis in UVM (76).However, the actual role of mast cells in UVM and the specific mechanism still need a lot of experiments to explain.The highrisk group has higher infiltration of CD8+T cells.It has been found that the major histocompatibility complex II (MHC II) UVM vaccine can activate tumor-specific CD8+T cells and CD4 +T cells, thereby killing tumor cells but does not affect normal cells, suggesting that the high-risk group receiving the MHC II UVM vaccine may have better efficacy (77).DC cocultured with UVM cells showed reduced expression of CD1a and CD83, which failed to activate T cells for immune response, revealing the possibility of tumor-pulsed dendritic cell vaccines as therapeutic measures for UVM (78).In addition, the degree of infiltration of resting dendritic cells in the high-risk group was higher than that in low-risk group, but there was no significant difference in activated dendritic cells, which may imply that UVM in high-risk group can change the status of dendritic cells to achieve immune escape.
With the increasing attention to tumor immune checkpoint, immunotherapy based on immune checkpoint inhibitors also has a good performance in clinical practice.But for UVM, immune checkpoint therapy has little effect (79).Some studies have pointed out that the retrospective data show that the response rate is very low (80).Therefore, there is an urgent need to analyze the expression of immune checkpoint genes in patients to identify patients who can benefit from immune checkpoint therapy.We found that the expression of most immune checkpoints was higher in the high-risk group compared with the low-risk group, suggesting that the high-risk group is more likely to benefit from ICB therapy.However, due to the rarity of UVM, the sample size limits the applicability and generality of almost all relevant studies, and there is still a need to increase the sample and conduct randomized prospective trials (81).In general, SMGs model can accurately identify UVM patients who are sensitive to ICB therapy, which has guiding significance for clinicians' immunotherapy strategy.Given the influence of immune cells and immune checkpoints on immune function, we further explored the differences in immune function between high and low-risk groups.We found that HLA expression was higher in the high-risk group than in the low-risk group.It is generally assumed that low HLA expression will allow tumor cells to escape from CTL-mediated lysis, resulting in metastasis.However, a large number of studies have shown that the role of HLA in UVM is contrary to that of other common tumors.UVM cells reduce the expression of HLA antigens during liver metastasis through the blood route, thereby escaping the killing effect mediated by NK cells and cytotoxic T cells (82)(83)(84).This provides a new direction for the treatment of high-risk patients by increasing the expression of HLA antigens in UVM cells.Interestingly, we also found that the type II interferon response was also enhanced in the high-risk group.Bosch et al. developed an MHC II -matched vaccine that induces type II interferon secretion, thereby allowing the generation of a CD8 cell immune response in the eye (85) and may be more effective in high-risk patients.
Although our study has greater clinical implications for the prognostic assessment and selection of treatment options for patients with UVM, our study still has some limitations.First, our study is a retrospective study that needs to be validated in future prospective studies.Second, the mechanisms by which SMGs affect the prognosis of UVM patients need to be explored in more in vivo and in vitro experiments.Finally, the predominant race in the TCGA-UVM cohort was white, with a lack of data on Asians or blacks.This makes it critical to include other ethnic groups in future studies.
Technology Department of foreign (border) high-end talent introduction project (No: 23RCYJ0044).

2
FIGURE 2 Consensus clustering identified the molecular subtypes of SMGs.(A) Consensus clustering CDF for k = 2 to 9. (B) Relative change in area under the cumulative CDF curve for k = 2 to 9. (C) Consensus matrix for k = 2. (D) The different OS between Cluster A and Cluster B. (E) PCA plot .(F) Relationships between SMGs expression and clinicopathological parameters.(G) KEGG enrichment analysis of different clusters.* P < 0.05, ** P < 0.01, *** P < 0.001, ns>0.05.(H) Immune cell scores between Cluster A and Cluster B. (I) Different expressions of immune checkpoints between Cluster A and Cluster B.

3
FIGURE 3 Construction and validation of 4-SMGs Signature.(A) Ten-time cross-validation for tuning parameter selection in the LASSO model.(B) LASSO coefficient profiles.(C, D) Biomarker signature gene expression validation by SVM-RFE algorithm selection.(E) Two algorithmic venn diagram screening genes.(F) Correlation of 4-SMGs with riskscore.(G) Heatmap of risk factors in high-and low-risk patients.(H) Distribution of risk scores between low-and high-risk groups in the TCGA cohort.(I) Survival status of UVM patients in the low-and high-risk groups in the TCGA cohort.(J) KM curve compares the overall UVM patients between low-and high-risk groups in the TCGA cohort.(K) Time-dependent ROC curves analysis in the TCGA cohort.(L) PCA plot in the TCGA cohort.(M) Distribution of risk scores between low-and high-risk groups in the GEO cohort.(N) Survival status of UVM patients in the low-and high-risk groups in the GEO cohort.(O) KM curve compares the overall UVM patients between low-and high-risk groups in the GEO cohort.(P) Time-dependent ROC curves analysis in the GEO cohort.(Q) PCA plot in the GEO cohort.

4
FIGURE 4 Clinical correlation and survival analysis of SMGs in patients with UVM (A) heatmap showing the associations between the high-risk and low-risk score of patients with UVM and UVM related clinical characteristics in the dataset from TCGA.(B-E) The percentage Stacked histogram of clinical characteristics.The risk score based on 4-SMGs signature is a valuable marker for poor prognosis in various subgroups divided by clinicopathological characteristics.The SMGs could distinguish high-risk patients in a variety of subgroups divided by clinicopathological characteristics including (F, G) T stage, (H, I) clinical stage, (J, K) gender and (L, M) age.

5
FIGURE 5 Establishment of nomograms in combination with clinical characteristics.(A) Univariate and (B) multivariate COX regression analysis of the signature and different clinical feature.Green squares: univariate cox analysis of HR values for each variable.Red squares: multivariate cox analysis of HR values for each variable.(C) A nomogram combining the SMGs, age, gender, clinical stage, and T stage.(D) The calibration curve of the constructed nomogram of 1-year, 2-year and 3-year survival.(E) The nomogram's time-dependent ROC curves.

7
FIGURE 7 SMGs risk score predicts TME and immune cell infiltration.(A) Immune cell bubble map.(B) Differences in immune cell infiltration between high-and low-risk groups.(C) Immune checkpoint differences between high and low risk groups.(D) Immune cell and immune function ssGSEA scores between high-and low-risk groups.(E) The TME component analysis.(F) 4-SMGs predict immunotherapy response outcomes in patients.(G) Excellent predictive performance of ROC curve labeling model (H) Correlation between risk score and ICB response signature.(I) Correlation of risk scores with each step of the tumor immune cycle.(J) The plot of the difference in enrichment scores between the highrisk and low-risk groups on the immunotherapy prediction pathway.(K) The plot of differences between the high-risk and low-risk groups on each step of the cancer-immune cycle.* P <0.05; ** P <0.01; *** P <0.001.

9
FIGURE 9 Correlation of SMGs with tumor microenvironment.(A-D) Annotation of all cell types in UVM_GSE139829 and the percentage of each cell type.Percentages and expressions of (E) ARSH, (F) GBA2, (G) GLA and (H) GLB1.
(73,74)-risk groups.Activated naive B cells have been shown to promote Th1 polarization, thereby preventing further tumor growth (72).Plasma cells are the end functional state of the Bcell lineage and are involved in solid tumor progression.A large number of literature have shown that tumor-infiltrating plasma cells have a positive prognostic effect on cancer(73,74).The increased expression of naive B cells and plasma cells in the lowrisk group may indicate a better prognosis in the low-risk group.