Genome-wide DNA methylation analysis reveals a unique methylation pattern for pleural mesothelioma compared to healthy pleura and other lung diseases

Abstract

Background

Pleural mesothelioma (PM) is a rare and aggressive cancer type, typically diagnosed at advanced stages. Distinguishing PM from other lung diseases is often challenging. There is an urgent need for biomarkers that can enable early detection. Interest in the field of epigenetics has increased, particularly in the context of tumour development and biomarker discovery. This study aims to identify specific changes in DNA methylation from healthy pleural tissue to PM and to compare these methylation patterns with those found in other lung diseases.

Results

EPIC methylation array data (850 K) were generated for 11 PM and 29 healthy pleura in-house collected samples. This is the first time such a large dataset of healthy pleura samples has been generated. Additional EPIC methylation array data (850 K) for pleural mesothelioma and other lung-related diseases were downloaded from public databases. We conducted pairwise differential methylation analyses across all tissue types, which facilitated the identification of significantly differentially methylated CpG sites. Extensive differential methylation between PM and healthy pleura was observed, identifying 81,968 differentially methylated CpG sites across all genomic regions. Among these, five CpG sites located within four genes (MIR21, RNF39, SPEN and C1orf101) exhibited the most significant and pronounced methylation differences between PM and healthy pleura. Moreover, our analysis delineated distinct methylation patterns specific to PM subtypes. Finally, the methylation profiles of PM were distinctly different from those of other lung cancers, enabling accurate differentiation.

Conclusions

DNA methylation analyses provide a robust method for distinguishing PM from healthy pleural tissues, and specific methylation patterns exist within PM subtypes. These methylation differences underscore their importance in understanding disease progression and may serve as viable biomarkers or therapeutic targets. Moreover, differential methylation patterns between PM and other lung cancers highlights its diagnostic potential. These findings necessitate further translational studies to explore their clinical applications.

Introduction

Malignant mesothelioma is a rare and aggressive cancer type emerging from the mesothelium, which borders the surface of several organs, among which the lungs. Around 80% of all mesothelioma cases originate from the pleura, the mesothelial layer covering the lungs, and are called pleural mesothelioma (PM) [1]. PM is divided into three main subtypes based on its histomorphological growth pattern: epithelioid (55%), sarcomatoid (15%) and biphasic (30%), the latter having characteristics of both epithelioid and sarcomatoid PM.

The most important factor causing the development of PM is asbestos exposure, accounting for more than 80% of cases [2]. Inhaled asbestos fibres are deposited in the lungs and can cause chronic inflammation and frustrated phagocytosis, leading to DNA damage and tumour formation [3]. The latency period between first asbestos exposure and disease diagnosis is estimated to be 30–50 years. Describing the worldwide burden of PM is challenging, as there is invalid reporting and inaccurate diagnoses in many countries. At least 30,870 mesothelioma deaths worldwide were estimated in 2020, with the highest numbers in Northern Europe [4]. Due to the ban on the use of asbestos in several countries, the incidence rate has recently decreased for the first time [4]. Currently, the standard treatment has a palliative intent for most cases, and the prognosis is poor, with a median life expectancy of 10–12 months for untreated patients and 18 months for patients treated with immunotherapy [5].

The diagnosis of PM is difficult due to several reasons, including slow growth and non-specific presenting symptoms [6]. This results in an important diagnostic delay, often leading to diagnosis at an advanced disease stage. Moreover, PM is hard to differentiate from other benign or malignant lung-related diseases such as lung cancer. Currently, the initial diagnostic step for PM involves conducting a computed tomography (CT) scan [7, 8]. However, a chest CT scan is only able to visualize the tumour once it reaches a certain size [2]. The shape of a PM tumour also impedes visualization, because of its diffuse and non-spherical growth pattern along the lungs [9]. Moreover, conventional imaging methods, such as chest CT scans, are unreliable in distinguishing PM tumours from other benign asbestos-related lesions [2]. The histological examination of a biopsy sample by pathologists is the final challenging step towards the diagnosis of PM [10]. Initially, the cellular origin is determined using a targeted panel of mesothelial (e.g. WT1 and calretinin) and epithelial (e.g. MOC-31 and claudin-4) immunomarkers, for which sarcomatoid PM samples are negative [10, 11]. After confirmation of the mesothelial origin, histologic morphology is assessed. However, the morphologic overlap between PM and benign mesothelial lesions complicates a reliable distinction. Therefore, additional immunohistochemical markers (e.g. for loss of BAP1 and MTAP expression) and fluorescence in situ hybridization (e.g. for homozygous deletion of CDKN2A) are used to assist this decision [10]. Unfortunately, the absence of these specific markers is not consistent across all PM cases, thereby reducing the sensitivity of the diagnostic process. Another difficulty is both inter- and intra-tumour heterogeneity [8]. This diversity appears in both morphological and molecular dimensions. Molecular analyses elucidate notable heterogeneity among patients, within distinct regions of a given tumour concerning different clonal compositions, and throughout the treatment trajectory [12]. All these facts stress the need for biomarkers for early detection of PM with high specificity and sensitivity and for possible treatment allocation in case of an actionable target [6].

The field of tumour genetics is promising for the identification of novel biomarkers. Genetic alterations, including mutations in suppressor genes, such as BAP1 (45.1%), CDKN2A (42.2%), CDKN2B (36.0%), NF2 (31.3%), and MTAP (27.3%), have been reported in PM [13]. However, mutations are not suitable as general biomarkers for PM, as PM exhibits a heterogeneous genetic landscape and a low somatic mutational burden. Furthermore, no oncogenic driver mutations have been identified [14]. This could be explained by the fact that asbestos is a non-mutagenic carcinogen [15]. Therefore, interest in the epigenome has grown for the elucidation of tumour development and the identification of biomarkers [3]. The epigenome consists of histone modifications (including acetylation, methylation and phosphorylation) and DNA alterations (including methylation) without changing the genetic nucleotide sequence. An important epigenetic modification that is considered to be a hallmark of cancer is DNA methylation [16, 17].

DNA methylation involves the addition of a methyl group on the fifth position of cytosine in a CpG dinucleotide context, resulting in 5-methylcytosine [16]. Regions with a high frequency of CpG sites are called CpG islands and are often located in promotor regions of genes. In general, global epigenetic reprogramming is observed in all kinds of tumour types [17]. Overall, a global loss of methylation is detected in tumour cells, leading to genomic instability [17]. In contrast, hypermethylation of specific CpG islands can lead to gene silencing or inactivation, as observed in tumour suppressor genes. Conversely, hypomethylation of a CpG island in a promotor region can lead to gene activation or overexpression, potentially involving oncogenes [16]. Until now, most research has focused on DNA methylation of individual genes in PM [18]. Epigenome-wide analyses on PM samples have been sparse until two research groups recently published DNA methylation analyses performed on a large cohort of PM patients [11, 19]. Bertero et al. aimed to discriminate PM from neoplastic and reactive histologic mimics, and Jurmeister et al. built a classification model to distinguish PM from chronic pleuritis, pleural carcinosis, and pleomorphic lung carcinomas. Neither research group included a control group comprising healthy pleural tissue samples. In this study we aimed to elucidate specific changes from healthy pleural tissue towards PM, necessitating the use of such a control group. Furthermore, variations in methylation patterns between PM and other lung-related diseases were studied. These alterations in DNA methylation patterns, when compared to those in pleural tissue and other lung diseases, hold the potential to serve as diagnostic biomarkers for PM.

Methods

Sample collection

PM tumour tissue samples were routinely collected between 2012 and 2021 by the biobank of the Antwerp University Hospital (UZA, Belgium). The UZA ethical committee approved our study and permitted the use of retrospective samples (Reference number 16/23/248). All tissue samples were fresh frozen and stored at -80 °C after collection until further use. Diagnosis and overall tumour percentage were verified and determined by a pathologist (D.P.) by histological examination of hematoxylin–eosin-stained sections. We used 11 fresh frozen PM tumour tissue samples from treatment naïve patients, collected between 2013 and 2020. All samples had an overall tumour cell percentage (TCP) between 40 and 80% (Suppl Table S1).

For this study, 29 healthy parietal pleural tissue samples were prospectively collected from treatment naïve patients after approval of the UZA ethical committee (EDGE number 002046) in the Antwerp University Hospital in 2021 and 2022. The samples were collected from patients who underwent thoracic surgery for another reason than mesothelioma (Suppl Table S1). All patients gave written informed consent. All tissue samples were fresh frozen and stored at -80 °C after collection until further use. The samples were verified by a pathologist (D.P.) by histological examination of hematoxylin–eosin-stained sections.

DNA extraction and methylation analysis

DNA was extracted from ten to fifteen 10 µm-sections from the tissue samples, depending on the tissue size, using the QIAamp DNA Micro Kit (Qiagen, Hilden, Germany) according to the manufacturer’s protocol. The DNA was stored at -20 °C until further usage. 500 ng of each DNA sample was bisulfite converted with an EZ DNA methylation kit (Zymo Research, California, USA) following the manufacturer’s protocol specifically for a downstream analysis with an Infinium Methylation Microarray. Genome-wide methylation profiles were obtained using an Infinium Methylation EPIC 850 K BeadChip Kit both v1.0 and v2.0 (Illumina, California, USA) according to the manufacturer’s protocol.

Methylation data processing and differential methylation analysis

Raw intensity array files were processed using the ChAMP (2.29.1) Bioconductor package [20, 21]. Methylation status was reported as Beta values, which range from 0 (indicating no methylation) to 1 (indicating full methylation). After data read-in, samples with more than 5% of their probed data missing were excluded. Underperforming probes were filtered out from the downstream analysis; this included control probes, X‐/Y‐chromosome probes, multihit probes, and probes with known single nucleotide polymorphisms (SNPs). The remaining probes with missing values were also removed and Beta values less than 0 were set at 0 and values above 1 were set at 1. To assess data quality, we calculated both the log2 median intensity ratios (for methylated and unmethylated signals) and the density of Beta values. To minimize technical discrepancies between Type-I and Type-II Illumina probes, BMIQ normalization was applied. Differential methylation was also conducted using ChAMP, which employs parametric linear mixed models to assess variations in methylation across different groups. The first criterion for identifying differentially methylated CpG sites (DMCs) was an adjusted P‐value (adj.P.Val) ≤ 0.05, corrected for multiple testing using the Benjamini–Hochberg method [22]. The delta beta for each probe was defined as the difference in mean Beta value between group A and group B. A positive delta beta indicates higher methylation in group A (hypermethylation), while a negative delta beta reflects lower methylation in group A (hypomethylation). DMCs with an absolute delta beta value ≤ 0.05 were filtered out to reduce false positives. Lastly, we assessed the significance of batch effects across the CpG sites by comparing the fit of an ordinary least square regression model (with group as independent variable) without a random effect to that of a linear mixed model, incorporating batch as a random effect and group as a fixed effect. For each CpG site, we then tested whether the random effect was significant using a likelihood ratio test. Under the null hypothesis that no batch effect exist, the p-values from the random effect should follow a uniform U(0,1) distribution. After calculating the significance of the random effect for all CpG sites, we compared the observed p-value distribution with the expected null distribution. CpG sites whose p-value distribution deviated from the null were subsequently excluded from the final DMP list. Differentially methylated regions (DMRs) were identified using the Bumphunter algorithm extension in ChAMP. The significance of DMRs was determined based on the harmonic mean of the individual CpG P-values (FDR-corrected). DMRs with a mean difference ≤ 0.05 were filtered out to reduce false positives. All genomic locations are reported using Genome Build 19/ GRCh37.

Publicly available datasets

The MESOMICS Methylation EPIC 850 K (EGAD00010002053) and the LungNENomics Methylation EPIC 850 K (EGAD00010001720) datasets were downloaded from the European Genome-phenome Archive (EGA) after a Data Access Agreement with the owners [23, 24]. All other used datasets were downloaded from the Gene Expression Omnibus (GEO) and are listed in Table 1. To maintain uniform data processing, we chose to use raw idat files over processed beta values. The datasets were then processed using the same steps described above.

Table 1 Datasets and samples used in this study

Statistical analyses

Subsequent analyses were carried out using the R software (version 4.3.1). For exploratory and visualization purposes, dimensionality reduction using t-distributed stochastic neighbour embedding (tSNE) clustering was performed using the Rtsne package with the 100,000 most variable probe Beta values. All adjusted P-values are two‐sided, and those ≤ 0.05 were considered statistically significant. All genomic annotations were done using the GRCh37/hg19 genome build.

Results

Generation of the methylation data matrix

We performed a meta-analysis, combining our in-house generated EPIC methylation dataset consisting of 11 PM tissue samples and 29 healthy pleural tissue samples with 10 online available datasets of EPIC methylation data from PM and other lung diseases (Table 1). Collecting healthy pleural samples is challenging and to the best of our knowledge, this is the first time such a large dataset of healthy pleura samples has been generated. After quality control, 5 PM samples (dataset GSE164269) and 1 LUAD sample (dataset GSE203061) were excluded due to an excess of failed CpG sites (cut-off failed fraction of 0.05). Of the 865,917 examined CpG sites on the EPIC methylation array, 551,236 CpG sites have a successful measurement for each of the samples and were retained for further analysis (Suppl Table S2). The final data matrix consisted of DNA methylation data for 551,236 CpG sites on 257 pleural mesothelioma (PM) samples, 32 healthy pleura (PL) samples, 7 chronic pleuritis (CP) samples, 90 lung adenocarcinoma (LUAD) samples, 77 lung squamous cell carcinoma (LUSC) samples, 56 lung carcinoid (LUCA) samples, 20 large cell neuroendocrine carcinoma (LCNEC) samples, and 44 lung adjacent normal (LAN) samples.

Differential methylation between pleural mesothelioma and healthy pleura

First, we explored the methylation data of the PM and the healthy pleural tissue using a t-distributed stochastic neighbour embedding (tSNE) plot. Figure 1 displays a nearly perfect division between healthy pleura and PM samples. The healthy pleura samples are clustered closely together, while the PM samples are more widespread. Three healthy pleura samples are visualized at the boundary of the PM cluster. We examined the association between tumour cell percentage (TCP) of each sample and the methylation pattern, but we found no correlation (Suppl Figure S1).

Fig. 1

t-distributed stochastic neighbour embedding (tSNE) plot of the methylation patterns based on the 100,000 most variable CpG sites of 257 PM samples and 32 healthy pleura samples

Differential methylation was identified in distinct genomic regions

Next, we investigated differential methylation patterns of individual CpG sites between PM tissue and healthy pleural tissue. Of the 551,236 examined CpG sites, we found 81,968 significantly differentially methylated CpG sites (DMCs). Of these, 50,346 are located in 14,745 unique protein-coding genes, 421 DMCs in 223 microRNA-coding genes, 1,178 DMCs in 569 lncRNA-coding genes and 30,023 in intergenic regions. For all DMCs, the difference in methylation level (delta beta) between the two groups was calculated. Of all DMCs, 63,463 CpG sites are hypomethylated in PM (negative delta beta), and 18,505 CpG sites are hypermethylated in PM (positive delta beta). The only genomic locations with clear hypermethylation, are within CpG islands, especially in the intergenic regions (IGRs), in the 1500 bp and 200 bp fragment before the transcription start site (TSS200), the 5’UTR region, and the first exon (Fig. 2). In all other regions, hypomethylation predominates, especially in the open sea regions (i.e. not in proximity of CpG islands). Delta beta values range between −0.549 and 0.568.

Fig. 2

Boxplot showing the delta beta distribution of each location category for all DMCs between PM and healthy pleura. Categories are based on genomic location (y-axis) and relation to a CpG island (colours). ExonBnd, exon boundaries; IGR, intergenic region; TSS, transcription start site; UTR, untranslated region

Most significant and differentially methylated CpG sites

To identify interesting CpG sites, we calculated two top 50 DMC lists, one ranked on adjusted P-value and one ranked on delta beta, as both parameters are essential for discrimination (Suppl Tables S3 and S4). Afterwards, we determined the overlap between both lists. This way, five CpG sites in four genes were identified: MIR21, RNF39, SPEN and C1orf101 (Fig. 3 and Table 2). The area under the ROC curve (AUC) for the discrimination between PM and healthy pleura for these DMCs ranges between 0.986 and 0.994.

Fig. 3

Volcano plot showing delta beta (x-axis) and adjusted P-value (y-axis) for all 81,968 DMCs between PM and healthy pleura. The five overlapping DMCs of the top 50 for delta beta (blue) and adjusted P-value (green) are coloured orange

Table 2 The five overlapping DMCs of the top 50 ranked on delta beta and the top 50 ranked by adjusted P-value, between pleural mesothelioma (PM) and healthy pleura (PL)

We examined these four genes in more detail (Fig. 4). For MIR21 (MicroRNA 21), all four examined CpG sites are hypomethylated and demonstrate a large delta beta (0.429, 0.482, 0.469 and 0.418). Two of them are located in the TSS200, the other two in the gene body. In RNF39 (Ring Finger Protein 39), 28 of the 54 examined CpG sites are differentially methylated. There are 14 DMCs with a delta beta higher than 0.2 (hypermethylated). All of them are located in a CpG island at the end of the gene body. Notably, all DMCs in the promotor region are hypomethylated. However, in that region, no large delta beta values were detected. In SPEN (Spen Family Transcriptional Repressor), 11 of the 36 examined CpG sites are differentially methylated. Most of the DMCs are hypomethylated and DMCs with the highest absolute delta beta are located at the end of the gene body. In the promotor region, there is only one significant CpG site. For C1orf101 also known as CATSPERE (Catsper Channel Auxiliary Subunit Epsilon), seven of the 16 examined CpG sites are differentially methylated. Most of the DMCs are hypomethylated and DMCs with the highest absolute delta beta are located at the end of the gene body. In the promotor region, there is only one significant CpG site.

Fig. 4

Scatter plot showing the delta beta (difference in the methylation level) between PM and healthy pleura for each examined CpG site in the genes A MIR21, B RNF39, C SPEN, and D C1orf101. The dark grey points indicate the significantly methylated DMCs, the light grey points indicate the non-significant CpG sites and the orange arrow indicates the genomic location and the direction of transcription

Comparison with COSMIC genes

We compared our list of DMCs with the COSMIC list (Catalogue of Somatic Mutations in Cancer) [25]. COSMIC presents a list of 743 genes with mutations causally implicated in cancer, including oncogenes, tumour suppressor genes, and fusion genes. Of this list, 677 genes are examined by the EPIC methylation array. Among the 81,968 DMCs we identified between PM and healthy pleura, 2,973 CpG sites are located in 556 COSMIC genes. For these genes, we calculated which ones have a high ratio of significant DMCs compared to the total examined CpG sites for that gene, excluding genes with 10 or fewer examined CpG sites. One gene has a ratio of significant DMCs to examined CpG sites exceeding 0.5, which is PDCD1LG2 (Programmed Cell Death 1 Ligand 2). Notably, all eight significant DMCs of PDCD1LG2 are hypomethylated and are located in all regions of the gene (5’UTR, TSS1500, TSS200, first exon, and gene body).

Differentially methylated regions

In addition to analysing DMPs, we extended our investigation to differentially methylated regions (DMRs) between PM and healthy pleura. Through our analysis, we identified a total of 3,078 DMRs. Of these, 1,494 regions are hypermethylated in PM, while 1,584 regions are hypomethylated. To enhance the specificity of our findings, we applied a minimum cut-off value of 0.1 to the mean differential methylation. This more stringent criterion resulted in a refined list of 1,053 DMRs. Among these, the five most significant DMRs overlap with the genes RNF39, SPEG, PISD, SEPT9, and PVT1, respectively (Table 3).

Table 3 Five most significant differentially methylated regions between pleural mesothelioma and healthy pleura

Differential methylation between pleural mesothelioma subtypes

We compared the methylation patterns of the three PM subtypes: epithelioid (n = 157), biphasic (n = 57) and sarcomatoid (n = 43). In the tSNE plot (Fig. 5), the subtypes are not grouped in clearly separate clusters. However, the epithelioid PM and the sarcomatoid PM clusters are notably distinct from each other, whereas the biphasic PM samples cluster between the other two groups. Moreover, we found 182,840 DMCs between epithelioid and sarcomatoid PM, 93,871 DMCs between epithelioid and biphasic PM, and 52,139 DMCs between biphasic and sarcomatoid PM (Fig. 6A). The majority of the DMCs are hypermethylated in epithelioid PM compared to biphasic and sarcomatoid PM. The largest amount of DMCs are differentially methylated between epithelioid and sarcomatoid PM, and a large part of these DMCs (n = 61,825) is also differentially methylated between epithelioid and biphasic PM (Fig. 6B). Of all DMCs between the subtypes, 9,157 are overlapping in each comparison, i.e. 9,157 CpG sites are significantly differentially methylated between each of the three groups.

Fig. 5

t-distributed stochastic neighbour embedding (tSNE) plot of the methylation patterns based on the 100,000 most variable CpG sites of 157 epithelioid PM samples, 57 biphasic PM samples and 43 sarcomatoid PM samples

Fig. 6

Number of DMCs between the three PM subtypes. a Barplot showing the numbers of DMCs higher methylated in each of the subtypes and each of the comparisons. b Venn diagram showing the overlapping DMCs in each of the comparisons. The darker the grey, the higher the amount of DMCs

Next, we compared the methylation level of the three subtypes separately with healthy pleura. We found 208,938 DMCs between pleura and epithelioid PM, 259,849 DMCs between pleura and biphasic PM, and 280,785 DMCs between pleura and sarcomatoid PM (Fig. 7A). The vast majority of these DMCs are hypermethylated in pleura compared to each of the PM subtypes. For the 9,157 overlapping DMCs between the subtypes, the distribution of the methylation levels for the three subtypes and healthy pleura is shown in a violin plot (Fig. 7B). This plot shows clearly that the methylation levels are highest in healthy pleura, lower in epithelioid PM, even lower in biphasic PM, and lowest in sarcomatoid PM.

Fig. 7

a Barplot showing the numbers of DMCs higher methylated in each of the subtypes or healthy pleura, and each of the comparisons. b Violin plot showing the distribution of the methylation levels in each of the subtypes and healthy pleura for the 9,157 overlapping DMCs between the subtypes

Differential methylation between pleural mesothelioma and other lung diseases

We compared the methylation pattern of PM with those from the other lung-related samples (PL, CP, LUAD, LUSC, LUCA, LCNEC, LAN), aiming to detect tissue and tumour type-specific methylation patterns. In the tSNE plot (Fig. 8), most of the groups form a clear cluster. In dimension 1, the pleura-related samples are perfectly distinguishable from lung-related samples, while dimension 2 discriminates based on the disease state. For each comparison, the number of DMCs was calculated (Suppl Table S5). More than 70,000 DMCs were identified for each comparison, except for the comparison with chronic pleuritis, where 13,834 DMCs were detected. Of all DMCs, 2,204 CpG sites are differentially methylated between PM and each of the other groups. In a violin plot showing the average methylation levels of those common DMCs for each of the groups (Fig. 9), it is evident that most of these DMCs are hypomethylated in PM, as the density of PM is much higher around a methylation level of 0.4, while in all other groups, the methylation level is denser for higher values.

Fig. 8

t-distributed stochastic neighbour embedding (tSNE) plot of the methylation patterns based on the 100,000 most variable CpG sites of 257 PM samples, 32 healthy pleura samples, 7 chronic pleuritic samples, 90 lung adenocarcinoma samples, 77 lung squamous cell samples, 56 lung carcinoid samples, 20 large cell neuroendocrine carcinoma samples, and 44 lung adjacent normal samples

Fig. 9

Violin plot showing the distribution of the methylation levels in PM and other lung-related samples for the 2,204 overlapping DMCs between all the groups. CP, Chronic pleuritis; LAN, Lung adjacent normal; LCNEC, Large Cell Neuroendocrine Carcinoma of the Lung; LUAD, Lung adenocarcinoma; LUCA, Lung carcinoid; LUSC, Lung squamous cell carcinoma; PM, Pleural Mesothelioma; PL, Pleura

Discussion

Pleural mesothelioma is a rare and rapidly fatal disease, with asbestos exposure standing out as the foremost risk factor contributing to its development. Previous research has demonstrated the presence of somatic mutations in PM [13]. However, PM exhibits a relatively low number of pathogenically significant mutations [26]. Therefore, the epigenome is now extensively being investigated. The precise mechanism behind the prominence of epigenetic alterations in PM remains elusive. However, it is widely recognized that chronic inflammation is a primary response to asbestos exposure. Notably, epigenetic modifications have been closely associated with inflammatory processes in other cancers [27]. This suggests that inflammation-related epigenetic changes may play an important role in various human cancers, including pleural mesothelioma.

Currently, the investigation of genome-wide methylation changes from healthy pleura to PM has been limited. Our study aims to further elucidate the epigenetic landscape of PM. Surprisingly, widespread differential methylation between PM and healthy pleura was observed, considering that we detected 81,968 differentially methylated CpG sites. These DMCs are located in all genomic regions, including noncoding RNA, which is perhaps not surprising as methylation also plays an important role in the regulation of noncoding RNA transcription, in addition to protein-coding genes [28]. MiRNAs and lncRNAs are regulators of cellular processes such as differentiation and proliferation, and aberrant methylation of these types of RNA can lead to cancer development [28]. Unfortunately, the noncoding genomic regions are underrepresented on the EPIC methylation array, despite being crucial in the epigenetic regulation. This is due to a selection process primarily focusing on genes implicated in cancer. This obstacle could potentially be addressed through future implementation of genome-wide ONT sequencing. Unexpectedly, a large number of DMCs was found in intergenic regions. It is still unclear whether these aberrations are tumour-causing or a consequence.

Actively transcribed genes in healthy human cells are known to have unmethylated promotors, especially when they contain a CpG island, and slightly methylated gene bodies [28]. In tumour cells, however, a global loss of DNA methylation is described in several studies. On the other hand, CpG islands and shores in gene promotors are often hypermethylated in tumoural DNA, resulting in gene silencing of tumour suppressor genes [28, 29]. Roughly 50% of all human genes harbour CpG islands, and are therefore susceptible to this type of aberrant silencing [30]. Moreover, hypomethylation of promotor regions without CpG island has been observed as well and can result in overexpression of oncogenes and proto-oncogenes, although this occurs less frequently [28]. This is in agreement with our observations (Fig. 2). When comparing healthy pleura to PM, the only hypermethylated locations are the CpG islands (in all gene locations and IGRs). Other locations are hypomethylated in all gene locations and IGRs.

Christensen et al. performed a genome-wide methylation analysis comparing PM samples with healthy pleura [15]. However, they investigated only 1505 CpG sites associated with 803 cancer-related genes using the Illumina GoldenGate BeadArray technology and compared 158 PM samples to only 18 healthy pleura samples. They identified 969 CpG sites in 646 genes that are aberrantly methylated in mesothelioma samples. Of these genes, two overlap with our Top 50 DMCs ranked on delta beta: FHIT and PLXDC2. Surprisingly, in our analysis, FHIT has a hypermethylated CpG site in PM (in the promotor region), while in Christensen’s study, FHIT has two hypomethylated CpG sites (one in the promotor and one in the body). The Fragile Histidine Triad Diadenosine Triphosphatase (FHIT) gene encodes for a triphosphate hydrolase that is involved in purine metabolism. In addition, FHIT is a tumour suppressor gene crucial for DNA repair, cell cycle regulation, and apoptosis. It has been suggested to be a target of lung carcinogens, such as tobacco smoke and asbestos [31]. FHIT has been described to have a reduced protein expression in PM [32, 33]. In childhood acute lymphoblastic leukaemia, FHIT is defined to have both promotor hypermethylation and reduced mRNA expression [34]. PLXDC2 encodes for Plexin Domain Containing-Protein 2 and is also known as TEM7R (tumour endothelial marker 7-related). In both studies, PLXDC2 is hypomethylated. Dysregulation of PLXDC2 is associated with cancer progression and metastasis via roles in angiogenesis, cell migration, and invasion, for example in gastric cancer [35]. However, PLXDC2 has never been linked with mesothelioma before.

We selected four genes for a more detailed examination. MIR21 encodes for a microRNA which is a small RNA fragment that regulates gene expression of several genes at post-transcriptional level [36]. MIR21 has been described earlier to be over-expressed in PM [37]. This corresponds to our finding that the TSS200 of the MIR21 gene is hypomethylated in PM [38]. MicroRNAs are described to be stable in body fluids and as such have potential as biomarkers [39]. Expression of MIR21 is also been described to be a potential prognostic biomarker for PM [40]. However, a recent review describes how MIR21 is claimed to be a predictive or prognostic biomarker for at least 29 diseases [39]. This knowledge undermines the potential of MIR21 as a specific biomarker for PM.

RNF39 encodes for Ring Finger Protein 39 which is suggested to play a role in an early phase of synaptic plasticity and has potential E3 ubiquitin ligase activity [41]. The altered expression of RNF39 is described to be a potential prognostic biomarker for cholangiocarcinoma and pancreatic cancer [42, 43]. The RNF39 gene is located in a significantly hypomethylated region in peripheral blood mononuclear cells of breast cancer patients with DOX-induced cardiotoxicity [44]. Further, little is known about the functions of this gene in cancer [43].

SPEN encodes for Spen Family Transcriptional Repressor which is a hormone-inducible transcriptional repressor and is mainly involved in X chromosome inactivation [45]. It is also known as SMART/HDAC1-associated repressor (SHARP) [46]. A recent study describes the potential of SPEN mutations as a predictive biomarker for immunotherapy in a pan-cancer analysis [45]. Another study describes SPEN as a tumour suppressor gene and a candidate predictive biomarker in ERα-positive breast cancers [47]. However, until now no association between this gene and PM has been described.

C1orf101 is also known as CATSPERE and encodes for Catsper Channel Auxiliary Subunit Epsilon. This is an auxiliary component of the CatSper complex, a complex involved in sperm cell hyperactivation. A recent study identified CATSPERE as a mediator of colorectal cancer susceptibility and progression [48]. Furthermore, no associations of this gene with cancer are described.

DNA methylation occurs early in tumour development and can be detected in body fluids [29]. Therefore, it is an ideal source for biomarkers. The aberrant methylation that we identified in this study, can be further investigated as a potential diagnostic biomarker for the detection of PM. The five CpG sites we elucidated (Table 2) have an AUC ranging between 0.986 and 0.994, and therefore have high biomarker potential. In addition to biomarkers, DNA methylation could also be a source of therapeutic targets. DNA methylation is reversible and methyltransferase inhibitors, such as 5’-azacytidine, are currently being used to treat several cancer types, including acute myeloid leukaemia [29, 49]. Understanding the functional roles of particular genes in the context of PM pathogenesis could offer valuable insights into the underlying molecular mechanisms driving this aggressive cancer and may hold promise for the development of novel targeted therapies.

We analysed our findings in the light of established knowledge about cancer in general and PM specifically. The COSMIC database, which catalogues mutations, copy number alterations and other genomic changes in cancer, provides a valuable reference for our study. Notably, of the cancer genes listed in COSMIC with CpG sites examined by the EPIC array, 82% (556/677) exhibited altered methylation in PM. This suggests that methylation changes in PM are more ubiquitous than mutations, extending across the entire genome rather than being confined to a few driver mutations.

In addition to analysing DMPs, we examined DMRs between PM and heathy pleura to gain a broader perspective on epigenetic regulation. While cancer cells are generally more hypomethylated than normal cells, as confirmed by our DMP analysis, we observed an almost equal number of hypermethylated and hypomethylated DMRs. This suggests a complex and nuanced landscape of epigenetic regulation in PM cells. Although global hypomethylation is a common feature, these results indicate that epigenetic changes in PM are not uniform. They involve both the silencing of key regulatory genes through hypermethylation and the activation or destabilization of other regions through hypomethylation. Notably, the five most significant DMRs overlap with the genes RNF39, SPEG, PISD, SEPT9, and PVT1. The DMR overlapping RNF39 corresponds to the end of the gene body (width of 1213 bp), where we also identified several DMPs (Fig. 4). Surprisingly, the DMR overlapping SEPT9 (Septin 9) is hypomethylated, even though SEPT9 is frequently described as a hypermethylated biomarker in colorectal cancer and other cancer types [50, 51]. Long noncoding RNA plasmacytoma variant translocation 1 (PVT1) has been identified as playing an important role in cancer development as oncogene [52]. However, the genes SPEG and PISD have not been previously linked to cancer.

When comparing the methylation patterns of the three PM subtypes, no explicit separation can be visualized by clustering. However, a clear trend is observed in the methylation levels of the subgroups and healthy pleura (Fig. 7). The highest methylation levels are detected in healthy pleura, followed by epithelioid PM, subsequently biphasic PM, and finally, sarcomatoid PM, in which the lowest methylation levels are observed. Increased hypomethylation observed in tumours is often associated with more aggressive and malignant disease phenotypes [53]. This hypomethylation can lead to the activation of oncogenes and to genomic instability, which contributes to tumour progression and aggressiveness. Studies have shown a correlation between the extent of hypomethylation in other tumours and the severity of the disease, including increased invasiveness, metastatic potential, and resistance to therapy [54, 55]. Global DNA hypomethylation is also associated with a detrimental prognosis in tumour patients [56]. This is consistent with our findings, as sarcomatoid PM is described to have more distant metastases than other PMs, and has the worst prognosis [57]. Moreover, differential methylation between the PM subgroups could potentially be used for PM subclassification. This was already demonstrated on central nervous system tumours, for which the application of DNA methylation-based classification has been demonstrated in a routine diagnostic setting [58].

Based on the methylation patterns, PM can be accurately distinguished from other lung-related diseases. For all comparisons, we reported more than 70,000 DMCs except when compared to chronic pleuritis. Chronic pleuritis, also known as chronic inflammation of the pleura, is not universally recognized as a precursor stage of mesothelioma. While chronic inflammation can sometimes precede the development of mesothelioma, it does not necessarily indicate a direct progression from chronic pleuritis to mesothelioma in all cases. The relationship between chronic pleuritis and mesothelioma is complex and may involve various contributing factors beyond inflammation alone. Therefore, it is important for individuals with chronic pleuritis and a history of asbestos exposure to undergo regular medical evaluation and monitoring for the potential development of mesothelioma. Although only 13,834 DMCs were identified between chronic pleuritis and PM, we believe that the small sample size of the chronic pleuritis group (n = 7) resulted in the lack of power to identify more DMCs. Studies with larger sample sizes are needed to validate the potential to discriminate PM from chronic pleuritis based on the methylation pattern. One remarkable observation is the differential methylation in the RNF39 gene between PM and chronic pleuritis. Five CpG sites of the top 10 DMCs ranked on delta beta are located in the RNF39 gene (Suppl Table S6). The end of the gene body is extremely hypermethylated in PM compared to chronic pleuritis (Suppl Figure S2), which is very similar to Fig. 4B in which PM is compared to healthy pleura.

In addition to biomarkers differentiating PM from healthy pleura, it could become even more interesting when biomarkers can be identified to distinguish PM from similar diseases to improve diagnostic accuracy. Especially differentiation from different types of lung cancer is important, as these can present with similar symptoms, such as chest pain, shortness of breath, and coughing. Additionally, both PM and lung cancer can result from exposure to carcinogens, such as asbestos. Moreover, radiographic findings can overlap between PM and other lung cancers, with features such as pleural thickening, pleural effusion, and mass lesions within the lung or pleura. Finally, these similar disease entities frequently show the same expression of pathologic immunohistochemical markers as PM, which complicates diagnosis and may even result in the wrong diagnosis. DNA methylation profiles are very robust and disease-specific and could potentially resolve all mentioned difficulties for the diagnosis of lung diseases, leading to appropriate treatment planning and management for the correct diagnosis.

Differences in genome-wide methylation patterns between PM and lung cancer were already described by several research groups. Goto et al. conducted a comparison of 6157 CpG islands in 20 PM samples with 20 LUAD samples using methylated CpG island amplification microarray analysis [59]. Across all samples, an average of 387 genes exhibited hypermethylation in PM, whereas 544 genes showed hypermethylation in LUAD. Among the most noteworthy hypermethylated genes in PM were TMEM30B, KAZALD1, and MAPK13. Notably, these genes were unmethylated in LUAD samples, confirming their specificity for PM [59]. Bertero et al. utilized EPIC methylation arrays to analyse the methylation patterns of 79 PM samples in comparison to 202 cases representing malignant and benign diagnostic mimics [19]. Employing both unsupervised hierarchical clustering and t-distributed stochastic neighbour embedding analysis, PM samples exhibited a distinct DNA methylation profile compared to other neoplastic and reactive mimics [19]. Jurmeister et al. utilized both 450 K and EPIC methylation arrays to compare the methylation patterns among 196 PM, 507 LUAD, 413 LUSC, and 17 CP samples [11]. Through the application of two machine learning algorithms, the study attained high accuracies using their support vector machine (97.8%), while their random forest model exhibited lower performance (89.5%), particularly in distinguishing PM from CP. Furthermore, differential methylation analysis uncovered promoter hypermethylation in PM specimens, implicating tumour suppressor genes such as BCL11B, EBF1, FOXA1, and WNK2 [11].

Conclusion

In conclusion, the analysis of DNA methylation patterns emerges as a promising avenue for differentiation between various tissue and tumour types. The substantial number of methylation alterations observed between healthy pleura and PM underscores the importance of these epigenetic changes in the pathogenesis of the disease. These alterations could be used as biomarkers or molecular targets for therapy. Additionally, the ability to distinguish between different subtypes of PM based on their unique methylation profiles offers valuable insights for personalized diagnostic and therapeutic interventions as well as prognosis. Finally, the distinct methylation signatures exhibited by PM compared to other lung cancers highlight the potential of methylation profiling as a diagnostic tool in the clinical setting. Translational studies need to be conducted to enable the utilization of this methylation signature in clinical settings.