Machine learning-based integrative analysis of methylome and transcriptome identifies novel prognostic DNA methylation signature in uveal melanoma

Ping Hou†, Siqi Bao†, Dandan Fan†, Congcong Yan, Jianzhong Su, Jia Qu and Meng ZhouImage

Corresponding authors: Meng Zhou, School of Ophthalmology & Optometry and Eye Hospital, School of Biomedical Engineering, Wenzhou Medical University, 270 Xueyuan Road, Wenzhou, Zhejiang 325003, P. R. China. E-mail: [email protected]; Jia Qu, School of Ophthalmology and Optometry, Eye Hospital, Wenzhou Medical University, 270 Xueyuan Road, Wenzhou, Zhejiang 325003, P. R. China; State Key Laboratory of Ophthalmology, Optometry and Vision Science. 270 Xueyuan Road, Wenzhou, Zhejiang.


Uveal melanoma (UVM) is the most common primary intraocular human malignancy with a high mortality rate. Aberrant DNA methylation has rapidly emerged as a diagnostic and prognostic signature in many cancers. However, such DNA methylation signature available in UVM remains limited. In this study, we performed a genome-wide integrative analysis of methylome and transcriptome and identified 40 methylation-driven prognostic genes (MDPGs) associated with the tumorigenesis and progression of UVM. Then, we proposed a machine-learning-based discovery and validation strategy to identify a DNA methylation-driven signature (10MeSig) composing of 10 MDPGs (AZGP1, BAI1, CCDC74A, FUT3, PLCD1, S100A4, SCN8A, SEMA3B, SLC25A38 and SLC44A3), which stratified 80 patients of the discovery cohort into two risk subtypes
with significantly different overall survival (HR = 29, 95% CI: 6.7–126, P < 0.001). The 10MeSig was validated subsequently in an independent cohort with 57 patients and yielded a similar prognostic value (HR = 2.1, 95% CI: 1.2–3.7, P = 0.006).

Multivariable Cox regression analysis showed that the 10MeSig is an independent predictive factor for the survival of patients with UVM. With a prospective validation study, this 10MeSig will improve clinical decisions and provide new insights into the pathogenesis of UVM.

Key words: uveal melanoma; machine-learning; DNA methylation

Ping Hou is a graduate student at the School of Biomedical Engineering, Wenzhou Medical University. Her research interests include bioinformatics and cancer epigenetics.

Siqi Bao is a graduate student at the School of Biomedical Engineering, Wenzhou Medical University. Her research interests include bioinformatics and translational medicine.

Dandan Fan is a graduate student at the School of Biomedical Engineering, Wenzhou Medical University. Her research interests include bioinformatics and cancer epigenetics.

Congcong Yan is a graduate student at the School of Biomedical Engineering, Wenzhou Medical University. Her research interests include bioinformatics and disease systems biology.

Jianzhong Su is a Professor at the School of Biomedical Engineering, Wenzhou Medical University. His research interests include bioinformatics and cancer epigenetics.

Jia Qu is a Professor and the Director at the School of Optometry and Ophthalmology and Eye Hospital, Wenzhou Medical University. His research interest is ophthalmology.
Meng Zhou is an associate professor at the School of Biomedical Engineering, Wenzhou Medical University. His research interests include bioinformatics and computational precision medicine.


Uveal melanoma (UVM) is the most common primary intraocu- lar malignancy in adults and represents a matter of ∼85% of all ocular melanomas [1]. The melanoma arises from melanocytes existing in the uveal tract and most commonly derives from choroidal melanocytes (85–90%), but can also derive from the iris (3–5%) and the ciliary body (5–8%) [2, 3]. The annual incidence of UVM in Europe and the USA is ∼6 per million populations per year [4]. In China, the annual incidence is second only to retinoblastoma, ranking second in intraocular tumors. UVM metastasizes to the liver in ∼50% of patients due to a lack of effective treatment [5]. The prognosis for patients with liver metastases is poor, and median survival is less than six months [6]. Therefore, it urgently needs to improve our understanding of molecular mechanisms and cellular biology of UVM and identify prognostic factors at the molecular level.

DNA methylation is a crucial epigenetic modification of the genome and plays essential roles in regulating various cellular processes, including embryonic development, differentiation, X- chromosome inactivation, genomic imprinting [7–9]. Increasing evidence revealed tissue- and cell-type-specific effects of DNA methylation in the regulation of gene expression [10–12]. Vari- ation in DNA methylation has been widely observed in many human diseases, including autoimmune diseases, metabolic dis- orders and cancers, and has been recognized as a hallmark of cancers [13–15]. Recent advances in high-throughput arrays have facilitated the characterization of methylome in multiple cancer types and revealed that aberrant DNA methylation is a common event in tumorigenesis of most human cancers and is associated with cancer development, prognosis and treatment response [16–21].

The epigenetic alterations, including non-coding RNA expression, hyper-methylation of genes and histone modi- fication, have also been implicated in UVM tumorigenesis, progression and metastasis [22–24]. For example, TIMP3 expres- sion, which is regulated by methylation, plays an essential role in UVM development [25]. BAP1 loss correlates with a global DNA methylation state from UVM different subtypes [26]. RASSF1a methylation is a common epigenetic event during UVM development [27]. These studies implied the potential use of aberrant epigenetic patterns as novel biomarkers for UVM diagnosis and prognosis. However, genome-wide DNA methylation and their clinical utilization have not been fully investigated in UVM precision medicine. Here, this study aimed to explore the clinical utility of methylome and transcriptome for UVM prognosis prediction and further identify DNA methylation-driven biomarkers using the machine learning method.

Materials and Methods

Sample datasets used in the study Clinical information, methylation (Illumina HumanMethyla- tion450 BeadChip, v.2017-12-13) and transcriptome data (RNA- Seq, v.2017-09-14) of 80 UVM cases generated by The Cancer Genome Atlas (TCGA) were downloaded from UCSC Xena (https: //gdc. Other independent 57 UVM cases from GSE44295 (Zhang’s cohort) (Illumina HumanRef-8 v3.0 Expression BeadChip) were obtained from the Gene Expression Omnibus (GEO, https: //www.ncbi.nlm. 

Preprocessing and analysis of DNA methylation data
485 577 DNA methylation probes in the Illumina HumanMethyla- tion450 BeadChip array with at least one ‘NA’ across all samples were removed, and 394 475 probes were left for further analysis. Differential methylation analysis was performed to obtain the stage-related CpG sites between early-stage and late-stage using the R package ‘minfi’ [28]. Differences between group methy- lation medians (DGMB) were calculated, keeping only probes with significant changes (|DGMB| ≥ 0.1), and P-value <0.01 were defined as differential CpG sites. DNA methylation probes were mapped onto the human genome coordinates using the UCSC Xena probe Map derived from the GPL13534 record (https: //www. ncbi.nlm. cgi? acc=GPL13534, v.1.1).

Preprocessing and analysis of gene expression data
RNA-seq data from TCGA with zero values greater than 10% of the sample size were filtered. Differential expression analysis between early-stage and late-stage and between non-metastatic and metastatic samples were performed with the R package ‘DESeq2’ [29]. Those genes with |fold change| >1 and FDR adjusted P-value <0.01 were selected as differentially expressed genes. Hierarchical clustering was performed using the R package ‘pheatmap’.

Identification of methylation-driven signature (MeSig)

Machine-learning-based variable selection was carried out to identify methylation-driven prognostic signature using likelihood-based boosting in the Cox model as implemented in the R package ‘CoxBoost’ [30] as follows: Firstly, candidate methylation-driven prognostic factors were integrated with clinical covariates and survival data by likelihood-based boosting. Then, the CoxBoost was applied for fitting a Cox proportional hazards regression model with time-dependent covariates to define optimal methylation-driven prognostic signature. The number of boosting iterations was optimized via ten cross-validations.

The updated parameter βˆ vector could be obtained by formula 1, as follows:βˆk,j(update) j ∈ Ikl∗ R package ‘ConsensusClusterPlus’, which could automatically select the number of clusters and is the most commonly used unsupervised clustering method [34]. The number of clustering was determined by the delta plot of the relative change in the area under the cumulative distribution function (CDF) curve. linear combination of coefficients and expression value of the MeSig was used to construct a risk score model. The optimal cutoff value of the risk score was calculated using the R package.

Assume that the actual estimate of an overall parameter vector βˆ being βˆk−1 = (βˆk1 , βˆk2 , ··· βˆkp )’ after step k − 1 of the algorithm and qk predefined candidate sets of features in step k with Ikl ⊂ {1, … , p},l = 1, … , qk. All parameters were updated ineach set simultaneously to determine best l∗ that improved the overall fitting most for updating βˆ. Finally, methylation-driven prognostic signatures were obtained when the optimal number of boosting steps was reached.

For the CoxBoost model, the optimal penalty was determined via ten cross-validations using the ‘optimCoxBoostPenalty’ func- tion in the R package ‘CoxBoost’, and then the number of boost- ing iterations was optimized via cross-validations using the ‘cv. CoxBoost’ function in the R package ‘CoxBoost’. The opti- mal numbers of methylation-driven prognostic genes (MDPGs) were determined using an automatic selection procedure in the CoxBoost algorithm.

Functional enrichment analysis

Functional enrichment analysis of Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway was performed using The Database for Annotation Visualization and Integrated Discovery (DAVID) (https: //david. ncifcrf.go v/, v6.8) for identifying biological processes (BP), GO terms and KEGG pathways that are over-represented in the given gene list [31, 32].

‘survminer’. We stratified the UVM patients into high- and low- risk groups according to the risk cutoff. Kaplan–Meier curve (Log-rank test) was used to visualize and compare the survival difference between the two risk groups. The time-dependent area under the curve (AUC) at 3 and 5 years was used to assess the prognostic performance of the models using the R package ‘timeROC’. All statistical calculations were performed using the R 3.6.0 statistical environment.


Identification of MDPGs - The workflow of this study is shown in Figure 1. We first performed differential methylation analysis between patients with early-stage and late-stage and identified a total of 3982 stage-related CpG sites. Of them, 2187 candidate prognostic CpGs were found to be significantly associated with overall survival (Figure 2A). The Manhattan plot showed that these associated sites were not evenly distributed across the human autosomes, and the highest density was on chromosome 17 (2.019824/mb) and 19 (1.758867/mb) and the lowest density was on chromosome 21 (0.3116566/mb) (Figure 2B). An unsupervised hierarchical cluster analysis of 2187 candidate prognostic CpG sites produced two distinct patient clusters, which were significantly correlated with the clinical stage (Chi-squared test P = 0.002) (Figure 2C). When 2187 candidate prognostic CpG sites were mapped to genomic compartments, we found that the majority of 2187 CpGs were enriched in promoter and gene body regions involved in 304 hypermethylated genes and 619 hypomethylated genes (Figure 2D and E). Compared with hypomethylated CpGs, hypermethylated CpGs were more enriched in promoter regions. With regard to CpG content and neighborhood context, the majority of 2187 candidate prognostic CpGs belongs to the shores and open sea areas of the genome. Compared with hypomethylated CpGs, more hypermethylated CpGs belong to the shore area of the genome. To identify BAI1 progression-related genes, we also performed differential expression analysis between patients with early-stage and late-stage, and between metastatic and non-metastatic patients. A total of 219 stage-related genes (including 188 upregulated and 31 downregulated genes) and 1013 metastasis-where N is total number of human genome background, M is number of genes belonging to the specified BP or pathway, n is the number of interested genes and k is the number of interested genes in the specified BP or pathway.

Statistical analysis

The conventional univariate Cox proportional hazards regres- sion model for overall survival data was implemented to identify variables associated with overall survival. Significant factors in univariate analysis were further subjected to a multivariate Cox regression analysis. Hazard ratios (HR) and corresponding 95% confidence interval (CI) were calculated in the Cox models. The consensus clustering analysis was implemented using the related genes (including 199 upregulated and 814 downregulated genes) were identified (Figure 2G and H). Finally, 40 hyperme- thylated and hypomethylated genes that were significantly associated with stage or metastasis were defined as MDPGs .

Functional characterization of MDPGs
To explore the functional roles of 40 MDPGs, we performed GO and KEGG pathway functional enrichment analysis for MDPGs. Results of GO analysis showed that 40 MDPGs are enriched in 7 GO BP, including nervous system development, cell adhesion, neural crest cell migration, axon guidance, cell surface receptor signaling pathway, cellular response to lipopolysaccharide and The workf ow of 10MeSig generation and validation. The workf low was implemented in three steps. Step 1: The MDPGs were identified by integrative analysis of methylome and transcriptome in the top solid-lined box. Step 2: The methylation-driven prognostic signature was established using the machine learning method. Step 3: The methylation-driven prognostic signature was validated in an independent cohort.

cellular response to epinephrine stimulus (Figure 3A). Results of KEGG analysis showed that 40 MDPGs are involved in protein digestion and absorption, and axon guidance. (Figure 3B). To further investigate the relationship between MDPGs and clinical features of UVM, we implemented consensus clustering analysis for 80 UVM patients in the TCGA dataset based on the expression pattern of 40 genes, which revealed three distinct patient clus- ters according to the change in CDF area (Figure 3C and D). The overall survivals among the three patient clusters were signifi- cantly different, and patients in the Cluster 3 had significantly poor outcome compared with those in the Cluster 1 and the Cluster 2 (log-rank P = 0.0046) (Figure 3E).

 Identification of MDPGs. (A) Venn diagram of candidate prognostic CpG sites. (B) Manhattan plot. The plot suggested a chromosomal location of 2187 candidate prognostic CpG sites. The CpG site with the highest signal intensity was labeled in each chromosome. (C) Heatmap of unsupervised hierarchical clustering of 2187 CpG sites. (D)The distribution of candidate prognostic CpG sites in all genomic compartments: promoter, body, 5’UTR, 3’UTR and 1stExon. (E) The box plot about the number of hyper- and hypo-methylation genes of 2187 CpG sites. (F) The distribution proportion of candidate prognostic CpG sites in all neighborhood contexts classified in shore, open-sea, shelf and island. Shores are considered regions 0–2 kb from CpG islands, shelves are regions 2–4 kb from CpG islands and open sea regions are isolated CpG sites distant from CpG islands (>4kb). Islands are DNA regions with GC content accounting for at least 55% of the nucleotides and an expected/observed CpG ratio of 0.65 (>500 bp). (G) Volcano plot of differentially expressed genes between early-stage and late-stage. (H) Volcano plot of differentially expressed genes between non-metastatic and metastatic. (I) Venn diagram plot among stage-related, metastasis-related and methylation-driven genes.

Development and validation of MeSig

We performed a boosting machine learning algorithm for sig- nature selection from 40 MDPGs and identified a MeSig, includ- ing 10 MDPGs (AZGP1, BAI1, CCDC74A, FUT3, PLCD1, S100A4, SCN8A, SEMA3B, SLC25A38 and SLC44A3) [hereafter referred to novel DNA methylation-driven signature (10MeSig)]. Then, the 10MeSig was transformed into a risk scoring model by linear combination of the expression of the 10MeSig and weighted by relative coefficients in the multivariate Cox regression as follows: 10MeSig = (−0.0096420) × expression value of AZGP1 +

 Functional characterization of MDPGs. (A) Significantly enriched GO BP. (B) Significantly enriched KEGG pathways. (C) CDF plot. (D) Consensus clustering heatmap of 40 MDPGs. (E) Kaplan–Meier survival curves of patients among different clusters.

0.0227310 × expression value of BAI1 + 0.0584049 × expression value of CCDC74A + (−0.2563306) × expression value of FUT3 + 0.0124631 × expression value of PLCD1 + 0.0005874 × expression value of S100A4 + (−0.3786607) × expression value of SCN8A
+ (−0.0132431) × expression value of SEMA3B + (−0.0333145)
× expression value of SLC25A38 + (−0.0645225) × expression
value of SLC44A3. The optimal risk cutoff value of the 10MeSig stratified 80 patients into the high-risk group (n = 30) and low- risk group (n = 50) with different overall survival. As shown in Figure 4A, patients in the high-risk group had conspicuously more poor overall survival than those in the low-risk group (log-rank test P-value<0.001). The time-dependent ROC analysis showed that the 10MeSig achieved AUC of 0.95 and 1.0 for predicting 3- and 5-year overall survival (Figure 4B).

To further investigate the impact of DNA methylation on gene regulation in UVM, we first investigate the expression and DNA methylation pattern of the 10MeSig in the high-risk group and low-risk group stratified by the 10MeSig. As shown in Figure 5A, four patterns of expression and DNA methylation involved in the 10MeSig were observed based on the intersection between DNA methylation and gene expression, including
hypermethylated-upregulated (hyper-up), hypermethylated- downregulated (hyper-down), hypomethylated-upregulated (hypo-up) and hypomethylated-downregulated (hypo-down) genes. Four MDPGs (AZGP1, FUT3, PLCD1 and SLC25A38)
are hyper-down genes. Two MDPGs (BAI1 and S100A4) are hypo-up genes, three MDPGs (SCN8A, SEMA3B and SLC44A3) are hypomethylated-downregulated genes and one MDPG (CCDC74A) is a hyper-up gene (Figure 5A).

Correlation analysis has been widely used to examine the relationship between methylation and gene expression [37, 38]. Therefore, we further examined the impact of DNA methylation on gene expression by conducting Spearman’s rank correlation analysis and found that seven CpG–gene pairs exhibited a sig- nificant negative correlation and four CpG–gene pairs showed a significant positive correlation (Figure 5B).

Independent validation of the 10MeSig
To further validate the relationship between MDPGs and clinical features of UVM, 35 genes of MDPGs were mapped in the inde- pendent Zhang’s cohort. Then, consensus clustering analysis

The 10MeSig was independent of clinical and pathological characteristicsThe univariate analysis showed that the 10MeSig was signifi- cant with OS both in the TCGA-UVM (HR = 29, 95% CI: 6.7–126, P < 0.001) and Zhang’s cohorts (HR = 2.1, 95% CI: 1.2–3.7, P = 0.007).

To further examine whether the 10MeSig was independent of other clinical and pathological factors, we performed multi- variable Cox proportional hazards analysis including individual clinical variables with the 10MeSig in each dataset. As shown in Figure 7A, in the TCGA-UVM cohort, the 10MeSig (HR = 173.17, 95% CI: 21.41–1400.8, P < 0.001), stage IV (HR = 18.48, 95% CI: 1.57–
217.6, P = 0.02) and age (HR = 15.33, 95% CI: 2.90–81.1, P = 0.001)
were significantly associated with OS in the multivariate anal- ysis. In the independent Zhang’s cohort, only the 10MeSig main- tained a significant association with OS (HR = 2.2, 95% CI: 1.24– 4.1, P = 0.008) (Figure 7B). These results suggested that 10MeSigis an independent predictive factor for the survival of patients with UVM.


UVM is the most common form of eye cancer. Although there have been increasing efforts in UVM research in the recent years, the UVM is still associated with poor prognosis and survival rates have not been changed over the past years [39, 40]. Clinical, morphological and genetic features have been used as prognos- tic factors or tools for guiding follow-up and making adjuvant therapy decisions [41]. Increasing evidence has suggested that aberrant DNA methylation was one of the hallmarks of cancers and associated with cancer development, prognosis and treat- ment response. Aberrant DNA methylation has rapidly emerged as a diagnostic and prognostic signature in many cancers. How- ever, such DNA methylation signature available in UVM remains limited.

In this study, we performed a genome-wide integrative analysis of methylome and transcriptome and identified 40 methylation-driven genes that may be associated with the tumorigenesis and progression of UVM. Of them, eight genes (RCVRN, RHO, CYCS, EDN3, CTNNB1, EDNRB, SAG and ROBO1) have
been confirmed to be involved in UVM. For example, the EDNRB gene was downregulated in metastatic UVM and associated with early clinical metastasis and poor survival [42]. CTNNB1 and EDNRB were melanocyte differentiation genes, and their expression losses were correlated with a worse prognosis [43– 45]. Besides, 17 of 40 methylation-driven genes (HNMT, KIF6, SLC25A38, SULT1C4, PLCD1, TNXB, SLC44A3, ACSF2, PAPPA2, RBPJL, IL12RB2, WARS, RAB1, MSC, MRC2, FUT3 and CCDC74A) have been reported to be associated with other eye diseases [46]. Functional analysis of 40 methylation-driven genes also revealed significantly enriched in GO BP and pathways involved in the pathogenesis of diverse cancers and blinding eye diseases [47]. Furthermore, the expression pattern of 40 methylation-driven genes could group UVM patients into two patient classes with significantly different outcomes. This evidence supported the association of 40 methylation-driven genes with the tumorige nesis and prognosis of UVM.

To further investigate the clinical application of these 40 methylation-driven genes, we performed machine-learning- based feature selection and identified a MeSig composing of 10 MDPGs (AZGP1, BAI1, CCDC74A, FUT3, PLCD1, S100A4, SCN8A, SEMA3B, SLC25A38 and SLC44A3) (10MeSig), which stratified UVM patients of TCGA cohort into two risk groups with significantly different overall survivals. The prognostic value of the 10MeSig was further validated in an external independent UVM cohort with a different platform. Even though only nine out of 10 MDPGs in the 10MeSig were present in the independent validation cohort, the 10MeSig also revealed a significant association with survival and had the ability to distinguish between UVM patients with good and poor outcome. Furthermore, the 10MeSig still remained a significant association with survival even after adjusted by clinical and pathological characteristics in both cohorts. These results demonstrated the effectiveness and robustness of the 10MeSig as an independent prognostic factor in UVM.

However, some limitations need to be noted. The discovery of the DNA methylation signature was based on in silico analysis, and the potential mechanism involved in tumorigenesis and prognosis of UVM was unclear and needed for experimental verification. Second, independent validation of the 10MeSig was challenged by the lack of available patient cohort with clinical information.

In summary, using a machine learning-based discovery and validation strategy, we had identified and validated a 10MeSig in two relatively large patient cohorts as an independent predictive

Key Points

⦁ A total of 40 novel methylation-driven prognostic genes (MDPGs) were identified through genome-wide integrative analysis of methylome and transcriptome.

⦁ A machine learning-based discovery and validation
strategy was proposed to identify a DNA methylation- driven signature (10MeSig) composing of 10 MDPGs.

⦁ With a prospective validation study, the 10MeSig will
improve clinical decisions and provide new insights into the pathogenesis of UVM.

The funders had no roles in study design, data collection and analysis, decision to publish or preparation of the manuscript.

Scientific Research Foundation for Talents of Wenzhou Med- ical University (QTJ18029, QTJ18023, QTJ18024).

Ethics approval and consent to participate
Not applicable

Consent for publication
Not applicable

Authors’ contributions
M.Z. and J.Q. designed the study; P.H. and S.Q.B. developed the methodology; P.H., D.D.F., S.Q.B., C.C.Y. and J.Z.S. performed data analysis. M.Z., J.Q. and P.H. drafted the manuscript. All authors read and approved the final manuscript.

1. Chang AE, Karnell LH, Menck HR. The National Can- cer Data Base report on cutaneous and noncutaneous melanoma: a summary of 84,836 cases from the past decade. The American College of Surgeons Commission on Cancer and the American Cancer Society. Cancer 1998;83: 1664–78.
2. Krantz BA, Dave N, Komatsubara KM, et al. Uveal melanoma: epidemiology, etiology, and treatment of primary disease. Clin Ophthalmol 2017;11:279–89.
3. Derrien AC, Rodrigues M, Eeckhoutte A, et al. Germline MBD4 mutations and predisposition to uveal melanoma. J Natl Cancer Inst 2020. MJ, Shields CL, Cebulla CM, et al. Uveal melanoma. Nat Rev Dis Primers 2020;6:24.
4. Jin B, Zhang P, Zou H, et al. Verification of EZH2 as a drug- gable target in metastatic uveal melanoma. Mol Cancer 2020; 19:52.
5. Salarian M, Yang H, Turaga RC, et al. Precision detection of liver metastasis by collagen-targeted protein MRI contrast agent. Biomaterials 2019;224:119478.
6. Greenberg MVC, Bourc’his D. The diverse roles of DNA methylation in mammalian development and disease. Nat Rev Mol Cell Biol 2019;20:590–607.
7. Meng H, Cao Y, Qin J, et al. DNA methylation, its mediators and genome integrity. Int J Biol Sci 2015;11:604–17.
8. Richardson B, Yung R. Role of DNA methylation in the regulation of cell function. J Lab Clin Med 1999;134: 333–40.
9. Gutierrez-Arcelus M, Ongen H, Lappalainen T, et al. Tissue- specific effects of genetic and epigenetic variation on gene regulation and splicing. PLoS Genet 2015;11:e1004958.
10. Zhang B, Zhou Y, Lin N, et al. Functional DNA methylation differences between tissues, cell types, and across indi- viduals discovered using the M&M algorithm. Genome Res 2013;23:1522–40.
11. Bloushtain-Qimron N, Yao J, Snyder EL, et al. Cell type- specific DNA methylation patterns in the human breast. Proc Natl Acad Sci U S A 2008;105:14076–81.
12. Esteller M. Cancer epigenetics: DNA methylation and chro- matin alterations in human cancer. Adv Exp Med Biol 2003;532:39–49.
13. Dong L, Ma L, Ma GH, et al. Genome-wide Analysis Reveals DNA Methylation Alterations in Obesity Associated with High Risk of Colorectal Cancer. Sci Rep 2019;9:5100.
14. Saghafinia S, Mina M, Riggi N, et al. Pan-Cancer Landscape of Aberrant DNA Methylation across Human Tumors. Cell Rep 2018;25:1066, e1068–80.
15. Laird PW. The power and the promise of DNA methylation markers. Nat Rev Cancer 2003;3:253–66.
16. Micevic G, Theodosakis N, Bosenberg M. Aberrant DNA methylation in melanoma: biomarker and therapeutic opportunities. Clin Epigenetics 2017;9:34.
17. Hinoue T, Weisenberger DJ, Lange CP, et al. Genome-scale analysis of aberrant DNA methylation in colorectal cancer. Genome Res 2012;22:271–82.
18. Zhou M, Zhang Z, Bao S, et al. Computational recognition of lncRNA signature of tumor-infiltrating B lymphocytes with potential implications in prognosis and immunotherapy of bladder cancer. Brief Bioinform 2020.
19. Yan C, Zhang Z, Bao S, et al. Computational Methods and Applications for Identifying Disease-Associated lncRNAs as Potential Biomarkers and Therapeutic Targets. Mol Ther Nucleic Acids 2020;21:156–71.
20. Sun J, Zhang Z, Bao S, et al. Identification of tumor immune infiltration-associated lncRNAs for improving prognosis and immunotherapy response of patients with non-small cell lung cancer. J Immunother Cancer 2020;8:e000110.
21. Chen X, Wang J, Shen H, et al. Epigenetics, microRNAs, and carcinogenesis: functional role of microRNA-137 in uveal melanoma. Invest Ophthalmol Vis Sci 2011;52:1193–9.
22. Field MG, Kuznetsov JN, Bussies PL, et al. BAP1 Loss Is Associated with DNA Methylomic Repatterning in Highly Aggressive Class 2 Uveal Melanomas. Clin Cancer Res 2019;25:5663–73.
23. Li Y, Jia R, Ge S. Role of Epigenetics in Uveal Melanoma. Int J Biol Sci 2017;13:426–33.
24. van der Velden PA, Zuidervaart W, Hurks MH, et al. Expres- sion profiling reveals that methylation of TIMP3 is involved in uveal melanoma development. Int J Cancer 2003;106:472–9.
25. Robertson AG, Shih J, Yau C, et al. Integrative Analysis Identi- fies Four Molecular and Clinical Subsets in Uveal Melanoma. Cancer Cell 2017;32:204, e215–20.
26. Maat W, van der Velden PA, Out-Luiting C, et al. Epigenetic inactivation of RASSF1a in uveal melanoma. Invest Ophthal- mol Vis Sci 2007;48:486–90.
27. Aryee MJ, Jaffe AE, Corrada-Bravo H, et al. Minfi: a flexible and comprehensive Bioconductor package for the analy- sis of Infinium DNA methylation microarrays. Bioinformatics 2014;30:1363–9.
28. Love MI, Huber W, Anders S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol 2014;15:550.
29. Binder H, Allignol A, Schumacher M, et al. Boosting for high-dimensional time-to-event data with competing risks. Bioinformatics 2009;25:890–6.
30. Huang d W, Sherman BT, Lempicki RA. Bioinformat- ics enrichment tools: paths toward the comprehensive functional analysis of large gene lists. Nucleic Acids Res 2009;37:1–13.
31. Huang d W, Sherman BT, Lempicki RA. Systematic and integrative analysis of large gene lists using DAVID bioinfor- matics resources. Nat Protoc 2009;4:44–57.
32. Francesconi M, Remondini D, Neretti N, et al. Reconstruct- ing networks of pathways via significance analysis of their intersections. BMC Bioinformatics 2008;9(Suppl 4):S9.
33. Wilkerson MD, Hayes DN. ConsensusClusterPlus: a class dis- covery tool with confidence assessments and item tracking. Bioinformatics 2010;26:1572–3.
34. Rechache NS, Wang Y, Stevenson HS, et al. DNA methyla- tion profiling identifies global methylation differences and markers of adrenocortical tumors. J Clin Endocrinol Metab 2012;97:E1004–13.
35. Marzese DM, Hoon DS. Emerging technologies for studying DNA methylation for the molecular diagnosis of cancer. Expert Rev Mol Diagn 2015;15:647–64.

1. Cheng J, Wei D, Ji Y, et al. Integrative analysis of DNA methylation and gene expression reveals hepatocellu- lar carcinoma-specific diagnostic biomarkers. Genome Med 2018;10:42.
38. Xu W, Xu M, Wang L, et al. Integrative analysis of DNA methylation and gene expression identified cervical cancer- specific diagnostic biomarkers. Signal Transduct Target Ther 2019;4:55.
39. Lane AM, Kim IK, Gragoudas ES. Survival Rates in Patients After Treatment for Metastasis From Uveal Melanoma. JAMA Ophthalmol 2018;136:981–6.
40. Gorham JP, Szalai E, Wells JR, et al. Retinoinvasive Uveal Melanoma: Report of 2 Cases and Review of the Literature. Ocul Oncol Pathol 2017;3:292–5.
41. Nathan P, Cohen V, Coupland S, et al. Uveal Melanoma UK National Guidelines. Eur J Cancer 2015;51:2404–12.
42. Smith SL, Damato BE, Scholes AG, et al. Decreased endothelin receptor B expression in large primary uveal melanomas is
associated with early clinical metastasis and short survival.
Br J Cancer 2002;87:1308–13.
43. Onken MD, Ehlers JP, Worley LA, et al. Functional gene expression analysis uncovers phenotypic switch in aggres- sive uveal melanomas. Cancer Res 2006;66:4602–9.
44. Harbour JW, Onken MD, Roberson ED, et al. Frequent muta- tion of BAP1 in metastasizing uveal melanomas. Science 2010;330:1410–3.
45. Urtatiz O, Van Raamsdonk CD. Gnaq and Gna11 in the Endothelin Signaling Pathway and Melanoma. Front Genet 2016;7:59.
46. Porter LF, Saptarshi N, Fang Y, et al. Whole-genome methy- lation profiling of the retinal pigment epithelium of individ- uals with age-related macular degeneration reveals differ- ential methylation of the SKI, GTF2H4, and TNXB genes. Clin Epigenetics 2019;11:6.
47. Apte RS, Chen DS, Ferrara N. VEGF in Signaling and Disease: Beyond Discovery and Development. Cell 2019;176:1248–64.