Methylated CfDNA may distinguish between high- and intermediate-risk uveal melanoma: a pilot study

Abstract

Background

Uveal melanoma (UM) is a highly aggressive malignancy with a metastatic risk that depends on the molecular subclass. This subclass can be determined through molecular characterization of tumor-derived tissue. With eye-sparing treatments, tumor tissue is rarely available for molecular testing. We hypothesized that minimal invasive biomarkers such as methylated cell-free DNA (cfDNA) or circulating tumor DNA (ctDNA) can be used for prognosis and monitoring of patients.

Methods

Plasma cfDNA was isolated from healthy blood donors (HBDs, N = 19) and UM patients (N = 22). Plasma was collected at baseline (localized disease, N = 13) and during follow-up (metastatic disease, N = 9) from independent patients with high metastatic risk (HR, N = 11) (monosomy 3 and/or BAP1-mutated tumor) or intermediate metastatic risk (IR, N = 11) (disomy 3 and/or SF3B1-mutated tumor). Methylation signatures were determined using genome-wide LpnPI-based methylated DNA sequencing (MeD-seq). Samples with a CpG/reads ratio < 20% (N = 3) were excluded. IchorCNA was used to estimate the tumor fraction. cfDNA samples with detectable tumor fraction (N = 2) were analyzed separately from the other cfDNA samples without detectable tumor fraction (N = 18) to reduce noise in downstream analyses. Differentially methylated regions (DMRs) were identified between the following predefined subgroups: UM (N = 11) vs. HBDs (N = 19), and HR (N = 10) vs. IR (N = 7). To visualize clustering, principal component analysis (PCA) and hierarchical clustering was performed on the DMRs with fold change > 2.0. Gene set enrichment analysis (GSEA, Z-score > 2.0 and p < 0.05) was performed to evaluate biological relevance.

Results

Distinct clustering was observed between UM and HBDs samples, and between HR and IR samples, although outliers were present in the latter comparison. GSEA implicated eight canonical pathways including the S100 Family Signaling Pathway and RAF/MAP kinase cascade, which are linked to tumorigenesis and immune processes.

Conclusion

This pilot study reports on cfDNA methylation signatures that differentiates UM patients from HBDs, and may distinguish between intermediate and high risk UM subgroups, supporting its prognostic potential. However, its role in monitoring disease progression requires further validation. Independent replication studies are warranted to confirm our findings and evaluate the clinical applicability in UM.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12935-025-04093-2.

Introduction

Uveal melanoma (UM) is the most common primary intraocular malignancy, affecting 0.3 to 11.0 patients per million person-years [1]. It is generally assumed that UM metastasis often has already occurred prior to diagnosis [2]. Metastatic risk can be determined by assessing clinical parameters such as tumor size [3] or, more precisely, through molecular characterization of the primary tumor tissue [4, 5]. Currently, most patients receive eye-sparing treatments, such as proton therapy, stereotactic radiotherapy, and plaque brachytherapy [6]. The advantages of these methods are preservation of the eye and more peripheral vision [7]. Unfortunately, this strategy can lead to more false positives as the diagnosis is made clinically and can no longer be confirmed by histopathology on tissue acquired via enucleation [8]. Additionally, this also reduces the availability of sufficient (quality) tissue for molecular prognostication, as intraocular tumor biopsies are rarely performed due to the risk of ocular complications [9].

Melanocytes located in the iris, ciliary body or choroid, undergo malignant transformation predominantly through a series of events [10, 11]. First, an activating mutation occurs in one of four genes of the Gαq-pathway. These changes mostly occur at hotspot locations in G protein subunit alpha 11 (GNA11) and G protein subunit alpha q (GNAQ), phospholipase C beta 4 (PLCB4) or cysteinyl leukotriene receptor 2 (CYSLTR2) [12]. Mutations in these primary driver genes are mutually exclusive, and more than 95% of UM somatic mutations occur at these hotspots [13]. The risk of developing metastatic disease is related to secondary driver mutations and corresponding copy number variation (CNV) profiles [4, 14] Secondary driver mutations typically occur in three distinct genes: eukaryotic translation initiation factor 1 A X-linked (EIF1AX), splicing factor 3b subunit 1 (SF3B1), and BRCA1 associated protein 1 (BAP1). EIF1AX stimulates and stabilizes the ribosome and acts as a regulator of translation initiation [15, 16]. Patients harboring mutations in EIF1AX often show gain of chromosome 6p and are usually associated with a good prognosis [5, 17]. SF3B1 encodes a crucial part of the spliceosome, and mutations herein result in incorrect selection of branch points on precursor mRNAs [18]. SF3B1-mutated tumors present more and relatively smaller CNVs at the distal end of the chromosomes, such as loss of 1p, gain of 6p, gain of 8q, and loss of 11q. Patients with these tumors follow a bimodal survival curve in which a subset of patients develop early metastases within five years, and others develop metastases after six years [19]. BAP1 mutated tumors mostly exhibit a concurrent loss of heterozygosity of chromosome 3. Patients with these tumors have the worst prognosis, as most patients develop metastases within five years [17].

Since diagnosis confirmation and prognostication based on molecular characterization is not possible due to the lack of primary tumor tissue, minimally invasive alternatives are needed. The clinical applicability of minimally invasive strategies in UM has been evaluated in several studies [2, 20–23]. One promising strategy is the characterization of cell-free DNA (cfDNA), which can contain tumor-specific genetic and epigenetic alterations, making it a valuable biomarker for UM detection and monitoring [2, 21, 24]. The concentrations of cfDNA from patients with cancer are often elevated compared to those from healthy controls [25, 26]. In patients, cfDNA contains DNA originating from a mixture of healthy cells (e.g. immune and endothelial cells) and a variable level of contribution from cancer cells, which depends on cancer type and disease status [27]. The latter is also known as circulating tumor DNA (ctDNA). This ctDNA can be distinguished from cfDNA using the tumor-derived CNV profiles or somatic mutations, which are absent in cfDNA derived from healthy cells [28]. The ctDNA quantification can be used for assessing minimal residual disease by evaluating the fraction of mutated DNA particles in the blood [20, 29, 30].

Cell-free DNA contains methylated sequences, and both normal and disease associated methylation signatures have been used in the clinic in prognosis and monitoring using minimally invasive approaches [31]. DNA methylation is an essential epigenetic modification that is tissue-specific and defines cellular properties. DNA methylation occurs at different genomic regions, and its effects vary depending on the location. For example, methylation at the promoter region typically represses gene expression by inhibiting the binding of transcription factors, silencing the gene [32]. In contrast, gene body methylation, which occurs in the coding region of genes and non-coding regions of gene body, is often associated with actively transcribed genes [33]. For instance, methylation of gene body CpG islands has been associated with increased expression levels of oncogenes in hepatocellular carcinoma [34]. Cancer cells often exhibit genome-wide hypomethylation of CpG islands on oncogenes, and genome-wide hypermethylation of CpG islands on tumor suppressor genes [35]. Methylation signals can even be used to determine the tissue of origin [36].

Previously, we detected the presence of ctDNA in metastatic disease using digital PCR, but not in localized disease [2]. This could be due to either absence of ctDNA in localized disease or technical limitations in detecting low levels of ctDNA. Thus, our objective was to investigate alternative methods such as restriction-enzyme based methylation profiling, to detect tumor-derived cfDNA and evaluate its prognostic potential. Our approach is rooted in previous observations in primary UM, on methylation signatures able to distinguish between low, intermediate and high risk subclasses [37–39]. Moreover, the methylated DNA sequencing (MeD-seq) method has been successfully used to determine genome-wide DNA methylation patterns in cfDNA [40]. Therefore, in this study, we evaluated the prognostic potential of MeD-seq in plasma cfDNA using the samples of UM patients and healthy blood donors (HBD). We aimed to determine the circulating tumor fraction in cfDNA, explore differences in methylated loci in cfDNA between UM patients and HBDs, between patients with known metastasis and those without metastasis, and between intermediate- and high-risk patients, and investigate the potential biological relevance of our findings.

Methods

Study design and patient cohort

This study adheres to the tenets of the Declaration of Helsinki and has been approved by the Erasmus MC medical ethics committee (MEC-2009-375), The Netherlands, Human Research Ethics Committee protocols of Edith Cowan University (No. 11543 and No. 18957) and Sir Charles Gardner Hospital (No. 2013 − 246 and No. RGS0000003289), Western Australia. All patients provided written informed consent. In this two-center, retrospective cohort study, patients were recruited from the Rotterdam Ocular Melanoma Study Group (ROMS cohort) [14], and the Lions Eye Institute and Royal Perth Hospital, Western Australia (Perth cohort) [20]. Clinical and demographic data were collected for all patients. These included age, sex, date of diagnosis, largest basal diameter (LBD), apical height, tumor location, CNV profiles, mutational analysis and BAP1 immunohistochemistry (IHC). Patients with a BAP1 mutation, negative BAP1 IHC or monosomy 3 were classified as high metastatic risk, whereas the other patients with a SF3B1 mutation, positive BAP1 IHC or disomy 3 were classified as having intermediate metastatic risk. Metastatic disease was assessed at baseline using chest radiography, abdominal ultrasound, lactate dehydrogenase (LDH) test and γ-glutamyl transferase (GGT) test, followed by routine abdominal ultrasound, LDH-test and GGT test every 6 months. UM was staged according to the American Joint Committee on Cancer (AJCC) Staging Manual 8th edition [3].

Collection of blood samples for CfDNA extraction

Between 2015 and 2021, blood samples were collected from independent patients at baseline (n = 13) or during follow-up with metastatic disease (n = 9), and from healthy blood donors (HBDs, n = 19). Blood was collected in either EDTA (BD vacutainer systems, Plymouth, United Kingdom), Cell-Free DNA (Streck, La Vista, Nebraska, USA) or CellSafe Preservative Tubes (Menarini Silicon Biosystems, Bologna, Italy). All three tubes effectively stabilize ctDNA and genomic DNA if plasma is isolated within 6 h, as demonstrated by Kang et al. [41]. This protocol was followed in our study. The plasma was separated from blood using centrifugation as previously described [2, 24, 40]. The cfDNA was isolated using the QIAamp Circulating Nucleic Acid Kit (Qiagen, Hilden, Germany) according to the manufacturer’s instructions. The buffer and reagent volumes were adjusted to the respective input plasma volumes. The washing column was eluted twice to increase the cfDNA yield, as reported previously [20]. The cfDNA concentrations were measured using the QuantiT dsDNA Assay Kit, High Sensitivity (Thermo Fisher Scientific, Grand Island, NY, USA), as described by the manufacturer. The isolated cfDNA samples were stored at −80 ˚C until methylated DNA sequencing (MeD-seq) was performed. Sample selection was performed based on the mutation signature of the primary tumor.

Collection of primary tumor, neonatal melanocyte and leukocyte subset samples

Primary tumor samples were obtained from 8 patients who also provided blood and 22 unrelated patients who underwent enucleation as primary treatment and stored at −80 °C until DNA isolation. We included tumors with a BAP1 mutation (n = 10), SF3B1 mutation (n = 10) and EIF1AX mutation (n = 10). Additionally, five healthy foreskin-derived neonatal melanocytes, obtained from the NIGMS Human Genetic Cell Repository at the Coriell Institute for Medical Research were included as a control group. These were GM22250 (RRID: CVCL_1P46), GM22134 (RRID: CVCL_V263), GM22182 (RRID: CVCL_1P32), GM21807 (RRID: CVCL_H988), and GM21810 (RRID: CVCL_H992). Furthermore, leukocyte subsets, specifically neutrophils, monocytes, B lymphocytes, and T lymphocytes were isolated from HBDs. To obtain pure cell populations, whole blood collected in EDTA tubes was stained with directly conjugated mouse anti-human antibodies, including CD3-Pacific Blue, CD14-PE, CD20-FITC (BD Biosciences, Franklin Lakes, NJ), and CD16-APC (BioLegend, San Diego, CA). Cells were subsequently sorted in purity mode using a FACSDiscover S8 cell sorter (BD Biosciences). The gating strategy applied to identify and isolate specific cell populations was as follows: neutrophils (CD16⁺, SSChigh), monocytes (CD14⁺, SSCintermediate), B-lymphocytes (CD20⁺, SSClow), and T-lymphocytes (CD3⁺, SSClow). DNA was isolated as described previously [37, 42, 43] and then stored at −80 ˚C until MeD-seq was performed.

Methylated DNA sequencing (MeD-seq) sample processing

MeD-seq assays using the restriction enzyme LpnPI were performed as reported previously [38, 40]. In brief, 8 µL cfDNA isolated from plasma (input ranged 7.0 to 32.9 ng) or 8 µL DNA isolated from tumor tissue (input 50 ng) were digested with the restriction-enzyme LpnPI (New England Biolabs, Ipswich, Massachusetts, USA), yielding 32 bp (bp) around the fully methylated CpG sites. The ThruPLEX DNA-seq 96D kit (Rubicon Genomics, Takara Bio Europe, Saint-Germain-en-Laye, France) was used for sample preparation and samples were purified on a Pippin HT system with 3% agarose gel cassettes (Sage Science, Beverly, Massachusetts, USA). Libraries were multiplexed and sequenced using the Nextseq2000 (Illumina, San Diego, California, USA) for 50 bp single reads according to the manufacturer’s instructions. Samples were first sequenced until approximately 2 million (M) reads and continued to a total of ~ 20 M reads if the fraction of reads that passed the LpnPI filter was at least 20%. When the fraction of reads was between 20 and 30%, samples were further sequenced until ~ 30 M reads.

MeD-seq data processing

Data was processed using specifically created scripts in Python version 2.7.5 (Python Software Foundation available at http://www.python.org), as previously described [38, 40]. In short, dual indexed samples were demultiplexed to create Raw FASTQ files using bcl2fastq software version 2.20.0, available at https://support.illumina.com/sequencing/sequencing_software/bcl2fastq-conversion-software/downloads.html (Illumina, San Diego, California, USA). Raw FASTQ files were subjected to Illumina adaptor trimming, and reads were filtered based on LpnPI restriction site occurrence between 13 and 17 bp from either 5’ or 3’ end of the read. Reads passing the LpnPI filter were mapped to hg38 using Bowtie 2 [44]. Afterwards, the total number of obtained reads, number of reads that passed the filter, and number of duplicate reads were used as outcome measures for methylated CpG site counts. Gene and CpG island annotations were downloaded from Ensembl (https://www.ensembl.org). Genome wide individual LpnPI site scores were used to generate read count scores for the following annotated regions: transcription start site (TSS, 1 kb before and after), CpG island and gene body (1 kb after TSS until transcription end site (TES)).

Data analysis of differentially methylated regions

Differentially methylated regions (DMRs) were determined using a supervised approach by employing a genome-wide sliding window to detect sequentially differentially methylated LpnPI sites between two groups of the derivation cohort. Genome-wide read counts were normalized (RPM, reads per million) for coverage and compared using the chi-square test, with significance set at p < 0.05 and a Bonferroni correction for multiple testing. Neighboring significantly LpnPI sites were binned and reported. Overlap of genome-wide detected DMRs was reported for TSS, CpG island or gene body regions using annotations of the UCSC database (Hg38) [45]. The Z-score for differentially methylated regions was normalized to the total number of reads. DMRs with fold changes (FCs) ≥ 2.0 were considered statistically significant. Heatmaps and dendrograms of the Z-score were generated with the Python script ‘SciPy’ (version 3.11.6) using the ‘City Block’ distance and the ‘Complete’ method. An area under the curve (AUC) score was computed using Python module ‘sklearn’ (version 1.3.2). To identify clusters and potential outliers, principal component analysis (PCA) was performed by reducing the dimensionality of the data using the singular value decomposition function in R v.3.6.3. Clustering of leukocyte subsets were visualized using Python script ‘matplotlib’.

Analysis of CNVs, the tumor fraction and CfDNA levels

Mapped and sorted bam files were obtained through quality control (QC), preprocessing, and analysis pipelines. Background reads without a methylated CpG signal were selected, and subsequently filtered based on the read quality, ENCODE blacklist, and PCR duplicates. Next, background reads were counted in bins of size 1 Mb and analyzed using the IchorCNA model to simultaneously predict copy number states and estimate the (circulating) tumor (DNA) fraction with a detection threshold of 10% [46] (Fig. 1A). In our dataset, only UM relevant chromosomes were assessed, which are chromosomes 1, 3, 6 and 8 [42]. In addition to estimating the tumor fraction, we compared cfDNA levels between cfDNA samples of different subgroups: all patients at baseline (N = 11), patients at baseline with high metastatic risk (N = 4), patients at baseline with intermediate metastatic risk (N = 7), cfDNA of patients with metastatic disease during follow-up (N = 6) and HBDs N = 19).

Fig. 1.

Flowchart of the included and analyzed samples. A. Schematic overview of the MeD-seq pipeline and bioinformatic analyses [38]. First, (cf.) DNA is digested by LpnPI which recognizes methylated CpGs and cuts the DNA 16 bp up- and downstream leading to 32 bp fragments. This is followed by adaptor ligation, size selection, amplification, sequencing, and aligning to reference genome hg38. Methylated reads are filtered from background reads using the central methylated CpG and aligned prior to analysis. Methylated reads were used to perform PCA, heatmaps and gene set enrichment analyses (GSEA), whereas the background reads were used to perform copy number (CN) profiling and tumor fraction (TF) estimation. B. samples were collected from UM patients and baseline which included 4 high risk and 7 intermediate risk cases (total N = 11). An additional 6 samples were obtained from high risk patients at the time of metastatic progression. Nineteen cfDNA control samples were collected from HBDs. DNA from 30 primary UM tumor tissue was collected from, 10 each from patients with high, intermediate and low metastatic risk. Additionally, five foreskin-derived neonatal melanocytes were included as healthy tissue controls. C. Methylation signatures were created by determining differentially methylated regions (DMRs) between two groups of the derivation cohort. Next, these signatures were evaluated by generating Z-score of the read count from additional samples relative to the normalized read count of the DMRs in the derivation cohort. These Z-scores were then added as additional samples to the supervised clustering analysis of the derivation cohort C1. The ‘UM cfDNA methylation signature’ was established using the cfDNA of UM patients collected at baseline (N = 11) and cfDNA of HBDs (N = 19) and evaluated with leukocyte subsets (N = 17). C2. The ‘metastatic risk methylation signature’ was established using high risk (N = 10) and intermediate risk (N = 7) samples and evaluated with HBDs’ samples. C3. The ‘primary tumor DNA methylation signature’ compared the DMRs between the three tumor subgroups and evaluated with ctDNA + samples (N = 2) and neonatal melanocytes (N = 5)

Establishment of methylation signatures using differentially methylated regions

Samples with a CpG/reads ratio of < 20% (N = 3) did not pass quality control and were excluded. cfDNA samples with a tumor fraction > 10% (N = 2), hereafter referred to as ‘ctDNA + samples’, were analyzed separately to reduce noise from ctDNA in the downstream methylation signature, which are free from circulating tumor DNA. The remaining patients’ cfDNA samples (N = 17) and primary tumor DNA (N = 30) were categorized based on type of DNA (cfDNA or tissue DNA) and metastatic risk class (high risk, HR; intermediate risk, IR; low risk; LR) (Fig. 1B). Three methylation signatures were established using samples from the derivation cohort and evaluated with additional samples (Fig. 1C). Evaluation was performed by generating Z-scores of the read count from additional samples relative to the mean normalized read counts of the DMRs in the derivation cohort. Next, these Z-scores were then added as additional samples to the supervised clustering analysis of the derivation cohort.

Evaluation of biological relevance using gene set enrichment analysis and a literature search

Ingenuity Pathway Analysis (IPA) [47] (available at QIAGEN., https://digitalinsights.qiagen.com/IPA) was used to determine which pathways were impacted by the DMRs. Only the comparison between the eleven patients at baseline without measurable tumor fraction and the nineteen HBD controls resulted in a sufficient number of DMRs (N = 1448) to perform the gene set enrichment analysis (GSEA). The DMRs were filtered on FC ≥ 2.0 and an area under the curve (AUC) ≤ 0.2 or ≥ 0.8. The DMRs located on promoter regions and CpG islands were labeled as ‘downregulated’, whereas DMRs on gene body methylation were labeled as ‘upregulated’. Within IPA, mapped DMRs were used to run a Core Analysis with default parameters. Significant canonical pathways were identified using the P value of overlap, which was calculated using the right-tailed Fisher’s exact test, and the direction change of the pathway (Z-score). As thresholds, a negative log₁₀(P value) greater than 1.3 and a Z-score of 2.0. A non-systematic literature search was performed on 24-10-2024 using the search query “uveal melanoma AND (epigenetics OR DNA methylation or cfDNA or ctDNA)”.

Statistical analysis

The P values were calculated with the Student’s t-test or the Mann-Whitney U test for continuous variables and the chi-square or Fisher’s exact test for categorical values. All the statistical tests were two-sided, unless otherwise specified. Differences between groups were considered statistically significant at P < 0.05. The P value of overlap was adjusted with the Benjamini-Hochberg procedure [45]. Box- and jitter plots were created using the Chart Builder in SPSS. Venn diagrams were drawn using Adobe Illustrator 2025. Analyses were performed using Microsoft Excel version Professional Plus 2016 and IBM SPSS Version 26.0.

Results

Patient cohort

The baseline characteristics of patients with UM are displayed in Table S1. In the discovery cohort, 17 cfDNA, 19 HBDs’ cfDNA, and 30 primary tumor DNA samples were analyzed. No significant differences were found in age (P = 0.72), sex (P = 0.49), AJCC stage (P = 0.64), or metastatic risk (P = 0.63) between the patients from the ROMS and those from the Perth cohort (Table S2). Age (P = 0.12) and sex (P = 0.74) did not differ between patients and HBDs (Table S3). Also no significant age and sex differences were observed between the analyzed subgroups (Table 1), and additionally, AJCC T-stage for the primary tumors (Table S4). The difference in metastasis-free survival between intermediate risk and high risk UM patients did not reach statistical significance (P = 0.11, log-rank test).

Table 1.

Comparison of age and sex between the subgroups of the methylation signatures

Subgroups	Age (mean, years)	Sex (female %)	P-value (age)	P-value (sex)
cfDNA baseline UM (N = 11) vs. HBDs (N = 19)	58.8 vs. 50.7	50% vs. 37%	0.19	0.47
cfDNA: baseline UM (N = 11) vs. Metastatic UM (N = 6)	58.8 vs. 50.7	50% vs. 37%	0.19	0.47
cfDNA: high risk UM (N = 10) vs. Intermediate risk UM (N = 7)	59.9 (all)	40% vs. 57%	0.994	0.49
Primary tumor DNA: BAP1 (N = 10) vs. SF3B1 (N = 10) vs. EIF1AX (N = 10)	59.9 (all)	50% vs. 60% vs. 50%	0.994	0.88

Data quality, tumor fraction and CNV calling

The mean coverage was 2.12 (range: 1.43–4.84). Usually, MeD-seq has a coverage of approximately 50% of all CpG sites sequenced to a mean depth of 1.0–1.5 [38]. Among the 22 cfDNA samples, three samples did not pass quality control due to a CpG/reads ratio < 20.0%; thus, these samples were excluded from further analyses (TableS1). The cfDNA levels were highest in ctDNA + samples (mean 2.61 ng/µL), followed by localized disease samples from high-risk UM patients (mean 2.26 ng/µL), localized disease samples from intermediate-risk UM patients (mean 2.11 ng/µL), and samples from patients with metastatic disease (mean 1.67 ng/µL) (Fig. 2A). These levels were all significantly higher than the cfDNA levels of the HBDs (mean 0.80 ng/µL, P < 0.01). Using IchorCNA on the background reads, we simultaneously predicted the CNV profiles and tumor fractions of our patient samples. Two plasma cfDNA samples of patients with metastatic disease (SE22-4155 and SE22-4154) had estimated ctDNA fractions of 68% and 15%, respectively, whereas all other plasma samples had tumor fractions below the detection threshold of 10% (Fig. 2B). IchorCNA predicted a loss of chromosome 3, 6q, and 8p, and a gain of 8q in ctDNA + sample SE22-4155 (Fig. 3A). Loss of chromosome 3 and gain of 8q was also observed in primary tumor DNA (SE23-4314) derived from the same patient using IchorCNA (Fig. 3B) and confirmed via the SNP-array (Fig. 3C). The tumor fraction of the other ctDNA + sample (SE22-4154) was too low to reliably estimate the CNV profile. In contrast, one included patient provided two samples, a cfDNA sample and a primary tumor DNA sample that had a tumor fraction below the detection threshold of 10%, resulting in normal CNV profiles on IchorCNA and the SNP-array (Fig. S2).

Fig. 2.

cfDNA levels and tumor fraction Plots depicting cell-free DNA (cfDNA) concentrations and estimated tumor fraction in ctDNA + samples with metastatic disease (N = 2), localized disease of all UM (N = 11), localized high-risk UM (N = 4), localized intermediate-risk UM (N = 7), and metastatic UM (N = 6). The Mann-Whiteney U test was used to compare mean values of cfDNA concentrations. Abbreviations: ns. P > 0.05, * P ≤ 0.05, *** P ≤ 0.001. (A) The cfDNA concentrations of ctDNA + samples (mean 2.61 ng/uL), all localized disease (mean 2.16 ng/uL), localized disease of high-risk UM (mean 2.26 ng/uL), localized disease of intermediate-risk UM (mean 2.11 ng/uL), and metastatic UM patients (mean 1.67 ng/uL) were significantly higher than that of the HBDs (mean 0.80 ng/uL, P < 0.05). (B) The tumor fractions could be estimated using ichorCNA. ctDNA + samples had an estimated (circulating) tumor (DNA) fraction of 15% and 68%. All other samples had a tumor fraction below the detection threshold of 10%

Fig. 3.

Prediction of CNVs and tumor fraction using IchorCNA Here we display the CNV profiles of the only two cfDNA samples in our study with a tumor fraction above the detection threshold of 10%. (A) SE22-4155 is a cfDNA sample with metastatic disease. UM-related predicted CNVs using IchorCNA are loss of chromosome 3, 6q, 8p, and gain of 8q, with a tumor fraction of 68%. (B) SE23-4314 is a primary tumor DNA sample, derived from the same patient as SE22-4155. Only a loss of chromosome 3 and gain of 8q was observed using IchorCNA. (C) The SNP-array of the same sample, SE23-4314 displayed loss of chromosome 3, gain 4p, marginal gain 6p, loss 6q, and gain 8q. Noteworthy, CNVs such as loss of chromosome 10, 13, 16 and X have been observed in the cfDNA sample SE22-4155, but not in the IchorCNA and SNP-array of the tumor sample SE23-4314

Identification of clusters and outliers through principal component analysis.

PCA was performed on the cfDNA methylation signatures to identify clusters of subgroups (Fig. S3). Compared with the other cfDNA samples, the ctDNA + samples presented relatively high negative values of PC1 but were not identified as separate clusters or outliers. Nonetheless, those ctDNA + samples were analyzed separately to ensure a consistent cfDNA signal across the different downstream analyses. Most, but not all, cfDNA samples from UM patients can be distinguished from HBDs (Fig. S3A-C). No clusters were observed in the metastatic risk classes (Fig. S3D-F).

Established methylation signatures

The comparison of cfDNA from UM patients and HBDs revealed a methylation signature of 1,448 DMRs using a genome wide sliding window approach, with 402 DMRs above the cutoff of FC > 2.0, (Table S5). The 402 DMRs were used to visualize the distinct clustering of UM patients (N = 11) against HBDs (N = 19) (Fig. 4). To investigate the origin of this clustering, methylated read from leukocyte subsets were incorporated, revealing that these samples clustered closely with the cfDNA profiles of UM samples (Fig. S4). Second, 38 DMRs with 12 DMRs above the cutoff of FC > 2.0 (Table S6), were identified from the metastatic risk signature. Most samples from UM patients in the HR group (N = 10) clustered separately from those in the IR group (N = 7), although a heterogeneous cluster containing both HR and IR samples was also observed (Fig. 5). Here, most clustering was associated with the methylation of several CpG sites on a intergenic region chr1:161,441,815 − 161,465,530 that, due to the absence of a gene annotation was labeled “CpG island 3132”. Evaluation with HBDs samples resulted in poor clustering, indicating that these 38 DMRs are specific for distinguishing between HR and IR samples but are not suitable for differentiating UM patients from healthy controls (Fig. S5). Our primary tumor signature revealed a total of 181 DMRs that were unique to each of the primary tumor groups, of which 49 had FC > 2.0 (Table S7). With these 49 unique subgroup-specific DMRs, hierarchical clustering of BAP1 (N = 10), SF3B1 (N = 10), and EIF1AX-mutated tumor DNA (N = 10) revealed heterogeneous clusters, particularly among SF3B1 and EIF1AX samples, which often exhibit fewer CNVs. In contrast, BAP1-mutated samples, typically characterized by extensive CNVs, showed more distinct clustering (Fig. S6). Finally, the foreskin-derived neonatal melanocytes (N = 5) formed one cluster based on hypomethylated reads on chromosomes 3, 6, 9, 11, and 14, whereas the ctDNA + sample with a highly elevated tumor fraction (SE22-4155, tumor fraction 68%) clustered with BAP1-mutated tumors and the other ctDNA + samples with mildly elevated tumor fraction (SE22-4154, tumor fraction 15%) clustered with EIF1AX-mutated tumors (Fig. S7).

Fig. 4.

Heatmap and dendrogram of the UM cfDNA methylation signature: UM patients vs. HBDs Two distinct clusters were observed, one for the cfDNA of HBDs (left) and UM patients (right)

Fig. 5.

Heatmap and dendrogram of the cfDNA metastatic risk signature from HR UM patients vs. IR UM patients Three clusters were observed with one outlier (SE23-4304): cfDNA of UM patients with high metastatic risk (HR, left), patients with intermediate metastatic risk (IR, right), and a heterogenous cluster of high and intermediate risk UM

Biological relevance through impacted pathways and literature search

A total of 1,448 DMRs were filtered based on FC and AUC, resulting in 207 DMRs. The associated genes were then uploaded into IPA for GSEA (Fig. S8). After filtering the enriched pathways based on Z-score and -log(p-value), eight canonical pathways remained, each showing significant enrichment of genes associated with these pathways in the cfDNA of UM patients (Fig. 6). Additionally, we performed a literature search for genes related to methylation in UM (Fig. S9). The analyzed genes (N = 77) from 27 articles are summarized in Table S8. Next, these were categorized based on biological relevance and checked for their presence in our dataset. Unfortunately, no overlap was found between the genes from the literature search and the DMRs in our three established methylation signatures. Additionally, the DMRs from the three methylation signatures showed no overlap with each other (Fig. S10).

Fig. 6.

Enriched canonical pathways Bar chart of the eight significant associated canonical pathways from the gene set enrichment analysis using enriched genes based on differentially methylated data of UM cfDNA. The threshold was set on a Z-score > 2.0 and -log(p-value) > 1.3. The pathways with a negative Z-score are enriched in UM patients, depicted in blue

Discussion

Prognostic information for UM patients can be obtained through the analysis of the primary tumor. However, tumor tissue is scarce due to eye conserving treatments, and other ways to provide prognostic information are being explored. Blood based techniques such as the use of CTCs, cfDNA, ctDNA, and microRNAs, are under investigation but have yielded contradictory results [2, 13, 24, 48, 49]. In this retrospective pilot study, we aimed to identify cfDNA methylation signatures in UM patients and investigate the origins of the cfDNA. Using methylated cfDNA profiles, we were able to stratify patients into high- and intermediate metastatic risk subgroups, but validation is warranted. In addition to subgroup comparisons, we performed an analysis combining all UM patients versus healthy donors, which revealed a distinct methylation signature and clear separation between these groups. This finding underscores the potential clinical applicability of methylated cfDNA as a minimal invasive biomarker for detecting UM, even when tumor-derived cfDNA fractions are low.

cfDNA refers to fragments of DNA that circulate freely in the bloodstream, and originates from various sources, including normal cell turnover [50], apoptosis [51], necrosis [51], and active secretion by cells [52]. In cancer patients, these processes can lead to elevated levels of cfDNA. Earlier studies encountered challenges in detecting a tumor-derived cfDNA signal [2, 20, 21, 24, 29]. Alternatively, methylation-based cfDNA detection can provide a higher degree of specificity and sensitivity [53]. In recent studies [27, 54, 55], elevated cfDNA levels have been described in patients with cancer, which is consistent with our findings (Fig. 2A). Using the IchorCNA model [46] on methylated cfDNA, we simultaneously predicted CNVs and estimated the tumor fraction. No aberrant CNVs were observed in most of our samples, resulting in estimated tumor fractions below the detection threshold of 10% (Table S1). This finding was in agreement with our previous study, where ctDNA levels were undetectable in patients with localized UM, but ctDNA was found in 77% of those with metastatic disease [2]. While ctDNA mutation analysis has demonstrated prognostic value in UM, its sensitivity is limited in early-stage or localized disease due to low tumor fractions. In our cohort, most samples had tumor fractions below the detection threshold for CNV or mutation-based approaches. This highlights a potential advantage of methylation-based cfDNA profiling, which may capture tumor-associated or microenvironmental signals even when ctDNA levels are minimal. In our current cohort, we could accurately predict CNVs in one patient with metastatic disease. Monosomy 3 and gain of 8q were observed in both the methylated cfDNA and the primary tumor DNA using IchorCNA, and confirmed through SNP data. The low tumor fraction raises an important question about the origins of the observed methylation signature in the cfDNA of UM patients. This could imply that the methylation alterations potentially reflect signals originating from cells in the tumor microenvironment [56] or tumor-associated processes, such as an immune response through increased cell turnover [50], apoptosis or necrosis [51] from leukocytes [27]. The MeD-seq data of monocytes, neutrophils, B and T lymphocytes exhibited a relatively large overlap in differentially hypermethylated regions with the cfDNA of UM patients. This suggests that a substantial portion of the methylation signal may originate from immune cells rather than tumor DNA. These findings indicate that the cfDNA methylation signature likely reflects a combination of tumor-derived and immune-related signals, possibly linked to tumor–immune interactions or systemic immune responses. Future studies could include paired leukocyte profiling from UM patients to clarify the relative contributions of these sources.

In a previous proof of concept study, we identified distinct methylation patterns in the primary tumors of UM patients and their matched metastases [37]. The prognostic utility of methylated tumor DNA has been demonstrated by a recent study that was able to distinguish patients at diagnosis who developed metastases in the future with an AUC of 81% [57]. In the present study, we included primary tumor DNA as well as cfDNA samples, both with and without tumor fractions. The circulating tumor DNA fraction is correlated with advanced disease stage and increased tumor burden [58, 59]. Global methylation patterns via PCA revealed that DMRs located on the gene body are associated with transcription [33], discriminated between cfDNA samples from patients with metastatic UM and HBDs, and partially distinguished localized cfDNA samples from localized disease. In contrast, PCA plots of methylation in the promoter region, which are associated with gene silencing [32, 34], did not distinguish between subgroups due to differences in PC values. This could be explained by the silencing effect of DNA methylation in the promoter region, while the methylation of gene body is associated with gene upregulation, also known as the DNA methylation paradox [32].

Hierarchical clustering revealed that the observed DMRs have the potential to differentiate UM patients from HBDs, and also subgroups of molecular subclasses within UM patients. Notably, only a limited number of DMRs were utilized to separate intermediate-risk from high-risk UM patients. The majority of these DMRs were located within a 23 kb region on chr.1 labeled as CpG island 3132. Close inspection via the UCSC Genome Browser of this region revealed segmental duplications of more than 1 kb on chr1:161,442,834 − 161,457,633, chr1:161,442,897 − 161,450,261, chr1:161,450,270 − 161,457,642, and chr1:161,457,634 − 161,472,424 [60, 61]. These duplicated regions overlap with seven of the eleven DMRs identified on chr. 1 in the cfDNA metastatic risk signature. Subsequent inspection using Integrative Genomics Viewer (IGV, default settings for alignment and hg38 as reference) revealed a relatively high number of sequenced reads on some loci corresponding with the locations from the DMRs. The methylation sequencing approach employed in this study provided approximately 50% genome-wide coverage of CpG sites, which is insufficient for CNV analysis. Additionally, CNV detection using IchorCNA is only accurate on samples with a high tumor fraction and was performed at the individual sample level rather than across patient subgroups, limiting its applicability in this context. In contrast, the remaining four DMRs, located at chr1:161,441,815 − 161,441,828, chr1:161,441,951 − 161,442,352, chr1:161,442,400 − 161,442,476, and chr1:161,442,518 − 161,442,554, do not overlap with this region with segmental duplications and may represent true discriminative markers. Importantly, beyond the duplications, CpG island 3132 also exhibits high CpG density and overlaps with regulatory regions identified in the ReMap and GeneHancer databases [62, 63], which exhibit long-range genomic interactions across multiple cell types. These features suggest that CpG island 3132 may function as an enhancer element that is frequently methylated in the cfDNA of patients with high-risk uveal melanoma. The discriminative ability of methylation profiling using LpnPI-digested cfDNA has been demonstrated in identifying presymptomatic carriers of frontotemporal dementia [64], patients with progression of renal cell carcinoma [65], and myelodysplastic syndrome patients [66], although these findings have not been replicated. Studies using other cfDNA methylation profiling techniques have reported positive implementations in lung cancer [67, 68], colorectal cancer [68, 69], hepatocellular carcinoma [68, 70] and breast cancer [71].

To gain a better understanding of our cfDNA signal, we performed a GSEA using the 207 DMRs of the UM cfDNA methylation signature. Among the 458 canonical and 435 disease pathways, eight significant canonical pathways were identified. The first to sixth ranked pathways are dopamine-DARPP32 feedback in cAMP signaling, orexin signaling, role of osteoclasts in rheumatoid arthritis signaling, the role of NFAT in cardiac hypertrophy, pancreatic secretion signaling, and wound healing signaling, which are primarily involved in processes such as neuronal activity, sleep regulation, bone resorption, cardiac hypertrophy, digestive enzyme regulation, and tissue repair, respectively. The underlying pathways of these processes may be utilized by UM or could be attributed to the difference between patients and controls, such as a mean age difference of 9.2 years, even though this difference was not statistically significant. The seventh- and eighth-ranked pathways are related to the immune system and cancer, supporting our finding that the leukocytes cluster with the methylation-specific signal for UM patients. These pathways included the S100 family signaling pathway(-log(p-value) = 1.42, Z-score=−2.828) and the RAF/MAP kinase cascade (-log(p-value) = 1.41, Z-score=−2.00). Studies have shown that S100A13 is involved in promoting cell proliferation, migration, and invasion in UM cells; knockdown of S100A13 in UM cell lines has been found to inhibit these processes, indicating its role in tumor progression [72]. Additionally, S100 family proteins are known to be involved in various cellular processes, such as cell cycle regulation and differentiation, which are critical in cancer development [73]. The RAF/MAP kinase pathway, also known as the MAPK/ERK pathway, plays a crucial role in cell division, differentiation, and survival. In UM, mutations in genes such as GNAQ and GNA11, which are upstream regulators of the MAPK pathway, lead to its constitutive activation [74]. This activation promotes tumor growth and progression by driving the proliferation and survival of melanoma cells [75]. RASAL3, RASA3, and DLG2 were identified as part of the MAPK pathway, which showed significant enrichment in our analysis (Table S5). RASAL3 is an activating protein of Ras GTPase that is highly expressed in neutrophils that negatively regulates neutrophil activity to modulate the inflammatory response [76]. Similarly, RASA3 is also a GTPase activating protein of the GAP1 family and highly expressed in type 17 T-helper cells [77, 78]. DLG2 is hypothesized to play a role in apoptosis acting as a tumor suppressor gene for neuroblastoma [79] and bone tumors [80]. The RAF/MAP kinase cascade and the S100 family signaling pathway are interconnected through their roles in cell signaling and tumor progression. The activation of the MAPK pathway can influence the expression and activity of S100 proteins, which in turn can affect cellular processes such as proliferation and migration [81].

Our results indicate that cfDNA methylation profiles obtained via MeD-seq of the methylated reads can clearly distinguish healthy controls from patients, whereas CNV profiling on the background reads using IchorCNA can estimate the molecular subclasses in patients with metastatic disease and detectable tumor fraction. Overall, methylation and CNV profiling show promise for minimally invasive prognostication in UM patients when tumor tissue is unavailable due to eye-sparing treatments. However, several limitations warrant consideration. Tumor fraction estimation using IchorCNA on the methylated background reads may lead to false-negatives in EIF1AX cases, as it relies the presence of copy alterations, which are typically absent in these tumors. Targeted mutation analysis assays, such as digital PCR could be a viable alternative with increased sensitivity in detecting EIF1AX-mutated tumors [2, 21, 82]. Most importantly, only a limited number of samples were available for this pilot study, which was sufficient to establish several methylation signatures and evaluate them with additional samples, but our findings should ideally be replicated in an independent cohort. While the identified DMRs show promise for distinguishing between high- and intermediate-risk UM patients, these findings remain preliminary. Validation was not feasible within the scope of this work due to limited availability of additional cfDNA samples from UM patients, which is a common challenge in rare tumor research. Consequently, the proposed DMRs should be interpreted as hypothesis-generating rather than definitive classifiers. Nevertheless, the number of samples used to investigate the metastatic risk signature were greater than the minimum of five samples, but less than the ideal number of ten samples per subgroup [38]. As a result, the number of DMRs may have been limited, preventing a comprehensive GSEA of the DMRs including CpG island 3132. Future studies could prioritize replication in larger, independent cohorts and assess the reproducibility and predictive performance of these methylation signatures. Such efforts will be critical to determine whether these DMRs can ultimately serve as clinically useful biomarkers for risk stratification in UM. Additionally, they could also evaluate the specificity of these signatures by including cfDNA from patients with other ocular conditions or malignancies that may present with similar clinical features. Finally, our study obtained methylation signatures primarily originating from non-malignant blood cells of UM patients. This may explain the absence of tumor-derived markers reported in the literature within our dataset. Future studies should explore other sources of tumor-derived cfDNA such as aqueous or vitreous humor, as prior research suggests these fluids may have a higher tumor DNA content than plasma, potentially enhancing the sensitivity of MeD-seq analyses [21, 82].

Conclusion

In conclusion, we demonstrated the potential of MeD-seq in differentiating between subgroups of molecular subclasses within UM patients and distinguish them from healthy individuals based on their methylation signatures. However, due to our limited sample size, we were unable to fully explore the utility of cfDNA methylation as a potential biomarker in UM patients. Gene enrichment analyses suggested possible associations with the S100 and RAF/MAP pathways, which are linked to the immune processes and drive UM cell proliferation and survival. Although the results of this pilot study need to be independently replicated, we believe that genome-wide methylation profiling may hold potential for future clinical use in UM patients.

Abbreviations

AJCC

The American Joint Committee on Cancer

BAP1

BRCA1 associated protein 1

cfDNA

cell-free DNA

CNV

Copy Number Variation

ctDNA

circulating tumor DNA

CYSLTR2

Cysteinyl leukotriene receptor 2

DMR

Differentially Methylated Regions

EIF1AX

Eukaryotic translation Initiation Factor 1 A X-linked

GNA11

G protein subunit alpha 11

GNAQ

G protein subunit alpha q

HR

High metastatic risk

IHC

Immunohistochemistry

IR

Intermediate metastatic risk

LR

Low metastatic risk

MeD-seq

Methylated DNA sequencing

PLCB4

Phospholipase C beta 4

SF3B1

Splicing factor 3b subunit 1

TF

Tumor fraction

UM

Uveal melanoma

Author contributions

Conceptualization: A.d.K, E.K. and E.B; acquisition: E.K. and E.B.; next generation sequencing: W.F.J. v IJ.; analysis: M.W., D.P.d.B., R.G.B., D.H.; interpretation of data; M.W., D.P.d.B., R.G.B., J.V.; creation of new software used in the work: R.G.B., D.H., J.B.; drafted the work or substantively revised it: M.W., D.P.d.B., D.H., A.B.B., S.S.M., R.O.B.D.K., R.M.V., N.C.N., D.P., S.W., E.S.G, W.F.J.v.IJ., J.G., A.d.K, E.K. and E.B. All authors have approved the submitted version and have agreed both to be personally accountable for the author’s own contributions and to ensure that questions related to the accuracy or integrity of any part of the work, even those in which the author was not personally involved, are appropriately investigated, resolved, and the resolution documented in the literature.

Funding

CORR 7.2.0, Biomarker High Risk Uveal Melanoma.

Data availability

The non-identifiable personal information used in this study is available in this published article and its supplementary information.

Declarations

Ethics approval and consent to participate

This study adheres to the tenets of the Declaration of Helsinki and has been approved by the Erasmus MC medical ethics committee (MEC-2009-375), Human Research Ethics Committee protocols of Edith Cowan University (No. 11543 and No. 18957) and Sir Charles Gardner Hospital (No. 2013 − 246 and No. RGS0000003289), Western Australia.

Consent for publication

All named authors meet the International Committee of Medical Journal Editors (ICMJE) criteria for authorship for this manuscript, take responsibility for the integrity of the work as a whole, and have given final approval for the version to be published.

Competing interests

R.G. Boers, J. Boers, J. Gribnau and W.F.J. van IJcken are shareholders of Methylomics BV. All the other authors declare that they have no conflicts of interest.