Background

After the introduction of cisplatin in the treatment of testicular cancer (TC) in the late 1970s [1], this malignancy has become a model for curative treatment even in case of metastatic disease. Cisplatin-based chemotherapy (CBCT) has been integrated into standard treatment of several cancers in addition to TC, including gynecological, lung, bladder, and head and neck cancer [2]. For men with metastatic TC, three to four cycles of cisplatin in combination with etoposide and bleomycin (BEP) comprise the cornerstone in the treatment of metastatic disease [3], yielding 5-year disease-specific survival rates > 90% [4].

Due to the excellent prognosis and young age at diagnosis, TC survivors can expect to live for 30–50 years after successful treatment [5]. However, the very long-term relative survival among TC survivors is lower than among the age-matched population [6], primarily related to increased risks of second cancers and cardiovascular disease (CVD) [5]. Metabolic syndrome (MetS) is a well described late effect after TC treatment and is a possible mediator of both the increased risk of second cancers as well as CVD [7,8,9]. The prevalence of MetS in the general population differs according to MetS definition and increases by age, affecting about 20–25% of most Western populations [10, 11]. MetS including hypertension, increased body mass index (BMI), pre-diabetic biochemical serum changes, and/or hyperlipidemia is a constellation of risk factors for CVD [12]. An increased age-adjusted odds ratio (OR) for developing MetS after CBCT has been found when compared with patients treated with only surgery [9, 13, 14], although these results are not quite consistent [15].

Cisplatin exposure has been shown to result in drug-induced DNA hypermethylation both in vitro and in vivo [16,17,18,19]. In recent years, evidence for epigenetic changes predisposing to MetS has also been documented [20,21,22]; it is therefore plausible that these changes caused by CBCT could be involved in the development of MetS. Identification of differential DNA methylation (DNAm) in TC survivors that develop MetS compared to those who do not could provide a better understanding for the underlying mechanisms behind this serious late effect.

We hypothesize that epigenetic changes caused by CBCT render TC survivors susceptible for developing MetS later in life. The aim of this study was to (i) evaluate the potential long-term effect of CBCT on the epigenome in a cohort of Norwegian TC survivors, and (ii) identify possible associations between epigenetic changes and development of MetS. We also investigated if these changes in (i) and (ii) links to genes, functional networks, and pathways related to MetS symptoms.

Results

The basic characteristics of the TC survivors are outlined in Table 1. Median age at diagnosis ranged between 27 and 30 years, while median age at SII ranged between 47 and 51 years. Mean β methylation was 0.62 in all four groups.

Table 1 Characteristics at diagnosis and follow-up for patients (N = 279), split by treatment group. Median values and range are reported for the groups
Full size table

Study confounders

Principal component analyses of the DNAm did not show differences between the four groups (Additional file 1: Figure S1). Global methylation was not associated with CBCT or MetS (p > 0.05) (Additional file 1: Figure S2). However, we identified associations between CBCT and relative proportions of CD4+ T cells (p = 0.0001), and CD8+ T cells (p = 0.04). Testosterone was also significantly associated with MetS (p = 1.6 E-07). We found 3109 Bonferroni significant CpGs (cytosine nucleotide followed by a guanine nucletide) (p value < 0.01) associated with age and 229 Bonferroni significant CpGs (p value < 0.01) associated with smoking habits, including two smoking related genes (Additional file 1: Figure S3) (Fig. 1).

Fig. 1
figure 1

CONSORT flow diagram of included samples that were analyzed with the MethylationEPIC BeadChip (n = 279). Samples were from testicular cancer survivors divided into four groups according to CBCT and MetS status. a CBCT cisplatin-based chemotherapy. b MetS metabolic syndrome

Full size image

Long-term effects of cisplatin on DNA methylation

After adjusting for age, testosterone, smoking, and cell count, 35 CpG sites were associated with CBCT after False Discovery Rate (FDR) correction (Table 2). Of these, 13 CpG sites were significantly associated with CBCT after Bonferroni correction (Fig. 2a). Of the FDR-associated CpGs, 21 were located in Open Sea, three in a CpG island, and 11 in either CpG island shelf or shore. The different multivariate models showed similar results (Additional file 1: Figure S4B and Figure S5B). Nineteen annotated CBCT genes were found in the GENIUS database. One network-structure was identified for 13 genes, of which six were related to MetS. Each node had an average of 2.0 neighbors. CACNA1D, DIP2C, and GRHL1 had the highest network degrees and were all associated with MetS (Fig. 3a).

Table 2 Annotations for the 35 FDR significant CpGs (adjusted p values < 0.05) for cisplatin-based chemotherapy (CBCT) model
Full size table
Fig. 2
figure 2

Q-Q plots for a cisplatin-based chemotherapy (CBCT) model, methylation β value as the dependent and CBCT as the independent variable, adjusted for smoking, age, testosterone, and cell count. b Metabolic syndrome (MetS) model, MetS as the dependent and methylation β value as the independent variable, adjusted for CBCT, smoking and age

Full size image
Fig. 3
figure 3

Functional gene networks of cisplatin-based chemotherapy (a) and metabolic syndrome (b) related genes reconstructed using the GENIUS tool. Nodes represent genes and edges (arrows) show the directions of the interactions found. Size of nodes is proportional to a gene network degree (number of neighbors of a given gene in the network). A node color intensity represents significance from the differential methylation analysis (high intensity colors represent highly significant genes, adjusted p value for panel (a) and unadjusted p value for panel (b) shown). Nodes marked with black circles represent genes associated with any of the metabolic syndrome trait

Full size image

DNA methylation and risk of developing MetS

We could not identify MetS differentially methylated CpG sites after adjusting for age, smoking, CBCT, and multiple testing (FDR or Bonferroni) (Fig. 2b). We present the top 15 differentially methylated CpG sites (unadjusted p values) (Table 3) of which 11 CpGs were located on an open sea and four on the CpG island shores. None of the multivariate models showed epigenome-wide association study (EWAS) significant results, and their top hits differed (Additional file 1: Figure S5B). In addition, models for the individual MetS components (hypertension, cholesterol, waist circumference, fasting glucose, and triglycerides) did not give EWAS significant associations. There was no overlap between the top 2000 nominally significant CpGs for these five MetS component models and the MetS model (Additional file 1: Figure S6).

Table 3 The 15 CpG sites with lowest unadjusted p values for associations between DNA methylation (DNAm) and metabolic syndrome (MetS)
Full size table

We identified one network-structure for the nine MetS-associated genes found in the GENUS database (Fig. 3b).

Overlap between CBCT and MetS associations

There were no FDR or Bonferroni significantly differentially methylated CpG sites associated with both CBCT and MetS. However, the comparison of the top 2000 CpG sites associated with CBCT and MetS with unadjusted p values < 0.05 for both analyses identified ten common CpG sites (Table 4).

Table 4 Annotations for the nine overlapping CpGs for the cisplatin-based chemotherapy (CBCT) model and metabolic syndrome (MetS) model for the 2000 CpGs with the lowest unadjusted p values < 0.05
Full size table

Pathway enrichment for CBCT-associated CpGs

Genes in approximation to 78 differentially methylated CpG sites (FDR < 0.1) associated with CBCT were analyzed for gene enrichment to provide a functional interpretation of our results. We identified the “PTK6 Regulates RHO GTPases, RAS GTPase, and MAP kinase” Reactome pathway as significantly enriched (adjusted p value = 0.03). For GO biological process, we found the “cellular response to growth hormone stimulus” pathway significantly enriched (p value = 0.005). For GO cellular component, the “L-type voltage-gated calcium channel complex” pathway was significantly enriched (p value = 0.02). We did not find significant pathways for KEGG and GO molecular function.

Differentially methylated regions associated with CBCT and MetS

Using bumphunting, 419 regions (bumps) were identified; however, none were significantly associated with CBCT based on adjusted p values (data not presented). Neither did we identify significant hits when using DMRcate for the two model (data not presented).

We checked if genomic coordinates of the CpGs of interest were present as SNPs in GWAS Central database. We did not find any of the CBCT, MetS, and overlapping CpG sites from Tables 2, 3, and 4, respectively. In the EWAS Atlas database, we found one association with post-obese (cg07677157), and one association with high-saturated fatty acids diet (cg07677157) for CBCT-associated CpGs. From our CpG sites related to MetS, there was one hit cg06500161 (ABCG1), associated with MetS, BMI, and type 2 diabetes in this database. In the overlapping CpGs, we found cg27087650 associated with BMI.

Discussion

In this EWAS, we identified lasting CBCT-related effects in 35 differentially methylated CpG sites across the genome, on average 16 years after treatment. These effects may be attributed to the initial CBCT, and/or to the small amount of platinum persistent in serum post-treatment [5]. We found insulin- and body mass-related genes in proximity to the CBCT-associated hits, supporting our hypothesis that the CBCT-MetS relationship is linked to epigenetics. Interestingly, we found CpG sites in proximity to the gene ABCG1, which has been associated to body mass, triglycerides, HDL-C, atherosclerosis, and type 2 diabetes in EWAS [23,24,25], among our nominally significant MetS CpGs.

The gene closest to the CBCT− top hit, cg07677157, is RPSAP52, a gene linked to type 2 diabetes in genome-wide association studies (GWAS) [26]. The top annotated CBCT gene, CACNA1D, encodes voltage-dependent calcium channels, which regulate insulin release. Polymorphisms in CACNA1D are also associated with type 2 diabetes [27], linked to diastolic and systolic blood pressure [28], and ototoxicity [29]. Other CBCT-associated genes were ACOT7, implicated in the pathophysiology of type 2 diabetes [30] and atherosclerosis [31]; GRHL1, encoding a transcription factor involved in epithelial development and linked to several types of cancer, cardiovascular diseases, and type 1 diabetes; and TOM1L2, linked to body mass and type 2 diabetes in GWAS [32] (Table 2).

We found “PTK6 Regulates RHO GTPases, RAS GTPase, and MAP kinases” which is part of the RAS signal transduction pathway enriched with CBCT-associated methylation. The pathway regulates cell differentiation and plays a role in cell proliferation [33,34,35], which might be a relevant mechanism in relation to the increased risk of second cancer after CBCT in TC survivors [36]. The pathway “cellular response to growth hormone stimulus” is also enriched for CpGs associated to CBCT. The involvement of these oncogenic pathways is consistent with the cytotoxicity of CBCT, which is interesting considering the rather long time window between treatment and DNAm measurements.

Examples of MetS-associated genes include COLEC12, encoding a scavenger receptor involved in several functions associated with host defense; NCF2, for which increased expression has been observed in patients with insulin resistance [37]; and SMG7, playing a role in p53 function in response to DNA damage [38].

The possible relationship between CBCT-induced DNAm and susceptibility for developing MetS was analyzed by exploring the sequence of events separately, as there are no well-established tools for doing high-dimension mediation analysis that fit with our design. The overlap between the two models, including the top 2000 hits, was ten sites (Table 4). These might be spurious findings since associations were nominal. The small overlap might suggest two unrelated mechanisms leading to MetS which was supported by publicly available distinct EWAS (EWAS Atlas). Nonetheless, for both models, the majority of top CpGs was linked to factors pertaining to MetS. Additionally, among the genes that clustered in the network analysis, those with the highest network degree (CACNA1D, DIP2C, and GRHL1 for CBCT, and ABCG1 and NCF2 for MetS) were all associated with MetS. The network analyses suggest that CBCT and MetS-related effects may be mediated synergistically. We speculate that changes in methylation in these clustered genes might affect gene expression, thus increasing the likelihood of developing MetS. These findings support the notion that DNAm may mediate the effect of CBCT on MetS risk later in life. Validation and replication of the top results are needed in an independent cohort. Inclusion of a non-oncological control group, with and without MetS, will further elucidate if the findings are independent of tumor intrinsic factors, and if the MetS hits are related to CBCT.

The survivors in our study who developed MetS have somewhat lower testosterone levels than those without MetS, regardless of CBCT, thus low testosterone may have confounded the results. Previous studies have shown that lower total testosterone level is associated with higher risk of developing MetS [39]. Independently of the model used, testosterone did not markedly alter the significant associations. This indicates that the MetS DNAm association is not attributable to low testosterone levels.

Strengths of the study include a reliable and broadly characterized study population which has been followed for many years. In addition, we have considered the most important confounding factors by matching the groups. Furthermore, the epigenetic analyses have been performed on EPIC BeadChip arrays, which provides state-of-the art tool for epigenome-wide association analyses, covering over 850,000 CpG sites. The EPIC array represents a significant improvement compared to its predecessor, the HM450 array, with increased genome coverage of regulatory regions [40]. Finally, we used curated annotation resources and updated GRCh38/hg38 genome [41].

Some limitations of the present study should also be considered. Even though the total number of TC survivors was large, the sample size of those treated with surgery only and developed MetS was 34. Due to the limited sample size, we chose the minimally adjusted model for MetS. A power issue may thus contribute as to why there were no differentially DNA methylated CpG sites associated with MetS after adjusting for multiple testing. We were not able to distinguish between the effect of the initial CBCT and the effect of platinum residuals [5] as this would have needed a time-series analysis. It is not possible to rule out that some of the differential DNA methylation associated to CBCT could be related to tumor-specific intrinsic factors. However, TC stage markers of DNA methylation including HOXA9, RASSF1A, and SCGB3A1 [42] were not observed, indicating that potential confounding by such factors was unlikely. The possible impact of second primary cancer is negligible due to long time span between sampling and diagnosis.

Conclusions

Our results suggest that CBCT has long-term effects on the epigenome. Although we could not directly link the CBCT effects to the risk of developing MetS, it may still play a role in the development of MetS in TC survivors. This is supported by the observation that the differential DNAm occurs in genes related to MetS. Furthermore, our results contribute to a better understanding of the cellular mechanisms behind the development of MetS in TC survivors treated with CBCT. Although the influence of CBCT on the epigenome is plausible, validation of the observed differential methylation of specific CpGs is necessary. Our findings also indicate that other adverse effects of CBCT, such as ototoxicity, may be mediated by epigenetic changes. These topics could be subject to future studies, also encompassing other cancer forms using CBCT, and risk of second cancer. In terms of clinical perspective, our results may provide early identification of individuals with increased risk for development of MetS.

Methods

Study population and patient assessments

Participants were long-term survivors of unilateral TC diagnosed from 18 to 75 years of age, treated between 1980 and 1994. The original national cohort consisted of 1463 men (Caucasians) who participated in follow-up survey I (SI) at five Norwegian university hospitals during the period 1998–2002. In total, 990 males, younger than 60 years old at SI, were subsequently evaluated with regard to CVD and MetS in a second survey (SII) performed 2007–2008 [15]. A third survey (SIII) was performed in 2015–2016. Overall, 279 participants with MetS data obtained in SII, and frozen whole blood samples from either SI (n = 137), SII (n = 132), or SIII (n = 8) available for DNA analyses, were included in the present study. We included the samples that best fitted the matching criteria (see below). For those who had samples from more than one survey, we chose the DNA obtained at the earliest time point after diagnosis to capture as much of the CBCT related effects as possible.

Data from questionnaires, clinical examinations (including blood pressure and waist circumference measurements), and laboratory tests (including fasting blood glucose and blood lipid measurements) were retrieved from SII. Smoking status was classified from questionnaire data into three groups: never, former, and current smoker. Age was used as a continuous variable in all statistical analyses. All routine blood samples were analyzed at the Oslo University Hospital. Plasma levels of lipids and glucose were measured enzymatically. Serum levels of testosterone were determined using immunoassays. MetS was defined according to the National Cholesterol Education Program expert panel, as the presence of minimum three of the following five criteria: blood pressure ≥ 130/85 mmHg, HDL-cholesterol < 1.0 mmol/L, triglycerides ≥ 1.7 mmol/L, waist circumference > 102 cm, and fasting glucose ≥ 6.5 mmol/L [43, 44]. Epigenomic changes related to MetS were assumed to be present at the sampling time point. Data regarding initial tumor stage, histology, and treatment details were retrieved from medical records. Treatment details according to stage and histology [45], CVD risk, and morbidity data from SI and SII have been published previously [9, 15, 46, 47]. Cancer Registry follow-up showed that 51 survivors acquired a second primary cancer within the cohort follow-up time (SIII). These cancers were diagnosed on average 6 (SD = 3) years after sample donation.

From the 279 TC survivors included, 103 had surgery only (orchiectomy with or without retroperitoneal lymph node dissection for selected cases) due to stage I disease, whereas 176 had undergone both surgery and CBCT (all with metastatic disease) (Fig. 1). From the surgery only (CBCT−) and the treatment (CBCT+) groups, we selected similar fractions of patients with MetS. These four groups were frequency-matched on smoking habits and age at blood sampling, allowing no more than two years difference in mean age. However, complete matching was not possible due to limited number of patients categorized as CBCT+ and MetS−. From this 2 × 2 design, we analyzed the data using a CBCT model and a MetS model, where all cases were included in both models (Fig. 1). Comparing the top hits from the two models enables the identification of CpG sites associated with both CBCT and MetS. This study was approved by Regional Ethical Committee (REC) south east D, reference 2015/1332.

DNA methylation profiling

Genomic DNA was isolated from whole blood using standard chloroform–phenol extraction method. DNA concentration and purity of the DNA samples were analyzed using a NanoDrop ND-1000 (Thermo Fisher Scientific, Waltham, MA, USA). DNA from ten samples was isolated using QIAamp DNA Blood Mini Kit (Qiagen) and the Qiacube (Qiagen) according to manufacturer’s protocol. Of the 279 samples, four samples had only 300 ng (500 ng recommended); however, they showed good array quality in all control steps. Bisulfite conversion of the genomic DNA was done using the EZ DNA Methylation Kit (Zymo) and whole-genome DNAm were analyzed with the Infinium MethylationEPIC Kit (Illumina) according to manufacturer’s recommendations. This array covers 850,000 individual CpGs at CpG islands, RefSeq genes, ENCODE open chromatin, ENCODE transcription factor binding sites, and FANTOM5 enhancers sites. The 279 samples were randomized on three 96-well plates according to the four groups described.

The resulting raw data were analyzed using minfi v.1.20.2 in the R statistical environment v3.3.3 [48]. For details, see Supplementary method information. Two samples were excluded, one due to missing smoking information, and one being an outlier in the principal component analyses (Additional file 1: Figure S1). This resulted in a quality-controlled dataset of 277 samples and 862,400 CpG sites. CpG sites were mapped to the human genome (GRCh37/hg19) using the annotation file provided by the manufacturer (Illumina) [49] and further curated and translated to GRCh38/hg38 [41]. Additional information were retrieved from the UCSC genome browser [50, 51] and GeneCards (www.genecards.org) [52] and Gene2Function (http://www.gene2function.org) [53]. Relative proportion of cell types (B cells, CD4+ T cells, CD8+ T cells, natural killer cells, granulocytes, and monocytes) from the methylation profiles were estimated using the reference-based Houseman method [54, 55].

We deployed GENIUS (GEne Networks Inference Using Signatures) tool to predict local gene networks and key genes for biological functions [56]. The network was constructed using supervised machine learning method to find expression signatures. Input for the gene network was the FDR significant genes from the CBCT and MetS model. The network was visualized with Cytoscope 3.7.1 according to the nodes connectivity and degree [57].

Statistics

The methylation values were transformed to β values (between 0 and 1), representing the intensity of methylation [58]. We used a linear regression model to investigate if cell type composition was associated with CBCT, adjusting for age at blood sampling and smoking habits. We also tested if MetS was associated with testosterone levels, adjusting for age. The results indicated that testosterone and cell type (five out of six cell types, B cells was dropped) composition might be confounders. The robust linear CBCT model with DNA methylation status as the dependent variable was therefore adjusted for age, smoking, cell type, and testosterone. To investigate the associations between DNAm and MetS with the latter as the dependent variable, we used a generalized logistic regression model. The MetS model was adjusted for age, smoking, and CBCT. Models with the best fit were included, and additional regression models tested are available in the Supplementary information (Additional file 1: Figure S4 and Figure S5). Additionally, separate generalized logistic regression models with the five variables underlying the MetS diagnosis (hypertension, cholesterol, waist circumference, fasting glucose, and triglycerides) as dependent variable and with the same covariates were tested. To adjust for multiple testing, Benjamini and Hochberg’s FDR [59] and Bonferroni correction [60] were applied to all models. Differentially methylated CpG sites, identified in the CBCT model and the MetS model, were defined as the intersection of the top 2000 hits with the lowest unadjusted p values.

In order to identify differentially methylated regions (DMR), we used two approaches, bumphunter [61] and dmrcate [62,63,64]. Bumphunter (v.1.20.0) was run with 1000 permutations and the cutoff was set to 0.05, corresponding to 5% difference on the β values on the CBCT model. We ran the DMRcate function (v.1.14.0) with default settings (max gap 1000 nucleotides between two significant probes and DNAm as outcome) on the CBCT model and by flipping the function around with the dichotomous variable as outcome on the CBCT and MetS models.

Gene enrichment analysis and CpG characterization

CpG sites passing an FDR adjustment (p value < 0.1) were used to assess pathways enriched for differential DNAm. We employed Enrichr (http://amp.pharm.mssm.edu/Enrichr) [65, 66] a platform for KEGG [67], Reactome 2016 (v.62), and Gene Ontology (GO) 2018 (biological process, molecular function, and cellular component) pathways analysis [68, 69]. We performed an unweighted analysis, and reported p values are based on Fisher’s exact test.

We employed GWAS Central database (www.gwascentral.org) to evaluate if any of CpG sites of interest were previously reported as known SNPs [70]. We further scanned for associations between CpG sites of interest and known epigenome-wide associations from literature studies. We used the EWAS Atlas resource (https://bigd.big.ac.cn/ewas/index) [71]. This database features a large number of high-quality, manually curated, EWAS associations.